Battery swap station equipment fault diagnosis method and system based on edge calculation

Abstract

The invention relates to an edge calculation-based equipment fault diagnosis method and system for a battery swap station. The method comprises the following steps of: acquiring data and constructing a regional equipment model; deploying the regional equipment model at an edge computing node, and collecting operation time sequence data in real time; obtaining expected state data through the lightweight simulation model, and comparing the expected state data to generate a residual sequence; obtaining a causal violation event set through a lightweight causal graph model; generating a fault diagnosis result through the fault mode map; in conclusion, the regional equipment model comprising the lightweight causal graph model, the simulation model and the fault mode graph is constructed, real-time data acquisition and closed-loop correction are realized in combination with the edge computing nodes, and the dynamic threshold optimization and fault mode graph fusion mechanism is utilized. The problems that a traditional method is large in response delay, high in false alarm rate and insufficient in diagnosis precision are effectively solved, and the effects of improving fault diagnosis real-time performance, reducing cloud data transmission delay, improving diagnosis precision and reducing the false alarm rate are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery swap station equipment fault diagnosis, in particular to a battery swap station equipment fault diagnosis method and system based on edge computing. BACKGROUND

[0002] With the rapid development of the electric vehicle industry, as an important energy supply infrastructure, the reliability and efficiency of the battery swap station are increasingly concerned. The equipment system in the battery swap station (such as battery swap robots, battery compartments, cooling systems, etc.) has a complex structure and is operated under high load for a long time, which is prone to mechanical wear, electrical aging and other faults. Therefore, accurate fault diagnosis of the equipment is crucial to ensure the continuity and safety of the battery swap service.

[0003] At present, the fault diagnosis method in this field mainly relies on traditional threshold alarm and cloud data analysis. The traditional threshold alarm usually sets a fixed threshold in the equipment controller, and triggers an alarm when the sensor reading exceeds the threshold. Although this method is simple and direct, it is difficult to identify complex or gradual fault patterns, has a high false alarm rate, and cannot provide root cause information of the fault. On the other hand, with the popularization of Internet of Things technology, solutions that upload all device sensor data to the cloud for analysis have also been gradually applied. However, whether it is a simple threshold or cloud analysis, the existing methods generally have the problem of insufficient intelligence, i.e., they can only determine whether there is an anomaly, but cannot answer deeper questions such as "why is it abnormal" and "where is it abnormal". They lack understanding and reasoning ability of the fault mechanism, resulting in insufficient operation and decision support, and are difficult to adapt to the development trend of unmanned operation of the battery swap station. SUMMARY

[0004] In order to solve the above-mentioned defects, the present application provides a battery swap station equipment fault diagnosis method and system based on edge computing.

[0005] The above-mentioned invention purpose of the present application is achieved by the following technical scheme: A battery swap station equipment fault diagnosis method based on edge computing, comprising the steps of: obtaining system equipment information and historical operation data of a target battery swap station, and constructing a regional equipment model, the regional equipment model including a lightweight causal graph model, a lightweight simulation model and a fault mode atlas; deploying the regional equipment model on an edge computing node associated with the target battery swap station, and collecting real-time operation time series data of the target battery swap station through the edge computing node; obtaining a device control instruction and inputting it to the lightweight simulation model to obtain expected state data, and comparing the expected state data with the operation time series data to generate a residual sequence; inputting the residual sequence and the runtime sequence data into the lightweight causal graph model, and outputting a causal violation event set when the lightweight causal graph model identifies a violation event; inputting the residual sequence and the causal violation event set into the fault mode atlas, and enabling the fault mode atlas to generate and output a fault diagnosis result.

[0006] In a preferred example, the application can be further configured to: acquire system device information and historical operation data of the target battery swap station, and construct a regional device model, the regional device model including a lightweight causal graph model, a lightweight simulation model, and a fault mode atlas, and the steps include: constructing an initial causal graph model and an initial simulation model based on the system device information and the historical operation data; constructing a generative adversarial network with a generator and a discriminator, wherein the generator takes device control instructions as input, and the discriminator takes runtime sequence data and simulation data output by the generator as input; embedding the initial simulation model into the generator of the generative adversarial network to form an enhanced generative adversarial network with physical information enhancement; training the enhanced generative adversarial network through historical operation data, and freezing and extracting the lightweight simulation model in the generator as a lightweight simulation model after the training is completed; collecting discrimination probability distribution information output by the discriminator during the training process, and optimizing the initial causal graph model based on the discrimination probability distribution information to obtain a lightweight causal graph model; acquiring expert knowledge information, and constructing a fault mode atlas based on historical fault data and expert knowledge.

[0007] In a preferred example, the application can be further configured to: collect discrimination probability distribution information output by the discriminator during the training process, and optimize the initial causal graph model based on the discrimination probability distribution information to obtain a lightweight causal graph model, and the steps include: during the model training process, collecting discrimination probability distribution information output by the discriminator for historical runtime sequence data, and calculating an information entropy value thereof; identifying device operating conditions based on historical runtime sequence data, and associating and mapping the calculated information entropy value with the device operating conditions; acquiring a discretization level of the device operating conditions and the information entropy value forming the association mapping, and combining the acquired discretization levels to form a joint state space; filtering out historical data segments corresponding to each state of the state units in the joint state space based on the historical operation data; based on the filtered historical data segments, taking maximizing a fault detection rate and minimizing a false alarm rate as optimization objectives, and using an optimization algorithm to calculate optimal violation detection threshold values for each state unit; A lightweight dynamic threshold mapping table is constructed based on each state unit and its corresponding optimal violation detection threshold, and the dynamic threshold mapping table is integrated into the initial causal graph model to obtain a lightweight causal graph model.

[0008] In a preferred example, the application can be further configured to: the step of obtaining expert knowledge information and constructing a fault mode graph based on historical fault data and expert knowledge comprises the steps of: The expert knowledge information is parsed by a pre-trained large language model to extract fault entity features and entity relationship features, and key signal features are extracted based on historical fault data, the key signal features including time domain features, frequency domain features and time-frequency domain features; The extracted fault entity features, entity relationship features and key signal features are fused and aligned to generate a fault representation vector; An initial skeleton of the fault mode graph is constructed based on the fault entity features and entity relationship features, and the initial skeleton is embedded and learned based on the fault representation vector to obtain an optimized skeleton; Strong association rules between the key signal features and the fault entity features are mined from the fault representation vector by a pre-set association rule mining algorithm, and strong association rules with a confidence higher than a pre-set threshold are added as new edges to the optimized skeleton to obtain the fault mode graph.

