An order whole-process evidence chain construction system and method for heavy truck maintenance

Abstract

The application discloses an order whole-process evidence chain construction system and method for heavy truck maintenance, belongs to the technical field of computers, and through the introduction of a dual-role digital twin engine, real-time collection and alignment of physical operation and simulation data, generation of difference evolution path, and comparison of measured and simulated states after maintenance to obtain global difference, and then combined with the model training evidence of federated learning, the evidence graph of the maintenance order is constructed and stored. Thus, the problems of independent storage of multi-source data, lack of logical confirmation relationship in the heavy truck maintenance scene are solved, the precise verification of the maintenance effect, the clear traceability of the model parameter source, the clear proof of the maintenance abnormal cause and effect relationship, and the comprehensive confirmation of the multi-source evidence are realized, and the integrity of the maintenance evidence chain is improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a system and method for constructing a chain of evidence for the entire order process in heavy truck repair. Background Technology

[0002] In heavy-duty truck repair, the complete documentation of repair orders is crucial for warranty assessment, liability tracing, and dispute resolution. With the increasing intelligence of vehicles, the amount of multi-source data generated during repairs is growing richer. How to integrate this scattered data into a complete chain of evidence with legal probative value has become a widely discussed technical direction within the industry.

[0003] Existing maintenance evidence construction methods mainly rely on multimodal acquisition devices to record raw information such as video and sensor readings. However, the data from multiple sources are stored independently, lacking logical corroboration between them.

[0004] How to construct an evidence chain that can verify repair effectiveness, trace parameter sources, prove causality, and corroborate multiple sources of evidence is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] This application provides a system and method for constructing a complete evidence chain for heavy truck repair orders, the technical solution of which is as follows: On the one hand, a method for constructing a complete evidence chain for heavy truck repair orders is provided, the method comprising: In response to the initiation of a maintenance order for the target heavy truck, the dual-role digital twin engine of the target heavy truck is initialized. The dual-role digital twin engine includes a prediction instance and a benchmark instance. The prediction instance is used to simulate and predict the simulation state based on the maintenance operation, and the benchmark instance is used to solidify the standard maintenance process and standard performance benchmark. During the maintenance process, physical operation evidence stream and twin simulation evidence stream are collected in real time. Based on the standard maintenance process in the benchmark instance, each physical operation evidence unit in the physical operation evidence stream and the corresponding twin simulation evidence unit in the twin simulation evidence stream are time-aligned and semantically mapped to obtain multiple related evidence pairs and the initial difference evolution path corresponding to each related evidence pair. After maintenance is completed, the measured performance data of the target heavy truck is obtained as the measured state. The measured state is compared with the simulation state output by the predicted instance after all maintenance operations are completed to obtain the global difference degree. Based on the multiple related evidence pairs and the initial difference evolution path corresponding to each related evidence pair, a difference evidence node containing the global difference degree, the difference evolution path chain and the abnormal link identifier is generated. Obtain the model training evidence chain generated by each repair station during the federated learning process. The model training evidence chain includes local model parameters and corresponding generation proofs. Based on the model training evidence chain, the multiple associated evidence pairs, and the difference evidence nodes, construct the evidence graph of the repair order, and store the evidence graph using a triple hash chain. Attached Figure Description

[0006] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0007] Figure 1 This is a schematic diagram of the implementation environment for a method for constructing a chain of evidence for the entire order process in heavy-duty truck repair, as provided in an embodiment of this application. Figure 2 This is a flowchart of a method for constructing a complete evidence chain for heavy truck repair orders, provided in an embodiment of this application. Figure 3 This is a flowchart of another method for constructing a complete evidence chain for heavy truck repair orders, provided in an embodiment of this application. Figure 4 This is a schematic diagram of a system for constructing a chain of evidence for the entire order process in heavy-duty truck repair, as provided in an embodiment of this application. Figure 5 This is a schematic diagram of the structure of a server provided in an embodiment of this application. Detailed Implementation

[0008] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0009] In this application, the terms "first," "second," etc., are used to distinguish identical or similar items with essentially the same function. It should be understood that there is no logical or temporal dependency between "first," "second," and "nth," nor are there any restrictions on quantity or execution order.

[0010] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0011] Figure 1This is a schematic diagram illustrating the implementation environment of a method for constructing a complete evidence chain for heavy-duty truck repair orders, as provided in an embodiment of this application. (See attached diagram.) Figure 1 The implementation environment may include node 110 and system 140.

[0012] Node 110 is connected to system 140 via a wireless or wired network. Optionally, node 110 can be a smartphone, tablet, laptop, desktop computer, etc., but is not limited to these. Node 110 has an application installed and running that supports the construction of a complete chain of evidence for the entire order process for heavy truck repair.

[0013] System 140 is a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms. System 140 can provide background services for applications running on node 110.

[0014] In heavy truck repair scenarios, the complete documentation of repair orders is crucial for warranty assessment, liability tracing, and dispute resolution. Traditional repair evidence construction methods primarily rely on multimodal data acquisition devices to record raw information, resulting in independent storage of multiple data sources with a lack of logical corroboration, making it difficult to construct a complete chain of evidence with judicial probative value.

[0015] To address this, this application proposes a method for constructing a complete evidence chain for heavy-duty truck repair orders, see [link to relevant documentation]. Figure 2 The methods include: 201. In response to the initiation of a maintenance order for the target heavy truck, initialize the dual-role digital twin engine of the target heavy truck. The dual-role digital twin engine includes a predictive instance and a baseline instance. The predictive instance is used to simulate and predict the simulation state based on the maintenance operation, and the baseline instance is used to solidify the standard maintenance process and standard performance benchmark.

[0016] 202. During the maintenance process, the physical operation evidence stream and the twin simulation evidence stream are collected in real time. Based on the standard maintenance process in the benchmark instance, the physical operation evidence unit in the physical operation evidence stream and the corresponding twin simulation evidence unit in the twin simulation evidence stream are time-aligned and semantically mapped to obtain multiple related evidence pairs and the initial difference evolution path corresponding to each related evidence pair.

[0017] 203. After the maintenance is completed, the measured performance data of the target heavy truck is obtained as the measured state. The measured state is compared with the simulation state output by the predicted instance after all maintenance operations are completed to obtain the global difference degree. Based on multiple related evidence pairs and the initial difference evolution path corresponding to each related evidence pair, difference evidence nodes containing global difference degree, difference evolution path chain and abnormal link identifier are generated.

[0018] 204. Obtain the model training evidence chain generated by each repair station during the federated learning process. The model training evidence chain includes local model parameters and corresponding generation proofs. Based on the model training evidence chain, multiple related evidence pairs and differential evidence nodes, construct the evidence graph of the repair order and store the evidence graph using a triple hash chain.

[0019] For ease of understanding, the following explains some key terms in this embodiment: A dual-role digital twin engine is an integrated digital modeling system designed to simultaneously perform predictive simulation and standard reference functions. By digitally modeling physical heavy-duty trucks, this engine can simulate the operating status and maintenance processes of heavy-duty trucks in a virtual environment and provide standardized process benchmarks.

[0020] The predictive instance is a dynamic component in a dual-role digital twin engine, configured to simulate the state changes of a heavy-duty truck during maintenance in real time, based on actual or planned maintenance operation instructions. By receiving input parameters and operation sequences, the predictive instance can output predicted simulation states of the heavy-duty truck at different maintenance stages, reflecting the impact of maintenance operations on the truck's performance.

[0021] The benchmark instance is a static component in the dual-role digital twin engine, used to solidify standard maintenance procedures and performance benchmarks for specific heavy-duty truck models. This instance serves as an authoritative reference, providing an unchanging basis for the standardized assessment of maintenance operations and the calculation of performance differences.

[0022] Physical operation evidence stream refers to the sequence of actual physical operation data collected in real time during heavy truck maintenance through various sensors and recording devices. This data can include tool usage records, operator actions, environmental parameters, etc., directly reflecting the real situation at the maintenance site.

[0023] The twin simulation evidence stream refers to the real-time sequence of heavy-duty truck simulation state data generated by the predictive instance of the digital twin engine during simulated maintenance operations. This data stream corresponds to the physical operation evidence stream in time and is used to reflect the state changes of heavy-duty trucks under ideal or predicted conditions.

[0024] A correlated evidence pair refers to a logical combination formed by matching a physical operation evidence unit in the physical operation evidence stream with a corresponding twin simulation evidence unit in the twin simulation evidence stream through temporal alignment and semantic mapping. This combination represents the correspondence between the actual execution of a specific maintenance operation and the digital twin prediction.

[0025] The initial difference evolution path refers to a trajectory generated for each associated evidence pair, based on the initial deviation between the simulation state reflected by the twin simulation evidence unit and the standard performance benchmark in the benchmark instance, reflecting the potential difference of the operation link changing over time.

[0026] Global variance refers to the degree of deviation between the overall maintenance effect and the expected effect, which is quantified by comparing the measured performance data (measured state) of the heavy truck with the simulated state output by the digital twin engine prediction instance after the heavy truck maintenance is completed.

[0027] A difference evidence node is a structured data unit generated to encapsulate key difference information during the maintenance process. This node includes the global difference level, the difference evolution path chain throughout the entire maintenance process, and identifiers of anomalous points identified during the analysis.

[0028] The difference evolution path chain refers to a complete trajectory that reflects the dynamic changes in the performance or condition differences of heavy trucks throughout the entire maintenance cycle, formed by connecting all relevant evidence pairs during the maintenance process and their corresponding initial difference evolution paths in accordance with the maintenance time sequence.

[0029] Anomaly identification refers to a key maintenance operation or state deviation point within the discrepancy evidence node that explicitly identifies the cause of or propagation of the global discrepancy. This identification helps to quickly pinpoint the root cause of problems during the maintenance process.

[0030] The model training evidence chain refers to the data chain generated by each maintenance station during the federated learning process to ensure the transparency and traceability of the model training process. This evidence chain typically includes local model parameters, training data hashes, training process hyperparameters, and corresponding generation proofs.

[0031] An evidence graph is a graphical representation of the entire process of a repair order. This graph uses nodes to represent various evidence units (such as physical operations, simulation states, discrepancy information, and model training parameters) and edges to represent the logical relationships between these units (such as corroboration, tracing, and causality), thus constructing a comprehensive and traceable evidence network.

[0032] Triple hash chain evidence storage refers to a multi-layered blockchain evidence storage mechanism. It constructs a physical operation evidence chain, a twin simulation evidence chain, and a differential evidence chain, links their respective root hashes, and writes the overall hash into the blockchain to ensure the integrity, immutability, and traceability of repair evidence.

[0033] The embodiments of this application provide a method for constructing a complete chain of evidence for heavy truck repair orders.

[0034] When a heavy-duty truck repair order is initiated, a dual-role digital twin engine needs to be initialized for the target heavy-duty truck. This process may include: after receiving the repair order initiation signal, the system loads a general heavy-duty truck digital twin model template from a pre-set digital twin model library. Based on the specific model, configuration information, and historical repair records of the target heavy-duty truck, this general template is parameterized to accurately represent the heavy-duty truck currently undergoing repair. On this basis, the configured digital twin model is instantiated into two independent but related components: a prediction instance and a baseline instance. The prediction instance is configured to receive repair operation commands in real time and simulate the state changes of the heavy-duty truck during the repair process based on these commands, outputting a predicted simulation state. For example, when simulating the replacement of a component, the prediction instance calculates and updates the relevant performance parameters of the heavy-duty truck based on the component's physical characteristics and installation process. The baseline instance is set to store and solidify the standard repair process flow and corresponding standard performance benchmarks for this model of heavy-duty truck, serving as a static reference for subsequent comparisons. For example, the baseline instance may contain the standard replacement time, standard installation torque, and standard performance indicators to be achieved after replacement for a certain component.

[0035] During maintenance operations, the system continuously collects evidence data of actual maintenance operations from the physical world, forming a physical operation evidence stream. For example, sensors installed on maintenance tools can record each use of the tool, its duration, and the force applied. Alternatively, cameras can record the operator's action sequence and convert it into structured data. Simultaneously, the predictive instance of the digital twin engine generates corresponding simulation state data in real time based on the received maintenance operation instructions, forming a twin simulation evidence stream. For example, when the predictive instance simulates the disassembly process of a component, it generates virtual state parameters of the component before and after disassembly. To establish the correlation between physical operations and simulation states, the system performs an initial match between each operation unit in the physical operation evidence stream and a simulation unit in the twin simulation evidence stream that is temporally close. For example, if the timestamp of a physical operation unit and the timestamp of a twin simulation unit are within a preset time window, they are considered potentially related. By analyzing the content of these matched units, such as operation type and involved components, and referring to the standard maintenance processes stored in the baseline instance, semantic association is confirmed, thereby forming multiple pairs of related evidence. For example, if the physical operation unit describes "changing the oil filter," while the twin simulation unit simulates "the system state after changing the oil filter," they can be semantically mapped. For each pair of related evidence, the system generates an initial difference evolution path based on the preliminary difference between the simulation state reflected by the twin simulation evidence unit and the standard performance benchmark in the baseline instance, reflecting the potential deviation of that operation. For example, if the simulation state shows that the oil pressure is slightly lower than the standard benchmark, this deviation will be recorded as part of the path.

[0036] Once all maintenance work on the heavy truck is completed, the system retrieves its actual operating performance data as the measured state. For example, it reads the engine's actual operating parameters, fuel efficiency, and emissions data through the On-Board Diagnostics (OBD) system. This measured state is then sent to a comparison module for comprehensive comparison with the simulation state output by the digital twin engine prediction instance after all maintenance operations are completed. This comparison quantifies the overall deviation between the actual maintenance effect and the expected simulation effect, thus obtaining a global difference degree. For example, if the measured fuel efficiency is lower than the simulation prediction value, a negative difference degree will be generated. Based on this, the system integrates multiple related evidence pairs generated during the maintenance process and their corresponding initial difference evolution paths, connecting these paths in chronological order to form a difference evolution path chain that runs through the entire maintenance process. Combining the global difference degree, the system generates a difference evidence node. This difference evidence node is designed to include the global difference degree, the complete difference evolution path chain, and anomaly markers identified during the maintenance process. For example, if the simulation state of a certain operation deviates from the standard benchmark, and this deviation is related to the global difference degree, then that operation will be marked as an anomalous operation.

[0037] To ensure the transparency of data sources and model training during the maintenance process, the system obtains the model training evidence chains generated by each maintenance station participating in federated learning. These evidence chains typically include model parameters trained locally at each maintenance station, along with proof of the parameter generation process, such as hash values ​​of the model parameters, fingerprints of the training dataset, and training hyperparameters. The system utilizes the obtained model training evidence chains, multiple related evidence pairs generated during the maintenance process, and previously generated discrepancy evidence nodes to construct a comprehensive maintenance order evidence graph. This graph graphically represents the complex relationships between maintenance operations, simulation predictions, discrepancy analysis, and model training. For example, physical operation evidence nodes can be connected to twin simulation evidence nodes through a "corroboration" relationship, discrepancy evidence nodes can be connected to physical operation nodes that caused the anomalies through a "causal" relationship, and model training evidence nodes can be connected to physical operation nodes that used their training data through a "source" relationship. To ensure the integrity, immutability, and traceability of the evidence graph, the system uses a triple hash chain for storage. This could include organizing the physical operational evidence, twin simulation evidence, and discrepancy evidence in the graph into separate hash chains, linking their respective root hashes, and writing the overall hash into the blockchain, thereby providing strong judicial proof.

[0038] This application introduces a dual-role digital twin engine to collect and align physical operation and simulation data in real time, generating differential evolution paths. After maintenance, it compares the measured and simulated states to obtain global differences. Then, combined with federated learning model training evidence, it constructs and stores an evidence graph of maintenance orders. This solves the problems of independent storage of multi-source data and lack of logical corroboration in heavy truck maintenance scenarios. It achieves accurate verification of maintenance effects, clear traceability of model parameter sources, explicit proof of causal relationships in maintenance anomalies, and comprehensive corroboration of multi-source evidence, thus improving the integrity and judicial probative value of the maintenance evidence chain.

[0039] In some of the embodiments described above in this application, an initial dual-role digital twin engine is proposed to initiate maintenance orders and set up predictive instances and baseline instances. However, in its implementation, there are shortcomings in ensuring the consistency of the initial state and establishing an effective difference calculation mechanism to support the subsequent construction of the evidence chain. Specifically, the acquisition of the base model is inaccurate, the instantiation process lacks standardized references, and the initial state configuration is disconnected from the baseline, resulting in unreliable subsequent difference calculations and affecting the integrity and credibility of the evidence chain.

[0040] In this regard, this application further proposes a specific method for initializing the dual-role digital twin engine of the target heavy truck, see [link to relevant documentation]. Figure 3 The method includes: 301. Obtain the vehicle identification number of the target heavy truck.

[0041] The vehicle identification number (VIN) is a unique identifier for a vehicle, which can be used to accurately identify the vehicle's model, configuration, and other information.

[0042] 302. Based on the vehicle identification code, the system obtains a basic digital twin model from the cloud that matches the specific target heavy truck.

[0043] The basic digital twin model is a digital description of the heavy-duty truck formed during the design and manufacturing stages. It is comprehensive, typically including the vehicle's structural parameters, performance parameters, and standard maintenance procedures. For example, the vehicle identification number (VIN) can be automatically read via the on-board diagnostic (OBD) system interface or obtained by scanning the QR code / barcode on the vehicle's nameplate. The cloud service then queries a pre-established vehicle database based on the received VIN and returns the corresponding basic digital twin model data package. Alternatively, maintenance personnel can input the VIN via a handheld terminal, and the system will send a request to the manufacturer or a third-party data platform to obtain a basic digital twin model containing the vehicle's factory configuration, historical maintenance records, and standard process parameters for the corresponding batch. This step ensures that the subsequent construction of digital twin instances is based on accurate and personalized vehicle data, avoiding initialization deviations caused by model mismatches.

[0044] 303. Instantiate the basic digital twin model into the prediction instance and the baseline instance respectively.

[0045] During instantiation, the predictive instance inherits only the structural and performance parameters from the base digital twin model, and its primary responsibility is to simulate and predict the simulation state based on maintenance operations. The baseline instance, on the other hand, inherits the structural and performance parameters, as well as the standard maintenance process, from the base digital twin model, and locks the standard maintenance process and standard performance benchmark in the baseline instance as static references. For example, the system can create two independent digital twin objects in memory. The predictive instance object copies only the structural and performance attributes from the base model, while the baseline instance object copies all attributes and marks its standard maintenance process and standard performance benchmark as unmodifiable or read-only to ensure they remain unchanged throughout the maintenance cycle. Another implementation approach is to define a predictive instance class and a baseline instance class using the inheritance mechanism in object-oriented programming (OOP), both inheriting from the base digital twin model class. The baseline instance class additionally includes logic for locking the standard process and performance, ensuring they remain unchanged throughout the maintenance cycle, thus providing a stable and unchanging reference benchmark for dynamic comparisons during the maintenance process.

[0046] 304. Based on the standard performance benchmark locked in the benchmark instance, configure the initial state parameters for the prediction instance.

[0047] This configuration aims to ensure that the initial operating state of the predictive instance is consistent with the standard performance benchmark represented by the baseline instance before maintenance operations begin, thereby eliminating the impact of initial deviations on subsequent difference calculations.

[0048] 305. Establish a difference calculation interface between the predicted instance and the baseline instance. This difference calculation interface is used to calculate in real time the offset of the simulation state output by the predicted instance relative to the standard performance benchmark in the baseline instance during the maintenance process, and use this offset as the input of the initial difference evolution path.

