Deep learning based adaptive diagnosis system and method for electrical control circuit fault

Abstract

The application discloses a deep learning-based electrical control circuit fault self-adaptive diagnosis system and method, which comprises a data acquisition module, a preprocessing module, a feature extraction module, a diagnosis model module, a self-adaptive optimization module and a visual interactive module. Based on the improved Transformer architecture, the local feature perception ability of the self-attention mechanism is enhanced by embedding the time convolution operator, high-dimensional time sequence features are extracted, a double-channel deep residual network is constructed, the electrical features and the thermal-mechanical features are analyzed respectively, the fault classification results are output after fusion, the model parameters are updated through online incremental learning, the pre-training parameters are transferred based on the migration learning controller to adapt to the new fault sample distribution characteristics, and the network depth and the convolution kernel size are dynamically optimized by using the genetic algorithm to improve the generalization ability. The application has the advantages that the improved Transformer architecture and the self-adaptive optimization mechanism are combined with the time convolution and the migration learning, high-precision and multi-dimensional information fusion is realized, and new fault types can be quickly adapted.

Description

Technical Field

[0001] This invention relates to deep learning, and in particular to an adaptive fault diagnosis system and method for electrical control circuits based on deep learning. Background Technology

[0002] Electrical control circuits are an indispensable part of modern industrial automation systems, widely used in various mechanical equipment, production lines, and transportation vehicles. With the development of electrical control technology, the complexity of circuits has gradually increased, leading to a corresponding increase in the likelihood of malfunctions. Electrical control circuit failures typically cause equipment downtime or performance degradation, posing risks to production efficiency and safety. Therefore, timely and accurate diagnosis of circuit faults and the implementation of appropriate repair measures are of paramount importance.

[0003] Most current adaptive fault diagnosis systems for electrical control circuits are based on rule engines or shallow machine learning, relying on manual feature extraction and threshold setting. While these methods can identify common faults, they have significant shortcomings: First, they lack the ability to analyze complex nonlinear signals, resulting in a high false detection rate; second, the models have poor generalization capabilities, requiring redesign of the rule base when the circuit structure or load changes, leading to high maintenance costs; and third, traditional algorithms are insufficient in modeling the global correlation of time-series features, making it difficult to capture the dynamic coupling relationships of multi-sensor data and easily missing early latent faults. Summary of the Invention

[0004] To improve existing adaptive fault diagnosis systems and methods for electrical control circuits, this paper presents a deep learning-based adaptive fault diagnosis system and method for electrical control circuits. It integrates an improved version of the TRANSFORMER architecture with a temporal convolutional network and captures long-term temporal dependencies through a multi-head self-attention mechanism, significantly reducing the reliance on human experience and full data, and providing a more cost-effective solution for complex industrial scenarios.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A deep learning-based adaptive fault diagnosis system for electrical control circuits includes:

[0007] The data acquisition module includes a high-precision current / voltage sensor array, an infrared thermal imaging unit, and a vibration sensor group, which are used to acquire the operating sequence signals, temperature field distribution data, and mechanical vibration spectrum of the electrical control circuit in real time.

[0008] The preprocessing module includes a signal denoising unit and a data alignment unit. The signal denoising unit uses a wavelet threshold denoising algorithm to perform nonlinear filtering on the time-series signal and applies an adaptive contrast enhancement algorithm to process the infrared image.

[0009] The feature extraction module includes an improved Transformer architecture, in which a time convolution operator is embedded in a self-attention mechanism to form a time sequence encoder with local feature perception capability.

[0010] The diagnostic model module includes a dual-channel deep residual network and an online incremental learning unit, the dual-channel deep residual network processes electrical features and thermal-mechanical features respectively to obtain electrical control circuit fault diagnosis results.

[0011] The adaptive optimization module includes a transfer learning controller and a model structure dynamic adjustment unit, the transfer learning controller selects pre-trained model parameters according to the distribution characteristics of new fault samples, and the dynamic adjustment unit optimizes the network depth and convolution kernel size through a genetic algorithm.

[0012] The visual interaction module includes a fault thermal map generator and a diagnostic decision tree interpreter, which can display fault positioning results in real time and provide explainable analysis of diagnostic basis.

[0013] Preferably, the data acquisition module specifically includes:

[0014] A high-precision current / voltage sensor array captures dynamic time sequence signals of the electrical control circuit in real time through a multi-channel synchronous sampling technology, completes signal filtering, analog-to-digital conversion and data calibration, synchronously encapsulates time-stamped original signals and working condition parameters through a high-speed communication interface, and transmits them to the preprocessing module for further processing.

