Multi-dimensional data joint diagnosis method and device for hidden fault of handcart circuit breaker

Abstract

The application provides a multi-dimensional data joint diagnosis method and device for hidden faults of handcart circuit breakers, relates to the field of handcart circuit breaker detection, and comprises the following steps: constructing a chaotic fractal feature extraction model, performing chaotic feature extraction, fractal feature extraction and core feature screening on multi-dimensional data through the chaotic fractal feature extraction model, and obtaining a core feature set; mapping the core feature set into a hidden fault feature vector through a meta-learning network; capturing probability features and spatial features of hidden faults according to the hidden fault feature vector through a capsule network, constructing a capsule feature vector through the probability features and the spatial features; inputting the capsule feature vector into a feature fusion network to perform semantic feature extraction, feature weight adjustment and feature fusion, and obtaining a fusion feature vector; inputting the fusion feature vector into a hidden fault prediction network to identify hidden faults of handcart circuit breakers, and obtaining a hidden fault diagnosis result.

Description

Technical Field

[0001] This application relates to the field of circuit breaker testing, and in particular to a multi-dimensional data-based diagnostic method and apparatus for latent faults in circuit breakers. Background Technology

[0002] As a critical control and protection device in the power system, the operating status of the handcart circuit breaker directly affects the safety and stability of the power grid. Latent faults in handcart circuit breakers refer to faults that are difficult to identify in a timely manner using traditional detection methods. Long-term accumulation of these faults can easily lead to serious accidents such as equipment malfunction, failure to operate, or even burnout. Common types of latent faults include contact oxidation, coil aging, mechanism jamming, core corrosion, insufficient contact pressure, insufficient excitation, loose components, and environmentally induced anomalies—a total of eight categories. Latent faults are difficult to identify in a timely manner using traditional detection methods due to their weak early signals, highly nonlinear evolution process, and insignificant differences from normal operating conditions. The long-term accumulation of these faults can easily lead to serious accidents such as equipment malfunction, failure to operate, or even burnout, causing huge economic losses and social impact. Currently, there are many shortcomings in the diagnostic technology for latent faults in handcart circuit breakers, resulting in insufficient accuracy in diagnosing these faults.

[0003] First, traditional feature extraction methods often focus on linear features of single-dimensional data, making it difficult to capture the chaotic evolution patterns and complex data structures associated with latent faults. This results in weak fault features being masked by noise, leading to extremely low identification accuracy. Second, existing diagnostic models often only focus on the probability features of fault occurrence, ignoring the spatial distribution information of features. This makes it difficult to accurately distinguish between latent faults with similar structures, resulting in a high misjudgment rate. Third, the diagnostic results do not consider the bias caused by uneven sample distribution and lack a scientific calibration mechanism. As a result, the diagnostic accuracy in practical applications is easily affected by the data distribution and fluctuates, making it difficult to meet the power system's demand for accurate early warning of latent faults. Summary of the Invention

[0004] To address the aforementioned problems, in a first aspect, this invention provides a multi-dimensional data-based joint diagnosis method for latent faults in handcart circuit breakers, comprising: Multi-dimensional data of the handcart circuit breaker is collected by a distributed sensor array. A chaotic fractal feature extraction model is constructed. The chaotic fractal feature extraction model is used to extract chaotic features, fractal features and core features from the multi-dimensional data to obtain a set of core features. A latent fault diagnosis model is constructed, which includes a meta-learning network, a capsule network, a feature fusion network, and a latent fault prediction network. The core feature set is mapped to a latent fault feature vector through a meta-learning network; Capsule networks capture the probabilistic and spatial features of latent faults based on latent fault feature vectors, and construct capsule feature vectors using these probabilistic and spatial features. The capsule feature vector is input into the feature fusion network for semantic feature extraction, feature weight adjustment and feature fusion to obtain the fused feature vector. The fused feature vectors are input into the latent fault prediction network to identify latent faults in the handcart circuit breaker and obtain the latent fault diagnosis results.

[0005] Optionally, the step of collecting multi-dimensional data of the handcart circuit breaker through a distributed sensor array includes: The mechanical vibration data, electromagnetic induction data, contact resistance data, and environmental correlation data of the handcart circuit breaker are collected by a distributed sensor array, forming multi-dimensional data.

[0006] Optionally, the step of extracting chaotic features, fractal features, and core features from multi-dimensional data using a chaotic fractal feature extraction model to obtain a core feature set includes: The chaotic feature extraction module of the chaotic fractal feature extraction model calculates the Lyapunov exponent of mechanical vibration data, electromagnetic induction data, contact resistance data and environmental correlation data, and uses the Lyapunov exponent as a chaotic feature. The fractal feature extraction module of the chaotic fractal feature extraction model calculates the fractal dimension of each data in multi-dimensional data and uses the fractal dimension as the fractal feature. Chaotic features and fractal features are used as candidate features to obtain latent fault label features. The mutual information between the candidate features and the latent fault label features is calculated by K-nearest neighbor estimation. The candidate features are sorted in descending order according to the mutual information, and a preset number of candidate features at the top of the sort are selected as core features. All core features constitute the core feature set.

