Vehicle analysis method based on trajectory data and driving behaviors and readable storage medium

Abstract

The invention discloses a vehicle analysis method based on trajectory data and driving behaviors. The method comprises the following steps: generating a standardized vehicle full-time-domain operation state sequence through multi-source heterogeneous data of an operating vehicle; three types of features are extracted based on the sequence, and an operation intensity atlas reflecting the actual use load of the vehicle is constructed through fusion; mapping the atlas to a preset vehicle core part loss model, calculating the accumulated damage degree and residual life of the parts, and reasoning and outputting a vehicle real-time health state evaluation result through a customized vertical big language model; generating a vehicle operation efficiency quantitative index in combination with the operation intensity map and the cross-vehicle-type and cross-line standardized energy efficiency reference; and finally, outputting a maintenance execution scheme by generating an optimal personalized maintenance period and a maintenance item list. According to the invention, accurate quantification of the operation load of the commercial vehicle, scientific analysis of the operation efficiency and dynamic optimization of the maintenance strategy are realized, and the refinement and intelligence level of motorcade management is improved.

Description

Technical Field

[0001] This invention relates to the field of vehicle operation management and intelligent maintenance technology, and in particular to a vehicle analysis method based on trajectory data and driving behavior. Background Technology

[0002] As the core carriers of logistics and commercial operations, the operational efficiency, equipment health status, and maintenance rationality of commercial vehicles directly impact a company's operating costs, transportation safety, and task delivery efficiency. Currently, the field of commercial vehicle management still faces numerous technical pain points and industry challenges: vehicle load assessment relies solely on single indicators such as mileage, failing to comprehensively consider the coupled effects of micro-driving behaviors (rapid acceleration / deceleration, idling), macro-operating characteristics (load rate, empty-run rate), and external environments (road slope, congestion), making it difficult to accurately reflect the actual usage load of vehicles and leading to a disconnect between component wear assessment and actual operating conditions; health status assessments of core vehicle components often employ fixed threshold alarms or post-fault diagnosis models, lacking cumulative damage calculations and remaining life predictions based on operating conditions, and lacking sufficient multi-source data fusion and reasoning capabilities, failing to identify system-level coupled fault risks where a single indicator does not exceed limits, easily triggering sudden failures; energy efficiency evaluation of commercial vehicles lacks standardized benchmarks across vehicle types and routes, and there is no unified comparison scale for energy consumption data under different vehicle types and road conditions, making it difficult to accurately identify inefficient operating vehicles and abnormal driving patterns, resulting in a lack of targeted operational efficiency optimization; traditional maintenance strategies use fixed mileage / The time-based, scheduled maintenance model, without dynamic optimization based on the actual health status of vehicles, operational efficiency, and business task schedules, is prone to problems such as over-maintenance leading to resource waste or under-maintenance causing malfunctions. At the same time, maintenance plans and operational tasks are likely to conflict, affecting the delivery of transportation tasks.

[0003] Furthermore, while some existing technologies attempt to combine vehicle data for health assessment or efficiency analysis, these are mostly single-dimensional applications, such as optimizing driving behavior or diagnosing component failures. Simultaneously, the application of large-scale AI models in vehicle maintenance and diagnosis often involves direct invocation of general models, lacking specialized fine-tuning and standardized interactive design specific to the vehicle domain. This leads to issues such as missing professional knowledge and unexecutable output results, failing to meet the actual needs of refined and intelligent management of commercial vehicles. Therefore, there is an urgent need for an integrated vehicle analysis method that can fuse multi-source heterogeneous data to achieve precise quantification of vehicle operating load, intelligent assessment of health status, scientific quantification of operational efficiency, and dynamic optimization of maintenance strategies, addressing the current industry pain points in commercial vehicle management. Summary of the Invention

[0004] The vehicle analysis method and computer-readable storage medium based on trajectory data and driving behavior provided by this invention achieve refined, intelligent and efficient management of operating vehicles by constructing a full-process technical system that includes multi-source data standardization processing, operation load map construction, intelligent component health assessment, quantitative analysis of operation efficiency, and dynamic maintenance strategy generation.

[0005] Specifically, the first aspect of this invention discloses a vehicle analysis method based on trajectory data and driving behavior, characterized by comprising the following steps: Multi-source heterogeneous data of operating vehicles are acquired, and after multi-scale spatiotemporal alignment and data cleaning, a standardized vehicle full-time domain operating status sequence is generated. Based on the vehicle's full-time operating state sequence, three types of features are extracted: micro-driving behavior, macro-operation, and external environment. These three types of features are then fused to construct an operating intensity map that reflects the actual usage load of the vehicle. The operational intensity map is mapped to a preset vehicle core component loss model to calculate the cumulative damage and remaining life of the components. Multi-source heterogeneous data is fused and reasoned through a customized vertical large language model to output the real-time health status assessment results of the vehicle. By combining the aforementioned operational intensity map with standardized energy efficiency benchmarks across vehicle types and routes, a multi-dimensional weighted calculation is performed to generate a quantitative index of vehicle operating efficiency, identifying inefficient operating vehicles and abnormal driving modes. Based on the real-time health status assessment results of the vehicle and the quantitative index of operational efficiency, combined with the operational task scheduling constraints, the optimal personalized maintenance cycle and maintenance item list are generated through dynamic programming algorithm, and a dynamic maintenance execution plan is output. Specifically, the extraction of three types of features—micro-driving behavior, macro-operation, and external environment—based on the vehicle's full-time operating state sequence includes: A time-series analysis was performed on the vehicle's full-time operating state sequence. By setting acceleration and engine speed thresholds, micro-driving behavior features were extracted, including the frequency of rapid acceleration and deceleration and the duration of long-term idling. By combining vehicle dispatch records, macro-operational characteristics are calculated, and the vehicle load rate and empty running rate are quantified to obtain the macro-operational data. By associating with a geographic information system, matching vehicle trajectory locations, analyzing external environmental features, and statistically analyzing road slope and traffic congestion levels, the external environmental features are obtained.

[0006] In a second aspect, the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions for a vehicle analysis method based on trajectory data and driving behavior.

[0007] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0008] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A flowchart of a vehicle analysis method based on trajectory data and driving behavior provided in an embodiment of this disclosure.

[0010] Figure 2 A flowchart of a method for acquiring multi-source heterogeneous data of operating vehicles and generating standardized sequences, provided in an embodiment of this disclosure.

[0011] Figure 3 A flowchart illustrating the method for constructing and running intensity maps provided in this embodiment of the disclosure.

[0012] Figure 4 A flowchart illustrating the method for outputting real-time vehicle health status assessment results provided in this embodiment of the disclosure.

[0013] Figure 5 A flowchart illustrating the method for generating a quantitative index of vehicle operating efficiency provided in this embodiment of the disclosure.

[0014] Figure 6 A flowchart of the output dynamic maintenance execution scheme method provided in the embodiments of this disclosure.

[0015] Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present disclosure. Detailed Implementation

[0016] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0017] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0018] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0019] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0020] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0021] See Figure 1 As shown in the figure, this disclosure presents a method for quantifying vehicle operating efficiency and dynamic maintenance based on trajectory data and driving behavior analysis, specifically disclosing: Step S1. Acquire multi-source heterogeneous data from operating vehicles, including vehicle trajectory data, CAN bus operation data, service scheduling data, and historical maintenance records. Through multi-scale spatiotemporal alignment and data cleaning, obtain a standardized sequence of vehicle's full-time-domain operating status. See also Figure 2 The process of acquiring multi-source heterogeneous data from operating vehicles and generating standardized sequences is achieved through the following steps: Step S101: Construct a multi-source data acquisition network that integrates hardware and software. Specifically, firstly, the BeiDou / GPS dual-mode positioning module integrated in the vehicle-mounted intelligent terminal (T-BOX) is used to capture vehicle trajectory data in real time at a sampling frequency of 1Hz (including latitude, longitude, altitude, heading angle, and timestamp). Secondly, the underlying vehicle operating data is deeply read through the CAN bus gateway, specifically including engine speed, instantaneous torque, fuel consumption rate, brake pedal opening, and diagnostic fault codes (DTCs). Simultaneously, the system connects to the enterprise's TMS transportation management system via a RESTful API interface to synchronously obtain business scheduling data (such as task order number, cargo weight, origin and destination information, and designated routes). Finally, the fleet's historical maintenance database is accessed via ODBC database connection to extract historical maintenance records (including the last maintenance mileage, replacement parts list, and repair category), thus completing the collection of all raw data.

