Intelligent early warning and remote operation and maintenance method for express cabinet equipment failure

Abstract

The application relates to an express cabinet equipment fault intelligent early warning and remote operation and maintenance method, which comprises the following steps: acquiring the physical topology connection relationship of a target express cabinet as domain knowledge constraint; performing time sequence cause and effect discovery based on the constraint to generate an initial health cause and effect diagram; fine-tuning parameters from a pre-constructed knowledge base according to a deployment scene to obtain a scene-adapted health cause and effect diagram; collecting data in real time and generating a real-time cause and effect diagram based on the same constraint; identifying abnormalities and early warning by comparing the graph structure difference between the real-time diagram and the scene-adapted health cause and effect diagram; after early warning, virtual intervention simulation can be carried out in digital twinning, and remote instructions are issued according to the results, and the model is updated according to the feedback. The application effectively improves the early warning accuracy and positioning accuracy of complex and intermittent faults, realizes predictive maintenance and remote closed-loop operation and maintenance, and significantly reduces operation and maintenance cost and equipment downtime.

Description

Technical Field

[0001] This application relates to the technical field of express delivery locker operation and maintenance, and in particular to an intelligent early warning and remote operation and maintenance method for express delivery locker equipment failure. Background Technology

[0002] Currently, fault warnings for express delivery locker equipment mainly rely on rule-based threshold judgments or traditional time series analysis models. Common methods include: setting fixed thresholds for single sensors (such as current or vibration) to trigger over-limit alarms; or using deep learning models such as LSTM to learn from historical sensor data sequences to predict future data trends and identify abnormal patterns. These methods achieve fault detection to some extent, but generally remain at the level of analyzing the correlation of data appearances.

[0003] However, the aforementioned existing technologies have significant drawbacks: First, they cannot reveal the deep, time-delayed causal mechanisms among multiple variables within the device, resulting in high false alarm and false negative rates for intermittent, complex faults caused by complex causal chains; second, the models lack integration with differences in device physical structure and deployment scenarios, leading to poor generalization ability, requiring a long data accumulation period in new environments or on new devices, and failing to achieve accurate component-level fault tracing; finally, the early warning and operation and maintenance links are disconnected, making it difficult to form an intelligent closed loop from perception, diagnosis to intervention, resulting in low operation and maintenance efficiency.

[0004] To this end, this application proposes an intelligent early warning and remote operation and maintenance method for express cabinet equipment failures. The aim is to build an adaptive causal perception model by integrating physical constraints and data-driven approaches, so as to achieve more accurate, forward-looking and closed-loop verifiable intelligent operation and maintenance. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a method for intelligent early warning and remote operation and maintenance of express delivery locker equipment malfunctions, employing the following technical solution:

[0006] A method for intelligent early warning and remote operation and maintenance of express delivery locker equipment includes the following steps:

[0007] S1. Obtain domain knowledge data of the target express cabinet, wherein the domain knowledge data includes at least the physical topology connection relationship determined based on the device circuit and mechanical structure;

[0008] S2. Construct causal discovery constraints based on the physical topology connection relationship, and perform time-series causal discovery under the constraints based on the initial operation sensor data of the target express cabinet to generate the initial health causal graph of the cabinet; the nodes of the health causal graph are sensor variables, and the edges represent causal dependencies between variables with time delay.

[0009] S3. Based on the deployment scenario information of the target express cabinet, migrate the corresponding scenario adaptation parameters from the pre-built cross-scenario causal knowledge base, quickly fine-tune the initial health causal graph, and generate a scenario-adaptive health causal graph.

[0010] S4. During the real-time operation of the target express cabinet, real-time sensor data is continuously collected in short time windows, and real-time temporal causal discovery is performed based on the physical topology connection constraints to generate a real-time causal graph.

[0011] S5. By comparing the graph structure differences between the real-time causal graph and the scene-adapted health causal graph, abnormal changes in the causal structure are identified.

[0012] S6. When the identified graph structure difference exceeds the preset threshold, generate a fault warning message.

[0013] Preferably, the method for constructing the pre-built cross-scenario causal knowledge base in step S3 includes the following steps:

[0014] Obtain historical health operation data of multiple sample express lockers from various typical deployment scenarios;

[0015] For each sample cabinet, temporal causal discovery is performed based on its respective physical topology connection constraints to generate a sample health causal graph.

