Power distribution network early fault diagnosis and prediction method based on transient signal feature mining and transfer learning

Abstract

The invention relates to the technical field of fault diagnosis and prediction, in particular to a power distribution network early fault diagnosis and prediction method based on transient signal feature mining and transfer learning. The method comprises the following steps: acquiring transient signals of key equipment of the power distribution network in normal and slightly abnormal states, and preprocessing the transient signals; extracting a multi-resolution feature vector based on the preprocessed transient signal, a wavelet packet, multi-scale Fourier and empirical mode decomposition, inputting the multi-resolution feature vector to a sparse auto-encoder, and extracting a key feature subset representing a fault stage; a feature distribution difference metric between the source domain device and the target device is constructed based on the key feature subset. According to the method, multi-resolution analysis means such as a wavelet packet, multi-scale Fourier transform and empirical mode decomposition are fused, and impact, oscillation and frequency spectrum characteristics reflecting slight abnormality of equipment are deeply excavated from transient voltage / current signals.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis and prediction technology, and more specifically, to a method for early fault diagnosis and prediction of distribution networks based on transient signal feature mining and transfer learning. Background Technology

[0002] In modern power distribution networks, with the increasing prominence of distributed energy access, diversified load types, and aging equipment, the system operating environment is becoming increasingly complex. Early faults (such as partial discharge, poor contact, and insulation degradation) exhibit characteristics such as weak signals, short duration, and strong nonlinearity and non-stationarity. These faults often do not immediately cause tripping or obvious anomalies in their initial stages, but if they are not identified and intervened in time, they can easily evolve into serious faults, causing power outages or even equipment damage. Traditional fault diagnosis methods based on steady-state signals or threshold alarms are difficult to effectively capture these transient and weak early abnormal features, and rely on a large amount of labeled data, resulting in severely insufficient generalization ability when facing different models, operating conditions, or newly commissioned equipment. In addition, historical fault samples of target equipment are scarce in actual operation, while source domain equipment (such as laboratory or operational equipment) has abundant labeled data, but the feature distribution shifts due to equipment differences and environmental changes, significantly degrading the performance of direct transfer models. Therefore, this paper proposes a method for early fault diagnosis and prediction in power distribution networks based on transient signal feature mining and transfer learning. Summary of the Invention

[0003] The purpose of this invention is to provide a method for early fault diagnosis and prediction of distribution networks based on transient signal feature mining and transfer learning, so as to solve the problems mentioned in the background art, such as weak and difficult-to-distinguish early fault signals of distribution networks, large differences in feature distribution among equipment, and scarcity of labeled samples, which lead to low sensitivity and poor generalization ability of traditional diagnostic methods.

[0004] To achieve the above objectives, the present invention aims to provide a method for early fault diagnosis and prediction in distribution networks based on transient signal feature mining and transfer learning, comprising the following steps: S1. Acquire transient signals of key equipment in the distribution network under normal and slightly abnormal conditions, and preprocess the transient signals. S2. Based on the preprocessed transient signal, multi-resolution feature vectors are extracted based on wavelet packet, multi-scale Fourier and empirical mode decomposition, and the multi-resolution feature vectors are input into a sparse autoencoder to extract the key feature subsets representing the fault stage. S3. Construct a feature distribution difference measure between source domain devices and target devices based on key feature subsets, and use the maximum mean difference transfer model to perform feature alignment to compensate for device differences and operating condition offsets. S4. Based on the key features after feature alignment, a continuous temporal feature sequence is constructed. A long short-term memory network model is used to predict the future short-term failure risk trend. At the same time, a support vector machine model is constructed to identify the minor anomaly types and health status of the current equipment. S5. By integrating the outputs of the Long Short-Term Memory Network model and the Support Vector Machine model, early fault diagnosis results and future risk levels of the current equipment are generated.

[0005] As a further improvement to this technical solution, step S2, which extracts multi-resolution feature vectors based on wavelet packets, multi-scale Fourier transform, and empirical mode decomposition, includes the following steps: S2.1. Divide the preprocessed transient signal into segments of equal length, and perform windowing, detrending, and normalization on each segment; S2.2 Perform wavelet packet decomposition on each signal segment to form a wavelet packet feature vector, which includes at least node energy, peak value, energy ratio and entropy; S2.3 Perform Fast Fourier Transform on the signal segments under different window lengths to extract multi-scale spectral feature vectors; S2.4 Perform empirical mode decomposition on the signal segment to generate eigenmode feature vectors, which include at least the energy, sharpness, instantaneous frequency and Teager energy operator index of each eigenmode function; S2.5. The wavelet packet feature vector, multi-scale spectral feature vector, and intrinsic mode feature vector are concatenated and fused, and then standardized to generate the final unified multi-resolution feature vector.

[0006] As a further improvement to this technical solution, in step S2, the multi-resolution feature vector is input into a sparse autoencoder to extract a subset of key features representing the fault stage, including the following steps: S2.6 Input the multi-resolution feature vector into the sparse autoencoder and normalize the multi-resolution feature vector to unify the dimensions. S2.7 In a sparse autoencoder, the input features are mapped to a low-dimensional sparse hidden layer, and the number of activated neurons is limited by sparsity constraints. S2.8. Map the low-dimensional sparse hidden layer representation back to the original feature space through the decoder, and optimize it by minimizing the reconstruction error and sparsity constraints; S2.9. Select N features based on the activation of hidden layer neurons and sparse weights to form a subset of key features, and then normalize them.