[0049] For example, initial state parameter configuration can be accomplished by directly assigning the various metrics of the standard performance benchmark in the baseline instance to the corresponding state variables of the prediction instance. The difference calculation interface can be a standalone software module that periodically retrieves the current simulation state from the prediction instance and compares it numerically with the fixed standard performance benchmark in the baseline instance to calculate a multi-dimensional offset vector. Alternatively, the initial state parameter configuration of the prediction instance can be based on a rule engine, automatically generating the initial runtime configuration of the prediction instance according to the standard performance benchmark defined in the baseline instance. The difference calculation interface can be an API that allows external systems to call it to obtain real-time difference data. Internally, this interface is implemented as Euclidean distance calculation of state vectors, percentage deviation calculation, or weighted difference calculation based on domain-specific knowledge.

[0050] Through the above technical solution, this application ensures that the initialization process of the dual-role digital twin engine is highly accurate and standardized when a repair order is initiated. Obtaining an accurate basic digital twin model from the cloud using the vehicle identification code avoids initialization deviations caused by model mismatch, providing a reliable data foundation for subsequent simulation and comparison. Instantiating the basic model into prediction and benchmark instances, and locking the standard repair process and standard performance benchmark of the benchmark instance as static references, effectively separates dynamic prediction from static standards, ensuring the stability and authority of the comparison benchmark. Based on the locked standard performance benchmark, initial state parameters are configured for the prediction instance, and a difference calculation interface is established, enabling the prediction instance to start in an initial state consistent with the standard and to calculate the offset between the simulation state and the standard benchmark in real time and accurately during the repair process. These offsets serve as inputs to the initial difference evolution path, providing real-time, reliable, and traceable difference data for subsequently constructing the entire evidence chain of the repair order, greatly enhancing the integrity and credibility of the evidence chain, thereby solving problems such as inaccurate acquisition of the basic model, lack of standardized references in the instantiation process, and disconnection between the initial state configuration and the benchmark.

[0051] In some of the solutions described above in this application, initial state parameters are configured for the predicted instance to ensure that the predicted instance is in a state consistent with the standard performance benchmark of the baseline instance before the maintenance operation begins. However, in its implementation, since the standard performance benchmark contains complex parameters in multiple dimensions, if it is not accurately extracted and mapped, mismatch or omission between the initialization fields and parameter items may occur, and the initial value setting may be inaccurate, thereby affecting the consistency of the initial running state of the predicted instance, and thus reducing the reliability of subsequent maintenance simulation and the accuracy of difference calculation.

[0052] In response, this application further proposes a method for configuring initial state parameters for the prediction instance based on a standard performance benchmark locked in the benchmark instance. The method includes: Extract benchmark performance parameters from multiple dimensions from this standard performance benchmark.

[0053] Get the multiple initialization fields defined by the initialization interface of the prediction instance, each of which corresponds to a state variable in the prediction instance.

[0054] The benchmark performance parameters of the multiple dimensions are mapped to the corresponding initialization fields in the prediction instance, and each initialization field is assigned an initial value that matches the benchmark performance parameter.

[0055] Based on this initial value, the initial operating state of the predicted instance is synchronized, so that the predicted instance is in a state consistent with the standard performance benchmark of the baseline instance before the maintenance operation begins.

[0056] Specifically, extracting multi-dimensional benchmark performance parameters from the standard performance benchmark refers to identifying and obtaining the specific parameters constituting its performance characteristics from the standard performance benchmark fixed in the benchmark instance. These parameters may cover multiple aspects of heavy truck operation, such as engine speed, oil pressure, coolant temperature, brake system pressure, tire pressure, and battery voltage. These parameters can be extracted in various ways. For example, predefined parsing rules can be used to accurately identify and extract the name, unit, and standard value of each performance parameter from structured data formats (such as XML, JSON, or database records). Alternatively, semantic analysis techniques combined with domain ontology knowledge can be used to automatically identify and extract key performance indicators from the standard performance benchmark description text. This step aims to ensure that all necessary performance benchmark data is comprehensively and accurately obtained from the static reference, providing a complete and reliable data foundation for subsequent parameter configuration.

[0057] This step involves obtaining multiple initialization fields defined by the initialization interface of the prediction instance. Each initialization field corresponds to a state variable within the prediction instance, specifying which external parameters the prediction instance can accept to set its internal state during startup or reset. These initialization fields are the configuration entry points exposed by the prediction instance, and each field is directly associated with a specific state variable within the prediction instance (such as a physical quantity or logical flag in the simulation model). For example, the prediction instance might provide initialization fields such as "engine_rpm_initial" and "oil_pressure_start". These fields can be obtained by consulting the prediction instance's API documentation or Software Development Kit (SDK) to clarify their definitions and expected data types. Alternatively, in a programming environment that supports reflection, the list of configurable initialization parameters for the prediction instance can be dynamically discovered through program introspection. This step aims to provide clear mapping targets for baseline performance parameter items, ensuring the accuracy and validity of subsequent parameter assignments.

[0058] Mapping the multiple benchmark performance parameters to their corresponding initialization fields in the prediction instance, and assigning each initialization field an initial value matching the benchmark performance parameter, involves establishing a one-to-one correspondence between the extracted benchmark performance parameters and the prediction instance's initialization fields, and assigning the benchmark parameter values ​​as initial values ​​to the corresponding fields. For example, the "standard engine speed" parameter value extracted from the standard performance benchmark is assigned to the "engine_rpm_initial" initialization field of the prediction instance. This mapping can be implemented using a pre-defined mapping table or configuration file, which explicitly specifies the correspondence between benchmark parameters and initialization fields. Alternatively, an automated matching algorithm based on naming conventions or semantic similarity can be used to find the best match between parameter names and field names. This step is crucial for parameter value transfer, ensuring that each state variable of the prediction instance receives an accurate initial value consistent with its standard performance benchmark.

[0059] Synchronizing the initial running state of the prediction instance based on these initial values, ensuring that the prediction instance is in a state consistent with the standard performance benchmark of the baseline instance before the maintenance operation begins, means activating the prediction instance after all initialization fields have been assigned matching initial values, allowing its internal simulation model to start running or enter a ready state based on these initial values. For example, the prediction instance's simulation engine will load these initial values ​​and may perform an internal state calculation or iteration to ensure that all relevant internal variables reach a stable state consistent with the initial values. This can be achieved by directly calling the prediction instance's start or reset method and passing in a configuration object containing all initial values. Alternatively, after receiving the initial values, the prediction instance will perform a brief "warm-up" or "stabilization" process to ensure that its simulation state fully converges to an ideal state consistent with the standard performance benchmark of the baseline instance before the maintenance operation begins. This step ensures the reliability of the prediction instance's simulation starting point, laying a solid foundation for subsequent real-time simulation and difference calculations.

[0060] Through the above technical solution, this application can systematically ensure the accurate configuration of the initial state parameters of the prediction instance. Extracting benchmark performance parameters from multiple dimensions from the standard performance benchmark ensures comprehensive parameter coverage, avoids omissions of key dimensions, and provides a complete data foundation for subsequent mapping. Obtaining multiple initialization fields defined by the prediction instance's initialization interface, each corresponding to a state variable in the prediction instance, clearly defines the specific interface of the state variables, providing a clear mapping target for the parameter items and effectively preventing initialization deviations caused by ambiguous interface definitions. Mapping the benchmark performance parameters of multiple dimensions to the corresponding initialization fields in the prediction instance, and assigning each initialization field an initial value matching the benchmark performance parameter item, ensures that each state variable obtains an accurate initial value through dimension-based mapping, thereby eliminating the risk of parameter item and field mismatch. Synchronizing the initial running state of the prediction instance based on the initial values ​​ensures that the prediction instance is in a state consistent with the standard performance benchmark of the benchmark instance before the maintenance operation begins. This completes the state synchronization operation, ensuring that the simulation starting point is completely consistent, and providing a reliable benchmark for real-time simulation during subsequent maintenance processes. Overall, this scheme improves the consistency of the initial running state of the predicted instances, thereby enhancing the reliability of subsequent maintenance simulations and the accuracy of difference calculations, and providing a solid and reliable simulation foundation for building a full-process evidence chain for heavy truck maintenance orders.

[0061] In some of the solutions mentioned above in this application, temporal alignment and semantic mapping are proposed to associate physical operation evidence and twin simulation evidence. However, in the process of implementation, how to ensure the accuracy of alignment and the reliability of mapping to avoid the breakage of the evidence chain or incorrect association is a technical problem that needs to be solved.

[0062] To address this, this application further proposes a method for performing temporal alignment and semantic mapping on each physical operation evidence unit in the physical operation evidence stream and the corresponding twin simulation evidence unit in the twin simulation evidence stream, based on the standard maintenance process in the benchmark instance, to obtain multiple associated evidence pairs and the initial difference evolution path corresponding to each associated evidence pair. See [link to relevant documentation] Figure 4 The method includes the following steps.

[0063] 401. Extract the first timestamp of each physical operation evidence unit in the physical operation evidence stream, and extract the second timestamp of each twin simulation evidence unit in the twin simulation evidence stream.

[0064] The first timestamp of the physical operation evidence unit refers to the precise point in time when the actual maintenance operation occurred or was recorded, such as the start or end time of the operation recorded by sensors, cameras, or manual input. The second timestamp of the twin simulation evidence unit refers to the precise point in time when the digital twin system simulates the operation or its result, such as the timestamp of the simulation model's output state. These timestamps form the basis for time-series alignment, ensuring the comparability of events in the physical and digital worlds across time.

[0065] 402. Based on the time difference between the first timestamp and the second timestamp and the standard operation sequence specified by the standard maintenance process in the benchmark instance, each physical operation evidence unit is paired with a twin simulation evidence unit that satisfies the timing constraints to obtain multiple candidate evidence pairs.

[0066] The timing constraint refers to the restriction imposed on the temporal relationship between physical operation evidence units and twin simulation evidence units when pairing them. This can be defined through a preset time window, for example, requiring the time difference between physical and simulation operations to be within an acceptable range. Alternatively, it can be flexibly set through dynamically adjusted time tolerances based on the complexity of the operation or environmental factors. Furthermore, it can be based on the standard operation sequence defined in standard maintenance procedures, such as requiring one operation to follow another and be completed within a specific timeframe, to determine whether the timing constraint is met. Only evidence units that meet these timing constraints will be considered for pairing, forming candidate evidence pairs, thereby initially screening evidence with a reasonable temporal correspondence.

[0067] 403. For each candidate evidence pair, analyze the operation action tags contained in the physical operation evidence unit, and extract the corresponding simulation operation tags from the twin simulation evidence unit.

[0068] Among them, operation action labels are semantic descriptions or identifiers of actual maintenance actions in the physical operation evidence unit, such as "replacing brake pads" and "tightening bolts." These labels can be extracted from maintenance record text using natural language processing technology or annotated using a pre-set coding system. Simulation operation labels are semantic descriptions or identifiers of operations performed or simulated by the simulation model in the twin simulation evidence unit, corresponding to the actual operation action labels. These labels are usually generated by the simulation system based on its internal logic and model definition.

[0069] 404. Perform consistency verification between the operation action label and the simulation operation label, and at the same time, compare the simulation process parameters contained in the twin simulation evidence unit with the standard process parameters corresponding to the operation action label in the benchmark instance item by item.

[0070] The consistency check involves comparing the operation action label and the simulation operation label to confirm whether they semantically represent the same maintenance operation. This can be achieved in various ways, such as direct string matching, semantic similarity calculation based on ontology or knowledge graphs, or classification using machine learning models. The purpose of the check is to ensure that the physical operation and the simulation operation correspond in content. The simulation process parameter refers to the specific parameter values ​​recorded in the twin simulation evidence unit that are used or generated during the simulation process, such as torque, temperature, pressure, and time. The standard process parameter refers to the standard or ideal range or target value of process parameters specified for a specific operation action label in the benchmark instance. Item-by-item comparison involves a detailed comparison of the simulation process parameters with the corresponding standard process parameters, for example, determining whether the simulation parameters fall within the allowable range of the standard parameters, or calculating their deviation from the standard values. This comparison aims to evaluate the standardization and accuracy of the simulation process.

[0071] 405. Based on the consistency verification results and the item-by-item comparison results, determine the mapping confidence between the physical operation evidence unit and the twin simulation evidence unit in the candidate evidence pair, and take the candidate evidence pair with the mapping confidence exceeding the preset threshold as the associated evidence pair.

[0072] The mapping confidence score is a quantitative indicator used to measure the reliability of the correspondence between physical operation evidence units and twin simulation evidence units. It can be calculated comprehensively based on factors such as the success or failure of consistency verification, semantic similarity score, and the degree of deviation between simulation process parameters and standard process parameters, for example, through weighted summation or decision tree models. The preset threshold is a pre-defined lower limit of confidence; only when the mapping confidence score is higher than this threshold will the candidate evidence pair be confirmed as a reliable associated evidence pair, thereby ensuring the quality of evidence in subsequent analyses.

[0073] 406. Retrieve the standard performance benchmark associated with the operation action label in the associated evidence pair from the benchmark instance, compare the simulation state contained in the twin simulation evidence unit in the associated evidence pair with the standard performance benchmark dimension by dimension, calculate the deviation vector of the simulation state relative to the standard performance benchmark, and generate the initial difference evolution path corresponding to the associated evidence pair based on the deviation vector and the position of the associated evidence pair in the maintenance timeline.

[0074] The standard performance benchmark refers to the ideal performance index or state that a vehicle component or system should achieve after a specific maintenance operation or sequence of operations within the benchmark instance. For example, the braking distance and braking response time of the braking system after replacing brake pads. These benchmarks serve as reference standards for evaluating maintenance quality and effectiveness. The simulation state refers to the system or component state presented by the digital twin model after the simulation operation, as recorded in the twin simulation evidence unit. The deviation vector is a multi-dimensional numerical representation, with each dimension corresponding to a performance parameter of the standard performance benchmark and recording the difference between the simulation state and the standard performance benchmark in that parameter. For example, if the standard performance benchmark includes braking distance and braking response time, the deviation vector will include the difference between the simulated braking distance and the standard braking distance, as well as the difference between the simulated braking response time and the standard braking response time. The initial difference evolution path is a preliminary, structured data sequence that records the deviation of the simulation state from the standard performance benchmark at specific temporal points in the maintenance process. This path not only includes the deviation vector but also information on the time point or operational step in which the deviation occurred, providing foundational data for subsequent difference analysis and tracing.

[0075] Through the above technical solution, this application addresses the problem of ensuring the accuracy of alignment and the reliability of mapping during temporal alignment and semantic mapping to avoid broken evidence chains or erroneous associations. By accurately comparing the timestamps of physical operation evidence units and twin simulation evidence units, and combining this with the temporal constraints specified in standard maintenance procedures, candidate evidence pairs that are highly correlated in the time dimension can be initially screened, improving the accuracy of temporal alignment. For these candidate evidence pairs, consistency verification is performed on operation action labels and simulation operation labels, and simulation process parameters are compared item by item with standard process parameters, achieving dual verification at the semantic level. Based on this, the mapping confidence is calculated, thereby reliably identifying the truly corresponding related evidence pairs and effectively avoiding erroneous associations. By comparing the simulation state with the standard performance benchmark dimension by dimension, and combining this with the position of the related evidence pairs in the maintenance time sequence, an initial difference evolution path is generated, providing a structured and quantifiable basis for subsequent difference analysis and traceability, ensuring the integrity and traceability of the evidence chain. Overall, this solution, through a multi-dimensional and multi-level verification mechanism, greatly enhances the accuracy and reliability of the correlation between physical operations and digital twin simulation, laying a solid foundation for building a full-process evidence chain for heavy truck maintenance orders.

[0076] In some of the solutions described above in this application, it is proposed to pair physical operation evidence units with twin simulation evidence units to achieve time alignment. However, in the process of implementation, due to the complexity of maintenance operation timing and the difference in timestamps, the pairing may not be accurate enough, resulting in inaccurate generation of subsequent related evidence pairs and affecting the reliability of the evidence chain.

[0077] In response, this application further proposes a method for pairing each physical operation evidence unit with a twin simulation evidence unit that satisfies timing constraints, based on the time difference between the first and second timestamps and the standard operation sequence specified by the standard maintenance process in the benchmark instance, to obtain multiple candidate evidence pairs. Specifically, this method includes: This paper extracts multiple standard operation nodes, their sequence, and standard time intervals from a benchmark instance to construct a standard operation timing graph. This graph is a structured data model representing the sequence, dependencies, and expected durations or time intervals of various standard operations during heavy-duty truck maintenance. Its purpose is to provide an authoritative and standardized benchmark for evaluating the timing compliance of actual maintenance operations and to provide a basis for the precise alignment of physical operation evidence with twin simulation evidence. For example, it can be constructed using a directed acyclic graph (DAG), where each node represents a standard operation, edges represent the sequence of operations, and edge weights represent standard time intervals. The graph can be constructed based on expert experience, historical maintenance data analysis, or a digitized maintenance manual. Alternatively, it can be based on a rule engine or state machine, predefining a series of operation sequences and state transition conditions, with each operation node containing its expected execution time range and logical relationships with other operations.

[0078] For each physical operation evidence unit in the physical operation evidence stream, the operation action tags contained in that physical operation evidence unit are parsed, and the target standard operation node matching the operation action tag is located in the standard operation time sequence graph. The operation action tags contained in the physical operation evidence unit are semantic descriptions of the actual maintenance operation, such as "changing engine oil" or "tightening bolts." Locating the target standard operation node means matching the semantic tags of these actual operations with predefined standard operation nodes in the standard operation time sequence graph to determine the position and type of the current physical operation in the standard maintenance process. Its purpose is to establish a semantic association between the actual operation and the standard process, providing a foundation for subsequent time sequence alignment and difference analysis. For example, operation action tags can be parsed using keyword matching and natural language processing (NLP) techniques, and similarity calculations can be performed with the names or descriptions of standard operation nodes, selecting the node with the highest similarity as the target standard operation node. Alternatively, a mapping table between operation action tags and standard operation nodes can be pre-defined, and when a specific operation action tag is parsed, the mapping table can be directly queried to obtain the corresponding standard operation node.

[0079] Based on the position of the target standard operation node in the standard operation timing graph, the preceding and succeeding standard operation nodes that are directly related to the target standard operation node are determined, and the first standard time interval of the target standard operation node relative to its preceding standard operation node and the second standard time interval of its succeeding standard operation node relative to the target standard operation node are obtained. Its function is to construct an expected time context based on standard processes for the current physical operation evidence unit, thereby defining a reasonable "allowed time window" for filtering matching twin simulation evidence units. For example, in the standard operation timing graph, by traversing the incoming and outgoing edges directly connected to the target standard operation node, its preceding and succeeding nodes can be found, and the corresponding standard time intervals can be read from the edge attributes. Alternatively, if the graph is represented using an adjacency matrix or adjacency list, the adjacency information of the target node can be directly queried to obtain its predecessor and successor nodes and their associated time attributes.

[0080] Based on the first timestamp of the physical operation evidence unit, twin simulation evidence units whose second timestamps fall within an allowed time window are selected from the twin simulation evidence stream as candidate twin simulation evidence units. The lower limit of the allowed time window is the first timestamp minus the first standard time interval, and the upper limit is the first timestamp plus the second standard time interval. The physical operation evidence unit is then combined with each selected candidate twin simulation evidence unit to obtain multiple candidate evidence pairs corresponding to that physical operation evidence unit. Its purpose is to narrow the search range, improve pairing efficiency and accuracy, and allow for a certain time deviation between the actual operation and the simulation, thus enhancing the robustness of the system. For example, a timestamp comparison algorithm can be used to traverse all evidence units in the twin simulation evidence stream and determine whether their second timestamp satisfies the condition: first timestamp - first standard time interval <= second timestamp <= first timestamp + second standard time interval. Alternatively, the query function of a time series database can be used to directly retrieve twin simulation evidence units within a specified time window. Each selected candidate twin simulation evidence unit will be combined with the current physical operation evidence unit to form a candidate evidence pair.

