A method for extracting a fuse pre-fusing feature

Abstract

The present application belongs to the technical field of fuses, and particularly relates to a fuse pre-fusing feature extraction method, comprising: obtaining monitoring time series data of source domain and target domain fuses and extracting initial time series features; a feature extractor decouples the initial time series features into steady-state feature vectors and evolution feature vectors, and calculates subspace orthogonal loss of both; based on the two types of feature vectors, output prediction labels and their probability distribution, and calculate label prediction loss; output domain discrimination results and calculate domain discrimination loss; update the adversarial training network based on the label prediction loss, domain discrimination loss and subspace orthogonal loss; determine the gradient weight, and back-propagate the inverted domain discrimination loss gradient to update the feature extractor; use the trained adversarial training network to identify the pre-fusing features of the target domain fuse. The present application reduces the mutual interference of steady-state background and degradation information under cross-working conditions, and can improve the identification accuracy of the pre-fusing features.

Description

Technical Field

[0001] This invention relates to the field of fuse technology. More specifically, this invention relates to a method for extracting early warning features of fuse failure. Background Technology

[0002] Fuses are commonly used overcurrent protection components in power systems and electrical equipment. Their progressive degradation process before melting typically involves multiple stages, including temperature accumulation, material creep, localized melting, and pre-arc phase. Accurately extracting the precursor features of this process is crucial for predictive maintenance. Currently, fuse monitoring relies heavily on fixed threshold judgments or traditional time-frequency analysis methods. However, under conditions where normal current-carrying backgrounds dominate for extended periods, the signal changes in the precursor stage are weak and exhibit multi-scale non-stationary characteristics. These methods struggle to effectively capture early time-frequency changes, limiting the accuracy of precursor identification. Furthermore, complex industrial operating conditions, including ambient temperature, load fluctuations, and batch variations of components, lead to significant distribution shifts in monitoring data. Additionally, the high cost of acquiring full-lifecycle labeled data under the target operating condition means that directly applying a diagnostic model trained under one operating condition to a new condition often results in a significant decrease in recognition performance, failing to meet the needs of cross-condition early warning.

[0003] In recent years, deep learning and transfer learning methods have been gradually applied to the fields of equipment fault prediction and feature extraction. For example, Chinese patent application CN119827882A discloses a method for detecting slow-blow fuse faults based on phase voltage correlation features. This method collects three-phase voltages from multiple nodes upstream and downstream of the line, normalizes and performs sliding window processing on the collected phase voltages, and then inputs them into a neural network model. It utilizes the characteristic that the phase voltage may slightly drop when a fuse experiences slow blowing to achieve fault detection. This method has achieved certain detection results under specific operating conditions, providing a feasible approach for the intelligent identification of slow-blow fuse faults.

[0004] However, the aforementioned existing technologies still have the following shortcomings when practically applied to extract precursor features of fuse failure. Firstly, this method relies solely on the single feature of phase voltage, failing to fully utilize the complementary information from multi-source monitoring data such as current and temperature. Furthermore, it directly processes raw time-series data using ordinary neural networks, making it difficult to extract multi-scale weak amplitude time-frequency variations in the precursor stage, resulting in insufficient identification of precursor signals. Secondly, this method does not consider the data distribution shift caused by changes in operating conditions. When ambient temperature, load, or batch differences in fuses change, the model's recognition performance significantly degrades, failing to meet the requirements for cross-operating-condition generalized early warning.

[0005] Furthermore, existing deep learning-based fuse fault diagnosis methods typically treat features as a whole, failing to distinguish between relatively stable background information and gradually increasing abrupt evolution information in precursor signals. This lack of effective feature decoupling mechanisms leads to interference between information from different physical attributes. Existing domain adversarial transfer methods often feed back domain discriminant gradients in a fixed manner, failing to dynamically adjust the alignment strength of different subspaces based on the training process and feature distribution alignment status. Moreover, the input to the domain discriminator is only the feature vector output by the feature extractor, failing to integrate category conditional distribution information. This makes it difficult to achieve fine-grained cross-domain alignment at the category level, further limiting the recognition accuracy in multi-state precursor degradation scenarios. Summary of the Invention

[0006] To address the issue of insufficient accuracy in identifying pre-fuse features of fuses under various operating conditions, this invention provides a method for extracting pre-fuse features of fuses.

[0007] A method for extracting precursor features of a circuit breaker, comprising: S1, acquiring monitoring time-series data of circuit breakers in the source and target domains and extracting initial time-series features; S2, using a feature extractor of an adversarial training network to decouple the initial time-series features into a steady-state feature vector and an evolving feature vector, and calculating the subspace orthogonality loss between the two; S3, using a label predictor of the adversarial training network to output predicted labels and their probability distributions based on the two types of feature vectors, and calculating the label prediction loss; S4, using a domain discriminator of the adversarial training network to output a domain discrimination result based on the two types of feature vectors and the probability distribution, and calculating the domain discrimination loss; S5, updating the adversarial training network based on the label prediction loss, the domain discrimination loss, and the subspace orthogonality loss; wherein, the gradient weights assigned to the two types of feature vectors are determined according to the network iteration process and the subspace orthogonality loss, and the inverted gradient of the domain discrimination loss is backpropagated according to the gradient weights to update the feature extractor; S6, using the trained adversarial training network to identify precursor features of the circuit breaker in the target domain.

[0008] This invention converts monitoring data into initial time-frequency features and then uses an adversarial training network for cross-domain learning. The feature extractor decouples the initial features into steady-state and evolving vectors and applies an orthogonal loss to reduce the aliasing of the two types of information. The label predictor outputs a probability distribution, and the domain discriminator fuses this distribution with the two types of features to determine the source, achieving category-level cross-domain alignment. Gradient weights are dynamically allocated according to the iteration process and the orthogonal loss, and the inverted domain discriminant gradient is backpropagated to different feature branches according to the weights, achieving progressive alignment. This method, through feature decoupling and dynamic adjustment, reduces the mutual interference between steady-state background and degradation information under different operating conditions, improving the accuracy of precursor feature recognition.