[0007] As a further improvement to this technical solution, in step S3, a feature distribution difference metric between the source domain device and the target device is constructed based on a subset of key features, and a maximum mean difference migration model is used for feature alignment to compensate for device differences and operating condition offsets. This includes the following steps: S3.1. Use the key feature subsets as source domain device features respectively. Characteristics of the target device Input, and characteristics of source domain devices. Characteristics of the target device Standardize the process; S3.2 Calculate the statistical distribution difference measure of features between the source domain and the target domain to preliminarily assess the distribution differences of features across different devices; S3.3, Constructing a measure of feature distribution difference based on the maximum mean difference And measure the difference in feature distribution. As an optimization target for transfer learning; S3.4, Based on feature distribution difference measurement Construct a maximum mean difference transfer learning model to incorporate source domain device features. Knowledge transfer to target device features By minimizing the feature distribution difference measure Characteristics of the target equipment Align; S3.5, Output Alignment Target Device Characteristics .

[0008] As a further improvement to this technical solution, in S3.3, a measure of the difference in characteristic distribution is constructed based on the maximum mean difference. This includes the following steps: S3.31 Obtain the set of key features of the source domain With the set of key features of the target domain And normalize it; S3.32. Set the key features of the source domain using Gaussian radial basis kernel functions. With the set of key features of the target domain Mapping to the reproducing kernel Hilbert space to compute the mean embedding of the feature distribution; S3.33. Apply class balance coefficients to the mapped source domain features and target domain features respectively. The mean embedding of the source domain features is then calculated. Mean embedding of target domain features ; S3.34 Calculate the difference in mean embeddings, i.e., a measure of the difference in feature distributions. .

[0009] As a further improvement to this technical solution, in S3.4, a feature distribution difference measurement is used. Construct a transfer learning model to incorporate source domain device features Knowledge transfer to target device features By minimizing the feature distribution difference measure Characteristics of the target equipment Alignment involves the following steps: S3.41 Measure the difference in characteristic distribution By incorporating transfer learning loss function, a joint optimization objective is constructed. ; S3.42. Input both labeled samples from the source domain and unlabeled samples from the target domain into the maximum mean difference transfer model to make the feature mapping of the target domain closer to the feature distribution of the source domain. S3.43, Utilize transfer loss weights during training. Adaptive adjustment of distribution alignment strength; S3.44. After training convergence, the aligned target device features are output from the feature extraction layer of the maximum mean difference transfer model. .

[0010] As a further improvement to this technical solution, in step S4, a continuous temporal feature sequence is constructed based on the key features after feature alignment, and a long short-term memory network model is used to predict the future short-term fault risk trend, including the following steps: S4.1. Based on the timestamp order of the aligned key features, form a time-series feature sequence; S4.2, using a length of The sliding window segments the time-series feature sequence, with each segment containing... The multidimensional features of each continuous time step are used as a sequence input sample; S4.3. Denote the feature sequence of each sliding window as the input. and the future The fault state is set as the corresponding predicted target. And normalize the time-series feature sequences of each input by channel; S4.4 Organize the normalized temporal feature sequence into a three-dimensional tensor form and input it into the sequence input layer of the Long Short-Term Memory network model, and output the hidden state. ; S4.5, Hide state The input is fed into a fully connected prediction layer, where a linear transformation is used to calculate the short-term failure risk value for several future steps. .

[0011] As a further improvement to this technical solution, in step S4.5, the hidden state is... The input is fed into a fully connected prediction layer, where a linear transformation is used to calculate the short-term failure risk value for several future steps. This includes the following steps: S4.51, Hide state As input to the fully connected prediction layer, where, To hide the state dimension, it represents the temporal representation of the sliding window sequence; S4.52. Construct a multi-layer fully connected network, with each layer containing several neurons, and construct a linear transformation matrix. and bias vector Hidden state Mapped to output space ; S4.53, Use the Softmax function to adjust the output space. Convert to class probability ; S4.54. The class probabilities after processing by the fully connected network and the Softmax function. As a sliding window The predicted value of future short-term failure risk.

[0012] As a further improvement to this technical solution, in step S4, a support vector machine model is constructed to identify the minor anomaly types and health status of the current device, including the following steps: S4.6 Aligned Target Device Features As input to the support vector machine model; S4.7 Construct a training set using source domain labeled samples and some target domain labeled samples. The training set labels include the minor anomaly types and health status of the device. S4.8. Train the support vector machine model using the training set; S4.9 Align the target device features Input the trained support vector machine model and output the current state category of the device.

