Power cabinet fault prediction self-healing controller and deep learning method

Abstract

The invention discloses a power cabinet fault prediction self-healing controller and a deep learning method, particularly relates to the field of power distribution cabinet monitoring, and is used for solving the problems of misjudgment and maloperation caused by multi-mode signal asynchronization. The maintenance of the coupling relationship among the current, partial discharge and vibration signals in the fault sample expansion stage is enhanced, and the generated sample is prevented from deviating from the real synchronization feature; in the online reasoning stage, a sliding window-based probability fluctuation resolution amplitude and confidence imbalance metric value extraction strategy is introduced, a synchronous robustness coefficient is comprehensively generated and serves as a triggering basis of adaptive resampling and bypass switching, and a linkage closed structure from sample generation to prediction verification is formed; according to the invention, the dynamic coping capability for boundary drift and synchronization mismatch is realized, and the fault identification precision and the motion stability of the controller in a complex disturbance scene are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power distribution cabinet monitoring, more particularly, to a power cabinet fault prediction self-healing controller and a deep learning method. BACKGROUND

[0002] Under the compound working condition of photovoltaic intermittent grid connection and random parallel connection of charging piles, the outdoor power distribution cabinet needs a fault prediction self-healing controller to aggregate multi-mode transient signals such as current, partial discharge and vibration; in the offline training stage, a generative adversarial network is introduced to expand samples to alleviate the shortage of fault waveforms, but the generator only overwrites the single-channel current track, and does not reconstruct the subtle phase coupling and energy complementary relationship between channels. This unbalanced supplement makes the deep model form an idealized same-frequency feature boundary during training, and when encountering nanosecond-level asynchronous disturbances in the mutual inductor current and partial discharge channels during online operation, the healthy waveform is identified as a high-confidence fault and triggers the solid-state bypass action. Mis-isolation causes the bus to be unloaded instantaneously, generating a reverse voltage spike, and the contact arc quickly erodes the copper bar. Since the log only traces the single-channel prediction results, the mismatch logic is difficult to expose in routine playback, reflecting that there is a coupling blind area between the data supplement strategy and the multi-mode synchronous monitoring link, which is difficult to meet the accuracy requirements of power cabinet fault prediction and self-healing control.

[0003] To solve the above problems, a technical solution is provided. SUMMARY

[0004] To overcome the above-mentioned defects of the prior art, embodiments of the present application provide a power cabinet fault prediction self-healing controller and a deep learning method. A phase consistency constraint mechanism is constructed through multi-channel fusion caching, which strengthens the maintenance of the coupling relationship between current, partial discharge and vibration signals during the fault sample expansion stage, avoiding the deviation of generated samples from the true synchronous features. In the online inference stage, a probability fluctuation resolution amplitude and confidence imbalance metric value extraction strategy based on a sliding window is introduced, and a synchronous robustness coefficient is synthesized as the trigger basis for adaptive resampling and bypass switching, forming a linked closed structure from sample generation to prediction verification. The dynamic response capability to boundary drift and synchronization mismatch is realized, effectively improving the fault identification accuracy and controller action stability in complex disturbance scenarios, to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purposes, the present application provides the following technical solutions: S1: A time synchronization window reads multi-mode sensing streams and writes them into a fusion cache after adding a phase identifier, which is used to lock the transient alignment starting point of the current-partial discharge-vibration three channels; S2: Calculate the cross-channel coupling graph according to the fusion cache and extract the phase consistency vector, which is used as a generation constraint descriptor and updated with the frame stream; S3: Call the coupled constraint-based generative adversarial network to expand the synchronous fault sample based on the descriptor, and then write the qualified samples that pass the phase consistency threshold to the enhanced library; S4: Train the dilated convolution-attention network with the enhanced library and the original library, and freeze the coupled graph weight as the inference benchmark to stabilize the decision boundary; S5: The edge end checks each frame of fault prediction probability online with the inference benchmark as the ruler, extracts the fault prediction probability fluctuation resolution and confidence imbalance metric value through a double-layer sliding window, and generates a self-healing trigger threshold coefficient, which immediately adapts to re-sampling and writing the enhanced library and driving bypass switching once it leaves the safe interval.

[0006] In a preferred embodiment, step S1 includes the following: The power cabinet fault prediction self-healing controller reads the current signal, partial discharge signal and vibration signal through the time synchronization window, writes them into the fusion cache after adding the phase identifier, locks the transient alignment starting point of the three signals, sets the time synchronization window, extracts synchronized signal segments from the current sensor, partial discharge sensor and vibration sensor, applies Hilbert transform to each signal segment to calculate the phase identifier, integrates the signal segment and the phase identifier into a record, stores it in the fusion cache in time sequence, and realizes the alignment of the signals in time and phase through the time synchronization window and the phase identifier, locking the transient alignment starting point.

[0007] In a preferred embodiment, step S2 includes the following: Read the records in time sequence from the fusion cache, calculate the mutual information between the current signal and the partial discharge signal, the mutual information between the current signal and the vibration signal, and the mutual information between the partial discharge signal and the vibration signal, and use them as the edge weights of the cross-channel coupling graph, calculate the phase difference based on the phase identifier and normalize it to form a phase consistency vector, and combine the mutual information and the phase consistency vector into a generation constraint descriptor, which is dynamically updated with the time synchronization window.

[0008] In a preferred embodiment, step S3 includes the following: The coupled constraint-based generative adversarial network is used to expand the synchronous fault sample, and the generator receives random noise and the generation constraint descriptor to generate a multi-channel fault sample containing the current signal, the partial discharge signal and the vibration signal.

[0009] In a preferred embodiment, step S3 further includes the following: The generation constraint descriptor is composed of the cross-channel coupling graph and the phase consistency vector, and the coupled constraint loss and the phase consistency loss are used to optimize the generation process to ensure that the coupling relationship and phase characteristics of the generated sample are consistent with the real sample.

[0010] In a preferred embodiment, step S3 further comprises the following content: The discriminator evaluates the difference between the generated sample and the real sample to optimize the network parameters, calculates the phase consistency vector of the generated sample, and compares the Euclidean distance with the real sample to screen qualified samples and write them into the enhanced library to improve the sample quality.

[0011] In a preferred embodiment, step S4 comprises the following content: The original library and the enhanced library are merged to form a joint training set, which contains real collected current signals, partial discharge signals and vibration signals and their fault labels, as well as the expanded synchronous fault samples generated by the generative adversarial network; the network architecture is designed to consist of dilated convolution layers, attention mechanisms and coupled graph embedding layers, the dilated convolution layers extract long-range temporal features, the attention mechanisms enhance the attention to key coupling features, and the coupled graph embedding layers embed the cross-channel coupling graph into the network to guide feature fusion.