[0022] Step S102 involves preprocessing the collected raw data, specifically including removing noise data, eliminating duplicate records, and filling missing values. In practice, firstly, a Kalman filter algorithm is used to smooth the GPS trajectory coordinates, while setting physical limit thresholds (e.g., setting the instantaneous speed limit to 150 km / h) to identify and eliminate abnormal noise data caused by signal drift. Secondly, the uniqueness of all records is verified based on the combination key value of "Vehicle Identifier (VIN) + timestamp," automatically deleting duplicate data rows caused by network retransmission mechanisms. Finally, for short-term data loss due to signal blind spots (e.g., a missing window of less than 60 seconds), linear interpolation or Lagrange interpolation is used to fit and fill in the missing latitude, longitude, or engine speed data based on the effective values ​​of adjacent moments.

[0023] Step S103 involves performing multi-scale spatiotemporal alignment, mapping data from different sampling frequencies to a unified time reference axis. Specifically, firstly, a continuously sampled and stable GPS timestamp sequence (e.g., 1Hz) is selected as the main time axis. Secondly, for high-frequency CAN data (e.g., RPM or torque between 10Hz and 50Hz), a sliding time window algorithm is used for downsampling, calculating the statistical characteristic values ​​(e.g., mean, maximum) within the window as the reading for that second. For low-frequency service data (e.g., single task status), a zero-order hold is used for upsampling and filling, broadcasting the same status value to the corresponding time period. For example, the cumulative value of 20 instantaneous fuel injection data collected within one second is aligned to the latitude and longitude coordinates of that second and synchronously associated with the current "heavy load" task label, thus forming a multi-dimensional feature vector with strictly aligned time granularity.

[0024] Step S104 involves standardizing and encapsulating the aligned data to output a sequence of vehicle operating states across the entire time domain. Specifically, this involves first constructing a unified data dictionary and storage structure, defining a standardized set of fields including time index, spatial location, chassis condition, and business attributes; second, unifying the data's units of measurement and encoding, for example, unifying all speed units to km / h, converting discrete states such as "fully loaded / unloaded" into "1 / 0" numerical markers, and mapping fault codes to standard fault vectors; finally, using a structured data format (such as JSON or Parquet), concatenating the feature vectors from each time segment to generate a continuous sequence of vehicle operating states across the entire time domain. For example, a single record in the output sequence might be described as: {Time: "10:00:01", Coordinates: [113.2, 23.1], Status: [RPM 1200rpm, Torque 350N·m, Fuel Consumption 12L / h], Business: [Task ID_99, Load 30 tons]}.

[0025] Step S2. Based on the vehicle's full-time operating state sequence, extract micro-level driving behavior features (such as rapid acceleration / deceleration, prolonged idling), macro-level operational features (such as load factor, empty-running rate), and external environmental features (such as road gradient, congestion level) to construct an operating intensity map reflecting the vehicle's actual usage load. See also Figure 3 As shown, the construction of the running intensity map is achieved through the following steps: Step S201 involves performing time-series analysis on the operating state sequence. By setting acceleration and engine speed thresholds, micro-level driving behavior features are extracted, including the frequency of rapid acceleration and deceleration and the duration of prolonged idling. Specifically, firstly, based on the instantaneous speed field in the time series, the real-time acceleration sequence of the vehicle is calculated using the first-order difference method. Next, physical thresholds for rapid acceleration and deceleration are preset (e.g., setting the rapid acceleration threshold to >2.5 m / s² and the rapid deceleration threshold to <-2.5 m / s²). A sliding time window algorithm is used to scan the acceleration sequence, identifying and counting the number of events that continuously exceed the thresholds and last for a duration greater than a set value (e.g., 2 seconds), thus determining the frequency of rapid acceleration and deceleration. Simultaneously, idling conditions are determined by combining speed and engine speed data. The determination logic is set as "instantaneous speed equals 0 km / h and engine speed is in the idling range (e.g., 600 rpm-900 rpm)," and a time threshold is set (e.g., lasting more than 60 seconds). By traversing the entire time-domain sequence, the lengths of all segments that satisfy this logic and exceed the time threshold are accumulated to obtain the duration of prolonged idling. For example, if the algorithm detects that a vehicle’s speed increases from 10 km / h to 45 km / h in 3 seconds, it is considered a rapid acceleration behavior; if it detects that the engine of a vehicle is running continuously for 5 minutes while the vehicle is stationary, it is counted as a long idling time of 5 minutes.

[0026] Step S202: Combining business scheduling records, calculate macro-operational characteristics and quantify vehicle load rate and empty-run rate. Specifically, firstly, based on timestamp indexes, associate and map the vehicle's full-time-domain operating status sequence with TMS business scheduling data, identifying the task status corresponding to each time segment. Segments within the task execution window with cargo weight greater than zero are marked as "load status," while non-task periods or driving segments without cargo in the task are marked as "empty-running status." Secondly, statistical calculations are performed based on the marking results. The cumulative mileage under "empty-running status" is divided by the total mileage throughout the time period to obtain the empty-running rate. Simultaneously, the load rate is calculated using turnover logic, i.e., summing up (actual cargo weight × mileage for that segment) under all "load statuses" and then dividing by (vehicle rated load × total mileage). For example, a vehicle with a rated load capacity of 30 tons travels 200 kilometers with 24 tons of cargo in a transport mission. After unloading, it travels 50 kilometers empty to return to the depot. The system calculates that the empty run rate for this operation cycle is 20% (50km / 250km), and the overall load rate is 64% ((24 tons × 200km) / (30 tons × 250km)).

[0027] Step S203: Connect to a Geographic Information System (GIS), match vehicle trajectory locations, analyze external environmental features, and obtain road slope and traffic congestion level.

[0028] In practical implementation, a map matching algorithm (such as a matching algorithm based on a Hidden Markov Model) is first used to accurately map the cleaned vehicle GPS trajectory points to specific road segment IDs in the GIS road network, correcting the positioning error caused by satellite signal drift. Here, a path matching mechanism based on a Hidden Markov Model (HMM) can be introduced to address the positioning point drift phenomenon that occurs when vehicles pass under overpasses, in tunnels, or in densely populated areas of high-rise buildings (such as positioning points falling into non-road areas). By calculating the geometric distance and driving direction matching degree between the positioning point and surrounding road segments, the drifting point is forcibly mapped and absorbed to the road centerline with the highest probability. A probabilistic graphical model containing observed and implicit states is constructed; in a specific embodiment, the GPS positioning trajectory point sequence is defined as the observation sequence. ,in

[0029] Let t be the latitude and longitude coordinates; the candidate road segment sequence in the road network is defined as the hidden state sequence. ,in Let t be the actual road segment where the vehicle is located at time t. Indicated by The set of neighboring road segments centered at a radius r = 50m.

[0030] First, calculate the emission probability. Used to characterize positioning points Located on the road section The geometric likelihood is calculated using the following formula: ,in Indicates the location point To candidate road sections Vertical projection distance of the centerline The standard deviation of the positioning error is set (e.g., 10 meters). Next, the transition probability is calculated. Used to characterize the vehicle's movement from the previous time segment of the road Move to the current road segment The topological rationality is calculated using the following formula: ,in The great circle distance between two consecutive positioning points. For road segments in the road network topology arrive The shortest path length, These are the road network constraint parameters. Finally, the Viterbi algorithm is used for dynamic programming to solve for the global posterior probability. The largest optimal state sequence The recursive formula is: Observation points at each time point Vertical projection to the corresponding optimal road segment On the center line, the corrected high-precision trajectory points are obtained, thereby correcting the positioning deviation caused by the multipath effect.