[0009] Preferably, the extraction of initial time-series features includes: performing wavelet scattering transform and attention fusion processing on the monitoring time-series data: constructing a wavelet scattering network based on wavelet functions, performing wavelet convolution, modulus taking, and low-pass filtering operations on the monitoring time-series data sequentially to generate wavelet scattering coefficients of different orders, aligning them in the time dimension through cubic spline interpolation, and then concatenating them into a time-frequency feature matrix; inputting the time-frequency feature matrix into a channel attention network composed of a global average pooling layer and a linear mapping layer to calculate the attention weight coefficients corresponding to each channel; and performing element-wise multiplication of each channel of the time-frequency feature matrix with the corresponding attention weight coefficient to obtain the initial time-series features.

[0010] By extracting multi-scale time-frequency coefficients through wavelet scattering transform, we can effectively capture the weak non-stationary signal changes in the pre-fuse stage, while preserving the translation invariance and local features of the signal. By combining the channel attention mechanism to weight different time-frequency channels, we can enhance key feature components while suppressing redundant noise, thereby improving the expressive power and anti-interference performance of the initial time-series features.

[0011] Preferably, the step of decoupling the initial temporal features into a steady-state feature vector and an evolving feature vector using the feature extractor of the adversarial training network includes: extracting the initial temporal features using a shared backbone network containing one-dimensional convolutional layers and rectified linear units to obtain common latent features; calculating the temporal mean of the common latent features as a first intermediate feature, and calculating the first-order difference of the temporal dimension of the common latent features, flattening it to obtain a second intermediate feature; performing linear transformations on the first and second intermediate features using two fully connected branches respectively, and setting the output features obtained respectively as the steady-state feature vector and the evolving feature vector in sequence.

[0012] By extracting common latent features through a shared backbone network, the homology and information integrity of the two types of features are guaranteed; the mean of the time dimension is used to represent the global stationarity level of the signal, corresponding to steady-state information such as the working condition background; and the first-order time difference is used to represent the local change trend of the signal, corresponding to the evolution information of the degradation process, thus achieving effective separation of features with different physical attributes.

[0013] Preferably, the calculation of the subspace orthogonal loss of the two includes: calculating the inner product of the steady-state feature vector and the evolved feature vector after L2 normalization; taking the absolute value of the inner product as the basic orthogonal loss value, and averaging the basic orthogonal loss values ​​in the training batch as the subspace orthogonal loss.

[0014] Preferably, the label predictor of the adversarial training network outputs a predicted label and its probability distribution based on the two types of feature vectors, including: concatenating the steady-state feature vector and the evolutionary feature vector along the feature dimension to form a global joint feature vector, and inputting it into a feedforward neural network classifier containing a fully connected layer for hierarchical dimensionality reduction; applying a Softmax function at the output of the feedforward neural network classifier to convert the classification result into a probability distribution representing each category of the circuit breaker precursor state, and assigning the category number corresponding to the maximum probability value as the predicted label.

[0015] Preferably, the domain discriminator of the adversarial training network outputs a domain discrimination result based on the two types of feature vectors and the probability distribution, including: calculating the outer product of the probability distribution and the steady-state feature vector and the outer product of the probability distribution and the evolved feature vector, respectively, to obtain a first multidimensional joint feature tensor and a second multidimensional joint feature tensor; concatenating the first multidimensional joint feature tensor and the second multidimensional joint feature tensor along the channel, flattening them, performing dimensionality reduction and activation mapping operations, and outputting the domain discrimination result representing the source and target domains of the features.

[0016] Preferably, the calculation of label prediction loss includes: using a cross-entropy loss function to calculate the classification error between the probability distribution of the source domain training samples and the true labels of the samples, as the label prediction loss; the calculation of domain discrimination loss includes: based on the domain discrimination result, using a binary cross-entropy loss function to calculate the domain classification error, as the domain discrimination loss.

[0017] Using the binary cross-entropy loss function to measure the classification error of the domain discriminator provides an optimization objective for adversarial training. It maximizes the domain discrimination loss while minimizing the label prediction loss, thereby achieving domain invariance learning of features.

[0018] Preferably, determining the gradient weights assigned to the two types of feature vectors based on the network iteration process and the subspace orthogonal loss includes: calculating a decay coefficient within the range of 0 to 1 using an exponential decay function, based on the ratio of the current network parameter update iteration count to the maximum iteration count, combined with the normalized subspace orthogonal loss; using the decay coefficient as the first gradient weight assigned to the steady-state feature vector, and using the difference between the constant value 1 and the decay coefficient as the second gradient weight assigned to the evolving feature vector.

[0019] Preferably, the step of backpropagating the inverted domain discrimination loss gradient according to the gradient weights to update the feature extractor includes: in the backpropagation stage, inverting the sign of the gradient of the domain discrimination loss by using a gradient inversion layer with a scaling factor; separating the inverted gradient, multiplying it by the first gradient weight and the second gradient weight respectively, and using the weighted gradient information from both paths to backpropagate, correspondingly updating the branch parameters and backbone network parameters of the feature extractor that generate the steady-state feature vector and the evolving feature vector.

[0020] Preferably, the step of using the trained adversarial training network to identify the pre-fuse features of the target domain circuit breaker includes: freezing all network weight parameters of the feature extractor and the label predictor, and switching the network configuration to inference mode; acquiring real-time collected time-series monitoring data of the target domain circuit breaker, and after low-pass filtering and normalization preprocessing, forward propagating the data to the adversarial training network; obtaining the predicted label corresponding to the highest probability through the label predictor, and completing the online identification of the pre-fuse state.