[0016] By using meta-learning algorithms, shared scenario feature parameters are extracted from the health causal graphs of multiple samples in the same scenario, forming a scenario adaptation parameter set corresponding to different deployment scenarios, and summarizing them to build a cross-scenario causal knowledge base.

[0017] Preferably, it also includes a fault sample enhancement step:

[0018] On the scene-adapted health causal graph, one or more preset causal structure perturbation atomic operations are simulated to generate virtual fault causal graphs corresponding to different fault modes; the atomic operations include causal edge delay change, edge weight decay, abnormal edge insertion, or node state removal.

[0019] Based on the virtual fault cause-effect graph and its corresponding simulated sensor data sequence, the training dataset is expanded to optimize the generalization ability of the graph structure difference recognition model in step S5; wherein, the simulated sensor data sequence is obtained through a simulation model based on physical laws or a conditional generative adversarial network.

[0020] Preferably, the identification of anomalous changes in the causal structure in step S5 specifically includes:

[0021] The real-time causal graph is matched with a predefined causal feature subgraph corresponding to a specific parcel locker failure mode; the causal feature subgraph is a subgraph structure defined according to the failure mechanism and containing specific node and edge connection patterns.

[0022] If the match is successful, it is determined that a fault warning corresponding to the causal feature subgraph has been triggered.

[0023] Preferably, after step S6, the method further includes:

[0024] S7. After generating fault warning information, virtual intervention simulation is performed in the digital twin model of the target express cabinet according to the suspected fault cause indicated by the warning information.

[0025] S8. Based on the simulation results, generate and issue remote operation and maintenance instructions to the physical express cabinet;

[0026] S9. Receive feedback data after the physical express cabinet executes the remote operation and maintenance command, and update the scenario adaptation health cause-effect graph or the graph structure difference recognition model based on the feedback data.

[0027] Preferably, the domain knowledge data is stored in the digital twin model of the target express delivery locker; the physical topological connection relationship is provided by the digital twin model as a hard constraint or prior probability distribution for the temporal causal discovery process.

[0028] Preferably, the temporal causal discovery described in steps S2 and S4 employs a constraint-based temporal causal discovery algorithm, where the physical topological connectivity is used as an input constraint to limit or weight the possible causal edge search space.

[0029] Preferably, the identification of anomalies by comparing differences in graph structures in step S5 is specifically implemented through a lightweight graph neural network model. The input of this model is the difference graph between the real-time causal graph and the scene-adapted health causal graph, and the output is the fault type and probability.

[0030] Preferably, the short time window in step S4 is 1 to 10 minutes; the fault warning information in step S6 includes at least: warning time, description of the causal structure of the suspected fault, location of the associated physical component, and recommended handling measures.

[0031] In summary, this application includes at least one of the following beneficial technical effects:

[0032] 1. Significantly Improved Early Warning Accuracy and Reliability: By introducing the physical topology connection relationship of the express delivery lockers as a hard constraint, the temporal causal discovery algorithm is guided, effectively avoiding the generation of spurious causal relationships that violate physical laws in complex sensor data, thereby greatly reducing false alarms. Combined with scene-adaptive causal graph fine-tuning and precise matching of causal feature subgraphs, the system's accuracy in identifying intermittent and complex faults is improved to over 95%, while the false alarm rate is reduced to below 5%.

[0033] 2. Fault location achieves a leap from module level to component level: Based on causal feature subgraph matching and reverse tracing of digital twin models, the cause of the fault can be accurately located to a specific physical component (such as the lock cylinder motor of a specific cabinet door, the power interface on the main control board), replacing the traditional method that can only report vague information such as "door lock fault" or "main control fault", reducing the average on-site troubleshooting time from more than 1 hour to less than 10 minutes.

[0034] 3. The operation and maintenance model shifts from passive response to proactive prediction and closed-loop processing: By utilizing real-time causal discovery and health benchmark comparison within a short time window, early warnings can be issued before fault symptoms fully manifest. Combined with virtual intervention simulation in digital twins, diagnostics can be verified and safe and effective remote repair instructions can be generated, realizing a complete intelligent closed loop of "early warning-diagnosis-decision-verification-repair-learning". This allows more than 80% of minor faults to be resolved remotely without manual on-site visits, significantly reducing operation and maintenance costs and downtime.