[0031] Secondly, by overlaying a digital elevation model (DEM) or calling high-precision map data, the elevation difference between the start and end points of the matched road segment is extracted, and the longitudinal slope value of the road is calculated in combination with the road segment length. At the same time, real-time traffic situation data from online map service providers is called through API interfaces to obtain the average traffic speed and congestion level of the road segment at that timestamp. For example, the system matches the vehicle coordinates at a certain moment to "G4 Expressway K1024 segment", calculates that the road segment is "4% uphill" based on elevation data, and simultaneously obtains the real-time congestion index of the road segment as "4.5 (severe congestion)", thereby identifying that the vehicle is in a severe external condition of "high-load uphill climbing and frequent starts and stops".

[0032] Step S204: Integrating the aforementioned microscopic, macroscopic, and environmental characteristics, a multi-dimensional feature weighting algorithm is used to construct an operational intensity map reflecting the actual load of vehicles. Specifically, the extracted multi-source feature data, such as the frequency of rapid acceleration / deceleration, load rate, and road gradient, are first standardized using range normalization. The process involves setting physical limit thresholds for each characteristic indicator (e.g., setting the upper limit for emergency braking frequency to 5 times / minute and the upper limit for road gradient to 10%), and then using formulas. Differences in metrics such as frequency per minute, percentage, and degree are uniformly mapped to the dimensionless numerical range [0,1]. These are the original characteristic values ​​(such as the frequency of emergency braking, load rate, road gradient, etc.). The minimum possible value of this feature (often taken as the lower limit of the physical limit, such as the lower limit of the frequency of emergency braking being 0). The maximum possible value of this feature (physical limit, such as setting the maximum frequency of emergency braking to 5 times / minute, and the maximum road gradient to 10%). The normalized dimensionless values ​​are used to eliminate dimensional differences. Next, a feature weight matrix is ​​constructed. The analytic hierarchy process (AHP) is used to build the judgment matrix, or principal component analysis (PCA) is used to calculate the variance contribution rate, determining the weight coefficients of each feature's contribution to vehicle wear. For example, the weight of micro-driving behavior is set to 0.4, the weight of macro-operational characteristics to 0.3, and the weight of external environmental characteristics to 0.3. Then, a multidimensional linear weighted summation model is used to calculate the comprehensive load index for each moment of time, generating continuous time series data. , For a moment The comprehensive load index (values ​​range from 0 to 1, with higher values ​​indicating greater load). For a moment Normalized micro-driving behavior characteristics (such as normalized values ​​of emergency braking frequency). For a moment Normalized macro-operating characteristics (such as normalized load factor). For a moment Normalized external environmental characteristics (such as normalized road slope values). , , The weight coefficients corresponding to the features are then assigned. Finally, the sequence is visualized or stored in a structured manner to form an operational intensity map representing the vehicle's stress state throughout its entire lifecycle. For example, in a transportation task, the system identifies that the vehicle is under specific operating conditions at "14:00:00": emergency braking frequency is 4 times / minute (normalized value 0.8), load rate is 100% (normalized value 1.0), and road gradient is 8% (normalized value 0.8). The algorithm calculates the operational intensity index at that moment by substituting the weights. (1.0 full marks) and mark this period as "extremely high load zone" on the graph as a key input for subsequent evaluation of braking system and engine thermal decay.

[0033] Step S3. Map the operational intensity map to a preset vehicle core component wear model (such as an engine thermal load model or a braking system wear model), calculate the cumulative damage and remaining life of the components under different operating conditions, and output the real-time vehicle health status assessment results. See also Figure 4 The output shows the real-time health status assessment results of the vehicle, which is achieved through the following steps: Step S301: Call the preset vehicle core component loss model library, including engine thermal load model, braking system wear model and transmission component fatigue model. In specific implementation, based on the physical failure mechanism of the components and bench fatigue test data, the mathematical loss model of each key component is pre-constructed and calibrated, and it is encapsulated as a callable algorithm module and stored in the system background. Set refined model logic for different component characteristics: (1) The engine thermal load model uses the Arrhenius equation to describe the relationship between thermal stress and aging rate, and the formula is expressed as ,in, This refers to the engine's real-time operating temperature. For aging rate, It is a constant. For activation energy, As a gas constant, this model reflects how high temperatures exponentially accelerate the aging of seals and the failure of lubricating oil films, helping to quantify thermal stress damage. (2) The brake system wear model adopts Archard's wear law. ,in, For wear volume, For normal load, The sliding distance, For material hardness, The wear coefficient is given. (3) The fatigue model of the transmission component is constructed based on the material SN curve (stress-life curve) and Miner's linear cumulative damage theory, and the cumulative damage degree is defined. ,in To be at stress level The actual number of loops, This represents the fatigue life limit corresponding to this stress level. Finally, the index interface of the model library is configured to support the dynamic loading of corresponding calculation formulas and material constants based on vehicle model ID and component type.

[0034] Step S302 involves mapping the microscopic driving behaviors and environmental features in the operational intensity map to model input parameters, simulating the physical stress response of each component under specific operating conditions. Specifically, this involves first analyzing the time slice data in the operational intensity map to extract key state variables, including instantaneous acceleration, engine speed, load mass, and road gradient. Second, based on vehicle dynamics equations, these macroscopic state variables are converted into physical load boundary conditions acting on the component level. For example, using Newton's second law and the gradient resistance formula, the vehicle deceleration, total mass, and gradient angle are converted into the friction torque and heat flux of the braking system, or the high-load engine operating condition is mapped into the explosion pressure and thermal load within the cylinder. Finally, the converted boundary conditions are input into the physical loss model called in the first step, iteratively calculating the instantaneous stress amplitude or temperature field distribution borne by the component at that moment. For example, if the system reads that a vehicle is fully loaded (30 tons) and brakes suddenly (deceleration -3.5m / s²) on an 8% downhill section at a certain moment, the algorithm first calculates the total braking energy required to overcome the gravitational component and inertia; then, combined with the braking distribution ratio, it determines the frictional work that the front wheel brake discs need to bear; then, it inputs the thermal load model to simulate the thermal shock response of the brake disc surface temperature rising sharply from 150℃ to 500℃ within 3 seconds, which is used as the stress input for subsequent calculation of thermal fatigue crack propagation.

[0035] Step S303: Apply the fatigue damage accumulation rule, combined with the material properties of the components, to calculate the cumulative damage degree of each core component and predict its remaining service life. Specifically, firstly, Miner's Rule is introduced as the core algorithm framework, and the Rainflow Counting Method is used to perform cyclic counting on the complex stress-time history output from the second step, extracting several stress cycles with different amplitudes. Secondly, based on the material property data of the components, especially the SN curve (stress-life curve), the limit number of cycles at which fatigue failure occurs under each stress amplitude is queried. Next, the damage fraction caused by a single stress cycle is calculated. The algorithm linearly accumulates the damage scores from all cycles to obtain the current cumulative damage level (D) of the component. Finally, based on the current cumulative damage rate and the length of service, the remaining service life (RUL) of the component is estimated. For example, for the aforementioned brake disc that has undergone high-temperature emergency braking, the algorithm identifies a significant thermal stress cycle through rainflow counting. Consulting the SN curve of gray cast iron material, it is found that the limit life at this stress level is 10,000 cycles. Therefore, the damage increment caused by this braking is 0.0001. If the historical cumulative damage level recorded by the system has reached 0.85, the brake disc is determined to have entered the fatigue failure warning zone, and the remaining service life is predicted to be only about 1,500 similar braking operations.

[0036] Step S304: Integrate multi-source heterogeneous data, use large-scale artificial intelligence model (LLM) for comprehensive reasoning, and generate a real-time vehicle health status assessment report that includes health scores of key components, fault risk warnings, and maintenance suggestions.