[0009] In a preferred example, the application can be further configured to: the step of obtaining the device control instruction and inputting it into the lightweight simulation model, obtaining the expected state data, comparing the expected state data with the runtime sequence data, and generating a residual sequence comprises the steps of: The runtime sequence data with a pre-set time difference value is compared with the expected state data, and the difference is obtained, a compensation amount is generated by a PID controller based on the difference to correct the device control instruction to achieve closed-loop correction; The device control instruction sequence generated after closed-loop correction is input into the lightweight simulation model to obtain an expected state data sequence; The expected state data sequence and the runtime sequence data sequence are compared point by point in the time domain to generate a residual sequence.

[0010] In a preferred example, the application can be further configured to: the step of inputting the residual sequence and the runtime sequence data into the lightweight causal graph model, and outputting a causal violation event set when the lightweight causal graph model identifies a violation event comprises the steps of: The received runtime sequence data and residual sequence are distributed to the causal paths in the lightweight causal graph model according to a pre-defined configuration relationship; The causal path receiving the runtime sequence data and the residual sequence starts its calculation process, and when there is an exception in the calculation process of any causal path, corresponding violation event information is generated; The violation event information generated by all causal paths is collected and combined and output as a causal violation event set.

[0011] In a preferred example, the application can be further configured such that the step of the causal path receiving the runtime sequence data and the residual sequence starting its calculation process, and when there is an exception in the calculation process of any causal path, corresponding violation event information is generated, includes the steps of: The causal path extracts cause data and result data from the received runtime sequence data and residual sequence according to its predefined data requirements; Based on the cause data and the predefined constraint function, the predicted value of the result data is calculated, and the calculated predicted value is compared with the result data to obtain a difference value; The difference value is compared with the dynamic threshold value of the causal path, if the difference value is greater than the dynamic threshold value and the duration is greater than the preset time threshold, it is determined that the causal path has an exception, and violation event information is generated according to the determination result, the violation event information includes causal path identification, difference value and abnormal time stamp.

[0012] The second application object is achieved by the following technical scheme: An edge computing-based battery swap station equipment fault diagnosis system, comprising: A model construction module is configured to obtain system equipment information and historical operation data of a target battery swap station, and construct a regional device model, the regional device model including a lightweight causal graph model, a lightweight simulation model, and a fault mode atlas; A model deployment module is configured to deploy the regional device model on an edge computing node associated with the target battery swap station, and collect runtime sequence data of the target battery swap station in real time through the edge computing node; A data comparison module is configured to obtain a device control instruction and input it to the lightweight simulation model to obtain expected state data, and compare the expected state data with the runtime sequence data to generate a residual sequence; A first data input module is configured to input the residual sequence and the runtime sequence data to the lightweight causal graph model, and output a causal violation event set when the lightweight causal graph model identifies a violation event; A second data input module is configured to input the residual sequence and the causal violation event set to the fault mode atlas, so that the fault mode atlas generates and outputs a fault diagnosis result.

[0013] The application also relates to a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned battery swap station equipment fault diagnosis method based on edge computing when executing the computer program.

[0014] The application also relates to a computer readable storage medium storing a computer program, wherein the computer program implements the steps of the above-mentioned battery swap station equipment fault diagnosis method based on edge computing when executed by a processor.

[0015] To sum up, the application provides a battery swap station equipment fault diagnosis method and system based on edge computing, which constructs a regional equipment model comprising a lightweight causal graph model, a simulation model and a fault mode atlas, realizes real-time data acquisition and closed-loop correction in combination with an edge computing node, and effectively solves the problems of large response delay, high false alarm rate and insufficient diagnostic accuracy of traditional methods by utilizing dynamic threshold optimization and a fault mode atlas fusion mechanism model and expert knowledge, thereby having the effects of improving fault diagnosis real-time performance, reducing cloud data transmission delay, improving diagnostic accuracy and reducing false alarm rate. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of an embodiment of the battery swap station equipment fault diagnosis method based on edge computing of the application; Figure 2 is an implementation flowchart of step S10 in an embodiment of the battery swap station equipment fault diagnosis method based on edge computing of the application; Figure 3 is an implementation flowchart of step S15 in an embodiment of the battery swap station equipment fault diagnosis method based on edge computing of the application. DETAILED DESCRIPTION

[0017] The following will be described in detail below with reference to the accompanying drawings. Figures 1-3 The application will be further described in detail.

[0018] In an embodiment, as shown in the accompanying drawings, Figure 1 The application discloses a battery swap station equipment fault diagnosis method based on edge computing, which specifically comprises the following steps: S10: Obtain system equipment information and historical operation data of a target battery swap station, and construct a regional equipment model, wherein the regional equipment model comprises a lightweight causal graph model, a lightweight simulation model and a fault mode atlas; In this embodiment, the system device information is configuration data describing the static properties and topological relationships of physical devices in the target battery swap station, and includes device model, specification parameters, installation location, and mechanical, electrical, and control connection relationships between devices, etc., providing a bottom framework and constraint conditions for model construction; the historical operation data is a time series data set recorded by the battery swap station in the past period of time, sorted by timestamp, and usually includes device control instruction sequence, sensor readings (such as current, voltage, temperature, position), and associated device state and alarm logs; the regional device model is a multi-dimensional analysis framework integrating device physical properties and operation rules, and can specifically adopt a modular modeling method to decompose the device into independent functional units and establish a causal relationship network between the units; the lightweight causal graph model is an optimized causal relationship reasoning structure, which can be constructed by using a Bayesian network or a structural equation model, and is used for detecting logical violations between data; the lightweight simulation model is a simplified mathematical model that retains the core dynamic characteristics, such as a differential equation system after order reduction, and is used for predicting the expected state of the device under a given control instruction; the fault mode atlas is a knowledge base containing fault characteristics and associated rules, and can specifically use a graph database to store fault entities and their multi-dimensional associated relationships.