[0081] Through the aforementioned technical solution, this application, when pairing physical operation evidence units with twin simulation evidence units, no longer relies solely on simple timestamp comparison, but introduces temporal constraints based on standard maintenance processes. Specifically, by constructing a standard operation timing graph, the sequential order and standard time intervals of each operation are clarified, providing precise semantic and temporal references for pairing. After locating the target standard operation node, a dynamic allowable time window can be preset based on the standard process, enabling physical operation evidence units to be matched with twin simulation evidence units within a reasonable time deviation range. This method solves the problem of inaccurate pairing caused by the complexity of maintenance operation timing and timestamp differences, improving the accuracy and reliability of associated evidence pairs. This lays a solid foundation for subsequent difference analysis, evidence graph construction, and evidence chain preservation, ensuring the judicial probative value of the entire maintenance order evidence chain.

[0082] In some of the embodiments described above in this application, an initial difference evolution path is proposed to track the evolution of differences during the maintenance process. However, in its implementation, there are shortcomings in how to accurately quantify the deviation and structurally represent its evolution over time to support subsequent global analysis and anomaly identification. Specifically, the deviation vector is treated as a single data point, lacking multidimensional decomposition and anomaly level mapping, making it difficult to quantify and compare the deviation evolution process. Furthermore, the deviation is not closely linked to the maintenance time sequence, failing to form a structured path over time, thus affecting the accurate tracing of abnormal processes.

[0083] To address this, this application further proposes a method for generating an initial difference evolution path corresponding to a pair of related evidence based on the deviation vector and the position of the related evidence pair in the maintenance timeline. This method includes: performing multidimensional feature decomposition on the deviation vector to obtain deviation components in multiple preset dimensions; mapping each deviation component in the preset dimension to its corresponding anomaly level interval to obtain an anomaly level value for each preset dimension; assigning a timestamp to the anomaly level value based on the position of the related evidence pair in the maintenance timeline, and encapsulating the timestamp and the anomaly level value into a difference node of the related evidence pair; obtaining the difference nodes of other related evidence pairs in the same maintenance process, and concatenating the difference nodes with the difference nodes of other related evidence pairs according to the order of their timestamps to obtain the initial difference evolution path.

[0084] Specifically, during the maintenance process, after comparing the simulation state contained in the twin simulation evidence unit with the standard performance benchmark in the benchmark instance dimension by dimension, a deviation vector is calculated. This deviation vector may contain multiple interrelated or independent performance index deviations. To analyze these deviations more precisely, this application performs multidimensional feature decomposition on the deviation vector. Multidimensional feature decomposition is a technique for converting high-dimensional data into a low-dimensional representation, such as using principal component analysis (PCA), independent component analysis (ICA), or domain-specific decomposition algorithms. Through decomposition, the original complex deviation vector can be broken down into multiple deviation components in a pre-defined dimension that are more physically independent and easier to interpret. For example, a deviation vector containing multiple parameters such as temperature, pressure, and vibration can be decomposed into "thermal management system deviation components," "power output system deviation components," etc., thereby making the understanding and quantification of deviations more in-depth and specific, avoiding the one-sidedness that may be caused by single-dimensional analysis.

[0085] To standardize and quantify these decomposed deviation components, this application maps the deviation components on each preset dimension to corresponding anomaly level intervals to obtain anomaly level values ​​for each preset dimension. The anomaly level interval can be a pre-defined numerical range, such as a rating range of 0 to 100, or a set of discrete level labels, such as "normal," "slightly abnormal," "moderately abnormal," and "severely abnormal." The mapping method can be based on preset threshold rules, such as classifying a deviation component as "moderately abnormal" when it exceeds a certain value. Alternatively, it can use statistical methods, such as determining the level based on the standard deviation of historical data distribution. Or, it can use a machine learning classification model for automatic mapping. This mapping achieves standardized quantification of different types and degrees of deviation, facilitating comparisons across operational stages and intuitive judgment of anomaly severity.

[0086] Building upon this foundation, to precisely correlate deviation information with specific maintenance timelines, this application assigns a timestamp to the anomaly level value based on the position of the correlated evidence pair within the maintenance timeline, and encapsulates the timestamp and anomaly level value into a difference node for the correlated evidence pair. The position of the correlated evidence pair within the maintenance timeline can be determined by the timestamp of its corresponding physical operation evidence unit or twin simulation evidence unit, such as the operation start time, end time, or intermediate time point. The difference node is a structured data unit, which can be a data object, a JSON structure, or a database record. It encapsulates the anomaly level value, the corresponding timestamp, and a unique identifier for the correlated evidence pair. This encapsulation ensures that each deviation information is precisely correlated with a specific maintenance timeline position, providing a solid temporal context for subsequent tracing and analysis.

[0087] To establish a coherent trajectory of deviation evolution, this application obtains the difference nodes of other related evidence pairs during the same maintenance process and concatenates these difference nodes with the difference nodes of the other related evidence pairs according to their timestamp order to obtain the initial difference evolution path. Obtaining other difference nodes can be achieved by querying temporary storage or a database during the maintenance process. The concatenation method can be to construct an ordered list, linked list, or time series array, ensuring that all difference nodes are arranged in chronological order of their timestamps.

[0088] Through the above technical solution, this application can capture the evolution of deviations during the maintenance process more precisely. Multidimensional feature decomposition of the deviation vector breaks down the complex overall deviation into multiple independently analyzable dimensions, making the understanding and quantification of deviations more in-depth and specific, avoiding the one-sidedness that may result from single-dimensional analysis. Mapping these decomposed deviation components to a unified anomaly level range achieves standardized quantification of different types and degrees of deviations, facilitating comparisons across operational stages and intuitive judgment of anomaly severity. Furthermore, by assigning timestamps to anomaly level values ​​and encapsulating them as difference nodes, it ensures that each deviation information is precisely associated with a specific maintenance timeline, providing a solid temporal context for subsequent tracing and analysis. Concatenating these difference nodes in timestamp order forms a structured, continuous initial difference evolution path, clearly demonstrating the dynamic evolution of deviations during maintenance, greatly improving the diagnostic capability for maintenance quality problems, and providing an accurate and traceable data foundation for subsequent global difference calculation, anomaly identification, and causal tracing.

[0089] In some of the solutions mentioned above in this application, a global difference degree is obtained by comparing the measured state with the simulated state to verify the maintenance effect. However, in this process, direct comparison may not be able to fully take into account the influence of historical deviations in each link of the maintenance process, resulting in an inaccurate global difference degree and affecting the reliability of the evidence chain.

[0090] To address this, this application further proposes a method for comparing the measured state with the simulated state output by the predicted instance after all maintenance operations are completed, to obtain the global difference. The specific steps include: Feature extraction and layer-by-layer comparison of the measured state and the simulated state are performed to obtain a multi-dimensional original difference feature map.

[0091] The initial difference evolution path corresponding to multiple related evidence pairs generated during the maintenance process is obtained. The original difference feature map is spatially matched with the initial difference evolution path to obtain the maintenance operation link to which each difference feature point in the original difference feature map belongs. Based on the deviation amount corresponding to each maintenance operation link in the initial difference evolution path, the difference feature value at the corresponding position in the original difference feature map is weighted and corrected to obtain the corrected multidimensional difference feature map.

[0092] The modified multidimensional difference feature map is subjected to dimensionality reduction and aggregation processing to extract the difference components between the measured state and the simulated state on multiple preset evaluation dimensions.

[0093] The contribution weight of each difference component to the overall maintenance quality is calculated based on the preset fault propagation model, and the multiple difference components are weighted and fused according to the contribution weight to obtain the global difference.

[0094] Specifically, in the process of extracting features and performing layer-by-layer comparisons on the measured and simulated states to obtain multi-dimensional original difference feature maps, feature extraction aims to transform the raw, heterogeneous measured and simulated data into unified, structured feature representations for effective comparison. For example, methods based on deep learning models (such as Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs)) can be used to automatically extract high-dimensional features from time-series sensor data or image data. Alternatively, signal processing techniques (such as Fourier transforms and wavelet transforms) can be used to extract frequency domain or time-frequency domain features from time-domain data. Layer-by-layer comparison compares these extracted features at different levels of abstraction to identify differences. For example, hierarchical matching algorithms can be used to compare macroscopic system-level performance indicators down to microscopic component-level parameters. Alternatively, difference measurement functions (such as Euclidean distance and cosine similarity) can be used to calculate the differences between feature vectors at different levels. The resulting original difference feature map can be a multi-channel image, a multi-dimensional vector, or a tensor, where each element or channel represents an initial difference at a specific dimension, time point, or component.

[0095] In the process of acquiring multiple correlation evidence pairs generated during maintenance and their corresponding initial difference evolution paths, the original difference feature map is spatially matched with these initial difference evolution paths to determine the maintenance operation steps to which each difference feature point in the original difference feature map belongs. Based on the deviation of each maintenance operation step in the initial difference evolution path, the difference feature values ​​at corresponding positions in the original difference feature map are weighted and corrected to obtain the corrected multidimensional difference feature map. During this process, the initial difference evolution path records the historical deviation of the simulation state of each operation step relative to the standard performance benchmark. The purpose of spatial matching is to associate the difference information in the original difference feature map with the maintenance operation steps. For example, timestamp alignment can be used to match the time axis in the original difference feature map with the time nodes in the initial difference evolution path, thereby determining the maintenance operation step corresponding to each difference feature point. Alternatively, semantic label matching can be used to semantically associate specific feature regions or feature points in the original difference feature map with the marked maintenance operation steps in the initial difference evolution path. Weighted correction adjusts the difference feature values ​​at corresponding positions in the original difference feature map based on the historical deviation of each maintenance operation step in the initial difference evolution path. For example, a linear weighting method can be used to multiply the difference feature values ​​by a correction coefficient proportional to the historical deviation. Alternatively, the deviation can be mapped using a non-linear function (such as the sigmoid function) to generate a correction factor, which is then combined with the difference feature values. This correction can eliminate or reduce the interference of historical deviations on the current difference assessment, making the corrected multidimensional difference feature map more accurately reflect the actual difference situation.

[0096] In the process of dimensionality reduction and aggregation of the corrected multidimensional difference feature map to extract the difference components between the measured state and the simulated state across multiple preset evaluation dimensions, the dimensionality reduction and aggregation process aims to simplify the data and extract key, representative difference information from the complex difference feature map. For example, statistical dimensionality reduction methods such as Principal Component Analysis (PCA) and t-SNE can be used to project the high-dimensional feature map into a low-dimensional space, retaining the most important difference information. Alternatively, feature aggregation can be performed using pooling layers or fully connected layers based on convolutional neural networks (CNNs) to extract high-level semantic features. The preset evaluation dimensions can be key aspects of heavy-duty truck maintenance quality assessment, such as engine performance, braking system safety, fuel economy, and component wear. The extracted difference components are the quantitative differences between the measured state and the simulated state across these specific evaluation dimensions.

[0097] In the process of calculating the contribution weight of each difference component to the overall maintenance quality based on a pre-defined fault propagation model, and then weighting and fusing multiple difference components according to this contribution weight to obtain the global difference, the fault propagation model is used to describe the mechanism by which faults in various components, systems, or maintenance operation links of a heavy truck influence and propagate each other. For example, this model can be a causal graph or Bayesian network built based on expert knowledge, where nodes represent different components or operation links, and directed edges represent the direction and intensity of fault propagation. Alternatively, it can be a machine learning model trained on historical maintenance data to learn the correlation between different fault characteristics and overall maintenance quality. Through this model, the contribution weight of each difference component to the overall maintenance quality can be calculated, for example, by determining the weight based on the number of downstream components affected by a certain difference component or its criticality in the fault propagation path. Alternatively, sensitivity analysis can be used to assess the degree of impact of changes in each difference component on the overall maintenance quality indicators. These difference components are weighted and merged according to their contribution weights. For example, a linear weighted summation method is used, where each difference component is multiplied by its contribution weight and then summed to obtain a comprehensive global difference score. This global difference score can comprehensively and accurately reflect the overall quality of this maintenance.

[0098] Through the above technical solution, this application can solve the problem of inaccurate global difference degree caused by insufficient consideration of the historical deviations in each stage of the maintenance process when directly comparing the measured state and the simulated state. By introducing the initial difference evolution path generated during the maintenance process, spatial location matching and weighted correction are performed on the original difference feature map, enabling the calculation of global difference degree to trace and integrate historical deviation information, thereby obtaining a more accurate difference assessment after historical correction. On this basis, combined with a preset fault propagation model, the actual contribution of different dimension difference components to the overall maintenance quality is further quantified, ensuring that the global difference degree not only reflects the magnitude of the difference, but also its actual impact on the performance and safety of the heavy truck. This method enables the global difference degree to more accurately reflect the maintenance effect, improves the reliability and persuasiveness of the maintenance order evidence chain, and provides more solid technical support for warranty determination, liability tracing, and dispute resolution.

[0099] In some of the solutions mentioned above in this application, the initial difference evolution path corresponding to multiple related evidence pairs generated during the maintenance process is proposed to support the calculation of global difference degree. However, in this process, due to the data of the related evidence pairs being scattered and the temporal order being disordered, it is difficult to efficiently and accurately organize and sort the path data, resulting in the lack of structure and orderliness of path information in subsequent comparison operations, which affects the accuracy and reliability of global difference degree calculation.

[0100] To address this, this application further proposes a method for obtaining the initial difference evolution path corresponding to multiple pairs of related evidence generated during the maintenance process. This method includes: traversing multiple pairs of related evidence generated during the maintenance process; extracting path data corresponding to each pair of related evidence, whereby the path data includes the position information of the related evidence pair in the maintenance timeline and the deviation of the simulation state at that position relative to the standard performance baseline; sorting the extracted path data according to the maintenance operation sequence based on the position information of each pair of related evidence in the maintenance timeline to generate a path data sequence organized according to time series; and determining this path data sequence as the initial difference evolution path corresponding to the multiple pairs of related evidence.

[0101] Specifically, when obtaining the initial difference evolution paths corresponding to multiple related evidence pairs generated during the maintenance process, it is necessary to traverse the multiple related evidence pairs generated during the maintenance process and extract the corresponding path data from each related evidence pair. Here, "traversal" refers to systematically accessing each generated related evidence pair to ensure that all relevant data points are considered and to avoid omitting any key information that may affect the difference evolution analysis. For example, this can be achieved by iterating through the set of related evidence pairs stored in a database or memory, or by capturing them in real time through an event-driven mechanism when the related evidence pairs are generated. The extracted "path data" describes the key information of the related evidence pair during the maintenance process, and it contains at least two core elements: First, the "position information of the related evidence pair in the maintenance timeline," which can be the operation start timestamp, operation end timestamp, operation step ID, or sequence number in the standard maintenance process, used to identify the specific position of the evidence pair in the entire maintenance timeline. Second, the "deviation of the simulation state at the position relative to the standard performance benchmark," which is usually a numerical vector or scalar that quantifies the degree of difference between the simulation state output by the predicted instance and the standard performance benchmark fixed in the benchmark instance at that specific maintenance timeline point. The extracted data can be encapsulated into a structure or object as an independent path data unit.

[0102] Based on the positional information of each associated evidence pair in the maintenance timeline, the extracted path data are sorted according to the maintenance operation sequence, thereby generating a path data sequence organized according to time series. This step aims to address the potential disorder problem of the original path data. Since associated evidence pairs may be generated at different time points or in a non-linear manner, the directly extracted data may not possess strict temporal order. By utilizing the positional information in the maintenance timeline contained in the path data, such as timestamps or operation step IDs, standard sorting algorithms (such as quicksort and mergesort) can be used to sort these path data in ascending order. If the positional information is an operation step ID, it can be sorted according to the step order in a predefined standard maintenance process flow diagram. In this way, the originally scattered and potentially chaotic path data is reconstructed into a logically clear and temporally continuous sequence, accurately reflecting the actual evolution trajectory of differences during the maintenance process.

[0103] This path data sequence is identified as the initial difference evolution path corresponding to the multiple pairs of related evidence. This operation formally names and confirms the aforementioned sorting results, meaning that the ordered path data sequence represents a complete record of how the deviation of the simulation state from the standard performance baseline gradually accumulates and changes from the start to the end of maintenance. This sequence can be encapsulated into a specific data structure, such as a list or array, and used as structured input for subsequent calculations of global difference, analysis of difference evolution path chains, and identification of anomalous links.

[0104] Through the above technical solution, this application solves the problems of data dispersion and temporal disorder in the acquisition of initial difference evolution paths. By systematically traversing and extracting the key path data of each associated evidence pair, the integrity of the information is ensured. Furthermore, based on the location information in the maintenance timeline, these path data are precisely sorted to generate a structured, time-series-based path data sequence, thereby eliminating the interference of data disorder on subsequent analysis. This ordered initial difference evolution path provides accurate and reliable input for subsequent comparison of measured and simulated states to obtain the global difference degree, improving the accuracy and reliability of global difference degree calculation, and laying a solid data foundation for the construction of the entire evidence chain of heavy truck maintenance orders.

[0105] In some of the solutions mentioned above in this application, the original difference feature map is matched with the initial difference evolution path to determine the maintenance operation link to which the difference point belongs. However, in this process, there may be problems such as inaccurate mapping of spatial coordinates and time axis and imprecise matching, which makes it impossible to accurately identify abnormal links.

[0106] To address this, this application further proposes a method for spatially matching the original difference feature map with the initial difference evolution path to obtain the maintenance operation stage to which each difference feature point in the original difference feature map belongs. The method includes: parsing the initial difference evolution path, extracting multiple location nodes and a maintenance operation stage identifier corresponding to each location node from the initial difference evolution path, where each location node corresponds to a specific location in the maintenance timeline; obtaining the spatial coordinates of each difference feature point in the original difference feature map, mapping these spatial coordinates to the time axis of the initial difference evolution path, and determining the distance relationship between each difference feature point and the multiple location nodes; based on this distance relationship, matching the nearest location node to each difference feature point, and assigning the maintenance operation stage identifier corresponding to that location node to the difference feature point, thereby obtaining the maintenance operation stage to which each difference feature point in the original difference feature map belongs.

[0107] Specifically, the initial difference evolution path is parsed to extract multiple location nodes and their corresponding maintenance operation step identifiers. Each location node corresponds to a specific position in the maintenance timeline. Here, "parsing" refers to reading and interpreting the generated initial difference evolution path data structure to identify the key information it contains. One implementation is that if the initial difference evolution path is stored as structured data (e.g., JSON, XML, or database records), the system can directly extract predefined location nodes and corresponding maintenance operation step identifiers through field names or path expressions. For example, the path data might contain a series of timestamps and associated maintenance operation step names or IDs; these timestamps are the location nodes, and the step names or IDs are the maintenance operation step identifiers. Another implementation is that if the initial difference evolution path exists as a serialized data stream, the system can use a specific parser or pattern matching algorithm to identify specific markers or data segments in the data stream that represent location nodes and maintenance operation step identifiers. These location nodes have clear positioning in the maintenance timeline, providing a time reference for subsequent difference feature point matching.

[0108] The spatial coordinates of each differential feature point in the original differential feature map are obtained. These spatial coordinates are then mapped to the time axis of the initial differential evolution path to determine the distance relationship between each differential feature point and the multiple location nodes. The original differential feature map is typically multi-dimensional, and its differential feature points may exist in image pixel coordinates, 3D point cloud coordinates, or other spatial forms. Mapping these spatial coordinates to the time axis aims to align the spatial dimension of the differential information with the time series of the maintenance process. One implementation is that if the original differential feature map originates from a video or image sequence, its spatial coordinates (e.g., pixel coordinates) can be associated with the timestamps of the image frames, which in turn can be synchronized with the maintenance time series. In this way, the spatial coordinates of the differential feature points can be converted into their corresponding time points in the maintenance time series, thereby determining their temporal distances to the various location nodes in the initial differential evolution path. Another implementation is that if the differential feature points originate from sensor data or 3D scan data, their spatial coordinates may be directly or indirectly associated with the acquisition time. For example, using the timestamps of the sensor data, the spatial coordinates can be mapped to a specific moment in the maintenance time series, thereby calculating the temporal distance to the location nodes. Distance relationships can be absolute differences over time or relative distances after normalization.

