Adjustable circuit fault diagnosis method based on variational auto-encoder
Through the hierarchical structure and dynamic adaptive mechanism of the variational autoencoder, the shortcomings of weak anomaly detection and fault propagation path analysis in existing automotive circuit fault diagnosis methods are solved, efficient and accurate circuit fault diagnosis is achieved, and the operating safety and maintenance efficiency of smart cars are improved.
Patent Information
- Application Number
- CN202510833755.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-20
- Publication Date
- 2025-10-17
AI Technical Summary
Existing automotive circuit fault diagnosis methods are insufficient in detecting weak anomalies, lack adaptive diagnostic capabilities, and have difficulty in effectively analyzing fault propagation paths, resulting in insufficient diagnostic accuracy and reliability, making it difficult to meet the complex circuit system requirements of smart cars.
A hierarchical structure based on variational autoencoders is adopted. Through regional and system-level variational autoencoder models, multi-region voltage, current and temperature data are combined to perform fine-grained feature extraction and global feature analysis. Dynamic thresholds and adaptive parameter optimization are introduced, and the inter-region interaction matrix is dynamically updated to achieve accurate fault location and propagation assessment.
It significantly improves the sensitivity and accuracy of circuit fault diagnosis, enhances the adaptability and robustness in different environments, can accurately trace and evaluate the fault propagation path, and improves the operational safety and maintenance efficiency of the circuit system.
Smart Images
Figure CN120804968A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of circuit, more particularly, to a variable circuit fault diagnosis method based on variational autoencoder. BACKGROUND
[0002] With the rapid development of automotive electronics technology, the electronic and electrical systems of modern vehicles are becoming increasingly complex, covering engine management systems, body control systems, braking systems, sensor networks, and intelligent auxiliary driving systems. These circuit systems not only improve the intelligence, safety, and comfort of vehicles, but also bring complex fault diagnosis challenges. How to efficiently and accurately detect potential faults in automotive electronic systems to ensure stable vehicle operation has become an important technical problem faced by the automotive industry.
[0003] Currently, automotive circuit fault diagnosis mainly relies on traditional on-board diagnostic systems and expert systems for fault detection. The OBD system monitors various electrical signals in real time through sensors and determines whether a fault has occurred based on preset threshold rules. However, there are significant shortcomings when facing complex working conditions: on the one hand, the OBD system usually monitors voltage, current, and temperature signals based on preset thresholds, making it difficult to effectively identify weak and short-term abnormal signals, which can lead to missed or false alarms; on the other hand, the OBD system is limited in fault code analysis, making it difficult to accurately locate the fault area, and can only provide preliminary warning information, requiring maintenance personnel to rely on experience for further investigation.
[0004] In recent years, deep learning-based automotive circuit fault diagnosis methods have gradually been applied. Through large-scale data training, deep learning models can improve the accuracy of anomaly detection to some extent. However, existing methods still have the following problems: first, most deep learning methods use a single model structure, which cannot balance local circuit features and system-level global associations, limiting the accuracy of fault diagnosis; second, existing methods lack adaptive ability and cannot dynamically adjust according to the running state of the vehicle under different working conditions, leading to a decline in the reliability of diagnosis results in different environments; third, existing methods have weak analysis capabilities for fault propagation paths, making it difficult to effectively capture the cascading effects of circuit faults among multiple systems, such as sensor faults that may affect engine management systems or intelligent driving systems. Existing methods often fail to model this complex relationship, affecting the accurate positioning of faults.
[0005] In summary, the prior art has the problems of insufficient weak anomaly detection capability, lack of adaptive diagnosis capability and difficulty in effectively analyzing fault propagation path in the field of automobile circuit fault diagnosis, which limits its application in intelligent automobile maintenance. Therefore, a new method is needed, which can combine deep learning and generative model to realize accurate fault diagnosis of complex automobile circuit system through multi-level feature extraction, adaptive parameter adjustment and inter-regional interaction analysis, and improve the safety and stability of vehicle operation. SUMMARY
[0006] The purpose of the present application is to overcome the problems of insufficient weak anomaly detection capability, lack of adaptive diagnosis capability and difficulty in effectively analyzing fault propagation path in the prior art automobile circuit fault diagnosis field, and provide an adjustable circuit fault diagnosis method based on variational autoencoder. The present application can efficiently cope with the fault diagnosis demand in complex circuit system, significantly improve the accuracy, adaptability and reliability of diagnosis, and provide strong support for the operation safety and maintenance efficiency of circuit system.