[0007] Optionally, the meta-learning network includes a feature embedding layer, a meta-feature aggregation layer, and a mapping output layer. The mapping of the core feature set to latent fault feature vectors through the meta-learning network includes: The core feature set is divided into multiple task feature sets according to the type of latent fault. The task feature sets are non-linearly transformed through the feature embedding layer to obtain task meta-feature vectors. The meta-feature aggregation layer uses an attention mechanism to perform association weight aggregation on the task meta-feature vectors to obtain aggregated meta-feature vectors. The aggregated meta-feature vectors are then mapped to latent fault feature vectors through the mapping output layer.

[0008] Optionally, the capsule network includes an input capsule layer, a convolutional capsule layer, and an output capsule layer. The capsule network captures the probabilistic and spatial features of latent faults based on the latent fault feature vector, and constructs a capsule feature vector using these probabilistic and spatial features, including: The initial capsule feature vector is obtained by increasing the feature dimension of the latent fault feature vector through the input capsule layer. By performing multi-channel convolution operations on the initial capsule feature vector through convolutional capsule layers, local fault feature vectors are obtained. The local fault features in the local fault feature vectors are used as primary capsules to obtain a set of primary capsules. The softmax function of the output capsule layer generates a set of coupling coefficients, and the primary capsule set is weighted and adjusted using the set of coupling coefficients to obtain the final capsule set. Calculate the L2 norm and dimensionality component of each final capsule in the final capsule set. Use the L2 norm as the probabilistic feature of the latent fault and the dimensionality component as the spatial feature of the latent fault. Use the probabilistic feature and the spatial feature as the capsule feature of the final capsule. The capsule feature vector is composed of all the capsule features.

[0009] Optionally, the feature fusion network includes dilated convolutional layers and attention layers. The step of inputting the capsule feature vector into the feature fusion network for semantic feature extraction, feature weight adjustment, and feature fusion to obtain a fused feature vector includes: High-dimensional semantic features of capsule feature vectors are obtained by extracting high-dimensional semantic feature vectors through dilated convolutional layers; The capsule feature vector is optimized by adjusting the weights of the capsule feature vector through the attention mechanism of the attention layer. The high-dimensional semantic feature vector and the optimized capsule feature vector are summed element-wise with weights to obtain the fused feature vector.

[0010] Optionally, the latent fault prediction network includes a feature enhancement layer, a classification decision layer, and a result calibration layer. The step of inputting the fused feature vector into the latent fault prediction network to identify latent faults in the handcart circuit breaker and obtain latent fault diagnosis results includes: By performing deep semantic mining and dimensionality compression on the fused feature vector through the feature enhancement layer, a low-dimensional high-resolution feature vector is obtained. By using a multi-layer fully connected network in the classification decision layer, low-dimensional high-resolution feature vectors are mapped to initial latent fault probability vectors. The calibration layer uses a Bayesian posterior probability correction algorithm to calibrate the initial latent fault probability vector to obtain the final latent fault probability vector. The latent fault category with the highest probability in the final latent fault probability vector is taken as the latent fault diagnosis result.

[0011] Secondly, the present invention provides a multi-dimensional data joint diagnosis device for latent faults in handcart circuit breakers, used to implement the aforementioned multi-dimensional data joint diagnosis method for latent faults in handcart circuit breakers, the device comprising: The core feature set acquisition module is used to collect multi-dimensional data of the handcart circuit breaker through a distributed sensor array, construct a chaotic fractal feature extraction model, and perform chaotic feature extraction, fractal feature extraction and core feature selection on the multi-dimensional data through the chaotic fractal feature extraction model to obtain the core feature set. The latent fault diagnosis model construction module is used to build a latent fault diagnosis model, which includes a meta-learning network, a capsule network, a feature fusion network, and a latent fault prediction network. The latent fault feature vector acquisition module is used to map the core feature set into latent fault feature vectors through a meta-learning network; The capsule feature vector acquisition module is used by the capsule network to capture the probabilistic and spatial features of latent faults based on the latent fault feature vector, and to construct the capsule feature vector through the probabilistic and spatial features. The fusion feature vector acquisition module is used to input the capsule feature vector into the feature fusion network for semantic feature extraction, feature weight adjustment and feature fusion to obtain the fusion feature vector; The latent fault diagnosis result acquisition module is used to input the fused feature vector into the latent fault prediction network to identify latent faults in the handcart circuit breaker and obtain the latent fault diagnosis result.

[0012] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the multi-dimensional data joint diagnosis method for latent faults of the handcart circuit breaker.

[0013] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the multi-dimensional data joint diagnosis method for latent faults of handcart circuit breakers.