[0037] In practical implementation, the first step is to construct a vertical-category large language model based on an industrial knowledge graph and a maintenance expert experience base for fine-tuning. This is achieved through the following steps: First, a domain-specific instruction dataset is constructed. Natural language processing techniques are used to extract entity relationships from vehicle maintenance manuals, DTC fault code parsing libraries, and historical expert diagnostic cases to build a knowledge graph of "fault phenomenon - root cause analysis - repair solution," generating tens of thousands of instruction tuning samples. Second, an open-source base model with a suitable number of parameters (such as Llama) is selected. 3. Or ChatGLM), employing Low-Rank Adaptation (LoRA) technology for fine-tuning, specifically injecting a trainable matrix of rank r (e.g., r=8) into the attention module of the Transformer layer, freezing the pre-trained weights, and optimizing the model's reasoning ability regarding automotive engineering terminology and fault causal chains through supervised learning; simultaneously, designing a hierarchical prompt engineering template, clearly divided into "System Instruction Layer" (defining the role of senior maintenance expert), "Context Layer" (filling in vehicle ID, cumulative mileage, and historical maintenance records), "Observation Layer" (filling in component cumulative damage, sensor timing features, and current fault codes), and "Format Layer" (specifying the JSON structure and scoring criteria). For example, in implementation, the system fills "DTC: P0087 (low fuel rail pressure)" and "high-pressure fuel pump cumulative damage: 0.92" into the Prompt template. Based on the knowledge learned through fine-tuning, the model infers that "the high-pressure fuel pump has an extremely high risk of mechanical failure" and outputs the suggestion: "Replace the high-pressure fuel pump and fuel metering unit immediately."

[0038] Specifically, the AI ​​model used for real-time vehicle health status assessment in this technology is a customized vertical language model for the field of vehicle maintenance and diagnosis. It adopts a bottom-up four-layer architecture design, with each layer sequentially supporting and empowering the model. The bottom layer provides basic reasoning capabilities for the model, the middle layer endows the model with vehicle-specific attributes and customizes interaction rules, and the top layer provides a reliability guarantee for the model's output. The four-layer structure works together to achieve fusion reasoning of multi-source heterogeneous vehicle data, root cause analysis of faults, and generation of executable maintenance suggestions. The key technical points of each layer and the connection relationships between layers are as follows: The first layer is the open-source foundational large model layer, which serves as the basic inference capability layer for vertical-category large models. Technically, it selects open-source large language models with moderate parameter counts, such as Llama 3 and ChatGLM, as foundational models, directly reusing their Transformer core architecture. It retains the multi-head self-attention mechanism and feedforward neural network core modules, and achieves general natural language understanding, text generation, and logical reasoning capabilities based on pre-training results. At the same time, the pre-trained weights at the bottom layer of the foundational model are completely frozen without modification, effectively reducing development costs and time. This layer serves as the basic support layer for the entire vertical-category large model, providing a general algorithm architecture and basic inference capabilities for the upper domain fine-tuning layer. It is the technical carrier of the domain fine-tuning layer, and all customized modifications to the domain fine-tuning layer are based on the Transformer architecture of this layer.

[0039] The second layer is the domain fine-tuning layer, which is the core layer for professional empowerment, transforming the general base model into a professional vehicle repair and diagnosis model. Its core technologies include three aspects: First, constructing a knowledge carrier specific to the vehicle repair domain. This involves extracting entity relationships from vehicle repair manuals, DTC fault code parsing libraries, and expert historical diagnostic cases using natural language processing technology. This builds an industrial knowledge graph centered on fault phenomena, root cause analysis, and repair solutions, and generates tens of thousands of structured instruction fine-tuning samples based on this graph, representing "vehicle data input - professional diagnostic output." Second, employing low-rank adaptation (LoRA) technology for lightweight, targeted fine-tuning. A trainable sparse matrix with rank r=8 is injected into the attention module of the Transformer layer of the first-layer base model as a dedicated domain knowledge learning module. Third, using supervised learning, the newly added LoRA trainable matrix is ​​trained separately to optimize the model's understanding and reasoning ability regarding automotive engineering terminology, fault causal chains, and vehicle diagnostic logic, without altering the pre-training weights of the first-layer base model. This layer is a customized and lightweight modification based on the first layer of open source base model. It inherits the general reasoning capabilities of the first layer and endows it with professional understanding of the vehicle maintenance field. At the same time, it provides a domain knowledge reasoning kernel for the third layer of layered Prompt engineering template layer, ensuring that the professional vehicle data input to the third layer can be effectively identified and analyzed by the model.

[0040] The third layer is the hierarchical Prompt engineering template layer, which is the standardized interaction layer for the vertical category's large model. Its core design uses a fixed four-layer structured Prompt template as a standardized interaction interface between the vertical category's large model and the vehicle data system, enabling the adaptation and conversion of vehicle hard data with the large model. This template is a configurable structured framework with fixed slots at each level. The system instruction layer pre-sets fixed text, defining the role of a senior operational vehicle maintenance expert and clarifying the diagnostic rules and inference boundaries for vehicle health assessment. The context layer sets up structured data slots for filling in static vehicle data such as vehicle ID, cumulative mileage, historical maintenance records, and engine oil type. The observation data layer sets up structured data slots for filling in real-time dynamic vehicle data such as component cumulative damage, sensor timing characteristics, current fault codes, and driving conditions. The output specification layer pre-sets fixed format requirements, forcing the model to output structured data conforming to the JSON Schema definition, including health scores (0-100). The system includes mandatory fields such as risk level, natural language diagnostic conclusion, and maintenance suggestion list. A data adaptation interface has also been developed to format vehicle dynamic operation data, static archive data, and component cumulative damage calculation results, automatically mapping them to corresponding slots in the template to form a complete standardized inference context. This layer serves as the standardized input / output interface for the vertical category's large model. It connects upwards to the vehicle data acquisition and calculation system and downwards to input formatted vehicle professional data into the second-layer domain fine-tuning layer to trigger its professional inference capabilities. Simultaneously, it receives the inference results from the second-layer domain fine-tuning layer, constrains the output format according to the requirements of the output specification layer, and provides standardized data to be verified for the fourth-layer rule verification layer.

[0041] The fourth layer is the rule validation layer, which serves as the bottom layer for the reliability of the vertical category's large-scale model results. Deployed as an independent validation layer at the model output, it filters out illusory outputs from the large model through a multi-dimensional validation mechanism, ensuring the accuracy, logic, and executability of the evaluation results. If the output fails validation, a regeneration request with error messages is triggered until all constraints are met. Specific validation mechanisms include: a structured data validator based on JSON Schema and regular expressions, which verifies whether the structured data output by the model contains all required fields, whether the data type is correct, and whether the format is compliant, automatically correcting minor format errors and triggering model regeneration for missing fields / incorrect data types; an entity link validation mechanism based on the enterprise vehicle bill of materials (BOM) and standard maintenance time library, which compares the parts involved in the maintenance suggestions generated by the model with the actual SKUs in the database, eliminating non-existent parts and parts with mismatched names and models, thus blocking illusory outputs at the entity level; and physical / physical link validation mechanisms with embedded hard-coded rule bases and causal validation algorithms. The business logic rule engine pre-defines physical and business logic constraints within the vehicle diagnostic domain. Its core components include strong correlation constraints between health scores and risk levels, causal closed-loop constraints between fault phenomena and repair suggestions, and matching constraints between repair suggestions and operational vehicle maintenance standards. Outputs that violate these rules trigger model regeneration. This layer serves as the final output layer of the entire vertical model, directly receiving the standardized model inference results from the third-layer hierarchical Prompt engineering template layer. It performs multi-dimensional verification and correction on these results, ultimately outputting a rigorously cleaned and directly executable real-time vehicle health status assessment report, providing reliable decision-making support for downstream fleet management and maintenance scheduling systems.