[0012] In a preferred embodiment, step S4 comprises the following content: During training, the cross-entropy loss is minimized, the weights of the coupled graph embedding layer are frozen, and only the parameters of the dilated convolution layer, the attention mechanism and the classification layer are updated to ensure that the model follows the real physical relationship between signals, and the decision boundary is stabilized by freezing the weights of the coupled graph embedding layer.

[0013] In a preferred embodiment, step S4 comprises the following content: The double-layer sliding window structure is composed of an inner sliding window and an outer sliding window, the inner sliding window calculates the instantaneous difference of fault prediction probability to capture short-term fluctuations, the outer sliding window accumulates the peak intensity of the difference to evaluate long-term trends, the fluctuation resolution amplitude is calculated by piecewise integration to reflect the fluctuation activity of fault prediction probability, and the confidence imbalance metric value is calculated by logarithmic entropy normalization of the confidence contrast array to reflect the imbalance degree of the confidence distribution.

[0014] In a preferred embodiment, step S5 comprises the following content: The self-healing trigger threshold coefficient is calculated by the minimum circumscribed ellipse method to calculate the vector module length from the joint scatter density center of the fluctuation resolution amplitude and the confidence imbalance metric value to the origin, and when the self-healing trigger threshold coefficient exceeds the safety threshold, the adaptive resampling is performed to write the enhanced library and drive the bypass switch.

[0015] The technical effects and advantages of the power cabinet fault prediction self-healing controller and the deep learning method of the application are as follows: This invention achieves collaborative modeling of current, partial discharge, and vibration channels during fault sample expansion by constructing a multi-channel synchronization constraint mechanism driven by fusion cache and a generation control strategy based on phase consistency. This effectively avoids the misjudgment boundary misalignment problem caused by single-channel training. In the online phase, the fault prediction probability output by inference is further incorporated into a sliding window dynamic monitoring framework. The fluctuation resolution amplitude reveals the sensitive amplitude of the model response, and the confidence imbalance metric reflects the prediction skewness risk. These two parameters work together to generate a synchronization robustness coefficient, serving as the core basis for triggering adaptive resampling and bypass switching. This constructs a complete prediction self-healing pathway driven by coupled sensing, with closed-loop linkage between generation enhancement and online verification. This not only improves the resistance of anomaly identification to transient disturbances but also enables the model to adapt to boundary drift and data heterogeneity in real-time deployment, significantly enhancing the discrimination stability and self-healing execution accuracy under complex operating conditions. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the power cabinet fault prediction and self-healing controller and deep learning method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: Figure 1 The present invention provides a power cabinet fault prediction and self-healing controller and a deep learning method, including: S1: The time synchronization window reads the multimode sensor stream, adds a phase identifier, and writes it to the fusion buffer to lock the transient alignment start point of the current-partial discharge-vibration three channels.

[0019] S2: Calculate the cross-channel coupling graph based on the fusion cache and extract the phase consistency vector. The vector is used as the generation constraint descriptor and updated with the frame stream.

[0020] S3: Invoke the adversarial generative network with coupling constraints to expand the synchronization fault samples based on the descriptor. After generation, qualified samples are selected according to the phase consistency threshold and written into the enhancement library.

[0021] S4: Train an extended convolutional-attention network jointly with the augmented and original libraries, and freeze the coupling graph weights as inference benchmarks to stabilize the decision boundary.

[0022] S5: The edge end checks the fault prediction probability of each frame online with the inference benchmark as the ruler, extracts the fluctuation resolution amplitude of the fault prediction probability and the confidence imbalance metric value through a double-layer sliding window, and generates a self-healing trigger threshold coefficient. Once it deviates from the safe interval, it immediately adapts to the enhanced library and drives the bypass switch.

[0023] As the core equipment of power distribution and protection in the power system, power cabinets need to cope with the complex working conditions of intermittent photovoltaic grid connection and random parallel charging piles in outdoor environments. In this scenario, power cabinets face the challenges of severe load fluctuations and complex electrical environments, which may lead to power outages, equipment damage, and even safety accidents due to overload, short circuit, or insulation failure. Traditional fault detection relies on single signal (such as current or voltage) monitoring, which is difficult to adapt to the complexity of multi-modal transient signals, and the scarcity of fault data limits the training effect of deep learning models. Therefore, a method based on power cabinet fault prediction self-healing controller and deep learning is proposed, which aggregates multi-modal transient signals such as current, partial discharge (partial discharge), and vibration, and uses deep learning technology to realize fault prediction and self-healing control.

[0024] This method improves fault prediction accuracy through time synchronization and coupling analysis of multi-modal signals, and triggers self-healing actions in real time on the edge. Its operation includes time-synchronized signal acquisition, cross-channel coupling feature extraction, generative adversarial network expansion of fault samples, deep learning model training, and edge real-time decision-making. Compared with traditional methods, this scheme not only can predict potential faults, but also can take self-healing measures such as load adjustment or fault isolation before the fault occurs, thereby improving power supply reliability and equipment safety.

[0025] In the power cabinet fault prediction self-healing controller and deep learning method, step S1 is the basic link of multi-modal signal processing. In view of the complex transient characteristics of outdoor power cabinets under the combined working conditions of photovoltaic and charging piles, step S1 ensures the alignment of current, partial discharge, and vibration signals in time and phase through time synchronization window and phase identification, providing accurate input data for subsequent step S2 to calculate cross-channel coupling graph and extract phase consistency vector.

[0026] When outdoor power distribution cabinets face the complex working conditions of photovoltaic intermittent grid connection and random parallel charging piles, the load fluctuates dramatically and the electrical environment is complex, and the fault features are often hidden in the subtle coupling of multi-modal signals. The reason for collecting current signals, partial discharge signals and vibration signals is that they can fully represent the running state of the power distribution cabinet: current signals directly reflect load changes and electrical faults such as short circuits or overloads; partial discharge signals are sensitive to insulation degradation and potential breakdown risk, and can provide early warning of insulation faults; vibration signals capture mechanical looseness or structural abnormalities to avoid failures caused by physical damage. The combination of the three creates a multi-dimensional feature space that covers multiple main fault modes such as electrical, insulation and mechanical. Compared with other signals, they have higher sensitivity and specificity, significantly improving the accuracy and robustness of fault prediction. In addition, these three types of signal acquisition technology is mature, the sensor cost is low and the deployment difficulty is small, which is suitable for long-term stable operation in outdoor environment. Other signals such as temperature or sound, although have certain reference value, have lower sensitivity or are easily disturbed by the environment, and have less complementarity and practicality than the above three. Therefore, based on the importance of fault diagnosis, the complementarity between signals and engineering practicality, it is a reasonable and efficient choice to prioritize collecting current signals, partial discharge signals and vibration signals.