[0013] As a further improvement to this technical solution, in step S5, the outputs of the Long Short-Term Memory Network model and the Support Vector Machine model are integrated to generate the early fault diagnosis results and future risk level of the current device, including the following steps: S5.1. Use the current device status category output by the support vector machine model as diagnostic information, and use the future short-term fault risk value output by the long short-term memory network model. As predictive information; S5.2. Using a weighted average layer, the current equipment status category is compared with the future short-term failure risk value. Combined, a fused index is generated. ; S5.3. Based on the set risk threshold, the merged indicators... Classification; S5.4 Output the early fault diagnosis results of the current equipment and the corresponding future risk level.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention relates to a method for early fault diagnosis and prediction in distribution networks based on transient signal feature mining and transfer learning. By integrating multi-resolution analysis techniques such as wavelet packet analysis, multi-scale Fourier transform, and empirical mode decomposition, it deeply mines the impact, oscillation, and spectral features reflecting minor equipment anomalies from transient voltage / current signals. Furthermore, it automatically selects the most discriminative subset of key features for the fault stage using a sparse autoencoder. Based on this, a transfer learning model based on maximum mean difference (MMD) and incorporating a class balancing mechanism is introduced. This effectively aligns the feature distributions between different devices, significantly mitigating the performance degradation caused by differences in equipment models, operating condition deviations, or sample class imbalances. Thus, it can achieve high-precision early fault type identification and health status assessment even when target equipment labeled data is scarce.

[0015] 2. This invention relates to a method for early fault diagnosis and prediction in distribution networks based on transient signal feature mining and transfer learning. It utilizes Support Vector Machines (SVM) for accurate classification of current minor anomalies, while simultaneously employing Long Short-Term Memory Networks (LSTM) to model aligned temporal key features and predict short-term fault risk trends. Furthermore, by weighted fusion of the two model outputs, a comprehensive risk index with both real-time diagnostic accuracy and trend prediction foresight is generated, triggering tiered early warnings based on preset thresholds. This mechanism not only promptly detects potential faults in their nascent stages but also dynamically assesses their evolutionary risks, providing maintenance personnel with scientific and reliable decision-making support and significantly improving the safety and intelligence level of distribution network operation. Attached Figure Description

[0016] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

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

[0018] Example: Please refer to Figure 1 As shown, this embodiment provides a method for early fault diagnosis and prediction of distribution networks based on transient signal feature mining and transfer learning, including the following steps: S1. Acquire transient signals (including transient voltage and current signals) of key equipment in the power distribution network under normal and slightly abnormal conditions, and preprocess the transient signals.

[0019] S2. Based on the preprocessed transient signal, multi-resolution feature vectors are extracted based on wavelet packet, multi-scale Fourier and empirical mode decomposition, and the multi-resolution feature vectors are input into a sparse autoencoder to extract the key feature subsets representing the fault stage. In this embodiment, multi-resolution feature vectors are extracted based on wavelet packets, multi-scale Fourier transform, and empirical mode decomposition, including the following steps: S2.1. The preprocessed transient signal is divided into segments of equal length, and each segment is windowed, detrended, and normalized to establish a unified input for subsequent feature extraction. Specifically, the preprocessed transient signal is divided into several segments of equal length according to a preset time length or number of sampling points. A window function (such as a Hanning window or a Hamming window) is first applied to each segment to reduce boundary effects. Then, the segment signal is detrended (e.g., removing linear or low-frequency drift components) to eliminate slow-varying interference. Finally, the detrended segments are normalized to map the signal amplitude to a uniform dimension interval (such as [0,1] or a normalized value with a mean of 0 and a variance of 1), ensuring that each segment has a consistent amplitude range and statistical characteristics, providing a standardized input for subsequent multi-resolution feature extraction. S2.2. Perform L-level wavelet packet decomposition on each signal segment to form a wavelet packet feature vector, which includes at least node energy, peak value, energy ratio, and entropy, to characterize the impulse and low-to-mid-frequency components. Specifically, for each signal segment, first select a wavelet basis function (such as Daubechies) and set the decomposition level L. Decompose the segment signal through L-level wavelet packet decomposition to obtain the sub-node coefficients of different frequency bands. Calculate the energy of each node coefficient as a signal strength feature. Extract the node peak value to reflect instantaneous extreme value information. Calculate the node energy ratio (the proportion of node energy to total energy) to characterize the energy distribution of each frequency band. Further calculate the node entropy (such as Shannon entropy) to reflect the uncertainty or complexity of the signal. Concatenate the energy, peak value, energy ratio, and entropy features of all nodes to form a wavelet packet feature vector, which is used to characterize the impulse and frequency characteristics of the signal at multiple scales. S2.3. Perform Fast Fourier Transform (FFT) on signal segments under different window lengths to extract multi-scale spectral feature vectors, including at least the dominant frequency, harmonic energy, high-frequency energy, and spectral centroid, to characterize the frequency energy distribution. Specifically, for each signal segment, select multiple analysis windows of different lengths (such as short and long windows) to segment the segment. Apply a window function to each window and then perform a Fast Fourier Transform (FFT) to obtain the corresponding frequency domain amplitude spectrum. Identify the frequency corresponding to the largest amplitude value in the spectrum as the dominant frequency, calculate the harmonic energy of each order as the harmonic energy feature, and statistically analyze the spectral energy above a set threshold or a specific frequency band as the high-frequency energy index. Calculate the spectral centroid (frequency-weighted average) to characterize the energy center location. Concatenate the dominant frequency, harmonic energy, high-frequency energy, and spectral centroid features obtained under different window lengths to form a multi-scale spectral feature vector, which is used to describe the frequency distribution characteristics of the signal at multiple time scales. S2.4. Perform Empirical Mode Decomposition (EMD) on the signal segment to generate an Eigenmode Feature Vector (EMV), which includes at least the energy, sharpness, instantaneous frequency, and Teager energy operator index for each Eigenmode Function (IMF) to capture impulse and oscillation modes. Specifically, for each signal segment, EMD is first applied to decompose the signal into several IMFs and residuals. Each IMF represents the oscillation mode of the signal at a specific time scale. The energy of each IMF is calculated to reflect its amplitude intensity, the sharpness (Kurtosis) is calculated to measure the peak characteristics of the signal, the instantaneous frequency is calculated using Hilbert transform or instantaneous phase derivative to describe local frequency changes, and the instantaneous energy change is calculated using the Teager energy operator (TEO) to capture nonlinear or rapid impulse characteristics. The energy, sharpness, instantaneous frequency, and Teager energy operator index of each IMF are integrated to form a complete Eigenmode Feature Vector, which is used to characterize the multi-scale oscillation and impulse modes of the signal. S2.5. The wavelet packet feature vector, multi-scale spectral feature vector, and intrinsic mode feature vector are concatenated and fused, and then standardized to generate a final unified multi-resolution feature vector, providing input for subsequent feature compression and fault identification.