[0042] The vertical language model used for vehicle health status assessment in this solution requires targeted training. The training employs a lightweight, domain-specific fine-tuning training mode. The core objective is to enable the open-source model, which originally only possesses general natural language understanding and logical reasoning capabilities, to acquire professional knowledge in the field of vehicle maintenance and diagnosis, identify specialized vehicle data, and establish causal reasoning logic between faults and repairs, adapting to the specific scenario of multi-source heterogeneous data fusion and reasoning for vehicles. The model's training utilizes Low-Rank Adaptation (LoRA) lightweight fine-tuning technology. During training, all pre-trained weights of the open-source base model are frozen, and only a trainable sparse matrix with rank r=8 is injected into the attention module of its Transformer layer. This matrix serves as a dedicated module for domain knowledge learning, and supervised learning optimization is performed only on this small matrix. This approach preserves the general reasoning capabilities of the base model while significantly reducing the number of training parameters. The model's specific training involves supervised learning based on a vehicle repair domain-specific dataset. Before training, a domain-specific instruction dataset needs to be constructed. Entity relationships are extracted from vehicle repair manuals, DTC fault code parsing libraries, and expert historical diagnostic cases using natural language processing techniques. This constructs an industrial knowledge graph centered on "fault phenomenon - root cause analysis - repair solution." Based on this knowledge graph, tens of thousands of structured instruction fine-tuning samples of "vehicle professional data input → standardized diagnostic output" are generated as the core training material. Subsequently, a lightweight fine-tuning training architecture is built, determining that only the newly added LoRA matrix is ​​to be trained. The constructed domain instruction fine-tuning samples are then input into the training architecture, and the LoRA matrix is ​​optimized through supervised learning. This allows the model to master automotive engineering terminology, fault causal chain logic, and vehicle diagnostic reasoning rules, achieving the transformation from a general model to a professional vehicle repair and diagnostic model.

[0043] Secondly, the cumulative damage values ​​of each component calculated in the third step, real-time sensor readings, and vehicle file data are dynamically mapped to the corresponding slots in the Prompt template to form a complete inference context, which is then input into the fine-tuned large model. In practice, the following steps are taken: First, a data adaptation interface is developed to capture real-time dynamic operating data of the vehicle (such as engine speed, oil pressure, and brake disc temperature) and static archive information (such as the last maintenance time and oil type). The cumulative damage of each component output in step S3 (such as $D_{brake}=0.7$) is formatted and filled into the "Observation" layer of the Prompt template. Second, the filled Prompt is input into the large model inference terminal. The model uses a multi-head self-attention mechanism to deeply encode the input sequence, calculate the weight relationship between different sensor data features, mine the implicit correlation between multi-source heterogeneous data, and identify system-level coupling risks that are difficult to detect by a single threshold method. Next, the chain-of-thought inference mode is triggered to guide the model to analyze the complex fault mode under the combined effect of multiple variables based on physical common sense and maintenance knowledge base. For example, it can deduce the nonlinear decay effect of "long-term high engine load" superimposed with "edge rise in coolant temperature" on "oil lubrication performance". Finally, the model generates a JSON-compliant output based on the inference results. The structured output defined by the schema includes a comprehensive health score (0-100 points), risk level classification (such as "healthy", "sub-healthy", "high risk"), fault diagnosis conclusions in natural language, and a list of specific maintenance suggestions. For example, during a long-distance transport, the system detected that the vehicle's "engine load rate is consistently >85%" and "coolant temperature is 98℃ (not exceeding the alarm threshold of 105℃)". However, combined with the weak signal of "increased fluctuations in oil pressure sensor", the model infers that "although a single indicator is not exceeded, the continuous high heat load has led to thermal decay of oil viscosity, increasing the risk of lubrication failure", and thus outputs a score of "72 points", a risk level of "medium-high risk", and suggests "stopping at the nearest service area to cool down and checking the oil quality".

[0044] Finally, a deterministic rule validation layer (Guardrails) is set up to perform multi-dimensional post-processing and logical verification of the model output, ensuring the accuracy and executability of the evaluation results. In specific implementation, firstly, a structured data validator is deployed, using JSON Schema or regular expressions to enforce constraints on the model output format, automatically detecting and correcting missing fields or incorrect data types, ensuring that the output content can be parsed by downstream systems. Secondly, an entity link validation mechanism based on the vehicle bill of materials (BOM) and standard maintenance time library is constructed, comparing the "suggested replacement parts" generated by the model with the actual SKUs in the database, eliminating non-existent part names or mismatched models, effectively preventing the "illusionary" output of the large model. Simultaneously, a rule engine for consistency between physical common sense and business logic is embedded, such as setting a strong constraint that "the risk level must be high when the health score is <60," and verifying the causal closed loop between fault phenomena and maintenance suggestions (e.g., "overheating water" should not match the suggestion of "replacing tires"). If the output fails validation, the system will automatically trigger a regeneration request with error messages until all constraints are met, finally outputting a rigorously cleaned health status assessment report. For example, the system constructs the following Prompt input: "[Vehicle Information] Heavy truck, mileage 150,000 km; [Status Data] Brake disc cumulative damage 0.88 (threshold 1.0), recent emergency braking frequency 5 times / hour, ABS sensor signal intermittent loss; [Task] Assess health status and provide recommendations." The model outputs after inference: "{Health Score: 62, Risk Level: High, Diagnosis: Brake disc is approaching fatigue limit, and abnormal ABS signal may lead to anti-lock braking failure, posing a serious safety hazard; Recommendations: [1. Replace the front axle brake disc immediately; 2. Check the left front wheel speed sensor wiring harness and gear ring; 3. It is recommended to suspend operation for comprehensive maintenance]}." Step S4. Combining the aforementioned operational intensity map with standardized energy efficiency benchmarks across vehicle types and routes, perform multi-dimensional weighted calculations to generate a vehicle operational efficiency quantification index, used to identify inefficient operating vehicles or abnormal driving patterns. See also Figure 5 The generation of the vehicle operation efficiency quantification index is achieved through the following steps: Step S401 involves constructing a standardized energy efficiency benchmark library across vehicle models and routes. Big data statistical methods are used to eliminate differences in physical parameters between different vehicle models and basic road conditions, establishing a unified energy efficiency evaluation standard. Specifically, this involves first aggregating massive amounts of historical operational data from the fleet over a long period. Based on vehicle physical attributes (such as weight, displacement, and fuel type) and route environmental characteristics (such as plain highways, mountain national roads, and urban congestion), multi-dimensional clustering is performed to establish a statistical distribution model of energy consumption under specific operating conditions. Second, "fuel consumption per ton per 100 kilometers" or "equivalent energy consumption per unit of work" is introduced as a core evaluation indicator. Regression analysis is used to calculate the difficulty coefficients for different routes (e.g., 1.2 for mountain routes and 1.0 for plain routes), correcting and normalizing the original energy consumption data to eliminate energy consumption bias caused by objective road conditions. Finally, based on the corrected data, energy efficiency benchmarks for each sub-cluster are calculated (e.g., the 20th percentile of the statistical distribution is taken as the excellent benchmark, and the median as the average benchmark), and stored in the benchmark library for future reference. For example, for data on a 49-ton heavy truck driving on mountainous routes, the system first converts its fuel consumption of 35L per 100 kilometers into energy consumption per unit load, then divides it by the difficulty coefficient of the mountainous route of 1.2, and converts it into the equivalent energy consumption under standard plain conditions, so that it can be compared with the energy efficiency level of light trucks driving on plains under the same scale.

[0045] Step S402 involves inputting the operational intensity map data of the vehicle to be evaluated into the benchmark library and calculating its relative energy consumption level and efficiency deviation under the same operating load. Specifically, the operational intensity map of the vehicle to be evaluated is first analyzed to extract feature vectors representing operational intensity, including average load rate, road slope factor, and congestion duration percentage, while simultaneously obtaining the actual fuel consumption during that period. Secondly, using the K-Nearest Neighbors (KNN) algorithm or multidimensional lookup table method, the benchmark operating condition cluster with the closest Euclidean distance to the current feature vector is retrieved from the standardized energy efficiency benchmark library to obtain the standard energy efficiency reference value for that operating condition (such as the median P50 value and the excellent P20 value). Finally, an efficiency deviation calculation model is constructed, comparing the actual energy consumption with the standard energy efficiency reference value to calculate the relative energy consumption ratio. Reasonable deviations caused by force majeure (such as extreme weather) are eliminated, quantifying the efficiency drop purely caused by vehicle performance degradation or driving behavior. For example, the system reads that a vehicle's actual fuel consumption per 100 kilometers under "heavy load + hilly" conditions is 38L, while the average fuel consumption under the same conditions in the benchmark database is 34L. The system calculates that the relative energy consumption level is 111.7%, which means there is an efficiency negative deviation of 11.7%. The system initially judges that this deviation may be due to the driver's failure to make reasonable use of inertial coasting.

