Drive control circuit with fault detection function

Abstract

The present invention provides a drive control circuit with a fault detection function, which relates to the field of fault detection technology, including: obtaining historical state parameters, performing time-frequency domain analysis through wavelet transform, and constructing a heterogeneous neural network for fault judgment; real-time acquisition of operating parameters, multi-domain feature extraction, and determination of the fault type based on a dynamic time warping algorithm and fuzzy logic reasoning; constructing a time series prediction network with a multi-layer attention mechanism, combining an integrated learning model to output multi-dimensional warning indicators, and generating hierarchical warning information. The present invention can accurately detect and predict faults in the drive control circuit.

Description

Technical Field

[0001] The present invention relates to the technical field of fault detection, and in particular to a drive control circuit with a fault detection function. Background Art

[0002] The existing technology lacks the ability to comprehensively analyze the multi-domain characteristics of the drive control circuit. It usually only focuses on the time domain or frequency domain characteristics, ignoring the interrelationships between different domains, resulting in incomplete feature extraction and difficulty in accurately reflecting the overall state of the system.

[0003] Existing fault diagnosis models generally use a single neural network structure, which is difficult to effectively process the multi-scale and multi-modal data in the drive control circuit, limiting the model's ability to identify complex fault modes, especially when facing new or complex faults, the diagnostic accuracy is low;

[0004] Traditional methods lag behind in fault warning and lack the ability to predict fault development trends and conduct multi-dimensional risk assessments, which results in the inability to detect potential fault risks in a timely manner and makes it difficult to provide effective decision support for equipment maintenance and fault prevention.

[0005] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0006] The embodiment of the present invention provides a drive control circuit with a fault detection function, which can at least solve some of the problems existing in the prior art.

[0007] An embodiment of the present invention provides a drive control circuit with fault detection, comprising:

[0008] Acquire the state parameters of the drive control circuit in the historical working process, calibrate the health index of the drive control circuit based on the change trend of the state parameters, perform time-frequency domain analysis through wavelet transform, extract characteristic parameters in different frequency bands, obtain multi-domain feature vectors, build a multi-layer cascade heterogeneous neural network based on long short-term memory network, convolutional neural network and deep belief network, add the multi-domain feature vectors corresponding to the historical state parameters of the drive control circuit to the heterogeneous neural network, judge whether a fault occurs and the fault mode corresponding to the circuit fault and associate it with the event label, repeatedly judge and iterate the parameters of the heterogeneous neural network, and build a circuit fault detection model based on the model parameters obtained by training;

[0009] Arrange sensors at the detection nodes of the drive control circuit to collect the operating parameters of the drive control circuit in real time, perform multi-domain feature extraction on the operating parameters, obtain real-time feature vectors and add them to the circuit fault detection model to diagnose whether the drive control circuit has a fault risk and the fault type corresponding to the fault risk, perform similarity calculation between the real-time feature vectors and preset fault features, calculate the distance between feature sequences based on a dynamic time warping algorithm, determine the membership of the fault type in combination with fuzzy logic reasoning, update the fault probability distribution through recursive Bayesian estimation, and construct a multidimensional fault feature space. In the multidimensional fault feature space, extract key feature dimensions based on the kernel principal component analysis method, use an improved deep autoencoder to perform feature dimensionality reduction and reconstruction, and dynamically classify fault modes in combination with an adaptive K-means clustering algorithm;

[0010] A time series prediction network with a multi-layer attention mechanism is constructed, and the historical data sequence in the multi-dimensional fault feature space is input into the time series prediction network. The long-range dependency between features is captured by the self-attention layer, and the dynamic change law corresponding to the historical data sequence is extracted in combination with the gated recurrent unit. The fault state transition probability is estimated based on the variational inference method to obtain the fault development trend prediction result, which is input into an integrated learning model based on random forest and gradient boosting tree, and multi-dimensional warning indicators are output. An adaptive warning threshold is designed based on the multi-dimensional warning indicators, and a hierarchical warning rule is constructed in combination with an expert knowledge base to generate hierarchical warning information. The hierarchical warning information is pushed to a preset terminal through an encrypted channel, and the warning process data is stored in a distributed database.

[0011] In an optional embodiment,

[0012] The state parameters of the drive control circuit in the historical working process are obtained, and based on the change trend of the state parameters, the health index of the drive control circuit is calibrated and time-frequency domain analysis is performed through wavelet transform, and characteristic parameters in different frequency bands are extracted to obtain multi-domain feature vectors. A multi-layer cascade heterogeneous neural network is constructed based on a long short-term memory network, a convolutional neural network and a deep belief network. The multi-domain feature vector corresponding to the historical state parameters of the drive control circuit is added to the heterogeneous neural network, and it is determined whether a fault occurs and the fault mode corresponding to the circuit fault is associated with the event label, and the parameters of the heterogeneous neural network are repeatedly determined and iterated. The circuit fault detection model is constructed based on the model parameters obtained by training, including:

[0013] Acquire state parameters of the drive control circuit during historical operation, wherein the state parameters include voltage parameters, current parameters, temperature parameters, power parameters, and performance parameters;

[0014] Based on the change trend of the state parameter, a multi-layer health evaluation system is used to calibrate the health index of the drive control circuit, wherein the multi-layer health evaluation system includes a single index evaluation layer, a combined index evaluation layer and a comprehensive health evaluation layer;

[0015] The state parameters are analyzed in the time and frequency domains by using an improved wavelet packet transform algorithm, eight-scale decomposition is performed by using an optimal basis selection algorithm and the sym6 wavelet basis function, noise is suppressed by an adaptive threshold method, characteristic parameters are extracted from each sub-band obtained by decomposition, the extracted mean and standard deviation are used as time domain features, the band energy and centroid frequency are used as frequency domain features, and the wavelet energy entropy and singular value distribution are used as time-frequency joint features to construct a multi-domain feature vector;

[0016] Constructing a multi-layer cascade heterogeneous neural network including a feature extraction layer, a feature fusion layer and a decision layer, wherein the feature extraction layer includes a first branch consisting of a three-layer bidirectional long short-term memory network, a second branch consisting of a convolutional neural network with a channel attention mechanism and a spatial attention mechanism, and a third branch consisting of a deep belief network introducing a contrastive divergence algorithm;

[0017] Input the multi-domain feature vector into the heterogeneous neural network, extract features through the first branch, the second branch and the third branch respectively, use the attention mechanism in the feature fusion layer to weight the features extracted by the three branches, use the feature selection algorithm to remove redundant features, perform deep feature fusion through the multi-head self-attention mechanism, input the fused features into the support vector machine classifier, the random forest classifier and the XGBoost classifier in the decision layer, use the ensemble learning method to dynamically weight and fuse the output results of the three classifiers, determine whether a fault occurs and the fault mode corresponding to the circuit fault, and associate the fault mode with the event label;

[0018] The periodic change of learning rate is controlled by a cyclic learning rate adjustment mechanism, a hybrid loss function is constructed by combining cross entropy loss and focal loss, the batch size is adaptively adjusted according to the change of training loss value, the parameters of the heterogeneous neural network are continuously iterated and optimized through an incremental learning mechanism, and a circuit fault detection model is constructed based on the model parameters obtained through training.

[0019] In an optional embodiment,

[0020] Combining cross entropy loss and focal loss to construct a hybrid loss function is shown in the following formula:

[0021] ;

[0022] in, L mix represents the mixed loss value, N represents the total number of samples,yes Indicates i The true labels of samples, Indicates i The predicted probability of a sample, α represents the contribution weight coefficient, c Represents the focus parameter.

[0023] In an optional embodiment,

[0024] Arranging sensors at detection nodes of the drive control circuit, collecting operating parameters of the drive control circuit in real time, performing multi-domain feature extraction on the operating parameters, obtaining real-time feature vectors and adding them to the circuit fault detection model includes:

[0025] Arrange various types of sensors at different detection nodes of the drive control circuit, including voltage sensors, current sensors, temperature sensors and vibration sensors, measure the bus voltage through the voltage sensor connected in parallel to the DC bus, determine the current signal through the Hall current sensor, measure the circuit temperature through the thermocouple temperature sensor, and measure the vibration signal through the acceleration sensor, and use the collected information as the multi-source heterogeneous operation parameters of the drive control circuit;

[0026] For the voltage data and current data in the operating parameters, the multi-layer wavelet packet is used to decompose them into multiple sub-bands, and the energy, mean and root mean square corresponding to each sub-band are extracted as non-stationary eigenvectors. For the temperature parameters, the principal component analysis algorithm is used to perform centering and normalization processing, and the covariance matrix and the eigenvalues ​​of the covariance matrix are calculated at the same time. Feature selection is performed according to the cumulative contribution rate to obtain the principal component features and construct the principal component matrix, and finally a slowly varying feature matrix is ​​obtained. For the vibration signal, the original vibration signal is decomposed into intrinsic mode functions and residuals by empirical mode decomposition, and each intrinsic mode function is subjected to Hilbert transformation to obtain the instantaneous phase and instantaneous amplitude, and the instantaneous frequency standard deviation and the instantaneous amplitude mean are extracted as vibration features. The vibration features, the non-stationary eigenvector and the slowly varying feature matrix are combined to obtain a real-time feature vector and add the real-time feature vector to a preset circuit fault detection model;

[0027] Common faults are applied to the drive control circuit, wherein the common faults include open circuit, breakdown, capacitor short circuit and phase loss, an oscilloscope is used to record the voltage and current corresponding to each fault as fault waveform characteristics, an infrared camera is used to record the device surface temperature distribution, the fault level corresponding to the degree of device damage is determined through thermal imaging, and a fault label is determined based on the fault level, the fault waveform characteristics and the device surface temperature distribution.