[0015] An infrared thermal imaging unit captures temperature field distribution data by real-time scanning of the surface of the electrical equipment, integrates the time-sequenced temperature matrix with the equipment working condition parameters, and constructs a dynamic thermal feature data set for streaming to the preprocessing module.

[0016] A vibration sensor group acquires vibration analog signals in real time and converts them into digital signals, generates time-domain waveform data after filtering, amplification and analog-to-digital conversion, and transmits them to the preprocessing module.

[0017] Preferably, the preprocessing module specifically includes:

[0018] A signal denoising unit performs nonlinear filtering on time sequence signals using a wavelet threshold denoising algorithm, specifically including:

[0019] The original signals based on the data acquisition module are decomposed into subbands of different frequencies through multi-scale wavelet transform, and the approximation coefficients and detail coefficients are extracted for wavelet decomposition, the formula being:

[0020]

[0021]

[0022] wherein, is the original signal, is the wavelet base function, is the scale function, , are the approximation and detail coefficients of the jth layer respectively, and k is the position index;

[0023] Based on the extracted approximation and detail coefficients, the signal and noise are distinguished, and nonlinear filtering is performed to suppress noise;

[0024] Based on the processed wavelet coefficients, the denoised signal is reconstructed;

[0025] The data alignment unit processes the infrared image through an adaptive contrast enhancement algorithm, specifically including:

[0026] The local mean and standard deviation of the infrared image are calculated by block;

[0027] Based on the standard deviation, the gain factor is dynamically designed, and each sub-block is subjected to nonlinear contrast stretching to avoid block effects by fusing overlapping areas;

[0028] Based on global normalization, the standard dynamic range is obtained, which can enhance the features of the fault area while suppressing noise and over-enhancement.

[0029] Preferably, the feature extraction module specifically includes:

[0030] The improved Transformer architecture specifically includes:

[0031] Based on the data processed by the preprocessing module, one-dimensional causal convolution is performed, the number of convolution output channels is d, and local features are obtained;

[0032] The local features output by time convolution are added to the original data embedding element by element to form fusion features;

[0033] Query (Q), key (K), and value (V) matrices are generated from the fusion data through linear transformation, and the query (Q), key (K), and value (V) are divided into h heads, and the dimension of each head is ;

[0034] The attention outputs of each head are spliced, and they are projected back to the original input dimension through a linear layer;

[0035] The output of multi-head attention is added to the original input, which retains the initial information, alleviates gradient disappearance, stabilizes the training process, adjusts the output distribution, and accelerates convergence;

[0036] Nonlinear transformation is introduced through the feedforward network, and the refined features are obtained position by position;

[0037] The output result of the feedforward network is added to the normalized attention result, and the output distribution is stabilized through layer normalization as the final output of the encoder layer;

[0038] Multiple encoder layers are repeatedly stacked to extract high-order time sequence features layer by layer, and the parameters of each layer are independent and the local features are gradually fused.

[0039] Preferably, the diagnostic model module specifically comprises:

[0040] The double-channel deep residual network model avoids gradient disappearance in deep network based on residual connection after obtaining electrical characteristic data and thermal-mechanical characteristic data, and uses the following residual block structure for each channel:

[0041]

[0042] wherein, is a convolution operation, is a convolution kernel;

[0043] After processing electrical characteristics and thermal-mechanical characteristics in two channels respectively, the outputs of the two channels are fused to obtain the preliminary fault diagnosis result through a fully connected layer;

[0044] The online incremental learning unit uses an online gradient descent method to update the model step by step, and the loss function is updated based on cross-entropy loss for each new data batch, which is in the form of:

[0045]

[0046] wherein, is a true label, is the predicted output of the model, and N is the total number of data;

[0047] Based on the output of the double-channel deep residual network and the adaptability of the incremental learning unit, the diagnosis result is divided by the probability of the model output to obtain the final electrical control circuit fault diagnosis result and identify the fault type.

[0048] Preferably, the adaptive optimization module specifically comprises:

[0049] The transfer learning controller calculates the distribution difference between the new sample and the source domain sample by extracting the feature vector of the new fault sample, and selects the parameter transfer strategy;

[0050] The model structure dynamic adjustment unit optimizes the structure of the neural network, including the network depth and the size of the convolution kernel, by a genetic algorithm.

[0051] The migration learning selects a suitable model and performs fine-tuning, the dynamic adjustment unit optimizes the network structure, and the model is further fine-tuned according to new samples and structures until the diagnosis effect reaches the expectation.