[0046] Step S403: Establish a multi-dimensional weighting system to calculate the relative energy consumption level, load task completion rate, and driving behavior smoothness using weighted calculations, generating a unique quantitative index for vehicle operating efficiency. In practice, firstly, the Analytic Hierarchy Process (AHP) or the Entropy Weight Method is used to determine the weight coefficients of each evaluation dimension to balance cost, efficiency, and safety objectives. For example, the weight of relative energy consumption level is set to 0.4, load task completion rate to 0.4, and driving behavior smoothness to 0.2. Secondly, the index data is standardized and positively processed. Inverse indicators such as "relative energy consumption level" (smaller values ​​are better) are transformed into positive scores of [0, 100] through linear transformation, and positive indicators such as "load task completion rate" are directly mapped. Finally, a linear weighted summation model is used to aggregate the scores of each dimension to generate the final quantitative index. For example, a vehicle has a relative energy consumption level of 110% (i.e., 10% over-consumption, which scores 80 points after algorithm mapping), a load-bearing task completion rate of 100% (score of 100 points), and a driving behavior smoothness score of 90 points. Based on the aforementioned weights, the quantitative index of the vehicle's operating efficiency is calculated as $80\times0.4 + 100\times0.4 + 90\times0.2 = 90$. This index directly reflects the vehicle's comprehensive efficiency in cost control and task delivery.

[0047] Step S404 involves horizontal ranking and threshold determination based on the quantitative index to accurately identify inefficient operating vehicles and correlate abnormal driving patterns to assist in targeted management. Specifically, firstly, a fleet-level energy efficiency ranking and grading evaluation mechanism is constructed. The quantitative indexes of all vehicles' operating efficiency are arranged in descending order of value, and grading thresholds are set (e.g., an index below 80 is defined as "inefficient operation," and above 95 is defined as "benchmark operation") to quickly identify the vehicles at the bottom of the rankings requiring close attention. Secondly, for the selected inefficient vehicles, a drill-down attribution analysis of abnormal patterns is performed. Statistical methods (such as Pearson correlation coefficient analysis) or association rule mining algorithms are used to correlate the vehicle's inefficiency index with the micro-driving behavior characteristics extracted in step S2 (such as frequency of rapid acceleration / deceleration, idling time, and speeding percentage) in multiple dimensions to identify the core causal patterns leading to inefficiency. Finally, based on the attribution results, targeted management strategies are generated, and a comprehensive management report is output, including driver behavior correction suggestions (such as reducing idling), vehicle hardware inspection prompts (such as checking tire pressure or air filter), and route scheduling optimization schemes. For example, when the system conducted a weekly evaluation of a logistics fleet, it identified that vehicle A's efficiency index was only 72 points (threshold 80 points), ranking last. Through correlation analysis, it was found that the vehicle's "frequency of rapid acceleration" was 200% higher than the fleet average, and it showed a strong positive correlation with "instantaneous high fuel consumption" (correlation coefficient 0.92). Based on this, the system determined that the main reason for the vehicle's inefficiency was "aggressive driving style leading to fuel waste," and automatically generated management suggestions: "It is recommended to conduct predictive driving training for the driver of this vehicle, focusing on assessing smooth start operation."

[0048] Step S5. Based on the vehicle's real-time health status assessment results and operational efficiency quantification index, and under the constraint of ensuring operational task scheduling, use a dynamic programming algorithm to generate the optimal personalized maintenance cycle and maintenance item list, and output a dynamic maintenance execution plan. This is achieved through the following steps: Step S501 integrates the real-time vehicle health status assessment results with the operational efficiency quantification index, and simultaneously obtains the vehicle's scheduled operational tasks as a time window constraint. Specifically, firstly, the vehicle health status assessment report output in step S3 (including the predicted remaining life (RUL) of key components and the failure risk level) and the operational efficiency quantification index generated in step S4 are read through the system's internal data bus to establish a current "health-efficiency" dual-dimensional status profile of the vehicle. Secondly, the scheduling database of the TMS transportation management system is called via API to retrieve the task schedule table for the target vehicle within a future preset period (e.g., 30 days), analyzing the start and end times, priorities, and route mileage of the tasks, and identifying non-operational periods between tasks as "candidate maintenance time windows." Finally, the vehicle status profile and candidate time windows are time-series aligned and structured to construct a decision input set that includes the urgency of maintenance needs and time feasibility constraints. For example, the system reads that the remaining life of a heavy truck's braking system is less than 2,000 kilometers (high risk), and the efficiency index drops to 75 points due to braking lag; at the same time, TMS data shows that the vehicle will carry out inter-provincial transportation in the next 3 days and stand by at the base from the 4th to the 5th day; based on this, the system generates input constraints: maintenance must be intervened before the cumulative driving of 2,000 kilometers, and the feasible time window is locked as "the 4th-5th day".

[0049] Step S502: Construct a multi-objective dynamic programming model, setting the joint optimization objectives as minimizing maintenance costs, minimizing safety risks, and minimizing the impact on operational tasks. In specific implementation, first define the model's decision variables and state space, setting the decision variables as whether to perform maintenance actions within the candidate time window $t$. (Values ​​0 or 1) and a specific set of maintenance items (e.g., changing engine oil or brake pads), the state space is set to the remaining lifespan of each critical component of the vehicle at time t. and cumulative damage Secondly, a multi-objective loss function J is constructed, which includes direct maintenance costs, implicit failure risk costs, and operational interruption penalty costs. The formula is expressed as follows: J =min t is the time step index (such as day or hour). For the planning cycle; In time The set of maintenance and repair projects to be performed (such as changing engine oil, brake pads, etc.); Direct maintenance costs include spare parts costs and labor costs; Key components of the vehicle at all times Remaining Useful Life; The failure probability function based on remaining lifetime typically increases as the remaining lifetime decreases; Indirect losses caused by the malfunction, such as rescue costs and cargo loss; Loss of transport capacity or penalties (such as fines for delayed delivery) due to maintenance taking up operating time. Weighting coefficients are used to balance the relative importance of the three objectives (cost, risk, and delay).

[0050] Finally, constraints are introduced, including "decision variables". Indicates whether it is within a time window Maintenance and repair services, and a collection of specific maintenance and repair projects. "Any component" The model ensures the feasibility of a solution by specifying that the maintenance interval must not fall below a safety threshold (e.g., 500 km) and that the maintenance duration must not exceed the candidate time window length. For example, for a truck that needs to replace its clutch disc, the model calculates the following: Solution A (maintenance during the break on day 3): direct cost 2000 yuan, risk cost 100 yuan, delay cost 0 yuan, total score 2100; Solution B (maintenance during the break on day 8): direct cost 2000 yuan, risk cost 800 yuan (due to approaching the lifespan limit), delay cost 3000 yuan (penalty), total score 5800. Based on this, the model selects Solution A as the target solution.

[0051] Step S503: Solve the model under the premise of meeting the scheduling constraints to determine the optimal maintenance intervention time (personalized maintenance cycle) and the specific damaged parts to be handled at that time (maintenance item list). In specific implementation, firstly, a dynamic programming algorithm or a heuristic search algorithm (such as a genetic algorithm) is used to solve the multi-objective optimization model constructed in the second step, searching for the global optimal solution in all feasible solution spaces that meet the "parts' remaining life safety threshold" and "maintenance duration is less than the idle window"; secondly, calculate the comprehensive cost function value of each candidate maintenance scheme, weigh the loss of the remaining value of the parts (premature maintenance) against the failure risk cost (late maintenance), and lock the time node with the lowest total cost as the optimal maintenance intervention time; finally, based on the cumulative damage status of the vehicle corresponding to this time, automatically generate a maintenance item list to be performed, including vulnerable parts to be replaced and potential hazards to be inspected. For example, the system simulates the operation plan of a logistics vehicle for the next 30 days and finds that if maintenance is carried out on the 10th day, although it will not affect the delivery task, the brake pads will still have 20% of their remaining lifespan, resulting in a waste of resources; if it is postponed to the 25th day, although the lifespan will be fully utilized, the probability of failure during the journey will surge to 15%; finally, the algorithm finds that the task gap period on the 18th day is the best time (with the lowest overall cost) and outputs "replace the rear axle brake pads and replace the oil filter" as the specific items for this maintenance.