[0109] Based on this, and using this distance relationship, the nearest location node is matched for each difference feature point, and the corresponding maintenance operation step identifier is assigned to the difference feature point, thus obtaining the maintenance operation step to which each difference feature point belongs in the original difference feature map. The matching process aims to find the closest known maintenance operation step in terms of maintenance time sequence for each difference feature point. One implementation is to calculate the distance between the time point mapped to the time axis of each difference feature point in the original difference feature map and the time points of all location nodes in the initial difference evolution path, and then select the location node with the smallest distance as the matching result. For example, Euclidean distance or Manhattan distance can be used to measure the temporal proximity. Another implementation is to use a spatial index structure (such as a KD tree or R tree) to store location nodes to improve matching efficiency, thereby accelerating the process of nearest neighbor search for each difference feature point. Once the nearest location node is found, the maintenance operation step identifier associated with that location node (e.g., "replace brake pads", "check engine oil level", etc.) is assigned to the difference feature point.

[0110] Through the above technical solution, this application can accurately associate each difference feature point in the original difference feature map with a specific operation step in the maintenance timeline. This precise matching solves the problems of inaccurate spatial coordinate and time axis mapping and imprecise matching in traditional methods, ensuring that each difference point can be accurately assigned to the maintenance operation step that generated it. This is crucial for subsequent weighted correction of difference feature values, as it ensures that the correction weights can be accurately applied to the difference feature points generated by the corresponding maintenance operation steps, thereby improving the accuracy and effectiveness of the correction. This helps to more accurately assess the global difference between the measured state and the simulation state, providing a more reliable and refined data foundation for constructing the entire process evidence chain of heavy truck maintenance orders.

[0111] In some of the solutions mentioned above in this application, the original difference feature map is corrected according to the initial difference evolution path to improve the accuracy of the global difference degree. However, in this process, how to specifically implement the weighted correction based on the deviation amount to ensure that the corrected feature map more realistically reflects the differences in the maintenance operation process and avoid the distortion of the correction result due to the deviation amount not being directly related to the difference feature point is a problem that needs to be solved.

[0112] To address this, this application further proposes a method to weight and correct the difference feature values ​​at corresponding positions in the original difference feature map based on the deviation amounts corresponding to each maintenance operation step in the initial difference evolution path, thereby obtaining a corrected multidimensional difference feature map. This method specifically includes the following steps: The process involves obtaining the deviation amounts corresponding to each maintenance operation step in the initial difference evolution path and generating a correction weight coefficient positively correlated with the deviation amount for each maintenance operation step. Obtaining the deviation amounts for each maintenance operation step in the initial difference evolution path aims to quantify the degree of deviation between each specific maintenance operation step and the standard performance benchmark. These deviation amounts can be directly parsed from the initial difference evolution path, and may have been pre-calculated during path generation by comparing the simulation state with the standard performance benchmark. Alternatively, after obtaining the initial difference evolution path, the deviations within each maintenance operation step can be aggregated and calculated based on detailed data contained in the path (e.g., simulation state, standard performance benchmark, timestamps, etc.), for example, by calculating the average, maximum, or weighted average to obtain the representative deviation amount for that step. Various mapping methods can be used when generating the correction weight coefficients positively correlated with the deviation amounts for each maintenance operation step. For example, a linear function can be used to map the deviation amount to the correction weight coefficient, ensuring that the larger the deviation amount, the larger the weight coefficient. Alternatively, nonlinear mapping methods such as exponential, logarithmic, or piecewise functions can be used to more precisely reflect the impact of the deviation amount on the correction strength. Alternatively, a lookup table of deviation and correction weight coefficient can be predefined, and the corresponding correction weight coefficient can be directly queried based on the obtained deviation.

[0113] Based on the maintenance operation steps to which each difference feature point belongs in the original difference feature map, the target maintenance operation step corresponding to each difference feature point is determined. This step aims to establish the association between the fine-grained difference feature points in the original difference feature map and the more macroscopic maintenance operation steps in the initial difference evolution path. Specifically, each difference feature point may contain or be associated with timestamp information. By comparing the timestamp of the difference feature point with the time interval of each maintenance operation step in the initial difference evolution path, the maintenance operation step to which the difference feature point belongs can be determined. Furthermore, if the original difference feature map has a spatial dimension, and the maintenance operation steps can be mapped to that spatial dimension to some extent, the maintenance operation step to which the difference feature point belongs can be determined through spatial coordinate mapping.

[0114] The core of weighted correction is to find the target corrected weight coefficient corresponding to the target maintenance operation step from the corrected weight coefficients. This target corrected weight coefficient is then multiplied by the original difference feature value of the difference feature point to obtain the corrected difference feature value. This step directly applies the degree of deviation of a specific maintenance operation step to its associated difference feature point. When searching for the target corrected weight coefficient, the maintenance operation step identifier can be used as the key, and the corrected weight coefficient as the value, constructing a hash table or dictionary for fast lookup. The most direct way to multiply the target corrected weight coefficient by the original difference feature value is to perform a simple multiplication operation to directly scale the original difference feature value.

[0115] Based on the corrected differential eigenvalues ​​of all differential feature points, a multi-dimensional feature map is reconstructed, and this reconstructed multi-dimensional feature map serves as the corrected multi-dimensional differential feature map. This step ensures that all corrected differential eigenvalues ​​can be integrated back into a unified feature representation. If the original differential feature map is a matrix or tensor, all corrected differential eigenvalues ​​are refilled into the new matrix or tensor according to their original spatial or temporal positions. If the differential feature map is represented by a more abstract data structure, the values ​​in that data structure are directly updated.

[0116] Through the above technical solution, this application ensures that the correction weighting coefficient objectively reflects the actual degree of deviation, avoiding the influence of subjective factors on the correction process. Simultaneously, by accurately mapping each difference feature point to its corresponding maintenance operation step, targeted local correction is achieved, avoiding over-correction of irrelevant areas or dilution of key issues. Multiplying the target correction weighting coefficient by the original difference feature value of the difference feature point allows the corrected difference feature value to more realistically reflect the differences in maintenance operation steps, effectively avoiding distortion of correction results caused by deviations not being directly associated with difference feature points. The reconstructed corrected multidimensional difference feature map provides a more accurate and reliable input for subsequent global difference degree calculation, improving the accuracy of maintenance effect evaluation and the effectiveness of diagnosis.

[0117] In some of the solutions mentioned above in this application, a weighted average of the contribution of each difference component to the overall maintenance quality is calculated based on a preset fault propagation model to obtain the global difference. However, if the calculation process is not specified, the contribution weight may be inaccurate, the fault propagation effect may be ignored, and the accuracy of the global difference may be affected.

[0118] To address this, this application further proposes a method for calculating the contribution weight of each difference component to the overall maintenance quality based on a pre-defined fault propagation model. This method includes: obtaining difference components on multiple pre-defined evaluation dimensions and labeling each difference component with a corresponding maintenance operation step identifier; inputting each difference component and its corresponding maintenance operation step identifier into the fault propagation model, which contains a directed graph of fault propagation between multiple maintenance operation steps, where nodes represent maintenance operation steps and directed edges represent fault propagation directions; calculating the fault impact range of each difference component propagating to other steps based on the fault propagation model and the directed graph; and calculating the contribution weight of each difference component to the overall maintenance quality based on the fault impact range of each difference component.

[0119] Specifically, when acquiring the difference components across multiple preset evaluation dimensions and labeling each difference component with a corresponding maintenance operation step identifier, this step aims to establish a correlation between abstract performance difference data and specific maintenance operation steps, providing refined input for subsequent fault propagation analysis. The difference component measures the degree of deviation between the measured and simulated states on a specific evaluation dimension; for example, it could be a difference in engine power, braking performance, or fuel efficiency. The maintenance operation step identifier indicates which specific maintenance step (such as replacing parts, adjusting parameters, or cleaning) the difference component might be associated with. One implementation method is to simultaneously record the position of the original data from which the difference component originates in the maintenance timeline when generating the difference component, and map this position to the corresponding maintenance operation step according to a preset maintenance flowchart, thereby automatically generating the maintenance operation step identifier. Another implementation method is to manually or semi-automatically prompt maintenance personnel or an expert system when the system detects that a difference component exceeds a threshold, allowing them to manually or assistedly select the corresponding maintenance operation step identifier for that difference component based on experience or diagnostic rules.

[0120] When each difference component and its corresponding maintenance operation step identifier are input into the fault propagation model, the model contains a directed graph of fault propagation between multiple maintenance operation steps. Nodes in this graph represent maintenance operation steps, and directed edges represent the direction of fault propagation. The fault propagation model is the core; it uses a directed graph to depict the potential fault impact and transmission relationships between different operation steps in the heavy truck maintenance process. Inputting the difference components and their step identifiers allows the model to simulate the diffusion path and impact range of faults in the maintenance process based on this information. Nodes represent specific maintenance operations, such as "changing engine oil," "adjusting brakes," and "checking the circuit," while directed edges indicate that an abnormality in one step may cause a problem in another step; for example, "improper brake adjustment" may lead to "degraded braking performance." One implementation approach is to construct the fault propagation directed graph based on domain expert knowledge, using methods such as expert interviews, fault tree analysis, and Failure Mode and Effects Analysis (FMEA) to identify the causal relationships and impact paths between various heavy truck components and maintenance operations. Another implementation approach is to learn and construct the fault propagation directed graph using historical maintenance data. For example, by analyzing fault reports, maintenance logs, and performance test results from a large number of maintenance cases, and using techniques such as graph neural networks and association rule mining, fault propagation patterns and directed edges between different maintenance operation stages can be automatically discovered and established.

[0121] This fault propagation model calculates the impact range of each maintenance operation step's difference component on other steps based on the directed fault propagation graph. This step utilizes the constructed fault propagation model to quantify the impact of each local difference on the overall maintenance system performance. The fault impact range refers to the set of other maintenance operation steps that a particular maintenance operation step's difference component can affect through the directed fault propagation graph, along with the strength of its impact. This helps identify critical, cascading fault sources. One implementation method is to use graph traversal algorithms (such as depth-first search or breadth-first search) to simulate propagation on the directed fault propagation graph. Starting from the maintenance operation step node where the difference occurs, traversing along the directed edges, and calculating the impact value on downstream nodes based on the edge weights (representing propagation strength) and node attributes (representing the sensitivity of the step), thereby determining the impact range. Another implementation method is to use matrix operations to represent the directed fault propagation graph as an adjacency matrix, and simulate the iterative propagation process of the fault through matrix multiplication or exponentiation, calculating the cumulative impact of each difference component on other steps after multiple propagation steps.

[0122] Based on the fault impact range of each difference component, the contribution weight of each difference component to the overall maintenance quality is calculated. The goal is to assign a weight to each difference component that reflects its degree of influence on maintenance quality. Difference components with larger impact ranges and higher impact strengths should have larger contribution weights, thus playing a more important role in subsequent global difference calculations. One approach is to directly use or convert the quantified value of the fault impact range (e.g., the number of affected nodes, cumulative impact strength, the reciprocal of the propagation path length, etc.) as the contribution weight, or through a mapping function. For example, a larger impact range corresponds to a higher weight. Another approach is to use the analytic hierarchy process (AHP) or fuzzy comprehensive evaluation method, combining expert scoring and impact range data, to comprehensively assess the contribution weight of each difference component.

[0123] By employing the aforementioned technical solution, the calculation of the global difference degree can fully consider the potential fault propagation and mutual influence between different maintenance operation stages, enabling each difference degree component to obtain a more accurate contribution weight during weighted fusion. This avoids the errors that may arise from simple averaging or fixed weight allocation in traditional methods, improving the accuracy and objectivity of the global difference degree assessment. For example, a seemingly minor operational deviation, if located at a critical node in the fault propagation path, may have a chain reaction impacting multiple subsequent stages. This solution can accurately identify and quantify this impact, thereby assigning a higher weight to the deviation, allowing the global difference degree to more realistically reflect the overall maintenance quality. This provides a more solid and reliable foundation for subsequent generation of difference evidence nodes, identification of abnormal stages, and construction of evidence maps for maintenance orders, enhancing the judicial probative value and traceability of the entire evidence chain.

[0124] In some of the embodiments described above in this application, a global difference degree is obtained by comparing the measured state with the simulated state to quantify the difference between the actual state and the predicted state after maintenance. However, in this process, direct overall comparison may ignore the accumulation and evolution of local deviations during maintenance, resulting in an inaccurate global difference degree and affecting the accurate identification of abnormal links.

[0125] To address this, this application proposes a method for constructing a full-process evidence chain for heavy-duty truck maintenance orders. The method involves comparing the measured state with the simulated state output by the predicted instance after all maintenance operations are completed to obtain a global difference degree. The method includes: acquiring the initial difference evolution path corresponding to multiple related evidence pairs generated during the maintenance process; extracting multiple key time-series nodes and the node deviation amount corresponding to each key time-series node from these initial difference evolution paths; decomposing the measured state and the simulated state into multiple measured state segments and multiple simulated state segments corresponding to the multiple key time-series nodes; comparing the measured state segment corresponding to each key time-series node with the simulated state segment to generate a node difference degree at that key time-series node; and correcting the node difference degree based on the node deviation amount corresponding to the key time-series node to obtain a corrected node difference degree; and aggregating the corrected node difference degrees corresponding to the multiple key time-series nodes according to the maintenance sequence, using the aggregation result as the global difference degree between the measured state and the simulated state.

[0126] Specifically, the process involves acquiring the initial difference evolution path corresponding to multiple pairs of related evidence generated during the maintenance process. From this initial difference evolution path, multiple key time-series nodes and the corresponding node deviation for each key time-series node are extracted. These key time-series nodes correspond to maintenance operation steps where state deviations occur during the maintenance process. This step aims to identify specific time points or operation steps that affect maintenance quality from the deviation records of the entire maintenance process and quantify the degree of deviation in these steps. This helps to decompose complex global difference problems into more easily analyzed and localized local problems. In one implementation, key time-series nodes can be determined by data analysis of the initial difference evolution path, for example, identifying points with high deviation change rates, points where deviations exceed preset thresholds, or key checkpoints preset based on expert experience. The node deviation is then directly read from the path data. In another implementation, machine learning algorithms, such as time series anomaly detection models, can be used to analyze the initial difference evolution path, automatically identify statistically significant anomalies as key time-series nodes, and extract their corresponding deviations.

[0127] The measured and simulated states are decomposed into multiple measured state segments and multiple simulated state segments corresponding to multiple key time-series nodes. This step aims to segment the continuous, overall measured and simulated state data according to key time-series nodes, thereby achieving localized and refined comparison. This decomposition allows subsequent difference calculations to focus on state changes directly related to key operational steps, avoiding interference from irrelevant data. In one implementation, the measured and simulated state data streams can be truncated into multiple segments by setting a time window at or near each key time-series node. Each segment corresponds to a key time-series node and contains state data for a period of time before and after that node. In another implementation, data for the corresponding time period can be extracted from the measured and simulated state data based on the start and end times of the maintenance operation represented by the key time-series node to form state segments. For example, if the key time-series node represents "replacing component X", then the state data from the start to the end of the component X replacement operation is extracted.

[0128] For each key time-series node, the measured state segment corresponding to that node is compared with the simulated state segment to generate a node difference degree at that key time-series node. This node difference degree is then corrected based on the node deviation amount corresponding to that key time-series node, resulting in a corrected node difference degree. This step aims to quantify the difference between the actual and simulated states at each key time-series node and adjust the initial difference degree by considering the accumulated deviations (node ​​deviation amount) during the maintenance process, making it more accurately reflect the real problem of that stage. In one implementation, the comparison can be performed using various data analysis methods, such as calculating the Euclidean distance, Manhattan distance, or correlation coefficient between corresponding parameters of two state segments to generate the node difference degree. During correction, the node deviation amount can be used as a weight or offset to weight or adjust the initial node difference degree. In another implementation, a model-based approach can be used, for example, training a difference assessment model, taking the measured and simulated state segments as input, and outputting the node difference degree. During correction, the node deviation amount can be used as an additional input feature of the model, allowing the model to learn how to adjust the current difference degree based on historical deviations.

[0129] The corrected node differences corresponding to multiple key time-series nodes are aggregated according to the maintenance sequence, and the aggregated result is used as the global difference between the measured state and the simulated state. This step aims to integrate the local, corrected difference information at each key time-series node according to their chronological order in the maintenance process, thereby forming a global indicator that can comprehensively reflect the quality of the entire maintenance process. This aggregation method can reflect the cumulative effect and propagation law of deviations. In one implementation, aggregation can be performed through weighted averaging, summation, or integration based on time series. For example, each corrected node difference can be assigned a weight, which can be determined based on the importance of the node, its duration, or its influence in the fault propagation model, and then a weighted sum can be performed. In another implementation, a hierarchical aggregation model can be constructed to progressively aggregate the corrected differences of adjacent key time-series nodes upwards to obtain the global difference. For example, the node differences within the same maintenance stage can be aggregated first, and then the differences between different maintenance stages can be aggregated.

[0130] Through the above technical solution, this application no longer simply performs an overall comparison, but decomposes the measured and simulated states into segments corresponding to key time-series nodes, and calculates the local difference degree for each segment. More importantly, when calculating the local difference degree, the accumulated node deviation during the maintenance process is introduced for correction, thereby eliminating the interference of historical accumulated deviations on the current difference degree assessment, so that the difference degree of each key link can more realistically and accurately reflect the actual problem of that link. By aggregating these corrected local difference degrees according to the maintenance time sequence, a global difference degree that considers both local details and reflects the overall cumulative effect can be formed. This refined and corrected difference degree calculation method improves the accuracy of the global difference degree, making the subsequent identification and location of abnormal links more accurate, thereby solving the problem that direct overall comparison may ignore the accumulation of local deviations, and providing a more reliable quality assessment basis for the construction of the evidence chain of the entire process of heavy truck maintenance orders.

[0131] In some of the solutions mentioned above in this application, the measured state and the simulated state are decomposed into segments to calculate the global difference. However, in this process, the decomposition method may fail to strictly correspond to the key maintenance operation links due to inaccurate time interval definition or inaccurate segment matching, thereby affecting the reliability of the difference calculation and the integrity of the evidence chain.

[0132] To address this, this application further proposes a method for decomposing the measured state and the simulation state into multiple measured state segments and multiple simulation state segments corresponding to the multiple key timing nodes. Specifically, this includes: determining the time interval between two adjacent key timing nodes based on their sequential order in the maintenance timeline, and defining each time interval as a state segment interval corresponding to the next key timing node; extracting data segments from the measured state that correspond to each state segment interval in time, treating each extracted data segment as a measured state segment, and labeling each measured state segment with a corresponding key timing node identifier; extracting data segments from the simulation state that correspond to each state segment interval in time, treating each extracted data segment as a simulation state segment, and labeling each simulation state segment with a corresponding key timing node identifier; pairing the labeled multiple measured state segments and multiple simulation state segments according to the key timing node identifiers to generate multiple state segment pairs corresponding one-to-one with the multiple key timing nodes, each state segment pair containing one measured state segment and one simulation state segment.

[0133] Based on the sequential order of these key time nodes in the maintenance timeline, the time interval between two adjacent key time nodes is determined. This aims to discretize the continuous maintenance process into stages related to key operations, ensuring that the division of subsequent state segments corresponds to the logical progress of the maintenance operations. For example, by traversing the list of identified key time nodes, the timestamp of each node can be used as the start or end point of the interval. If key time nodes T1, T2, and T3 exist, then the intervals [T1, T2) and [T2, T3) can be determined. Alternatively, the interval boundaries can be dynamically adjusted by analyzing the time intervals between key time nodes and combining them with preset minimum or maximum interval lengths to accommodate the duration differences of different maintenance operations.