[0027] S1.1, multi-modal sensor stream reading: The power cabinet fault prediction self-healing controller first performs a multi-modal sensor stream reading step.

[0028] The current signal, partial discharge signal and vibration signal of the power cabinet are collected in real time by the current sensor, partial discharge sensor and vibration sensor respectively. These signals are continuous time series data, which reflect the electrical and mechanical states of the power cabinet under different operating conditions. In order to ensure the accuracy of subsequent analysis, the controller extracts synchronized signal segments from the current sensor, partial discharge sensor and vibration sensor channels. The extraction process requires that the signal segments be aligned in time to provide consistent data input for subsequent signal coupling analysis and fault prediction.

[0029] S1.2, time synchronization window setting: After the multi-modal sensor stream reading is completed, the time synchronization window setting step is entered.

[0030] The controller sets a fixed time period as the time synchronization window, for example, the time length is 100 milliseconds. In this time synchronization window, the same number of sampling points are extracted from the current signal, partial discharge signal and vibration signal. Assuming the sampling frequency is 10000 times per second, each time synchronization window contains 1000 sampling points. In this way, the controller ensures the time consistency of the current signal, partial discharge signal and vibration signal in each time synchronization window. The setting of the time synchronization window aims to provide a unified time reference for subsequent phase analysis and signal coupling calculation.

[0031] S1.3, Phase identification calculation: After completing the time synchronization window setup, the phase identification calculation step is performed.

[0032] For each current signal segment, partial discharge signal segment, and vibration signal segment within a time synchronization window, the respective phase identification is calculated to represent the phase relationship between the signals. The calculation process first applies the Hilbert transform to each signal segment to generate the corresponding complex signal. Then, the instantaneous phase value of each sampling point is extracted from the complex signal. Finally, the average of all instantaneous phase values is calculated to obtain the phase identification of the signal segment. The phase identification reflects the overall phase characteristics of the signal segment and is used to analyze the phase consistency between the current signal, partial discharge signal, and vibration signal.

[0033] S1.4, Write to fusion cache: After the phase identification calculation is completed, the write to fusion cache step is performed.

[0034] The controller integrates the signal data and phase identification within each time synchronization window into a record and stores it in the fusion cache in chronological order. Each record contains the start timestamp of the time synchronization window, the current signal segment, the partial discharge signal segment, the vibration signal segment, and the corresponding phase identification. Through this structured storage method, the fusion cache retains the original current signal, partial discharge signal, and vibration signal while adding phase information. This data organization method facilitates subsequent steps to retrieve and process records in chronological order, providing complete data support for cross-channel signal coupling analysis.

[0035] S1.5, Lock transient alignment starting point: After completing the write to fusion cache, the lock transient alignment starting point step is performed.

[0036] Using the time synchronization window and phase identification, the current signal, partial discharge signal, and vibration signal are aligned in time and phase. The time synchronization window ensures the synchronization of the three signals in time, while the phase identification provides a reference standard for the phase relationship between the signals. By aligning time and phase, the controller locks the transient alignment starting point of the three signals. This alignment method can accurately capture the specific phase and amplitude relationship between the signals, thereby improving the accuracy and reliability of fault prediction.

[0037] As current signals, partial discharge signals, and vibration signals are key data reflecting the state of power cabinets, they need to be consistent in time and phase to accurately analyze the coupling relationship between signals. High-precision synchronization and alignment of multi-modal signals are achieved through time synchronization window and phase identification. This method significantly improves the accuracy of signal analysis and provides high-quality input data for deep learning models. At the same time, the integration of phase information enhances the ability to capture subtle relationships between signals, thereby improving the robustness and adaptability of fault prediction. Ultimately, the application of this technology logic reduces the risk of power cabinet misoperation and operation and maintenance costs, and improves the reliability and safety of the power system.

[0038] Step S1 achieves transient alignment of multi-modal signals through time synchronization window and stores them in fusion cache after adding phase identification, laying a data foundation for subsequent processing. However, relying solely on synchronization signals cannot fully represent the phase coupling and energy complementary characteristics between signals, especially under nanosecond-level non-synchronous disturbances. Single-channel sample generation can easily lead to model misjudgment. Therefore, step S2 focuses on extracting cross-channel coupling graphs and phase consistency vectors from the fusion cache to quantify these relationships and provide constraints for fault sample generation.

[0039] Step S2 generates cross-channel coupling graphs and phase consistency vectors through mutual information calculation and phase difference extraction, successfully representing the coupling characteristics of current, partial discharge, and vibration signals under complex working conditions. Dynamic updating of generated constraint descriptors provides precise constraints for fault sample generation in step S3, improving sample quality and effectively alleviating the problem of model misjudgment caused by single-channel generation.

[0040] S2.1, read fusion cache data: Read records in time order from the fusion cache, each record containing the start timestamp of the time synchronization window, the sampling sequence of the current signal, the phase identification of the current signal, the sampling sequence of the partial discharge signal, the phase identification of the partial discharge signal, the sampling sequence of the vibration signal, and the phase identification of the vibration signal. The reading process ensures that all signal sampling sequences and their corresponding phase identifications are arranged in the order of time synchronization windows. Reading records in time order ensures the consistency of signal sampling sequences and phase identifications in time sequence, providing accurate data sources for subsequent calculation of cross-channel coupling graphs and extraction of phase consistency vectors.

[0041] S2.2, calculate cross-channel coupling graph: The mutual information method is used to quantify the dependency between the current signal and the partial discharge signal, the current signal and the vibration signal, and the partial discharge signal and the vibration signal. The specific process is as follows: first, the mutual information between the current signal sampling sequence and the partial discharge signal sampling sequence is calculated. The calculation method is based on the joint probability distribution of the two signal sampling sequences and the marginal probability distribution of each signal sampling sequence. The joint probability distribution is estimated by counting the discrete values of the two signal sampling sequences. Then the marginal probability distribution of each signal sampling sequence is calculated, and the dependency between the two signals is determined. Then, the mutual information between the current signal sampling sequence and the vibration signal sampling sequence, and the mutual information between the partial discharge signal sampling sequence and the vibration signal sampling sequence are calculated in the same way. The three mutual information values calculated are used as the weights of the corresponding edges in the cross-channel coupling graph to form a fully connected graph containing three nodes, which correspond to the current signal, the partial discharge signal and the vibration signal respectively.