[0052] Step S504 outputs a dynamic maintenance execution plan containing specific execution time, work content, and required resources, achieving collaborative optimization of maintenance plans and operational tasks. In practice, firstly, based on the optimal maintenance strategy obtained in step three, a structured electronic maintenance work order is generated, detailing the suggested entry time window, estimated work hours, required repair item codes, and corresponding operating procedures. Secondly, through an API interface, it links with the enterprise's inventory management system (WMS) and maintenance resource scheduling system to automatically check and pre-lock required spare parts (such as specific models of filters and brake pads), while matching available workstations and qualified technicians at repair stations to ensure the availability of maintenance resources. Finally, the generated complete plan is pushed to fleet managers, drivers, and designated repair stations via mobile terminals. After the maintenance work is completed, the vehicle health record and the prediction model for the next maintenance cycle are updated based on actual feedback data. For example, the system ultimately outputs a dynamic maintenance execution order: "It is recommended that the vehicle go to the North District Repair Center on November 15th from 08:00 to 12:00 (non-operational off-peak period) to perform the 'replacement of rear axle brake pads' and 'inspection of turbocharger' items. The required spare parts (item P-2024) have been reserved, and the estimated time is 3.5 hours." In this way, the safety hazards of the vehicle are accurately eliminated without affecting the long-distance transportation task the next day.

[0053] This invention solves the problems of noise, missing data, and spatiotemporal misalignment in multi-source data such as trajectory, CAN bus, service scheduling, and historical maintenance by constructing a hardware and software collaborative multi-source data acquisition network and combining Kalman filtering for noise reduction, interpolation completion, multi-scale spatiotemporal alignment, and standardized encapsulation technology. It generates a vehicle full-time domain operation status sequence with unified time granularity and unified dimensions, breaks down "data silos," and provides a precise and standardized data source for all subsequent analysis stages, greatly improving the utilization value of vehicle operation data.

[0054] This invention extracts three core features: micro-driving behavior, macro-operation, and external environment. It uses range standardization to eliminate dimensional differences, combines the analytic hierarchy process (AHP) to determine feature weights, and constructs an operational intensity map. This enables dynamic and accurate quantification of vehicle operating load. For the first time, it incorporates multi-dimensional factors of "people-vehicle-road-goods" into the load assessment system, abandoning the limitations of traditional single-mileage assessment. It truly reflects the actual usage load of vehicles under different operating conditions and provides realistic operating condition inputs for calculating the wear and tear of core components.

[0055] This invention constructs dedicated loss models for engines, braking systems, and transmission components based on the physical failure mechanisms of parts. Combining Miner's linear cumulative damage theory and rainflow counting method, it calculates the cumulative damage degree and remaining life of components, achieving quantitative assessment of component loss. Simultaneously, it constructs a customized vertical language model for vehicle maintenance, and through LoRA lightweight fine-tuning, hierarchical Prompt engineering, and multi-dimensional rule verification, it achieves fusion reasoning of multi-source heterogeneous data. This model can identify system-level coupled fault risks that are difficult to detect using a single threshold method, and outputs a structured assessment report containing health scores, fault warnings, and maintenance suggestions. This shifts vehicle fault diagnosis from post-event handling to pre-event warning, significantly reducing the probability of sudden faults and improving the driving safety of commercial vehicles.

[0056] This invention constructs a standardized energy efficiency benchmark library that integrates vehicle physical attributes and route environmental characteristics. It introduces the core indicator of "fuel consumption per ton per 100 kilometers" and corrects energy consumption data by combining route difficulty coefficients, establishing a unified energy efficiency evaluation scale and solving the problem of not being able to compare energy efficiency across different vehicle models and road conditions. By generating a quantitative index of operational efficiency through multi-dimensional weighted calculations and combining attribution analysis, it accurately identifies inefficient operating vehicles and abnormal driving modes, providing targeted management strategies for driver behavior correction, vehicle hardware optimization, and route scheduling adjustments. This effectively reduces vehicle operating energy consumption and improves the overall operational efficiency of the fleet.

[0057] This invention integrates vehicle health status assessment results, operational efficiency quantification index, and operational task scheduling constraints to construct a multi-objective dynamic programming model aimed at minimizing maintenance costs, safety risks, and operational impact. Through algorithmic solutions, it determines the optimal timing for maintenance intervention and the maintenance project list, achieving a shift from "periodic maintenance" to "state-based, personalized dynamic maintenance." Simultaneously, it links with inventory management and maintenance resource scheduling systems to ensure the availability of maintenance resources, enabling deep collaboration between maintenance plans and operational tasks. This avoids resource waste caused by over-maintenance and the risk of malfunctions due to under-maintenance, while effectively mitigating conflicts between maintenance and transportation tasks. This significantly improves the scientific rigor and feasibility of maintenance strategies and reduces overall enterprise operating costs.

[0058] This invention forms a closed-loop technology process from multi-source data acquisition and processing, to operational load quantification, health assessment, efficiency analysis, and then to dynamic maintenance plan generation, execution, and data feedback. All analysis steps are completed automatically, and maintenance plans can be directly pushed to relevant terminals and linked to resource systems, significantly reducing the cost of manual intervention. At the same time, the integrated application of technologies such as vertical large language models, dynamic programming algorithms, and machine learning enables intelligent decision-making in vehicle management, improving the precision of fleet operation management and overall operational efficiency.

[0059] The technical solution of this invention can dynamically load the corresponding loss model and parameters according to vehicle model ID and component type. The energy efficiency benchmark library can be continuously iterated and optimized based on fleet historical data. The Prompt template and rule validation layer of the vertical large language model can be flexibly configured, which can adapt to different types of operating vehicles such as heavy trucks, light trucks, and buses, as well as different route scenarios such as plain expressways, mountain national highways, and urban congestion. It has good versatility and scalability and can be widely used in various logistics and transportation and commercial fleet management companies.

[0060] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

[0061] A computer device according to embodiments of the present disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0062] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the computer device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory, causing the computer device to perform all or part of the steps of the learning outcome prediction method based on learning behavior data mining of the foregoing embodiments of this disclosure.

[0063] like Figure 7 This is a schematic diagram of a computer device provided for an embodiment of the present disclosure. It illustrates a structural schematic diagram suitable for implementing the computer device in the embodiments of the present disclosure. Figure 7 The computer device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0064] like Figure 7As shown, a computer device may include a processor (such as a central processing unit, graphics processing unit, etc.), which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) or programs loaded from storage devices into random access memory (RAM). The RAM also stores various programs and data required for the operation of the computer device. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0065] Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard drives; and communication devices. Communication devices allow the computer device to communicate wirelessly or wiredly with other devices (such as edge computing devices) to exchange data. Although a computer device with various devices is illustrated, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0066] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processor, all or part of the steps of the learning outcome prediction method based on learning behavior data mining according to embodiments of this disclosure are performed.