[0021] By freezing network weights and switching to inference mode, and disabling random deactivation during the training phase, the stability of the model output is ensured. Real-time monitoring data is preprocessed in the same way as during the training phase and then input into the network to quickly output the prediction results of the pre-fuse state, thus realizing online real-time monitoring and early warning of fuses in the target domain of industrial sites.

[0022] The technical solution of the present invention has the following beneficial technical effects: This invention employs a wavelet scattering transform and channel attention fusion mechanism to extract multi-scale time-frequency features from multi-source monitoring time-series data. The features are decoupled into steady-state and evolving feature vectors, and subspace orthogonal constraints are applied. An adversarial training network containing a label conditional domain discriminator is constructed. A dynamic gradient weight allocation strategy based on iterative process and orthogonal loss is combined to achieve progressive feature alignment. This scheme can effectively capture weak non-stationary signal changes in the pre-fuse stage, achieve pure separation of features with different physical attributes, and complete category-level fine-grained cross-domain feature alignment. This overcomes the problems of insufficient pre-fuse feature recognition, weak cross-condition generalization ability, feature aliasing, and poor stability of the adversarial alignment process in existing technologies, significantly improving the accuracy of pre-fuse feature extraction in cross-domain scenarios. Attached Figure Description

[0023] Figure 1 A flowchart of a method for extracting features indicating impending fuse failure; Figure 2 A schematic diagram of the monitoring time series data window; Figure 3 A schematic diagram of gradient weight allocation; Figure 4 This is a performance comparison chart for different models. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0025] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0026] This invention discloses a method for extracting features indicating impending fuse failure, referring to... Figure 1 This includes the following steps: S1. Extract monitoring time series data and output initial time series features.

[0027] The purpose of this step is to acquire fuse operating status data from complex industrial environments and extract multi-scale time-frequency features that can reflect subtle changes in the precursory phase of fuse failure. Since the signal changes of fuses in the precursory phase are usually small in amplitude and easily masked by stable operating conditions, environmental noise, and load fluctuations, directly inputting the original time-domain signal into the diagnostic model is difficult to effectively capture multi-scale weak amplitude changes, which can easily lead to insufficient expression of precursory features.

[0028] Therefore, this embodiment introduces wavelet scattering transform and channel attention mechanism to extract high-frequency abrupt change information while preserving low-frequency steady-state information, and highlights effective time-frequency channels related to circuit breaker precursors through attention weighting. The specific implementation process is as follows: Acquire monitoring time-series data of fuses in both the source and target domains. The source domain data can be simulated environment or historical operating condition monitoring data with real-world labels indicating impending fuse failure, while the target domain data can be real-time monitoring data from actual industrial environments that lack labels or have insufficient labels. In one embodiment, the monitoring time-series data includes a one-dimensional time series of at least one of current, voltage, and temperature, acquired simultaneously.

[0029] Before inputting the data into the time-frequency decomposition and attention fusion module, the monitored time-series data can be low-pass filtered and normalized. For example, a Butterworth low-pass filter can be used to filter out high-frequency environmental noise, and the amplitude of each time series can be linearly normalized to a preset range of 0 to 1 to reduce the impact of differences in the dimensions of different acquisition channels on subsequent model training.

[0030] Subsequently, the monitoring time-series data is processed through a time-frequency decomposition and attention fusion module. Specifically, a two-layer wavelet scattering network is constructed based on the Morlet wavelet function. Wavelet convolution, modulus taking, and low-pass filtering operations are sequentially performed on the monitoring time-series data to generate wavelet scattering coefficients of order 0 to 2. Since the lengths of scattering coefficients of different orders may differ in the time dimension, the system performs cubic spline interpolation to align the wavelet scattering coefficients of different orders, and then concatenates the aligned scattering coefficients of each order along the channel dimension to form a time-frequency feature matrix.

[0031] In one implementation, the wavelet scattering coefficient can be expressed as: ; ; ; in, This represents the input monitoring timing signal; This represents the low-pass filter function; and Wavelet functions representing different scales; This represents the convolution operation; This represents the modulo operation. The above relationship is constructed based on the following: the 0th-order scattering coefficient mainly reflects the low-frequency smoothing component of the signal; the first-order scattering coefficient is used to characterize the local frequency response of the signal at different scales; and the second-order scattering coefficient further characterizes the envelope change and higher-order modulation information. By combining multiple-order scattering coefficients, the ability to express the weak time-frequency changes of the precursors to fuse failure can be improved to a certain extent.

[0032] The concatenated time-frequency feature matrix is ​​input into the channel attention network. Let the time-frequency feature matrix be: ; in, Indicates the number of channels; This represents the length of the time dimension. First, each channel is compressed along the time dimension using a global average pooling layer to obtain the channel description vector: ; in, Indicates the first The first channel in the Feature values ​​at each time location; Indicates the first Global description value for each channel; global description value for each channel Combining to form a channel description vector The purpose of this relationship is to compress the response intensity of each channel over the entire time window into a single statistic, so that the importance of that channel can be calculated subsequently.

[0033] Furthermore, the channel description vector Inputting a linear mapping layer yields intermediate activation features: ; The channel attention weights are then obtained through dimensionality upscaling and the Sigmoid activation function: ; in, , These represent the weight matrices of the linear mapping layer; , These represent the bias vectors respectively; Represents the ReLU activation function; This represents the Sigmoid activation function. The construction of this relationship is based on the following: by using a mapping structure that first reduces the dimensionality and then increases it, the correlation between different channels can be learned; the Sigmoid function constrains the output between 0 and 1, making it usable as a channel-level weight coefficient.

[0034] Finally, the initial temporal features are obtained by multiplying each channel of the time-frequency feature matrix element-wise with its corresponding attention weight: ; in, Indicates the first Attention weights for each channel; This represents the initial time-series features after weighting. The purpose of this relationship is to enhance the effective time-frequency channels related to circuit breaker precursors, while reducing the interference of redundant and noisy channels on subsequent model training.