[0014] The present invention has the following beneficial effects: 1. A chaotic fractal feature extraction model captures the chaotic evolution patterns and complex structures of multi-dimensional data, and then uses mutual information filtering to remove redundant features, accurately mining the weak and nonlinear features of hidden faults, significantly improving the identification of hidden fault features; a meta-learning network effectively adapts to scenarios with scarce hidden fault samples by aggregating features across fault tasks, improving diagnostic performance; a capsule network simultaneously captures probabilistic and spatial features, quantifying the probability of fault occurrence through L2 norm and preserving feature spatial distribution information through dimensional components, achieving accurate differentiation of similar faults; a feature fusion network highlights key features while capturing long-range semantic associations, making the fused features more discriminative; a hidden fault prediction network corrects initial prediction bias through Bayesian posterior probability, avoiding diagnostic bias caused by uneven sample distribution, greatly improving the diagnostic accuracy of hidden faults in handcart circuit breakers.

[0015] 2. The capsule network consists of an input capsule layer, a convolutional capsule layer, and an output capsule layer. The input capsule layer achieves feature upscaling, providing sufficient dimensionality support for subsequent in-depth fault feature mining and avoiding the loss of key information in low-dimensional features. The multi-channel convolution of the convolutional capsule layer can accurately capture local fault features, and the set of primary capsules formed by splitting makes the fault feature expression more refined and comprehensive. The output capsule layer generates coupling coefficients through the softmax function, realizing weighted optimization of primary capsules, highlighting the contribution of features strongly correlated with faults, and suppressing redundant interference. At the same time, it extracts probabilistic features and spatial features, quantifying the probability of fault occurrence and retaining feature spatial distribution information, perfectly solving the pain point of difficulty in distinguishing similar latent faults, and providing highly recognizable capsule feature vectors for subsequent accurate diagnosis.

[0016] 3. The latent fault prediction network consists of a feature enhancement layer, a classification decision layer, and a result calibration layer. The feature enhancement layer, through residual blocks and bottleneck layer structures, achieves both deep semantic mining and dimensionality compression, reducing computational complexity while retaining key fault information. The low-dimensional, high-resolution feature vector output lays the foundation for accurate classification. The classification decision layer, with its multi-layer fully connected network and regularization mechanism, effectively avoids overfitting and stably maps feature vectors to initial latent fault probability vectors, initially quantifying the probability of various fault occurrences. The result calibration layer's Bayesian posterior probability correction algorithm, combined with prior fault probability, corrects initial prediction bias, perfectly solving the diagnostic bias problem caused by uneven sample distribution, significantly improving the reliability and accuracy of diagnostic results, and making fault category judgments more consistent with actual operating scenarios. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a structural diagram of the device according to an embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0020] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0021] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0022] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0023] Reference Figure 1 This invention provides a multi-dimensional data-based joint diagnosis method for latent faults in handcart circuit breakers, including: The S100 collects multi-dimensional data from the handcart circuit breaker using a distributed sensor array, constructs a chaotic fractal feature extraction model, and then uses this model to extract chaotic features, fractal features, and core features from the multi-dimensional data to obtain a set of core features.

[0024] In some embodiments, the acquisition of multi-dimensional data of the handcart circuit breaker via a distributed sensor array includes: The mechanical vibration data, electromagnetic induction data, contact resistance data, and environmental correlation data of the handcart circuit breaker are collected by a distributed sensor array, forming multi-dimensional data.

[0025] In some embodiments, high-precision piezoelectric accelerometers are used to collect mechanical vibration data. These accelerometers are distributed and installed in key mechanical parts of the circuit breaker, such as the closing and opening mechanism boxes, frame guide rails, and contact seats. The sampling rate is set to 1kHz-10kHz to continuously collect vibration signals during the circuit breaker's opening and closing operations and steady-state operation. The vibration amplitude, frequency spectrum, and transient impact waveforms are recorded to capture vibration anomalies caused by latent faults such as component wear, jamming, and loosening.

[0026] Electromagnetic induction data is acquired using current and voltage sensors, which are installed at the power input terminals of the closing and opening coils, and near the electromagnet core, respectively. The sampling rate is set to 5kHz-20kHz, simultaneously acquiring the current change curves and core excitation voltage signals during coil switching, while also recording electromagnetic induction intensity distribution data to accurately capture electromagnetic signal distortions caused by latent faults such as coil aging, core corrosion, and insufficient excitation.

[0027] Contact resistance data is collected using a micro-ohm contact resistance sensor, which is installed on the connection face of the moving and stationary contacts and at the busbar connector. A four-terminal measurement method is used to eliminate lead resistance interference. The sampling interval is set to 10 seconds / time to continuously collect the real-time value of the contact resistance, while recording the resistance fluctuation data under changes in current load, and identifying abnormal increases in resistance caused by latent faults such as contact oxidation, insufficient pressure, and poor contact.

[0028] Temperature and humidity sensors, dust concentration sensors, and air pressure sensors are installed inside the circuit breaker cabinet, around its exterior, and in the control room. Sampling intervals are set to 5 minutes per sampling, collecting ambient temperature, humidity, dust content, and atmospheric pressure data as environmental correlation data. Sudden changes in environmental parameters are also recorded to provide data support for analyzing equipment malfunctions induced by environmental factors.

