Self-healing type fault diagnosis method and system for intelligent power distribution box

Abstract

The application discloses a self-recovery type fault diagnosis method and system for an intelligent distribution box, realizes closed-loop intelligent operation and maintenance from abnormal sensing, root cause positioning, self-recovery decision and continuous optimization, is real-time synchronized with a digital twin model and a physical distribution box, improves diagnosis accuracy and response speed, introduces causal driving reasoning to replace traditional threshold alarm, significantly improves fault positioning accuracy, and a self-recovery strategy is preformed in a digital twin body to avoid secondary influence of misoperation on the system; and the system has online learning capability and can continuously improve diagnosis and decision performance with the elapse of operation time.

Description

Technical Field

[0001] This invention relates to the field of power distribution box fault detection, and in particular to a self-healing fault diagnosis method and system for intelligent power distribution boxes. Background Technology

[0002] Traditional fault diagnosis of distribution boxes mainly relies on manual inspections or simple threshold alarms, which suffers from problems such as delayed response, high false alarm rates, and inability to pinpoint the root cause. While some existing intelligent diagnostic systems can achieve preliminary anomaly detection, they are mostly in a "sensing-alarm" mode, lacking a deep understanding of the causes of faults, and are unable to autonomously generate and verify repair strategies, making it difficult to achieve true "self-healing." In addition, once the system model is deployed, it is difficult to update, and it cannot adapt to new challenges brought about by equipment aging and environmental changes.

[0003] In summary, a self-healing fault diagnosis method and system for intelligent distribution boxes is needed to address the shortcomings of existing technologies. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a self-healing fault diagnosis method and system for intelligent distribution boxes, aiming to solve the aforementioned problems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a self-healing fault diagnosis method for intelligent distribution boxes, comprising the following steps:

[0006] Step S1: Establish a virtual model. Create a virtual model that corresponds one-to-one with the physical power distribution box in the edge computing unit or cloud, and synchronize sensor data in real time.

[0007] Step S2: Multimodal anomaly perception, real-time monitoring of electrical parameters, capture of hotspot distribution and partial discharge acoustic signatures, and use a deep learning model to perform joint anomaly scoring on multi-source signals;

[0008] Step S3: Cause-and-effect fault tracing. Based on the power distribution topology and physical laws, a structured cause-and-effect graph is constructed. When an anomaly is detected, counterfactual reasoning is performed to determine whether the node is the root fault point.

[0009] Step S4: Self-healing strategy generation. For the identified root fault points, several repair strategies are automatically generated. The effect of the strategies is pre-rendered in the digital twin, the impact on other loops is evaluated, and the optimal strategy is selected for execution.

[0010] Step S5: Continuous learning and model evolution. Feedback the results of each real fault handling to the digital twin, update the causal graph weights and anomaly detection, and perform online system evolution.

[0011] Optionally, step S1 is implemented in the following manner:

[0012] Step A1: Geometric topology modeling. Construct a high-precision 3D model of the distribution box according to the actual dimensions and at the level of detail, and integrate BIM family parameters for topology interaction;

[0013] Step A2: Establish a multiphysics coupling model, assign physical semantic attributes to each component, establish circuit topology, heat conduction model and mechanical stress model, and upload normal data, time-triggered data and infrared thermal data for real-time synchronization with sensor data.

[0014] Optionally, step S2 is implemented in the following manner:

[0015] Step B1: Multi-source heterogeneous data acquisition and preprocessing.

[0016] Electrical signal processing employs sliding window mean filtering and wavelet denoising to extract and standardize features;

[0017] Thermal imaging processing involves non-uniformity correction and background temperature compensation of the original infrared frame, converting it into a thermal map tensor, and extracting the maximum temperature difference in key areas.

[0018] Partial discharge signal processing employs bandpass filtering to remove power frequency interference, extracts the discharge pulse envelope through envelope detection, calculates peak energy and pulse repetition rate frame by frame, and outputs a voiceprint feature sequence.