[0052] Preferably, the migration learning controller specifically comprises:

[0053] A pre-trained model library containing multiple device models is constructed, and each model establishes a hyper-spherical decision boundary in the parameter space;

[0054] The WASSERSTEIN distance between the new fault data distribution and each pre-trained model is calculated;

[0055] The three models closest to the new fault data are selected for model parameter interpolation, and the interpolation weight is determined according to the reciprocal of the relative distance;

[0056] The bottom feature extraction layer is frozen during migration, and only the upper classifier is fine-tuned.

[0057] Preferably, the model structure dynamic adjustment unit specifically comprises:

[0058] The network depth is evolved by automatically increasing or decreasing the number of residual blocks based on the change rate of the validation set accuracy, and the increase or decrease decision formula is:

[0059]

[0060] wherein, is the change amount of the number of residual blocks, is the adjustment coefficient, is the difference between the current stage and the previous stage validation set accuracy, reflecting the change of the model performance, is the standard deviation of the validation set accuracy, used for normalizing the adjustment amount;

[0061] The convolution kernel size is optimized by genetic algorithm evolution selection in the candidate set;

[0062] The number of channels of each layer is dynamically adjusted based on the information entropy value of the feature map.

[0063] Preferably, the visual interaction module specifically comprises:

[0064] The heat map generator uses gradient class activation mapping technology to superimpose the fault feature area focused by the neural network to the circuit topology map in a color mapping manner, real-time labels the abnormal current / voltage signal distribution, and accurately locates the fault element;

[0065] The diagnostic decision tree interpreter explains the fault classification probability output by the model into understandable logical reasoning chains through rule extraction and path backtracking, dynamically displays the decision-making process of "signal distortion -> feature extraction -> hierarchical judgment", and correlates the historical fault case library.

[0066] Further, the electrical control circuit fault adaptive diagnosis method based on deep learning comprises:

[0067] Through a high-precision sensor array, an infrared thermal imaging and a vibration sensor group, time sequence signals, temperature field distribution and mechanical vibration spectrum data of the circuit are synchronously collected.

[0068] A wavelet threshold denoising algorithm is used to denoise the time sequence signals, data alignment is used to eliminate the time sequence deviation of the multi-sensor, and an adaptive contrast enhancement is used to optimize the infrared image quality.

[0069] Based on the improved Transformer architecture, a time convolution operator is embedded to enhance the local feature perception ability of the self-attention mechanism, and high-dimensional time sequence features are extracted.

[0070] A double-channel deep residual network is constructed to analyze electrical features and thermal-mechanical features respectively, and the fusion output is used to output the fault classification result, and the model parameters are updated through online incremental learning.

[0071] Based on the transfer learning controller, pre-training parameters are transferred to adapt to the distribution characteristics of new fault samples, and a genetic algorithm is used to dynamically optimize the network depth and convolution kernel size to improve the generalization ability.

[0072] A fault thermal map is generated to locate the abnormal area, and a visual analysis and an interpretable report are provided for the diagnosis basis.

[0073] Compared with the prior art, the advantages of the present application are:

[0074] The improved TRANSFORMER architecture, by embedding a temporal convolution operator into its self-attention mechanism, effectively perceives local features in time-series signals, improving the accuracy and robustness of fault diagnosis. This is particularly beneficial for complex electrical control circuits, where it accurately captures the temporal patterns and evolution of faults. The dynamic model adjustment mechanism, combined with a transfer learning controller and a genetic algorithm, automatically optimizes the network structure based on different fault samples and changing environmental conditions, such as adjusting network depth and convolution kernel size, thus enabling rapid adaptation to new fault types. This adaptive capability allows the system to continuously learn and improve the accuracy of fault diagnosis. Especially when facing unknown or complex faults, the system can effectively leverage existing knowledge through transfer learning, avoiding the inefficiency and misdiagnosis risks of traditional methods. Furthermore, the system uses a dual-channel deep residual network to separate electrical and thermo-mechanical features, further improving the accuracy of multi-dimensional information fusion and providing a more comprehensive and accurate analysis for fault localization. Attached Figure Description

[0075] Figure 1 This is a schematic diagram of the system proposed in this invention;

[0076] Figure 2 This is a schematic diagram of the method proposed in this invention;

[0077] Figure 3 This is a schematic diagram of the Transformer architecture proposed in this invention;

[0078] Figure 4 This is a schematic diagram of the transfer learning controller proposed in this invention;

[0079] Figure 5 This is a schematic diagram of the dynamic adjustment unit of the model structure proposed in this invention;

