Multi-channel feature extraction and fault diagnosis method and system for on-load tap changers

Abstract

This invention discloses a multi-channel feature extraction and fault diagnosis method and system for on-load tap changers, addressing the problems of strong vibration signal noise interference, difficulty in extracting early weak fault features, and insufficient fault identification accuracy in existing on-load tap changer mechanical condition monitoring. The method includes the following steps: S1. Multi-channel vibration signal acquisition and sample construction; S2. Phase space reconstruction and multi-channel attractor tensor construction; S3. Tensor singular value decomposition and singular subspace expansion; S4. Amplitude-preserving non-convex tensor low-rank separation and noise reduction; S5. Multi-order feature reconstruction and determination of the optimal reconstruction order; S6. Generation of time-domain feature samples and time-frequency feature tensors; S7. Bi-branch multi-scale attention feature learning; S8. Cross-domain feature fusion and mechanical condition recognition. This invention can effectively enhance weak fault features under complex background noise conditions, improve the separability between samples of different mechanical conditions, and significantly improve the accuracy and stability of mechanical condition recognition.

Description

Technical Field

[0001] This invention relates to the field of on-load tap changer fault diagnosis technology, specifically to a method and system for multi-channel feature extraction and fault diagnosis of on-load tap changers. Background Technology

[0002] With the development of multi-sensor testing methods, multi-channel vibration signal fusion analysis has gradually become an important direction for OLTC mechanical fault diagnosis. However, most existing signal processing methods still model data in vector or matrix form, making it difficult to maintain the inherent coupling relationship of multi-channel signals in time, space, and dynamic structure. Tensors, as a natural expression of multi-dimensional data, can describe multi-channel information and its inherent correlations within a unified framework, thus providing a new approach for feature modeling of complex mechanical vibration signals. Existing tensor decomposition methods, such as typical multinomial decomposition (CPD), can extract effective information from multi-channel data to a certain extent, but their rank selection is difficult, and their ability to express nonlinear and non-stationary dynamic characteristics is limited. In contrast, combining phase space reconstruction with tensor singular value decomposition can complete the separation of multi-channel feature subspaces while recovering the nonlinear dynamic structure of the signal, making it more suitable for handling complex vibration responses during OLTC switching processes.

[0003] Nevertheless, two prominent problems remain in the feature extraction of OLTC vibration faults based on tensor decomposition. First, noise is typically distributed across multiple singular subspaces. While traditional tensor robust principal component analysis methods based on convex threshold contraction can achieve a certain degree of noise reduction and separation, their uniform contraction mechanism easily weakens high-amplitude fault features, leading to excessive suppression of early, weak fault information during the noise reduction process, thus affecting subsequent diagnostic results. Second, the selection of the reconstruction order after tensor decomposition has a decisive impact on the feature extraction results. However, existing methods often rely on empirical settings or use difference spectra and inflection point criteria for order selection, making them susceptible to interference components and lacking specific characterization of impact-type fault features, thus making it difficult to stably determine the optimal reconstruction order.

[0004] On the other hand, existing OLTC mechanical fault diagnosis methods are mostly based on manual experience features or single feature domain analysis, often using only one type of information from the time domain, frequency domain, or local statistics for identification, making it difficult to simultaneously consider the temporal evolution characteristics and frequency band energy distribution characteristics of the fault impact. For multiple types of mechanical faults, especially early weak faults and similar fault types, single feature representation methods are prone to insufficient inter-class discrimination, thus limiting the accuracy of fault identification. With the development of intelligent diagnostic methods, how to effectively combine high-quality fault feature extraction results with deep discriminative models has become an important research direction for improving the performance of OLTC mechanical condition monitoring. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for multi-channel feature extraction and fault diagnosis of on-load tap changers, which solves the problems of strong vibration signal noise interference, difficulty in extracting early weak fault features, and insufficient fault identification accuracy in the existing mechanical condition monitoring of on-load tap changers.

[0006] The technical solution adopted by the present invention to solve its technical problem is: a method for multi-channel feature extraction and fault diagnosis of on-load tap changers, including the following steps.

[0007] S1. Multi-channel vibration signal acquisition and sample construction.