[0028] In an optional embodiment,

[0029] Diagnose whether the drive control circuit has a fault risk and the fault type corresponding to the fault risk, calculate the similarity between the real-time feature vector and the preset fault feature, calculate the distance between the feature sequences based on the dynamic time warping algorithm, determine the membership of the fault type in combination with fuzzy logic reasoning, update the fault probability distribution through recursive Bayesian estimation, and construct a multidimensional fault feature space. In the multidimensional fault feature space, extract key feature dimensions based on the kernel principal component analysis method, use an improved deep autoencoder to perform feature dimensionality reduction and reconstruction, and combine the adaptive K-means clustering algorithm to dynamically classify the fault mode, including:

[0030] The circuit fault detection model includes a one-dimensional convolutional neural network and a long short-term memory network. Based on the real-time feature vector, local spatial feature extraction is performed through a one-dimensional convolutional layer, and the activation function and maximum pooling downsampling are combined to cascade two long short-term memory neural networks to learn the long-term dependency in the real-time feature vector, and obtain the probability distribution of the real-time feature vector in ten fault states, and use the probability distribution as the fault risk and the corresponding fault type corresponding to the drive control circuit;

[0031] For each fault type, in combination with the pre-acquired fault label, the similarity between the real-time feature vector and the fault features in the pre-set historical fault library is calculated, the distance between the real-time feature vector and the feature sequence in the historical fault library is calculated by a dynamic regularization algorithm, the optimal matching path is generated and used as the similarity, a fuzzy set is defined, and an inference rule is formulated and inference is performed based on the similarity to obtain a membership function corresponding to the current fault type, based on the membership function, the dynamic probability distribution of the fault risk and the membership function are updated by a recursive Bayesian estimation algorithm, and the multi-dimensional fault feature space is repeatedly updated and constructed;

[0032] In the multidimensional fault feature space, the key feature dimensions corresponding to each fault type and the real-time feature vector are extracted by the kernel principal component analysis algorithm, and the feature dimension reduction and reconstruction are performed in combination with the improved deep autoencoder to generate deep fault features. For the deep fault features, clustering is performed in the multidimensional fault feature space by an adaptive K-means clustering algorithm combined with a clustering effectiveness index to obtain a dynamic fault classification result.

[0033] In an optional embodiment,

[0034] Construct a time series prediction network with a multi-layer attention mechanism, input the historical data sequence in the multi-dimensional fault feature space into the time series prediction network, capture the long-range dependency between features through the self-attention layer, extract the dynamic change law of different fault features in combination with the gated recurrent unit, estimate the fault state transition probability based on the variational inference method, and obtain the fault development trend prediction result, input the fault development trend prediction result into an integrated learning model based on random forest and gradient boosting tree, output multi-dimensional warning indicators, design an adaptive warning threshold based on the multi-dimensional warning indicators, construct hierarchical warning rules in combination with the expert knowledge base, generate hierarchical warning information, push the hierarchical warning information to the preset terminal through an encrypted channel, and store the warning process data in a distributed database, including:

[0035] Construct a time series prediction network including a multi-layer attention mechanism, add the historical data sequence in the multi-dimensional fault feature space to the time series prediction network, combine the pre-acquired dynamic fault classification results, determine the long-range dependency between different fault features through the self-attention layer, based on the long-range dependency, combine the gated recurrent unit to determine the hidden state of the current time step and update the state, repeat the update and cyclic encoding until the last time step is reached, and obtain the feature encoding corresponding to the hidden state of the last time step and the dynamic change law corresponding to the historical data sequence;

[0036] Based on the feature coding, latent variables are defined by a variational inference method, and the posterior distribution is approximated in combination with the variational distribution. The variational lower bound is maximized with the goal of maximizing the logarithmic marginal likelihood of the observed data. The variational distribution is parameterized by a neural network in combination with the prior distribution of the latent variables that satisfies the normal distribution, and the sampling values ​​of the latent variables are obtained. The observed data are reconstructed by generating a network and the latent variables are sampled by introducing random noise variables. The state transition probability within a unit time interval is obtained in combination with the loss function. The latent variable encoding is obtained in combination with the posterior distribution sampling and the state transfer function is recursively applied to obtain the latent state sequence and decode it to obtain the fault development trend prediction result.

[0037] The fault development trend prediction result is added to the integrated learning model based on random forest and gradient boosting tree. By integrating the prediction results of random forest module and gradient boosting tree module, a multidimensional early warning index is obtained. For each dimension of the multidimensional early warning index, the relative deviation is determined according to the distribution frequency of the historical data sequence in different intervals. The corresponding adaptive early warning threshold is set based on the relative deviation. A hierarchical early warning rule is constructed in combination with the pre-acquired expert knowledge base. For the current operating parameters, the corresponding hierarchical early warning information is generated and pushed to the pre-set early warning terminal through an encrypted channel, and the early warning process data is stored in a pre-set distributed database.

[0038] In an optional embodiment,

[0039] The calculation formula corresponding to the variational lower bound is as follows:

[0040] ;

[0041] Among them, ELBO( i , f ) represents the value of the variational lower bound, i Describes the decoder parameters, f represents the encoder parameters, is the variational posterior distribution, which means that given the observed data x Then, the hidden variable z The distribution of To generate a model, it means that given latent variables z After that, the model generates observation data x The probability of D KL Represents KL divergence, which is used to measure the variational posterior distribution With prior distribution p ( z ), p ( z ) represents the prior distribution of latent variables, Indicates that observations are generated under the posterior distribution x The log-likelihood of .

[0042] In the present invention, by performing wavelet transform on the state parameters of the drive control circuit, time-frequency domain feature extraction is realized, and a multi-domain feature vector is obtained, which can more comprehensively characterize the health state of the circuit and provide rich information for subsequent fault diagnosis. By arranging sensors at the detection nodes to collect operating parameters in real time and using the trained fault detection model for diagnosis, it is possible to accurately determine whether the circuit has a fault risk and the corresponding fault type. Combined with similarity calculation, dynamic time warping, fuzzy logic reasoning and other methods, the fault probability distribution can be dynamically updated to achieve adaptive classification of fault modes. A timing prediction network with a multi-layer attention mechanism is introduced, and the long-range dependency between features is captured through the self-attention layer. The dynamic change law of historical data is extracted by combining the gated recurrent unit, and the variational inference method is used to estimate the fault state transition probability, thereby predicting the development trend of the fault. Measures can be taken in time before the fault occurs to avoid serious accidents. In summary, the present invention realizes the full process intelligence of the drive control circuit from fault diagnosis, prediction to early warning, which can significantly improve the system reliability and safety, reduce unexpected downtime and maintenance costs, and has important application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1A schematic diagram of the operation flow of a drive control circuit with a fault detection function according to an embodiment of the present invention; DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0046] Figure 1 FIG. 1 is a schematic diagram of the operation flow of a drive control circuit with a fault detection function according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0047] S1. Obtain the state parameters of the drive control circuit in the historical working process, calibrate the health index of the drive control circuit based on the change trend of the state parameters, and perform time-frequency domain analysis through wavelet transform, extract characteristic parameters in different frequency bands, obtain multi-domain feature vectors, build a multi-layer cascade heterogeneous neural network based on long short-term memory network, convolutional neural network and deep belief network, add the multi-domain feature vector corresponding to the historical state parameters of the drive control circuit to the heterogeneous neural network, judge whether a fault occurs and the fault mode corresponding to the circuit fault and associate it with the event label, repeatedly judge and iterate the parameters of the heterogeneous neural network, and build a circuit fault detection model based on the model parameters obtained by training;

[0048] The time-frequency domain analysis is a signal processing technology that considers the characteristics of the signal in time and frequency at the same time in order to more comprehensively understand the changes in non-stationary signals. It is often used in the processing of vibration signals and sound signals. The deep belief network is a deep learning model composed of multiple stacked restricted Boltzmann machines. It learns data features through layer-by-layer unsupervised training and is widely used in fields such as image processing and speech recognition. The failure mode refers to the description of abnormal states or behaviors that may occur in a device or system during operation, including the type of failure, its manifestation and its consequences, which helps to identify and manage potential problems in advance. The event label is a label used to identify and classify specific events or states in data analysis. By assigning labels to observations in a data set, fault detection, data mining and machine learning tasks can be effectively trained and evaluated.