[0019] Step B2: Timestamp all modal data, use a sliding window to synchronize the buffer in the edge gateway to align the modal samples, and design a dedicated encoder for each modality to extract high-level semantic features;

[0020] Step B3: Construct training data. Normal samples are constructed using the operating data of newly installed or overhauled distribution boxes. Abnormal samples are obtained through accelerated aging experiments or historical fault databases. Random offsets are added to the heatmaps, Gaussian white noise is added to the soundprints, and a multimodal fusion module is constructed.

[0021] Step B4: Reconstruct the original modal input through the decoder, calculate the weighted reconstruction error, and output the anomaly score through the discriminator trained with normal samples. Finally, perform a joint score by the weighted reconstruction error and the anomaly score.

[0022] Optionally, the weighted reconstruction error is calculated in step B4 in the following way:

[0023] ,

[0024] In the formula, The original electrical parameter signal, For the reconstructed electrical parameter signals, This is the raw infrared thermal image data. For the reconstructed infrared thermal image data, These are the original discharge acoustic signature characteristics. For the reconstructed voiceprint features, For electrical mode weights, For thermal imaging mode weights, For acoustic modal weights;

[0025] The output anomaly score is determined in the following way:

[0026] If the output during testing is a confidence score p that indicates a normal distribution, then the anomaly score is: S disc =1-p;

[0027] The joint score is obtained through the following methods:

[0028] ,

[0029] In the formula, S rec For the weighted reconstruction error, S disc For abnormal scoring, For fusion weighting coefficients.

[0030] Optionally, in step S3, the determination of whether the node is a fundamental fault point is made in the following way:

[0031] Step C1: Through high-confidence alarms from the multimodal anomaly perception layer, output anomaly observation nodes and generate anomaly event triggers and candidate variable sets;

[0032] Step C2: Construct a structured causal graph, using each observable or modifiable variable as a graph node and electrical causality, thermal causality, and aging causality as edges, and automatically construct a directed acyclic graph.

[0033] Step C3: Identify the potential root cause candidate set, and use causal graph topology analysis to trace back upstream from the observed anomaly to find all non-observed variables that can affect at least one observed node, and generate the candidate set;

[0034] Step C4: Perform counterfactual reasoning to simulate a corrective intervention for each candidate root cause:

[0035] Using the current real-world environmental parameters as evidence, we run a physical model or probabilistic reasoning to obtain counterfactual predictions. If all anomalies disappear, the candidate root cause is the root cause; if the anomalies still exist, the candidate root cause is excluded.

[0036] Step C5: Root cause confidence quantification and ranking. Calculate the counterfactual elimination rate for each candidate root cause, sort the candidate set in descending order of counterfactual elimination rate, and output the structured root cause analysis results.

[0037] Optionally, the counterfactual elimination rate in step C5 Calculated in the following way:

[0038] In the formula, This represents the true anomaly observation vector. These are the predicted values ​​under counterfactual simulation.

[0039] Optionally, step S4 is implemented in the following manner:

[0040] Step D1: Input the root cause of the fault and the system status. The root cause diagnosis module outputs the root cause analysis results. Input the current full status of the distribution box and the operation and maintenance constraints.

[0041] Step D2: Generate a set of candidate repair strategies. Input the current topology and the faulty node, use the hybrid strategy library of rules and models to generate several feasible solutions, and output topology reconstruction suggestions.

[0042] Step D3: Pre-simulate the evaluation and quantitative scoring of each strategy, perform simulation in the digital twin model for each strategy, and calculate the comprehensive utility score using weighted multi-quasi-component decision-making;

[0043] Step D4: Sort the strategies in descending order of comprehensive utility scores, and apply decision gating rules to select the optimal strategy to execute.