[0029] The data acquisition terminal receives data from various sensors in real time, encapsulates it in a standardized format of timestamp + sensor number + data type to form structured multi-dimensional data, stores it in the local database and uploads it synchronously to the diagnostic system backend.

[0030] In some embodiments, the step of extracting chaotic features, fractal features, and core features from multi-dimensional data using a chaotic fractal feature extraction model to obtain a core feature set includes: The chaotic feature extraction module of the chaotic fractal feature extraction model calculates the Lyapunov exponent of mechanical vibration data, electromagnetic induction data, contact resistance data and environmental correlation data, and uses the Lyapunov exponent as a chaotic feature. The fractal feature extraction module of the chaotic fractal feature extraction model calculates the fractal dimension of each data in multi-dimensional data and uses the fractal dimension as the fractal feature. Chaotic features and fractal features are used as candidate features to obtain latent fault label features. The mutual information between the candidate features and the latent fault label features is calculated by K-nearest neighbor estimation. The candidate features are sorted in descending order according to the mutual information, and a preset number of candidate features at the top of the sort are selected as core features. All core features constitute the core feature set.

[0031] In some embodiments, the data of each type in the multi-dimensional data are first preprocessed. A 5th-order Butterworth filter is used to remove power frequency interference and high-frequency noise, and then Z-score normalization is applied to normalize the data to the [0,1] interval. The data embedding dimension m and time delay τ are determined using the CC algorithm, and a phase space is constructed based on the embedding dimension m and time delay τ. The evolution of neighboring trajectories in the phase space is tracked based on the Wolf criterion, and the mean of the trajectory separation rate is calculated as the Lyapunov exponent for each data point. The calculation formula is:

[0032] Where N represents the total number of samples in a single class of data; Let represent the Euclidean distance between the i-th trajectory in the phase space at time t and its neighboring trajectories; Indicates the interval time Then, the Euclidean distance between the i-th trajectory and its corresponding neighboring trajectories. A single data type corresponds to one core chaotic feature, and multi-dimensional data generates a total of four chaotic features.

[0033] The box-counting method is used to calculate the fractal dimension. For each type of preprocessed data, an N×N grid covering the data value range is constructed, and the number of non-empty boxes covering the data points at different grid scales is counted. A linear fit is performed on ln(1 / ε) and lnN(ε), and the slope of the fitted line is the fractal dimension D, calculated using the following formula: in, For grid scale, This represents the number of non-empty boxes. Each data type corresponds to one fractal feature, and a total of four fractal features are generated from multi-dimensional data.

[0034] The latent fault label features adopt a one-hot encoding method to label the fault type corresponding to the data, such as contact oxidation, coil aging, etc., for a total of 8 latent faults. In the K-nearest neighbor estimation, K is set to 7, and the mutual information between each candidate feature and the label feature is calculated based on Euclidean distance. The calculation formula is as follows:

[0035] in, This represents the mutual information between candidate feature X and fault label feature Y; The marginal information entropy represents the candidate feature X; The edge information entropy represents the fault label feature Y; This represents the joint information entropy of X and Y. After sorting in descending order of mutual information values, a preset number is determined based on retaining a cumulative mutual information ratio of ≥85%, ultimately resulting in the core feature set.

[0036] S200 constructs a latent fault diagnosis model, which includes a meta-learning network, a capsule network, a feature fusion network, and a latent fault prediction network.

[0037] In some embodiments, the core purpose of the meta-learning network is to transform high-dimensional core features into latent fault feature vectors with cross-fault task relevance; the core purpose of the capsule network is to extract fault probability features and spatial distribution features from latent fault features to construct capsule feature vectors; the core purpose of the feature fusion network is to fuse high-dimensional semantic features and key optimization features to improve fault feature recognition; and the core purpose of the latent fault prediction network is to mine deep semantics from fused features and output accurate fault diagnosis results.

[0038] The S300 uses a meta-learning network to map the core feature set into latent fault feature vectors.

[0039] In some embodiments, the meta-learning network includes a feature embedding layer, a meta-feature aggregation layer, and a mapping output layer. The mapping of the core feature set to a latent fault feature vector via the meta-learning network includes: The core feature set is divided into multiple task feature sets according to the type of latent fault. The task feature sets are non-linearly transformed through the feature embedding layer to obtain task meta-feature vectors. The meta-feature aggregation layer uses an attention mechanism to perform association weight aggregation on the task meta-feature vectors to obtain aggregated meta-feature vectors. The aggregated meta-feature vectors are then mapped to latent fault feature vectors through the mapping output layer.

[0040] In some embodiments, eight types of latent faults in the circuit breaker are first identified, including contact oxidation, coil aging, mechanism jamming, core corrosion, insufficient contact pressure, insufficient excitation, loose components, and environmentally induced anomalies. The core feature set is divided according to the fault type-sample mapping relationship, with each fault type corresponding to one task feature set, forming a total of eight task feature sets. Each task feature set is split using a support set + query set, with the support set accounting for 30% of the samples in that type and the query set accounting for 70%. The sample dimension of a single task feature set is the number of samples × the number of core features.