[0008] The original vibration signals of multiple channels during a single switching process of the on-load tap changer are collected to obtain the time-domain sequence corresponding to each vibration measurement channel; the collected signals are then truncated, aligned, and normalized.

[0009] S2. Phase space reconstruction and multichannel attractor tensor construction.

[0010] The original vibration signals are reconstructed in phase space to obtain the corresponding attractor matrices; then the attractor matrices of each channel are stacked along the channel dimension to construct a multi-channel attractor tensor.

[0011] S3. Tensor Singular Value Decomposition and Singular Subspace Expansion.

[0012] Tensor singular value decomposition is performed on the multichannel attractor tensor to decompose the original vibration signal into different singular subspaces, thereby obtaining the corresponding tensor singular value tensor and characteristic components of each order.

[0013] S4. Amplitude-preserving non-convex tensor low-rank separation noise reduction.

[0014] A low-rank separation strategy for amplitude-preserving non-convex tensors is introduced to perform low-rank feature separation and sparse interference suppression on attractor tensors, thereby obtaining denoised low-rank feature tensors.

[0015] S5. Multi-order feature reconstruction and determination of the optimal reconstruction order.

[0016] The denoised low-rank feature tensor is expanded according to the singular value order, and reconstructed sequentially using the first k order components to obtain the reconstructed feature signals corresponding to each order. Furthermore, the singular value kurtosis index of the tensor and its relative rate of change are constructed. By searching for the candidate position with the largest absolute value of the relative rate of change, the optimal reconstruction order is adaptively determined, thereby outputting a pure fault feature signal.

[0017] S6. Generation of time-domain feature samples and time-frequency feature tensors.

[0018] The fault feature signal obtained in step S5 is used as the input sample of the diagnostic model, and a short-time Fourier transform is performed on it to construct the corresponding time-frequency feature tensor.

[0019] S7. Two-branch multi-scale attention feature learning.

[0020] A dual-branch, multi-scale attention fusion fault diagnosis model is constructed, and the reconstructed time-domain and time-frequency features are deeply mined to achieve intelligent identification of the mechanical state of on-load tap changers.

[0021] S8. Cross-domain feature fusion and mechanical state recognition.

[0022] Gated fusion is performed on the discriminative features extracted from the time-domain branch and the time-frequency branch to construct a fused deep representation, which is then input into the classification layer. The output is the posterior probability of the mechanical state corresponding to the on-load tap changer and the final diagnostic category.

[0023] Furthermore, in step S1, the original vibration signal sequence of the c-th channel acquired by the on-load tap changer during one switching process. (1); where C s This indicates the total number of vibration measurement channels; N represents the number of sampling points for a single channel signal. Indicates the transpose operation; This represents the amplitude of the c-th channel at the Nth sampling time.

[0024] Furthermore, in step S2, phase space reconstruction is performed on the signals of each channel to obtain the attractor matrix of the c-th channel. (2); where, Indicates a time delay; Indicates the embedding dimension; Let represent the effective sampling length after reconstruction, and satisfy: (3).

[0025] Furthermore, in step S2, the multichannel attractor tensor (4); where, This represents the tensor construction operation that stacks along the channel dimension; This represents the third-order real-valued tensor space.

[0026] Furthermore, in step S3, the formula is used. (5) Perform tensor singular value decomposition on the attractor tensor T; where U represents the left orthogonal tensor and V represents the right orthogonal tensor; Represents a singular value tensor; Represents t-product; This indicates the transpose of a tensor.

[0027] Furthermore, in step S4, an amplitude-preserving non-convex tensor low-rank separation strategy is introduced to perform amplitude-preserving denoising on the attractor tensor, and an optimization model is established. (6); where, This represents the low-rank feature tensor after denoising; Represents the sparse disturbance tensor; Tensor The qth singular value; This represents the total number of singular values; Represents the generalized nonconvex penalty function; Indicates the sparse constraint weights; This represents the L1 norm.

[0028] Furthermore, in step S5, the denoised low-rank feature tensor is... Expanding by singular value order and reconstructing using the first k components, we obtain the k-th order feature tensor: (7); where, This represents the feature tensor reconstructed from the first k singular components; The tensor component corresponding to the q-th singular value is represented by k; k represents the reconstruction order; then... By performing inverse tensor transformation and inverse phase space mapping, the corresponding k-th order reconstructed feature signal is obtained. .