[0044] Optionally, step S5 is implemented in the following manner:

[0045] The multimodal observation data before the failure, the root cause diagnosis results, and the repair strategies and their effects are stored in the time series experience base. Historical normal samples and recent failure samples are randomly sampled from the experience base and mixed with the newly collected samples for training to dynamically optimize the causal graph structure and weights.

[0046] A self-healing fault diagnosis system for intelligent distribution boxes, employing the aforementioned self-healing fault diagnosis method for intelligent distribution boxes, includes a digital twin modeling module, a multimodal perception and anomaly detection module, a causal reasoning and root cause tracing module, a self-healing strategy generation and decision-making module, and a continuous learning and model evolution module.

[0047] The digital twin modeling module is used to build a high-precision virtual model that corresponds one-to-one with the physical distribution box. It integrates multi-physics coupling models of geometric topology, circuit topology, heat conduction, and mechanical stress, and synchronizes normal data, time-triggered data and infrared thermal images from sensors in real time to achieve virtual-real mapping.

[0048] The multimodal perception and anomaly detection module is used to collect and preprocess electrical parameters, infrared thermal imaging and partial discharge acoustic signatures, perform noise reduction, feature extraction and time alignment on each modal data, use a deep learning model to perform multimodal fusion, and output a joint anomaly score.

[0049] The causal reasoning and root cause tracing module is used to construct a directed acyclic graph based on the topology and physical laws of the power distribution system. After an anomaly is triggered, it traces the potential root cause upstream from the observation node. Through counterfactual reasoning simulation, it verifies the root fault point and outputs a root cause candidate list sorted by the counterfactual elimination rate.

[0050] The self-healing strategy generation and decision-making module is used to receive root cause diagnosis results and current system status, generate multiple feasible repair solutions from the hybrid strategy of rule base and AI model, pre-enact the execution effect of each strategy in digital twin, evaluate the impact on other loops, calculate the comprehensive utility score using weighted multi-criteria decision-making, and select the optimal strategy to execute.

[0051] The continuous learning and model evolution module is used to store the full process data of each real failure in the time series experience library, dynamically sample historical normal and abnormal samples, mix them with new data for training, and update the causal graph structure, weights and anomaly detection model parameters online to enable the system to self-evolve.

[0052] The beneficial effects of this invention are:

[0053] 1. In this invention, the method realizes closed-loop intelligent operation and maintenance from anomaly perception, root cause localization, self-healing decision-making and continuous optimization. By synchronizing the digital twin model with the physical distribution box in real time, the accuracy of diagnosis and response speed are improved. Causal-driven reasoning is introduced to replace the traditional threshold alarm, which significantly improves the accuracy of fault location. The self-healing strategy is pre-rehearsed in the digital twin to avoid secondary impacts on the system caused by misoperation. The system has online learning capabilities and can continuously improve its diagnostic and decision-making performance over time.

[0054] 2. In this invention, the weighted reconstruction error and discriminator confidence are fused to take into account both reconstruction bias and distribution anomaly. Modal weights and fusion coefficients are introduced to dynamically adjust the importance of each modality according to the scenario and output a quantitative anomaly score, which facilitates the setting of threshold alarms or graded responses and improves the intelligence level of the system.

[0055] 3. In this invention, the system is designed to facilitate engineering deployment and iterative upgrades through modular design. Each module has a clear function and data flow, supporting an edge-cloud collaborative architecture. The overall system has high real-time performance, high accuracy, and high autonomy, and is suitable for smart grids, industrial power distribution, and other scenarios. It provides a complete technical path for the transformation of power distribution equipment from passive maintenance to proactive self-healing. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of a method flow of the present invention.

[0057] Figure 2 This is a schematic diagram of step S1 of the present invention.

[0058] Figure 3This is a schematic diagram of step S2 of the present invention.

[0059] Figure 4 This is a schematic diagram of step S3 of the present invention.

[0060] Figure 5 This is a schematic diagram of step S4 of the present invention.