[0041] The feature embedding layer employs a two-layer fully connected neural network structure, with each layer including a BatchNorm layer and a Dropout layer to prevent overfitting. The LeakyReLU activation function enhances non-linear expressive power. The support set and query set samples of each task feature set are input into the embedding layer, and the process proceeds through the following steps: Input → Fully Connected 1 → BatchNorm → LeakyReLU → Dropout → Fully Connected 2 → BatchNorm → LeakyReLU, outputting the task-specific meta-features corresponding to each core feature. , where t is the number of the fault task set, i is the number of the core feature, and the task meta-feature vector is composed of all task meta-features.

[0042] A multi-head scaling dot product attention mechanism is adopted. For each task meta-feature, a query vector Q, a key vector K, and a value vector V are generated through three independent fully connected layers. The attention weights of each task meta-feature are obtained through Softmax normalization. The task meta-features are input into the attention heads to obtain the output features. The outputs of eight attention heads are concatenated and fused through a single fully connected layer to obtain the aggregated meta-feature vector, realizing feature association and aggregation across fault types. The expression is as follows: in, This represents the aggregated meta-feature corresponding to the i-th task meta-feature. This represents the attention weight of the k-th attention head on the i-th task meta-feature. This represents the output feature of the k-th attention head.

[0043] The output layer is a single fully connected network with an input dimension equal to the aggregate feature vector dimension and an output dimension equal to the latent fault feature vector dimension. No activation function is added. L2 regularization is used to constrain the parameters and prevent overfitting. The final expression for the output latent fault features is:

[0044] in, This represents the latent fault feature corresponding to the i-th aggregate element feature. This represents the weight matrix of the fully connected network that maps the output layer. This represents the bias vector of the fully connected network that maps the output layer.

[0045] The S400 capsule network captures the probabilistic and spatial features of latent faults based on the latent fault feature vector, and constructs capsule feature vectors through the probabilistic and spatial features.

[0046] In some embodiments, the capsule network includes an input capsule layer, a convolutional capsule layer, and an output capsule layer. The capsule network captures the probabilistic and spatial features of latent faults based on latent fault feature vectors, and constructs capsule feature vectors using the probabilistic and spatial features, including: The initial capsule feature vector is obtained by increasing the feature dimension of the latent fault feature vector through the input capsule layer. By performing multi-channel convolution operations on the initial capsule feature vector through convolutional capsule layers, local fault feature vectors are obtained. The local fault features in the local fault feature vectors are used as primary capsules to obtain a set of primary capsules. The softmax function of the output capsule layer generates a set of coupling coefficients, and the primary capsule set is weighted and adjusted using the set of coupling coefficients to obtain the final capsule set. Calculate the L2 norm and dimensionality component of each final capsule in the final capsule set. Use the L2 norm as the probabilistic feature of the latent fault and the dimensionality component as the spatial feature of the latent fault. Use the probabilistic feature and the spatial feature as the capsule feature of the final capsule. The capsule feature vector is composed of all the capsule features.

[0047] In some embodiments, the input capsule layer adopts a single-layer fully connected network structure, and the input is a latent fault feature vector. The goal is to upscale low-dimensional features to a high-dimensional space suitable for convolutional operations. This involves setting the weight matrix of the fully connected layers. Bias vector The activation function used is ReLU, and BatchNorm and Dropout layers are added to prevent overfitting. .

[0048] The convolutional capsule layer employs a 1D multi-channel convolutional structure with 32 convolutional kernels, each kernel having a size of [missing information]. The stride is set to 1. The initial capsule feature vector is converted into a pseudo-sequence format, input into the convolutional layer, and output as a multi-channel feature map. This feature map is then split by channel, with each channel corresponding to one local fault feature vector. A transformation operation further splits each channel feature vector into 32 local fault features of dimension 32. Each local fault feature serves as a primary capsule, ultimately forming a set of primary capsules. ,in This represents the k-th primary capsule.

[0049] The number of output capsule layers corresponds to the number of latent fault categories, and each output capsule needs to be associated with all primary capsules through a coupling coefficient. First, a coupling coefficient matrix is ​​constructed. Through the softmax function The function normalizes each row of the matrix, generating a set of coupling coefficients. , For the final capsule index, For primary capsule index, satisfying ,Right now Then, a weighted summation is used to map the primary capsule to the final capsule: ,in This represents the i-th final capsule, forming the final capsule set. .

[0050] For each final capsule Calculate the L2 norm The formula is , This is the notation for L2 norm calculation. Normalizing the L2 norm to the [0,1] interval reflects the probability of occurrence of the corresponding fault category for that capsule. The final capsule is directly selected. 32-dimensional elements as spatial features This characterizes the spatial distribution pattern of fault features. Each final capsule's feature vector is a concatenated vector of probabilistic and spatial features. The feature vectors of the eight final capsules are concatenated in order of fault category to form the final capsule feature vector. .