[0019] S20: Deploy the regional device model on the edge computing node associated with the target battery swap station, and collect real-time time series data of the target battery swap station through the edge computing node; In this embodiment, the edge computing node is a near-end computing device deployed on the site of the battery swap station, used to carry out the operation of the regional device model, and to realize real-time data collection and processing, ensuring low delay and high reliability of the diagnosis process; the real-time time series data is the device operation parameter record collected by the sensor in real time and arranged in chronological order, which is used as the original signal input for fault diagnosis.

[0020] S30: Obtain the device control instruction and input it to the lightweight simulation model to obtain the expected state data, and compare the expected state data with the real-time time series data to generate a residual sequence; In this embodiment, the device control instruction is a command signal issued by the input end to drive the battery swap device to perform an action (such as moving, grabbing, and charging), and is the input basis for the simulation model to calculate the expected state; the expected state data is the operation parameter data of the device in the ideal fault-free state calculated by the lightweight simulation model according to the received control instruction, and is used for comparison with the actual measured value; the residual sequence is the difference sequence between the actual real-time time series data of the device and the expected state data output by the simulation model, which is used to quantify the deviation between the actual behavior of the system and the ideal model, and is a key indicator for identifying abnormalities.

[0021] S40: inputting the residual sequence and the runtime sequence data into the lightweight causal graph model, and outputting a causal violation event set when the lightweight causal graph model identifies a violation event; In this embodiment, the violation event is an event and a record thereof indicating that a predefined causal constraint or physical rule in the device system is violated, and the violation event at least includes a specific causal path identifier that is violated, a timestamp of the violation occurrence, and an index for quantifying a violation severity (such as a residual size); and the causal violation event set is a structured result output by the lightweight causal graph model, which records in detail which preset causal relationship is violated, a degree of violation, and a time of occurrence, and provides a clue for final fault location.

[0022] S50: inputting the residual sequence and the causal violation event set into the fault mode atlas, so that the fault mode atlas generates and outputs a fault diagnosis result.

[0023] In this embodiment, the fault diagnosis result is a conclusive information finally output by the fault mode atlas, and the fault diagnosis result includes a fault type, an occurrence location, a confidence degree, and a processing suggestion.

[0024] Specifically, first, a regional device model is established by using device information and historical data, so as to ensure that the regional device model can reflect actual operation characteristics of the device, and the regional device model includes a lightweight causal graph model, a lightweight simulation model, and a fault mode atlas; after the regional device model is deployed on an edge node, real-time data streams are continuously collected, device control instructions are synchronously received, and expected state data is input into the lightweight causal graph model to generate; a residual sequence of the expected state data and a measured value reflects a deviation between an actual state and an ideal state; the lightweight causal graph model is used to analyze a correlation between the residual sequence and runtime data, and an abnormal event set that violates a preset causal relationship is identified; finally, a fault feature mode stored in the fault mode atlas is combined, an association rule between the abnormal event and a known fault type is matched, and a diagnosis result with a clear direction is output.

[0025] Compared with the prior art, the traditional method relies on a single data source for threshold judgment, while the present scheme realizes multi-dimensional fault feature extraction by fusing simulation prediction, causal reasoning, and knowledge graph technology; the existing cloud analysis is difficult to process high-frequency data in real time, and the present scheme uses edge computing nodes for local processing, thereby effectively reducing response delay; meanwhile, the traditional fault atlas is mostly a static knowledge base, while the present scheme enhances the real-time reasoning ability of the atlas by dynamic residual analysis and causal event detection.

[0026] By the technical solution, the application can effectively distinguish between normal fluctuations of equipment and real fault signals, reduce false positive rates, detect and locate abnormal sources through causal violation events, improve the explainability of diagnosis results, realize localized real-time processing through edge computing nodes to avoid delay problems caused by cloud transmission, and enhance the recognition ability of composite faults through multi-dimensional feature matching of fault mode maps.

[0027] In an embodiment, as shown in FIG. 1, step S10 includes: Figure 2 S11: Based on system equipment information and historical operation data, an initial causal graph model and an initial simulation model are constructed; In this embodiment, the initial causal graph model is a graph structure model established based on the physical causal relationship of the equipment, which can be implemented by using a Bayesian network or a structural equation model, and is used to represent the causal dependence relationship between the equipment state variables; the initial simulation model is a numerical simulation model established based on the physical equation of the equipment, which can be implemented by using a differential algebraic equation or a state space equation, and is used to simulate the expected behavior of the equipment under the control instruction.

[0028] S12: An adversarial generative network with a generator and a discriminator is constructed, wherein the generator takes the equipment control instruction as input, and the discriminator takes the operation time series data and the simulation data output by the generator as input; In this embodiment, the adversarial generative network is a deep learning architecture containing a generator and a discriminator, which can be implemented by combining a convolutional neural network and a fully connected layer, and the output authenticity of the simulation model is improved through adversarial training; the generator is a module responsible for data generation in the adversarial generative network, which takes the equipment control instruction as input, and aims to generate simulation data that is difficult to distinguish from real sensor data in statistical features and dynamic characteristics; the discriminator is a module responsible for data identification in the adversarial generative network, which takes the real operation time series data and the simulation data output by the generator as input, and aims to accurately judge the authenticity of the input data, and the output probability distribution of the discriminator reflects the degree of realism of the generated data.

[0029] S13: The initial simulation model is embedded into the generator of the adversarial generative network to form a physically information-enhanced enhanced generative adversarial network; In this embodiment, the physically information-enhanced enhanced generative adversarial network takes the physical simulation model as the core component of the generator, which can be implemented by model embedding and parameter freezing technology, so that the simulation data output by the generator conforms to the physical law.

[0030] S14: The enhanced generative adversarial network is trained through historical operation data, and after the training is completed, the lightweight simulation model in the generator is frozen and extracted as a lightweight simulation model; ​In this embodiment, the lightweight simulation model is a high-performance simulation model extracted from the generator after optimization by the adversarial generation network and with fixed parameters. The lightweight simulation model has both the accuracy of physical mechanism and the authenticity of data-driven, and can efficiently and accurately predict the expected state of the device on the edge computing node.