[0020] Furthermore, the multi-resolution feature vectors are input into a sparse autoencoder to extract a subset of key features representing the fault stage, including the following steps: S2.6 Input the multi-resolution feature vector into the sparse autoencoder and normalize the multi-resolution feature vector to unify the dimensions (using standardization with a mean of 0 and a variance of 1 or scaling the features to a fixed interval (such as [0,1])) to provide standardized input for encoding training. The sparse autoencoder consists of an input layer, an encoder, a sparse hidden layer, a decoder, and an output layer. The input layer receives the normalized multi-resolution feature vector. The encoder maps the high-dimensional input to a low-dimensional space through several fully connected layers and applies sparse constraints in the bottom sparse hidden layer so that only a small number of neurons are activated to capture the most sensitive features. The decoder then maps the low-dimensional representation back to the output layer with the same dimension as the input through a symmetric fully connected layer. The output reconstructed feature vector is used to calculate the reconstruction error and optimize the network parameters, thereby ensuring that the hidden layer can effectively represent key feature information. S2.7 In a sparse autoencoder, the input features are mapped to a low-dimensional sparse hidden layer. The number of activated neurons is limited by sparse constraints, so that the hidden layer focuses on capturing the features most sensitive to fault detection. S2.8. The low-dimensional sparse hidden layer representation is mapped back to the original feature space through the decoder, and optimized by minimizing the reconstruction error and sparsity constraints to ensure that the hidden layer effectively represents the original feature information. Specifically, the low-dimensional representation of the sparse hidden layer is first input into the decoder, and mapped back to the original feature space layer by layer through a symmetric fully connected layer to generate a reconstruction vector with the same dimension as the input feature. During training, the degree of information retention is evaluated by calculating the reconstruction error (such as mean square error) between the input feature vector and the reconstruction vector. At the same time, the number of activated neurons in the hidden layer is limited by combining sparsity constraints (such as KL divergence or L1 regularization). The reconstruction error and sparsity constraints are jointly used to construct a loss function. The network weights and biases are iteratively optimized through backpropagation so that the decoder can accurately restore the original features while maintaining the sparsity of the hidden layer, thereby ensuring that the low-dimensional representation effectively represents the key information of the input features. S2.9. Select the N most sensitive features based on the activation of hidden layer neurons and sparse weights to form a key feature subset (based on the trained sparse autoencoder, the input features are mapped to the sparse hidden layer through the encoder, the activation value of each hidden neuron on the sample and its corresponding sparse weight are counted, and the input features corresponding to neurons with high activation frequency and large weight are identified as most sensitive to fault discrimination. Then, the top N features strongly correlated with these highly sensitive neurons are selected from the original features to form a key feature subset), and normalization is performed to provide highly discriminative input for transfer learning and fault prediction.