[0061] Figure 6 This is a schematic diagram of a system structure according to the present invention. Detailed Implementation

[0062] To more clearly illustrate the technical solutions in the embodiments of the invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] like Figures 1 to 5 As shown, a self-healing fault diagnosis method for intelligent distribution boxes includes the following:

[0064] Step S1: Establish a virtual model. Create a virtual model that corresponds one-to-one with the physical power distribution box in the edge computing unit or cloud, and synchronize sensor data in real time.

[0065] Using 3D software to build a high-precision 3D model of the distribution box:

[0066] Model according to actual product dimensions: circuit breakers, contactors, busbars, cable inlets, explosion-proof enclosures, etc.

[0067] It supports remote viewing of simplified models at various detail levels, and loading of high-level models during fault diagnosis; it integrates BIM family parameters, such as electrical connections and conduit interfaces in Revit, to ensure interactive topology.

[0068] Assign physical semantic attributes to each component, based on the Simscape Electrical or COMSOL interface:

[0069] Circuit topology: Relationship between node voltage and branch current;

[0070] Heat conduction model: Electric current generates heat, temperature rises, and heat diffuses to the outer shell;

[0071] Mechanical stress model, addressing issues such as loose bolts on explosion-proof surfaces.

[0072] Step S2: Multimodal anomaly perception, real-time monitoring of electrical parameters, capture of hotspot distribution and partial discharge acoustic signatures, and use a deep learning model to perform joint anomaly scoring on multi-source signals;

[0073] Implemented in the following ways:

[0074] Step B1: Multi-source heterogeneous data acquisition and preprocessing.

[0075] Electrical signal processing employs sliding window mean filtering and wavelet denoising to extract and standardize features;

[0076] Thermal imaging processing involves non-uniformity correction and background temperature compensation of the original infrared frame, converting it into a thermal map tensor, and extracting the maximum temperature difference in key areas.

[0077] Partial discharge signal processing employs bandpass filtering to remove power frequency interference, extracts the discharge pulse envelope through envelope detection, calculates peak energy and pulse repetition rate frame by frame, and outputs a voiceprint feature sequence.

[0078] Step B2: Timestamp all modal data, use a sliding window to synchronize the buffer in the edge gateway to align the modal samples, and design a dedicated encoder for each modality to extract high-level semantic features;

[0079] Step B3: Construct training data. Normal samples are constructed using the operating data of newly installed or overhauled distribution boxes. Abnormal samples are obtained through accelerated aging experiments or historical fault databases. Random offsets are added to the heatmaps, Gaussian white noise is added to the soundprints, and a multimodal fusion module is constructed.

[0080] Step B4: Reconstruct the original modal input through the decoder, calculate the weighted reconstruction error, and output the anomaly score through the discriminator trained with normal samples. Finally, perform a joint score by the weighted reconstruction error and the anomaly score.

[0081] The weighted reconstruction error is calculated as follows:

[0082] ,

[0083] In the formula, The original electrical parameter signal, For the reconstructed electrical parameter signals, This is the raw infrared thermal image data. For the reconstructed infrared thermal image data, These are the original discharge acoustic signature characteristics. For the reconstructed voiceprint features, For electrical mode weights, For thermal imaging mode weights, For acoustic modal weights;

[0084] The output anomaly score is determined in the following way:

[0085] If the output during testing is a confidence score p that indicates a normal distribution, then the anomaly score is: Sdisc =1-p;

[0086] The joint score is obtained through the following methods:

[0087] ,

[0088] In the formula, S rec For the weighted reconstruction error, S disc For abnormal scoring, For fusion weighting coefficients.

[0089] Step S3: Cause-and-effect fault tracing. Based on the power distribution topology and physical laws, a structured cause-and-effect graph is constructed. When an anomaly is detected, counterfactual reasoning is performed to determine whether the node is the root fault point.