[0029] Furthermore, in step S5, the optimal reconstruction order is adaptively determined using the tensor singular value kurtosis (TSVK) as an order selection index. Let the kurtosis of the k-th order reconstructed signal on the c-th channel be... (8); where, This is the time-domain sequence of the k-th order reconstructed signal on the c-th channel; for The mean; for Standard deviation; This represents the expectation operation; further, the tensor singular value kurtosis of the k-th order reconstructed signal is defined. (9); redefine the relative rate of change of TSVK for adjacent orders. (10); by calculating all orders And determine the position corresponding to the largest absolute value, and determine the optimal reconstruction order; Let: (11); where I represents the candidate position corresponding to the optimal reconstruction order; This represents the index that maximizes the objective function; when When, it indicates that the first-order reconstructed signal F I It exhibits more prominent impact characteristics, and in this case, the optimal reconstruction order is I; when When the optimal reconstruction order is I+1, it indicates that the reconstruction result of order I+1 is better. In this case, the optimal reconstruction order is taken as I+1. Finally, the feature signal determined by the optimal reconstruction order is used as the fault feature signal output by this invention.

[0030] Furthermore, in step S6, after determining the optimal reconstruction order I, the first-order extracted signal F is...I As input to the fault diagnosis model; let the single-action multi-channel time-domain feature sample matrix be truncated, aligned, and normalized. (12); where n3 is the number of vibration measurement channels; m t The length of a single action sample; This indicates a sequence of n3 rows and m... t D is a matrix space composed of real elements; a short-time Fourier transform is performed on D to construct the corresponding time-frequency feature tensor. (13); where, Indicates the short-time Fourier transform operation; m s For time frames; This represents the third-order real-valued tensor space.

[0031] Furthermore, in step S7, the time-domain branch uses multi-scale one-dimensional convolution and bidirectional gated recurrent units to model the fault impact pulse, envelope fluctuations and temporal evolution characteristics; the time-frequency branch uses two-dimensional convolution and channel-space joint attention mechanism to deeply mine the fault frequency band, harmonic structure and local energy accumulation region.

[0032] Furthermore, in step S7, in the temporal branch, multi-scale one-dimensional convolution is used to extract features from D in parallel; let the set of convolution kernel scales be... The temporal feature response extracted at the u-th convolution scale (14); where u represents the convolution scale index; This represents the length of the u-th one-dimensional convolution kernel; Indicates the kernel length is One-dimensional convolution operation; For batch normalization; The activation function is nonlinear; the outputs at each scale are concatenated to obtain a multi-scale representation in the time domain. (15); where, Indicates feature concatenation operation; This represents the output features corresponding to each convolution scale.

[0033] Furthermore, in step S7, a channel attention allocation mechanism is introduced in the time-domain branch to adaptively weight the contributions of multiple channels, resulting in a channel attention weight vector. (16); where, This represents the normalized exponential function; Indicates global average pooling; For activation functions; and The trainable parameter matrix; the weighted time-domain features are represented as: (17); where, This indicates an element-wise weighted operation based on the channel; subsequently, a bidirectional gated loop unit is used to perform... The temporal correlation is modeled to obtain the temporal discriminant vector. (18); where, This represents a bidirectional gated cyclic cell network; This represents pooling operations; in the time-frequency branch, the time-frequency feature tensor G is used as input, and two-dimensional convolution is used to extract the fault frequency band and local energy accumulation region: (19); where B represents the initial time-frequency features extracted by two-dimensional convolution; This represents a two-dimensional convolution operation.

[0034] Furthermore, in step S7, a channel-space joint attention mechanism is introduced into the time-frequency branch. First, the channel attention coefficients are constructed. (20); where, This represents the Sigmoid activation function; Represents a multilayer perceptron mapping; This represents the result after performing global average pooling on B; This represents the result after global max pooling of B; after channel weighting, we get: (21); where, This represents the time-frequency features after channel attention weighting; further, the spatial attention mapping is defined. (22); where, [;] indicates a 7×7 convolution operation; [;] indicates feature concatenation; This represents the average pooling operation; Represents max pooling operation; enhanced time-frequency features (23); and the time-frequency discrimination vector is obtained by pooling. (twenty four).