[0007] To solve the above technical problems, the technical scheme adopted by the present application is: An adjustable circuit fault diagnosis method based on variational autoencoder is provided, which comprises the following steps: S1: Collecting the original running data of multi-region voltage, current and temperature of the circuit system through sensors arranged in different regions of the circuit system to form a regionalized running data set covering all regions; S2: Data preprocessing of the regionalized running data set to obtain a regionalized standard running data set; S3: Constructing a hierarchical variational autoencoder model, which includes a plurality of regional-level variational autoencoder sub-models and a system-level variational autoencoder main model; S4: Training the hierarchical variational autoencoder model using the regionalized standard running data set, inputting the standard running data of each region into the corresponding regional-level variational autoencoder sub-model, optimizing the latent variable space of the regional-level variational autoencoder sub-model, and making the regional-level variational autoencoder sub-model represent the characteristics of the running data of each region; S5: Inputting the latent variables of the regional-level variational autoencoder sub-model into the system-level variational autoencoder main model, optimizing the global feature representation capability of the system-level variational autoencoder main model by minimizing the reconstruction error and the difference in latent variable distribution; S6: Regionalizing and segmenting the new running data of the circuit system, dividing the new running data into regionalized running data subsets corresponding to the regional-level variational autoencoder sub-models, and performing standardization processing on each regionalized running data subset to generate regionalized real-time standard running data subsets; S7: input the regionalized real-time standard running data subset into the corresponding regional level variational autoencoder sub-model, generate reconstructed data through the regional level variational autoencoder sub-model and calculate regional reconstruction error; S8: input the latent variable representation of all regional level variational autoencoder sub-models into the system level variational autoencoder main model, generate global reconstruction data through the system level variational autoencoder main model and calculate global reconstruction error; S9: if the regional reconstruction error of a certain region or the global reconstruction error of the system level variational autoencoder main model exceeds the respective preset threshold value, analyze the fault type and source in the region in combination with the latent variable distribution characteristics of the corresponding region, and optimize the fault location result through the system level variational autoencoder main model to output the diagnosis result.
[0008] Preferably, the step S1 specifically comprises the following steps: S11: arranging sensors in a plurality of regions of the circuit system, the sensors of each region being used to collect voltage data , current data and temperature data of the region, wherein i represents the region number, t represents the sampling time, and the voltage data , current data and temperature data are used to form a regional running data set of the region: ; wherein N is the number of regions divided in the circuit system, and represent the starting time and the ending time of sampling, respectively; S12: performing integrity check on the regional running data set of each region, and if a certain data item is missing at a certain time t, marking the corresponding data in the regional running data set as invalid.
[0009] Preferably, the step S2 comprises the following steps: S21. performing outlier rejection on the regional running data set , defining outliers as data points exceeding the preset threshold range, and generating a regional running data set without outliers after rejection; S22. performing normalization processing on the regional running data set without outliers, normalizing the voltage data, current data and temperature data of each region to the interval [0, 1], and generating a standardized regional running data set after normalization; S23. filtering out noise from the standardized regional running data set by using a low-pass filtering method to eliminate high-frequency noise interference, and generating a regionalized standard running data set after noise filtering .
[0010] Preferably, the step S3 comprises the following steps: S31: For each region i of the circuit system, design a region-level variational autoencoder submodel for extracting key features and potential abnormal signals from the region's operational data, which contains an encoder and a decoder: ; wherein, represents the circuit operational data of region i, including voltage, current and temperature signals, represents the latent variable distribution of region i, which is used to capture the low-dimensional features and potential abnormal information of the region's operational data, represents the dynamic weight vector related to region i, and are the parameters of the encoder and the decoder, respectively; S32: Design a system-level variational autoencoder main model for integrating the latent variable representations of all region-level submodels and capturing the interaction features between regions, which introduces a region interaction matrix to represent the relationship between the latent variables of each region: ; wherein, represents the correlation between the latent variables of region i and region j, reflecting the degree of coupling between the regions, is a similarity function used to quantify the relationship between the latent variables of region i and region j, implemented using a Gaussian kernel; The encoder and decoder of the system-level variational autoencoder main model are defined as: ; wherein, represents the set of all region latent variables, which is used to capture global features and inter-regional interaction information, and are the parameters of the main model encoder and decoder, respectively.
[0011] Preferably, the step S4 comprises the following steps: S41: Input the regionalized standard operational data set into the region-level variational autoencoder submodel, define the input data of region i as , which is used to initialize the parameters of the encoder and the decoder and , so that the initial state of the region-level variational autoencoder submodel reflects the basic features of the circuit operation; S42: using the input data and the dynamic weight vector of each region i training the region-level variational autoencoder sub-model, optimizing the parameters of the region-level variational autoencoder sub-model by minimizing the loss function; S43: after the training of the region-level variational autoencoder sub-model is completed, the reconstruction error of each region i is calculated to evaluate the feature extraction capability of the region-level variational autoencoder sub-model for the region operation data: ; wherein, and represent the input data and the reconstructed data of region i at time t, represents the Euclidean distance, which is used to quantify the reconstruction accuracy of the model for the circuit operation data; if the reconstruction error exceeds the set threshold, the dynamic weight vector or the dimension of the latent variable distribution is adjusted, and the region-level variational autoencoder sub-model is re-optimized.