[0035] For example, in a fuse monitoring scenario, the system collects multi-source signals such as current, voltage, and temperature at a sampling frequency of 10kHz, and constructs a monitoring time-series data window with a length of 1024 sampling points. After low-pass filtering and normalization, the window data is fed into a two-layer Morlet wavelet scattering network. The wavelet scattering network extracts the 0th-order coefficients reflecting the overall stationary level, the first-order coefficients reflecting the local frequency band response, and the second-order coefficients reflecting the envelope modulation changes. Since these scattering coefficients have different lengths in the time dimension, the system uses interpolation to uniformly align them to a fixed length of 16, and then splices them along the channel dimension to form a 37-channel time-frequency feature matrix. Figure 2As shown in the figure, this figure displays a typical waveform of the timing data window for fuse monitoring in an industrial field. The horizontal axis represents time, and the vertical axis represents the signal amplitude. The waveform exhibits the characteristics of the power frequency fundamental wave superimposed with high-frequency noise, which intuitively reflects the problem that the weak signal in the precursor stage of fuse failure is easily masked by the steady-state background and environmental noise. It is difficult to effectively extract the precursor features by directly using the original signal, thus verifying the necessity of the present invention to extract multi-scale time-frequency features using wavelet scattering transform and channel attention mechanism.

[0036] In this time-frequency feature matrix, some channels may correspond to stable frequency bands under normal load fluctuations, while others may contain weak abrupt changes caused by temperature rise disturbances, local melting, or changes in contact state during the precursor stage of melting. The channel attention network automatically generates weights based on the overall response of each channel within the time window and the relationships between channels, ensuring that channels related to precursor changes contribute more to subsequent features, while suppressing channels with weaker responses or those mainly composed of noise. The resulting initial time-series features retain the multi-scale time-frequency structure and provide a more effective input for decoupling subsequent steady-state and evolutionary features.

[0037] S2, decouples the output of two types of feature vectors and subspace orthogonal loss.

[0038] The purpose of this step is to decouple the steady-state operating condition information and the pre-fault evolution information mixed in the initial time-series features. The fuse monitoring signal contains both relatively stable background information such as ambient temperature and load level, and information that gradually increases with the degradation process. If these two types of information are directly aligned across domains as a whole feature, the operating condition background and fault evolution information can easily overlap, causing the fault features to be submerged by the background information, thus triggering negative transfer. By constructing parallel feature branches and applying subspace orthogonal loss, the network can be prompted to learn steady-state features and evolutionary features separately, improving the separability of feature representation. The specific implementation process is as follows: The initial temporal features are input into the feature extractor in the adversarial training network. This feature extractor includes a shared backbone network and two parallel feature mapping branches. The shared backbone network is used to extract general temporal features, and the two parallel branches are used to generate steady-state and evolutionary features, respectively.

[0039] First, a shared backbone network containing one-dimensional convolutional layers and rectified linear units is used to compute the initial temporal features, extracting common latent features. In one implementation, the first... The calculation of a one-dimensional convolution layer can be expressed as: ; in, Indicates the first Layer input features; Indicates the first Layer output features; and These represent the kernel weights and biases, respectively. This represents the convolution operation. This represents the batch normalization operation. The purpose of this relation is to extract local temporal patterns through convolution and to stabilize the feature distribution across different training batches through batch normalization.

[0040] Subsequently, two types of derivation processing are performed on the shared latent features. Let the shared latent features be... ,in Indicates channel dimension, This represents the number of time steps. The mean of the time dimension is calculated as the first intermediate feature: ; in, This is the first intermediate feature; Represents the shared hidden characteristic matrix In the The formula is based on the fact that the time mean can reflect the overall level and low-frequency stable components within the observation window, making it suitable for forming steady-state features.

[0041] Simultaneously, the first-order difference of the shared latent features in the time dimension is calculated and flattened to serve as the second intermediate feature: ; in, This is the second intermediate feature; This represents performing a first-order difference in the time dimension on the shared latent features, that is, subtracting the features from adjacent time steps; This means flattening a two-dimensional feature sequence into a one-dimensional vector according to channel priority. The basis for this relationship is that the difference between adjacent time steps can reflect the trend and degree of change of the signal over time, which is suitable for capturing the dynamic changes in the evolution process of the precursor to a circuit breaker.

[0042] Then, the first and second intermediate features are linearly transformed using two fully connected branches with the same output dimension to obtain the steady-state feature vector and the evolving feature vector: ; ; in, Represents the steady-state eigenvector; Represents the evolutionary eigenvector; , These represent the weight matrices for the corresponding branches; , These represent the bias vectors, respectively. With the above structure, the network can characterize the circuit breaker state from two perspectives: the global stationary level and the local trend of change.

[0043] After feature decoupling, the subspace orthogonality loss between the steady-state feature vector and the evolving feature vector is calculated. For a single sample, the basic orthogonality loss can be expressed as: ; in, Represents the basic orthogonal loss of a single sample, which is the absolute value of the cosine similarity between the two types of feature vectors after normalization, and is used to measure the degree of information overlap; and Let L2 norms represent the steady-state eigenvectors and the evolving eigenvectors, respectively. This represents the inner product operation of two vectors. The relationship is essentially the absolute value of the cosine similarity between two types of feature vectors after L2 normalization. Its construction is based on the fact that cosine similarity measures the correlation between two vectors along their directions; the closer this value is to 0, the more orthogonal the two feature directions are, and the less overlap in the information they express. Therefore, minimizing this value as a loss term can help the network reduce the correlation between steady-state features and evolutionary features.

[0044] For a training batch, the subspace orthogonality loss can be further expressed as: ; in, This represents the batch-level subspace orthogonality loss, which is the average of the basic orthogonality losses of all samples within the batch; Indicates the number of samples in the training batch; Indicates the first The basic orthogonal loss value corresponding to each sample. The role of this batch loss is to smooth the impact of abnormal fluctuations in individual samples on the training process, so that orthogonal constraints can stably participate in the backpropagation of the network.