[0042] The mutual information method can capture the nonlinear dependency between signals, and is more suitable for representing the complex signal characteristics of power cabinets under complex working conditions than traditional linear correlation analysis. The construction of the cross-channel coupling graph can intuitively reflect the coupling strength between multi-modal signals, providing important constraint information for subsequent fault sample generation, and ensuring that the generated samples meet the true signal relationship.

[0043] S2.3, extract the phase consistency vector: Based on the phase identifier of the current signal, the phase identifier of the partial discharge signal and the phase identifier of the vibration signal, the phase difference between each pair of signals is calculated. The specific process is as follows: first, the absolute difference between the current signal phase identifier and the partial discharge signal phase identifier is calculated to obtain the phase difference between the current signal and the partial discharge signal; then the absolute difference between the current signal phase identifier and the vibration signal phase identifier is calculated to obtain the phase difference between the current signal and the vibration signal; then the absolute difference between the partial discharge signal phase identifier and the vibration signal phase identifier is calculated to obtain the phase difference between the partial discharge signal and the vibration signal. The above three phase differences are divided by the maximum possible value of the phase difference, i.e. the circumference, to complete the normalization process, so that the value of each phase difference falls within the interval of 0 to 1. The three normalized phase differences are arranged in the order of current signal and partial discharge signal, current signal and vibration signal, and partial discharge signal and vibration signal to form a phase consistency vector.

[0044] The phase consistency vector can quantify the consistency degree of the signals in the phase, which is of great significance for power cabinet fault prediction, because the fault features are often closely related to the phase relationship between signals. Extracting the phase consistency vector provides phase constraints for fault sample generation, ensuring that the generated samples are consistent with the true signals in terms of phase characteristics, thereby improving the accuracy of fault prediction.

[0045] S2.4, update the generation constraint descriptor: The edge weights of the cross-channel coupling graph, i.e., the three calculated mutual information values, and the phase consistency vector, i.e., the three normalized phase differences, are combined to generate a constraint descriptor. The specific process is to arrange the three mutual information values and the three normalized phase differences in a predetermined order to form a descriptor containing six items of data, which is dynamically updated with the frame stream of the time synchronization window, and the mutual information and phase difference are recalculated based on the signal data in the current time synchronization window each time. The generated constraint descriptor integrates the coupling strength and phase consistency information between signals, providing comprehensive constraints for the generation of fault samples in step S3. Dynamically updating the generated constraint descriptor can adapt to the characteristics of signals changing over time, ensuring that the generated samples always reflect the true relationship between the current signals, thereby enhancing the robustness and adaptability of fault prediction.

[0046] Through the above processing logic, the coupling relationship and phase consistency between the current signal, partial discharge signal and vibration signal are accurately characterized, providing precise constraint conditions for fault sample generation. Compared with traditional single-channel sample generation methods, this multi-modal signal analysis can effectively reduce the possibility of model misjudgment and improve the accuracy and reliability of fault prediction. At the same time, the dynamically updated generated constraint descriptor adapts to the signal change characteristics of the power cabinet under complex working conditions, ensuring the adaptability of the prediction process to complex environments.

[0047] Step S2 generates a cross-channel coupling graph and a phase consistency vector through mutual information calculation and phase difference extraction, successfully characterizing the coupling characteristics of the current signal, partial discharge signal and vibration signal under complex working conditions. The dynamic update of the generated constraint descriptor provides precise constraints for the fault sample generation of step S3, improves the sample quality, effectively alleviates the model misjudgment problem caused by single-channel generation, and provides reliable data support for subsequent training and reasoning.

[0048] Step S2 calculates the cross-channel coupling graph and extracts the phase consistency vector based on the fusion buffer, which serves as the key constraint for the generation of the constraint descriptor for the generation of fault samples. However, relying solely on synchronous signals and coupling constraints cannot sufficiently expand the scarce fault samples, especially under complex working conditions. Single-channel sample generation is prone to model misjudgment and difficult to capture the phase coupling and energy complementarity between channels. Therefore, step S3 focuses on using an adversarial generative network with coupling constraints to expand synchronous fault samples based on the generated constraint descriptor, and filters qualified samples through a phase consistency threshold to improve sample quality and diversity, providing support for subsequent model training.

[0049] Step S3 successfully expands the synchronous fault samples through the band-coupled constrained generative adversarial network and phase consistency threshold screening, alleviating the problem of insufficient fault waveforms of outdoor power distribution cabinets under complex working conditions. The generated qualified samples are written into the enhanced library, not only preserving the subtle coupling relationship between current, partial discharge, and vibration signals, but also improving the diversity and representativeness of the samples, laying a solid foundation for the training of the subsequent expanded convolution-attention network, and effectively enhancing the accuracy and robustness of fault prediction.

[0050] S3.1, band-coupled constrained generative adversarial network: The generative adversarial network consists of a generator and a discriminator. The generator receives random noise and a generation constraint descriptor as input, processes it through an internal neural network, and outputs a multi-channel fault sample containing current signals, partial discharge signals, and vibration signals.

[0051] Random noise is derived from a normal distribution and is used to introduce randomness, while the generation constraint descriptor contains a cross-channel coupling graph and a phase consistency vector to guide the generation process.

[0052] The discriminator receives either a real fault sample or a generated fault sample and outputs a discrimination result representing the probability that the input sample is real.

[0053] When generating a multi-channel fault sample, the generator must ensure that the generated current signal, partial discharge signal, and vibration signal meet the requirements of the cross-channel coupling graph and phase consistency vector in the generation constraint descriptor.

[0054] The loss function consists of three parts: adversarial loss, coupling constraint loss, and phase consistency loss. Adversarial loss measures the degree to which the generated sample approximates the real sample distribution by comparing the discrimination results of the discriminator on real and generated samples.

[0055] Coupling constraint loss quantifies the consistency of the coupling strength between the generated current signal, partial discharge signal, and vibration signal with the real sample by calculating the mutual information difference between the generated sample and the real sample across channels.

[0056] Phase consistency loss ensures that the phase characteristics of the generated sample remain consistent with the real sample by comparing the differences in phase between the generated sample and the real sample.

[0057] The total loss is defined as the weighted sum of adversarial loss, coupling constraint loss, and phase consistency loss. By optimizing the total loss, the generator and discriminator are jointly trained to generate fault samples that are both realistic and consistent with multi-channel coupling characteristics.