[0049] In an optional embodiment,

[0050] The state parameters of the drive control circuit in the historical working process are obtained, and based on the change trend of the state parameters, the health index of the drive control circuit is calibrated and time-frequency domain analysis is performed through wavelet transform, and characteristic parameters in different frequency bands are extracted to obtain multi-domain feature vectors. A multi-layer cascade heterogeneous neural network is constructed based on a long short-term memory network, a convolutional neural network and a deep belief network. The multi-domain feature vector corresponding to the historical state parameters of the drive control circuit is added to the heterogeneous neural network, and it is determined whether a fault occurs and the fault mode corresponding to the circuit fault is associated with the event label, and the parameters of the heterogeneous neural network are repeatedly determined and iterated. The circuit fault detection model is constructed based on the model parameters obtained by training, including:

[0051] Acquire state parameters of the drive control circuit during historical operation, wherein the state parameters include voltage parameters, current parameters, temperature parameters, power parameters, and performance parameters;

[0052] Based on the change trend of the state parameter, a multi-layer health evaluation system is used to calibrate the health index of the drive control circuit, wherein the multi-layer health evaluation system includes a single index evaluation layer, a combined index evaluation layer and a comprehensive health evaluation layer;

[0053] The state parameters are analyzed in the time and frequency domains by using an improved wavelet packet transform algorithm, eight-scale decomposition is performed by using an optimal basis selection algorithm and the sym6 wavelet basis function, noise is suppressed by an adaptive threshold method, characteristic parameters are extracted from each sub-band obtained by decomposition, the extracted mean and standard deviation are used as time domain features, the band energy and centroid frequency are used as frequency domain features, and the wavelet energy entropy and singular value distribution are used as time-frequency joint features to construct a multi-domain feature vector;

[0054] Constructing a multi-layer cascade heterogeneous neural network including a feature extraction layer, a feature fusion layer and a decision layer, wherein the feature extraction layer includes a first branch consisting of a three-layer bidirectional long short-term memory network, a second branch consisting of a convolutional neural network with a channel attention mechanism and a spatial attention mechanism, and a third branch consisting of a deep belief network introducing a contrastive divergence algorithm;

[0055] Input the multi-domain feature vector into the heterogeneous neural network, extract features through the first branch, the second branch and the third branch respectively, use the attention mechanism in the feature fusion layer to weight the features extracted by the three branches, use the feature selection algorithm to remove redundant features, perform deep feature fusion through the multi-head self-attention mechanism, input the fused features into the support vector machine classifier, the random forest classifier and the XGBoost classifier in the decision layer, use the ensemble learning method to dynamically weight and fuse the output results of the three classifiers, determine whether a fault occurs and the fault mode corresponding to the circuit fault, and associate the fault mode with the event label;

[0056] The periodic change of learning rate is controlled by a cyclic learning rate adjustment mechanism, a hybrid loss function is constructed by combining cross entropy loss and focal loss, the batch size is adaptively adjusted according to the change of training loss value, the parameters of the heterogeneous neural network are continuously iterated and optimized through an incremental learning mechanism, and a circuit fault detection model is constructed based on the model parameters obtained through training.

[0057] The multi-layer health assessment system is a layered method used to comprehensively evaluate the health status of equipment or systems. It provides more accurate health assessment results through comprehensive analysis of indicators and models at different levels. The optimal basis selection algorithm is an algorithm used to select the best basis function in the feature selection process, aiming to improve the performance and efficiency of the model, reduce the dimension and enhance the generalization ability of the model by evaluating the effectiveness of different feature combinations. The eight-scale decomposition is a signal processing method that decomposes the signal into eight different scales to more deeply analyze its characteristics in different frequency and time ranges. It is often used for fault diagnosis and feature extraction. The cyclic learning rate adjustment mechanism is a strategy for dynamically adjusting the learning rate. It optimizes the convergence speed and stability of the model by continuously changing the learning rate during the training process, thereby improving the training effect of the deep learning model.

[0058] Obtain the state parameters of the drive control circuit during the historical working process, including voltage, current, temperature, power, performance and other parameters. Based on the change trend of the state parameters, a multi-layer health assessment system is used to calibrate the health of the drive control circuit. The assessment system includes a single indicator assessment layer, a combined indicator assessment layer and a comprehensive health assessment layer.

[0059] At the single indicator evaluation layer, the single parameters such as voltage, current, and temperature are evaluated respectively to obtain the health score of each parameter. For example, the health of the voltage parameter can be evaluated based on factors such as its stability and fluctuation range. If the voltage fluctuation is within ±5%, the score is 90 points, ±5%~±10% score is 80 points, and so on.

[0060] At the combined indicator evaluation layer, related single indicators are combined for comprehensive evaluation, such as combining voltage and current parameters to evaluate power health, and combining temperature and performance parameters to evaluate thermal management health;

[0061] At the comprehensive health evaluation layer, the overall health score of the drive control circuit is calculated based on the weights of each combined indicator. For example, the power health weight can be set to 0.4, the thermal management health weight can be set to 0.3, and the weights of other combined indicators can be set to 0.3. The final health score is obtained by weighted average.

[0062] The improved wavelet packet transform algorithm is used to analyze the state parameters in the time and frequency domains. The optimal basis selection algorithm is used to determine the optimal decomposition level. The sym6 wavelet basis function is selected for eight-scale decomposition. The adaptive threshold method is used to suppress noise and improve the signal-to-noise ratio of the signal. Feature parameters are extracted from each sub-band, including mean and standard deviation as time domain features, band energy and centroid frequency as frequency domain features, wavelet energy entropy and singular value distribution as time-frequency joint features, and finally a multi-domain feature vector is constructed.

[0063] For example, after wavelet packet decomposition of the voltage parameter, 256 sub-bands are obtained. The mean, standard deviation, band energy and other features of each sub-band are calculated to obtain a 1280-dimensional feature vector. The other parameters are processed in the same way, and finally a high-dimensional multi-domain feature vector is obtained. A multi-layer cascaded heterogeneous neural network is constructed, including a feature extraction layer, a feature fusion layer and a decision layer. The feature extraction layer consists of three parallel branches. The first branch is a three-layer bidirectional long short-term memory network for capturing temporal features. The second branch is a convolutional neural network with a channel attention mechanism and a spatial attention mechanism for extracting spatial features. The third branch is a deep belief network that introduces a contrastive divergence algorithm for learning deep feature representation.

[0064] The multi-domain feature vectors are input into the heterogeneous neural network for feature extraction. In the bidirectional long short-term memory network branch, the number of hidden layer neurons is set to 128, and the dropout rate is 0.5. In the convolutional neural network branch, three convolutional layers are used, and the convolution kernel sizes are 3x3, 5x5, and 7x7, respectively, and the number of channels is 64, 128, and 256. In the deep belief network branch, four hidden layers are set, and the number of neurons is 1024, 512, 256, and 128, respectively.

[0065] In the feature fusion layer, the attention mechanism is used to weight the features extracted by the three branches, calculate the importance score of each branch feature, and perform weighted summation of the features according to the score. The LASSO algorithm is used for feature selection to remove redundant features. The multi-head self-attention mechanism is used for deep feature fusion. Eight attention heads are set, and the dimension of each head is 64.

[0066] In the decision layer, the fused features are input into three classifiers, namely support vector machine (SVM), random forest (RF) and XGBoost. The support vector machine uses the RBF kernel function, the penalty factor C is set to 1, the number of trees of the random forest is set to 100, the maximum depth is 10, the learning rate of XGBoost is set to 0.1, and the maximum depth is 6. The output results of the three classifiers are dynamically fused using the weighted voting ensemble learning method. The weights are dynamically adjusted according to the accuracy of each classifier to determine whether a fault occurs and the specific mode corresponding to the fault, and the fault mode is associated with the event label.

[0067] During the training process, a cyclic learning rate adjustment mechanism is used to control the periodic change of the learning rate. The learning rate changes cyclically in the range of [0.0001, 0.1]. Each cycle includes three stages: warm-up, hold and decay. A hybrid loss function is constructed by combining cross entropy loss and focal loss. The focus parameter γ of the focal loss is set to 2. The batch size is adaptively adjusted according to the change of the training loss value. When the loss decreases slowly, the batch size is increased, otherwise it is reduced. The network parameters are continuously iterated and optimized through the incremental learning mechanism. Each time a new sample is added, only the last few layers of the network are fine-tuned, and the parameters of the previous layers are kept unchanged.