[0134] Each time interval is defined as a state segment interval corresponding to the next critical time sequence node. The purpose is to clarify the affiliation of each time interval and associate it with a specific critical time sequence node. This facilitates the logical binding of data within the interval with the maintenance operation or state change represented by the critical time sequence node in subsequent processing, enhancing the semantic integrity of the data segments. For example, a reference or identifier pointing to its "next critical time sequence node" can be added to each time interval object in the data structure. Alternatively, a naming convention can be adopted, such as naming the interval "segment_T_next critical time sequence node ID", thereby establishing a logical correspondence.

[0135] This method involves extracting data segments from the measured state data that correspond to each state segment interval in time. Each extracted data segment is considered a measured state segment. The goal is to accurately extract corresponding data subsets from a continuous measured state data stream based on a defined time interval. These data subsets constitute the measured state segments, representing the actual operation or performance of the heavy truck during a specific maintenance phase. For example, the query function of a time series database can be used to filter and extract data based on the start and end timestamps of the time interval. Alternatively, the data processing program can iterate through the measured state data points, determining whether the timestamp of each data point falls within the current state segment interval, and collecting these data to form segments.

[0136] Each measured state segment is labeled with a corresponding key time-series node identifier. The purpose is to attach a unique identifier to each extracted measured state segment, pointing to its corresponding key time-series node. This gives each segment clear contextual information, facilitating subsequent matching, indexing, and analysis, and ensuring data traceability. For example, a field can be added to the measured state segment data structure to store the key time-series node identifier (e.g., node ID or node name). Alternatively, the measured state segments can be stored in a hash table or dictionary with the key time-series node identifier as the key.

[0137] This method extracts data segments from the simulation state that correspond to each state interval in time. Each extracted data segment is considered a simulation state segment. The aim is to extract corresponding subsets of simulation data from the simulation state data stream output by the predictive instance, based on the same time interval. These simulation state segments represent the performance or behavior of heavy trucks simulated based on the predictive instance within a specific maintenance phase. For example, time-series querying or data traversal methods similar to those used for extracting measured state segments can be employed to extract data from the simulation data source. Alternatively, the API provided by the simulation platform can be used to specify a time range to obtain the corresponding simulation data.

[0138] Each simulated state segment is labeled with a corresponding key timing node identifier. The purpose of this is to attach this identifier to each simulated state segment, ensuring that the simulated data segments and the measured data segments logically belong to the same category, thus laying the foundation for subsequent accurate comparison. For example, a field can be added to the data structure of the simulated state segment to store the key timing node identifier. Alternatively, the same labeling mechanism as the measured state segment can be used to ensure consistency in the identifiers.

[0139] Multiple measured state segments and multiple simulated state segments, after being labeled, are paired according to key timing node identifiers. The aim is to achieve a one-to-one match between measured and simulated state segments based on the key timing node identifiers previously added to each segment. This is a crucial step in achieving accurate comparison, ensuring that segments from different sources but logically corresponding can be correctly associated. For example, the list of measured state segments can be traversed, and for each measured segment, a match can be found in the list of simulated state segments corresponding to its key timing node identifier. Alternatively, a hash map or dictionary can be used, storing the measured and simulated segments separately with the key timing node identifier as the key, and then matching can be performed through key-value lookup.

[0140] Multiple state segment pairs are generated, each corresponding one-to-one with the key time-series nodes. Each state segment pair contains one measured state segment and one simulated state segment, aiming to encapsulate successfully paired measured and simulated state segments into a single "state segment pair." This structured organization ensures that each key time-series node corresponds to a comparison unit containing both actual and simulated data, greatly simplifying subsequent difference calculations and analysis. For example, a "state segment pair" data structure can be defined, containing two members: a measured state segment and a simulated state segment, with one instance created for each key time-series node. Alternatively, the pairing results can be stored as a list or array, where each element is a tuple containing both a measured and a simulated segment.

[0141] Through the above technical solution, this application solves the problem that inaccurate time interval definitions or inaccurate segment matching in decomposition methods can lead to state segments failing to strictly correspond to key maintenance operation steps, thus affecting the reliability of difference calculation and the integrity of the evidence chain. Specifically, time intervals are determined based on the sequential order of key time nodes, and interval boundaries are defined using the node order to ensure that the time interval covers key maintenance steps and avoids segment deviations caused by time misalignment. The time interval is defined to correspond to the next key time node, clarifying interval attribution and enhancing the organization and traceability of state segments. Time-corresponding data segments are extracted from the measured state, directly acquiring data based on the actual time axis to ensure that the measured state segments truly reflect the state of that interval. These data segments are used as measured state segments, creating independent analysis units for easy subsequent comparison. Key time node identifiers are marked on the measured state segments, adding unique markers for easy identification and matching, preventing segment confusion. Similarly, the simulation state is segmented, defined, and labeled to ensure structural consistency between simulation and measured data. The labeled segments are paired according to key timing node identifiers, achieving precise matching based on these identifiers. This ensures that each measured segment is associated with the corresponding simulated segment within the same maintenance process. State segment pairs corresponding one-to-one with key timing nodes are generated, forming directly related comparison units. This allows the difference calculation to focus on specific operational steps, improving the accuracy of the global difference and the reliability of the evidence chain. This provides high-quality, high-precision input for subsequent correction of node differences based on the node deviation corresponding to key timing nodes, and for aggregating the corrected node differences corresponding to multiple key timing nodes according to the maintenance sequence to obtain the global difference between the measured and simulated states. This enhances the accuracy and reliability of the entire maintenance order evidence chain construction method.

[0142] In some of the solutions mentioned above in this application, a method is proposed to obtain the global difference by comparing the measured state with the simulated state. However, in this process, the correction of the node difference may not fully take into account the deviation of the key timing nodes and their importance in the maintenance timing, resulting in the corrected node difference being inaccurate, which in turn affects the accuracy and reliability of the global difference and cannot effectively reflect the impact of the actual deviation of the maintenance operation on the overall quality.

[0143] To address this, this application further proposes a method for correcting node dissimilarity based on node deviations corresponding to key time-series nodes, yielding a corrected node dissimilarity. This method includes: obtaining the node deviation corresponding to the key time-series node and inputting it into a correction coefficient mapping function; calculating a correction coefficient positively correlated with the node deviation using the correction coefficient mapping function; multiplying the node dissimilarity by the correction coefficient to obtain an initial corrected node dissimilarity; obtaining the position weight of the key time-series node in the initial dissimilarity evolution path, where the position weight is pre-set based on the importance of the key time-series node in the maintenance sequence; and multiplying the initial corrected node dissimilarity by the position weight to obtain the corrected node dissimilarity.

[0144] Specifically, in the above technical solution, it is necessary to obtain the node deviation amount corresponding to the key timing nodes. Key timing nodes are maintenance operation steps where state deviation occurs during maintenance, and their node deviation amount represents the degree of deviation of the simulation state of that step relative to the standard performance benchmark. Obtaining this deviation amount is the basis for subsequent corrections, as it directly reflects the degree of anomaly in that step. This node deviation amount can be directly read from the pre-stored initial difference evolution path data, showing the deviation value associated with a specific key timing node, or it can be dynamically calculated and obtained by monitoring the difference between the simulation state corresponding to the key timing node and the standard performance benchmark in the benchmark instance in real time.

[0145] The node deviation is input into the correction coefficient mapping function, which calculates a correction coefficient positively correlated with the node deviation. The correction coefficient mapping function is a predefined mathematical model or lookup table that converts the node deviation into a correction coefficient. The function is designed to ensure that the larger the deviation, the larger the generated correction coefficient, thus achieving dynamic weighting of the difference. For example, the correction coefficient mapping function can be a linear function, such as correction coefficient = k × node deviation + b (where k > 0). It can also be a piecewise function or a nonlinear function (such as an exponential function or a sigmoid function), assigning different correction strengths based on different intervals or severity of the node deviation. Furthermore, it can be a mapping function trained based on a machine learning model (such as a regression model), which learns the relationship between the node deviation and the actual correction requirement through historical data.

[0146] Multiplying the node difference by the correction coefficient yields the initially corrected node difference. The node difference is the original difference value obtained by comparing the measured state segment with the simulated state segment. Multiplying it by the correction coefficient aims to preliminarily weight and adjust the original difference based on the actual deviation of the key timing node, making it more reflective of the true anomaly of that node. This is usually achieved through direct numerical multiplication, for example, initially corrected node difference = node difference × correction coefficient. If the node difference is a multidimensional vector, the correction coefficient can be a scalar, scaling each component of the vector proportionally. Alternatively, the correction coefficient itself can be a vector or matrix, allowing for matrix multiplication or element-wise multiplication.

[0147] Based on this, the positional weight of the key time-series node in the initial differential evolution path is obtained. The positional weight is a numerical value used to quantify the importance or influence of the key time-series node in the entire maintenance process. Some maintenance stages (such as core component installation and key parameter debugging) have a much greater impact on maintenance quality than others, thus requiring higher weights. This weight is pre-set, reflecting domain expert knowledge or historical data analysis results. Positional weights can be manually set by domain experts based on factors such as the criticality and risk level of the maintenance process. Alternatively, they can be automatically generated by analyzing historical maintenance data and using statistical methods (such as principal component analysis and sensitivity analysis) or machine learning models (such as decision trees and random forests) to assess the degree of influence of different maintenance stages on maintenance quality. They can also be calculated using multi-criteria decision-making methods (such as the analytic hierarchy process) based on indicators such as the complexity of the maintenance operation, the required skill level, and the potential scope of fault impact.

[0148] The corrected node difference is obtained by multiplying the initial corrected node difference by the location weight. This is the step in correcting the node difference. By multiplying the initially corrected difference by the location weight, the strategic importance of this critical timing node in the entire maintenance process is further considered, so that the correction result not only reflects the magnitude of the node's deviation but also its impact on the overall maintenance quality. This is usually achieved by directly performing numerical multiplication operations, for example, corrected node difference = initial corrected node difference × location weight. If the initial corrected node difference is a multidimensional vector, the location weight can be a scalar, scaling each component of the vector proportionally. Alternatively, the location weight itself can be a vector or matrix, performing matrix multiplication or element-wise multiplication.

[0149] Through the above technical solution, this application solves the problem of inaccurate correction caused by the failure to fully consider the deviation of key timing nodes and their importance in the maintenance sequence when calculating the global difference degree by introducing a refined correction mechanism for node difference degree. Specifically, by obtaining the node deviation corresponding to the key timing node and inputting it into the correction coefficient mapping function to calculate the correction coefficient that is positively correlated with the deviation, the initial correction of the node difference degree can dynamically and accurately reflect the actual degree of abnormality of the node, avoiding the errors caused by the fixed correction method. On this basis, a pre-set position weight based on the importance of the key timing node in the maintenance sequence is further introduced to perform a second weighting on the node difference degree after the initial correction. This dual correction mechanism not only considers the actual magnitude of the node deviation, but also incorporates the criticality and influence of the node in the entire maintenance process, ensuring that the corrected node difference degree can more comprehensively and realistically reflect the impact of the maintenance operation on the overall maintenance quality. In this way, the aggregated global difference degree will be more accurate and reliable, thereby improving the accuracy and credibility of the maintenance order evidence chain and providing more solid technical support for subsequent warranty determination, liability tracing, and dispute resolution.

[0150] In some of the solutions mentioned above in this application, the generation of difference evidence nodes is proposed to integrate difference evidence in the maintenance process. However, in the implementation process, there are shortcomings in how to accurately identify the root cause abnormal links and transmission paths that lead to global differences, so as to ensure the causal relationship of the evidence chain is clear and traceable.

[0151] To address this, this application further proposes a method for generating differential evidence nodes. This method generates differential evidence nodes containing the global differential degree, the differential evolution path chain, and anomaly identifiers based on the multiple related evidence pairs and the initial differential evolution paths corresponding to each related evidence pair. Specifically, the method includes the following steps: The system concatenates the initial difference evolution paths corresponding to various related evidence pairs according to the maintenance timeline, generating an initial difference evolution path chain that runs through the entire maintenance process. Multiple path nodes and their corresponding node deviations are then extracted from this initial difference evolution path chain. This step aims to integrate the difference evolution information of various local stages in the maintenance process into a continuous, global difference evolution trajectory. For example, the system can assign a start and end timestamp to each initial difference evolution path, sort all paths based on these timestamps, and then logically connect or merge the ends of adjacent paths with the start of the next path to form a continuous time series. Path nodes can be defined as the start point, end point, or key time point within each initial difference evolution path, and node deviations are directly obtained from the corresponding initial difference evolution path. Alternatively, since the maintenance process typically consists of a series of discrete operational stages, the system can associate each initial difference evolution path with its corresponding operational stage identifier. During concatenation, a directed graph can be constructed, where nodes represent operational stages and edges represent the temporal relationships between operations. Each operational node can aggregate its internal initial differential evolution path information to form a higher-level path node, and calculate the comprehensive node deviation of that node, for example, by taking the average, maximum, or weighted sum.

[0152] The process involves determining the global deviation fit between the initial differential evolution path chain and the global difference degree, and then selecting candidate path nodes from among the multiple path nodes that exhibit a deviation propagation correlation with the global difference degree based on this global deviation fit. The purpose of this step is to quantify the relationship between detailed, time-series deviation information (initial differential evolution path chain) and the overall system and deviation (global difference degree), thereby identifying the specific links that contribute the most to the global difference degree. For example, statistical methods such as Pearson correlation coefficient, cosine similarity, or dynamic time warping (DTW) algorithms can be used to measure the similarity or correlation between the overall trend or key features of the initial differential evolution path chain and the global difference degree. The global deviation fit can be defined as a weighted combination of these correlation indicators. When selecting candidate path nodes, a threshold can be set; for example, when the deviation amount or its trend of change of a certain path node is consistent with the direction of change of the global difference degree and exceeds a certain level, it is marked as a candidate node. Another approach is to train a machine learning model, such as a regression or classification model, taking into account the features of the initial differential evolution path chain, such as node bias and rate of change, and outputting the predicted global differential degree. The global bias fit can be defined as the error or degree of matching between the model's predicted value and the actual global differential degree. Then, through model interpretability analysis, such as feature importance analysis or LIME / SHAP values, the path nodes that contribute the most to the global differential degree are identified as candidate nodes.

[0153] A causal tracing analysis is performed on the candidate path nodes. Based on the temporal relationship and deviation propagation direction between the candidate path nodes, the root cause node and the transmission node leading to the global difference are identified, and corresponding anomalous link identifiers are generated for the root cause node and the transmission node, respectively. This step is the core causal analysis step, aiming to distinguish the initial cause (root cause node) and subsequent influence or propagation (transmission node) from potential contributors, thereby providing a clear causal chain for the observed global deviation. For example, a series of causal inference rules can be preset, such as "If the deviation of node A increases before node B, and the deviation of node B also increases accordingly, then A may be the root cause or transmission node of B." Combining the experience knowledge base of maintenance experts, propagation patterns of different types of deviations are defined. By traversing the candidate path nodes and applying these rules, nodes without preceding causes are identified as root cause nodes, and nodes that pass the deviation from the root cause node to subsequent links are identified as transmission nodes. The anomalous link identifier can be a predefined code or descriptive text. Furthermore, candidate path nodes and their temporal relationships can be constructed into a directed acyclic graph (DAG). Using Bayesian networks or other causal inference algorithms, the conditional dependencies and information flow between nodes can be analyzed to infer causal chains. For example, the causal influence or information entropy of each node can be calculated, identifying the node with the greatest influence and no preceding causal node as the root node, and those nodes that transmit influence along the causal chain as transmission nodes. Anomaly markers can be automatically generated based on the node's causal role and deviation type.

[0154] The initial difference evolution path chain, the global difference degree, and the anomaly identifiers of the root node and the transmission node are associated and encapsulated to generate the difference evidence node. This step integrates all key information related to the deviation and its causes into a structured "difference evidence node" as a comprehensive, self-contained record of the maintenance quality assessment. For example, all relevant information can be encapsulated as a JSON object, XML document, or database record. The JSON object can contain fields such as "global_difference":value, "difference_evolution_path_chain":[node1, node2, ...], "root_anomaly_identifiers":[id1, id2, ...], "transmission_anomaly_identifiers":[id3, id4, ...]. Each node or identifier can also be nested structured data. To enhance the immutability of the evidence, the above structured data can be hashed, and the hash value can be stored along with the original data (or its reference). The difference evidence node can contain a pointer to the original data storage location and a hash digest containing all key information, ensuring data integrity and verifiability.

[0155] Through the above technical solution, this application can systematically integrate and deeply analyze multi-source difference data generated during the maintenance process, thereby accurately identifying the root causes of global differences in maintenance quality and their propagation paths in the maintenance process. Specifically, by connecting the initial difference evolution paths corresponding to each pair of related evidence into an initial difference evolution path chain that runs through the entire maintenance process according to the maintenance timeline, and extracting path nodes and their deviations, a data foundation is laid for subsequent refined analysis. On this basis, by determining the degree of global deviation consistency between this path chain and the global difference, and thereby screening candidate path nodes that are associated with the propagation of deviations in the global difference, the scope of analysis is effectively focused, avoiding the waste of resources on irrelevant links. Furthermore, by performing causal tracing analysis on these candidate path nodes, combined with temporal relationships and deviation propagation directions, the root cause nodes and transmission nodes leading to global differences can be accurately identified, and corresponding abnormal link identifiers can be generated for them, making the causes and scope of impact of maintenance quality problems clear at a glance. By associating and encapsulating the initial difference evolution path chain, global difference degree, and abnormal link identifiers of root node and transmission node, structured and traceable difference evidence nodes are generated, which greatly improves the clarity and credibility of the causal relationship of the repair order evidence chain and provides strong technical support for subsequent warranty determination, liability tracing and dispute resolution.

[0156] In some of the embodiments described above in this application, candidate path nodes are selected based on global deviation consistency to identify nodes associated with deviation propagation. However, in its implementation, the selection may not be accurate enough and may not be able to effectively distinguish the key nodes that truly cause global differences, thereby affecting the accuracy of subsequent causal tracing and leading to errors in identifying maintenance anomalies or low efficiency.

[0157] In response, this application further proposes a method for selecting candidate path nodes from multiple path nodes that are associated with the global difference degree through deviation propagation based on the global deviation consistency. This method includes: Obtain the node deviation value corresponding to each path node, compare the node deviation value with the global deviation matching degree, and filter out the path nodes whose node deviation value is greater than the global deviation matching degree as the first candidate node set.

[0158] The path nodes in the first candidate node set are sorted according to the maintenance time sequence, and the gradient of the node deviation change between two adjacent path nodes is calculated.

[0159] Based on the gradient of the node deviation change, path nodes whose node deviation change gradient exceeds a preset gradient threshold are selected from the first candidate node set and used as the second candidate node set.

[0160] The path nodes in the second set of candidate nodes are selected as candidate path nodes that are associated with the deviation propagation of the global difference.

[0161] Specifically, in the step of obtaining the node deviation for each path node and comparing it with the global deviation consistency, selecting path nodes with node deviations greater than the global deviation consistency as the first candidate node set, this step aims to initially identify path nodes exhibiting deviations during maintenance. By comparing with the global deviation consistency, nodes with small deviations and little impact on the global difference can be effectively filtered out, thus focusing subsequent analysis on potential anomalies. The node deviation can be a scalar value representing the overall deviation between the simulation state at a node and the standard performance benchmark, for example, obtained by calculating the norm or weighted average of a multidimensional deviation vector. The global deviation consistency can be a normalized value or percentage reflecting the degree of matching between the initial difference evolution path chain and the global difference. In practice, numerical comparisons can be performed directly. Furthermore, if the node deviation is a multidimensional vector, the comparison can be made by using the vector norm (such as Euclidean distance or Manhattan distance) to compare with the global deviation consistency, or by weighted summing of the vector's dimensions and comparing with the global deviation consistency.