[0067] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

Claims

1. A vehicle analysis method based on trajectory data and driving behavior, characterized in that, Includes the following steps: Multi-source heterogeneous data of operating vehicles are acquired, and after multi-scale spatiotemporal alignment and data cleaning, a standardized vehicle full-time domain operating status sequence is generated. Based on the vehicle's full-time operating state sequence, three types of features are extracted: micro-driving behavior, macro-operation, and external environment. These three types of features are then fused to construct an operating intensity map that reflects the actual usage load of the vehicle. The operational intensity map is mapped to a preset vehicle core component loss model to calculate the cumulative damage and remaining life of the components. Multi-source heterogeneous data is fused and reasoned through a customized vertical large language model to output the real-time health status assessment results of the vehicle. By combining the aforementioned operational intensity map with standardized energy efficiency benchmarks across vehicle types and routes, a multi-dimensional weighted calculation is performed to generate a quantitative index of vehicle operating efficiency, identifying inefficient operating vehicles and abnormal driving modes. Based on the real-time health status assessment results of the vehicle and the quantitative index of operational efficiency, combined with the operational task scheduling constraints, the optimal personalized maintenance cycle and maintenance item list are generated through dynamic programming algorithm, and a dynamic maintenance execution plan is output. Specifically, the extraction of three types of features—micro-driving behavior, macro-operation, and external environment—based on the vehicle's full-time operating state sequence includes: A time-series analysis was performed on the vehicle's full-time operating state sequence. By setting acceleration and engine speed thresholds, micro-driving behavior features were extracted, including the frequency of rapid acceleration and deceleration and the duration of long-term idling. By combining vehicle dispatch records, macro-operational characteristics are calculated, and the vehicle load rate and empty running rate are quantified to obtain the macro-operational data. By associating with a geographic information system, matching vehicle trajectory locations, analyzing external environmental features, and statistically analyzing road slope and traffic congestion levels, the external environmental features are obtained.

2. The method according to claim 1, characterized in that, The steps, including multi-scale spatiotemporal alignment and data cleaning, generate a standardized vehicle full-time domain operating state sequence, specifically including: The multi-source heterogeneous data is preprocessed to obtain a multi-source heterogeneous standard dataset. The preprocessing specifically includes removing noisy data, removing duplicate records, and filling in missing values. Multi-scale spatiotemporal alignment is performed on the multi-source heterogeneous standard dataset to obtain a multi-source heterogeneous aligned dataset, which maps data with different sampling frequencies to a unified time reference axis. The multi-source heterogeneous aligned dataset is standardized and encapsulated to output a full-time-domain vehicle operating state sequence.

3. The method according to claim 2, characterized in that, The process of integrating three types of features and constructing an operational intensity map reflecting the actual usage load of vehicles specifically includes: The extracted three types of feature data are subjected to range standardization, and the physical limit threshold of the feature index is set to uniformly map them to the dimensionless numerical range of [0,1]. Construct a feature weight matrix, use the analytic hierarchy process (AHP) to construct a judgment matrix or use principal component analysis to calculate the variance contribution rate, and determine the weight coefficients of the features' contribution to vehicle wear. A multidimensional linear weighted summation model is used to calculate the comprehensive load index at each moment according to the time step, generating continuous time series data; The sequence is visualized or stored in a structured manner to form an operational intensity map characterizing the vehicle's stress state throughout its entire life cycle.

4. The method according to claim 3, characterized in that, The associated geographic information system, by matching vehicle trajectory locations, analyzing external environmental features, and statistically analyzing road gradients and traffic congestion levels to obtain the external environmental features, specifically includes: To address the issue of location point drift, a path matching mechanism based on a hidden Markov model is used. By calculating the geometric distance and driving direction matching degree between the location point and surrounding road segments, the drifting point is forcibly mapped and absorbed to the road centerline with the highest probability. By overlaying digital elevation models or calling high-precision map data, the elevation difference between the start and end points of the matching road segment is extracted, and the longitudinal slope value of the road is calculated in combination with the road segment length. By accessing real-time traffic data from online map service providers, the average traffic speed and congestion level of road segments at the specified timestamp can be obtained.

5. The method according to claim 1, characterized in that, The process of mapping the operational intensity map to a preset vehicle core component loss model, calculating the cumulative damage and remaining lifespan of the components, fusing multi-source heterogeneous data, and inferring through a customized vertical large language model to output the real-time vehicle health status assessment results specifically includes: The system calls upon a pre-defined vehicle core component wear model library, which includes engine thermal load models, braking system wear models, and transmission component fatigue models. The micro-driving behaviors and environmental features in the operational intensity map are mapped to model input parameters to simulate the physical stress response of each component under specific working conditions. Based on the physical stress response, the fatigue damage accumulation law is applied, and the material properties of the components are combined to calculate the cumulative damage degree of each core component and predict its remaining service life. Based on the cumulative damage and predicted remaining service life of the core components, a comprehensive reasoning process using a large artificial intelligence model is employed to generate a real-time vehicle health status assessment report that includes health scores for key components, fault risk warnings, and maintenance recommendations.

6. The method according to claim 5, characterized in that, The step of calling a preset vehicle core component wear model library, which includes engine thermal load models, braking system wear models, and transmission component fatigue models, specifically includes: The engine thermal load model formula is as follows: ,in, This refers to the engine's real-time operating temperature. For aging rate, It is a constant. For activation energy, It is the gas constant; The formula for the brake system wear model is as follows: ,in, For wear volume, For normal load, The sliding distance, For material hardness, The wear coefficient; The fatigue model of the transmission component is constructed based on the material SN curve and Miner's linear cumulative damage theory, defining the cumulative damage degree. ,in This represents the actual number of cycles at the stress level. This represents the fatigue life limit corresponding to the stress level.

7. The method according to claim 5, characterized in that, The process of generating a real-time vehicle health status assessment report, which includes key component health scores, fault risk warnings, and maintenance recommendations, based on the cumulative damage and predicted remaining service life of the core components and using a large-scale artificial intelligence model (LLM) for comprehensive reasoning, specifically includes: Construct a vertical-category large language model based on industrial knowledge graphs and maintenance expert experience bases with fine-tuning; The cumulative damage values ​​of the core components, real-time sensor readings, and vehicle file data are dynamically mapped to the corresponding slots of the Prompt template in the vertical language model to form a complete inference context, which is then input into the large language model. A deterministic rule validation layer is set up to perform multi-dimensional post-processing and logical verification of the model output, ensuring the accuracy and executability of the evaluation results; The large language model comprises an open-source base model layer, a domain fine-tuning layer, a hierarchical Prompt engineering template layer, and a rule validation layer. The large language model is trained using a lightweight, directional domain fine-tuning training mode, employing low-rank adaptive lightweight fine-tuning technology. During training, all pre-trained weights of the open-source base model are frozen, and only a trainable sparse matrix with rank r=8 is injected into the attention module of its Transformer layer. This trainable sparse matrix is ​​used as a dedicated module for domain knowledge learning.

8. The method according to claim 1, characterized in that, The process of combining the operational intensity map with standardized energy efficiency benchmarks across vehicle types and routes to perform multi-dimensional weighted calculations to generate a vehicle operation efficiency quantification index, and identifying inefficient operating vehicles and abnormal driving modes, specifically includes: Construct a standardized energy efficiency benchmark library that spans vehicle models and routes, and use big data statistical methods to eliminate the differences in physical parameters of different vehicle models and basic road conditions, thereby establishing a unified energy efficiency evaluation standard. Input the operational intensity map data of the vehicle to be evaluated into the standardized energy efficiency benchmark library to calculate its relative energy consumption level and efficiency deviation under the same working load. A multi-dimensional weighting system is established to calculate the relative energy consumption level, load task completion rate and driving behavior stability by weighting, and generate a quantitative index of vehicle operating efficiency. Based on the quantitative index, horizontal ranking and threshold determination are performed to accurately identify inefficient operating vehicles and correlate abnormal driving patterns.

9. The method according to claim 8, characterized in that, The step of generating an optimal personalized maintenance cycle and maintenance item list based on the vehicle's real-time health status assessment results, operational efficiency quantification index, and operational task scheduling constraints, and outputting a dynamic maintenance execution plan through a dynamic programming algorithm, specifically includes: The real-time health status assessment results of the vehicle are combined with the quantitative index of operational efficiency, and the scheduled operational tasks of the vehicle are obtained as time window constraints. Construct a multi-objective dynamic programming model, setting the joint optimization objectives as the lowest maintenance cost, the lowest safety risk, and the lowest impact on operational tasks; Solve the model while satisfying the scheduling constraints to determine the optimal time for maintenance intervention and the specific damaged components that need to be handled at the time of maintenance intervention. The output includes a dynamic maintenance execution plan that includes specific execution time, task content, and required resources, enabling collaborative optimization of maintenance plans and operational tasks.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the vehicle analysis method based on trajectory data and driving behavior as described in any one of claims 1-9.

Images