[0068] After multiple rounds of training, a circuit fault detection model is constructed based on the obtained model parameters. The model can monitor the operating status of the drive control circuit in real time, detect potential faults in time, and give specific fault modes.

[0069] In this embodiment, by obtaining state parameters of multiple dimensions such as voltage, current, temperature, power and performance, and calibrating them using a multi-layer health evaluation system, the health status of the drive control circuit can be comprehensively and systematically characterized, providing a reliable data basis for fault diagnosis. The state parameters are analyzed in the time and frequency domains using an improved wavelet packet transform algorithm. Through optimal basis selection, eight-scale decomposition, adaptive threshold denoising and other processing, the time domain, frequency domain and time-frequency joint features are extracted to construct a multi-domain feature vector, which can capture the dynamic changes and hidden patterns of the circuit state and enhance the richness and effectiveness of feature representation. By integrating the decision results of multiple classifiers through an integrated learning method, the advantages of different models can be taken into account, and the accuracy and robustness of fault diagnosis can be improved. The batch size can be adaptively adjusted according to the change in the training loss value, and the computational efficiency and learning effect can be dynamically balanced. The incremental learning mechanism supports continuous optimization and updating of the model, improves the adaptability and life cycle of the method. In summary, this embodiment realizes dynamic monitoring and diagnosis of the health status of the circuit. Through continuous optimization and incremental learning, it has good adaptability and scalability, can cope with changes in the circuit state in a complex environment, and provides a strong guarantee for the reliable operation of the system.

[0070] In an optional embodiment,

[0071] Combining cross entropy loss and focal loss to construct a hybrid loss function is shown in the following formula:

[0072] ;

[0073] in, L mix represents the mixed loss value, N represents the total number of samples, yes Indicates i The true labels of samples, Indicates i The predicted probability of a sample, α represents the contribution weight coefficient, c Represents the focus parameter.

[0074] In this embodiment, the contribution weight coefficient is introduced to balance the contribution of positive and negative samples to the loss function and improve the sensitivity of the model to minority categories. The cross entropy loss is conducive to improving the classification accuracy, while the focal loss helps to enhance the robustness of the model. The combination of the two loss functions can ensure the classification performance while improving the model's ability to deal with abnormal samples and adversarial samples, so that it has better generalization and adaptability. Two hyperparameters are introduced in the hybrid loss function, which can be adjusted according to the specific tasks and data characteristics to achieve the optimal training effect. It can be applied to various scenarios and problems. In summary, this embodiment can effectively improve the model training effect under the problem of category imbalance, improve the recognition rate of minority categories, accelerate the convergence speed, and enhance the generalization ability and adaptability of the model.

[0075] S2. Arrange sensors at the detection nodes of the drive control circuit, collect the operating parameters of the drive control circuit in real time, perform multi-domain feature extraction on the operating parameters, obtain real-time feature vectors and add them to the circuit fault detection model, diagnose whether the drive control circuit has a fault risk and the fault type corresponding to the fault risk, calculate the similarity between the real-time feature vectors and preset fault features, calculate the distance between feature sequences based on a dynamic time warping algorithm, determine the membership of the fault type in combination with fuzzy logic reasoning, update the fault probability distribution through recursive Bayesian estimation, and construct a multidimensional fault feature space. In the multidimensional fault feature space, extract key feature dimensions based on the kernel principal component analysis method, use an improved deep autoencoder to perform feature dimensionality reduction and reconstruction, and dynamically classify fault modes in combination with an adaptive K-means clustering algorithm;

[0076] The multi-domain feature extraction is a technology that extracts features from different data domains (such as time domain, frequency domain and time-frequency domain) to comprehensively analyze the characteristics of signals or data and improve the recognition and classification capabilities of the model. The dynamic time warping algorithm is an algorithm for comparing different time series. It minimizes the distance between time series through nonlinear alignment to keep them consistent in time. It is often used in speech recognition and pattern recognition. The recursive Bayesian estimation is an estimation method that combines observed data with continuously updated prior knowledge. It is suitable for state estimation of dynamic systems and can reflect system state changes in real time. The multi-dimensional fault feature space refers to a space containing multiple feature dimensions constructed in fault diagnosis, which can comprehensively describe the health status of the equipment and help identify and analyze fault modes. The kernel principal component analysis method is an extended principal component analysis technology that maps data to a high-dimensional space using a kernel function to capture nonlinear relationships and extract main features.

[0077] In an optional embodiment,

[0078] Arranging sensors at detection nodes of the drive control circuit, collecting operating parameters of the drive control circuit in real time, performing multi-domain feature extraction on the operating parameters, obtaining real-time feature vectors and adding them to the circuit fault detection model includes:

[0079] Arrange various types of sensors at different detection nodes of the drive control circuit, including voltage sensors, current sensors, temperature sensors and vibration sensors, measure the bus voltage through the voltage sensor connected in parallel to the DC bus, determine the current signal through the Hall current sensor, measure the circuit temperature through the thermocouple temperature sensor, and measure the vibration signal through the acceleration sensor, and use the collected information as the multi-source heterogeneous operation parameters of the drive control circuit;

[0080] For the voltage data and current data in the operating parameters, the multi-layer wavelet packet is used to decompose them into multiple sub-bands, and the energy, mean and root mean square corresponding to each sub-band are extracted as non-stationary eigenvectors. For the temperature parameters, the principal component analysis algorithm is used to perform centering and normalization processing, and the covariance matrix and the eigenvalues ​​of the covariance matrix are calculated at the same time. Feature selection is performed according to the cumulative contribution rate to obtain the principal component features and construct the principal component matrix, and finally a slowly varying feature matrix is ​​obtained. For the vibration signal, the original vibration signal is decomposed into intrinsic mode functions and residuals by empirical mode decomposition, and each intrinsic mode function is subjected to Hilbert transformation to obtain the instantaneous phase and instantaneous amplitude, and the instantaneous frequency standard deviation and the instantaneous amplitude mean are extracted as vibration features. The vibration features, the non-stationary eigenvector and the slowly varying feature matrix are combined to obtain a real-time feature vector and add the real-time feature vector to a preset circuit fault detection model;

[0081] Common faults are applied to the drive control circuit, wherein the common faults include open circuit, breakdown, capacitor short circuit and phase loss, an oscilloscope is used to record the voltage and current corresponding to each fault as fault waveform characteristics, an infrared camera is used to record the device surface temperature distribution, the fault level corresponding to the degree of device damage is determined through thermal imaging, and a fault label is determined based on the fault level, the fault waveform characteristics and the device surface temperature distribution.

[0082] The cumulative contribution rate refers to the proportion of the variance explained by each principal component to the total variance in principal component analysis, which is used to determine the number of principal components to be selected to ensure that sufficient information is retained. The principal component matrix is ​​a matrix generated by principal component analysis, in which each column represents a principal component, contains the main features of the original data, and is used for dimensionality reduction and feature extraction. The Hilbert transform is a signal processing technology used to extract instantaneous frequency and instantaneous amplitude from real signals, and analyze signal characteristics by generating an analytical form of the signal. The instantaneous phase refers to the phase information of the signal at a given moment, which reflects the phase change of the signal and can be used to analyze the modulation characteristics and change trend of the signal. The fault label is a label used to identify and classify specific faults in the fault diagnosis process, helping the model to identify different fault modes in training and prediction;

[0083] Arrange various types of sensors at different detection nodes of the drive control circuit, including voltage sensors, current sensors, temperature sensors and vibration sensors. Specifically, connect the voltage sensor in parallel to the DC bus to measure the bus voltage, install Hall current sensors at the input and output ends of the circuit to determine the current signal, paste thermocouple temperature sensors on the surface of the circuit board and the surface of key components to measure the circuit temperature, and fix the acceleration sensor on the edge of the circuit board to measure the vibration signal. These sensors collect circuit operation parameters in real time to form multi-source heterogeneous data;

[0084] The collected operating parameters are subjected to multi-domain feature extraction, and the voltage and current data are processed. The signal is decomposed into 5-layer wavelet packets using the db4 wavelet basis function to obtain 32 sub-bands. The energy, mean, and root mean square are calculated for each sub-band to form a 96-dimensional non-stationary feature vector. For example, at a certain moment, the energy of the first sub-band of the voltage signal is 0.5, the mean is 220V, and the root mean square is 225V.