[0080] Figure 6 This is an architecture diagram of the electronic devices in this solution;

[0081] Figure 7 This is a schematic diagram of the computer-readable storage medium structure in this scheme. Detailed Implementation

[0082] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0083] See Figure 1 As shown, the deep learning-based adaptive fault diagnosis system for electrical control circuits includes:

[0084] The data acquisition module includes a high-precision current / voltage sensor array, an infrared thermal imaging unit, and a vibration sensor group, which are used to collect real-time timing signals, temperature field distribution data, and mechanical vibration spectrum of the electrical control circuit;

[0085] The preprocessing module includes a signal denoising unit and a data alignment unit. The signal denoising unit uses a wavelet threshold denoising algorithm to perform nonlinear filtering on the timing signal and applies an adaptive contrast enhancement algorithm to process the infrared image.

[0086] The feature extraction module includes an improved Transformer architecture, which embeds a time convolution operator in the self-attention mechanism to form a timing encoder with local feature perception capability.

[0087] The diagnostic model module includes a dual-channel deep residual network and an online incremental learning unit. The dual-channel deep residual network processes electrical features and thermal-mechanical features respectively to obtain electrical control circuit fault diagnosis results.

[0088] The adaptive optimization module includes a transfer learning controller and a model structure dynamic adjustment unit. The transfer learning controller selects pre-trained model parameters according to the distribution characteristics of new fault samples, and the dynamic adjustment unit optimizes the network depth and convolution kernel size through a genetic algorithm.

[0089] The visualization interaction module includes a fault thermal map generator and a diagnostic decision tree interpreter, which can display fault positioning results in real time and provide explainable analysis of diagnostic evidence.

[0090] Referring to Figure 1 As shown, the data acquisition module specifically includes:

[0091] The high-precision current / voltage sensor array captures dynamic timing signals of the electrical control circuit in real time through multi-channel synchronous sampling technology, completes signal filtering, analog-to-digital conversion, and data calibration, synchronously encapsulates the original signals with time stamps and working condition parameters through a high-speed communication interface, and transmits them to the preprocessing module for further processing.

[0092] The infrared thermal imaging unit captures temperature field distribution data by real-time scanning of the surface of the electrical equipment, integrates the timing temperature matrix with the equipment working condition parameters, and constructs a dynamic thermal feature data set for streaming transmission to the preprocessing module.

[0093] The vibration sensor group converts the vibration analog signal into a digital signal in real time, generates time-domain waveform data after filtering, amplification, and analog-to-digital conversion, and transmits it to the preprocessing module.

[0094] It can be understood that the sensor data is usually high-dimensional and needs to be associated and integrated with the working condition data of the electrical equipment. The large amount of data and complex association may cause delay in data processing. Data dimension reduction and feature extraction techniques are adopted to optimize the data processing algorithm. In addition, a real-time data stream processing framework such as APACHE KAFKA can be introduced to ensure the real-time performance and processing efficiency of the data.

[0095] Referring to Figure 1 As shown in the figure, the preprocessing module specifically includes:

[0096] The signal denoising unit adopts a wavelet threshold denoising algorithm to perform nonlinear filtering on the time series signal, specifically including:

[0097] The original signal based on the data acquisition module is decomposed into subbands of different frequencies through multi-scale wavelet transform, and the approximation coefficients and detail coefficients are extracted for wavelet decomposition, and the formula is:

[0098]

[0099]

[0100] wherein, is the original signal, is a wavelet basis function, is a scale function, , are the approximation coefficients and detail coefficients of the jth layer respectively, and k is a position index;

[0101] Based on the extracted approximation coefficients and detail coefficients, the signal and noise are distinguished and nonlinearly filtered to suppress noise;

[0102] The denoised signal is reconstructed based on the processed wavelet coefficients;

[0103] The data alignment unit processes the infrared image through an adaptive contrast enhancement algorithm, specifically including:

[0104] The local mean and standard deviation of the infrared image are calculated by block;

[0105] Based on the standard deviation, a gain factor is dynamically designed, and each subblock is nonlinearly stretched in contrast to avoid block effects by fusing overlapping areas;

[0106] Based on global normalization, a standard dynamic range is obtained to enhance the features of the fault area while suppressing noise and over-enhancement.

[0107] Specifically, in practical applications, a soft threshold or a hard threshold method is usually used to process wavelet coefficients, the soft threshold can provide a smoother signal reconstruction, and the hard threshold can better preserve important detail information. Therefore, selecting a suitable threshold method and strategy is crucial for improving the denoising effect and preserving signal characteristics. In addition, multiple decomposition levels can be combined to process noise in different frequency bands, improving the overall performance of denoising, especially in the processing of complex signals.