[0012] Preferably, in the step S42, the loss function is minimized as follows: ; wherein, represents the reconstruction capability of the region-level variational autoencoder sub-model for the input data of region i, which is used to restore the detailed characteristics of the circuit operation data, represents the difference between the latent variable distribution and the prior distribution , which is used to constrain the consistency of the latent variable space distribution, is a dynamically adjusted regularization coefficient, which is used to balance the data reconstruction and the consistency of the latent variable distribution according to the complexity of the region operation data.
[0013] Preferably, the step S5 comprises the following steps: S51: collecting the latent variables generated by each region-level variational autoencoder sub-model to form a latent variable set Z; S52: taking the latent variable set Z as the input of the system-level variational autoencoder main model, combining the region interaction matrix to train the system-level variational autoencoder main model, and optimizing the loss function; S53: evaluating the feature extraction capability of the model for the overall operation state of the circuit system by calculating the global reconstruction error of the system-level variational autoencoder main model , and the reconstruction error is defined as: ; wherein, X(t) and respectively represent the input and reconstruction result of global operation data at time t; If the reconstruction error exceeds a set threshold, adjust the regional interaction matrix or the latent variable dimension of the system-level variational autoencoder master model, and re-optimize the training process; S54: Finally output the optimized system-level variational autoencoder master model, including the encoder parameters , decoder parameters and optimized global latent variable distribution Z.
[0014] Preferably, in the step S52, the loss function is: ; wherein, represents the reconstruction ability of the system-level variational autoencoder master model to the regionalized operation data X, represents the difference between the latent variable distribution and the prior distribution p(Z), used to constrain the consistency of the global latent variable distribution, is a regularization coefficient, used to adjust the trade-off between global reconstruction ability and latent variable distribution consistency.
[0015] Preferably, the S9 comprises the following steps: S91: Compare the regional-level reconstruction error and the system-level reconstruction error with the corresponding dynamic threshold and , and the dynamic threshold is calculated by combining the regional characteristics and the global operation state:
[0016] ; wherein, and are respectively used for the adjustment coefficients of the regional level and the system level, reflecting the weight of the abnormal sensitivity under the specific regional or global operation condition, is a global weight factor, used to balance the influence of the average signal value and the signal fluctuation range, and respectively represent the standard deviation of the regional i and the global operation data, used to quantify the signal fluctuation range; When the reconstruction error meets one of the following conditions, it is determined that there is an anomaly: ; S92: For the abnormal region i exceeding the threshold, extract the feature of the latent variable distribution from the regional-level variational autoencoder sub-model, analyze the type of abnormal signal, and analyze and the range of variation determines the fault type; S93: Combine the hidden variable set Z of the system-level variational autoencoder master model and the regional interaction matrix to optimize the fault location result of the abnormal region, and update the dynamic weight of the interaction matrix: ; wherein, represents the dynamic interaction weight of the abnormal region i and other regions j at iteration t+1, represents the Euclidean distance between the hidden variable features of the abnormal region and other regions, which is used to measure the degree of association between regions; S94: Generate multi-level fault location and diagnosis results according to the regional hidden variable features and the optimized interaction matrix.
[0017] Preferably, in the step S94, the diagnosis result includes the identification result of the fault region, the classification information of the fault type, the range of fault propagation and the affected other regions, and the overall evaluation result of the system-level running state Compared with the prior art, the present application has the following advantages: (1) The present application constructs a hierarchical structure of regional-level variational autoencoder sub-model and system-level variational autoencoder master model. The regional-level sub-model extracts fine-grained features from circuit operation data, captures the hidden variable distribution characteristics of local operation data, and accurately identifies subtle abnormal signals within the region. The system-level master model further extracts global features and analyzes the interaction between regions by integrating the hidden variable representation of all regions, which can simultaneously consider local features and global features, significantly improving the sensitivity and accuracy of circuit fault diagnosis.
[0018] (2) The present application introduces a dynamic threshold and adaptive parameter optimization mechanism during model training and diagnosis. By dynamically adjusting the regional-level reconstruction error threshold and the global reconstruction error threshold based on regional characteristics and global running state, the model can automatically optimize the diagnosis strategy according to different circuit operation conditions. At the same time, the dynamic update of the regional interaction matrix effectively models the fault propagation path and correlation strength between regions, significantly improving the robustness of fault location and the adaptability in variable scenarios, which can maintain consistent diagnosis effect in different circuit environments.
[0019] (3) The present invention is based on the in-depth analysis of latent variable distribution. It distinguishes different types of fault signals by extracting the mean and variance of regional latent variables. It dynamically updates the interaction weights between abnormal regions and other regions in combination with the regional interaction matrix in the system-level model, and accurately analyzes the path and impact range of fault propagation. The multi-level fault location and propagation analysis mechanism effectively makes up for the shortcomings of traditional methods in modeling inter-regional correlations in complex circuit systems, and can achieve accurate tracing and propagation evaluation of faults. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 This is a flow chart of the adjustable circuit fault diagnosis method based on variational autoencoder of the present invention. DETAILED DESCRIPTION
[0021] The present invention will be further described below in conjunction with specific embodiments.