[0162] In the steps of sorting the path nodes in the first candidate node set according to the maintenance time sequence and calculating the gradient of the node deviation change between adjacent path nodes, sorting is to maintain the temporal logic of the maintenance process and ensure the correctness of subsequent gradient calculations. Calculating the gradient change is crucial for identifying the trend and rate of change of the deviation over time, which is essential for determining whether the deviation is accumulating, worsening, or suddenly appearing. Sorting can employ standard timestamp sorting algorithms, such as bubble sort, quicksort, or merge sort, arranging the path nodes in ascending order based on the timestamp information contained within them. The gradient of the node deviation change can be obtained by calculating the ratio of the difference in deviation between adjacent path nodes to the time interval. For example, for node A and node B (B follows A), the gradient can be expressed as (deviation B - deviation A) / (timestamp B - timestamp A). ​​If the node deviation is multidimensional, the gradient can be calculated separately for each dimension, or the norm change gradient of the deviation vector can be calculated.

[0163] In the step of selecting path nodes from the first candidate node set whose deviation change gradient exceeds a preset gradient threshold, based on the node deviation change gradient, this step aims to further refine the candidate nodes, focusing on those nodes with drastic deviation changes. Dramatic changes often indicate a potential problem in a maintenance operation or an abnormal event that has led to the rapid accumulation or spread of deviations. The preset gradient threshold can be an empirical value, set through historical maintenance data analysis or expert knowledge; for example, if the gradient value exceeds a certain percentage change rate, it is considered a drastic change. During screening, the calculated gradient value (which can be an absolute value representing the magnitude of change) of each node can be directly compared with the preset gradient threshold. If the gradient value is greater than the threshold, the node is selected. Furthermore, the gradient threshold can also be dynamically adjusted, for example, adaptively adjusted based on the overall deviation level of the current maintenance order or the characteristics of the heavy truck model, to improve the flexibility and accuracy of the screening.

[0164] The path nodes in this second set of candidate nodes are considered as candidate path nodes that are associated with the propagation of the global dissimilarity. This is the step in determining candidate nodes that are associated with the propagation of the global dissimilarity.

[0165] Through the above technical solution, this application adopts a two-stage screening mechanism. Initial screening is performed based on the consistency between node deviation and global deviation, focusing on nodes with deviations and avoiding interference from irrelevant nodes to ensure high relevance of the screening starting point. Secondary screening is performed by calculating the gradient of node deviation changes and setting a preset gradient threshold, further refining key nodes with drastic deviation changes. This screening method, combining deviation magnitude and rate of change, can more accurately identify candidate path nodes that are truly associated with the propagation of deviations in relation to global differences, thereby improving the accuracy and efficiency of subsequent causal tracing. This helps to more accurately locate abnormal links in heavy-duty truck maintenance scenarios, providing more reliable evidence for maintenance quality assessment, responsibility tracing, and dispute resolution.

[0166] In some of the embodiments described above in this application, a causal tracing analysis of candidate path nodes is proposed to identify the source of deviation. However, in the implementation process, there may be problems such as inaccurate identification or inability to effectively distinguish between root nodes and transmission nodes, resulting in insufficient causal proof of the evidence chain and affecting the accuracy of maintenance responsibility tracing.

[0167] In response, this application further proposes a causal tracing analysis of the aforementioned candidate path nodes. Based on the temporal relationship and deviation propagation direction among the candidate path nodes, the root cause node and propagation node leading to the global dissimilarity are identified. This analysis process includes: Based on the order of the candidate path nodes in the maintenance sequence, determine the set of preceding nodes and the set of succeeding nodes corresponding to each candidate path node.

[0168] Obtain the node deviation value corresponding to each candidate path node, and determine the deviation propagation direction based on the deviation change trend between the node deviation values ​​of each preceding node in the preceding node set and the node deviation value of the current candidate path node. The deviation propagation direction points in the direction where the node deviation value increases.

[0169] Candidate path nodes that have no preceding nodes or whose node deviations for all preceding nodes are less than a preset benchmark threshold are selected from the candidate path nodes and used as the root node.

[0170] Other candidate path nodes besides the root node are identified as undetermined propagation nodes. From these undetermined propagation nodes, those whose node deviation can be traced back to the root node along the deviation propagation direction are selected as propagation nodes.

[0171] Specifically, in determining the set of preceding and succeeding nodes for each candidate path node, this step aims to establish the temporal dependencies between the candidate path nodes, providing a temporal basis for subsequent causal tracing. By clearly defining the "preceding" and "following" nodes in the maintenance sequence for each node, a directed graph or sequence can be constructed to analyze the propagation path of deviations. For example, by traversing the list of candidate path nodes and identifying the one or more preceding nodes on the timeline as preceding nodes and the one or more following nodes as succeeding nodes based on each node's timestamp or index position in the maintenance process, a maintenance process graph can be pre-constructed, containing all possible maintenance operation steps and their standard temporal relationships. During causal tracing, candidate path nodes are mapped to corresponding steps in this graph, and then their sets of preceding and succeeding nodes are determined based on the graph's topology.

[0172] The core of this step, which involves obtaining the node deviation for each candidate path node and determining the deviation propagation direction based on the trend of deviation changes between the node deviations of each preceding node in the preceding node set and the node deviation of the current candidate path node (pointing towards an increase in node deviation), lies in quantifying the degree of state anomaly (node ​​deviation) of each candidate path node and dynamically identifying how deviations evolve and propagate in the maintenance process by comparing the deviation changes between adjacent nodes. Determining the deviation propagation direction is crucial for causal analysis, indicating where the anomaly begins to intensify. For example, the node deviation can be a scalar value, obtained by calculating the Euclidean distance, Manhattan distance, or weighted distance between the simulated state and the standard performance baseline. The deviation change trend can be obtained by calculating the difference or ratio between the deviation of the current node and the deviation of its preceding nodes. If the difference or ratio is positive and exceeds a certain threshold, the deviation is considered to be increasing, thus determining the propagation direction. Alternatively, the node deviation can also be a multi-dimensional vector representing deviations in different performance dimensions. The deviation change trend can be determined by comparing the changes in the deviation vector of the current node and the deviation vector of the preceding nodes in each dimension. When the deviation in one or more dimensions increases, it is determined to be the direction of deviation propagation.

[0173] When selecting candidate path nodes from among candidate path nodes that have no preceding nodes or whose node deviations for all preceding nodes are less than a preset baseline threshold as the root source node, this step aims to identify the initial cause or starting point leading to the global variability. The root source node is the point where the deviation first appears or deviates from the normal range; it is independent of previous anomalies, or the severity of previous anomalies is insufficient to constitute propagation. For example, the system can identify all candidate path nodes without preceding nodes, which are naturally potential root sources. Then, for candidate path nodes with preceding nodes, the system checks the node deviations of all its preceding nodes. If the deviations of all preceding nodes are below a preset baseline threshold (e.g., indicating slight fluctuations or an acceptable range), the current node is considered the node where the deviation first appears, i.e., the root source node. Alternatively, statistical methods can be used to perform a distribution analysis of the node deviations of all candidate path nodes to identify those nodes with deviations above the average level and whose preceding node deviations are within the normal fluctuation range.

[0174] The process involves identifying candidate path nodes other than the root cause node as potential propagation nodes, and then selecting those nodes from this list whose deviation can be traced back to the root cause node along the deviation propagation direction. This step aims to identify those links in the maintenance process that further amplify or propagate the root cause deviation. Propagation nodes are key intermediate links in the evolution of deviation from the root cause node to the global difference; they are important components of the deviation propagation path. For example, all non-root cause candidate path nodes are marked as potential propagation nodes. Then, for each potential propagation node, the system traces backward along the previously determined deviation propagation direction (i.e., the direction of increasing deviation). If a series of consecutive nodes with increasing deviation can be traced back to an identified root cause node, then the potential propagation node is confirmed as a propagation node. Alternatively, a deviation propagation graph can be constructed, where nodes are candidate path nodes and edges represent deviation propagation directions. Starting from each root node, perform a depth-first search or breadth-first search along the directed edges. All non-root nodes that can be reached from the root node are marked as transit nodes.

[0175] Through the above technical solution, this application solves the problems of inaccurate identification and unclear distinction between root cause nodes and transmission nodes when identifying deviation sources by performing refined causal tracing analysis on candidate path nodes, thereby improving the accuracy of maintenance responsibility tracing and the causal relationship proving power of the evidence chain. Specifically, by determining the set of preceding and succeeding nodes for each node according to the chronological order of candidate path nodes in the maintenance sequence, this application lays a solid temporal foundation for the analysis of deviation propagation paths. This ensures the logical coherence of causal analysis, avoids misjudgments caused by temporal confusion, and makes the subsequent determination of deviation propagation direction more accurate. On this basis, by obtaining the node deviation amount corresponding to each candidate path node and determining the deviation propagation direction (pointing to the direction of increasing deviation amount) according to the deviation change trend between preceding nodes and the current node, this application can dynamically capture the evolution trajectory of deviation in the maintenance process. This dynamic comparison mechanism enables the system to accurately identify the aggravation point and propagation path of deviation, providing a key basis for the subsequent identification of root cause and transmission nodes. Furthermore, by selecting nodes from candidate path nodes that have no preceding nodes or whose node deviations are all less than a preset benchmark threshold as root source nodes, this application can accurately locate the initial source of the deviation. This effectively avoids misjudging nodes that are merely part of the deviation propagation process as the root source, ensuring accurate identification of the problem's starting point. Other candidate path nodes besides the root source node are identified as undetermined propagation nodes, and nodes whose node deviations can be traced back to the root source node along the deviation propagation direction are selected as propagation nodes. This application constructs a complete causal chain from the root source to the global difference. This not only enhances the completeness and verifiability of the evidence chain but also enables maintenance personnel to clearly understand how the deviation gradually spreads from the initial stage and affects maintenance quality, thus providing strong data support for accurate responsibility determination and process improvement. Through the above technical solutions, this application can provide more accurate and reliable causal evidence for the evidence map of maintenance orders, enabling rapid and accurate location of specific abnormal operation stages when global differences occur, greatly improving the efficiency and fairness of heavy truck maintenance warranty determination, responsibility tracing, and dispute resolution.

[0176] In some of the embodiments described above in this application, a model training evidence chain is proposed to integrate federated learning data into the evidence chain. However, in its implementation, there are shortcomings in how to ensure the credible correlation between model parameters, training data and hyperparameters, and how to organize this evidence to support the aggregation process of federated learning. Specifically, the model training process lacks transparency and traceability, making it difficult to verify the evidence chain in terms of data source and training configuration, thus affecting the overall credibility.

[0177] To address this, this application further proposes a method for obtaining the model training evidence chain generated by each repair station during the federated learning process. This method includes: after each round of local training in the federated learning process, obtaining the model parameters generated by the local training at each repair station, and recording the training data hash value and training process hyperparameters corresponding to the model parameters for each repair station. These hyperparameters include the training round and the learning rate. A hash calculation is performed on the model parameters of each repair station to obtain the parameter hash value corresponding to that model parameter. This parameter hash value is then associated and encapsulated with the training data hash value and the training process hyperparameters to generate a generation proof corresponding to each repair station. The model parameters, parameter hash values, and generation proofs corresponding to each repair station are organized according to the training round to obtain a local training evidence sub-chain for each repair station. The local training evidence sub-chains corresponding to each repair station are synchronously associated according to the aggregation round of the federated learning process to obtain the model training evidence chain.

[0178] Specifically, after each round of local training in federated learning, the system retrieves the model parameters generated by each repair station's local training. These model parameters refer to the internal configuration data such as weights and biases learned by the locally trained models (e.g., deep neural networks, support vector machines) at each repair station during the federated learning process. These parameters are typically obtained after local training iterations through model serialization or parameter extraction interfaces; for example, the model parameters can be exported in JSON, Protobuf, or binary file formats. Simultaneously, the system records the training data hash value corresponding to these model parameters and the training process hyperparameters for each repair station. The training data hash value is a unique identifier obtained by hashing the original data used for local model training. Its purpose is to ensure the integrity and immutability of the training data and provide unique proof of the data source. The generation method can use standard hash algorithms such as SHA-256 and MD5 to calculate the entire content of the training dataset or its metadata. The training process hyperparameters refer to parameters pre-set during model training that are not learned by the model itself, such as training epochs, learning rate, batch size, and optimizer type. Recording these hyperparameters aims to provide transparency about the training environment and configuration, ensuring the reproducibility and traceability of the training process.

[0179] For each repair station, the model parameters are hashed to obtain the corresponding parameter hash value. Hash calculation of model parameters refers to performing a hash operation on the locally trained model parameters to obtain a fixed-length, unique parameter hash value. This step aims to provide a tamper-proof digital fingerprint for the model parameters themselves, ensuring that the model parameters have not been maliciously modified during transmission or storage. Commonly used hash algorithms include SHA-256 and SHA-3. This parameter hash value is then associated and encapsulated with the training data hash value and the training process hyperparameters to generate a generation certificate for each repair station. Association and encapsulation refers to logically binding the parameter hash value, training data hash value, and training process hyperparameters together to form a unified data structure or record. This encapsulation can be achieved by constructing a JSON object, an XML document, or a custom data structure, ensuring that this information is logically closely related and together constitutes a complete evidentiary unit. This generation certificate, after association and encapsulation, serves as a comprehensive and credible credential for each repair station to complete one round of local training. It contains the fingerprints of the model parameters, the fingerprints of the training data, and detailed training configuration information, providing a foundation for subsequent auditing and verification.

[0180] Based on this, the model parameters, parameter hash values, and generated proofs corresponding to each repair station are organized according to the training rounds, resulting in a local training evidence subchain for each repair station. Organizing by training round means arranging and storing the model parameters, parameter hash values, and generated proofs generated by each repair station in a series of training rounds in an ordered manner according to their corresponding training rounds (e.g., round 1, round 2... round N). This organization method can adopt a linked list structure, array, or database records to ensure that the training history of each repair station has a clear temporal sequence, facilitating traceability and auditing. This local training evidence subchain refers to the sequence of evidence (model parameters, parameter hash values, generated proofs) generated by all local training rounds of a repair station from the start of federated learning to the current moment, linked together in chronological order. It represents the complete and verifiable training history of a single repair station.

[0181] The local training evidence subchains corresponding to each maintenance station are synchronously associated according to the aggregation rounds of federated learning to obtain the model training evidence chain. Synchronous association refers to integrating or linking the latest evidence units corresponding to that aggregation round from the local training evidence subchains of all participating maintenance stations in each aggregation round. This association can be achieved by sharing aggregation round identifiers, timestamps, or aggregation hash values, ensuring the integrity of evidence in the global aggregation process. The model training evidence chain is a complete evidence record of the entire federated learning process; it contains the local training history of all participating maintenance stations and the association information of this history across different aggregation rounds.

[0182] Through the above technical solution, this application can solve the problem of lack of transparency and traceability in model training during federated learning. By promptly acquiring and recording model parameters, training data hash values, and training process hyperparameters after each round of local training, the integrity of training results and the transparency of training configuration are ensured. Hash calculations are performed on model parameters and associated with training data hash values ​​and hyperparameters to generate tamper-proof proofs, thereby tightly binding scattered evidence points and greatly improving the credibility of the evidence. Furthermore, these pieces of evidence are organized into local training evidence sub-chains according to training rounds, providing each repair station with a clear and auditable training history. By synchronously associating the local training evidence sub-chains of each repair station according to the aggregation rounds of federated learning, a model training evidence chain that runs through the entire federated learning process is constructed, enabling the global aggregation process of federated learning to have comprehensive evidence integration capabilities. This not only enhances the transparency and traceability of the model training stage in the repair order evidence graph but also provides a solid data foundation for subsequent repair effect verification, parameter source tracing, and causal relationship proof, thereby improving the judicial probative value and credibility of the entire repair order evidence chain.

[0183] Among the solutions proposed in this application, an evidence graph is constructed to integrate the entire process evidence of maintenance orders. However, in its implementation, there are shortcomings in how to specifically implement the logical connections between evidence nodes to ensure consistency between actual operations and simulation predictions, the connection between data sources for model training and maintenance operations, and the proof of causal relationships between global differences and specific links. Specifically, existing evidence graph construction lacks detailed processing of evidence unit mapping, dynamic establishment mechanisms for edge relationships, and precise location of abnormal links, making it difficult for the evidence chain to effectively support quality assurance determination and liability tracing.

[0184] In response, this application further proposes a method for constructing an evidence graph of the repair order based on the training of the evidence chain, the multiple related evidence pairs, and the differential evidence nodes using the model. This method includes: The physical operation evidence units and twin simulation evidence units in the multiple associated evidence pairs are mapped to physical operation evidence nodes and twin simulation evidence nodes, respectively. The difference evidence node is also mapped to a difference evidence node. Simultaneously, the local model parameters and corresponding generated proofs of each repair station are extracted from the model training evidence chain as model training evidence nodes, constructing an evidence node set. The physical operation evidence nodes and twin simulation evidence nodes are the basic components of the evidence graph, representing the actual physical operation events and the simulated operation events of the digital twin system during the repair process, respectively. They can be structured data objects containing information such as timestamps, operation types, and operation parameters. For example, a physical operation evidence node can record the start time, end time, operator ID, and tools used for the operation of "changing the oil filter." A twin simulation evidence node records the performance changes predicted by the digital twin model and the simulation time when simulating "changing the oil filter." The difference evidence node encapsulates information such as the global difference degree after repair, the difference evolution path chain, and the abnormal link identifier. It is crucial for assessing repair quality and locating anomalies. It can be a composite data structure, for example, containing a numerical value representing the overall repair quality deviation, a sequence of data describing how the deviation evolves over time, and one or more identifiers pointing to specific abnormal operation links. The model training evidence node represents the results and proof of the local model training at each repair station during the federated learning process. It can contain the hash values ​​of local model parameters, training data hash values, training hyperparameters (such as training epochs, learning rate), and cryptographic proofs generated from this information. For example, a model training evidence node can record the unique identifier of the model parameters and the tamper-proof proof of the training process after a repair station trains a fault diagnosis model for a specific component. The purpose of constructing the evidence node set is to unify and abstract the evidence units from different sources and of different types into nodes in the graph, laying the foundation for establishing the relationships (edges) between nodes. This can be achieved by defining a unified node data structure or interface. For example, each node contains metadata such as a unique ID, type identifier, content hash, and timestamp.

[0185] This paper analyzes the semantic mapping relationship between physical operation evidence units and twin simulation evidence units in multiple related evidence pairs. Based on this semantic mapping relationship, corroborating edges are established between the corresponding physical operation evidence nodes and twin simulation evidence nodes. These corroborating edges are used to represent the logical consistency between the actual operation and the simulation prediction in the same maintenance operation. The semantic mapping relationship refers to the correspondence between physical operation evidence units and twin simulation evidence units in terms of content, meaning, and function. For example, "replacing the oil filter" in the physical operation evidence unit is semantically equivalent to the simulated "oil filter replacement" operation in the twin simulation evidence unit. This mapping relationship can be analyzed using a predefined rule base, ontology matching, or natural language processing techniques based on machine learning. Establishing corroborating edges is a directed or undirected edge connecting physical operation evidence nodes and twin simulation evidence nodes. Its function is to explicitly indicate that the events represented by these two nodes are semantically mutually corroborating and corresponding. For example, when the system identifies that the physical operation "replacing brake pads" and the simulation operation "brake pad replacement simulation" are highly matched in terms of time and operation type, it establishes a confirmation edge between the two corresponding nodes, indicating that the actual operation and the simulation prediction are logically consistent in this step.