[0035] 4. Solves the challenges of cold start and sample scarcity in new equipment / scenarios: Through cross-scenario causal knowledge bases and meta-learning transfer technology, newly deployed express lockers can quickly acquire early warning capabilities adapted to their scenarios without a long initial data accumulation period. Simultaneously, the physical mechanism-based fault sample augmentation technology can effectively expand the training set and improve the model's generalization ability to recognize rare and unknown fault patterns when real fault data is scarce.

[0036] 5. It has achieved continuous accumulation of operation and maintenance knowledge and adaptive evolution of the system: Every early warning, remote intervention and its feedback results are used for online fine-tuning of the scenario-adaptive health cause-effect graph or identification model, enabling the system to continuously learn from actual operation and maintenance experience, and continuously optimize its early warning threshold, cause-effect benchmark and diagnostic logic, thereby achieving continuous autonomous evolution of operation and maintenance capabilities and extending the life cycle and value of the entire early warning operation and maintenance system. Attached Figure Description

[0037] Figure 1 This is a flowchart of a method for intelligent early warning and remote operation and maintenance of express locker equipment faults according to Embodiment 1 of this application;

[0038] Figure 2This is a flowchart of the method for constructing a cross-scenario causal knowledge base in Embodiment 2 of this application. Detailed Implementation

[0039] The following is in conjunction with the appendix Figure 1 and Figure 2 This application will be described in further detail.

[0040] Example 1

[0041] This application discloses an intelligent early warning and remote operation and maintenance method for express delivery locker equipment malfunctions. (Refer to...) Figure 1 A method for intelligent early warning and remote operation and maintenance of express delivery locker equipment faults includes the following steps:

[0042] S1. Constructing a digital twin and physical topology constraints: Obtaining domain knowledge data of the target express cabinet, wherein the domain knowledge data includes at least the physical topology connection relationship determined based on the device circuit and mechanical structure; in this embodiment, the domain knowledge data is stored in the digital twin model of the target express cabinet; the physical topology connection relationship is provided by the digital twin model as a hard constraint or prior probability distribution for the subsequent temporal causal discovery process;

[0043] S2. Generate an initial health causal graph: Construct causal discovery constraints based on the physical topology connection relationship, and perform time-series causal discovery under the constraints based on the initial operation sensor data of the target express cabinet to generate an initial health causal graph for the cabinet; the nodes of the health causal graph are sensor variables, and the edges represent causal dependencies between variables with time delay.

[0044] S3. Fine-tuning of causal graph based on scenario knowledge base: Based on the deployment scenario information of the target express cabinet, the corresponding scenario adaptation parameters are migrated from the pre-built cross-scenario causal knowledge base to quickly fine-tune the initial health causal graph and generate a scenario-adapted health causal graph.

[0045] S4. Real-time Causal Graph Discovery: During the real-time operation of the target express cabinet, real-time sensor data is continuously collected in a short time window, and real-time temporal causal discovery is performed based on the physical topology connection constraints to generate a real-time causal graph; the short time window is 1 to 10 minutes; the temporal causal discovery adopts a constraint-based temporal causal discovery algorithm, and the physical topology connection is used as an input constraint to limit or weight the possible causal edge search space;

[0046] S5. Causal Structure Anomaly Identification: By comparing the graph structure differences between the real-time causal graph and the scene-adapted healthy causal graph, abnormal changes in the causal structure are identified. Specifically, this anomaly identification is achieved through a lightweight graph neural network model. The model's input is the difference graph between the real-time causal graph and the scene-adapted healthy causal graph, and its output is the fault type and probability. Identifying abnormal changes in the causal structure specifically includes matching the real-time causal graph with a predefined causal feature subgraph corresponding to a specific parcel locker fault mode. The causal feature subgraph is a subgraph structure defined according to the fault mechanism, containing specific node and edge connection patterns. If the match is successful, it is determined that a fault warning corresponding to the causal feature subgraph has been triggered.

[0047] S6. Generate fault warning information: When the identified graph structure difference exceeds a preset threshold, generate fault warning information; the fault warning information includes at least: warning time, causal structure description of the suspected fault, location of associated physical components and recommended handling measures.

[0048] S7. Digital Twin Virtual Intervention Simulation: After generating fault warning information, virtual intervention simulation is performed in the digital twin model of the target express cabinet according to the suspected fault cause indicated by the warning information.