[0045] For example, after the initial temporal feature matrix is ​​input into the feature extractor, the shared backbone network first extracts common latent features containing 256 channels and 16 time steps. If these common latent features are directly flattened globally and input into the classifier, the steady-state background and transient changes will be mixed in the same feature space, making it difficult to distinguish which information comes from environmental conditions and which comes from the evolution of circuit breaker precursors.

[0046] In this embodiment, the mean value is calculated along the time dimension to obtain a first intermediate feature that represents the overall level of the observation window; on the other hand, the difference between adjacent time steps is calculated to obtain a second intermediate feature that represents local fluctuations and state change trends. These two features are mapped to 128-dimensional steady-state feature vectors and evolutionary feature vectors via independent fully connected branches. If the two types of vectors still have a strong correlation in the early stages of training, the subspace orthogonal loss will penalize this correlation during backpropagation, allowing the network to gradually learn more separated feature representations.

[0047] S3. Based on the features, output the predicted label and the label prediction loss.

[0048] This step is used to output predicted labels and corresponding probability distributions for the circuit breaker precursor states based on the steady-state feature vector and the evolutionary feature vector. On one hand, this prediction result is used to complete the identification of the circuit breaker precursor states; on the other hand, this probability distribution also serves as input to the subsequent label conditional domain discriminator, enabling the cross-domain alignment process to incorporate category conditional information and avoid misalignment of features from different categories. The specific implementation process is as follows: steady-state eigenvectors and evolutionary eigenvectors By concatenating along the feature dimensions, a global joint feature vector is formed: ; in, This is the global joint feature vector; This represents a concatenation operation along the feature dimension. The joint feature vector simultaneously contains both the overall operating context information and the local degradation evolution information of the fuse.

[0049] The global joint feature vector is input into a feedforward neural network classifier containing three fully connected layers for dimensionality reduction and classification mapping. In one implementation, a Dropout layer can be added after some of the fully connected layers, with a dropout rate set to 0.5, to reduce the risk of overfitting in small-sample training scenarios. The last layer of the classifier outputs the non-standardized prediction scores for each category, denoted as: ; in, This indicates the total number of categories of warning signs before a circuit breaker is triggered; Indicates the first The unstandardized predicted scores for each category are further transformed into a probability distribution using the Softmax function. ; in, Indicates that the sample belongs to the first The probability of each category. This relationship is constructed based on the fact that the Softmax function can map any real-valued classification score to a non-negative probability and ensure that the sum of the probabilities of all categories is 1, thus facilitating the subsequent determination of the predicted label based on the maximum probability.

[0050] The predicted label can be represented as: ; in, This indicates the predicted category of the current sample, i.e., the category of the circuit breaker precursor state with the highest probability. Denotes the probability of taking. Largest category number For training samples with true labels in the source domain, the label prediction loss is calculated using the multi-class cross-entropy loss function: ; in, This represents the label prediction loss, which measures the deviation between the predicted probability distribution and the true label distribution. Indicates the number of training samples in the source domain; Indicates the first The source domain sample at the th th True one-hot encoding for each category, Indicates the corresponding sample number The predicted probabilities of each category; This is the natural logarithm function. The purpose of this relationship is to measure the deviation between the predicted probability distribution and the true label distribution. By minimizing this loss, the label predictor learns the discrimination rules for circuit breaker precursor states in the source domain samples.

[0051] For example, if the precursory states of a fuse are classified into five categories—normal operation, temperature rise warning, minor deformation, partial melting, and critical melting—consistent with the source domain labeling categories, that is... If the value is 5, the final layer of the label predictor outputs five classification scores. After transformation by the Softmax function, the system obtains a probability distribution for the five states. If the probability corresponding to slight deformation is the highest, the current monitoring window is determined to be in a slightly deformed state. This probability distribution is not only used to output predicted labels, but also enters the subsequent label conditional domain discriminator, enabling the domain discrimination process to incorporate state category information, rather than simply performing a general mixing and alignment of the source and target domains.

[0052] S4. Output the domain discrimination results and domain discrimination loss based on the distribution.

[0053] Traditional domain adaptation methods typically align only the overall feature distribution, which can easily lead to the problem of features from different categories being incorrectly grouped together; that is, features of class A in the source domain are incorrectly aligned with features of class B in the target domain. This step introduces a label conditionalization mechanism, inputting the class probability distribution output by the label predictor along with steady-state and evolutionary features into the domain discriminator. This allows the discriminator to be aware of class conditions when determining the source domain of features, thus facilitating finer-grained class-level cross-domain alignment. The specific implementation process is as follows: The label conditionalization domain discriminator takes the steady-state feature vector, the evolving feature vector, and the probability distribution vector output by the label predictor as input. Specifically, it calculates the outer product of the probability distribution vector and the steady-state feature vector, and the outer product of the probability distribution vector and the evolving feature vector, respectively. ; ; in, The steady-state joint feature tensor fuses class probabilities and steady-state features; The evolutionary joint feature tensor fuses class probabilities and evolutionary features; This represents the probability distribution vector output by the label predictor. Represents the steady-state eigenvector; Represents the evolutionary eigenvector. This represents the outer product operation. Through the outer product operation, each class probability can be applied to each feature dimension, thereby forming a high-order feature representation that includes the relationship between class conditions and feature dimensions.

[0054] Subsequently, the system respectively and Add a channel dimension to form a first multidimensional joint feature tensor and a second multidimensional joint feature tensor. Concatenate the two along the channel dimension and flatten them into a one-dimensional vector: ; in, The flattened domain discriminant input vector; This indicates a channel splicing operation. This indicates a flattening operation. The flattened vector is input into a fully connected domain discriminant network, which outputs the domain classification result. ; in, Indicate the domain discrimination output result; , The weight matrix in the domain discriminant network, , This indicates the corresponding bias. This represents the Sigmoid activation function. The purpose of this expression is to map high-dimensional categorical conditional features to domain probability values ​​between 0 and 1. For example, it can be agreed that an output value closer to 0 indicates that the sample is more likely to originate from the source domain, and an output value closer to 1 indicates that the sample is more likely to originate from the target domain.