[0021] S3. Construct a feature distribution difference measure between source domain devices and target devices based on key feature subsets, and use the maximum mean difference transfer model to perform feature alignment to compensate for device differences and operating condition offsets. In this embodiment, a feature distribution difference metric between the source domain device and the target device is constructed based on a subset of key features, and a maximum mean difference transfer model is used for feature alignment to compensate for device differences and operating condition offsets. This includes the following steps: S3.1. Use the key feature subsets as source domain device features respectively. Characteristics of the target device Input, and characteristics of source domain devices. Characteristics of the target device Standardization is performed to ensure that the characteristic dimensions of different devices are consistent, providing a unified basis for subsequent measurement of distribution differences. S3.2 Calculate the statistical distribution difference measure of the features between the source domain and the target domain, such as the mean, variance, and covariance matrix, to preliminarily assess the distribution difference of the features across different devices; S3.3 Constructing a measure of feature distribution difference based on maximum mean difference (MMD) And measure the difference in feature distribution. As an optimization objective for transfer learning, it is used to guide feature alignment and model training; Among them, the measure of feature distribution difference is constructed based on the maximum mean difference (MMD). This includes the following steps: S3.31 Obtain the set of key features of the source domain With the set of key features of the target domain And normalize them to ensure that the characteristic dimensions of different devices are consistent; S3.32. Set the key features of the source domain using Gaussian radial basis kernel functions. With the set of key features of the target domain Mapping to the Regenerating Kernel Hilbert Space (RKHS) to compute the mean embedding of the feature distribution; where, when mapping features through the Gaussian radial basis kernel function, a high-dimensional mapping is not explicitly constructed, but the kernel function is used to directly compute the inner product of any two feature vectors in the Regenerating Kernel Hilbert Space (RKHS), which is equivalent to implicitly mapping the original features to a high-dimensional or even infinite-dimensional space, in which the nonlinear relationship becomes linearly additive; S3.33. Apply class balance coefficients to the mapped source domain features and target domain features respectively. The mean embedding of the source domain features is then calculated. Mean embedding of target domain features In the formula, The number of key feature samples in the source domain. For the source domain Key feature vectors (such as feature samples from different devices). Let Gaussian radial basis kernel function be used. This is the source domain feature sample class balance coefficient, used to compensate for the impact of source domain fault sample class imbalance on mean embedding. The number of key feature samples in the target domain. For the target domain Key feature vectors It is the class balance coefficient of the target domain feature samples, used to reduce the bias caused by class imbalance or insufficient samples in the target domain; Category balance coefficient It is calculated based on the number of samples in each category and is used to balance the influence of different categories when embedding by mean. Specifically, for each category... Its balance coefficient is defined as ,in The total number of samples, The total number of categories, For category The number of samples; through this coefficient, the class with fewer samples gets a higher weight in the mean embedding calculation, while the class with more samples gets a relatively lower weight, thus ensuring that the contribution of each class to the mean embedding is relatively balanced and avoiding the bias caused by class imbalance; In this invention, a class balance coefficient is applied to the mapped source domain features and target domain features respectively. This is to compensate for the imbalance in the distribution of sample classes in the source or target domain. In actual power distribution network equipment, the number of samples for different fault types can vary significantly. If the mean embedding is directly calculated, the class with the larger sample size will dominate the feature mean, causing the mean embedding to be biased towards the majority class, resulting in a decrease in the discrimination ability of the transfer learning model for minority class faults. Introducing a class balance coefficient addresses this issue. This method weights the samples of each category during the mean embedding calculation, making the contributions of different categories to the mean embedding relatively balanced, thus more accurately reflecting the central distribution of each category in the RKHS space. The method adjusts the sample categories by weighting them with a category balance coefficient during the mean embedding stage, which significantly reduces the negative impact of class imbalance on feature distribution alignment, ensures the ability of the transfer learning model to identify minority class faults, and enhances the stability and robustness of cross-device feature alignment. Compared with the traditional unweighted MMD method, it has higher diagnostic accuracy and generalization ability in scenarios with imbalanced fault type distribution. S3.34 Calculate the difference in mean embeddings, i.e., a measure of the difference in feature distributions. , denoted by the vector norm in the reproducing kernel Hilbert space (RKHS), which is naturally defined by the inner product in the RKHS; S3.4, Based on feature distribution difference measurement Construct a maximum mean difference transfer learning model to incorporate source domain device features. Knowledge transfer to target device features By minimizing the feature distribution difference measure Characteristics of the target equipment Alignment is performed to make its distribution consistent with or similar to the characteristics of the source domain; Furthermore, based on feature distribution difference measurement Construct a transfer learning model to incorporate source domain device features Knowledge transfer to target device features By minimizing the feature distribution difference measure Characteristics of the target equipment Alignment involves the following steps: S3.41 Measure the difference in characteristic distribution By incorporating transfer learning loss function, a joint optimization objective is constructed. (To balance source domain task learning with cross-domain feature distribution alignment): ; in, For source domain supervised classification loss, the cross-entropy loss function is used to measure the classification error of source domain samples. These are the transfer loss weights, used to balance knowledge learning and distribution alignment; S3.42. Input both labeled samples from the source domain and unlabeled samples from the target domain into the maximum mean difference transfer model, and continuously optimize the model parameters through backpropagation to minimize the difference while ensuring the classification performance in the source domain. This gradually brings the target domain feature mapping closer to the source domain feature distribution; The Maximum Mean Difference (MMD) transfer model consists of an input layer, a feature extraction intermediate layer, an MMD computation layer, and an output layer. The input layer receives labeled features from the source domain and features from the target domain, with each node corresponding to a key feature dimension. The feature extraction intermediate layer performs nonlinear mapping on the input features through several fully connected layers, extracting higher-order representations and applying class balance coefficients to compensate for sample imbalance. In the MMD computation layer, features are implicitly mapped to the Regenerating Kernel Hilbert Space (RKHS) using a Gaussian kernel, and the squared distance between the mean embeddings of the source and target domains is calculated as the transfer loss. The output layer generates aligned target domain features for subsequent fault classification or risk prediction, achieving cross-device feature distribution alignment and transfer. S3.43, Utilize transfer loss weights during training. Adaptive adjustment of distribution alignment strength (by setting an initial value) The value is dynamically adjusted based on the differences in source domain classification accuracy and target domain distribution during training iterations, and is appropriately increased when the source domain classification performance is stable. Strengthen target domain feature alignment, while reducing it when source domain classification performance degrades or target domain alignment is too strong. To maintain the accuracy of the source domain task, the model adaptively balances knowledge transfer and source domain task learning, so that it maintains overall classification accuracy while gradually narrowing the difference in feature distribution between the source and target domains. This allows the model to gradually reduce the mean embedding distance between the target domain features and the source domain features in RKHS while maintaining the accuracy of source domain classification. S3.44. After training convergence, the aligned target device features are output from the feature extraction layer of the maximum mean difference transfer model. Its distribution is related to the source region. More consistent, used to improve the accuracy of fault diagnosis and prediction for target equipment; S3.5, Output Alignment Target Device Characteristics Alignment features are used for subsequent fault classification and time-series prediction model training to achieve cross-device generalization capability.