[0049] S8. Issue remote maintenance instructions: Based on the simulation results, generate and issue remote maintenance instructions to the physical parcel locker;

[0050] S9. Feedback-based model update: Receive feedback data after the physical express cabinet executes the remote operation and maintenance instructions, and update the scenario-adaptive health cause-effect graph or the graph structure difference recognition model based on the feedback data.

[0051] The intelligent early warning and remote operation and maintenance method for parcel locker equipment provided in this application deeply integrates the physical mechanism knowledge of parcel lockers (the physical topological connections in the digital twin model) with multi-dimensional operational data to construct a scenario-adaptive and causal-aware intelligent early warning and operation and maintenance system. Compared with traditional early warning methods based on thresholds or single data models, this solution achieves three major breakthrough improvements:

[0052] 1. Significantly improved early warning accuracy: Through temporal causal discovery under physical constraints, it can capture the inherent causal transmission mechanism of equipment, greatly improving the accuracy of identifying intermittent and complex faults and significantly reducing the false alarm rate (the applicant's actual test showed that the identification accuracy rate of traditional methods has increased from 70% to over 96%, and the false alarm rate has decreased from 25% to below 5%).

[0053] 2. A qualitative leap in operation and maintenance efficiency: It has achieved a transformation in operation and maintenance mode from passive response after a fault to proactive prediction before a fault occurs, from fuzzy positioning at the module level to precise tracing at the component level, and from manual on-site troubleshooting to remote closed-loop repair. The average fault troubleshooting time has been greatly shortened (the applicant's actual test showed that it has been shortened from more than 1 hour to less than 10 minutes).

[0054] 3. Significantly enhanced system adaptability: Through cross-scenario knowledge transfer and continuous online learning, the early warning model can quickly adapt to different deployment environments, solving the pain point of traditional methods requiring a long data accumulation period in new scenarios and on new devices. At the same time, the ability to generalize the identification of rare faults is improved through virtual sample enhancement.

[0055] This embodiment takes "Target Express Locker A" deployed in a residential community as an example to illustrate the whole process of its intelligent fault early warning and remote operation and maintenance.

[0056] Step S1, obtaining the domain knowledge data of the target parcel locker, specifically involves:

[0057] The domain knowledge data is pre-constructed and stored in the digital twin model corresponding to the target parcel locker A. This digital twin model is a virtual image created based on the parcel locker A's precise BOM (Bill of Materials), circuit schematics, mechanical assembly drawings, and embedded software control logic. The key domain knowledge data is the physical topology connections.

[0058] This relationship is concretized as an adjacency list. For example, from the circuit schematic, it is found that: "The +12V power output terminal on the main control board (denoted as node PWR_12V)" is directly connected to the power pin of the driver chip of the lock motor of cabinet door No. 2 on the third floor (denoted as node MOTOR_3_2_PWR) through a wire. From the mechanical transmission chain, it is found that: "The rotational motion of the lock cylinder (abstracted as node LOCK_ROTATION)" directly drives the "open / closed status sensor of the cabinet door (node ​​DOOR_STATE)" through gears. Conversely, "The ambient temperature and humidity sensor (node ​​TEMP_HUMID)" and "The core voltage of the main control board (node ​​CORE_V)" are not physically directly connected. These connections and isolation relationships are encoded into a priori knowledge matrix, where matrix element (i,j)=1 indicates that a causal edge from variable i to variable j is allowed, and =0 indicates that it is prohibited.

[0059] This step, by constructing a digital twin model and encoding physical topological connections into a prior knowledge matrix, provides crucial domain knowledge constraints for the subsequent causal discovery process. This design fundamentally avoids the generation of false causal relationships that violate physical laws in complex sensor data by general causal discovery algorithms, ensuring the physical interpretability of the discovered causal relationships and laying the first solid foundation for accurate early warning.

[0060] Step S2, generating the initial health causal graph, specifically involves:

[0061] Based on the physical topology connections obtained in step S1 (i.e., the prior knowledge matrix), causal discovery constraints are constructed. Multi-dimensional sensor data of the target express cabinet A, which has been running continuously for 7 days under known healthy conditions, including current, voltage, vibration, temperature and humidity, door opening and closing logs, are collected to form the initial operating sensor data.