[0051] The S500 inputs the capsule feature vector into the feature fusion network for semantic feature extraction, feature weight adjustment, and feature fusion to obtain the fused feature vector.

[0052] In some embodiments, the feature fusion network includes dilated convolutional layers and attention layers. The step of inputting the capsule feature vector into the feature fusion network for semantic feature extraction, feature weight adjustment, and feature fusion to obtain a fused feature vector includes: High-dimensional semantic features of capsule feature vectors are obtained by extracting high-dimensional semantic feature vectors through dilated convolutional layers; The capsule feature vector is optimized by adjusting the weights of the capsule feature vector through the attention mechanism of the attention layer. The high-dimensional semantic feature vector and the optimized capsule feature vector are summed element-wise with weights to obtain the fused feature vector.

[0053] In some embodiments, a parallel multi-scale dilated convolutional structure is employed to adapt to the capsule feature vector. Each branch has 16 1D convolutional kernels with dilation rates of 1, 2, and 3, and a stride of 1. The capsule feature vector is converted to a standard input format, and each branch performs independent convolution computation. The outputs of the three branches are then processed. Channel concatenation is performed to obtain multi-scale feature maps, which are then compressed to 264 dimensions through a 1×1 convolutional layer. After ReLU activation and BatchNorm processing, a high-dimensional semantic feature vector is output. .

[0054] capsule feature vector Perform global average pooling to obtain a 1-dimensional channel statistics vector. Channel statistical vectors are processed through a two-layer fully connected network. A non-linear transformation is performed, with ReLU+Sigmoid as the activation function, and the output channel attention weights are used. Multiply the capsule feature vector by the attention weights element-wise, channel by channel: ,in The Hadamard product is used to obtain the optimized capsule feature vector. Highlight the contribution of key fault characteristic channels.

[0055] Feature fusion is achieved using learnable weight coefficients, with two trainable parameters set. and ,satisfy The fusion formula is: ,in and The algorithm employs gradient descent for optimization, using the cross-entropy loss function for fault diagnosis to dynamically balance the importance of high-dimensional semantic features and optimized capsule features. The final output is a fused feature vector. This is used as input for subsequent latent fault prediction networks.

[0056] The S600 integrates feature vectors into a latent fault prediction network to identify latent faults in the handcart circuit breaker and obtain latent fault diagnosis results.

[0057] In some embodiments, the latent fault prediction network includes a feature enhancement layer, a classification decision layer, and a result calibration layer. The step of inputting the fused feature vector into the latent fault prediction network to identify latent faults in the handcart circuit breaker and obtain latent fault diagnosis results includes: By performing deep semantic mining and dimensionality compression on the fused feature vector through the feature enhancement layer, a low-dimensional high-resolution feature vector is obtained. By using a multi-layer fully connected network in the classification decision layer, low-dimensional high-resolution feature vectors are mapped to initial latent fault probability vectors. The calibration layer uses a Bayesian posterior probability correction algorithm to calibrate the initial latent fault probability vector to obtain the final latent fault probability vector. The latent fault category with the highest probability in the final latent fault probability vector is taken as the latent fault diagnosis result.

[0058] In some embodiments, the feature enhancement layer adopts a residual block + bottleneck layer structure to adapt and fuse feature vectors. Each residual block contains three 1D convolutional layers. The first 1×1 convolutional layer compresses the input dimension from 2^64 to 64, reducing computation. The second 3×1 convolutional layer mines deep semantic associations, using GELU as the activation function. The third 1×1 convolutional layer restores the dimension, and overfitting is prevented through BatchNorm and Dropout layers. Two concatenated residual blocks are set up, and finally, global average pooling is used to compress the feature dimension to 64, outputting a low-dimensional, high-resolution feature vector. .

[0059] The classification decision layer adopts a two-layer fully connected network structure. The first layer has an input dimension of 64 and an output dimension of 32, with LeakyReLU activation function and L2 regularization. The second layer has an input dimension of 32 and an output dimension of 8, with Softmax activation function, mapping the low-dimensional high-resolution feature vector to the initial latent fault probability vector. Each element in the vector Let the initial probability of a sample belonging to the k-th type of fault be expressed as follows: .

[0060] Based on historical fault statistics of the circuit breaker, the prior probabilities of eight types of faults were determined. If no historical data is available, a uniform prior is used. The initial probabilities output by the classification decision layer are used. Construct the likelihood function, calibrate the initial probability according to Bayes' theorem, and calculate the formula as follows:

[0061] in, Let be the final posterior probability of the k-th type of latent fault; Let be the likelihood function, representing the observed low-dimensional high-resolution feature vector under the condition that the input data belongs to the k-th type of fault. The probability is directly taken from the initial probability. .

[0062] The final latent fault probability vector is composed of all the final posterior probabilities. The fault category corresponding to the element with the highest probability in the final latent fault probability vector is taken as follows: As the final hidden fault diagnosis result, the final hidden fault diagnosis result corresponds to one of the eight preset hidden faults of the handcart circuit breaker, namely contact oxidation, coil aging, mechanism jamming, iron core corrosion, insufficient contact pressure, insufficient excitation, loose components, and environmentally induced abnormalities.