[0186] The generation proof in the model training evidence chain is matched with the physical operation evidence units at corresponding time positions in the multiple associated evidence pairs. Based on the matching results, a tracing edge is established between the model training evidence node and the successfully matched physical operation evidence node. This tracing edge is used to characterize the data source association between the model training process and the specific maintenance operation. The generation proof is a key component in the model training evidence chain, containing information such as model parameters, training data hash values, and training hyperparameters. The matching process involves comparing this proof information with the physical operation evidence units to determine whether a particular physical operation evidence unit (e.g., sensor data or diagnostic results) was used as a data source for a specific model training or is associated with that model training process. Matching can be achieved by comparing timestamps, data types, data source identifiers, or through hash value verification. Establishing a tracing edge is a directed edge connecting the model training evidence node and the physical operation evidence node. It explicitly indicates that the data or information of a certain physical operation evidence node was used or influenced the model training process represented by a certain model training evidence node. For example, if a diagnostic model at a repair shop uses a specific batch of sensor data during training, and this sensor data corresponds to a physical operation evidence unit, then a traceability edge can be established between the corresponding model training evidence node and the physical operation evidence node, indicating that the physical operation data is one of the data sources for model training.

[0187] Based on the difference evolution path chain contained in the difference evidence node, the target physical operation evidence node corresponding to the abnormal link identifier is identified along the difference evolution path chain. A causal edge is established between the difference evidence node and the target physical operation evidence node. This causal edge represents the causal relationship between the global difference degree and the specific maintenance operation link, thereby generating an evidence graph of the maintenance order containing the evidence node set, the corroborating edge, the tracing edge, and the causal edge. The difference evolution path chain is one of the core components of the difference evidence node. It is a sequence that records the evolution of the deviation of the simulation state of each link in the maintenance process relative to the standard performance benchmark. For example, it may be a time series, with each point containing a timestamp and a performance deviation vector detected at that time point. The process of identifying the target physical operation evidence node involves searching for physical operation evidence nodes in the constructed physical operation evidence node set that correspond to these abnormal links in terms of time, type, or content, based on the abnormal link identifier indicated in the difference evolution path chain. For example, if the difference evolution path chain indicates an abnormality in the "replace brake fluid" link, the system will search for nodes related to "replace brake fluid" in the physical operation evidence node set. Establishing a causal edge is a directed edge connecting a difference evidence node to a target physical operation evidence node. It explicitly indicates that a specific physical operation is the cause of or influences the global difference. For example, if a difference evidence node shows a deviation in brake performance after repair, and the difference evolution path chain traces back to the physical operation of "improper brake pad installation," then a causal edge can be established between the difference evidence node and the physical operation evidence node corresponding to "improper brake pad installation," indicating that this operation is the direct cause of the performance deviation.

[0188] Through the aforementioned technical solution, this application constructs a comprehensive set of evidence nodes by uniformly mapping physical operation evidence units, twin simulation evidence units, difference evidence nodes, and local model parameters and generated proofs in the model training evidence chain to different types of evidence nodes, laying the foundation for subsequent correlation analysis. Based on this, by analyzing the semantic mapping relationship between physical operation evidence units and twin simulation evidence units, and establishing corroboration edges, the logical consistency between actual operations and simulation predictions in the same maintenance operation step is ensured, effectively avoiding data silos and achieving mutual corroboration of multi-source data. Simultaneously, by matching the generated proofs in the model training evidence chain with the physical operation evidence units at corresponding temporal positions, and establishing tracing edges, the data source correlation between the model training process and specific maintenance operations is clarified, making the source of model parameters traceable and enhancing model credibility. Furthermore, by identifying the target physical operation evidence nodes corresponding to abnormal steps based on the difference evolution path chain and establishing causal edges, the causal relationship between global difference and specific maintenance operation steps is clearly revealed, thereby accurately locating the root cause of maintenance quality problems and greatly enhancing the ability to trace responsibility and diagnose faults. Overall, by introducing a graph structure and defining three types of key edges, this application elevates previously scattered or only preliminarily related evidence information into a highly interconnected, interpretable, and traceable evidence system, greatly enhancing the integrity and probative value of the evidence chain. This provides a powerful and judicially admissible evidence chain for quality assurance determination, liability tracing, and dispute resolution.

[0189] In response, this application further proposes a method for identifying target physical operation evidence nodes corresponding to abnormal links by following the differential evolution path chain contained in the differential evidence nodes. This method addresses the problem that when identifying specific target physical operation evidence nodes, there may be a lack of accurate screening mechanisms and location association methods based on differential evolution path chains, leading to inaccurate or inefficient anomaly localization and an inability to effectively trace back to the actual maintenance operation.

[0190] Specifically, the method includes: extracting multiple path nodes and the node deviation amount corresponding to each path node from the differential evolution path chain; selecting path nodes whose node deviation amounts exceed a preset anomaly threshold from the multiple path nodes as anomalous path nodes; obtaining the location information of each anomalous path node in the maintenance timeline, and retrieving a physical operation evidence unit that matches the location information from multiple associated evidence pairs based on the location information; and using the physical operation evidence node corresponding to the retrieved physical operation evidence unit as the target physical operation evidence node corresponding to the anomalous link identifier.

[0191] The purpose of this study is to extract multiple path nodes and their corresponding node deviations from the differential evolution path chain, aiming to obtain detailed deviation information for each key point in the chain. This differential evolution path chain is a sequential record of deviation evolution throughout the maintenance process. Each path node represents a specific state or operational step in the maintenance timeline and contains the node deviation for that step. Extracting this information is fundamental for subsequent accurate analysis and anomaly localization. For example, the data structure of the differential evolution path chain, such as a linked list or array, can be traversed to access each path node sequentially and parse its node identifier, timestamp, and corresponding node deviation. Alternatively, if the differential evolution path chain is stored in graph form, a graph traversal algorithm (such as depth-first search or breadth-first search) can be used to start from the initial node, sequentially access all path nodes along the path direction, and read the node deviation stored in its attributes when accessing each node.

[0192] The process of selecting path nodes from multiple path nodes whose deviation exceeds a preset anomaly threshold as abnormal path nodes aims to focus on critical aspects that may lead to maintenance quality issues. By setting a preset anomaly threshold, normal and acceptable fluctuations during the maintenance process can be effectively filtered out, thus concentrating analytical resources on truly problematic anomalies. For example, a fixed value can be set as the preset anomaly threshold; when the node deviation exceeds a certain percentage (e.g., 5%) or absolute value (e.g., a specific unit deviation of a performance parameter), the node is considered an abnormal path node. The system iterates through all path nodes, directly comparing the deviation of each node. Furthermore, the preset anomaly threshold can be dynamically adjusted based on historical maintenance data or the performance fluctuation range of a specific heavy-duty truck model. For instance, a statistical threshold (e.g., mean plus two standard deviations) can be calculated based on the deviation distribution of historical normal maintenance, or a machine learning model can be used to dynamically determine the threshold based on contextual information.

[0193] Obtaining the location information of each anomalous path node in the maintenance timeline, and retrieving the physical operation evidence unit matching that location information from multiple related evidence pairs, is a crucial step in linking anomalies in the digital twin space with actual operations in the physical world. Each anomalous path node carries its specific location information in the maintenance timeline (e.g., a timestamp or operation sequence number), which allows for precise backtracking to the corresponding physical operation evidence unit. For example, each anomalous path node typically contains a timestamp or time interval. The system can use this timestamp as a search key to find physical operation evidence units in the database or index structure storing all related evidence pairs whose timestamps are close to or fall within the time interval of the anomalous path node. To improve efficiency, physical operation evidence units can be pre-sorted by timestamp or a time index can be created. Alternatively, if the operation in the maintenance process has a clear sequence number or stage identifier, the location information of the anomalous path node can contain the corresponding operation sequence number. The system can then directly retrieve physical operation evidence units with the same operation sequence number from the related evidence pairs based on this sequence number.

[0194] The physical operation evidence nodes corresponding to the retrieved physical operation evidence units are used as target physical operation evidence nodes corresponding to the anomaly identifier. This aims to directly map the identified anomalies to specific physical operation nodes in the evidence graph, thus providing clear direction for subsequent causal tracing and responsibility determination. Once a physical operation evidence unit is retrieved, its corresponding physical operation evidence node in the evidence graph can be directly marked or referenced as the target physical operation evidence node corresponding to the anomaly identifier. This can be achieved by storing a reference or ID pointing to the physical operation evidence node in the difference evidence node. Alternatively, a new association edge (e.g., named "anomaly association" or "causal association") can be established between the difference evidence node and the retrieved physical operation evidence node in the evidence graph, explicitly indicating that the physical operation evidence node is direct operational evidence leading to or reflecting the anomaly.

[0195] Through the aforementioned technical solution, this application can accurately filter out abnormal path nodes with deviations from a comprehensive chain of differential evolution paths by setting anomaly thresholds, avoiding excessive focus on non-critical fluctuations and thus improving the accuracy and efficiency of anomaly localization. Furthermore, by utilizing the positional information of abnormal path nodes in the maintenance timeline, this application can efficiently and accurately retrieve matching physical operation evidence units from multiple related evidence pairs, achieving a close correlation between deviation information in the digital twin space and actual operations in the physical world. Mapping the retrieved physical operation evidence units to target physical operation evidence nodes allows for the establishment of a clear causal relationship between differential evidence nodes and specific physical operation evidence nodes, providing precise identification of abnormal links and tracing basis for constructing the evidence map of maintenance orders. This greatly enhances the completeness and credibility of the evidence map, enabling more detailed and robust evidentiary support for warranty determination, liability tracing, and dispute resolution in heavy-duty truck maintenance scenarios.

[0196] In some of the solutions mentioned above in this application, evidence graphs are stored to ensure the integrity and immutability of the evidence chain. However, in the process of implementation, the nodes of the evidence graph may be tampered with, and the temporal relationship between the nodes may be disrupted, resulting in a decrease in the credibility of the evidence chain.

[0197] To address this, this application proposes a method for storing evidence graphs using a triple hash chain. Specifically, it includes: generating a physical operation evidence chain based on the content of each physical operation evidence node in the evidence graph and the temporal relationship between nodes; each physical operation evidence node in the physical operation evidence chain containing its content hash and a pointer hash pointing to the previous physical operation evidence node; generating a twin simulation evidence chain based on the content of each twin simulation evidence node in the evidence graph and the temporal relationship between nodes; each twin simulation evidence node in the twin simulation evidence chain containing its content hash and a pointer hash pointing to the previous twin simulation evidence node; generating a difference evidence chain based on the content of each difference evidence node in the evidence graph and the temporal relationship between nodes; and linking the root hash of the physical operation evidence chain, the root hash of the twin simulation evidence chain, and the root hash of the difference evidence chain to generate the overall hash of the evidence graph, and writing this overall hash into the blockchain.

[0198] Specifically, when generating a physical operation evidence chain, the content of each physical operation evidence node can include action tags for maintenance operations, operation timestamps, operator identities, tool information used, and key parameter readings (such as torque and pressure). This content is an objective record of the maintenance process. The temporal relationship between nodes refers to the order in which these physical operation evidence nodes occur during the actual maintenance process. This is usually sorted by timestamps, but can also be determined according to a pre-defined standard maintenance process. The physical operation evidence chain is a chain-like data structure, its core being the use of cryptographic hash functions to ensure data integrity and the immutability of the temporal sequence. Each physical operation evidence node contains not only its own content hash but also a pointer hash pointing to its predecessor. The content hash is a unique digest obtained by hashing all the content data of the current node; any minor modification to the node's content will cause the content hash to change. The pointer hash is the result of hashing the complete hash of the previous node (including its content hash and pointer hash), cryptographically linking the current node to the previous node, thus forming an irreversible chain.

[0199] Similarly, when generating a twin simulation evidence chain, the content of the twin simulation evidence nodes can include simulation operation tags, simulation process parameters, simulation status outputs, simulation timestamps, etc. These are predictive data generated by the digital twin engine during the simulation of maintenance operations. The temporal relationship between nodes is also based on the chronological order of the simulation process. The twin simulation evidence chain is constructed in the same way as the physical operation evidence chain, with each twin simulation evidence node containing its content hash and a pointer hash pointing to the previous twin simulation evidence node. This ensures the integrity of the simulation data and its immutability along the simulation timeline.

[0200] Furthermore, when generating the difference evidence chain, the content of each difference evidence node can include global difference degree, difference evolution path chain, and anomaly identifiers. These are key information obtained by comparing and analyzing the physical operations and simulation results after maintenance. The temporal relationship between nodes reflects the evolution order of the difference analysis results during the maintenance process. The construction of the difference evidence chain also follows the same principle: each difference evidence node contains its content hash and a pointer hash pointing to the previous difference evidence node. This ensures the accuracy of the difference analysis results and the immutability of its evolution process.

[0201] When generating the overall hash of the evidence graph, the root hash is the starting hash value of each hash chain, representing the overall state of that chain. For example, the root hash of the physical operation evidence chain could be the hash value of one of its physical operation evidence nodes, or it could be the root hash obtained by aggregating the hashes of all nodes using methods such as Merkle trees. The linking operation refers to combining the root hashes of these three independent evidence chains (physical operation evidence chain, twin simulation evidence chain, and difference evidence chain), for example, through string concatenation or hash operations, to generate a unified overall hash representing the state of the entire evidence graph.

[0202] Write the overall hash to the blockchain. Blockchain is a distributed ledger technology characterized by its difficulty in tampering with data once written, and its timestamping and decentralized nature. Writing the overall hash of the evidence graph to the blockchain means that this overall hash is permanently recorded on a public, transparent, and immutable distributed ledger, accompanied by a timestamp. The writing method can be through a smart contract call, adding the overall hash as part of the transaction data to the chain, or directly as part of the block header.

[0203] Through the aforementioned technical solution, this application enhances the integrity and tamper-proof capability of the heavy-duty truck repair order evidence graph by introducing a triple hash chain evidence storage mechanism. Specifically, the independent construction of the physical operation evidence chain, the twin simulation evidence chain, and the difference evidence chain, utilizing content hashing and pointer hashing, ensures that the content of each type of evidence node is not tampered with, and that their logical relationship in the repair sequence is solidified. Any modification to the content of a single evidence node or its position in the chain will result in a hash value mismatch, which will be detected immediately, effectively preventing the tampering of local evidence. Furthermore, the root hashes of these three evidence chains are linked to generate the overall hash of the evidence graph, providing a unified, cryptographically secure fingerprint that comprehensively reflects the overall state and consistency of the entire repair order evidence graph. Writing this overall hash into the blockchain, utilizing the distributed, immutable, and timestamped characteristics of the blockchain, provides an authoritative and traceable third-party evidence storage for the entire evidence graph, greatly enhancing the judicial probative value and credibility of the evidence graph. This enables the entire chain of evidence for repair orders to provide highly credible and undeniable evidence when facing warranty determination, liability tracing, and dispute resolution, thereby solving the problem of reduced credibility caused by tampering with nodes in the evidence graph and disruption of the time sequence.

[0204] The following example will provide a more detailed explanation of the above technical solution: Suppose a heavy-duty truck enters repair shop A for major repairs due to engine failure. After the repair order is initiated, the system initializes the truck's dual-role digital twin engine. Specifically, the system obtains the truck's vehicle identification number (VIN) and, based on this, retrieves its basic digital twin model from the cloud. This model includes the engine's structural parameters, performance parameters, and standard repair procedures. The system instantiates this basic model into a predictive instance and a baseline instance. The predictive instance inherits the structural and performance parameters and is used to simulate the state after the repair operation. The baseline instance inherits the structural parameters, performance parameters, and standard repair procedures, locking the standard repair procedures and standard performance benchmark as a static reference. To ensure the predictive instance is in a state consistent with the baseline instance's standard performance benchmark before repair begins, the system extracts multi-dimensional benchmark performance parameters from the standard performance benchmark and maps them to the predictive instance's initialization fields, assigning matching initial values. Simultaneously, a difference calculation interface is established between the predictive instance and the baseline instance to calculate the offset of the predictive instance's simulation state relative to the baseline instance's standard performance benchmark in real time. This initialization process solves the problems of lacking standard references and real-time simulation capabilities in existing technologies, laying the foundation for subsequent evidence construction.

[0205] During engine overhauls, such as the "piston ring replacement" operation, the system collects physical operation evidence streams and twin simulation evidence streams in real time. The physical operation evidence stream may include sensor data of the technician tightening bolts with a torque wrench, operation video clips, and manually entered component serial numbers by the technician; each data unit has a precise timestamp. Simultaneously, the predictive instance simulates the "piston ring replacement" operation in real time based on a preset maintenance operation sequence and parameters, generating a corresponding twin simulation evidence stream. This stream includes simulated torque values, piston ring installation status, and other simulation data, also with timestamps.

[0206] The system performs temporal alignment and semantic mapping on these evidence streams. It extracts standard operation nodes, such as "replacing piston rings," along with their sequence and standard time intervals from the baseline instance, constructing a standard operation temporal graph. For the physical operation evidence unit "replacing piston rings" in the physical operation evidence stream, the system parses its operation action label and locates the matching target standard operation node in the standard operation temporal graph. Based on the node's position in the graph, it determines its preceding and succeeding standard operation nodes and obtains the corresponding first and second standard time intervals. Using the first timestamp of the physical operation evidence unit, the system filters out twin-simulation evidence units in the twin-simulation evidence stream whose second timestamp falls within the allowed time window range as candidate twin-simulation evidence units, and combines them with the physical operation evidence units to form multiple candidate evidence pairs.

[0207] For each candidate evidence pair, the system parses the operation action tags (e.g., "install piston rings") in the physical operation evidence unit and extracts the corresponding simulation operation tags from the twin simulation evidence unit. The system performs consistency checks and compares the simulation process parameters (e.g., simulated torque values) contained in the twin simulation evidence unit with the standard process parameters (e.g., standard torque values) of the corresponding operation action tags in the benchmark instance. Based on the check and comparison results, the mapping confidence level is determined, and candidate evidence pairs with confidence levels exceeding a preset threshold are identified as associated evidence pairs. For example, if the torque value recorded in the physical operation is highly consistent with both the simulated torque value and the standard torque value, the confidence level is high.

[0208] For these correlated evidence pairs, the system retrieves the standard performance benchmark (e.g., standard range of engine cylinder pressure) associated with the "replace piston rings" operation from the baseline instance. The system compares the simulation state (e.g., simulated cylinder pressure) contained in the twin simulation evidence unit of the correlated evidence pair with the standard performance benchmark dimension-by-dimensional, calculating the deviation vector of the simulation state relative to the standard performance benchmark. For example, the simulated cylinder pressure may be slightly lower than the standard value. The system performs multidimensional feature decomposition on this deviation vector, obtaining deviation components in multiple preset dimensions, and maps each deviation component to a corresponding anomaly level interval. These anomaly level values, along with the position of the correlated evidence pair in the maintenance timeline, are encapsulated as difference nodes. By concatenating the difference nodes of other correlated evidence pairs in the same maintenance process, the initial difference evolution path corresponding to the "replace piston rings" operation is generated. This process solves the problem of the lack of logical corroboration relationships in multi-source data in existing technologies, realizes the correlation between physical operations and simulation predictions, and initially identifies potential deviations.

[0209] After maintenance, the heavy truck undergoes actual performance testing to obtain the measured performance data as the actual state. The system compares this measured state with the simulated state output by the predicted instance after all maintenance operations are completed to obtain the global difference degree. Specifically, the system extracts features and compares them layer by layer from the measured and simulated states to obtain a multi-dimensional original difference feature map. Simultaneously, the system acquires the initial difference evolution paths corresponding to multiple pairs of related evidence generated during the maintenance process. The system performs spatial location matching between the original difference feature map and the initial difference evolution paths to determine the maintenance operation stage to which each difference feature point in the original difference feature map belongs. For example, a difference feature point might be matched to the "replace piston rings" stage. Based on the deviation amount corresponding to each maintenance operation stage in the initial difference evolution path, the system weights and corrects the difference feature values ​​at the corresponding positions in the original difference feature map to obtain a corrected multi-dimensional difference feature map. The corrected multi-dimensional difference feature map is then subjected to dimensionality reduction and aggregation processing to extract the difference degree components between the measured and simulated states on multiple preset evaluation dimensions. Based on a pre-defined fault propagation model, the contribution weight of each difference component to the overall maintenance quality is calculated, and then weighted and fused to obtain the global difference degree. This comparison process is corrected by introducing an initial difference evolution path, making the calculation of the global difference degree more accurate and able to reflect the actual accumulation of deviations during the maintenance process, rather than simple state differences.