[0108] Referring to Figure 3 The feature extraction module specifically includes:

[0109] The improved Transformer architecture specifically includes:

[0110] Based on the data processed by the preprocessing module, one-dimensional causal convolution is performed, the number of convolution output channels is d, and local features are obtained;

[0111] The local features output by the time convolution are added to the original data embedding element by element to form fusion features;

[0112] The query (Q), key (K), and value (V) matrices are generated from the fusion data through linear transformation, and the query (Q), key (K), and value (V) are divided into h heads, and the dimension of each head is ;

[0113] The attention outputs of each head are spliced, and they are projected back to the original input dimension through a linear layer;

[0114] The output of multi-head attention is added to the original input to retain initial information, alleviate gradient disappearance, stabilize the training process, adjust the output distribution, and accelerate convergence;

[0115] Nonlinear transformation is introduced through a feedforward network to refine the features position by position;

[0116] The output results of the feedforward network are added to the normalized attention results, and the output distribution is stabilized through layer normalization, which is used as the final output of the encoder layer;

[0117] Multiple encoder layers are repeatedly stacked to extract high-order time sequence features layer by layer, and the parameters of each layer are independent, and the local features are gradually fused.

[0118] Specifically, in the multi-head attention mechanism, in addition to dividing the query (Q), key (K) and value (V) into multiple heads and performing parallel calculation, it is worth noting that the multi-head attention can enable the model to simultaneously focus on different parts of the input data in different subspaces, thereby capturing more rich temporal features. Each head will learn different weights when calculating, so it can capture the diversity and complexity of information in different attention subspaces. By concatenating the outputs of these heads and restoring the dimension through linear transformation, the expression ability of the model can be effectively enhanced.

[0119] Referring to Figure 1 As shown in the figure, the diagnostic model module specifically includes:

[0120] The dual-channel deep residual network model, after obtaining the electrical characteristic data and the thermal-mechanical characteristic data, avoids gradient disappearance in deep network based on residual connection, for each channel, uses the following residual block structure:

[0121]

[0122] wherein, is a convolution operation, is a convolution kernel;

[0123] After the two channels process the electrical characteristics and the thermal-mechanical characteristics respectively, the outputs of the two channels are fused, and the preliminary fault diagnosis result is obtained through a fully connected layer;

[0124] The online incremental learning unit uses an online gradient descent method to update the model step by step, and for each new data batch, the loss function is updated based on the cross-entropy loss, which is in the form of:

[0125]

[0126] wherein, is the true label, is the predicted output of the model, and N is the total number of data;

[0127] Based on the output of the dual-channel deep residual network and the adaptability of the incremental learning unit, the diagnostic result is divided by the probability of the model output, and the final electrical control circuit fault diagnosis result and the fault type are obtained.

[0128] It can be understood that in the training of deep network, especially in the deep residual network, the phenomenon of gradient disappearance or gradient explosion may occur, which leads to the ineffective updating of network parameters. A proper initialization method (such as XAVIER initialization or HE initialization) can be used to ensure the reasonable distribution of weights in the initial stage, or when the gradient exceeds the preset threshold, the gradient is clipped to prevent gradient explosion.

[0129] Referring to Figure 1 As shown in the figure, the adaptive optimization module specifically includes:

[0130] A transfer learning controller calculates the distribution difference between the new sample and the source domain sample by extracting the feature vector of the new fault sample, and selects a parameter transfer strategy;

[0131] A model structure dynamic adjustment unit optimizes the structure of the neural network, including the network depth and the size of the convolution kernel, through a genetic algorithm;

[0132] The transfer learning selects a suitable model and fine-tunes it, the dynamic adjustment unit optimizes the network structure, and the model is further fine-tuned according to the new sample and structure until the diagnosis effect reaches the expectation.

[0133] Referring to Figure 4 As shown in the figure, the transfer learning controller specifically includes:

[0134] A pre-trained model library containing multiple device models is constructed, and each model establishes a hyper-spherical decision boundary in the parameter space;

[0135] The WASSERSTEIN distance between the distribution of the new fault data and each pre-trained model is calculated;

[0136] The three models with the closest distance are selected for model parameter interpolation, and the interpolation weight is determined according to the reciprocal of the relative distance;

[0137] The bottom feature extraction layer is frozen during the transfer process, and only the upper classifier is fine-tuned.