[0022] S4. Based on the key features after feature alignment, a continuous temporal feature sequence is constructed. A long short-term memory network model is used to predict the future short-term failure risk trend. At the same time, a support vector machine model is constructed to identify the minor anomaly types and health status of the current equipment. In this embodiment, a continuous temporal feature sequence is constructed based on the key features after feature alignment. A long short-term memory network model is then used to predict future short-term failure risk trends, including the following steps: S4.1 Based on the timestamp order of the aligned key features, arrange the features according to the sampling time and interpolate and complete the abnormal intervals using the linear interpolation method to form a continuous time series feature sequence with equal time steps; S4.2, using a length of The sliding window segments the time-series feature sequence, with each segment containing... The multidimensional features of each continuous time step are used as a sequence input sample; S4.3. Denote the feature sequence of each sliding window as the input. and the future The fault state is set as the corresponding predicted target. And normalize the time-series feature sequences of each input by channel; S4.4. The normalized temporal feature sequence is organized into a three-dimensional tensor according to the structure of "time step × feature dimension" and input into the sequence input layer of the Long Short-Term Memory (LSTM) network model. In each time step, the LSTM updates the state of the current feature through the input gate, forget gate, and output gate, thereby capturing the long-term dependency of the fault event in the time dimension and outputting the hidden state. (The final hidden state of LSTM after the entire sequence processing is completed) (This is used as a time-series representation of the sequence for subsequent risk prediction). The Long Short-Term Memory (LSTM) network model consists of an input layer, a hidden layer, and an output layer. The input layer receives a multi-dimensional temporal feature sequence arranged in time steps, with each input node corresponding to a feature dimension. The hidden layer consists of several LSTM units, each containing an input gate, a forget gate, and an output gate. These units are used to selectively update and retain historical information, thereby capturing long-term and short-term dependencies and generating a hidden state vector as a temporal representation during sequence processing. The output layer is typically a fully connected layer that maps the hidden state to the prediction space, outputting a future short-term fault risk value or classification probability, thus enabling the prediction of equipment health status and potential fault trends. S4.5, Hide state The input is fed into a fully connected prediction layer, where a linear transformation is used to calculate the short-term failure risk value for several future steps. ; Among them, the hidden state The input is fed into a fully connected prediction layer, where a linear transformation is used to calculate the short-term failure risk value for several future steps. This includes the following steps: S4.51, Hide state As input to the fully connected prediction layer, where, To hide the state dimension, it represents the temporal representation of the sliding window sequence; S4.52. Construct one or more fully connected networks, each containing several neurons, and construct a linear transformation matrix. and bias vector Hidden state Mapped to output space : ; S4.53, Use the Softmax function to adjust the output space. Convert to class probability ; S4.54. The class probabilities after processing by the fully connected network and the Softmax function. As a sliding window The predicted value of future short-term failure risk.

[0023] Furthermore, a support vector machine model is constructed to identify the minor anomaly types and health status of the current device, including the following steps: S4.6 Aligned Target Device Features The features are used as input to the support vector machine model and are standardized (using the Z-score method) to ensure stable model training. S4.7 Construct a training set using source domain labeled samples and some target domain labeled samples. The training set labels include the minor anomaly types and health status of the device. S4.8. Train a Support Vector Machine (SVM) model using the training set. The SVM model consists of an input layer, a kernel mapping (hidden intermediate layer), and an output layer. The input layer receives normalized key feature vectors, with each node corresponding to a feature dimension. A kernel function (such as a linear kernel, radial basis function kernel, or polynomial kernel) implicitly maps the input features to a high-dimensional feature space, acting as an intermediate layer to enhance the linear separability of samples. The output layer maps the input to category labels or health status prediction results based on the trained support vectors and the optimal hyperplane, thus determining the type of minor equipment anomalies and their health status. No; Support Vector Machine (SVM) training maximizes the margin between samples of different classes by finding the optimal hyperplane. First, the input features are mapped to a high-dimensional feature space through a kernel function. Then, a constrained optimization problem is constructed: minimizing the regularized classification loss while ensuring that the support vectors are correctly classified. By solving this convex quadratic optimization problem, the support vectors and weight coefficients are obtained, and the position and orientation of the classification hyperplane are determined. After training, the model can project new input features onto one side of the hyperplane to achieve classification of minor anomalies or healthy states. At the same time, the kernel function mapping ensures effective identification of samples with non-linear distributions. S4.9 Align the target device features Input the trained support vector machine model and output the current device state category, including specific minor anomaly types or healthy status.