[0035] Furthermore, in step S8, a gating fusion mechanism is constructed to adaptively couple the dual-branch features, and a gating weight vector is defined for dual-branch feature fusion. (25); where, Represents the trainable parameter moments of the fusion gate layer; Represents the bias vector of the fusion gate layer; This indicates that the time-domain discriminant vector With time-frequency discriminant vector The data is spliced ​​together; the resulting deep representation is: (26); where h represents the fused deep feature vector; This represents the trainable parameter matrix of the fusion mapping layer; Represents the bias vector of the fusion mapping layer; This represents the weighting coefficient that complements the gating weights; This represents the element-wise interaction term between the time-domain discriminant vector and the time-frequency discriminant vector; in the output layer, the fused feature h is input into the classifier to obtain the posterior probability vector corresponding to each mechanical state category. (27); where, This represents the trainable parameter matrix of the classification layer; This represents the bias vector of the classification layer; the corresponding fault identification result is: (28); where, This represents the probability that the sample belongs to the j-th type of mechanical state; Indicates taking The category index that reaches the maximum value; This is the final diagnostic category.

[0036] Furthermore, in step S8, cross-entropy loss and center constraint loss are introduced to form a joint optimization objective function. (29); where, Represents the cross-entropy loss term; Indicates the central constraint loss term; This represents the weighting coefficient used to balance the contributions of the two types of losses; (30); (31); v represents the batch sample size; v represents the batch sample index. Here, j represents the number of mechanical state categories; j represents the category index. Let be the label value of the v-th sample for the j-th class; c represents the center vector of the class to which the v-th sample belongs in the fused feature space; v This represents the true class label of the v-th sample; This represents the square of the L2 norm.

[0037] This invention also provides a multi-channel feature extraction and fault diagnosis system for on-load tap changers, used to execute a multi-channel feature extraction and fault diagnosis method for on-load tap changers. The system includes a multi-channel vibration sensing module, a signal conditioning module, a synchronization triggering module, a multi-channel data acquisition module, a data processing and system control module, a data storage module, and a communication module. The multi-channel vibration sensing module includes multiple vibration sensors, which are installed on the operating mechanism housing, transmission mechanism connection, drive motor mounting base, transmission bearing housing, and switching switch support structure of the on-load tap changer, used to acquire vibration signals from multiple measurement points during a single switching process of the on-load tap changer. The input terminal of the signal conditioning module is connected to the output terminal of the multi-channel vibration sensing module, used to perform charge conversion, amplification, filtering, anti-aliasing processing, and impedance matching on the vibration signals of each channel. The synchronization triggering module is connected to the action control circuit, auxiliary contacts, drive motor current detection unit, or position switch of the on-load tap changer, used to acquire the on-load tap changer's... The system initiates a switching action and outputs a synchronous trigger signal to the multi-channel data acquisition module. The multi-channel data acquisition module, connected to the signal conditioning module and the synchronous trigger module, synchronously acquires and converts the vibration signals of each channel to digital data under the action of the synchronous trigger signal, forming multi-channel raw vibration data. The data processing and system control module, connected to the multi-channel data acquisition module, receives and processes the multi-channel raw vibration data, performing signal interception, alignment, normalization, phase space reconstruction, tensor feature extraction, amplitude-preserving noise reduction, multi-order feature reconstruction, time-frequency feature generation, cross-domain feature fusion, and mechanical condition identification, outputting the mechanical condition diagnosis results of the on-load tap changer. The data storage module, connected to the data processing and system control module, stores the raw vibration data, feature data, diagnostic model parameters, historical diagnostic results, and condition assessment records. The communication module, connected to the data processing and system control module, uploads the mechanical condition diagnosis results to the condition monitoring platform, substation backend system, or maintenance terminal.