[0090] Exception event triggering and candidate variable set generation

[0091] Input: High-confidence alarms from the multimodal anomaly perception layer, such as Sanomaly>τ.

[0092] Output: A set of anomaly observation nodes O = {O1, O2, ..., O...} k},For example:

[0093] O1: "Busbar A phase temperature = 92°C (out of limit)"

[0094] O2: "Sudden increase in acoustic signature energy of circuit breaker QF3"

[0095] O3: "Circuit current harmonic distortion rate = 18%".

[0096] Constructing a structured causal graph, automatically building a directed acyclic graph based on power distribution topology and physical mechanisms:

[0097] Each observable / manipulated variable is represented as a node in the graph, including: device state nodes and physical quantity nodes. Edges represent causal relationships, including electrical causality, thermal causality, and aging causality.

[0098] Identify a potential root cause candidate set and use causal graph topology analysis to trace upstream from observed anomaly O:

[0099] Identify all unobservable variables that can affect at least one Oi ∈ O, i.e., latent variables that cannot be directly measured but can be assumed.

[0100] Candidate set C = {c1, c2, ..., cm}, for example:

[0101] c1 = Loose terminals; c2 = Mechanical aging of circuit breaker; c3 = Deterioration of cable insulation.

[0102] Perform counterfactual reasoning to simulate a "remedial" intervention in the digital twin for each candidate root cause cj ∈ C:

[0103] The counterfactual question format is: "If cj is in a normal state, will the current observation O still occur?"

[0104] Specific operations:

[0105] Fixed background conditions: using current real-world environmental parameters as evidence for identification;

[0106] Apply intervention: Force cj = normal in the SCM or digital twin simulation model, such as setting the contact resistance to 0.1mΩ;

[0107] Forward simulation: Run the physical model (electro-thermal coupling simulation) or probabilistic inference (BN inference) to obtain counterfactual predictions;

[0108] Comparison results: If all anomalies disappear, such as when the temperature drops to 60°C and the voiceprint energy returns to zero, then cj is the root cause; if the anomalies still exist, then they are excluded.

[0109] Calculate the counterfactual elimination rate for each candidate root cause, sort the candidate set in descending order of counterfactual elimination rate, and output the structured root cause analysis results.

[0110] Counterfact elimination rate Calculated in the following way:

[0111] In the formula, This represents the true anomaly observation vector. These are the predicted values ​​under counterfactual simulation.

[0112] If CER≈1.0, the intervention almost completely eliminates the abnormality, then it is a high-confidence root cause; if CER≈0.0, the intervention is ineffective, then it is a non-root cause.

[0113] Step S4: Self-healing strategy generation. For the identified root fault points, several repair strategies are automatically generated. The effect of the strategies is pre-rendered in the digital twin, the impact on other loops is evaluated, and the optimal strategy is selected for execution.

[0114] Implemented in the following ways:

[0115] Step D1: Input the root cause of the fault and the system status. The root cause diagnosis module outputs the root cause analysis results. Input the current full status of the distribution box and the operation and maintenance constraints.

[0116] Step D2: Generate a set of candidate repair strategies. Input the current topology and the faulty node, use the hybrid strategy library of rules and models to generate several feasible solutions, and output topology reconstruction suggestions.

[0117] Step D3: Pre-simulate the evaluation and quantitative scoring of each strategy, perform simulation in the digital twin model for each strategy, and calculate the comprehensive utility score using weighted multi-quasi-component decision-making;

[0118] Step D4: Sort the strategies in descending order of comprehensive utility scores, and apply decision gating rules to select the optimal strategy to execute.

[0119] Step S5: Continuous learning and model evolution, feeding back the results of each real fault handling to the digital twin, updating the causal graph weights and anomaly detection, and performing online system evolution;

[0120] The multimodal observation data before the failure, the root cause diagnosis results, and the repair strategies and their effects are stored in the time series experience base. Historical normal samples and recent failure samples are randomly sampled from the experience base and mixed with the newly collected samples for training to dynamically optimize the causal graph structure and weights.