[0024] S5. By integrating the outputs of the Long Short-Term Memory Network Model and the Support Vector Machine Model, early fault diagnosis results and future risk levels of the current equipment are generated. In this embodiment, generating early fault diagnosis results and future risk levels for the current device includes the following steps: S5.1. Use the current device status category output by the support vector machine model as diagnostic information, and use the future short-term fault risk value output by the long short-term memory network model. As predictive information; S5.2. Using a weighted average layer, the current equipment status category is compared with the future short-term failure risk value. Combined, a fused index is generated. ; S5.3. Based on the set risk threshold, the merged indicators... Risk levels can be categorized, for example: low risk, medium risk, and high risk. This is achieved by first analyzing historical equipment operating data or determining the range of risk indicators based on expert experience, and then dividing these into several risk levels, such as... Low risk Medium risk High risk, among which, and The threshold is set; in actual prediction, the fusion index obtained from the current sliding window is used. Compared with the threshold, if Then it is judged as low risk, if Then it is judged as medium risk, if This is then classified as high risk, thus mapping continuous risk values ​​to discrete risk levels, which facilitates equipment status monitoring and early warning triggering; S5.4 Output the early fault diagnosis results of the current device (from SVM state classification) and the corresponding future risk level (from fusion indicators). determination).

[0025] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for early fault diagnosis and prediction in distribution networks based on transient signal feature mining and transfer learning, characterized in that, Includes the following steps: S1. Acquire transient signals of key equipment in the distribution network under normal and slightly abnormal conditions, and preprocess the transient signals. S2. Based on the preprocessed transient signal, multi-resolution feature vectors are extracted based on wavelet packet, multi-scale Fourier and empirical mode decomposition, and the multi-resolution feature vectors are input into a sparse autoencoder to extract the key feature subsets representing the fault stage. S3. Construct a feature distribution difference measure between source domain devices and target devices based on key feature subsets, and use the maximum mean difference transfer model to perform feature alignment to compensate for device differences and operating condition offsets. S4. Based on the key features after feature alignment, a continuous temporal feature sequence is constructed. A long short-term memory network model is used to predict the future short-term failure risk trend. At the same time, a support vector machine model is constructed to identify the minor anomaly types and health status of the current equipment. S5. By integrating the outputs of the Long Short-Term Memory Network model and the Support Vector Machine model, early fault diagnosis results and future risk levels of the current equipment are generated.

2. The method for early fault diagnosis and prediction of distribution networks based on transient signal feature mining and transfer learning according to claim 1, characterized in that: In step S2, multi-resolution feature vectors are extracted based on wavelet packets, multi-scale Fourier transform, and empirical mode decomposition, including the following steps: S2.

1. Divide the preprocessed transient signal into segments of equal length, and perform windowing, detrending, and normalization on each segment; S2.2 Perform wavelet packet decomposition on each signal segment to form a wavelet packet feature vector, which includes at least node energy, peak value, energy ratio and entropy; S2.3 Perform Fast Fourier Transform on the signal segments under different window lengths to extract multi-scale spectral feature vectors; S2.4 Perform empirical mode decomposition on the signal segment to generate eigenmode feature vectors, which include at least the energy, sharpness, instantaneous frequency and Teager energy operator index of each eigenmode function; S2.

5. The wavelet packet feature vector, multi-scale spectral feature vector, and intrinsic mode feature vector are concatenated and fused, and then standardized to generate the final unified multi-resolution feature vector.

3. The method for early fault diagnosis and prediction of distribution networks based on transient signal feature mining and transfer learning according to claim 1, characterized in that: In step S2, the multi-resolution feature vector is input into the sparse autoencoder to extract a subset of key features representing the fault stage, including the following steps: S2.6 Input the multi-resolution feature vector into the sparse autoencoder and normalize the multi-resolution feature vector to unify the dimensions. S2.7 In a sparse autoencoder, the input features are mapped to a low-dimensional sparse hidden layer, and the number of activated neurons is limited by sparsity constraints. S2.

8. Map the low-dimensional sparse hidden layer representation back to the original feature space through the decoder, and optimize it by minimizing the reconstruction error and sparsity constraints. S2.

9. Select N features based on the activation of hidden layer neurons and sparse weights to form a subset of key features, and then normalize them.

4. The method for early fault diagnosis and prediction of distribution networks based on transient signal feature mining and transfer learning according to claim 1, characterized in that: In step S3, a feature distribution difference metric between the source and target devices is constructed based on a subset of key features, and a maximum mean difference transfer model is used for feature alignment to compensate for device differences and operating condition offsets. This includes the following steps: S3.