[0022] Example 1 like Figure 1 As shown, a method for diagnosing faults in an adjustable circuit based on a variational autoencoder comprises the following steps: S1: Sensors placed in different areas of the circuit system collect the original operating data of voltage, current, and temperature in multiple areas of the circuit system to form a regionalized operating data set covering all areas; S2: preprocess the regionalized operation dataset to obtain the regionalized standard operation dataset; S3: Construct a hierarchical variational autoencoder model, which includes multiple region-level variational autoencoder sub-models and a system-level variational autoencoder main model; S4: Use the regionalized standard operating data set to train the hierarchical variational autoencoder model. Input the standard operating data of each region into the corresponding regional variational autoencoder sub-model, optimize the latent variable space of the regional variational autoencoder sub-model, and make the regional variational autoencoder sub-model represent the characteristics of the operating data of each region. S5: Input the latent variables of the region-level variational autoencoder sub-model into the system-level variational autoencoder main model, and optimize the global feature representation capability of the system-level variational autoencoder main model by minimizing the reconstruction error and the difference in latent variable distribution; S6: Perform regional segmentation on the new operating data of the circuit system, dividing the new operating data into regionalized operating data subsets corresponding to the regional-level variational autoencoder sub-models, and performing standardization processing on each regionalized operating data subset to generate a regionalized real-time standard operating data subset; S7: input the regionalized real-time standard running data subset into the corresponding regional level variational autoencoder sub-model, generate reconstructed data through the regional level variational autoencoder sub-model and calculate regional reconstruction error; S8: input the latent variable representation of all regional level variational autoencoder sub-models into the system level variational autoencoder main model, generate global reconstructed data through the system level variational autoencoder main model and calculate global reconstruction error; S9: if the regional reconstruction error of a certain region or the global reconstruction error of the system level variational autoencoder main model exceeds the respective preset threshold, analyze the fault type and source in the region in combination with the latent variable distribution characteristics of the corresponding region, and optimize the fault location result through the system level variational autoencoder main model to output the diagnosis result.
[0023] In one of the embodiments, the step S1 specifically comprises the following steps: S11: arrange sensors in multiple regions of the circuit system, and the sensors of each region are used to collect voltage data , current data and temperature data of the region, wherein i represents the region number, t represents the sampling time, and the voltage data , current data and temperature data are used to form a regional running data set of the region: wherein N is the number of regions divided in the circuit system, and represent the starting time and the ending time of sampling, respectively; S12: perform integrity check on the regional running data set of each region, and if a certain data item is missing at a certain time t, mark the corresponding data in the regional running data set as invalid.
[0024] In one of the embodiments, the step S2 comprises the following steps: S21. perform outlier rejection on the regional running data set , define the outlier as a data point exceeding the preset threshold range, and generate a regional running data set without outliers after rejection; S22. perform normalization processing on the regional running data set without outliers, normalize the voltage data, current data and temperature data of each region to the interval [0, 1], and generate a standardized regional running data set after normalization; S23. Adopting a low-pass filtering method to filter out noise from the standardized regional operation data set, eliminating high-frequency noise interference, and generating a noise-filtered regional standardized operation data set after filtering .
[0025] Embodiment 2 The difference from Embodiment 1 is that the step S3 comprises the following steps: S31: For each region i of the circuit system, a region-level variational autoencoder sub-model is designed to extract key features and potential abnormal signals of the regional operation data, which contains an encoder and a decoder: ; Wherein, represents the circuit operation data of region i, including voltage, current and temperature signals, represents the latent variable distribution of region i, which is used to capture the low-dimensional features and potential abnormal information of the regional operation data, represents the dynamic weight vector related to region i, and are the parameters of the encoder and the decoder, respectively; S32: A system-level variational autoencoder main model is designed to integrate the latent variable representations of all region-level sub-models and capture the interaction features between regions. The system-level variational autoencoder main model introduces a region interaction matrix to represent the relationship between the latent variables of each region: ; Wherein, represents the correlation between the latent variables of region i and region j, reflecting the degree of coupling between the regions, is a similarity function used to quantify the relationship between the latent variables of region i and region j, realized by a Gaussian kernel; The encoder and decoder of the system-level variational autoencoder main model are defined as: ; Wherein, represents the set of all region latent variables, which is used to capture global features and inter-regional interaction information, and are the parameters of the main model encoder and decoder, respectively.