[0058] S3.2, expansion of synchronous fault samples based on generation constraint descriptor: The generator generates multi-channel fault samples according to random noise and generation constraint descriptors. The random noise is sampled from a normal distribution as the initial input of the generation process, and the generation constraint descriptors are extracted from the fusion cache of step S2, containing mutual information and phase consistency vectors, for constraining the generation results.

[0059] The generation process maps the random noise and the generation constraint descriptors to multi-channel fault samples containing current signals, partial discharge signals and vibration signals through a neural network. During training, the coupling constraint loss and the phase consistency loss are coupled to ensure that the coupling relationship and the phase characteristics of the generated current signals, partial discharge signals and vibration signals are consistent with the real samples.

[0060] The training process adopts an alternating optimization method. First, the generator is fixed, and the discriminator is optimized to accurately distinguish between real samples and generated samples. Then, the discriminator is fixed, and the generator is optimized to make the generated samples closer to the real sample distribution. The discriminator evaluates the difference between the generated samples and the real samples, calculates the total loss, and updates the network parameters of the generator and the discriminator through backpropagation. Training continues until the total loss converges and the distribution of the generated samples approaches the distribution of the real samples.

[0061] S3.3, screening qualified samples according to the phase consistency threshold: The phase consistency vector of the generated multi-channel fault samples is calculated, specifically including the phase difference between the current signal and the partial discharge signal, the phase difference between the current signal and the vibration signal, and the phase difference between the partial discharge signal and the vibration signal in the generated sample.

[0062] The phase consistency vector of the generated sample is compared with the phase consistency vector of the real fault sample, and the Euclidean distance between the two is calculated. The Euclidean distance represents the degree of deviation of the generated sample from the real sample in terms of phase characteristics. If the Euclidean distance between the phase consistency vector of the generated sample and the phase consistency vector of the real sample is less than the preset phase consistency threshold, the generated sample is determined to be a qualified sample. The qualified sample is written into the augmented library as the basis for using the training data in step S4. Through the phase consistency threshold screening, it is ensured that the generated sample is highly consistent with the real sample in terms of phase characteristics, thereby improving the sample quality and diversity.

[0063] The adversarial generative network with coupling constraints guides the generation process by generating constraint descriptors, ensuring that the generated multi-channel fault samples reflect the true coupling characteristics between current signals, partial discharge signals, and vibration signals. This method can effectively alleviate the problem of insufficient fault waveforms of outdoor distribution cabinets under complex working conditions, providing more available data support for fault prediction. The phase consistency threshold screening mechanism further ensures the consistency of the generated samples with the real samples in terms of phase characteristics by comparing the Euclidean distance of the phase consistency vector, thereby improving the authenticity and reliability of the samples. The generated high-quality fault samples increase the diversity and representativeness of the training data, enabling the subsequent fault prediction model to better adapt to complex signal characteristics and improve the accuracy and robustness of the prediction. This technical logic fully utilizes the internal relationship of multi-modal signals and solves the contradiction between data scarcity and model performance.

[0064] Step S1 aligns the multi-modal signals in the time synchronization window and stores the aligned signals in the fusion cache. Step S2 calculates the cross-channel coupling graph and extracts the phase consistency vector based on the data in the fusion cache, forming the generation constraint descriptor to provide guidance information for the subsequent steps. Step S3 uses the adversarial generative network with coupling constraints to expand the synchronous fault samples with the generation constraint descriptor as the input condition, and selects qualified samples through the phase consistency threshold to ensure that the generated samples meet the coupling characteristics of the real signals. The generated high-quality samples are written to the enhanced library to provide diversified and reliable data support for the model training in step S4. The entire process from signal alignment, feature extraction to sample generation and screening forms a complete technical chain, gradually solving the problem of insufficient fault waveforms, and providing a solid data foundation for the optimization of the fault prediction model.

[0065] Step S2 calculates the cross-channel coupling graph and extracts the phase consistency vector based on the fusion cache as the generation constraint descriptor to ensure multi-modal consistency during sample generation. Step S3 uses the adversarial generative network with coupling constraints to expand the synchronous fault samples and selects qualified samples to write to the enhanced library, solving the problem of insufficient fault samples. However, relying solely on high-quality enhanced samples is still insufficient to cope with the dynamic coupling relationship between multi-modal signals under complex working conditions, and a fault prediction model with long-range dependence capture capability and stable decision boundary stability needs to be further constructed. Therefore, step S4 focuses on training the dilated convolution-attention network with the enhanced library and the original library, and freezes the coupling graph weight as the inference benchmark to improve the robustness and prediction accuracy of the model under nanosecond-level asynchronous disturbances.

[0066] Step S4 successfully constructs a fault prediction model capable of capturing the complex coupling relationship between current, partial discharge, and vibration by jointly training the dilated convolution-attention network using the multi-modal signal samples of the augmented library and the original library. By freezing the cross-channel coupling graph weight as the inference benchmark, the model's decision boundary stability is ensured under the photovoltaic intermittent grid connection and charging pile parallel composite working conditions, effectively improving the accuracy and robustness of fault prediction, providing a reliable model foundation for step S5 edge real-time inference.

[0067] S4.1, Construction process of joint training set: By merging the samples in the original library with the samples in the augmented library, a joint training set is formed. The original library contains current signals, partial discharge signals, and vibration signals collected from actual working conditions, as well as the corresponding fault labels. The augmented library contains synchronous fault samples generated by step S3 and their corresponding fault labels. The purpose of constructing the joint training set is to integrate the advantages of real data and generated data for jointly training the model to improve the model's ability to identify fault features. The merged joint training set not only contains fault samples under real working conditions, but also contains diversified fault samples extended by the generative adversarial network, thereby increasing the diversity and representativeness of the data. This approach effectively addresses the problem of insufficient fault samples, providing comprehensive data support for the model, thereby improving the accuracy and robustness of fault prediction.