[0121] The method of this invention realizes closed-loop intelligent operation and maintenance from anomaly perception, root cause localization, self-healing decision-making and continuous optimization. By synchronizing the digital twin model with the physical distribution box in real time, it improves the accuracy of diagnosis and response speed. It introduces causal-driven reasoning to replace the traditional threshold alarm, which significantly improves the accuracy of fault location. The self-healing strategy is pre-rehearsed in the digital twin to avoid secondary impacts on the system caused by misoperation. The system has online learning capabilities and can continuously improve its diagnostic and decision-making performance over time.

[0122] The system integrates two indicators: weighted reconstruction error and discriminator confidence, taking into account both reconstruction bias and distribution anomalies. It introduces modal weights and fusion coefficients, which can dynamically adjust the importance of each modality according to the scenario and output quantitative anomaly scores. This facilitates the setting of threshold alarms or graded responses, thereby improving the system's intelligence level.

[0123] like Figure 6 As shown, a self-healing fault diagnosis system for intelligent distribution boxes adopts the self-healing fault diagnosis method for intelligent distribution boxes, including a digital twin modeling module, a multimodal perception and anomaly detection module, a causal reasoning and root cause tracing module, a self-healing strategy generation and decision-making module, and a continuous learning and model evolution module.

[0124] The digital twin modeling module is used to build a high-precision virtual model that corresponds one-to-one with the physical distribution box. It integrates multi-physics coupling models of geometric topology, circuit topology, heat conduction, and mechanical stress, and synchronizes normal data, time-triggered data and infrared thermal images from sensors in real time to achieve virtual-real mapping.

[0125] The multimodal perception and anomaly detection module is used to collect and preprocess electrical parameters, infrared thermal imaging and partial discharge acoustic signatures, perform noise reduction, feature extraction and time alignment on each modal data, use a deep learning model to perform multimodal fusion, and output a joint anomaly score.

[0126] The causal reasoning and root cause tracing module is used to construct a directed acyclic graph based on the topology and physical laws of the power distribution system. After an anomaly is triggered, it traces the potential root cause upstream from the observation node. Through counterfactual reasoning simulation, it verifies the root fault point and outputs a root cause candidate list sorted by the counterfactual elimination rate.

[0127] The self-healing strategy generation and decision-making module is used to receive root cause diagnosis results and current system status, generate multiple feasible repair solutions from the hybrid strategy of rule base and AI model, pre-enact the execution effect of each strategy in digital twin, evaluate the impact on other loops, calculate the comprehensive utility score using weighted multi-criteria decision-making, and select the optimal strategy to execute.

[0128] The continuous learning and model evolution module is used to store the full process data of each real failure in the time series experience library, dynamically sample historical normal and abnormal samples, mix them with new data for training, and update the causal graph structure, weights and anomaly detection model parameters online to enable the system to self-evolve.

[0129] The system's modular design facilitates engineering deployment and iterative upgrades. Each module has clearly defined functions and a clear data flow, supporting an edge-cloud collaborative architecture. The overall system boasts high real-time performance, high accuracy, and high autonomy, making it suitable for smart grids, industrial power distribution, and other scenarios. It provides a complete technical path for the transformation of power distribution equipment from passive maintenance to proactive self-healing.

[0130] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions or improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A self-healing fault diagnosis method for intelligent distribution boxes, characterized in that, Includes the following steps: Step S1: Establish a virtual model. Create a virtual model that corresponds one-to-one with the physical power distribution box in the edge computing unit or cloud, and synchronize sensor data in real time. Step S2: Multimodal anomaly perception, real-time monitoring of electrical parameters, capture of hotspot distribution and partial discharge acoustic signatures, and use a deep learning model to perform joint anomaly scoring on multi-source signals; Step S3: Cause-and-effect fault tracing. Based on the power distribution topology and physical laws, a structured cause-and-effect graph is constructed. When an anomaly is detected, counterfactual reasoning is performed to determine whether the node is the root fault point. Step S4: Self-healing strategy generation. For the identified root fault points, several repair strategies are automatically generated. The effect of the strategies is pre-rendered in the digital twin, the impact on other loops is evaluated, and the optimal strategy is selected for execution. Step S5: Continuous learning and model evolution. Feedback the results of each real fault handling to the digital twin, update the causal graph weights and anomaly detection, and perform online system evolution.