[0085] For the temperature parameter, the original data is centered, that is, the mean is subtracted, and normalized, and the data is scaled to the interval [-1, 1]. The covariance matrix of the normalized data is calculated, and the eigenvalues ​​of the matrix are solved. The principal components are selected according to the principle of 95% cumulative contribution rate, and the principal component matrix is ​​constructed. For example, the first three principal components are selected to obtain a 3x1 principal component matrix as a slowly varying characteristic matrix;

[0086] For vibration signals, the original signal is decomposed into 10 intrinsic mode functions and 1 residual using the empirical mode decomposition method. Each intrinsic mode function is subjected to Hilbert transform to obtain the instantaneous phase and instantaneous amplitude. The instantaneous frequency standard deviation and instantaneous amplitude mean of each intrinsic mode function are calculated to obtain a 20-dimensional vibration eigenvector. For example, the instantaneous frequency standard deviation of the first intrinsic mode function is 0.2, and the instantaneous amplitude mean is 1.5.

[0087] The non-stationary eigenvector, slowly varying eigenmatrix and vibration eigenvector are combined to obtain a 119-dimensional real-time eigenvector, and the real-time eigenvector is added to a pre-set circuit fault detection model for subsequent fault diagnosis;

[0088] Common faults are imposed on the drive control circuit, including open circuit, breakdown, capacitor short circuit and phase loss. An oscilloscope is used to record the voltage and current waveforms under each fault. For example, in the case of an open circuit fault, the voltage waveform is spike-shaped and the current waveform is intermittent. An infrared camera is used to take thermal images of the circuit board and record the surface temperature distribution of the device. For example, in the case of a breakdown fault, the temperature of a power device reaches 120°C, which is significantly higher than the normal operating temperature.

[0089] The degree of device damage is determined based on the thermal imaging image, and the fault level is divided into three levels: slight, moderate and severe. The fault label is determined based on the fault waveform characteristics and temperature distribution. For example, in an experiment, voltage spikes and current discontinuities were observed, and the power device temperature reached 120°C, which was determined to be a severe open circuit fault.

[0090] In this embodiment, the method of acquiring multi-source heterogeneous data can reflect the real-time working status of the circuit from different dimensions and provide a rich information source for fault diagnosis. Principal component analysis can remove noise and redundant information in temperature data, extract low-dimensional features that best characterize temperature change trends, and reflect the long-term health status of the circuit. The fusion of multi-source heterogeneous features can describe the circuit status from different time and space scales, improve the accuracy and reliability of fault diagnosis, and construct fault labels based on fault waveforms, temperature distributions, and fault levels. It can provide high-quality sample data for model training and improve the accuracy and comprehensiveness of fault diagnosis. In summary, this embodiment can significantly improve the health monitoring and fault diagnosis level of the drive control circuit, reduce false alarms and missed alarms, and provide reliable protection for the safe and stable operation of the system.

[0091] In an optional embodiment,

[0092] Diagnose whether the drive control circuit has a fault risk and the fault type corresponding to the fault risk, calculate the similarity between the real-time feature vector and the preset fault feature, calculate the distance between the feature sequences based on the dynamic time warping algorithm, determine the membership of the fault type in combination with fuzzy logic reasoning, update the fault probability distribution through recursive Bayesian estimation, and construct a multidimensional fault feature space. In the multidimensional fault feature space, extract key feature dimensions based on the kernel principal component analysis method, use an improved deep autoencoder to perform feature dimensionality reduction and reconstruction, and combine the adaptive K-means clustering algorithm to dynamically classify the fault mode, including:

[0093] The circuit fault detection model includes a one-dimensional convolutional neural network and a long short-term memory network. Based on the real-time feature vector, local spatial feature extraction is performed through a one-dimensional convolutional layer, and the activation function and maximum pooling downsampling are combined to cascade two long short-term memory neural networks to learn the long-term dependency in the real-time feature vector, and obtain the probability distribution of the real-time feature vector in ten fault states, and use the probability distribution as the fault risk and the corresponding fault type corresponding to the drive control circuit;

[0094] For each fault type, in combination with the pre-acquired fault label, the similarity between the real-time feature vector and the fault features in the pre-set historical fault library is calculated, the distance between the real-time feature vector and the feature sequence in the historical fault library is calculated by a dynamic regularization algorithm, the optimal matching path is generated and used as the similarity, a fuzzy set is defined, and an inference rule is formulated and inference is performed based on the similarity to obtain a membership function corresponding to the current fault type, based on the membership function, the dynamic probability distribution of the fault risk and the membership function are updated by a recursive Bayesian estimation algorithm, and the multi-dimensional fault feature space is repeatedly updated and constructed;

[0095] In the multidimensional fault feature space, the key feature dimensions corresponding to each fault type and the real-time feature vector are extracted by the kernel principal component analysis algorithm, and the feature dimension reduction and reconstruction are performed in combination with the improved deep autoencoder to generate deep fault features. For the deep fault features, clustering is performed in the multidimensional fault feature space by an adaptive K-means clustering algorithm combined with a clustering effectiveness index to obtain a dynamic fault classification result.

[0096] The maximum pooling downsampling is a dimensionality reduction technology commonly used in convolutional neural networks. It reduces the data dimension by selecting the maximum value in the feature map, thereby reducing the computational complexity and retaining the main features. The fuzzy set is a mathematical concept used to deal with uncertainty and ambiguity. It represents the degree of an element in a set by defining a membership function. It is commonly used in fuzzy logic and fuzzy control. The clustering effectiveness index is an indicator used to evaluate the quality of clustering results. Common ones include silhouette coefficient, Davies-Bouldin index, etc., which help judge the effect and rationality of clustering.

[0097] Collect the real-time operation data of the drive control circuit, including voltage, current, temperature and other parameters, pre-process and extract features of the collected data to obtain real-time feature vectors;

[0098] The real-time feature vector is input into a pre-trained circuit fault detection model, including a one-dimensional convolutional neural network and a long short-term memory network. The one-dimensional convolutional neural network contains three one-dimensional convolutional layers. The convolution kernel sizes of each layer are 3, 5, and 7, respectively, and the number of convolution kernels is 32, 64, and 128. Each convolutional layer is followed by a ReLU activation function and a maximum pooling layer with a pooling size of 2. The output feature map of the convolutional network is flattened and input into two cascaded long short-term memory networks. Each long short-term memory network contains 128 hidden units. Finally, through a fully connected layer and a Softmax function, the probability distribution of the real-time feature vector under 10 predefined fault states is output, and the fault type with the highest probability is taken as the fault risk judgment result of the current drive control circuit.

[0099] For each fault type, the similarity between the real-time feature vector and the fault features in the pre-established historical fault library is calculated, and the distance between the real-time feature vector and the historical fault feature sequence is calculated using the dynamic time warping algorithm. Specifically, a cumulative distance matrix is ​​constructed, and the optimal matching path is found through dynamic programming. The cumulative distance of the path is the similarity of the two sequences. Fuzzy sets are defined, and fuzzy reasoning rules are formulated based on the calculated similarity. For example, if the similarity is greater than 0.8, it is considered that there is a high possibility that it belongs to this fault type. If the similarity is between 0.5-0.8, it is considered that it may belong to this fault type. If the similarity is less than 0.5, it is considered that it is unlikely to belong to this fault type. The membership function of the current fault type is obtained through fuzzy reasoning;

[0100] Based on the membership function, the recursive Bayesian estimation algorithm is used to update the dynamic probability distribution of fault risk. The prior probability distribution is initialized according to historical data. Each time a new real-time feature vector is obtained, the likelihood probability is calculated based on the membership function, and the posterior probability distribution is updated according to the Bayesian formula. The fault probability distribution is continuously updated to construct a multi-dimensional fault feature space.

[0101] In the multidimensional fault feature space, the kernel principal component analysis method is used to extract the key feature dimensions. The original features are mapped to the high-dimensional feature space, in which principal component analysis is performed to extract the principal components, and the principal components with a cumulative contribution rate of 95% are selected as the key feature dimensions.

[0102] An improved deep autoencoder is used for feature dimensionality reduction and reconstruction. The autoencoder contains 5 layers. The number of neurons in the input layer and the output layer is equal to the original feature dimension. The number of neurons in the three hidden layers is 256, 128, and 64, respectively. Sparse constraints and denoising mechanisms are added during the training process to improve the robustness and generalization ability of the model. The key feature dimensions are input into the autoencoder to obtain the deep fault features after dimensionality reduction.

[0103] Perform cluster analysis on deep fault features and use an adaptive K-means clustering algorithm to automatically determine the optimal number of clusters. Set the range of cluster numbers, such as 2-10, perform K-means clustering for each cluster number, calculate clustering effectiveness indicators such as the silhouette coefficient, select the cluster number with the best indicator as the final result, and dynamically classify the fault mode based on the clustering results.