[0138] Referring to Figure 5 As shown in the figure, the model structure dynamic adjustment unit specifically includes:

[0139] The network depth is evolved by automatically increasing or decreasing the number of residual blocks based on the verification set accuracy rate change rate, and the increase or decrease decision formula is:

[0140]

[0141] Wherein, is the change amount of the number of residual blocks, is the adjustment coefficient, is the difference between the current stage and the previous stage verification set accuracy rate, reflecting the change of model performance, is the standard deviation of the verification set accuracy rate, used for normalizing the adjustment amount;

[0142] The convolution kernel size is optimized by genetic algorithm evolution selection in the candidate set;

[0143] The number of channels of each layer is dynamically adjusted based on the information entropy value of the feature map.

[0144] Referring to Figure 1 As shown in the figure, the visual interaction module specifically includes:

[0145] The heat map generator adopts gradient class activation mapping technology to superimpose the fault feature area focused by the neural network to the circuit topology map in a color mapping manner, to mark the abnormal current / voltage signal distribution in real time and accurately locate the fault element;

[0146] The diagnostic decision tree interpreter parses the fault classification probability output by the model into understandable logical reasoning chains through rule extraction and path backtracking, dynamically displays the decision-making process of "signal distortion → feature extraction → hierarchical judgment", and associates with the historical fault case library.

[0147] Referring to Figure 2 As shown in the figure, the deep learning-based electrical control circuit fault adaptive diagnosis method includes:

[0148] Step one: through a high-precision sensor array, an infrared thermal imaging and a vibration sensor group, time sequence signals, temperature field distribution and mechanical vibration spectrum data of the circuit are synchronously collected;

[0149] Step two: wavelet threshold denoising algorithm is used to denoise the time sequence signals, data alignment is used to eliminate the time sequence deviation of multiple sensors, and adaptive contrast enhancement is used to optimize the infrared image quality;

[0150] Step three: based on the improved Transformer architecture, time convolution operators are embedded to enhance the local feature perception ability of the self-attention mechanism, and high-dimensional time sequence features are extracted;

[0151] Step four: a double-channel deep residual network is constructed to analyze electrical features and thermal-mechanical features respectively, and after fusion, a fault classification result is output, and model parameters are updated through online incremental learning;

[0152] Step five: based on the transfer learning controller, pre-training parameters are transferred to adapt to the distribution characteristics of new fault samples, and the network depth and convolution kernel size are dynamically optimized by using genetic algorithm to improve the generalization ability;

[0153] Step six: a fault heat map is generated to locate the abnormal area, and a visual analysis and explainable report are provided for diagnosis basis by combining the decision tree interpreter.

[0154] Further, the method according to the embodiment of the application can also be implemented by means of Figure 6 the architecture of the electronic device as shown. As Figure 6As shown, the electronic device 500 can include a bus 501, one or more CPUs 502, a read-only memory (ROM) 503, a random access memory (RAM) 504, a communication port connected to a network 505, an input / output component 506, a hard disk 507, etc. The storage device in the electronic device 500, such as the ROM 503 or the hard disk 507, can store the deep learning based electrical control circuit fault adaptive diagnosis system and method provided in the present application. The electronic device 500 can also include a terminal interface 508. Of course, Figure 6 The architecture shown is only exemplary, and in implementing different devices, according to actual needs, some of the components shown can be omitted Figure 6 One or more components of the electronic device shown.

[0155] Figure 7 is a computer readable storage medium structure diagram provided by an embodiment of the present application. As Figure 7 As shown, the computer readable storage medium 600 according to an embodiment of the present application. The computer readable storage medium 600 stores computer readable instructions. When the computer readable instructions are run by the processor, the deep learning based electrical control circuit fault adaptive diagnosis system and method according to the embodiments of the present application described with reference to the above figures can be executed. The storage medium 600 includes, but is not limited to, for example, volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and cache memory, etc. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0156] It should be noted that the above-mentioned order of the embodiments of the present application is only for description, not representing the advantages and disadvantages of the embodiments. And the above describes the specific embodiments of the present application. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or can be advantageous.