[0055] Based on the domain discrimination results, the domain discrimination loss is calculated using the binary cross-entropy loss function: ; in, Domain discrimination loss is used to measure the ability of a domain discriminator to distinguish between the source domain and the target domain. This represents the total number of samples in the source and target domains involved in the domain discrimination. Indicates the first The true domain labels of each sample; This represents the domain discrimination output value of the corresponding sample. This relationship is used to measure the ability of the domain discriminator to distinguish between the source domain and the target domain. In subsequent adversarial training, this loss is used on the one hand to train the domain discriminator to improve its domain classification ability, and on the other hand, it acts inversely on the feature extractor through the gradient reversal layer, so that the feature extractor learns a cross-domain shared representation of the source domain that is more difficult to distinguish.

[0056] For example, if the label predictor outputs the probability distribution of five types of circuit breaker precursor states, and both the steady-state feature vector and the evolutionary feature vector are 128-dimensional, then after performing the outer product of the probability distribution with the two types of feature vectors respectively, two dimensions can be obtained. The joint feature tensor is used. This tensor doesn't simply concatenate class probabilities and features; instead, it establishes a correspondence between each class probability and each feature dimension. This allows the domain discriminator to more effectively determine the differences in feature distributions between the source and target domains under similar or identical conditions, rather than performing indiscriminate alignment of features across all categories. For example, for samples that both tend to trigger temperature rise warnings, the domain discriminator can determine their source domain differences based on the conditional features, thus aiding the subsequent feature extractor in performing cross-domain alignment under class conditions.

[0057] S5. Adversarial training network based on dynamic gradient weight updates.

[0058] This step is the core of cross-domain adversarial training. It uses label prediction loss, domain discrimination loss, and subspace orthogonality loss to jointly constrain the network. Furthermore, it dynamically allocates adversarial gradient weights between the steady-state and evolving branches based on the network's iteration progress and the subspace orthogonality loss, enabling the network to adjust the alignment strength between steady-state and evolving features at different training stages. This mechanism helps avoid training instability or negative transfer caused by fixed gradient weights, achieving a gradual alignment between steady-state and evolving features.

[0059] In each training iteration, the network calculates the label prediction loss based on the label predictor output. The domain discrimination loss is calculated based on the label-conditional domain discriminator output. The subspace orthogonality loss is calculated based on the steady-state eigenvectors and the evolutionary eigenvectors. The total loss function is a weighted sum of the label prediction loss, domain discrimination loss, and subspace orthogonality loss. The label prediction loss primarily ensures the state classification ability of the source domain samples; the domain discrimination loss aligns the feature distributions of the source and target domains; and the subspace orthogonality loss constrains the decoupling degree between steady-state and evolutionary features.

[0060] During the backpropagation phase, the gradient of the domain-discriminative loss is inverted when it passes through the gradient inversion layer. Let the scaling factor of the gradient inversion layer be... Then, for the gradient from the domain discriminative loss, its inverted form can be expressed as: ; in, This represents the gradient after processing by the gradient inversion layer. This represents all trainable parameters of the feature extractor. This represents the scaling factor of the gradient reversal layer, used to adjust the intensity of adversarial training, and is typically set to 1. This relationship is based on the following: the domain discriminator aims to minimize the domain discrimination loss to improve the ability to distinguish between the source and target domains; while the feature extractor, by receiving the inverse gradient, learns feature representations that can confuse the domain discriminator, thus achieving adversarial domain adaptation.

[0061] Furthermore, the system calculates the gradient weights assigned to the steady-state eigenvector and the evolving eigenvector based on the network iteration process and the subspace orthogonal loss. In one implementation, the first gradient weight assigned to the steady-state eigenvector can be expressed as: ; The second gradient weights assigned to the evolutionary eigenvectors can be expressed as: ; in, Indicates the current parameter update iteration count; Indicates the maximum number of iterations; This is a coefficient that controls the decay rate, usually taken as 5; The adjustment coefficient representing the subspace orthogonality loss is typically taken as 0.5; This represents the subspace orthogonality loss that is normalized and restricted to the range of 0 to 1; This means truncating the calculation result to a closed interval between 0 and 1. This represents the first gradient weight assigned to the steady-state feature branch; This represents the second gradient weight assigned to the evolutionary feature branch.

[0062] The construction of this relation is based on the fact that in the early stages of training, the network has not yet fully learned cross-domain shared representations, and the differences in steady-state background between the source and target domains are usually more pronounced. Therefore, by using a larger... Imparting higher adversarial gradient weights to the steady-state feature branch helps to preferentially mitigate the differences in background distribution between different domains. As training iterates, the exponential decay term gradually decreases. Correspondingly Increasing the weight of evolutionary feature branches gives them higher alignment weights in the later stages of training, thus focusing more on cross-domain adaptation of the degradation trend preceding circuit breaker. Simultaneously, introducing... The purpose is to adjust the alignment strength according to the feature decoupling state; when the steady-state features and evolutionary features are not yet fully orthogonal, a larger orthogonality loss will suppress the strong steady-state branch alignment and avoid premature domain alignment dominated by a single branch in the feature aliasing state.

[0063] During backpropagation of the domain discrimination loss, the gradient after being reversed by the gradient inversion layer is multiplied by the first gradient weight and the second gradient weight, respectively: ; ; in, This represents the weighted adversarial gradient backpropagated to the steady-state feature branch; This represents the weighted adversarial gradient propagated back to the evolutionary feature branch; and These represent the original gradients of the steady-state branch and the evolving branch after gradient reversal, respectively. The weighted gradients mentioned above are used to update the parameters of the fully connected branches that generate the steady-state feature vector and the evolving feature vector, and further affect the parameters of the shared backbone network.