[0031] S15: Collect the discrimination probability distribution information output by the discriminator in the training process, and optimize the initial causal graph model through the discrimination probability distribution information to obtain a lightweight causal graph model. In this embodiment, the discrimination probability distribution information is the probability judgment result of the discriminator on the real data and the generated data, which can be implemented by a probability vector output by a softmax function, and is used to reflect the deviation degree of the simulation model from the actual running data. The lightweight causal graph model is a causal graph model optimized by the discrimination probability distribution information, and the threshold of the constraint function, the weight of the causal edge and other parameters of the lightweight causal graph model are adaptively adjusted according to the confidence of the simulation model in different states, so that the lightweight causal graph model has more accurate violation event detection capability.

[0032] S16: Obtain expert knowledge information, and construct a fault mode graph based on the historical fault data and the expert knowledge.

[0033] In this embodiment, the expert knowledge information is unstructured fault diagnosis experience knowledge provided by a domain expert, which usually exists in the form of text in a maintenance manual, a fault report and expert experience, and includes fault phenomena, possible causes, treatment methods and the like, and is an important information source for constructing a fault knowledge base. The historical fault data is a running time series data segment recorded when a device fails and with a clear fault label, and records various parameter changes of the device before and after the fault occurs, and is a core training sample for establishing the association between fault features and modes.

[0034] Specifically, in the model construction phase, first, an initial causal graph model and a simulation model based on physical equations are established according to the physical characteristics of the device; then, an adversarial generation network is constructed, and the initial simulation model is used as the core component of the generator, so that the simulation data output by the generator has physical law constraints; in the training process, the discriminator continuously compares the real running time series data with the simulation data output by the generator, and forces the generator to improve the simulation accuracy through the adversarial training mechanism; after the training is completed, the parameters of the optimized simulation model in the generator are frozen, and are extracted as a lightweight simulation model for subsequent diagnosis; at the same time, the probability distribution information output by the discriminator is collected, which reflects the matching degree of the simulation model and the actual data under different working conditions, and is used to dynamically adjust the abnormal detection threshold of the causal graph model; finally, a fault mode graph is constructed in combination with the expert knowledge, forming a knowledge base containing fault features and association rules.

[0035] Through the above technical solutions, this application realizes the physical law constraints of the simulation model in the data-driven training process, avoiding the non-physical output that may be generated by the pure data-driven method; by dynamically optimizing the detection threshold of the causal graph model by discriminant probability distribution, the fault detection mechanism can adapt to different operating conditions and significantly reduce the false alarm rate; the adversarial training strategy with enhanced physical information effectively improves the generalization ability of the simulation model under complex operating conditions, and provides high-precision benchmark data for subsequent residual analysis and fault diagnosis.

[0036] In one embodiment, such as Figure 3 As shown, step S15 includes: S151: During model training, collect the discrimination probability distribution information output by the discriminator for historical runtime sequence data, and calculate its information entropy value; In this embodiment, the discriminant probability distribution information is the output of the discriminator in the generative adversarial network judging the authenticity of the input data. It is usually a probability value or probability distribution, which characterizes the goodness of fit or degree of difference between the generated data and the real data. The information entropy value is a quantitative index calculated based on the discriminant probability distribution. It is used to measure the uncertainty or disorder of the discriminator's output. The higher the entropy value, the more uncertain the discriminator is in judging the authenticity of the data. It usually corresponds to the working conditions of complex system operation or poor data quality.

[0037] S152: Identify the operating conditions of the equipment based on historical runtime sequence data, and associate and map the calculated information entropy value with the operating conditions of the equipment; In this embodiment, the equipment operating condition refers to the operating state mode of the equipment within a specific time window. Specifically, it can be achieved by dividing historical operating data into states using a clustering algorithm, such as "idle standby", "rated operation", "peak load", etc., to characterize different working modes of the equipment and establish a correlation with the information entropy value. The correlation mapping is the operation of establishing a correspondence between the calculated information entropy value and the identified equipment operating condition, forming a "operating condition-entropy value" data pair, which is used for subsequent analysis of the uncertainty characteristics under different operating conditions.

[0038] S153: Obtain the discretization level of the equipment operating conditions and information entropy values ​​that form the association mapping, and combine the obtained discretization levels to form a joint state space; In this embodiment, the equipment operating condition refers to the operating state mode of the equipment within a specific time window. Specifically, it can be achieved by dividing historical operating data into states using a clustering algorithm to establish a correlation with the information entropy value. The joint state space is a multi-dimensional set of states formed by combining the discretized operating condition level and the information entropy level. Specifically, it can be constructed using the Cartesian product method to divide the fine-grained space of the equipment operating state.

[0039] S154: filtering out historical data segments corresponding to each state unit in the joint state space based on historical operation data; In this embodiment, the state unit is a basic unit in the joint state space, which is uniquely determined by a specific operating condition level and an entropy value level, and is used to identify the running state characteristics of the system at a certain moment.

[0040] S155: based on the filtered historical data segments, an optimal violation detection threshold is calculated for each state unit by using an optimization algorithm with the optimization goal of maximizing the fault detection rate and minimizing the false alarm rate; In this embodiment, the optimal violation detection threshold is a residual anomaly judgment boundary set for a specific state unit, which can be realized by multi-objective optimization solution using genetic algorithm or particle swarm algorithm, and is used to balance between fault detection rate and false alarm rate.

[0041] S156: based on each state unit and its corresponding optimal violation detection threshold, a lightweight dynamic threshold mapping table is constructed, and the dynamic threshold mapping table is integrated into the initial causal graph model to obtain a lightweight causal graph model.

[0042] In this embodiment, the dynamic threshold mapping table is a lightweight data structure for storing the optimal threshold corresponding to each state unit, which can be realized by using hash table or key-value pair database, and is used to match the current state in real time and obtain the corresponding detection threshold.