1. Use the key feature subsets as source domain device features respectively. Characteristics of the target device Input, and characteristics of source domain devices. Characteristics of the target device Standardize the process; S3.2 Calculate the statistical distribution difference measure of features between the source domain and the target domain to preliminarily assess the distribution differences of features across different devices; S3.3, Constructing a measure of feature distribution difference based on the maximum mean difference And measure the difference in feature distribution. As an optimization target for transfer learning; S3.4, Based on feature distribution difference measurement Construct a maximum mean difference transfer learning model to incorporate source domain device features. Knowledge transfer to target device features By minimizing the feature distribution difference measure Characteristics of the target device Align; S3.5, Output Alignment Target Device Characteristics .

5. The method for early fault diagnosis and prediction of distribution networks based on transient signal feature mining and transfer learning according to claim 4, characterized in that: In S3.3, a measure of the difference in characteristic distribution is constructed based on the maximum mean difference. This includes the following steps: S3.31 Obtain the set of key features of the source domain With the set of key features of the target domain And normalize it; S3.

32. Set the key features of the source domain using Gaussian radial basis kernel functions. With the set of key features of the target domain Mapping to the reproducing kernel Hilbert space to compute the mean embedding of the feature distribution; S3.

33. Apply class balance coefficients to the mapped source domain features and target domain features respectively. The mean embedding of the source domain features is then calculated. Mean embedding of target domain features ; S3.34 Calculate the difference in mean embeddings, i.e., a measure of the difference in feature distributions. .

6. The method for early fault diagnosis and prediction of distribution networks based on transient signal feature mining and transfer learning according to claim 4, characterized in that: In S3.4, the feature distribution difference measurement is used. Construct a transfer learning model to incorporate source domain device features. Knowledge transfer to target device features By minimizing the feature distribution difference measure Characteristics of the target device Alignment involves the following steps: S3.41 Measure the difference in characteristic distribution By incorporating transfer learning loss function, a joint optimization objective is constructed. ; S3.

42. Input both labeled samples from the source domain and unlabeled samples from the target domain into the maximum mean difference transfer model to make the feature mapping of the target domain closer to the feature distribution of the source domain. S3.43, Utilize transfer loss weights during training. Adaptive adjustment of distribution alignment strength; S3.

44. After training convergence, the aligned target device features are output from the feature extraction layer of the maximum mean difference transfer model. .

7. The method for early fault diagnosis and prediction of distribution networks based on transient signal feature mining and transfer learning according to claim 1, characterized in that: In step S4, a continuous temporal feature sequence is constructed based on the key features after feature alignment. A long short-term memory network model is then used to predict future short-term failure risk trends, including the following steps: S4.

1. Based on the timestamp order of the aligned key features, form a time-series feature sequence; S4.2, using a length of The sliding window segments the time-series feature sequence, with each segment containing... The multidimensional features of each continuous time step are used as a sequence input sample; S4.

3. Denote the feature sequence of each sliding window as the input. and the future The fault state is set as the corresponding predicted target. And normalize the time-series feature sequences of each input by channel; S4.4 Organize the normalized temporal feature sequence into a three-dimensional tensor form and input it into the sequence input layer of the Long Short-Term Memory network model, and output the hidden state. ; S4.5, Hide state The input is fed into a fully connected prediction layer, where a linear transformation is used to calculate the short-term failure risk value for several future steps. .

8. The method for early fault diagnosis and prediction of distribution networks based on transient signal feature mining and transfer learning according to claim 7, characterized in that: In S4.5, the hidden state is... The input is fed into a fully connected prediction layer, where a linear transformation is used to calculate the short-term failure risk value for several future steps. This includes the following steps: S4.51, Hide state As input to the fully connected prediction layer, where, To hide the state dimension, it represents the temporal representation of the sliding window sequence; S4.

52. Construct a multi-layer fully connected network, with each layer containing several neurons, and construct a linear transformation matrix. and bias vector Hidden state Mapped to output space ; S4.53, Use the Softmax function to adjust the output space. Convert to class probability ; S4.

54. The class probabilities after processing by the fully connected network and the Softmax function. As a sliding window The predicted value of future short-term failure risk.

9. The method for early fault diagnosis and prediction of distribution networks based on transient signal feature mining and transfer learning according to claim 6, characterized in that: In step S4, a support vector machine model is constructed to identify the minor anomaly types and health status of the current device, including the following steps: S4.6 Aligned Target Device Features As input to the support vector machine model; S4.7 Construct a training set using source domain labeled samples and some target domain labeled samples. The training set labels include the minor anomaly types and health status of the device. S4.

8. Train the support vector machine model using the training set; S4.9 Align the target device features Input the trained support vector machine model and output the current state category of the device.

10. The method for early fault diagnosis and prediction of distribution networks based on transient signal feature mining and transfer learning according to claim 1, characterized in that: In step S5, the outputs of the Long Short-Term Memory Network model and the Support Vector Machine model are fused to generate early fault diagnosis results and future risk levels for the current device, including the following steps: S5.

1. Use the current device status category output by the support vector machine model as diagnostic information, and use the future short-term fault risk value output by the long short-term memory network model. As predictive information; S5.

2. Using a weighted average layer, the current equipment status category is compared with the future short-term failure risk value. Combined, a fused index is generated. ; S5.

3. Based on the set risk threshold, the merged indicators... Classification; S5.4 Output the early fault diagnosis results of the current equipment and the corresponding future risk level.

Images