[0062] A constraint-based temporal causal discovery algorithm (an improved version of the PCMCI+ algorithm in this embodiment) is used for processing. The algorithm takes the prior knowledge matrix as a hard constraint input, which means that when searching for possible causal relationships, the algorithm will completely exclude connections marked as 0 in the matrix (i.e., physically unconnected variable pairs). Under the constraints, the algorithm analyzes the statistical dependencies with time delays between sensor data and automatically generates an initial health causal graph.

[0063] This is a directed weighted graph. Nodes represent various sensor variables, such as grid input voltage (Vin), lock cylinder motor current (I_motor), and cabinet vibration acceleration (Acc). Edges represent causal dependencies between variables. For example, an edge from Vin to I_motor with a weight of 0.85 and a delay of 200ms indicates that grid voltage fluctuations will affect the motor current approximately 200ms later. Another edge from I_motor to Acc with a weight of 0.72 and a delay of 50ms indicates that motor operation directly causes cabinet vibration.

[0064] This step performs temporal causal discovery under physical constraints. The generated initial health causal graph not only reflects the statistical correlations between sensor data but also reveals the physical causal dependencies between variables with time delays. This enables the system to understand the causal logic of normal device operation, rather than simply memorizing data patterns, providing a physically meaningful benchmark for subsequent anomaly detection.

[0065] Step S3, based on the scene knowledge base, fine-tunes and generates a scene-adapted health cause-effect graph, specifically as follows:

[0066] Based on the deployment scenario information of the target parcel locker, corresponding scenario adaptation parameters are migrated from a pre-built cross-scenario causal knowledge base to quickly fine-tune the initial health causal graph, generating a scenario-adapted health causal graph. The deployment scenario information for target locker A is obtained as "old residential area". The set of scenario adaptation parameters matching this scenario is queried from the pre-built cross-scenario causal knowledge base.

[0067] The retrieved parameter set is applied to fine-tune the initial health causal graph obtained in step S2. This parameter set contains the parameter adjustments for typical causal edges in this scenario. For example, the knowledge base indicates that in the "old residential area" scenario, the normal fluctuation range of the weight of the causal edge "Vin→I_motor" should be corrected to [0.75, 0.92]. The system updates the health baseline interval of this edge accordingly. After fine-tuning, the scenario-adapted health causal graph is obtained, which more accurately reflects the health operating baseline of the target cabinet in its specific deployment environment.

[0068] This step enables the health cause-effect graph to quickly adapt to the current deployment environment within hours rather than weeks by migrating scenario adaptation parameters from a pre-built cross-scenario knowledge base. This solves the cold start problem of traditional methods that require long-term data accumulation in new scenarios, while ensuring the scenario relevance of the early warning benchmark and significantly reducing the false alarm rate caused by scenario differences.

[0069] Step S4, generating the real-time cause-effect graph, specifically involves:

[0070] During the real-time operation of the express delivery locker A, the edge computing module continuously collects real-time sensor data streams in 5-minute time windows (the short time window is 1 to 10 minutes; preferably 5 minutes in this embodiment). Also based on the physical topology constraints of S1, a lightweight real-time causal discovery algorithm (such as an online variant of Fast Causal Inference) is used to analyze the data in each time window and quickly generate a real-time causal graph reflecting the current short-term operating status.

[0071] This step uses a short time window of 1-10 minutes for real-time causal discovery, achieving near real-time causal perception of equipment operating status. This high-frequency, lightweight causal analysis can capture transient abnormal causal patterns that traditional methods cannot detect, providing the possibility for early warning and increasing the lead time for fault warnings from minutes to ten minutes.

[0072] Step S5, comparing and identifying causal structural anomalies, specifically involves:

[0073] This step utilizes a lightweight Graph Neural Network (GNN) model for anomaly detection. This GNN model is trained with input a difference graph between the real-time cause-effect graph and the scene-adapted healthy cause-effect graph (obtained by calculating edge additions / reductions, weight changes, and latency drift). Its output is a fault probability vector. This fault probability vector represents the likelihood that the real-time cause-effect graph belongs to various predefined fault modes. The maximum probability value in this vector is defined as the comprehensive anomaly score, used for subsequent threshold determination.