[0043] Specifically, in the training process of the generative adversarial network, the probability distribution information output by the discriminator is collected and its information entropy value is calculated; by discretizing the device operating condition and the information entropy value, a joint state space covering multiple operating states is constructed; for each state unit, data segments conforming to the state are filtered out from the historical data, and the optimal violation detection threshold under the state is solved by using an optimization algorithm; finally, the formed dynamic threshold mapping table is integrated into the causal graph model, so that the causal reasoning process can automatically match the corresponding detection threshold according to the real-time operating state; for example, when the device is in a high load operating condition and the information entropy value is low, the system automatically calls the strict threshold corresponding to the state unit for anomaly detection.

[0044] Through the above technical solutions, the application solves the problem of poor adaptability of the traditional fixed threshold detection method under complex operating conditions, and realizes adaptive anomaly detection based on the operating state of the device; through the construction of the joint state space and the application of the optimization algorithm, the accuracy and reliability of fault detection are significantly improved while ensuring the lightweight of the model, which provides effective support for real-time fault diagnosis in the edge computing environment.

[0045] In an embodiment, step S16 comprises: S161: Analyze the expert knowledge information by the pre-trained large language model, extract the fault entity features and entity relationship features, and extract the key signal features based on the historical fault data, wherein the key signal features include time domain features, frequency domain features, and time-frequency domain features; In this embodiment, the pre-trained large language model is a natural language processing model trained by large-scale text data, which can be implemented by BERT or GPT model based on Transformer architecture, and is used to automatically extract structured features from unstructured expert knowledge text; the fault entity features are semantic labels describing the types of device faults, such as battery overvoltage, mechanical jam, etc. entity name, which can be extracted from expert knowledge by named entity recognition technology; the entity relationship features are high-dimensional vectorized representations extracted from expert knowledge by language model to represent the relationship between fault entities, including “cause”, “show as”, “occur in” and the like, and the vector representation defines the logical association between entities; the key signal features are quantized feature vectors extracted from device historical fault data by signal processing technology and capable of representing fault states, which include time domain features, frequency domain features, and time-frequency domain features, wherein the time domain features (such as mean, variance, kurtosis) represent signal amplitude statistical characteristics, the frequency domain features (such as spectral peak, frequency center of gravity) represent signal energy distribution, and the time-frequency domain features (such as wavelet coefficient) represent signal frequency band energy variation law with time.

[0046] S162: Fuse and align the extracted fault entity features, entity relationship features, and key signal features to generate a fault representation vector; In this embodiment, the fusion and alignment is to map the semantic features obtained by text analysis and the numerical features obtained by signal analysis to a unified vector space, which can be implemented by attention mechanism or feature splicing; the fault representation vector is a comprehensive vector obtained by fusion and alignment and uniformly representing a certain fault mode or state, and the fault representation vector contains both text semantic information and numerical signal feature information, which is the basis for subsequent graph construction and reasoning.

[0047] S163: Construct an initial skeleton of the fault mode graph based on the fault entity features and the entity relationship features, and perform embedding learning on the initial skeleton based on the fault representation vector to obtain an optimized skeleton; In this embodiment, the initial skeleton of the fault mode graph is a preliminary graph for reflecting the logical structure of fault knowledge, which is constructed only based on fault entity features and entity relationship features. The initial skeleton contains basic elements such as entity nodes and relationship edges, but lacks data-driven details and verification. The embedding learning is a process of optimizing the vectorization representation of entities and relationships in the initial skeleton by using fault representation vectors through graph representation learning algorithms such as TransE and Node2Vec, so that the graph can better represent the fault modes existing in the historical data. The optimized skeleton is the fault mode graph after embedding learning optimization, and the vectorization representation of its nodes and edges is more accurate, which can better support subsequent similarity calculation and reasoning.

[0048] S164: Through a preset association rule mining algorithm, strong association rules between key signal features and fault entity features are mined from the fault representation vectors, and strong association rules with a confidence higher than a preset threshold are added to the optimized skeleton as new edges to obtain the fault mode graph.

[0049] In this embodiment, the association rule mining algorithm is an algorithm for discovering the association relationship between frequent item sets in the data set, which can be implemented by using Apriori or FP-Growth algorithm, and is used to establish the causal relationship between signal features and fault entities. The strong association rule is a rule discovered by the association rule mining algorithm and whose confidence and support exceed the preset threshold. The strong association rule reflects the stable and reproducible corresponding relationship between signal features and fault entities.

[0050] Specifically, the expert knowledge information is converted into structured features through a pre-trained language model, for example, the text description of "abnormal battery temperature causing contactor welding" is parsed into an "abnormal battery temperature" entity and its "causing" relationship. Meanwhile, signal features are extracted from historical fault data, for example, the standard deviation of the temperature sensor increases and the current frequency spectrum energy concentrates when a contactor fault occurs. After the fusion of the two types of features, a vector containing semantic and numerical information is formed, for example, the "contactor welding" entity and the temperature standard deviation feature are spliced into a multi-dimensional vector. When the initial skeleton is constructed, the entities are used as nodes and the entity relationships are used as edges to form the basic structure of the graph, for example, "battery overvoltage-causing-relay sticking" forms an edge. In the embedding learning process, the fault representation vector is used to adjust the connection weight between nodes, for example, the edge weight is updated through a graph neural network to reflect the influence of signal features on fault association strength. In the association rule mining stage, for example, if it is found that the co-occurrence probability of temperature standard deviation greater than the threshold and contactor welding exceeds 95%, a new edge of "temperature fluctuation anomaly-associated-contactor welding" is added, so that the graph can automatically capture the data-driven fault association mode.

[0051] Through the technical solution, the application realizes deep fusion of expert experience and data characteristics, solves the problem of semantic knowledge and numerical characteristic split in traditional fault diagnosis, avoids subjective deviation of manual labeling by analyzing unstructured text through a language model to generate standardized features, enhances completeness of fault representation by signal time-frequency domain feature extraction, and enables the atlas to have self-evolution ability by dynamic association rule mining to adapt to new fault modes generated by changes in device operating environment. In summary, the scheme significantly improves the recognition accuracy of the fault mode atlas for complex faults, for example, in complex working conditions where electrical abnormalities and mechanical wear exist simultaneously, the atlas that fuses multiple source features can accurately trace the fault root cause.