[0210] The system generates difference evidence nodes containing global difference degree, difference evolution path chain, and anomalous link identifiers based on multiple related evidence pairs and their corresponding initial difference evolution paths. The system concatenates the initial difference evolution paths corresponding to each related evidence pair according to the maintenance timeline, generating an initial difference evolution path chain that runs through the entire maintenance process. The system determines the global deviation consistency between this path chain and the global difference degree, and filters candidate path nodes that have a deviation propagation association with the global difference degree. For example, it filters path nodes whose node deviation is greater than the global deviation consistency and whose node deviation change gradient exceeds a preset threshold. Causal tracing analysis is performed on these candidate path nodes to identify the root cause node and the transmission node that led to the global difference degree, and corresponding anomalous link identifiers are generated for them. For example, "improper piston ring installation" might be identified as the root cause node, and "insufficient cylinder pressure" as the transmission node. The initial difference evolution path chain, global difference degree, and anomalous link identifiers of the root cause node and transmission node are associated and encapsulated to generate difference evidence nodes. This difference evidence node provides key information for verifying repair effectiveness and proving causality, solving the pain point of difficulty in tracing the root cause of problems in existing technologies.

[0211] Furthermore, the system acquires the model training evidence chain generated by repair station A during the federated learning process. After each round of local training in the federated learning process, the model parameters generated by repair station A are recorded, along with the training data hash value and training process hyperparameters (such as training rounds and learning rate). These model parameters are hashed to obtain parameter hash values, which are then encapsulated with the training data hash value and training process hyperparameters to form a generation proof. This information is organized into a local training evidence sub-chain for repair station A according to the training rounds, and synchronously associated with the sub-chains of other repair stations to form a complete model training evidence chain. This provides a reliable basis for subsequent parameter source tracing.

[0212] The system constructs an evidence graph for maintenance orders based on the model training evidence chain, multiple related evidence pairs, and discrepancy evidence nodes. Physical operation evidence units and twin simulation evidence units are mapped to physical operation evidence nodes and twin simulation evidence nodes, respectively; discrepancy evidence nodes are mapped to discrepancy evidence nodes; and local model parameters and generated proofs in the model training evidence chain serve as model training evidence nodes, collectively forming the evidence node set. The system analyzes the semantic mapping relationship between physical operation evidence units and twin simulation evidence units in related evidence pairs, establishing corroboration edges between corresponding physical operation evidence nodes and twin simulation evidence nodes to represent the logical consistency between actual operations and simulation predictions in the same maintenance operation stage. The generated proofs in the model training evidence chain are matched with the physical operation evidence units at corresponding temporal positions in the related evidence pairs. Based on the matching results, tracing edges are established between model training evidence nodes and successfully matched physical operation evidence nodes to represent the data source association between the model training process and specific maintenance operations. Based on the difference evolution path chain contained in the difference evidence nodes, the system identifies the target physical operation evidence node corresponding to the abnormal link identifier along the path chain (e.g., the physical operation evidence node corresponding to "improper piston ring installation"). A causal edge is established between the difference evidence node and the target physical operation evidence node to represent the causal relationship between the global difference degree and the specific maintenance operation link. This generates an evidence graph of the maintenance order containing a set of evidence nodes, as well as corroborating edges, tracing edges, and causal edges.

[0213] To ensure the immutability of the evidence graph, the system uses a triple hash chain for evidence storage. A physical operation evidence chain is generated based on the content and time sequence of each physical operation evidence node in the evidence graph. A twin simulation evidence chain is generated based on the content and time sequence of the twin simulation evidence nodes. A difference evidence chain is generated based on the content and time sequence of the difference evidence nodes. Each node in each chain contains its own content hash and a pointer hash pointing to the previous node. The root hashes of the physical operation evidence chain, the twin simulation evidence chain, and the difference evidence chain are linked to generate the overall hash of the evidence graph, and this overall hash is written to the blockchain.

[0214] Using the above method, all key information regarding the overhaul process of the heavy-duty truck engine at repair station A is integrated into a logically connected, traceable, and verifiable evidence map, which is then stored using blockchain technology. This solves the problems of independent storage of multi-source data and lack of logical corroboration in existing technologies, enabling verification of repair effects, tracing of parameter sources, proof of causality, and corroboration of multi-source evidence. This provides strong judicial proof for warranty determination, liability tracing, and dispute resolution.

[0215] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0216] Figure 5 This is a schematic diagram of a system for constructing a complete evidence chain for heavy-duty truck repair orders, as provided in an embodiment of this application. See also... Figure 5 The system includes: Initialization module 501 is used to initialize the dual-role digital twin engine of the target heavy truck in response to the initiation of a maintenance order. The dual-role digital twin engine includes a prediction instance and a benchmark instance. The prediction instance is used to simulate and predict the simulation state based on the maintenance operation, and the benchmark instance is used to solidify the standard maintenance process and standard performance benchmark.

[0217] The acquisition module 502 is used to acquire physical operation evidence stream and twin simulation evidence stream in real time during the maintenance process, and perform temporal alignment and semantic mapping on each physical operation evidence unit in the physical operation evidence stream and the corresponding twin simulation evidence unit in the twin simulation evidence stream based on the standard maintenance process in the benchmark instance to obtain multiple related evidence pairs and the initial difference evolution path corresponding to each related evidence pair.

[0218] The acquisition module 503 is used to acquire the measured performance data of the target heavy truck as the measured state after the maintenance is completed, compare the measured state with the simulation state output by the predicted instance after all maintenance operations are completed to obtain the global difference degree, and generate difference evidence nodes containing the global difference degree, difference evolution path chain and abnormal link identifier based on the multiple related evidence pairs and the initial difference evolution path corresponding to each related evidence pair.

[0219] Module 504 is used to obtain the model training evidence chain generated by each repair station during the federated learning process. The model training evidence chain includes local model parameters and corresponding generation proofs. Based on the model training evidence chain, the multiple related evidence pairs and the difference evidence nodes, the evidence graph of the repair order is constructed, and the evidence graph is stored using a triple hash chain.

[0220] It should be noted that the above-described system for constructing a complete evidence chain for heavy-duty truck repair orders is only illustrated by the division of the functional modules described above. In practical applications, the functions described above can be assigned to different functional modules as needed, that is, the internal structure of the computer equipment can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the system for constructing a complete evidence chain for heavy-duty truck repair orders provided in the above-described embodiments and the method embodiment for constructing a complete evidence chain for heavy-duty truck repair orders belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0221] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including a computer program that can be executed by a processor to complete the method for constructing a complete evidence chain for heavy truck repair orders in the above embodiments. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, or optical data storage device, etc.

[0222] In an exemplary embodiment, a computer program product or computer program is also provided, which includes program code stored in a computer-readable storage medium. The processor of a computer device reads the program code from the computer-readable storage medium and executes the program code, causing the computer device to execute the above-described method for constructing a chain of evidence for the entire order process for heavy truck repair.

[0223] In some embodiments, the computer program involved in the present application embodiments may be deployed and executed on a computer device, or executed on multiple computer devices located in one location, or executed on multiple computer devices distributed in multiple locations and interconnected through a communication network. Multiple computer devices distributed in multiple locations and interconnected through a communication network may constitute a blockchain system.

[0224] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0225] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An order whole-process evidence chain construction method for heavy truck maintenance, characterized in that, The method includes: In response to the initiation of a maintenance order for the target heavy truck, the dual-role digital twin engine of the target heavy truck is initialized. The dual-role digital twin engine includes a prediction instance and a benchmark instance. The prediction instance is used to simulate and predict the simulation state based on the maintenance operation, and the benchmark instance is used to solidify the standard maintenance process and standard performance benchmark. During the maintenance process, physical operation evidence stream and twin simulation evidence stream are collected in real time. Based on the standard maintenance process in the benchmark instance, each physical operation evidence unit in the physical operation evidence stream and the corresponding twin simulation evidence unit in the twin simulation evidence stream are time-aligned and semantically mapped to obtain multiple related evidence pairs and the initial difference evolution path corresponding to each related evidence pair. After maintenance is completed, the measured performance data of the target heavy truck is obtained as the measured state. The measured state is compared with the simulation state output by the predicted instance after all maintenance operations are completed to obtain the global difference degree. Based on the multiple related evidence pairs and the initial difference evolution path corresponding to each related evidence pair, a difference evidence node containing the global difference degree, the difference evolution path chain and the abnormal link identifier is generated. Obtain the model training evidence chain generated by each repair station during the federated learning process. The model training evidence chain includes local model parameters and corresponding generation proofs. Based on the model training evidence chain, the multiple associated evidence pairs, and the difference evidence nodes, construct the evidence graph of the repair order, and store the evidence graph using a triple hash chain.

2. The method according to claim 1, characterized in that, The initialization of the dual-role digital twin engine for the target heavy truck includes: Obtain the vehicle identification number of the target heavy truck, and obtain the basic digital twin model of the target heavy truck from the cloud based on the vehicle identification number. The basic digital twin model includes structural parameters, performance parameters, and standard maintenance processes. The basic digital twin model is instantiated into the prediction instance and the benchmark instance, respectively. The prediction instance inherits the structural parameters and performance parameters in the basic digital twin model, and the benchmark instance inherits the structural parameters, performance parameters and standard maintenance process in the basic digital twin model. The standard maintenance process and standard performance benchmark in the benchmark instance are locked as static references. Based on the standard performance benchmark locked in the benchmark instance, initial state parameters are configured for the prediction instance, and a difference calculation interface is established between the prediction instance and the benchmark instance. The difference calculation interface is used to calculate in real time the offset of the simulation state output by the prediction instance relative to the standard performance benchmark in the benchmark instance during the maintenance process, as the input of the initial difference evolution path.

3. The method according to claim 1, characterized in that, The process involves performing temporal alignment and semantic mapping on each physical operation evidence unit in the physical operation evidence stream and the corresponding twin simulation evidence unit in the twin simulation evidence stream, based on the standard maintenance process in the benchmark instance, to obtain multiple associated evidence pairs and the initial difference evolution path corresponding to each associated evidence pair, including: Extract the first timestamp of each physical operation evidence unit in the physical operation evidence stream, and extract the second timestamp of each twin simulation evidence unit in the twin simulation evidence stream. Based on the time difference between the first timestamp and the second timestamp and the standard operation sequence specified by the standard maintenance process in the benchmark instance, pair each physical operation evidence unit with a twin simulation evidence unit that meets the timing constraint conditions to obtain multiple candidate evidence pairs. For each candidate evidence pair, the operation action tags contained in the physical operation evidence unit are parsed, and the corresponding simulation operation tags are extracted from the twin simulation evidence unit. The consistency of the operation action tags and the simulation operation tags is verified. At the same time, the simulation process parameters contained in the twin simulation evidence unit are compared item by item with the standard process parameters in the benchmark instance corresponding to the operation action tags. Based on the consistency verification results and the item-by-item comparison results, the mapping confidence between the physical operation evidence unit and the twin simulation evidence unit in the candidate evidence pair is determined. Candidate evidence pairs with a mapping confidence exceeding a preset threshold are designated as the associated evidence pairs. Retrieve the standard performance benchmark associated with the operation action label in the associated evidence pair from the benchmark instance, compare the simulation state contained in the twin simulation evidence unit in the associated evidence pair with the standard performance benchmark dimension by dimension, calculate the deviation vector of the simulation state relative to the standard performance benchmark, and generate the initial difference evolution path corresponding to the associated evidence pair based on the deviation vector and the position of the associated evidence pair in the maintenance timeline.

4. The method according to claim 1, characterized in that, The step of comparing the measured state with the simulated state output by the predicted instance after all maintenance operations are completed to obtain the global difference degree includes: Feature extraction and layer-by-layer comparison are performed on the measured state and the simulated state to obtain a multi-dimensional original difference feature map; The initial difference evolution path corresponding to multiple related evidence pairs generated during the maintenance process is obtained. The original difference feature map is spatially matched with the initial difference evolution path to obtain the maintenance operation link to which each difference feature point in the original difference feature map belongs. Based on the deviation amount corresponding to each maintenance operation link in the initial difference evolution path, the difference feature value at the corresponding position in the original difference feature map is weighted and corrected to obtain the corrected multidimensional difference feature map. The modified multidimensional difference feature map is subjected to dimensionality reduction and aggregation processing to extract the difference components between the measured state and the simulated state on multiple preset evaluation dimensions. The contribution weight of each difference component to the overall maintenance quality is calculated based on the preset fault propagation model, and the multiple difference components are weighted and fused according to the contribution weight to obtain the global difference.

5. The method according to claim 1, characterized in that, The step of comparing the measured state with the simulated state output by the predicted instance after all maintenance operations are completed to obtain the global difference degree includes: The initial difference evolution path corresponding to multiple pairs of related evidence generated during the maintenance process is obtained. Multiple key time nodes and the node deviation amount corresponding to each key time node are extracted from the initial difference evolution path. The key time nodes correspond to the maintenance operation links in which state shifts occur during the maintenance process. The measured state and the simulation state are decomposed into multiple measured state segments and multiple simulation state segments corresponding to the multiple key timing nodes; For each key timing node, the measured state segment corresponding to the key timing node is compared with the simulated state segment to generate the node difference degree at the key timing node. The node difference degree is then corrected based on the node deviation corresponding to the key timing node to obtain the corrected node difference degree. The corrected node differences corresponding to the multiple key timing nodes are aggregated according to the maintenance timing, and the aggregation result is used as the global difference between the measured state and the simulation state.

6. The method according to claim 1, characterized in that, The step of generating differential evidence nodes containing the global differential degree, differential evolution path chain, and anomaly identifier based on the multiple pairs of associated evidence and the initial differential evolution path corresponding to each pair of associated evidence includes: According to the maintenance sequence, the initial difference evolution paths corresponding to each pair of related evidence are connected in series to generate an initial difference evolution path chain that runs through the entire maintenance process, and multiple path nodes and the node deviation amount corresponding to each path node are extracted from the initial difference evolution path chain. Determine the global deviation fit between the initial difference evolution path chain and the global difference degree, and select candidate path nodes that have a deviation propagation association with the global difference degree from the plurality of path nodes based on the global deviation fit. A causal tracing analysis is performed on the candidate path nodes. Based on the temporal relationship and deviation propagation direction between the candidate path nodes, the root node and the transmission node that cause the global difference are identified, and corresponding abnormal link identifiers are generated for the root node and the transmission node respectively. The initial difference evolution path chain, the global difference degree, and the abnormal link identifiers of the root node and the transmission node are associated and encapsulated to generate the difference evidence node.

7. The method according to claim 1, characterized in that, The acquisition of the model training evidence chain generated by each maintenance station during the federated learning process includes: After each round of local training in federated learning is completed, the model parameters generated by local training at each repair station are obtained, and the training data hash value corresponding to the model parameters and the training process hyperparameters are recorded for each repair station. The training process hyperparameters include the number of training rounds and the learning rate. The model parameters of each repair station are hashed to obtain the parameter hash value corresponding to the model parameters. The parameter hash value is then associated and encapsulated with the training data hash value and the training process hyperparameters to generate a generation proof for each repair station. The model parameters, parameter hash values, and generated proofs corresponding to each repair station are organized according to the training rounds to obtain the local training evidence subchain corresponding to each repair station. The local training evidence subchains corresponding to each maintenance station are synchronously associated according to the aggregation rounds of federated learning to obtain the model training evidence chain.

8. The method according to claim 1, characterized in that, The process of constructing the evidence graph of the repair order based on the trained evidence chain of the model, the multiple associated evidence pairs, and the differential evidence nodes includes: The physical operation evidence unit and twin simulation evidence unit in the multiple associated evidence pairs are mapped to physical operation evidence nodes and twin simulation evidence nodes, respectively, and the difference evidence nodes are mapped to difference evidence nodes. At the same time, the local model parameters and corresponding generated proofs of each maintenance station are extracted from the model training evidence chain as model training evidence nodes to construct an evidence node set. The semantic mapping relationship between the physical operation evidence unit and the twin simulation evidence unit in the multiple related evidence pairs is analyzed. Based on the semantic mapping relationship, a corroboration edge is established between the corresponding physical operation evidence node and the twin simulation evidence node. The corroboration edge is used to characterize the logical consistency between the actual operation and the simulation prediction in the same maintenance operation link. The generated proof in the model training evidence chain is matched with the physical operation evidence unit at the corresponding time position in the multiple associated evidence pairs. Based on the matching result, a traceability edge is established between the model training evidence node and the successfully matched physical operation evidence node. The traceability edge is used to characterize the data source association between the model training process and the specific maintenance operation. Based on the difference evolution path chain contained in the difference evidence nodes, the target physical operation evidence node corresponding to the abnormal link identifier is identified along the difference evolution path chain, and a causal edge is established between the difference evidence node and the target physical operation evidence node. The causal edge is used to characterize the causal relationship between the global difference degree and the specific maintenance operation link, thereby generating an evidence map of the maintenance order containing the evidence node set, the corroborating edge, the tracing edge, and the causal edge.

9. The method according to claim 1, characterized in that, The process of storing the evidence graph using a triple hash chain includes: Based on the content of each physical operation evidence node in the evidence graph and the temporal relationship between the nodes, a physical operation evidence chain is generated. Each physical operation evidence node in the physical operation evidence chain contains the content hash of the node and the pointer hash pointing to the previous physical operation evidence node. Based on the content of each twin simulation evidence node in the evidence graph and the temporal relationship between the nodes, a twin simulation evidence chain is generated. Each twin simulation evidence node in the twin simulation evidence chain contains the content hash of the node and the pointer hash pointing to the previous twin simulation evidence node. Based on the content of each differential evidence node in the evidence graph and the temporal relationship between the nodes, a differential evidence chain is generated. Each differential evidence node in the differential evidence chain contains the content hash of the node and the pointer hash pointing to the previous differential evidence node. The root hash of the physical operation evidence chain, the root hash of the twin simulation evidence chain, and the root hash of the difference evidence chain are linked to generate the overall hash of the evidence graph, and the overall hash is written into the blockchain.

10. A system for constructing a complete evidence chain for heavy-duty truck repair orders, characterized in that, The system includes: An initialization module is used to initialize the dual-role digital twin engine of the target heavy truck in response to the initiation of a maintenance order. The dual-role digital twin engine includes a prediction instance and a benchmark instance. The prediction instance is used to simulate and predict the simulation state based on the maintenance operation, and the benchmark instance is used to solidify the standard maintenance process and standard performance benchmark. The acquisition module is used to acquire physical operation evidence stream and twin simulation evidence stream in real time during the maintenance process, and perform temporal alignment and semantic mapping on each physical operation evidence unit in the physical operation evidence stream and the corresponding twin simulation evidence unit in the twin simulation evidence stream based on the standard maintenance process in the benchmark instance to obtain multiple related evidence pairs and the initial difference evolution path corresponding to each related evidence pair. The acquisition module is used to acquire the measured performance data of the target heavy truck as the measured state after the maintenance is completed, compare the measured state with the simulation state output by the predicted instance after all maintenance operations are completed to obtain the global difference degree, and generate difference evidence nodes containing the global difference degree, difference evolution path chain and abnormal link identifier based on the multiple related evidence pairs and the initial difference evolution path corresponding to each related evidence pair. The module is used to obtain the model training evidence chain generated by each repair station during the federated learning process. The model training evidence chain includes local model parameters and corresponding generation proofs. Based on the model training evidence chain, the multiple associated evidence pairs and the difference evidence nodes, the evidence graph of the repair order is constructed, and the evidence graph is stored using a triple hash chain.

Images