2. The self-healing fault diagnosis method for intelligent distribution boxes according to claim 1, characterized in that, Step S1 is implemented in the following manner: Step A1: Geometric topology modeling. Construct a high-precision 3D model of the distribution box according to the actual dimensions and at the level of detail, and integrate BIM family parameters for topology interaction; Step A2: Establish a multiphysics coupling model, assign physical semantic attributes to each component, establish circuit topology, heat conduction model and mechanical stress model, and upload normal data, time-triggered data and infrared thermal data for real-time synchronization with sensor data.

3. The self-healing fault diagnosis method for intelligent distribution boxes according to claim 1, characterized in that, Step S2 is implemented in the following manner: Step B1: Multi-source heterogeneous data acquisition and preprocessing. Electrical signal processing employs sliding window mean filtering and wavelet denoising to extract and standardize features; Thermal imaging processing involves non-uniformity correction and background temperature compensation of the original infrared frame, converting it into a thermal map tensor, and extracting the maximum temperature difference in key areas. Partial discharge signal processing employs bandpass filtering to remove power frequency interference, extracts the discharge pulse envelope through envelope detection, calculates peak energy and pulse repetition rate frame by frame, and outputs a voiceprint feature sequence. Step B2: Timestamp all modal data, use a sliding window to synchronize the buffer in the edge gateway to align the modal samples, and design a dedicated encoder for each modality to extract high-level semantic features; Step B3: Construct training data. Normal samples are constructed using the operating data of newly installed or overhauled distribution boxes. Abnormal samples are obtained through accelerated aging experiments or historical fault databases. Random offsets are added to the heatmaps, Gaussian white noise is added to the soundprints, and a multimodal fusion module is constructed. Step B4: Reconstruct the original modal input through the decoder, calculate the weighted reconstruction error, and output the anomaly score through the discriminator trained with normal samples. Finally, perform a joint score by the weighted reconstruction error and the anomaly score.

4. The self-healing fault diagnosis method for intelligent distribution boxes according to claim 3, characterized in that, In step B4, the weighted reconstruction error is calculated in the following way: ， In the formula, The original electrical parameter signal, For the reconstructed electrical parameter signals, This is the raw infrared thermal image data. For the reconstructed infrared thermal image data, These are the original discharge acoustic signature characteristics. For the reconstructed voiceprint features, For electrical mode weights, For thermal imaging mode weights, For acoustic modal weights; The output anomaly score is determined in the following way: If the output during testing is a confidence score p that indicates a normal distribution, then the anomaly score is: S disc =1-p; The joint score is obtained through the following methods: ， In the formula, S rec For the weighted reconstruction error, S disc For abnormal scoring, For fusion weighting coefficients.

5. The self-healing fault diagnosis method for intelligent distribution boxes according to claim 1, characterized in that, In step S3, the determination of whether the node is a fundamental fault point is made in the following way: Step C1: Through high-confidence alarms from the multimodal anomaly perception layer, output anomaly observation nodes and generate anomaly event triggers and candidate variable sets; Step C2: Construct a structured causal graph, using each observable or modifiable variable as a graph node and electrical causality, thermal causality, and aging causality as edges, and automatically construct a directed acyclic graph. Step C3: Identify the potential root cause candidate set, and use causal graph topology analysis to trace back upstream from the observed anomaly to find all non-observed variables that can affect at least one observed node; Step C4: Perform counterfactual reasoning to simulate a corrective intervention for each candidate root cause: Using the current real-world environmental parameters as evidence, we run a physical model or probabilistic reasoning to obtain counterfactual predictions. If all anomalies disappear, the candidate root cause is the root cause; if the anomalies still exist, the candidate root cause is excluded. Step C5: Root cause confidence quantification and ranking. Calculate the counterfactual elimination rate for each candidate root cause, sort the candidate set in descending order of counterfactual elimination rate, and output the structured root cause analysis results.