[0052] In an embodiment, step S30 includes: S31: comparing the runtime sequence data with the preset time difference value with the expected state data, and obtaining a difference value, generating a compensation amount based on the difference value through a PID controller, and correcting the device control instruction to realize closed-loop correction; In the embodiment, the preset time difference value is the alignment offset of the runtime sequence data and the expected state data in the time dimension, which can be dynamically matched by using a sliding window mechanism, for example, a 500 millisecond window is set to eliminate the timing error between sensor acquisition delay and model calculation delay; the difference value is the numerical difference between the actual runtime sequence data of the device and the expected state data output by the simulation model at the same time, which directly reflects the deviation between the actual behavior of the system and the prediction of the ideal model; the PID controller is a closed-loop control unit constructed based on a proportional-integral-derivative algorithm, which can realize dynamic compensation by adjusting three parameters such as the proportional coefficient, the integral time constant, and the derivative time constant, for example, the proportional coefficient is set to 0.8 to achieve fast response; the compensation amount is a correction value calculated by the PID controller for adjusting the original device control instruction, which aims to offset system errors or unknown disturbances, so that the prediction of the simulation model is closer to the response of the actual system; the closed-loop correction is a process of continuously optimizing the control instruction through a feedback mechanism, which can use an iterative correction strategy, for example, the instruction is updated every 200 milliseconds to balance the calculation overhead and correction accuracy.

[0053] S32: inputting the device control instruction sequence generated after closed-loop correction to the lightweight simulation model to obtain an expected state data sequence; In the embodiment, the device control instruction sequence is a series of new device control commands arranged in time sequence generated after being corrected by the closed-loop correction link, and the device control instruction sequence already contains the adjustment amount added to compensate for model errors and external disturbances; the expected state data sequence is a sequence of predicted values of the device state parameters in a period of time arranged in time sequence calculated by the lightweight simulation model after receiving the corrected device control instruction sequence.

[0054] S33: performing point-by-point comparison in time domain between the expected state data sequence and the runtime sequence data sequence to generate a residual sequence.

[0055] In this embodiment, the point-by-point comparison is an element-level difference calculation of the two data sequences under the same time reference, which can specifically adopt the Euclidean distance or the dynamic time warping algorithm, for example, data alignment with millisecond-level timestamps as indexes.

[0056] Specifically, when the expected state data is generated after the device control instruction is processed by the lightweight simulation model, due to the dynamic characteristics such as mechanical inertia and sensor delay in the actual device operation, directly comparing the expected data with the real-time collected runtime sequence data may cause error accumulation; therefore, data time sequence alignment is achieved by introducing a preset time difference, for example, the real-time collected current signal is delayed for 300 milliseconds and then matched with the expected voltage data output by the simulation; the PID controller generates a compensation amount according to the comparison difference, for example, when it is detected that the mechanical arm position deviation exceeds 2 millimeters, the control current parameter of the servo motor is automatically adjusted; the control instruction sequence after closed-loop correction is re-input into the simulation model to generate an expected state data sequence closer to the actual working condition; finally, through time-domain point-by-point comparison, for example, the temperature and pressure values of each sampling point are subtracted, a residual sequence reflecting the system deviation is formed.

[0057] Through the above technical solution, the present application effectively solves the residual calculation distortion problem caused by time sequence mismatch in traditional fault diagnosis, and improves the accuracy of the expected state data through a dynamic closed-loop correction mechanism; the generated residual sequence can more accurately represent the deviation between the actual state of the device and the ideal state, providing high-quality input data for subsequent causal analysis.

[0058] In an embodiment, step S40 includes: S41: distributing the received runtime sequence data and the residual sequence to the causal paths in the lightweight causal graph model according to a predefined configuration relationship; In this embodiment, the predefined configuration relationship is a mapping rule between the runtime sequence data and the causal paths established in advance, which can be implemented by a configuration file or a rule engine, and is used to guide the flow direction of data in the causal graph; the causal path is a logical link representing the causal relationship between specific device variables in the causal graph, which can be constructed by connecting nodes by directed edges, and each path corresponds to the causal logic of a device operating state.

[0059] S42: the causal path receiving the runtime sequence data and the residual sequence starts its calculation process, and generates corresponding violation event information when there is an abnormality in the calculation process of any causal path. In the embodiment, the computing process is a series of standardized and automatically executed data processing and logical judgment steps encapsulated in the internal causal path, which starts from receiving the distributed raw data and ends with outputting the structured diagnostic result (violation event information or no event output), and the core purpose is to determine whether the specific causal relationship monitored by the path is violated.

[0060] S43: Collect all violation event information generated by the causal path and combine and output as a causal violation event set.

[0061] In the embodiment, the violation event information is the structured diagnostic result output by a single causal path after determining that it has an abnormality. The violation event information at least includes which path the abnormality occurs in (causal path identifier), how big the abnormality is (difference value), and when the abnormality occurs (abnormality timestamp).

[0062] Specifically, the runtime sequence data and the residual sequence are allocated to the corresponding causal path through a pre-defined configuration relationship. Each causal path calculates the input data according to its internal logic; when the calculation result of a causal path deviates significantly from the expected state and exceeds the dynamic threshold, the system determines that the path has an abnormality, and generates violation event information including the path identifier, the difference value, and the timestamp; the violation events of all paths are collected to form a causal violation event set, which reflects the abnormal state of multiple causal relationship links in the device operation, and provides multi-dimensional abnormal clues for subsequent fault diagnosis.

[0063] Through the above technical solution, the application realizes fine monitoring of the operation state of the battery swap station device, improves the comprehensiveness and accuracy of fault detection through multi-causal path collaborative analysis, and reduces the false positive rate caused by environmental interference through the dynamic threshold mechanism, thereby providing a reliable basis for quickly locating the fault root.