[0038] The beneficial effects of this invention are: this invention can effectively enhance weak fault characteristics and improve the separability between samples of different mechanical states under complex background noise conditions, and significantly improve the accuracy and stability of mechanical state identification. The method of this invention has effectiveness, superiority and engineering application value in on-load tap changer mechanical state monitoring and fault diagnosis. Attached Figure Description

[0039] Figure 1 This is a flowchart of the diagnostic method of the present invention;

[0040] Figure 2 Here is the confusion matrix diagram corresponding to A1;

[0041] Figure 3 Here is the confusion matrix diagram corresponding to A2;

[0042] Figure 4 Here is the confusion matrix diagram corresponding to A3;

[0043] Figure 5 Here is the confusion matrix diagram corresponding to A4;

[0044] Figure 6 This is the confusion matrix diagram corresponding to the method of the present invention. Detailed Implementation

[0045] This invention aims to propose a multi-channel feature extraction and fault diagnosis method and system for on-load tap changers. It fully leverages the advantages of phase space reconstruction and tensor decomposition in modeling nonlinear, multi-channel signals, and achieves accurate extraction of weak fault features through amplitude-preserving noise reduction and adaptive order selection mechanisms. Furthermore, it combines a multi-branch deep fusion diagnostic strategy to accurately identify the mechanical state of the OLTC (On-Load Tap Changer). In the fault feature extraction stage, this invention introduces a fault feature extraction algorithm based on phase space reconstruction, tensor singular value decomposition, amplitude-preserving non-convex tensor low-rank separation, and tensor kurtosis adaptive order selection. By constructing a multi-channel attractor tensor, separating the main fault feature subspace, suppressing noise interference, and adaptively determining the optimal reconstruction order, it effectively enhances and accurately extracts early weak fault features of the on-load tap changer. In the fault diagnosis stage, a dual-branch multi-scale attention fusion fault diagnosis algorithm is introduced. The extracted time-domain reconstruction features and time-frequency distribution features are used as joint inputs to mine fault impact evolution features and frequency band energy distribution features, respectively. A gated fusion mechanism is used to achieve cross-domain information collaborative representation, ultimately achieving accurate identification of the mechanical state of the on-load tap changer. First, the phase space of the multi-channel original vibration sequence is reconstructed to build a multi-channel attractor tensor capable of characterizing nonlinear dynamic behavior. Based on this, tensor singular value decomposition and amplitude-preserving non-convex tensor low-rank separation are combined to achieve noise reduction and separation of fault features and reconstruction of principal features. Then, the optimal reconstruction order is determined using the impact-sensitive tensor kurtosis index, outputting a clean fault feature signal, which is further input into a bi-branch multi-scale attention fusion diagnostic model to achieve intelligent identification of mechanical fault states. Figure 1 As shown, the on-load tap changer multi-channel feature extraction and fault diagnosis method of the present invention specifically includes the following steps.

[0046] S1. Multi-channel vibration signal acquisition and sample construction.

[0047] During a single switching operation of an on-load tap changer (OLTC), vibration sensors deployed at different locations synchronously acquire multi-channel raw vibration signals to obtain the time-domain sequence corresponding to each vibration measurement channel. The acquired signals are then truncated, aligned, and normalized to form multi-channel vibration samples for subsequent feature extraction and fault diagnosis.

[0048] The on-load tap changer collects C during a single switching process. s The original vibration sequence of each vibration measurement channel, the signal of the c-th channel is represented as: (1). In the formula, This represents the original vibrational sequence of the c-th channel; C s This indicates the total number of vibration measurement channels; N represents the number of sampling points for a single channel signal. Indicates the transpose operation; This represents the amplitude of the c-th channel at the Nth sampling time.

[0049] S2. Phase space reconstruction and multichannel attractor tensor construction.

[0050] The original vibration signal sequences of each vibration measurement channel are reconstructed in phase space to obtain the corresponding attractor matrix. Then, the attractor matrices of each channel are stacked along the channel dimension to construct a multi-channel attractor tensor. Through step S2, the vibration signals from multiple measurement points and the nonlinear dynamic structure of the signals are unified into a single tensor expression, providing a foundation for subsequent tensor decomposition and feature separation.

[0051] To preserve the nonlinear dynamic structure of the vibration signal, phase space reconstruction is performed on the signals of each channel to obtain the attractor matrix of the c-th channel: (2). In the formula, This represents the attractor matrix obtained by reconstructing the c-th channel; Indicates a time delay; Indicates the embedding dimension; Let represent the effective sampling length after reconstruction, and satisfy: (3). Compared with the original number of sampling points N and the embedding dimension and time delay Together they determine the number of rows in the matrix after the phase space is reconstructed.