[0104] In this embodiment, real-time features are deeply learned from the two dimensions of space and time to obtain the probability distribution of the circuit under ten fault states, realize fault risk assessment and fault type diagnosis, and obtain the membership function of the current fault type through fuzzy reasoning, which can quantify the uncertainty and confidence level of the fault diagnosis result. Through repeated iterative updates, a multi-dimensional fault feature space can be constructed to comprehensively characterize the feature distribution and evolution laws of different fault types. The high-dimensional features are compressed into a low-dimensional space through the encoder, and then the original features are reconstructed through the decoder. The deep representation of the features can be learned to generate more abstract and advanced deep fault features. In summary, this embodiment can significantly improve the intelligence level of circuit fault diagnosis, realize real-time assessment of fault risks, accurate identification of fault types and dynamic tracking of fault evolution laws, and provide strong technical support for circuit health management and predictive maintenance.

[0105] S3. Construct a time series prediction network with a multi-layer attention mechanism, input the historical data sequence in the multi-dimensional fault feature space into the time series prediction network, capture the long-range dependency between features through the self-attention layer, extract the dynamic change law corresponding to the historical data sequence in combination with the gated recurrent unit, estimate the fault state transition probability based on the variational inference method, and obtain the fault development trend prediction result, input the fault development trend prediction result into the integrated learning model based on random forest and gradient boosting tree, output multi-dimensional warning indicators, design adaptive warning thresholds based on the multi-dimensional warning indicators, construct hierarchical warning rules in combination with the expert knowledge base, generate hierarchical warning information, push the hierarchical warning information to the preset terminal through an encrypted channel, and store the warning process data in a distributed database.

[0106] The time series prediction network is a deep learning model specially used for processing time series data. It can learn the time series features in the data and predict future values. It is widely used in finance, meteorology and other fields. The long-range dependency relationship refers to the fact that in time series or sequence data, the value at the current moment may be affected by the value at an earlier moment. This relationship is crucial for modeling and prediction, especially when processing long sequences. The variational inference method is a Bayesian inference method for approximating the posterior distribution. It optimizes the posterior distribution by introducing a variational lower bound, thereby solving the computationally complex probability model. It is often used in generative models in deep learning. The multidimensional warning indicator refers to an indicator generated by comprehensively evaluating the health status of a system or equipment from multiple dimensions, which can reflect potential risks and help decision makers to make effective warnings. The hierarchical warning rule is a rule system formulated according to different risk levels and health states, aiming to hierarchically manage equipment or systems and take corresponding measures in a timely manner according to changes in warning indicators. The hierarchical warning information refers to a specific warning notification generated based on the hierarchical warning rule, which usually includes the severity of the fault, recommended countermeasures, etc., and is used to guide maintenance and management.

[0107] In an optional embodiment,

[0108] Construct a time series prediction network with a multi-layer attention mechanism, input the historical data sequence in the multi-dimensional fault feature space into the time series prediction network, capture the long-range dependency between features through the self-attention layer, extract the dynamic change law of different fault features in combination with the gated recurrent unit, estimate the fault state transition probability based on the variational inference method, and obtain the fault development trend prediction result, input the fault development trend prediction result into an integrated learning model based on random forest and gradient boosting tree, output multi-dimensional warning indicators, design an adaptive warning threshold based on the multi-dimensional warning indicators, construct hierarchical warning rules in combination with the expert knowledge base, generate hierarchical warning information, push the hierarchical warning information to the preset terminal through an encrypted channel, and store the warning process data in a distributed database, including:

[0109] Construct a time series prediction network including a multi-layer attention mechanism, add the historical data sequence in the multi-dimensional fault feature space to the time series prediction network, combine the pre-acquired dynamic fault classification results, determine the long-range dependency between different fault features through the self-attention layer, based on the long-range dependency, combine the gated recurrent unit to determine the hidden state of the current time step and update the state, repeat the update and cyclic encoding until the last time step is reached, and obtain the feature encoding corresponding to the hidden state of the last time step and the dynamic change law corresponding to the historical data sequence;

[0110] Based on the feature coding, latent variables are defined by a variational inference method, and the posterior distribution is approximated in combination with the variational distribution. The variational lower bound is maximized with the goal of maximizing the logarithmic marginal likelihood of the observed data. The variational distribution is parameterized by a neural network in combination with the prior distribution of the latent variables that satisfies the normal distribution, and the sampling values ​​of the latent variables are obtained. The observed data are reconstructed by generating a network and the latent variables are sampled by introducing random noise variables. The state transition probability within a unit time interval is obtained in combination with the loss function. The latent variable encoding is obtained in combination with the posterior distribution sampling and the state transfer function is recursively applied to obtain the latent state sequence and decode it to obtain the fault development trend prediction result.

[0111] The fault development trend prediction result is added to the integrated learning model based on random forest and gradient boosting tree. By integrating the prediction results of random forest module and gradient boosting tree module, a multidimensional early warning index is obtained. For each dimension of the multidimensional early warning index, the relative deviation is determined according to the distribution frequency of the historical data sequence in different intervals. The corresponding adaptive early warning threshold is set based on the relative deviation. A hierarchical early warning rule is constructed in combination with the pre-acquired expert knowledge base. For the current operating parameters, the corresponding hierarchical early warning information is generated and pushed to the pre-set early warning terminal through an encrypted channel, and the early warning process data is stored in a pre-set distributed database.

[0112] The log marginal likelihood refers to the logarithmic value of the marginal likelihood of the observed data in the Bayesian model, which is usually used for model comparison and selection. The larger the log marginal likelihood is, the better the model fits the data. The state transfer function is a function that describes the change of the system state over time. It defines how the system state transfers from one state to another at a given time point, and is the core part of dynamic system modeling.

[0113] A time series prediction network with a multi-layer attention mechanism is constructed, which consists of an input layer, multiple self-attention layers, a gated recurrent unit layer and an output layer. The input layer receives the historical data sequence in the multi-dimensional fault feature space. The input of each time step is a multi-dimensional vector. The self-attention layer is used to capture the long-range dependencies between different features. The attention weight is obtained by calculating the similarity between the query vector, the key vector and the value vector, and then the weighted feature representation is obtained. The gated recurrent unit layer is used to extract the dynamic change law of the sequence data, including the update gate and the reset gate, and selectively retains the historical information. The output layer is used to generate the prediction results.

[0114] Input the historical data sequence into the constructed time series prediction network. Assuming that the length of the historical data sequence is 100 and each time step contains 10 features, the input data shape is (100, 10). The correlation between different features is calculated through the self-attention layer to obtain the attention weight matrix. For example, for the first time step, the following attention weights may be obtained: [0.2, 0.1, 0.3, 0.05, 0.15, 0.05, 0.05, 0.05, 0.03, 0.02], indicating that the third feature is most relevant to the current prediction. Multiply the attention weight by the original feature and sum them to obtain the weighted feature representation.

[0115] The weighted features are input into the gated recurrent unit. Assuming that the hidden state dimension is 64, the parameters of the gated recurrent unit include: update gate weight matrix (10×64), reset gate weight matrix (10×64), candidate hidden state weight matrix (10×64), and the hidden state of the current time step is calculated. For example, for the first time step, a 64-dimensional hidden state vector may be obtained: [0.1, -0.2, 0.3, ..., 0.5]. Repeat until all time steps are processed, and finally a hidden state sequence with a length of 100 and each element being a 64-dimensional vector is obtained.

[0116] Based on the hidden state of the last time step, the variational inference method is used to estimate the fault state transition probability, and the hidden variable z is defined. Assuming its dimension is 32, an encoder network is constructed to map the last hidden state to the mean vector μ and variance vector σ of the hidden variable z. For example, μ=[0.1, 0.2, ..., -0.3] and σ=[0.01, 0.02, ..., 0.03] may be obtained. The hidden variable z is sampled from the distribution, and a decoder network is constructed to map z back to the observation space to obtain the reconstructed observation data. At the same time, random noise ε~N(0, I) is introduced, and the hidden variable is perturbed by z+ε. The loss function is defined, including the reconstruction error and the KL divergence. The state transition probability matrix P within the unit time interval is obtained by optimizing the loss function.

[0117] The estimated state transition probability matrix P is used to perform multi-step prediction on the latent variable z. Specifically, z_t+1=P·z_t+ε is repeatedly applied, where ε is random noise. Assuming that the next 10 time steps are predicted, 10 32-dimensional latent state vectors are obtained. These latent state vectors are input into the decoder network to obtain the corresponding observation space prediction results, that is, the fault development trend prediction results. For example, 10-dimensional feature prediction values ​​for the next 10 time steps may be obtained.