[0026] In one embodiment, the step S4 comprises the following steps: S41: Input the regional standardized operation data set into the region-level variational autoencoder sub-model, and define the input data of region i as parameters for initializing the encoder and the decoder and making the initial state of the region-level variational autoencoder sub-model reflect the basic characteristics of the circuit operation; S42: using the input data and the dynamic weight vector of each region i training the region-level variational autoencoder sub-model, optimizing the parameters of the region-level variational autoencoder sub-model by minimizing the loss function; S43: after the training of the region-level variational autoencoder sub-model is completed, the reconstruction error of each region i is calculated to evaluate the feature extraction capability of the region-level variational autoencoder sub-model for the region operation data: ; wherein, and represent the input data and the reconstructed data of region i at time t, represents the Euclidean distance, which is used to quantify the reconstruction accuracy of the model for the circuit operation data; if the reconstruction error exceeds a set threshold, the dynamic weight vector or the dimension of the latent variable distribution is adjusted, and the region-level variational autoencoder sub-model is re-optimized.
[0027] In one of the embodiments, in the step S42, the loss function is minimized as follows: ; wherein, represents the reconstruction capability of the region-level variational autoencoder sub-model for the input data of region i, which is used to restore the detailed characteristics of the circuit operation data, represents the difference between the latent variable distribution and the prior distribution , which is used to constrain the consistency of the latent variable space distribution, is a dynamically adjusted regularization coefficient, which is used to balance the data reconstruction and the consistency of the latent variable distribution according to the complexity of the region operation data.
[0028] In one of the embodiments, the step S5 comprises the following steps: S51: collecting the latent variables generated by each region-level variational autoencoder sub-model to form a latent variable set Z; S52: taking the latent variable set Z as the input of the system-level variational autoencoder main model, combining the region interaction matrix to train the system-level variational autoencoder main model, and optimizing the loss function; S53: Evaluate the feature extraction capability of the model on the overall operation state of the circuit system by calculating the global reconstruction error of the system-level variational autoencoder master model , the reconstruction error is defined as: ; wherein X(t) and represent the input and reconstruction result of the global operation data at time t, respectively; If the reconstruction error exceeds the set threshold, adjust the regional interaction matrix or the hidden variable dimension of the system-level variational autoencoder master model, and re-optimize the training process; S54: Finally output the optimized system-level variational autoencoder master model, including the encoder parameters , decoder parameters and the optimized global hidden variable distribution Z.
[0029] In one embodiment, in the step S52, the loss function is: ; wherein represents the reconstruction capability of the system-level variational autoencoder master model on the regional operation data X, represents the difference between the hidden variable distribution and the prior distribution p(Z), used to constrain the consistency of the global hidden variable distribution, is a regularization coefficient, used to adjust the trade-off between global reconstruction capability and hidden variable distribution consistency.
[0030] In one embodiment, the S9 comprises the following steps: S91: Compare the regional-level reconstruction error and the system-level reconstruction error with the corresponding dynamic threshold and , the dynamic threshold is calculated by combining the regional characteristics and the global operation state:
[0031] ; wherein and are the adjustment coefficients for regional and system levels, respectively, reflecting the weight of abnormal sensitivity under specific regional or global operation conditions, is a global weight factor, used to balance the influence of average signal value and signal fluctuation range, and represent the standard deviation of regional i and global operation data, respectively, used to quantify the signal fluctuation range; When the reconstruction error meets one of the following conditions, it is determined to be abnormal: ; S92: For abnormal regions i exceeding the threshold, extract the characteristics of latent variable distribution from the regional variational autoencoder sub-model, analyze the type of abnormal signal, and analyze the and The range of change determines the fault type; S93: Combine the latent variable set Z of the system-level variational autoencoder main model and the regional interaction matrix to optimize the fault location results of the abnormal area and update the dynamic weight of the interaction matrix: ; in, represents the dynamic interaction weight between abnormal region i and other regions j at iteration t+1, The Euclidean distance between the hidden variable characteristics of the abnormal region and other regions is used to measure the degree of association between regions; S94: Generate multi-level fault location and diagnosis results based on regional latent variable characteristics and the optimized interaction matrix.
[0032] In one embodiment, in step S94, the diagnosis result includes the identification result of the fault area, classification information of the fault type, the scope of fault propagation and other affected areas and the overall evaluation result of the system-level operating status.
[0033] Example 3 In January 2025, while cruising on a highway, an autonomous driving test vehicle experienced intermittent power drops in its power system. While no obvious alarm appeared on the driver monitoring interface, the vehicle's speed dropped by approximately 15% within a short period of time before returning to normal. Because traditional on-board diagnostic systems only set fault detection criteria based on static thresholds, the problem was not identified as a serious fault by the system. However, the testing team discovered that the problem occurred three times over the next two hours, initially suspecting an abnormality in the power supply module of the power system, but unable to determine the specific fault point. To this end, engineers decided to apply the present invention's circuit fault diagnosis method based on a hierarchical variational autoencoder structure with regionalized fault location to conduct an in-depth analysis of the problem.
[0034] The test vehicle's circuit system mainly consists of six functional modules, including a power control unit, a battery management system, a sensor fusion module, an autonomous driving computing unit, an auxiliary power module, and an on-board communication unit. The modules are interconnected through a high-voltage bus and a low-voltage control bus to ensure the normal operation of the power system, battery pack, and autonomous driving system.