[0064] like Figure 3 As shown in the figure, the gradient weight changes of the steady-state feature branch and the evolutionary feature branch during the entire training process are illustrated. The figure shows a dynamic change pattern in which the steady-state weight is higher in the early stage of training and the evolutionary weight gradually increases in the later stage. This is consistent with the gradient weight allocation strategy proposed in this invention, which reflects the design idea of ​​progressive cross-domain alignment. It avoids the problem of insufficient or over-alignment caused by fixed weights and effectively improves the stability of the adversarial training process.

[0065] For example, if the maximum number of iterations is set to 10,000, in the early stages of training, the current iteration number is small, and the exponential decay term is close to 1. At this time, the steady-state feature branch obtains a higher first gradient weight. This setting allows the model to prioritize handling the stable background shift between the source and target domains caused by temperature environment, load baseline, or equipment batch differences. As training progresses, the first gradient weight decreases according to the exponential decay law, and the second gradient weight increases accordingly. The model training focus gradually shifts to cross-domain adaptation of the evolution features of circuit breaker precursors.

[0066] If the subspace orthogonality loss remains high at a certain stage, it indicates a strong coupling between steady-state and evolving features. Continuing to prioritize the steady-state branch for alignment in this case would further exacerbate feature aliasing. In this situation, the weight allocation formula appropriately reduces the dominance of the steady-state branch, allowing the network to continue domain adaptation while further promoting feature decoupling through orthogonal loss constraints. Thus, the model can gradually complete cross-domain feature adaptation by first stabilizing background alignment and then strengthening evolutionary alignment, rather than applying a fixed proportion of adversarial gradients to the two branches throughout the entire training process.

[0067] S6. Utilize the characteristics of circuit breaker precursors identified by the output of adversarial networks.

[0068] This step is used to apply the trained feature extractor and label predictor to the target domain circuit breaker monitoring data after the network training is completed, so as to realize the extraction of circuit breaker precursor features and state recognition under the target operating conditions without the need for a large amount of labeled data in the target domain.

[0069] The network is considered converged when the total loss value decreases below a preset threshold over several consecutive evaluation periods (e.g., less than one-thousandth over ten consecutive evaluation periods), or when the current iteration count reaches the maximum iteration count. After convergence, the parameters of the feature extractor and label predictor are frozen, and the model is switched to inference mode. For example, the batch normalization layer uses the accumulated statistical mean and variance from the training phase, and the random deactivation operation of the Dropout layer is disabled to ensure stable inference output.

[0070] In practical applications, real-time monitoring time-series data of the target domain fuse is acquired and sequentially processed through low-pass filtering, normalization, wavelet scattering transform, and channel attention fusion to obtain initial time-series features. These initial time-series features are then input into a trained feature extractor to obtain steady-state and evolving feature vectors. The label predictor outputs the probability distribution of each fuse-breaking precursor state category based on these two feature vectors and selects the category with the highest probability as the prediction result for the current monitoring window.

[0071] For example, in a field operating environment, the system continuously collects current, voltage, or temperature sequences of fuses in a target domain and slices them according to a preset window length. Each data window, after preprocessing and time-frequency attention fusion, is input into the trained network model. If the probability distribution output by the label predictor shows the highest probability for the local melting category, the system can identify that window as a precursor to local melting; if the probability for the temperature rise warning category is the highest, it can output a temperature rise warning. Since this identification process relies on the steady-state and evolutionary features that have completed cross-domain adaptation, as well as the fuse precursor discrimination rules learned from the source domain, it can also utilize the source domain training knowledge to assist in the identification of fuse precursors under target operating conditions even when the target domain lacks sufficient labeled samples.

[0072] To verify the effectiveness of the technical solution of this invention, a cross-domain operating condition monitoring dataset for high-voltage fuses was used. The source domain contains simulated environmental time-series samples labeled with five types of pre-fuse warning states, while the target domain contains unlabeled time-series samples from actual industrial environments. The sampling rate was uniformly set to 10kHz, and the length of a single sample window was 1024 sampling points, consistent with the parameter settings in the aforementioned embodiments. The Adam optimizer was used during training, with a batch size of 64 and a maximum parameter update iteration count of 10,000.

[0073] To verify the function of each core module, multiple sets of comparative models were set up. The complete scheme includes wavelet scattering and channel attention fusion, steady-state / evolutionary feature decoupling, subspace orthogonal loss, label-conditionalized domain discriminator, and dynamic gradient weight allocation strategy; the comparative schemes respectively remove the channel attention mechanism, remove the steady-state and evolutionary feature decoupling structure, replace the label-conditionalized domain discriminator with a normal domain discriminator, and replace the dynamic gradient weight allocation with fixed weight allocation.

[0074] like Figure 4 As shown in the figure, this figure compares the classification accuracy and F1 score of the complete scheme of this invention with various comparative models on the target domain test set. The results show that the complete scheme of this invention has the best performance. After removing each core module, the performance decreases to varying degrees, with the performance decrease being the most significant after removing the feature decoupling module. Experimental results show that the combination of steady-state and evolutionary feature decoupling and subspace orthogonal loss helps to reduce the impact of feature aliasing, the label conditional domain discriminator helps to enhance the cross-domain alignment ability under class conditions, the dynamic gradient weight allocation strategy helps to improve the stability of the adversarial training process, and the channel attention mechanism helps to enhance key time-frequency channels and suppress redundant channels. The above results indicate that the method described in this embodiment can extract the pre-fuse features of the target domain fuse more stably in cross-operating scenario scenarios.