6. The self-healing fault diagnosis method for intelligent distribution boxes according to claim 5, characterized in that, Counterfact elimination rate in step C5 Calculated in the following way: In the formula, This represents the true anomaly observation vector. These are the predicted values ​​under counterfactual simulation.

7. The self-healing fault diagnosis method for intelligent distribution boxes according to claim 1, characterized in that, Step S4 is implemented in the following manner: Step D1: Input the root cause of the fault and the system status. The root cause diagnosis module will output the root cause analysis results. Input the current full status of the distribution box and the operation and maintenance constraints. Step D2: Generate a set of candidate repair strategies. Input the current topology and the faulty node, use the hybrid strategy library of rules and models to generate several feasible solutions, and output topology reconstruction suggestions. Step D3: Pre-simulate the evaluation and quantitative scoring of each strategy, perform simulation in the digital twin model for each strategy, and calculate the comprehensive utility score using weighted multi-quasi-component decision-making; Step D4: Sort the strategies in descending order of comprehensive utility scores, and apply decision gating rules to select the optimal strategy to execute.

8. The self-healing fault diagnosis method for intelligent distribution boxes according to claim 1, characterized in that, Step S5 is implemented in the following manner: The multimodal observation data before the failure, the root cause diagnosis results, and the repair strategies and their effects are stored in the time series experience base. Historical normal samples and recent failure samples are randomly sampled from the experience base and mixed with the newly collected samples for training to dynamically optimize the causal graph structure and weights.

9. A self-healing fault diagnosis system for intelligent distribution boxes, employing the self-healing fault diagnosis method for intelligent distribution boxes as described in any one of claims 1-8, characterized in that, It includes modules for digital twin modeling, multimodal perception and anomaly detection, causal reasoning and root cause analysis, self-healing strategy generation and decision-making, and continuous learning and model evolution. The digital twin modeling module is used to build a high-precision virtual model that corresponds one-to-one with the physical distribution box. It integrates multi-physics coupling models of geometric topology, circuit topology, heat conduction, and mechanical stress, and synchronizes normal data, time-triggered data and infrared thermal images from sensors in real time to achieve virtual-real mapping. The multimodal perception and anomaly detection module is used to collect and preprocess electrical parameters, infrared thermal imaging and partial discharge acoustic signatures, perform noise reduction, feature extraction and time alignment on each modal data, use a deep learning model to perform multimodal fusion, and output a joint anomaly score. The causal reasoning and root cause tracing module is used to construct a directed acyclic graph based on the topology and physical laws of the power distribution system. After an anomaly is triggered, it traces the potential root cause upstream from the observation node. Through counterfactual reasoning simulation, it verifies the root fault point and outputs a root cause candidate list sorted by the counterfactual elimination rate. The self-healing strategy generation and decision-making module is used to receive root cause diagnosis results and current system status, generate multiple feasible repair solutions from the hybrid strategy of rule base and AI model, pre-enact the execution effect of each strategy in digital twin, evaluate the impact on other loops, calculate the comprehensive utility score using weighted multi-criteria decision-making, and select the optimal strategy to execute. The continuous learning and model evolution module is used to store the full process data of each real failure in the time series experience library, dynamically sample historical normal and abnormal samples, mix them with new data for training, and update the causal graph structure, weights and anomaly detection model parameters online to enable the system to self-evolve.

Images