[0064] In an embodiment, step S42 includes: S421: The causal path extracts cause data and result data from the received runtime sequence data and residual sequence according to its pre-defined data requirements; In the embodiment, the predefined data requirement is an input data specification configured for the causal path when it is constructed, which specifies specific parameter types that need to be extracted from the original data stream for the path to complete its causal analysis task, for example: "motor current" is required as cause data, and "motor speed" is required as result data; the cause data is the input data as the "cause" in a causal path, and the cause data is usually an independent variable or a precursor variable that can drive the system state to change, such as a control instruction, an output state of an upstream device, etc.; the result data is the input data as the "effect" in a causal path, and the result data is usually a dependent variable or a successor variable that is caused or affected by the cause data, such as the running state of a downstream device, a feedback value of a controlled object, etc.

[0065] S422: Calculate a predicted value of the result data based on the cause data and a predefined constraint function, and compare the calculated predicted value with the result data to obtain a difference value; In the embodiment, the constraint function is a mathematical expression describing the physical relationship between the cause data and the result data, which can be implemented by a transfer function or an empirical formula based on a device dynamics equation, and is used to establish a causal relationship model under normal working conditions; the predicted value is a theoretical value of the result data under ideal conditions calculated by inputting the cause data into the constraint function; and the difference value is a deviation between the predicted value and the actual result data, which can be implemented by calculating the Euclidean distance or the relative error percentage, and is used to quantify the abnormality degree of the causal path.

[0066] S423: Compare the difference value with a dynamic threshold of the causal path, if the difference value is greater than the dynamic threshold and the duration is greater than a preset time threshold, determine that the causal path is abnormal, and generate a violation event information according to the determination result, the violation event information includes a causal path identifier, a difference value and an abnormal time stamp.

[0067] In the embodiment, the dynamic threshold is a determination boundary that is self-adaptively adjusted according to the device running condition, which can be implemented by a multiple of the sliding window standard deviation based on historical data statistical analysis, and is used to adapt to the detection sensitivity requirement under different working conditions; the time threshold is a minimum time length of the abnormal state, which can be set to 1.5-2 times of the response delay time of the device, and is used to exclude false judgments caused by transient interference; the causal path identifier is a number or code used to uniquely identify a specific causal path in a lightweight causal graph model, and the causal path identifier ensures that each generated violation event can be accurately traced back to its source.

[0068] Specifically, when new time series data is generated during the operation of the device, the predefined causal path first extracts the cause variable and the result variable matching the logical relationship thereof from the data stream; for example, in the battery charging control path, the charging current value is the cause data, and the battery temperature change rate is the result data; the cause data is calculated by the constraint function to generate a theoretical prediction value, which represents the range of result data that should be observed under normal state of the device; the actual collected result data is compared with the prediction value in real time, and if the difference value exceeds the dynamic threshold corresponding to the current working condition and the duration reaches the preset standard, it is determined that the causal path has an abnormality; for example, when the charging current instruction is 50A, according to the thermodynamic model, the battery temperature rise rate should be in the range of 0.8-1.2 ℃ / min, and if the actual detected temperature rise rate exceeds 1.5 ℃ / min for 3 minutes, a violation event record is triggered.

[0069] Through the above technical solution, the application can accurately identify abnormal events that violate the established causal logic during the operation of the device, effectively distinguish between real device failures and temporary interference; through the dual judgment mechanism of dynamic threshold and time threshold, the false positive rate is significantly reduced, while the timeliness of fault detection is ensured; the generated violation event information contains specific causal path identification and abnormal time characteristics, providing accurate input features for subsequent fault mode matching, which helps to quickly locate the fault root cause.

[0070] It should be understood that the size of the serial number of each step in the above embodiments does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the application.

[0071] In an embodiment, a battery swap station device fault diagnosis system based on edge computing is provided, which corresponds one-to-one to the above-mentioned embodiment of a battery swap station device fault diagnosis method based on edge computing. The battery swap station device fault diagnosis system based on edge computing comprises: a model construction module configured to obtain system device information and historical operation data of a target battery swap station, and construct a regional device model, the regional device model comprising a lightweight causal graph model, a lightweight simulation model, and a fault mode atlas; a model deployment module configured to deploy the regional device model on an edge computing node associated with the target battery swap station, and collect runtime series data of the target battery swap station in real time through the edge computing node; a data comparison module configured to obtain a device control instruction and input it into the lightweight simulation model to obtain expected state data, and compare the expected state data with the runtime series data to generate a residual sequence; The first data input module is configured to input the residual sequence and the runtime sequence data into the lightweight causal diagram model, and output a set of causal violation events when the lightweight causal diagram model identifies a violation event. The second data input module is configured to input the residual sequence and the set of causal violation events into the fault mode atlas, and enable the fault mode atlas to generate and output a fault diagnosis result.

[0072] The specific limitations of the edge computing-based battery swap station equipment fault diagnosis system can refer to the limitations of the edge computing-based battery swap station equipment fault diagnosis method described above, and will not be repeated here. Each module in the edge computing-based battery swap station equipment fault diagnosis system described above can be realized by software, hardware, or a combination thereof, in whole or in part. Each module described above can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each module.

[0073] In an embodiment, a computer device is provided, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements an edge computing-based battery swap station equipment fault diagnosis method when executing the computer program.

[0074] In an embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement an edge computing-based battery swap station equipment fault diagnosis method.