[0118] The fault development trend prediction results are input into the integrated learning model. The model contains two sub-models: random forest and gradient boosting tree. For random forest, 100 decision trees are constructed, and the maximum depth of each tree is 10. For gradient boosting tree, 300 iterations are used, and the learning rate is 0.1. The two sub-models predict the input data respectively, and then the prediction results are weighted averaged to obtain the final multi-dimensional early warning indicators. For example, 5-dimensional early warning indicators may be obtained: [0.8, 0.6, 0.3, 0.9, 0.4].

[0119] For the obtained multi-dimensional warning indicators, adaptive warning thresholds are set based on the distribution of historical data, the historical data is binned, the data distribution frequency of each interval is calculated, and the degree of deviation of the current warning indicator relative to each interval is calculated. For example, for the first dimension 0.8, assuming that it falls in the 90-95 percentile interval of the historical data, the relative deviation is 0.9. Based on this, the warning threshold is set, such as setting a relative deviation greater than 0.8 as high risk, 0.5-0.8 as medium risk, and less than 0.5 as low risk.

[0120] Combined with the expert knowledge base, hierarchical warning rules are constructed. The expert knowledge base contains judgment criteria and processing methods for various fault types. Multidimensional warning indicators are combined with warning thresholds and expert rules to generate hierarchical warning information. For example, the following warning information may be obtained: "Equipment A has a high-risk fault, it is recommended to shut down and repair immediately; equipment B has a medium-risk fault, it is recommended to closely monitor the operating status" and so on.

[0121] The generated graded warning information is pushed to the pre-set warning terminal, such as the staff's mobile device, through the SSL encrypted channel. At the same time, the warning process data (including original features, prediction results, warning indicators, etc.) is stored in a distributed database, and a multi-copy storage strategy is adopted to ensure the reliability and availability of the data.

[0122] In this embodiment, by loop encoding until the last time step, the dynamic change law of the historical data sequence and the final feature encoding are obtained, providing high-quality feature representation for subsequent fault trend prediction, and mapping the hidden state back to the observation space through the decoder to obtain the prediction result of the fault development trend. The uncertainty of the modeled data can be modeled, and the generative model can characterize the inherent mechanism of fault evolution, improve the interpretability and generalization ability of the prediction, and ensemble learning can integrate the advantages of different models to improve the stability and accuracy of the early warning indicators. The graded early warning can take different response measures according to the severity of the fault risk, improve the pertinence and effectiveness of the early warning, and through the mining and analysis of the early warning process data, the early warning model and decision rules can be further optimized to achieve continuous improvement and intelligent upgrading of the early warning system. In summary, this embodiment can detect the fault signs of the circuit early, accurately predict the fault development trend, and provide graded early warning information, so as to provide strong support for the safe operation and preventive maintenance of the circuit, and minimize the losses and impacts caused by the fault.

[0123] In an optional embodiment,

[0124] The calculation formula corresponding to the variational lower bound is as follows:

[0125] ;

[0126] Among them, ELBO( i , f ) represents the value of the variational lower bound, i Describes the decoder parameters, f represents the encoder parameters, is the variational posterior distribution, which means that given the observed data x Then, the hidden variable z The distribution of To generate a model, it means that given latent variables z After that, the model generates observation data x The probability of D KL Represents KL divergence, which is used to measure the variational posterior distribution With prior distribution p ( z ), p ( z ) represents the prior distribution of latent variables, Indicates that observations are generated under the posterior distribution x The log-likelihood of .

[0127] In this embodiment, by maximizing the variational lower bound, the value of the marginal likelihood can be indirectly improved, thereby improving the model's ability to fit the data. The end-to-end training method can make the encoder and decoder adapt to each other, improving the effect of feature extraction and data generation. The continuous latent space is conducive to interpolation, smoothing, sample generation and other operations, which improves the generalization ability and interpretability of the model. By introducing the prior distribution, the variational autoencoder can impose prior knowledge and constraints on the latent variables, so that the latent variables have clear physical meaning and statistical characteristics. In summary, this embodiment can establish a connection between the generative model and the posterior inference, learn a continuous and interpretable latent space, improve the quality of data reconstruction and sample generation, and provide a framework with clear principles and concise implementation, which has achieved remarkable results in image generation, speech synthesis, anomaly detection and other fields.

[0128] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A drive control circuit with a fault detection function, characterized in that: include: Acquire the state parameters of the drive control circuit in the historical working process, calibrate the health index of the drive control circuit based on the change trend of the state parameters, perform time-frequency domain analysis through wavelet transform, extract characteristic parameters in different frequency bands, obtain multi-domain feature vectors, build a multi-layer cascade heterogeneous neural network based on long short-term memory network, convolutional neural network and deep belief network, add the multi-domain feature vectors corresponding to the historical state parameters of the drive control circuit to the heterogeneous neural network, judge whether a fault occurs and the fault mode corresponding to the circuit fault and associate it with the event label, repeatedly judge and iterate the parameters of the heterogeneous neural network, and build a circuit fault detection model based on the model parameters obtained by training; Arrange sensors at the detection nodes of the drive control circuit to collect the operating parameters of the drive control circuit in real time, perform multi-domain feature extraction on the operating parameters, obtain real-time feature vectors and add them to the circuit fault detection model to diagnose whether the drive control circuit has a fault risk and the fault type corresponding to the fault risk, perform similarity calculation between the real-time feature vectors and preset fault features, calculate the distance between feature sequences based on a dynamic time warping algorithm, determine the membership of the fault type in combination with fuzzy logic reasoning, update the fault probability distribution through recursive Bayesian estimation, and construct a multidimensional fault feature space. In the multidimensional fault feature space, extract key feature dimensions based on the kernel principal component analysis method, use an improved deep autoencoder to perform feature dimensionality reduction and reconstruction, and dynamically classify fault modes in combination with an adaptive K-means clustering algorithm; Construct a time series prediction network with a multi-layer attention mechanism, input the historical data sequence in the multi-dimensional fault feature space into the time series prediction network, capture the long-range dependency between features through the self-attention layer, extract the dynamic change law corresponding to the historical data sequence in combination with the gated recurrent unit, estimate the fault state transition probability based on the variational inference method, and obtain the fault development trend prediction result, input the fault development trend prediction result into an integrated learning model based on random forest and gradient boosting tree, output multi-dimensional warning indicators, design an adaptive warning threshold based on the multi-dimensional warning indicators, construct hierarchical warning rules in combination with the expert knowledge base, generate hierarchical warning information, push the hierarchical warning information to the preset terminal through an encrypted channel, and store the warning process data in a distributed database; The state parameters of the drive control circuit in the historical working process are obtained, and based on the change trend of the state parameters, the health index of the drive control circuit is calibrated and time-frequency domain analysis is performed through wavelet transform, and characteristic parameters in different frequency bands are extracted to obtain multi-domain feature vectors. A multi-layer cascade heterogeneous neural network is constructed based on a long short-term memory network, a convolutional neural network and a deep belief network. The multi-domain feature vector corresponding to the historical state parameters of the drive control circuit is added to the heterogeneous neural network, and it is determined whether a fault occurs and the fault mode corresponding to the circuit fault is associated with the event label, and the parameters of the heterogeneous neural network are repeatedly determined and iterated. The circuit fault detection model is constructed based on the model parameters obtained by training, including: Acquire state parameters of the drive control circuit during historical operation, wherein the state parameters include voltage parameters, current parameters, temperature parameters, power parameters, and performance parameters; Based on the change trend of the state parameter, a multi-layer health evaluation system is used to calibrate the health index of the drive control circuit, wherein the multi-layer health evaluation system includes a single index evaluation layer, a combined index evaluation layer and a comprehensive health evaluation layer; The state parameters are analyzed in the time and frequency domains by using an improved wavelet packet transform algorithm, eight-scale decomposition is performed by using an optimal basis selection algorithm and the sym6 wavelet basis function, noise is suppressed by an adaptive threshold method, characteristic parameters are extracted from each sub-band obtained by decomposition, the extracted mean and standard deviation are used as time domain features, the band energy and centroid frequency are used as frequency domain features, and the wavelet energy entropy and singular value distribution are used as time-frequency joint features to construct a multi-domain feature vector; Constructing a multi-layer cascade heterogeneous neural network including a feature extraction layer, a feature fusion layer and a decision layer, wherein the feature extraction layer includes a first branch consisting of a three-layer bidirectional long short-term memory network, a second branch consisting of a convolutional neural network with a channel attention mechanism and a spatial attention mechanism, and a third branch consisting of a deep belief network introducing a contrastive divergence algorithm; Input the multi-domain feature vector into the heterogeneous neural network, extract features through the first branch, the second branch and the third branch respectively, use the attention mechanism in the feature fusion layer to weight the features extracted by the three branches, use the feature selection algorithm to remove redundant features, perform deep feature fusion through the multi-head self-attention mechanism, input the fused features into the support vector machine classifier, the random forest classifier and the XGBoost classifier in the decision layer, use the ensemble learning method to dynamically weight and fuse the output results of the three classifiers, determine whether a fault occurs and the fault mode corresponding to the circuit fault, and associate the fault mode with the event label; The periodic change of learning rate is controlled by a cyclic learning rate adjustment mechanism, a hybrid loss function is constructed by combining cross entropy loss and focal loss, the batch size is adaptively adjusted according to the change of training loss value, the parameters of the heterogeneous neural network are continuously iterated and optimized through an incremental learning mechanism, and a circuit fault detection model is constructed based on the model parameters obtained through training.