[0068] S4.2, Architecture design process of dilated convolution-attention network: When designing the network architecture, the dilated convolution layer, attention mechanism, and coupling graph embedding layer are jointly constructed. The dilated convolution layer expands the receptive field by increasing the dilation rate of the convolution kernel, which is used to extract long-range temporal features of current signals, partial discharge signals, and vibration signals. The attention mechanism uses self-attention mechanism to calculate the correlation between different channel features, enhancing the attention ability to key coupling features. The coupling graph embedding layer embeds the cross-channel coupling graph network generated in step S2 as prior knowledge to guide feature fusion (such as weighted summation), ensuring that the model can learn the coupling relationship between current signals, partial discharge signals, and vibration signals. The reason for designing this architecture is that there are complex temporal and coupling characteristics between multi-modal signals, and separate feature extraction cannot fully capture these relationships. The dilated convolution layer expands the receptive field, the attention mechanism highlights key features, and the coupling graph embedding layer introduces prior knowledge, all of which work together to significantly improve the model's recognition accuracy of fault features. This structured design fully utilizes the complementarity of multi-modal signals, enhancing the accuracy and adaptability of fault prediction.

[0069] S4.3, Calculation process of training target and loss function: The training target is set to minimize the cross-entropy loss. The cross-entropy loss quantifies the prediction accuracy of the model by comparing the difference between the predicted failure probability and the true failure label. During training, a batch of sample data is randomly selected from the joint training set and input into the expanded convolution-attention network to calculate the failure prediction probability output by the model. Then, the cross-entropy loss value is calculated according to the difference between the predicted probability and the true label, and the parameters of the expanded convolution layer, attention mechanism and classification layer are updated using this loss value. The calculation logic is first to determine the probability distribution deviation of each class of failure based on the predicted probability and the true label, then to sum all categories to get the total loss value, and finally to adjust the network parameters through back propagation to reduce the loss value. Using cross-entropy loss can effectively measure the prediction error in the classification task, guiding the model to learn clear fault classification boundaries. This method can improve the model's ability to distinguish fault types and ensure that the prediction results are highly consistent with the true situation.

[0070] S4.4, the process of freezing the coupling graph weight: During training, the weights of the coupling graph embedding layer are kept fixed, and only the parameters of the expanded convolution layer, attention mechanism and classification layer are updated. The weights of the coupling graph embedding layer come from the cross-channel coupling graph generated in step S2, representing the mutual information relationship between the current signal, partial discharge signal and vibration signal. The processing logic is to embed the weights of the cross-channel coupling graph into the network as prior knowledge, and only adjust the parameters of other layers during each parameter update, while the weights of the coupling graph embedding layer remain unchanged. The reason for freezing the weights is that the cross-channel coupling graph reflects the true physical relationship between the signals. If the weights are allowed to change, the model may deviate from the correct coupling characteristics due to the enhancement of the distribution deviation of the library sample or the noise interference in the original library. Fixing the weights can ensure that the model always follows the true relationship between the signals, improves the robustness to nanosecond-level asynchronous disturbances, reduces the probability of misjudgment, and thus guarantees the consistency and reliability of the prediction results under different working conditions.

[0071] S4.5, the technical implementation process of stabilizing the decision boundary: By freezing the weights of the coupling graph embedding layer, the stability of the decision boundary is achieved. During training, the expanded convolution-attention network learns the fault features on the joint training set, and the coupling graph embedding layer provides fixed signal coupling relationship as a constraint, so that the model maintains the consistency of the decision boundary during parameter updating. The processing logic is to use the fixed coupling graph weights to guide feature fusion to ensure that the fault prediction results output by the model do not fluctuate greatly due to changes in data distribution. The reason for implementing this technology is that outdoor power distribution cabinets have complex working conditions and signals are easily disturbed. If the decision boundary is unstable, it may lead to fault misjudgment or omission. A stable decision boundary enables the model to accurately identify fault features under complex working conditions, reduces prediction errors, improves the practicality and reliability of the power cabinet fault prediction self-healing controller, and reduces operation and maintenance costs and potential risks.

[0072] Step S1 aligns the transient of current signal, partial discharge signal and vibration signal through time synchronization window, step S2 calculates cross-channel coupling graph and extracts phase consistency vector, step S3 generates synchronous fault samples by using the adversarial generation network with coupling constraint and constructs the enhanced library. Step S4 constructs the joint training set based on the original library and the enhanced library, designs and trains the dilated convolution-attention network, and freezes the weight of the coupling graph embedding layer to ensure that the model can capture the complex coupling relationship of multi-modal signals while maintaining the stability of the decision boundary. The whole process starts from signal alignment, gradually completes feature extraction, sample generation, model training and inference, forms a complete technical chain, and systematically solves the technical problem of fault prediction of outdoor distribution cabinet under complex working conditions, and provides support for efficient and reliable self-healing control.

[0073] Step S4 trains the dilated convolution-attention network with the enhanced library and the original library, and freezes the coupling graph weight as the inference benchmark, laying the foundation for fault prediction. However, the trained model may not be able to fully guarantee the accuracy of the prediction and the timeliness of the response when facing nanosecond-level asynchronous disturbances in real-time operation, so an online verification and adaptive adjustment mechanism needs to be introduced at the edge to improve the robustness and self-healing ability of fault prediction. Step S5 focuses on this demand, and through real-time monitoring of the synchronization drift and confidence imbalance of the input stream, adaptive resampling and bypass switching are triggered to ensure the stable operation of the distribution cabinet under complex working conditions.

[0074] The edge performs online verification on each frame of fault prediction probability based on the inference benchmark, extracts the fault prediction probability fluctuation resolution and confidence imbalance metric value through double-layer sliding window, generates a self-healing trigger threshold coefficient, and once the threshold coefficient deviates from the safe interval, adaptive resampling and bypass switching are performed to realize real-time monitoring and self-healing response.

[0075] S5.1, online verification and fault prediction probability acquisition: The edge receives real-time multi-modal signal streams, including current signals, partial discharge signals and vibration signals. These signal streams are divided into multiple frames according to the time synchronization window, and each frame contains the sampling data of the current signal, the partial discharge signal and the vibration signal. Each frame of data is input into the trained dilated convolution-attention network, and the network outputs the prediction probability of the frame being faulty after processing. The prediction probability ranges from 0 to 1. The prediction probability represents the confidence of the model in judging the current frame data as faulty, and the higher the value, the more confident the model is that the current frame has a fault. The reason for obtaining the fault prediction probability is to monitor the running state of the distribution cabinet in real time, and use the model prediction result to discover potential fault hazards in time. It can issue an early warning before the fault occurs, take preventive measures to avoid the fault from further expanding, thereby improving the stability and safety of the distribution cabinet.