[0035] When the vehicle is driving on a highway section in cruise mode, the power of the power system decreases by 15% in a short time, and the autonomous driving computing unit receives an abnormal signal, but the vehicle does not have emergency braking or safety strategy intervention. The traditional fault diagnosis method is based on fixed threshold setting, and this power fluctuation is not recognized as a serious abnormality, only a "power system short-term fluctuation" log information is recorded.
[0036] In order to accurately identify the fault source and its influence range, the test team uses the method of the application to analyze the current, voltage and temperature operation data of the power control unit, battery management system and other related modules in real time.
[0037] The data recording system of the test vehicle collected more than 1728000 operation data (20 times sampling per second) in 24 hours, covering power system voltage, current, temperature, module state information and autonomous driving computing unit feedback signal. In the preprocessing process, the system performs outlier rejection, normalization and noise filtering on all data to form a standardized operation data set.
[0038] The operation data of each module is input to the corresponding regional-level variational autoencoder sub-model for feature learning, generating hidden variable distribution and calculating reconstruction error. The power control unit appears at 9:02, with a hidden variable mean shift of 2.1%, a standard deviation increase of 17%, and a reconstruction error of 0.056, exceeding the set dynamic threshold of 0.042. At the same time, the hidden variable of the battery management system (BMS) also appears abnormal fluctuation, with a mean shift of 1.4% and a reconstruction error of 0.048.
[0039] Further analysis of the hidden variable interaction matrix of the system-level variational autoencoder master model found that the interaction weight between the power control unit and the battery management system increased from 0.38 in normal state to 0.81, indicating that the abnormal fluctuation of the power system is highly related to the voltage control abnormality of the battery management system. By analyzing the current signal of the battery management system, it is found that the charge and discharge control circuit of the battery management system appears a current reverse fluctuation within 5 seconds before the abnormal time point, and it is speculated that the fault may be caused by the abnormal work of the DC-DC conversion module in the battery management system, which limits the short-time power output of the power system.
[0040] Combined with historical data analysis, the test team confirms that the DC-DC conversion module of the vehicle has been running for more than 40000 hours, and the charge and discharge conversion efficiency has decreased by 3.5% in recent period. Through further inspection, it is found that the power MOSFET in the module has performance degradation signs, which leads to the decrease of conversion efficiency, and further affects the power output stability of the power control unit.
[0041] In addition, the method of the present application optimizes the model with 12000 training data, while the traditional method is only based on fixed historical threshold setting, lacking dynamic adaptability to real-time state, resulting in low fault detection accuracy. In addition, in the fault propagation path analysis, the method accurately identifies the coupling relationship between the power control unit and the battery management system, and correctly evaluates the fault propagation influence range, while the traditional method cannot establish dynamic connection between regions and can only perform local analysis based on single module data, resulting in high false alarm rate.
[0042] As can be seen from the present embodiment, the method of the present application has shown significant advantages in complex circuit system fault diagnosis. Its hierarchical feature extraction, adaptive dynamic threshold and regional interaction analysis enable it to accurately capture subtle abnormal signals in the circuit system, quickly locate the fault source and evaluate the propagation influence range, which not only improves the accuracy of fault detection in intelligent vehicle circuit system, but also optimizes the maintenance strategy, improves the operation reliability and safety of the vehicle, and provides more efficient circuit diagnosis solution for autonomous driving and intelligent electric vehicles.
[0043] The present application constructs a hierarchical structure of regional-level variational autoencoder sub-model and system-level variational autoencoder main model, the regional-level sub-model extracts fine-grained features from circuit operation data, captures the hidden variable distribution characteristics of local operation data, and accurately identifies subtle abnormal signals in the region, while the system-level main model integrates the hidden variable representation of all regions, further extracts global features and analyzes the interaction between regions, which can simultaneously consider local features and global features, significantly improving the sensitivity and accuracy of circuit fault diagnosis.
[0044] The present application introduces dynamic threshold and adaptive parameter optimization mechanism in model training and diagnosis process, dynamically adjusts regional-level reconstruction error threshold and global reconstruction error threshold by combining regional characteristics and global running state, so that the model can automatically optimize the diagnosis strategy according to different circuit running conditions, at the same time, the dynamic update of regional interaction matrix effectively models the fault propagation path and correlation strength between regions, significantly improves the robustness of fault location, and the adaptability in variable scene is obviously enhanced, which can maintain consistent diagnosis effect in different circuit environments.
[0045] Based on the deep analysis of hidden variable distribution, the present application distinguishes different types of fault signals by extracting the mean and variance of regional hidden variables, dynamically updates the interaction weight between abnormal regions and other regions by combining the regional interaction matrix in the system-level model, accurately analyzes the path and influence range of fault propagation, and the multi-level fault location and propagation analysis mechanism effectively makes up for the deficiency of traditional methods in modeling the correlation between regions in complex circuit system, which can realize accurate tracing and propagation evaluation of faults.