[0075] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. An edge computing-based battery swap station equipment fault diagnosis method, characterized in that, The method comprises the steps of: obtaining system equipment information and historical operation data of a target battery swap station, and constructing a regional equipment model, the regional equipment model comprising a lightweight causal diagram model, a lightweight simulation model and a fault mode atlas; deploying the regional equipment model on an edge computing node associated with the target battery swap station, and collecting real-time operation time series data of the target battery swap station through the edge computing node; obtaining a device control instruction and inputting the device control instruction into the lightweight simulation model to obtain expected state data, and comparing the expected state data with the operation time series data to generate a residual sequence; inputting the residual sequence and the operation time series data into the lightweight causal diagram model, and outputting a causal violation event set when the lightweight causal diagram model identifies a violation event; inputting the residual sequence and the causal violation event set into the fault mode atlas, so that the fault mode atlas generates and outputs a fault diagnosis result. 2.The method of claim 1, wherein: The step of obtaining system equipment information and historical operation data of a target battery swap station, and constructing a regional equipment model, the regional equipment model comprising a lightweight causal diagram model, a lightweight simulation model and a fault mode atlas, comprises the steps of: constructing an initial causal diagram model and an initial simulation model based on system equipment information and historical operation data; constructing a generative adversarial network with a generator and a discriminator, wherein the generator takes a device control instruction as input, and the discriminator takes operation time series data and simulation data output by the generator as input; embedding the initial simulation model into the generator of the generative adversarial network to form a physical information enhanced enhanced generative adversarial network; training the enhanced generative adversarial network through historical operation data, and freezing the lightweight simulation model in the generator after training and extracting it as a lightweight simulation model; collecting discrimination probability distribution information output by the discriminator during training, and optimizing the initial causal diagram model based on the discrimination probability distribution information to obtain a lightweight causal diagram model; obtaining expert knowledge information, and constructing a fault mode atlas based on historical fault data and expert knowledge. 3.The method of claim 2, wherein: The step of collecting discrimination probability distribution information output by the discriminator during training, and optimizing the initial causal diagram model based on the discrimination probability distribution information to obtain a lightweight causal diagram model, comprises the steps of: during model training, collecting discrimination probability distribution information output by the discriminator for historical operation time series data, and calculating an information entropy value thereof; identifying device operating conditions based on historical operation time series data, and associating and mapping the calculated information entropy value with the device operating conditions; obtaining a discretization level of the device operating conditions and the information entropy value forming the association mapping, and combining the obtained discretization levels to form a joint state space; based on historical operation data, filtering out historical data segments corresponding to each state unit in the joint state space; based on the filtered historical data segments, calculating an optimal violation detection threshold for each state unit by using an optimization algorithm with the optimization objective of maximizing fault detection rate and minimizing false alarm rate; and A lightweight dynamic threshold mapping table is constructed based on each state unit and the corresponding optimal violation detection threshold, and the dynamic threshold mapping table is integrated into the initial causal graph model to obtain a lightweight causal graph model.

4. The edge computing-based battery swap station equipment fault diagnosis method of claim 2, characterized in that: The step of acquiring expert knowledge information and constructing a fault mode graph based on historical fault data and expert knowledge comprises the following steps: The expert knowledge information is parsed by a pre-trained large language model to extract fault entity features and entity relationship features, and key signal features are extracted based on historical fault data, wherein the key signal features include time domain features, frequency domain features and time-frequency domain features; The extracted fault entity features, entity relationship features and key signal features are fused and aligned to generate a fault representation vector; An initial skeleton of the fault mode graph is constructed based on the fault entity features and entity relationship features, and the initial skeleton is embedded and learned based on the fault representation vector to obtain an optimized skeleton; Strong association rules between the key signal features and the fault entity features are mined from the fault representation vector by a pre-set association rule mining algorithm, and strong association rules with a confidence higher than a pre-set threshold are added as new edges to the optimized skeleton to obtain the fault mode graph.

5. The edge computing-based battery swap station equipment fault diagnosis method according to claim 1, characterized in that: The step of acquiring the device control instruction and inputting it into the lightweight simulation model to obtain expected state data, and comparing the expected state data with the runtime sequence data to generate a residual sequence comprises the following steps: The runtime sequence data with a pre-set time difference value is compared with the expected state data, and a difference value is obtained, and a compensation amount is generated by a PID controller based on the difference value to correct the device control instruction to realize closed-loop correction; The device control instruction sequence generated after closed-loop correction is input into the lightweight simulation model to obtain an expected state data sequence; The expected state data sequence and the runtime sequence data sequence are compared point by point in the time domain to generate a residual sequence.

6. The edge computing-based battery swap station equipment fault diagnosis method according to claim 1, characterized in that: The step of inputting the residual sequence and the runtime sequence data into the lightweight causal graph model and outputting a causal violation event set when the lightweight causal graph model identifies a violation event comprises the following steps: According to a pre-defined configuration relationship, the received runtime sequence data and residual sequence are distributed to the causal paths in the lightweight causal graph model; The causal path receiving the runtime sequence data and the residual sequence starts its calculation process, and generates corresponding violation event information when the calculation process of any causal path is abnormal; The violation event information generated by all causal paths is collected and combined to output as a causal violation event set.

7. The edge computing-based battery swap station equipment fault diagnosis method according to claim 6, characterized in that: The step of receiving the runtime sequence data and the residual sequence by the causal path and starting its calculation process, and generating corresponding violation event information when the calculation process of any causal path is abnormal comprises the following steps: The causal path extracts cause data and result data from the received runtime sequence data and residual sequence according to its pre-defined data requirements; The predicted value of the result data is calculated based on the cause data and a pre-defined constraint function, and the calculated predicted value is compared with the result data to obtain a difference value; The difference value is compared with a dynamic threshold value of the causal path, if the difference value is greater than the dynamic threshold value and the duration is greater than a preset time threshold value, it is determined that the causal path has an anomaly, and event information against the anomaly is generated according to the determination result, the event information against the anomaly includes a causal path identifier, a difference value and an anomaly timestamp.

8. An edge computing-based battery swap station equipment fault diagnosis system, characterized in that, The method comprises the steps of: The model construction module is configured to obtain system equipment information and historical operation data of a target battery swap station, and construct a regional equipment model, the regional equipment model comprising a lightweight causal graph model, a lightweight simulation model and a fault mode atlas; The model deployment module is configured to deploy the regional equipment model to an edge computing node associated with the target battery swap station, and collect real-time operation time series data of the target battery swap station through the edge computing node; The data comparison module is configured to obtain a device control instruction and input the device control instruction into the lightweight simulation model to obtain expected state data, and compare the expected state data with the operation time series data to generate a residual sequence; The first data input module is configured to input the residual sequence and the operation time series data into the lightweight causal graph model, and output a causal violation event set when the lightweight causal graph model identifies a violation event; The second data input module is configured to input the residual sequence and the causal violation event set into the fault mode atlas, so that the fault mode atlas generates and outputs a fault diagnosis result.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The computer program is executed by the processor to implement the steps of the battery swap station equipment fault diagnosis method based on edge computing according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the steps of the battery swap station equipment fault diagnosis method based on edge computing according to any one of claims 1-7.

Images