[0075] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for extracting precursor features of a fuse, characterized in that, include: S1. Obtain the monitoring timing data of the fuses in the source and target domains and extract the initial timing features; S2. The initial temporal features are decoupled into steady-state feature vectors and evolving feature vectors using the feature extractor of the adversarial training network, and the subspace orthogonality loss of the two is calculated. S3. The label predictor of the adversarial training network outputs the predicted label and its probability distribution based on the two types of feature vectors, and calculates the label prediction loss. S4. The domain discriminator of the adversarial training network outputs a domain discrimination result based on the two types of feature vectors and the probability distribution, and calculates the domain discrimination loss. S5. Update the adversarial training network based on the label prediction loss, domain discrimination loss, and subspace orthogonal loss; wherein, the gradient weights assigned to the two types of feature vectors are determined according to the network iteration process and the subspace orthogonal loss, and the inverted domain discrimination loss gradient is backpropagated according to the gradient weights to update the feature extractor; S6. Use the trained adversarial network to identify the pre-fuse features of the target domain fuse.

2. The method for extracting precursor features of a fuse according to claim 1, characterized in that, The extraction of initial temporal features includes: performing wavelet scattering transform and attention fusion processing on the monitored temporal data. A wavelet scattering network is constructed based on wavelet functions. Wavelet convolution, modulus taking, and low-pass filtering are performed sequentially on the monitoring time series data to generate wavelet scattering coefficients of different orders. After time dimension alignment by cubic spline interpolation, the coefficients are spliced ​​into a time-frequency feature matrix. The time-frequency feature matrix is ​​input into a channel attention network consisting of a global average pooling layer and a linear mapping layer to calculate the attention weight coefficients corresponding to each channel. The initial time-series features are obtained by performing element-wise multiplication of each channel of the time-frequency feature matrix with the corresponding attention weight coefficients.

3. The method for extracting precursor features of a fuse according to claim 1, characterized in that, The step of decoupling the initial temporal features into steady-state feature vectors and evolutionary feature vectors using the feature extractor of the adversarial training network includes: using a shared backbone network containing one-dimensional convolutional layers and rectified linear units to extract the initial temporal features and obtain common latent features; The mean of the time dimension of the common latent features is calculated as the first intermediate feature, and the first difference of the time dimension of the common latent features is calculated, flattened, and used as the second intermediate feature. Two fully connected branches are used to perform linear transformations on the first intermediate feature and the second intermediate feature, respectively, and the resulting output features are set as the steady-state feature vector and the evolution feature vector, respectively.

4. The method for extracting precursor features of a fuse according to claim 1, characterized in that, The calculation of the subspace orthogonality loss of the two includes: calculating the inner product of the steady-state feature vector and the evolution feature vector after L2 normalization; The absolute value of the inner product is taken as the basic orthogonal loss value, and the average value of the basic orthogonal loss values ​​in the training batch is taken as the subspace orthogonal loss.

5. The method for extracting precursor features of a fuse according to claim 1, characterized in that, The label predictor of the adversarial training network outputs predicted labels and their probability distributions based on the two types of feature vectors, including: concatenating the steady-state feature vector and the evolved feature vector along the feature dimension to form a global joint feature vector, and inputting it into a feedforward neural network classifier containing a fully connected layer for hierarchical dimensionality reduction; The Softmax function is applied to the output of the feedforward neural network classifier to convert the classification result into a probability distribution of each category representing the pre-circuit breaker state, and the category number corresponding to the maximum probability value is assigned as the prediction label.

6. The method for extracting precursor features of a fuse according to claim 1, characterized in that, The domain discriminator of the adversarial training network outputs a domain discrimination result based on the two types of feature vectors and the probability distribution, including: calculating the outer product of the probability distribution and the steady-state feature vector and the outer product of the probability distribution and the evolved feature vector, respectively, to obtain a first multidimensional joint feature tensor and a second multidimensional joint feature tensor; The first multidimensional joint feature tensor and the second multidimensional joint feature tensor are concatenated along the channel, flattened, and then dimensionality reduction and activation mapping operations are performed to output the domain discrimination result representing the source and target domains of the features.

7. The method for extracting precursor features of a fuse according to claim 1, characterized in that, The calculation of the label prediction loss includes: using the cross-entropy loss function to calculate the classification error between the probability distribution of the source domain training samples and the true labels of the samples, as the label prediction loss; The calculation of the domain discrimination loss includes: calculating the domain classification error using a binary cross-entropy loss function based on the domain discrimination result, and using this error as the domain discrimination loss.

8. The method for extracting precursor features of a fuse according to claim 1, characterized in that, The step of determining the gradient weights assigned to the two types of feature vectors based on the network iteration process and the subspace orthogonal loss includes: calculating and limiting the decay coefficient to the range of 0 to 1 using an exponential decay function based on the ratio of the current network parameter update iteration number to the maximum iteration number, combined with the normalized subspace orthogonal loss; The decay coefficient is used as the first gradient weight assigned to the steady-state feature vector, and the difference between the constant value 1 and the decay coefficient is used as the second gradient weight assigned to the evolution feature vector.

9. The method for extracting precursor features of a fuse according to claim 8, characterized in that, The step of backpropagating the inverted domain discrimination loss gradient according to the gradient weight to update the feature extractor includes: in the backpropagation stage, reversing the sign of the gradient of the domain discrimination loss by using a gradient inversion layer with a scaling factor. The inverted gradients are separated and multiplied by the first gradient weight and the second gradient weight respectively. The weighted gradient information from the two paths is then used to backtrack and update the branch parameters and backbone network parameters of the feature extractor that generate the steady-state feature vector and the evolving feature vector.

10. The method for extracting precursor features of a fuse according to claim 1, characterized in that, The step of using the trained adversarial network to identify the circuit breaker precursor features of the target domain circuit breaker includes: freezing all network weight parameters of the feature extractor and the label predictor, and switching the network configuration to inference mode. The system acquires real-time monitoring time-series data of the target domain fuse, performs low-pass filtering and normalization preprocessing, and then propagates the data forward to the adversarial training network. The label predictor obtains the predicted label corresponding to the highest probability, thus completing the online identification of the pre-fuse state of the fuse.

Images