[0046] In the specific contents of the foregoing specific embodiments, each technical feature can be combined arbitrarily without contradiction. In order to make the description simple, all possible combinations of the foregoing technical features are not described, but as long as the combinations of the technical features do not contradict, they should be considered as the scope of the present disclosure.
[0047] Obviously, the above embodiments of the present application are merely exemplary and are not intended to limit the implementation of the present application. Based on the above description, other different forms of changes or variations can be made by those of ordinary skill in the art. Here, it is not necessary and impossible to exhaust all the embodiments. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the claims of the present application.
Claims
1. A method for diagnosing faults in adjustable circuits based on variational autoencoders, characterized in that: The method comprises the following steps: S1: Sensors placed in different areas of the circuit system collect the original operating data of voltage, current, and temperature in multiple areas of the circuit system to form a regionalized operating data set covering all areas; S2: preprocess the regionalized operation dataset to obtain the regionalized standard operation dataset; S3: Construct a hierarchical variational autoencoder model, which includes multiple region-level variational autoencoder sub-models and a system-level variational autoencoder main model; S4: Use the regionalized standard operating data set to train the hierarchical variational autoencoder model. Input the standard operating data of each region into the corresponding regional variational autoencoder sub-model, optimize the latent variable space of the regional variational autoencoder sub-model, and make the regional variational autoencoder sub-model represent the characteristics of the operating data of each region. S5: Input the latent variables of the region-level variational autoencoder sub-model into the system-level variational autoencoder main model, and optimize the global feature representation capability of the system-level variational autoencoder main model by minimizing the reconstruction error and the difference in latent variable distribution; S6: Perform regional segmentation on the new operating data of the circuit system, dividing the new operating data into regionalized operating data subsets corresponding to the regional-level variational autoencoder sub-models, and performing standardization processing on each regionalized operating data subset to generate a regionalized real-time standard operating data subset; S7: Input the regionalized real-time standard operation data subset into the corresponding regional-level variational autoencoder sub-model, generate reconstructed data through the regional-level variational autoencoder sub-model and calculate the regional reconstruction error; S8: Input the latent variable representations of all region-level variational autoencoder sub-models into the system-level variational autoencoder main model, generate global reconstruction data through the system-level variational autoencoder main model and calculate the global reconstruction error; S9: If the regional reconstruction error of a certain area or the global reconstruction error of the system-level variational autoencoder main model exceeds the respective preset thresholds, the type and source of the fault in the area are analyzed in combination with the latent variable distribution characteristics of the corresponding area, and the fault location result is optimized through the system-level variational autoencoder main model to output the diagnosis result.
2. The adjustable circuit fault diagnosis method based on variational autoencoder according to claim 1, characterized in that: The step S1 specifically includes the following steps: S11: Arrange sensors in multiple areas of the circuit system. The sensors in each area are used to collect voltage data in that area. , current data and temperature data , where i represents the region number, t represents the sampling time, according to the voltage data , current data and temperature data Forming a regionalized operational dataset for the region : ； Where N is the number of regions divided in the circuit system, and Respectively represent the start and end time of sampling; S12: Perform integrity check on the regionalized operation data set of each region. If a data item is missing at a certain time t, the corresponding data in the regionalized operation data set is marked as invalid.
3. The adjustable circuit fault diagnosis method based on variational autoencoder according to claim 1, characterized in that: The step S2 comprises the following steps: S21. Run the dataset on the regionalization Perform outlier removal, defining outliers as data points that exceed a preset threshold range. After removal, a regionalized operating data set without outliers is generated; S22. Normalize the regionalized operating data set without outliers, normalizing the voltage data, current data, and temperature data of each region to the interval [0, 1]. After normalization, generate a standardized regional operating data set. S23. Use low-pass filtering to remove noise from the standardized regional operation dataset to eliminate high-frequency noise interference. After filtering, generate a noise-filtered regional standard operation dataset. .
4. The adjustable circuit fault diagnosis method based on variational autoencoder according to claim 1, characterized in that: The step S3 comprises the following steps: S31: For each region i of the circuit system, a region-level variational autoencoder sub-model is designed to extract key features and potential abnormal signals of the operating data in the region. The region-level variational autoencoder sub-model consists of an encoder and a decoder: ； in, Represents the circuit operation data of area i, including voltage, current and temperature signals, represents the latent variable distribution of region i, which is used to capture the low-dimensional features and potential abnormal information of regional operation data. represents the dynamic weight vector associated with region i, and are the parameters of the encoder and decoder respectively; S32: Design a system-level variational autoencoder main model to integrate the latent variable representations of all regional sub-models and capture the interaction characteristics between regions. The system-level variational autoencoder main model introduces the regional interaction matrix Characterize the relationship between latent variables in each region: ； in, represents the correlation between the latent variables of region i and region j, reflecting the degree of circuit operation coupling between regions. It is a similarity function used to quantify the latent variable relationship between regions i and j, and is implemented using a Gaussian kernel; The encoder and decoder of the system-level variational autoencoder main model are defined as: ； in, Represents the set of all regional latent variables, which is used to capture global features and interaction information between regions. and are the parameters of the main model encoder and decoder respectively.