[0074] Simultaneously, as a parallel and more precise diagnostic method, the system matches the real-time cause-effect graph with a predefined causal feature subgraph library. The subgraph library is defined based on the fault mechanism; for example, a motor stall fault feature subgraph is defined as follows: it must contain nodes I_motor and MOTOR_TEMP (motor temperature), have an edge from I_motor to MOTOR_TEMP, have a delay of <30 seconds, and the value of I_motor must continuously exceed the threshold Th1 for 10 seconds. If this pattern is matched in the real-time cause-effect graph using a subgraph isomorphism algorithm (such as VF2), a high-confidence warning for motor stall is directly triggered. This threshold can be determined statistically from the historical fault data of the target express delivery locker; for example, Th1 can be set to 5A and T1 to 0.9.

[0075] This step achieves a balance between high sensitivity and high specificity through a dual anomaly detection mechanism combining a lightweight GNN model and domain feature subgraph matching. The GNN model can detect unknown causal patterns of anomalies, while feature subgraph matching can accurately diagnose known typical faults. This design effectively improves the system's accuracy in identifying complex faults while maintaining low computational requirements on embedded hardware.

[0076] Step S6, generating fault warning information, specifically involves:

[0077] When the maximum value (i.e., the comprehensive anomaly score) in the fault probability vector output by the GNN model exceeds the threshold T1, or when a specific fault feature subgraph is matched, an anomaly is determined, and a fault warning message is generated. This message includes at least: the warning time (e.g., 2023-10-27 14:05:03), a description of the causal structure of the suspected fault (e.g., a significant increase in the causal edge between motor current and motor temperature was detected, consistent with a stalled rotor mode), the location of the associated physical component (cabinet 2 on the 3rd floor, lock cylinder motor), and recommended handling measures (immediately execute the motor self-test program remotely, prepare to replace the MF-003 motor).

[0078] The fault warning information generated in this step not only includes traditional time and component location, but also provides a fault mechanism description based on causal structure and targeted handling suggestions. This enables on-site maintenance personnel to quickly understand the nature of the fault, carry the correct spare parts and tools, and greatly shorten the average on-site handling time.

[0079] The virtual intervention simulation in step S7 of the digital twin is specifically as follows:

[0080] Upon receiving the warning, the system automatically initiates a simulation in the digital twin model of parcel locker A. For the suspected cause of motor stall, the motor load parameters are set to a stall state in the twin, and a high-fidelity physical simulation is run to observe whether the simulation data reproduces the causal pattern of the warning. Simultaneously, a reverse micro-motion command to the generator is simulated to observe whether the abnormal causal edges in the simulation disappear.

[0081] Step S8, issuing remote operation and maintenance instructions, specifically involves:

[0082] Simulation results show that reverse micro-motion of the motor can temporarily eliminate the abnormality. Based on this, the system generates a remote command, which, after security verification, is sent to physical parcel locker A via the Internet of Things channel to execute the motor's self-repair operation.

[0083] Step S9, updating the model based on feedback, specifically involves:

[0084] After the command is executed, the system collects real feedback data from cabinet A. If the data confirms that the abnormal causal pattern has been eliminated, the warning and intervention are marked as a successful case. The system uses this feedback data to fine-tune the scenario-adaptive health causal graph in S3 (e.g., slightly widening the current health range of the motor under specific operating conditions) and to perform online incremental learning on the lightweight GNN model in S5, optimizing its weights. Thus, the system achieves closed-loop learning and continuous optimization.

[0085] Steps S7-S9, through digital twin virtual intervention, remote command issuance, and feedback-driven model updates, form a complete intelligent closed loop encompassing perception, cognition, decision-making, action, and learning. This design enables most minor faults to be resolved through remote intervention, eliminating the need for on-site human intervention. Simultaneously, the system continuously evolves through each maintenance practice, resulting in exponentially improved maintenance efficiency over time.

[0086] Example 2: Method for constructing a cross-scenario causal knowledge base

[0087] Reference Figure 2 This section independently and in detail explains how a knowledge base is constructed from multi-scenario data through meta-learning. The method for constructing the cross-scenario causal knowledge base is as follows:

[0088] A1. Multi-scenario health data collection: Collect health operation data from 50 sample express delivery lockers in each of four typical scenarios: high-end office buildings, university campuses, open-air squares, and old residential communities for one month.

[0089] A2. Sample health cause-effect graph generation: For each sample cabinet, repeat steps S1-S2 (time-series cause-effect discovery based on its respective physical topology connection constraints) to generate its own sample health cause-effect graph.