[0063] Reference Figure 2 This invention provides a multi-dimensional data joint diagnosis device 20 for latent faults in handcart circuit breakers, used to realize a multi-dimensional data joint diagnosis method for latent faults in handcart circuit breakers. The device includes: The core feature set acquisition module 21 is used to collect multi-dimensional data of the handcart circuit breaker through a distributed sensor array, construct a chaotic fractal feature extraction model, and perform chaotic feature extraction, fractal feature extraction and core feature screening on the multi-dimensional data through the chaotic fractal feature extraction model to obtain the core feature set. The latent fault diagnosis model construction module 22 is used to construct a latent fault diagnosis model, which includes a meta-learning network, a capsule network, a feature fusion network, and a latent fault prediction network. The latent fault feature vector acquisition module 23 is used to map the core feature set into latent fault feature vectors through a meta-learning network; The capsule feature vector acquisition module 24 is used by the capsule network to capture the probabilistic and spatial features of latent faults based on the latent fault feature vector, and to construct the capsule feature vector through the probabilistic and spatial features. The feature vector fusion acquisition module 25 is used to input the capsule feature vector into the feature fusion network for semantic feature extraction, feature weight adjustment and feature fusion to obtain the fused feature vector. The latent fault diagnosis result acquisition module 26 is used to input the fused feature vector into the latent fault prediction network to identify latent faults in the handcart circuit breaker and obtain latent fault diagnosis results.

[0064] This application provides an electronic device, including a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements a multi-dimensional data joint diagnosis method for latent faults in handcart circuit breakers according to any of the above schemes.

[0065] Specifically, the processor may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor may also include onboard memory for caching purposes. The processor may be a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of this application.

[0066] Memory can be any medium capable of containing, storing, transmitting, propagating, or transmitting instructions. For example, memory can include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, instruments, or propagation media. Specific examples of memory include: magnetic storage devices such as magnetic tape or hard disk drives (HDDs); optical storage devices such as optical discs (CD-ROMs); and also random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0067] This application also provides a computer-readable medium storing a computer program that, when executed by a processor, implements a multi-dimensional data-driven joint diagnosis method for latent faults in handcart circuit breakers, as described in any of the above-described schemes. This computer-readable medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into that device / apparatus / system. The aforementioned computer-readable medium carries one or more programs, which, when executed, implement the methods described in the embodiments of this application.

[0068] According to embodiments of this application, a computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wired, optical fiber, radio frequency signals, etc., or any suitable combination thereof.

[0069] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments, but should be defined not only by the appended claims, but also by their equivalents. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A multi-dimensional data-based joint diagnosis method for latent faults in handcart circuit breakers, characterized in that, include: Multi-dimensional data of the handcart circuit breaker is collected by a distributed sensor array. A chaotic fractal feature extraction model is constructed. The chaotic fractal feature extraction model is used to extract chaotic features, fractal features and core features from the multi-dimensional data to obtain a set of core features. A latent fault diagnosis model is constructed, which includes a meta-learning network, a capsule network, a feature fusion network, and a latent fault prediction network. The core feature set is mapped to a latent fault feature vector through a meta-learning network; Capsule networks capture the probabilistic and spatial features of latent faults based on latent fault feature vectors, and construct capsule feature vectors using these probabilistic and spatial features. The capsule feature vector is input into the feature fusion network for semantic feature extraction, feature weight adjustment and feature fusion to obtain the fused feature vector. The fused feature vectors are input into the latent fault prediction network to identify latent faults in the handcart circuit breaker and obtain the latent fault diagnosis results.

2. The multi-dimensional data joint diagnosis method for latent faults in handcart circuit breakers according to claim 1, characterized in that, The multi-dimensional data of the handcart circuit breaker collected through the distributed sensor array includes: The mechanical vibration data, electromagnetic induction data, contact resistance data, and environmental correlation data of the handcart circuit breaker are collected by a distributed sensor array, forming multi-dimensional data.

3. The multi-dimensional data joint diagnosis method for latent faults in handcart circuit breakers according to claim 2, characterized in that, The process involves using a chaotic fractal feature extraction model to extract chaotic features, fractal features, and core features from multi-dimensional data, resulting in a core feature set, including: The chaotic feature extraction module of the chaotic fractal feature extraction model calculates the Lyapunov exponent of mechanical vibration data, electromagnetic induction data, contact resistance data and environmental correlation data, and uses the Lyapunov exponent as a chaotic feature. The fractal feature extraction module of the chaotic fractal feature extraction model calculates the fractal dimension of each data in multi-dimensional data and uses the fractal dimension as the fractal feature. Chaotic features and fractal features are used as candidate features to obtain latent fault label features. The mutual information between the candidate features and the latent fault label features is calculated by K-nearest neighbor estimation. The candidate features are sorted in descending order according to the mutual information, and a preset number of candidate features at the top of the sort are selected as core features. All core features constitute the core feature set.