[0076] S5.2, Double-layer sliding window setting: The double-layer sliding window structure is designed, in which the inner layer sliding window is used to capture the short-term fluctuations of the fault prediction probability, and the outer layer sliding window is used to analyze the long-term trend of the fault prediction probability. The length of the inner layer sliding window is relatively short, reflecting the instantaneous change of the fault prediction probability in a short time; the length of the outer layer sliding window is relatively long, covering data in a longer time period, and is used to evaluate the overall trend of the fault prediction probability. The reason for setting the double-layer sliding window is that when the power distribution cabinet is running under complex working conditions, the signal characteristics are complex, and both short-term fluctuations and long-term trends have important influence on fault prediction. By comprehensively considering the signal changes of different time scales, the accuracy and robustness of fault prediction can be improved, and misjudgment caused by relying on single time scale analysis can be avoided.

[0077] S5.3, Calculate the fluctuation resolution amplitude of fault prediction probability: In the outer layer sliding window, the instantaneous difference of the continuous fault prediction probability is calculated to generate a difference sequence, which reflects the change amplitude of the fault prediction probability between frames. Then, the difference sequence is segmented and integrated, and the peak intensity in the difference sequence is accumulated, and finally the fluctuation resolution amplitude is calculated. The fluctuation resolution amplitude represents the activity of the fault prediction probability fluctuation per unit time, and its dimension is the fault prediction probability per time. The reason for calculating the fluctuation resolution amplitude is that when a fault occurs, the fault prediction probability may fluctuate sharply, and by quantifying the fluctuation amplitude, the signs of fault occurrence can be captured. The dynamic change of the fault prediction probability can be monitored in real time, and abnormal fluctuations can be found in time, so as to improve the sensitivity and response speed of fault prediction.

[0078] S5.4, Calculate the confidence imbalance metric value: In the outer layer sliding window, a warning threshold is set to divide the fault prediction probability into a high confidence group and a low confidence group. The high confidence group contains fault prediction probability higher than the warning threshold, and the low confidence group contains fault prediction probability lower than the warning threshold. The average probability of the high confidence group and the low confidence group is calculated respectively to form a confidence contrast array. Based on the confidence contrast array, the logarithmic entropy is calculated, which reflects the uniformity of the confidence distribution. The confidence imbalance metric value is defined as 1 minus the logarithmic entropy, and the larger the value is, the more unbalanced the confidence distribution is.

[0079] Under normal working conditions, the fault prediction probability should remain relatively stable, and the confidence distribution is relatively balanced; while under abnormal working conditions, the confidence distribution may be unbalanced, and by quantifying the imbalance degree, the abnormal situation of model prediction can be identified. The confidence change of model prediction can be monitored, and the mismatch between model and data can be found in time, so as to improve the reliability of fault prediction.

[0080] S5.5, Generate self-healing trigger threshold coefficient: The fluctuation resolution amplitude and the confidence imbalance metric value are mapped to a two-dimensional orthogonal plane to form a coordinate point. By using the coordinate point set under the normal working condition, a minimum circumscribed ellipse is fitted to determine the center position of the safety interval. The vector module length from the current coordinate point to the center of the safety interval is calculated to obtain the self-healing trigger threshold coefficient. The self-healing trigger threshold coefficient represents the degree of deviation of the current state from the normal working condition. The larger the value, the more serious the deviation.

[0081] By comprehensively considering the fluctuation and confidence imbalance of the fault prediction probability, a reliable decision basis is constructed to avoid misjudgment caused by relying on a single indicator. It can accurately judge whether the input signal stream has double abnormalities of synchronous drift and confidence imbalance, thereby improving the accuracy and timeliness of self-healing action.

[0082] S5.6, adaptive resampling and write-back enhancement library: When the self-healing trigger threshold coefficient exceeds the preset safety threshold, high-frequency resampling is performed on the current frame and its adjacent frames to generate fine-grained data. The fine-grained data and its corresponding fault prediction probability are written back to the enhancement library for subsequent model optimization.

[0083] The reason for adaptive resampling is that in abnormal conditions, more detailed data can be obtained through high-frequency resampling to improve data quality. The reason for writing back to the enhancement library is that supplementing these data to the training data set can optimize the model performance. It can adjust the data acquisition strategy in real time, improve the adaptability of the model to abnormal working conditions, and thus enhance the accuracy of fault prediction.

[0084] S5.7, technical logic of driving bypass switching: When the self-healing trigger threshold coefficient exceeds the preset safety threshold, solid-state bypass switching is triggered to isolate the fault area and prevent further expansion of the fault. The reason for driving bypass switching is that by isolating the fault area in time, the power distribution cabinet and related equipment can be protected from further damage. The system can quickly respond to faults, reducing the impact of faults on the overall system, thereby improving the self-healing ability and safety of the power distribution cabinet.

[0085] The previous steps complete the transient alignment, feature extraction, fault sample generation, and model training of multi-modal signals, providing data and model support for subsequent fault prediction. On this basis, step S5 obtains the fault prediction probability through online verification and fault prediction probability, calculates the fluctuation resolution amplitude and the confidence imbalance metric value of the fault prediction probability by combining the double-layer sliding window setting, generates the self-healing trigger threshold coefficient, and realizes real-time monitoring and abnormality determination of the power distribution cabinet running state. After abnormality determination, adaptive resampling and write-back enhancement library optimize data and models, drive bypass switching to isolate faults, forming a complete technical chain from signal acquisition, fault prediction to self-healing control, ensuring the safe and stable operation of the power distribution cabinet under the composite working condition of photovoltaic intermittent grid connection and charging pile parallel connection.

[0086] Among them, the adaptive resampling write enhancement library and the driving bypass switching are a kind of technical logic integrated in the power cabinet fault prediction self-healing controller and the deep learning method, and the core is to realize the fault prediction and self-healing control of the system through real-time monitoring, data optimization and rapid response. The technical logic contains three key components: first, "adaptive resampling" means that the system dynamically adjusts the data acquisition strategy according to the abnormal conditions monitored in real time, so as to capture the fault characteristics under complex working conditions with higher frequency or accuracy; second, "write enhancement library" involves storing the data obtained by adaptive resampling and the corresponding fault prediction probability into the enhancement library as the data set for training the fault prediction model, and continuously supplementing high-quality data to optimize the model performance; finally, "driving bypass switching" triggers the protection mechanism after detecting the fault, isolates the fault area to prevent the fault from expanding, and ensures the safe operation of the system. The meaning is that this technical logic improves the accuracy of fault prediction through adaptive data acquisition, optimizes the robustness and adaptability of the model by using write enhancement library, and realizes rapid fault isolation by means of bypass switching, which together guarantees the stability and safety of the power cabinet under complex working conditions.