2. The drive control circuit with fault detection function according to claim 1, characterized in that: Combining cross entropy loss and focal loss to construct a hybrid loss function is shown in the following formula: Among them, L mix represents the mixed loss value, N represents the total number of samples, yi represents the true label of the i-th sample, represents the predicted probability of the i-th sample, α represents the contribution weight coefficient, and γ represents the focusing parameter.

3. The drive control circuit with fault detection function according to claim 1, characterized in that: Arranging sensors at detection nodes of the drive control circuit, collecting operating parameters of the drive control circuit in real time, performing multi-domain feature extraction on the operating parameters, obtaining real-time feature vectors and adding them to the circuit fault detection model includes: Arrange various types of sensors at different detection nodes of the drive control circuit, including voltage sensors, current sensors, temperature sensors and vibration sensors, measure the bus voltage through the voltage sensor connected in parallel to the DC bus, determine the current signal through the Hall current sensor, measure the circuit temperature through the thermocouple temperature sensor, and measure the vibration signal through the acceleration sensor, and use the collected information as the multi-source heterogeneous operation parameters of the drive control circuit; For the voltage data and current data in the operating parameters, the multi-layer wavelet packet is used to decompose them into multiple sub-bands, and the energy, mean and root mean square corresponding to each sub-band are extracted as non-stationary eigenvectors. For the temperature parameters, the principal component analysis algorithm is used to perform centering and normalization processing, and the covariance matrix and the eigenvalues ​​of the covariance matrix are calculated at the same time. Feature selection is performed according to the cumulative contribution rate to obtain the principal component features and construct the principal component matrix, and finally a slowly varying feature matrix is ​​obtained. For the vibration signal, the original vibration signal is decomposed into intrinsic mode functions and residuals by empirical mode decomposition, and each intrinsic mode function is subjected to Hilbert transformation to obtain the instantaneous phase and instantaneous amplitude, and the instantaneous frequency standard deviation and the instantaneous amplitude mean are extracted as vibration features. The vibration features, the non-stationary eigenvector and the slowly varying feature matrix are combined to obtain a real-time feature vector and add the real-time feature vector to a preset circuit fault detection model; Common faults are applied to the drive control circuit, wherein the common faults include open circuit, breakdown, capacitor short circuit and phase loss, an oscilloscope is used to record the voltage and current corresponding to each fault as fault waveform characteristics, an infrared camera is used to record the device surface temperature distribution, the fault level corresponding to the degree of device damage is determined through thermal imaging, and a fault label is determined based on the fault level, the fault waveform characteristics and the device surface temperature distribution.

4. The drive control circuit with fault detection function according to claim 1, characterized in that: Diagnose whether the drive control circuit has a fault risk and the fault type corresponding to the fault risk, calculate the similarity between the real-time feature vector and the preset fault feature, calculate the distance between the feature sequences based on the dynamic time warping algorithm, determine the membership of the fault type in combination with fuzzy logic reasoning, update the fault probability distribution through recursive Bayesian estimation, and construct a multidimensional fault feature space. In the multidimensional fault feature space, extract key feature dimensions based on the kernel principal component analysis method, use an improved deep autoencoder to perform feature dimensionality reduction and reconstruction, and combine the adaptive K-means clustering algorithm to dynamically classify the fault mode, including: The circuit fault detection model includes a one-dimensional convolutional neural network and a long short-term memory network. Based on the real-time feature vector, local spatial feature extraction is performed through a one-dimensional convolutional layer, and activation functions and maximum pooling downsampling are combined to cascade two long short-term memory neural networks to learn the long-term dependency in the real-time feature vector, and obtain the probability distribution of the real-time feature vector in ten fault states, and use the probability distribution as the fault risk and the corresponding fault type corresponding to the drive control circuit; For each fault type, in combination with the pre-acquired fault label, the similarity between the real-time feature vector and the fault features in the pre-set historical fault library is calculated, the distance between the real-time feature vector and the feature sequence in the historical fault library is calculated by a dynamic regularization algorithm, the optimal matching path is generated and used as the similarity, a fuzzy set is defined, and an inference rule is formulated and inference is performed based on the similarity to obtain a membership function corresponding to the current fault type, based on the membership function, the dynamic probability distribution of the fault risk and the membership function are updated by a recursive Bayesian estimation algorithm, and the multi-dimensional fault feature space is repeatedly updated and constructed; In the multidimensional fault feature space, the key feature dimensions corresponding to each fault type and the real-time feature vector are extracted by the kernel principal component analysis algorithm, and the feature dimension reduction and reconstruction are performed in combination with the improved deep autoencoder to generate deep fault features. For the deep fault features, clustering is performed in the multidimensional fault feature space by an adaptive K-means clustering algorithm combined with a clustering effectiveness index to obtain a dynamic fault classification result.

5. The drive control circuit with fault detection function according to claim 1, characterized in that: Construct a time series prediction network with a multi-layer attention mechanism, input the historical data sequence in the multi-dimensional fault feature space into the time series prediction network, capture the long-range dependency between features through the self-attention layer, extract the dynamic change law of different fault features in combination with the gated recurrent unit, estimate the fault state transition probability based on the variational inference method, and obtain the fault development trend prediction result, input the fault development trend prediction result into an integrated learning model based on random forest and gradient boosting tree, output multi-dimensional warning indicators, design an adaptive warning threshold based on the multi-dimensional warning indicators, construct hierarchical warning rules in combination with the expert knowledge base, generate hierarchical warning information, push the hierarchical warning information to the preset terminal through an encrypted channel, and store the warning process data in a distributed database, including: Construct a time series prediction network including a multi-layer attention mechanism, add the historical data sequence in the multi-dimensional fault feature space to the time series prediction network, combine the pre-acquired dynamic fault classification results, determine the long-range dependency between different fault features through the self-attention layer, based on the long-range dependency, combine the gated recurrent unit to determine the hidden state of the current time step and update the state, repeat the update and cyclic encoding until the last time step is reached, and obtain the feature encoding corresponding to the hidden state of the last time step and the dynamic change law corresponding to the historical data sequence; Based on the feature coding, latent variables are defined by a variational inference method, and the posterior distribution is approximated in combination with the variational distribution. The variational lower bound is maximized with the goal of maximizing the logarithmic marginal likelihood of the observed data. The variational distribution is parameterized by a neural network in combination with the prior distribution of the latent variables that satisfies the normal distribution, and the sampling values ​​of the latent variables are obtained. The observed data are reconstructed by generating a network and the latent variables are sampled by introducing random noise variables. The state transition probability within a unit time interval is obtained in combination with the loss function. The latent variable encoding is obtained in combination with the posterior distribution sampling and the state transfer function is recursively applied to obtain the latent state sequence and decode it to obtain the fault development trend prediction result. The fault development trend prediction result is added to the integrated learning model based on random forest and gradient boosting tree. By integrating the prediction results of random forest module and gradient boosting tree module, a multidimensional early warning index is obtained. For each dimension of the multidimensional early warning index, the relative deviation is determined according to the distribution frequency of the historical data sequence in different intervals. The corresponding adaptive early warning threshold is set based on the relative deviation. A hierarchical early warning rule is constructed in combination with the pre-acquired expert knowledge base. For the current operating parameters, the corresponding hierarchical early warning information is generated and pushed to the pre-set early warning terminal through an encrypted channel, and the early warning process data is stored in a pre-set distributed database.

6. The drive control circuit with fault detection function according to claim 5, characterized in that: The calculation formula corresponding to the variational lower bound is as follows: Among them, ELBO(θ, φ) represents the value of the variational lower bound, θ represents the decoder parameters, φ represents the encoder parameters, and q φ (z|x) is the variational posterior distribution, which represents the distribution of the latent variable z after given the observed data x, p θ (x|z) is the generative model, which indicates the probability of the model generating observation data x given the latent variable z, D KL Represents KL divergence, which is used to measure the variational posterior distribution q φ The difference between (z|x) and the prior distribution p(z), which represents the prior distribution of the latent variable, represents the log-likelihood of generating the observation x under the posterior distribution.

Images