[0090] A3. Meta-learning Extraction of Scene Adaptation Parameters: A meta-learning algorithm (in this embodiment, the Model-Independent Meta-Learning (MAML) framework) is applied for knowledge extraction. The inner loop learns to quickly adapt from a small number of sample causal graphs for each task in each scene; the outer loop optimizes a shared graph neural network encoder. Finally, a set of feature parameters (i.e., the initial optimization state of the encoder and scene-specific graph structure offsets) is extracted for each scene, forming a scene adaptation parameter set, which is then aggregated and stored in a database.

[0091] The cross-scenario causal knowledge base construction method provided in this embodiment extracts transferable scenario feature parameters from multi-scenario data through meta-learning, realizing the accumulation, sharing, and reuse of cross-scenario causal knowledge. This enables newly deployed express delivery lockers to stand on the shoulders of giants, quickly gain early warning capabilities adapted to the current environment, and shorten the maturity cycle of new equipment early warning models from the traditional 3-6 months to within 24 hours.

[0092] Example 3: Fault Sample Enhancement Method

[0093] To optimize the generalization ability of the graph neural network (GNN) model in step S5, especially when facing rare intermittent faults, this method also includes a fault sample augmentation step:

[0094] On the scene-adapted health causal graph, one or more preset causal structure perturbation atomic operations are simulated to generate virtual fault causal graphs corresponding to different fault modes; the atomic operation library is designed based on the fault mechanism of express cabinets and includes:

[0095] Causal edge delay changes: Simulates a slower response due to mechanical wear (e.g., the "current → vibration" edge delay increases from 50ms to 200ms).

[0096] Side weight decay: Poor analog contact causes signal weakening (e.g., the side weight of "power supply voltage → motherboard voltage" decays from 0.95 to 0.6).

[0097] Abnormal edge insertion: Simulate spurious coupling caused by electromagnetic interference (such as inserting a weak causal edge between ambient humidity and display drive current).

[0098] Node state removal: The analog sensor is completely disabled (e.g., the door magnetic switch node and all its associated edges are temporarily removed).

[0099] By combining these atomic operations, various realistic virtual fault cause-effect graphs can be generated in batches.

[0100] Secondly, a simulated sensor data sequence corresponding to the virtual fault cause-effect graph is generated. This embodiment provides two optional methods:

[0101] Method 1 (Simulation Model Based on Physical Laws): The abnormal causal logic defined in the virtual fault cause-effect graph (such as "abnormal increase in motor current → rapid rise in motor temperature") is converted into parameter adjustment instructions for the simulation model (such as the motor thermal model in MATLAB / Simulink). Running the simulation directly outputs multi-channel synchronous analog sensor data that conforms to the causal law.

[0102] Method 2 (Based on Conditional Generative Adversarial Network (CGAN): Construct a generative adversarial network (GAN) conditioned on the graph embedding vector of a virtual fault causal graph. The generator receives this conditional vector and random noise to generate simulated multidimensional time series; the discriminator determines whether the series represents real fault data. Through adversarial training, the generated data is made to faithfully reproduce the specified causal structure while possessing the statistical characteristics of real data.

[0103] Finally, by using a large number of generated {virtual fault cause-effect graphs, simulated sensor data sequences} pairs, combined with real fault data, an expanded training dataset is constructed for deep training and optimization of the lightweight GNN model in step S5, significantly improving its ability to identify unknown or complex faults.

[0104] The fault sample augmentation method provided in this embodiment significantly improves the model's generalization ability to rare and intermittent faults without relying on a large amount of real fault data. This solves the common problem of scarce fault samples in actual operation and maintenance, greatly improves the model's recognition rate of unknown fault types, and the generated virtual data conforms to physical laws, avoiding the false patterns that may be introduced by traditional data augmentation methods.

[0105] To verify the effectiveness of this method, the applicant conducted a three-month comparative experiment on a test set including 200 express delivery lockers in different scenarios. Compared with traditional threshold alarms and LSTM prediction methods, this method improved the accuracy of intermittent compound fault identification from 70% to 96.5%, and reduced the false alarm rate from 25% to 3.2%. The average fault location time was shortened from more than one hour at the module level to less than 10 minutes at the component level. 80% of minor faults can be resolved through remote intervention without the need for on-site personnel.

[0106] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions also fall within the scope of protection of the present invention.