4. The multi-dimensional data joint diagnosis method for latent faults in handcart circuit breakers according to claim 1, characterized in that, The meta-learning network includes a feature embedding layer, a meta-feature aggregation layer, and a mapping output layer. The mapping of the core feature set to latent fault feature vectors via the meta-learning network includes: The core feature set is divided into multiple task feature sets according to the type of latent fault. The task feature sets are non-linearly transformed through the feature embedding layer to obtain task meta-feature vectors. The meta-feature aggregation layer uses an attention mechanism to perform association weight aggregation on the task meta-feature vectors to obtain aggregated meta-feature vectors. The aggregated meta-feature vectors are then mapped to latent fault feature vectors through the mapping output layer.

5. The multi-dimensional data joint diagnosis method for latent faults in handcart circuit breakers according to claim 1, characterized in that, The capsule network comprises an input capsule layer, a convolutional capsule layer, and an output capsule layer. The capsule network captures the probabilistic and spatial features of latent faults based on latent fault feature vectors, and constructs capsule feature vectors using these probabilistic and spatial features, including: The initial capsule feature vector is obtained by increasing the feature dimension of the latent fault feature vector through the input capsule layer. By performing multi-channel convolution operations on the initial capsule feature vector through convolutional capsule layers, local fault feature vectors are obtained. The local fault features in the local fault feature vectors are used as primary capsules to obtain a set of primary capsules. The softmax function of the output capsule layer generates a set of coupling coefficients, and the primary capsule set is weighted and adjusted using the set of coupling coefficients to obtain the final capsule set. Calculate the L2 norm and dimensionality component of each final capsule in the final capsule set. Use the L2 norm as the probabilistic feature of the latent fault and the dimensionality component as the spatial feature of the latent fault. Use the probabilistic feature and the spatial feature as the capsule feature of the final capsule. The capsule feature vector is composed of all the capsule features.

6. The multi-dimensional data joint diagnosis method for latent faults in handcart circuit breakers according to claim 1, characterized in that, The feature fusion network includes dilated convolutional layers and attention layers. The process of inputting the capsule feature vector into the feature fusion network for semantic feature extraction, feature weight adjustment, and feature fusion to obtain a fused feature vector includes: High-dimensional semantic features of capsule feature vectors are obtained by extracting high-dimensional semantic feature vectors through dilated convolutional layers; The capsule feature vector is optimized by adjusting the weights of the capsule feature vector through the attention mechanism of the attention layer. The high-dimensional semantic feature vector and the optimized capsule feature vector are summed element-wise with weights to obtain the fused feature vector.

7. The multi-dimensional data joint diagnosis method for latent faults in handcart circuit breakers according to claim 1, characterized in that, The latent fault prediction network includes a feature enhancement layer, a classification decision layer, and a result calibration layer. The step of inputting the fused feature vector into the latent fault prediction network to identify latent faults in the handcart circuit breaker and obtain latent fault diagnosis results includes: By performing deep semantic mining and dimensionality compression on the fused feature vector through the feature enhancement layer, a low-dimensional high-resolution feature vector is obtained. By using a multi-layer fully connected network in the classification decision layer, low-dimensional high-resolution feature vectors are mapped to initial latent fault probability vectors. The calibration layer uses a Bayesian posterior probability correction algorithm to calibrate the initial latent fault probability vector to obtain the final latent fault probability vector. The latent fault category with the highest probability in the final latent fault probability vector is taken as the latent fault diagnosis result.

8. A multi-dimensional data joint diagnosis device for latent faults in a handcart circuit breaker, used to implement the multi-dimensional data joint diagnosis method for latent faults in a handcart circuit breaker as described in any one of claims 1 to 7, characterized in that, The device includes: The core feature set acquisition module is used to collect multi-dimensional data of the handcart circuit breaker through a distributed sensor array, construct a chaotic fractal feature extraction model, and perform chaotic feature extraction, fractal feature extraction and core feature selection on the multi-dimensional data through the chaotic fractal feature extraction model to obtain the core feature set. The latent fault diagnosis model construction module is used to build a latent fault diagnosis model, which includes a meta-learning network, a capsule network, a feature fusion network, and a latent fault prediction network. The latent fault feature vector acquisition module is used to map the core feature set into latent fault feature vectors through a meta-learning network; The capsule feature vector acquisition module is used by the capsule network to capture the probabilistic and spatial features of latent faults based on the latent fault feature vector, and to construct the capsule feature vector through the probabilistic and spatial features. The fusion feature vector acquisition module is used to input the capsule feature vector into the feature fusion network for semantic feature extraction, feature weight adjustment and feature fusion to obtain the fusion feature vector; The latent fault diagnosis result acquisition module is used to input the fused feature vector into the latent fault prediction network to identify latent faults in the handcart circuit breaker and obtain the latent fault diagnosis result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the multi-dimensional data joint diagnosis method for latent faults in handcart circuit breakers as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-dimensional data joint diagnosis method for latent faults of handcart circuit breakers as described in any one of claims 1 to 7.

Images