5. The adjustable circuit fault diagnosis method based on variational autoencoder according to claim 4, characterized in that: The step S4 comprises the following steps: S41: Regionalized Standard Operational Dataset Input into the region-level variational autoencoder sub-model, and define the input data of region i as , used to initialize the parameters of the encoder and decoder and , so that the initial state of the regional-level variational autoencoder sub-model reflects the basic characteristics of the circuit operation; S42: Using the input data and dynamic weight vector of each region i Train the region-level variational autoencoder sub-model and optimize its parameters by minimizing the loss function; S43: After the regional variational autoencoder sub-model training is completed, the reconstruction error of each region i is calculated , evaluate the feature extraction capability of the regional-level variational autoencoder sub-model for regional operation data: ； in, and denote the input data and reconstructed data of region i at time t, respectively. Represents the Euclidean distance, which is used to quantify the accuracy of the model's reconstruction of circuit operation data; If the reconstruction error If the value exceeds the set threshold, the dynamic weight vector or the latent variable distribution dimension is adjusted and the regional variational autoencoder sub-model is re-optimized.
6. The adjustable circuit fault diagnosis method based on variational autoencoder according to claim 5, characterized in that: In step S42, the minimization loss function is: ； in, Represents the region-level variational autoencoder sub-model for region i input data The reconstruction capability is used to restore the detailed characteristics of the circuit operation data. Represents the latent variable distribution With prior distribution The difference is used to constrain the distribution consistency of the latent variable space, is a dynamically adjusted regularization coefficient used to balance the consistency of data reconstruction and latent variable distribution according to the complexity of regional operation data.
7. The adjustable circuit fault diagnosis method based on variational autoencoder according to claim 5, characterized in that: The step S5 comprises the following steps: S51: The latent variables generated by each region-level variational autoencoder sub-model Collect and form a latent variable set Z; S52: Take the latent variable set Z as the input of the system-level variational autoencoder main model, combined with the regional interaction matrix Train the system-level variational autoencoder main model and optimize the loss function; S53: By calculating the global reconstruction error of the system-level variational autoencoder master model , the model’s ability to extract features of the overall operating state of the circuit system is evaluated, and the reconstruction error is defined as: ； Among them, X(t) and They represent the input and reconstruction results of the global operation data at time t respectively; If the reconstruction error Exceeding the set threshold, adjust the regional interaction matrix Or the hidden variable dimension of the main model of the system-level variational autoencoder to re-optimize the training process; S54: The final output is the optimized system-level variational autoencoder main model, including encoder parameters , decoder parameters And the optimized global latent variable distribution Z.
8. The adjustable circuit fault diagnosis method based on variational autoencoder according to claim 7, characterized in that: In step S52, the loss function is: ； in, represents the reconstruction capability of the system-level variational autoencoder main model for the regionalized operation data X, Represents the latent variable distribution The difference from the prior distribution p(Z) is used to constrain the distribution consistency of the global latent variables, is the regularization coefficient, which is used to adjust the trade-off between global reconstruction ability and latent variable distribution consistency.
9. The adjustable circuit fault diagnosis method based on variational autoencoder according to claim 7, characterized in that: The S9 comprises the following steps: S91: Reconstruction error at the regional level and system-level reconstruction error The corresponding dynamic threshold and In comparison, the dynamic threshold is calculated jointly by regional characteristics and global operating status: ； in, and The adjustment coefficients used for regional and system levels respectively reflect the weight of the sensitivity to abnormalities under specific regional or global operating conditions. is a global weight factor used to balance the influence of the average signal value and the signal fluctuation range. and They represent the standard deviation of the region i and global running data, respectively, and are used to quantify the signal fluctuation range; When the reconstruction error meets one of the following conditions, it is determined to be abnormal: ； S92: For abnormal regions i exceeding the threshold, extract the characteristics of latent variable distribution from the regional variational autoencoder sub-model, analyze the type of abnormal signal, and analyze the and The range of change determines the fault type; S93: Combine the latent variable set Z of the system-level variational autoencoder main model and the regional interaction matrix to optimize the fault location results of the abnormal area and update the dynamic weight of the interaction matrix: ； in, represents the dynamic interaction weight between abnormal region i and other regions j at iteration t+1, The Euclidean distance between the hidden variable characteristics of the abnormal region and other regions is used to measure the degree of association between regions; S94: Generate multi-level fault location and diagnosis results based on regional latent variable characteristics and the optimized interaction matrix.
10. The adjustable circuit fault diagnosis method based on variational autoencoder according to claim 9, characterized in that: In step S94, the diagnosis result includes the identification result of the fault area, classification information of the fault type, the scope of fault propagation and other affected areas and the overall evaluation result of the system-level operating status.