Claims

1. A method for intelligent early warning and remote operation and maintenance of express delivery locker equipment, characterized in that, Includes the following steps: S1. Obtain domain knowledge data of the target express cabinet, wherein the domain knowledge data includes at least the physical topology connection relationship determined based on the device circuit and mechanical structure; the domain knowledge data is stored in the digital twin model of the target express cabinet; the physical topology connection relationship is provided by the digital twin model as a hard constraint or prior probability distribution of the time-series causal discovery process; S2. Construct causal discovery constraints based on the physical topology connection relationship, and perform time-series causal discovery under the constraints based on the initial operation sensor data of the target express cabinet to generate the initial health causal graph of the cabinet; the nodes of the health causal graph are sensor variables, and the edges represent causal dependencies between variables with time delay. S3. Based on the deployment scenario information of the target express cabinet, migrate the corresponding scenario adaptation parameters from the pre-built cross-scenario causal knowledge base, quickly fine-tune the initial health causal graph, and generate a scenario-adaptive health causal graph. S4. During the real-time operation of the target express cabinet, real-time sensor data is continuously collected in short time windows, and real-time temporal causal discovery is performed based on the physical topology connection constraints to generate a real-time causal graph. S5. By comparing the graph structure differences between the real-time cause-effect graph and the scene-adapted health cause-effect graph, abnormal changes in the cause-effect structure are identified; wherein, the identification of abnormal changes in the cause-effect structure specifically includes: matching the real-time cause-effect graph with a predefined cause-effect feature subgraph corresponding to a preset express cabinet fault mode; the cause-effect feature subgraph is a subgraph structure defined according to the fault mechanism and containing a preset node and edge connection pattern; if the match is successful, it is determined that a fault warning corresponding to the cause-effect feature subgraph has been triggered; S6. When the identified graph structure difference exceeds the preset threshold, generate a fault warning message; S7. After generating fault warning information, virtual intervention simulation is performed in the digital twin model of the target express cabinet according to the suspected fault cause indicated by the warning information. S8. Based on the simulation results, generate and issue remote operation and maintenance instructions to the physical express cabinet; S9. Receive feedback data after the physical express cabinet executes the remote operation and maintenance command, and update the scenario adaptation health cause-effect graph or the graph structure difference recognition model based on the feedback data; The method further includes a fault sample enhancement step: On the scene-adapted health causal graph, one or more preset causal structure perturbation atomic operations are simulated to generate virtual fault causal graphs corresponding to different fault modes; the atomic operations include causal edge delay change, edge weight decay, abnormal edge insertion, or node state removal. Based on the virtual fault cause-effect graph and its corresponding simulated sensor data sequence, the training dataset is expanded to optimize the generalization ability of the graph structure difference recognition model in step S5; wherein, the simulated sensor data sequence is obtained through a simulation model based on physical laws or a conditional generative adversarial network; The temporal causal discovery described in steps S2 and S4 employs a constraint-based temporal causal discovery algorithm, where the physical topological connectivity is used as an input constraint to limit or weight the possible causal edge search space.

2. The method for intelligent early warning and remote operation and maintenance of express delivery locker equipment according to claim 1, characterized in that, The method for constructing the pre-built cross-scenario causal knowledge base described in step S3 includes the following steps: Obtain historical health operation data of multiple sample express lockers from various typical deployment scenarios; For each sample cabinet, temporal causal discovery is performed based on its respective physical topology connection constraints to generate a sample health causal graph. By using meta-learning algorithms, shared scenario feature parameters are extracted from the health causal graphs of multiple samples in the same scenario, forming a scenario adaptation parameter set corresponding to different deployment scenarios, and summarizing them to build a cross-scenario causal knowledge base.

3. The intelligent early warning and remote operation and maintenance method for express delivery locker equipment according to claim 1, characterized in that, The anomaly identification method described in step S5, which compares the differences in graph structures, is specifically implemented using a lightweight graph neural network model. The input of this model is the difference graph between the real-time causal graph and the scene-adapted health causal graph, and the output is the fault type and probability.

4. The intelligent early warning and remote operation and maintenance method for express delivery locker equipment according to claim 1, characterized in that, The short time window mentioned in step S4 is 1 to 10 minutes; the fault warning information mentioned in step S6 includes at least: warning time, description of the causal structure of the suspected fault, location of the associated physical components, and recommended handling measures.

Images