[0157] Each of the embodiments in the present specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0158] The above is only the preferred embodiment of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. A deep learning-based adaptive fault diagnosis system for electrical control circuits, characterized in that, include: The data acquisition module includes a high-precision current / voltage sensor array, an infrared thermal imaging unit, and a vibration sensor group, which are used to acquire the operating sequence signals, temperature field distribution data, and mechanical vibration spectrum of the electrical control circuit in real time. The preprocessing module includes a signal denoising unit and a data alignment unit. The signal denoising unit uses a wavelet threshold denoising algorithm to perform nonlinear filtering on the time-series signal and applies an adaptive contrast enhancement algorithm to process the infrared image. The feature extraction module includes an improved Transformer architecture that embeds a temporal convolution operator in the self-attention mechanism to form a temporal encoder with local feature awareness capabilities. The diagnostic model module includes a dual-channel deep residual network and an online incremental learning unit. The dual-channel deep residual network processes electrical features and thermo-mechanical features respectively to obtain fault diagnosis results of electrical control circuits. The adaptive optimization module includes a transfer learning controller and a model structure dynamic adjustment unit. The transfer learning controller selects pre-trained model parameters based on the distribution characteristics of new fault samples, and the dynamic adjustment unit optimizes the network depth and convolution kernel size through a genetic algorithm. The visualization and interaction module, including a fault heatmap generator and a diagnostic decision tree interpreter, can display fault location results in real time and provide interpretability analysis for diagnostic basis.

2. The deep learning-based adaptive fault diagnosis system for electrical control circuits according to claim 1, characterized in that, The data acquisition module specifically includes: A high-precision current / voltage sensor array is used to capture the dynamic timing signals of the electrical control circuit in real time through multi-channel synchronous sampling technology, complete signal filtering, analog-to-digital conversion and data calibration, and encapsulate the original timestamped signals and operating parameters synchronously through a high-speed communication interface and transmit them to the preprocessing module for further processing. The infrared thermal imaging unit scans the surface of electrical equipment in real time, simultaneously captures temperature field distribution data, integrates the time-series temperature matrix with equipment operating parameters, and constructs a dynamic thermal feature dataset which is then streamed to the preprocessing module. The vibration sensor group acquires vibration analog signals in real time and converts them into digital signals. After filtering, amplification and analog-to-digital conversion, time-domain waveform data is generated and transmitted to the preprocessing module.

3. The deep learning-based adaptive fault diagnosis system for electrical control circuits according to claim 1, characterized in that, The preprocessing module specifically includes: The signal denoising unit employs a wavelet threshold denoising algorithm to perform nonlinear filtering on the time-series signal, specifically including: The raw signal from the data acquisition module is decomposed into sub-bands of different frequencies using multi-scale wavelet transform. Approximation coefficients and detail coefficients are extracted for wavelet decomposition, as shown in the formula: in, The original signal, For wavelet basis functions, For scaling function, , are the approximation coefficient and detail coefficient of the j-th layer, respectively, and k is the position index; The extracted approximation coefficients and detail coefficients are used to distinguish between signal and noise, and nonlinear filtering is performed to suppress noise. Reconstruct the denoised signal based on the processed wavelet coefficients; The data alignment unit processes infrared images using an adaptive contrast enhancement algorithm, specifically including: The local mean and standard deviation of the infrared image are calculated by dividing it into blocks; Based on the dynamic design of the gain factor based on the standard deviation, nonlinear contrast stretching is performed on each sub-block to merge overlapping regions and avoid block effect; The standard dynamic range is obtained by global normalization, which enhances the characteristics of the fault region while suppressing noise and over-enhancement.

4. The deep learning-based adaptive fault diagnosis system for electrical control circuits according to claim 1, characterized in that, The feature extraction module specifically includes: The improved Transformer architecture specifically includes: Based on the data processed by the preprocessing module, a one-dimensional causal convolution is performed, with d output channels, to obtain local features. The local features output by the temporal convolution are added element-wise to the original data embedding to form fused features; A query (Q), key (K), and value (V) matrix is ​​generated from the fused data through linear transformation, and the query (Q), key (K), and value (V) matrix is ​​then divided into h heads, each head having a dimension of [missing value]. ; The attention outputs of each head are concatenated and projected back to the original input dimension through a linear layer; Adding the output of multi-head attention to the original input preserves initial information, alleviates gradient vanishing, stabilizes the training process, adjusts the output distribution, and accelerates convergence. Nonlinear transformations are introduced through a feedforward network to refine features position by position. The output of the feedforward network is added to the normalized attention result, and the output distribution is stabilized by layer normalization as the final output of the encoder layer. Multiple encoder layers are repeatedly stacked, and high-order temporal features are extracted layer by layer. The parameters of each layer are independent, and local features are gradually fused.