[0087] The above formulas are dimensionless values calculated, and the formulas are obtained by software simulation of a large amount of data to obtain a formula of the latest real situation, and the preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0088] It should be noted that the system of the present application can be deployed on the device itself to realize embedded application, or can be run on PC or other terminal with user interface, so as to meet various hardware environments and use requirements.

[0089] The above only describes some exemplary embodiments of the present application by way of illustration, and it is needless to say that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present application. Therefore, the above figures and descriptions are illustrative in nature and should not be understood as limiting the scope of protection of the claims of the present application.

[0090] It has to be noted that, in the present document, the terms "first", "second", etc. merely serve the purpose of differentiation and do not require or imply any kind of ordering or sequence of the entities or actions associated therewith. Furthermore, the terms "comprising", "containing", etc. are to be interpreted as non- exclusive in the sense that a process, method, article, or apparatus that comprises a list of elements is not necessarily limited to those elements, but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a", "contains... a", etc. does not, without further restriction, exclude the presence of additional identical elements in the process, method, article, or apparatus.

[0091] The above description is only specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A power cabinet fault prediction self-healing controller and deep learning method, characterized in that, The method comprises the steps of: S1: reading the multi-mode sensing stream in the time synchronization window and writing it into the fusion cache after adding the phase identification, so as to lock the transient alignment starting point of the current-local discharge-vibration three-channel; S2: calculating the cross-channel coupling graph and extracting the phase consistency vector according to the fusion cache, and updating the vector as a generation constraint descriptor with the frame stream; S3: calling the adversarial generation network with coupling constraints to expand the synchronous fault sample based on the descriptor, and generating the sample and writing it into the enhanced library after screening qualified samples according to the phase consistency threshold; S4: training the dilated convolution-attention network with the enhanced library and the original library, and freezing the coupling graph weight as an inference benchmark to stabilize the decision boundary; S5: the edge end checks the fault prediction probability of each frame online with the inference benchmark as the ruler, extracts the fault prediction probability fluctuation resolution and confidence imbalance metric value through a double-layer sliding window, and generates a self-healing trigger threshold coefficient, and once it deviates from the safe interval, it immediately adapts to the enhanced library and drives the bypass switch.

2. The power cabinet failure prediction self-healing controller and deep learning method according to claim 1, wherein, Step S1 includes the following contents: The power cabinet fault prediction self-healing controller reads the current signal, local discharge signal and vibration signal through the time synchronization window, and writes it into the fusion cache after adding the phase identification, so as to lock the transient alignment starting point of the three signals. In order to set the time synchronization window, the synchronous signal fragments are extracted from the current sensor, the local discharge sensor and the vibration sensor. Hilbert transform is applied to each signal fragment to calculate the phase identification. The signal fragment and the phase identification are integrated into a record, which is stored in the fusion cache in time sequence. The alignment of the signals in time and phase is realized through the time synchronization window and the phase identification, and the transient alignment starting point is locked.

3. The power cabinet failure prediction self-healing controller and deep learning method of claim 2, wherein, Step S2 includes the following contents: The records are read in time sequence from the fusion cache, the mutual information between the current signal and the local discharge signal, the mutual information between the current signal and the vibration signal, and the mutual information between the local discharge signal and the vibration signal are calculated, which are respectively used as the edge weight of the cross-channel coupling graph. The phase difference is calculated based on the phase identification and normalized to form a phase consistency vector. The mutual information and the phase consistency vector are combined into a generation constraint descriptor, which is dynamically updated with the time synchronization window.

4. The power cabinet failure prediction self-healing controller and deep learning method of claim 3, wherein, Step S3 includes the following contents: The adversarial generation network with coupling constraints is used to expand the synchronous fault sample. The generator receives random noise and generates a multi-channel fault sample containing current signal, local discharge signal and vibration signal.

5. The power cabinet failure prediction self-healing controller and deep learning method according to claim 4, wherein, Step S3 also includes the following contents: The generation constraint descriptor is composed of the cross-channel coupling graph and the phase consistency vector. The coupling constraint loss and the phase consistency loss are used to optimize the generation process to ensure that the coupling relationship and phase characteristics of the generated sample are consistent with the real sample. Step S3 also includes the following contents:

6. The power cabinet failure prediction self-healing controller and deep learning method of claim 5, wherein, The discriminator evaluates the difference between the generated sample and the real sample to optimize the network parameters, calculates the phase consistency vector of the generated sample, and compares the Euclidean distance with that of the real sample to screen qualified samples and write them into the enhanced library to improve the sample quality. Step S4 includes the following contents:

7. The power cabinet fault prediction self-healing controller and deep learning method according to claim 6, characterized in that, ​ The original library and the enhanced library are combined to form a joint training set, which contains real collected current signals, partial discharge signals and vibration signals and their fault labels, and synchronous fault samples extended by the generative adversarial network; the network architecture is designed to consist of dilated convolution layers, attention mechanisms and coupled graph embedding layers, the dilated convolution layers extract long-range temporal features, the attention mechanisms enhance the attention to key coupling features, and the coupled graph embedding layers embed the cross-channel coupling graph into the network to guide feature fusion.

8. The power cabinet fault prediction self-healing controller and deep learning method according to claim 7, characterized in that, Step S4 includes the following contents: During training, the cross-entropy loss is minimized, the coupled graph embedding layer weights are frozen, and only the dilated convolution layer, attention mechanism and classification layer parameters are updated to ensure that the model follows the real physical relationship between signals, and the decision boundary is stabilized by freezing the coupled graph embedding layer weights.

9. The power cabinet failure prediction self-healing controller and deep learning method of claim 8, wherein, Step S4 includes the following contents: The double-layer sliding window structure is composed of an inner sliding window and an outer sliding window, the inner sliding window calculates the instantaneous difference of fault prediction probability to capture short-term fluctuations, the outer sliding window accumulates the peak intensity of the difference to evaluate long-term trends, the fluctuation resolution amplitude is calculated by piecewise integration of the peak intensity of the difference sequence to reflect the fluctuation activity of the fault prediction probability, and the confidence imbalance measure is calculated by logarithmic entropy normalization of the confidence contrast array to reflect the imbalance degree of the confidence distribution.

10. The power cabinet fault prediction self-healing controller and deep learning method according to claim 9, characterized in that, Step S5 includes the following contents: The self-healing trigger threshold coefficient is calculated by the minimum circumscribed ellipse method to calculate the vector module length from the joint scatter density center of the fluctuation resolution amplitude and the confidence imbalance measure to the origin, and when the self-healing trigger threshold coefficient exceeds the safety threshold, the adaptive resampling and writing enhancement library is executed and the bypass switch is driven.

Images