5. The deep learning-based adaptive fault diagnosis system for electrical control circuits according to claim 1, characterized in that, The diagnostic model module specifically includes: A dual-channel deep residual network model, after acquiring electrical and thermo-mechanical feature data, avoids gradient vanishing in deep networks based on residual connections. For each channel, the following residual block structure is used: in, For convolution operations, For convolution kernel; After processing the electrical and thermo-mechanical characteristics in two channels respectively, the outputs of the two channels are fused, and the initial fault diagnosis results are obtained through the fully connected layer. The online incremental learning unit uses an online gradient descent method to progressively update the model. For each new batch of data, the loss function is updated based on cross-entropy loss, in the form of: in, For real labels, This represents the model's predicted output, where N is the total number of data points. Based on the output of the dual-channel network and the adaptability of the incremental learning unit, the diagnostic results are divided by the probability of the model output to obtain the final electrical control circuit fault diagnosis results and identify the fault type.

6. The deep learning-based adaptive fault diagnosis system for electrical control circuits according to claim 1, characterized in that, The adaptive optimization module specifically includes: A transfer learning controller, which calculates the distribution difference between the new sample and the source domain sample by extracting the feature vector of the new fault sample, and selects a parameter transfer strategy; A model structure dynamic adjustment unit optimizes the structure of the neural network, including network depth and convolution kernel size, through a genetic algorithm. Transfer learning selects a suitable model and fine-tunes it, dynamically adjusts the units to optimize the network structure, and further fine-tunes the model based on new samples and structures until the diagnostic effect reaches the expected level.

7. The deep learning-based adaptive fault diagnosis system for electrical control circuits according to claim 6, characterized in that, The transfer learning controller specifically includes: Construct a pre-trained model library containing multiple device models, and establish a hyperspherical decision boundary for each model in the parameter space; Calculate the Wasserstein distance between the distribution of new fault data and each pre-trained model; Select the three closest models for model parameter interpolation, and determine the interpolation weights based on the reciprocal of the relative distances; During the migration process, the bottom feature extraction layer is frozen, and only the upper classifier is fine-tuned.

8. The deep learning-based adaptive fault diagnosis system for electrical control circuits according to claim 6, characterized in that, The dynamic adjustment unit of the model structure specifically includes: The network deep evolution is performed by automatically increasing or decreasing the number of residual blocks based on the rate of change in validation set accuracy. The decision formula for increasing or decreasing the number of residual blocks is as follows: in, This represents the change in the number of residual blocks. To adjust the coefficient, This represents the difference between the validation set accuracy in the current stage and the previous stage, reflecting the change in model performance. The standard deviation of the validation set accuracy is used to normalize the adjustment amount. The convolution kernel size is optimized by using a genetic algorithm to select from the candidate set through evolution. The number of channels in each layer is dynamically adjusted based on the information entropy value of the feature map.

9. The deep learning-based adaptive fault diagnosis system for electrical control circuits according to claim 1, characterized in that, The visual interaction module specifically includes: The heatmap generator uses gradient activation mapping technology to overlay the fault feature regions of interest to the neural network onto the circuit topology diagram in a color mapping manner, thereby marking the distribution of abnormal current / voltage signals in real time and accurately locating faulty components. The diagnostic decision tree interpreter uses rule extraction and path backtracking to parse the fault classification probability output by the model into an understandable logical reasoning chain, dynamically displaying the decision process of "signal distortion → feature extraction → hierarchical determination", and linking it to a historical fault case library.

10. A deep learning-based adaptive fault diagnosis method for electrical control circuits, characterized in that, The deep learning-based adaptive fault diagnosis system for electrical control circuits as described in any one of claims 1-9 includes: The circuit's timing signals, temperature field distribution, and mechanical vibration spectrum data are simultaneously acquired through a high-precision sensor array, infrared thermal imaging, and vibration sensor group. Wavelet threshold denoising algorithm is used to denoise the time series signal, multi-sensor timing deviation is eliminated by data alignment, and adaptive contrast enhancement is used to optimize infrared image quality. Based on the improved Transformer architecture, a temporal convolution operator is embedded to enhance the local feature perception capability of the self-attention mechanism and extract high-dimensional temporal features; A dual-channel deep residual network is constructed to analyze electrical and thermo-mechanical features separately. The resulting fault classification results are then fused and the model parameters are updated through online incremental learning. Based on the transfer learning controller, the pre-trained parameters are transferred to adapt to the distribution characteristics of new fault samples. The genetic algorithm is used to dynamically optimize the network depth and convolution kernel size to improve the generalization ability. Generate a fault heatmap to locate abnormal areas, and combine it with a decision tree interpreter to provide visual analysis and interpretable reports that provide diagnostic basis.

Images