[0052] Stack the attractor matrices of each channel along the third dimension to construct a multichannel attractor tensor: (4). In the formula, Represents a multichannel attractor tensor; This represents the tensor construction operation that stacks along the channel dimension; This represents a third-order real-valued tensor space. Through step S3, the vibration information from multiple measurement points and the dynamic structure of the phase space can be unified into a single tensor expression.

[0053] S3. Tensor Singular Value Decomposition and Singular Subspace Expansion.

[0054] Tensor singular value decomposition (SVFD) is performed on the constructed multichannel attractor tensor to decompose the original vibration signal into different singular subspaces, yielding the corresponding singular value tensors and characteristic components of each order. The first few orders of singular components mainly carry the fault-related principal features, while the subsequent orders reflect more background noise and interference components.

[0055] Using formula (5) Perform tensor singular value decomposition on the attractor tensor T. Where U represents a left orthogonal tensor; V represents a right orthogonal tensor; Represents a singular value tensor; Represents t-product; This represents the tensor transpose. The components of the singular value tensor reflect the energy distribution of the original signal in different singular subspaces. The first few orders usually correspond to the main fault characteristics, while the subsequent orders contain more noise and interference components.

[0056] S4. Amplitude-preserving non-convex tensor low-rank separation noise reduction.

[0057] To address the issue of noise propagation in singular subspaces of various orders, which can easily lead to the attenuation of effective feature amplitudes, an amplitude-preserving non-convex tensor low-rank separation strategy is introduced. This strategy performs low-rank feature separation and sparse interference suppression on the attractor tensor to obtain a denoised low-rank feature tensor. This step suppresses noise while preserving the fault feature amplitude as much as possible, thereby improving the effectiveness of subsequent feature reconstruction.

[0058] To suppress the spread of noise in singular subspaces of all orders, an amplitude-preserving non-convex tensor low-rank separation strategy is further introduced to perform amplitude-preserving noise reduction on the attractor tensor, and the following optimization model is established: (6). In the formula, This represents the low-rank feature tensor after denoising; Represents the sparse disturbance tensor; Tensor The qth singular value; This represents the total number of singular values; Represents the generalized nonconvex penalty function; Indicates the sparse constraint weights; This represents the L1 norm. Unlike traditional convex kernel norm contraction, the amplitude-preserving nonconvex tensor low-rank separation strategy applies strong suppression to small singular values ​​and weaker contraction to large singular values ​​through nonconvexity penalty. Therefore, it can suppress noise while preserving the fault characteristic amplitude as much as possible, which is the key to this invention.

[0059] S5. Multi-order feature reconstruction and determination of the optimal reconstruction order.

[0060] The denoised low-rank feature tensor is expanded according to the singular value order, and reconstructed sequentially using the first k order components to obtain the reconstructed feature signals corresponding to each order. Furthermore, a tensor singular value kurtosis index and its relative rate of change are constructed. By searching for the candidate position with the largest absolute value of the relative rate of change, the optimal reconstruction order is adaptively determined, thereby outputting a clean fault feature signal.

[0061] The denoised low-rank feature tensor Expanding by singular value order and reconstructing using the first k components, we obtain the k-th order feature tensor: (7). In the formula, This represents the feature tensor reconstructed from the first k singular components; Let represent the tensor component corresponding to the q-th singular value; k represents the reconstruction order. Then, for... By performing inverse tensor transformation and inverse phase space mapping, the corresponding k-th order reconstructed feature signal is obtained. , This represents the fault characteristic signal characterized by the first k-order singular subspace.

[0062] To adaptively determine the optimal reconstruction order, tensor singular value kurtosis (TSVK) is proposed as an order selection metric. Let the kurtosis of the k-th order reconstructed signal in the c-th channel be: (8). In the formula, Let be the kurtosis value of the k-th order reconstructed signal in the c-th channel; This is the time-domain sequence of the k-th order reconstructed signal on the c-th channel; for The mean; for Standard deviation; This indicates the expected operation.

[0063] Since early mechanical failures typically correspond to non-stationary impact components, the more prominent the failure characteristics, the more sensitive the kurtosis index usually is. Further, the tensor singular value kurtosis of the k-th order reconstructed signal is defined as... (9). In the formula, The tensor singular value kurtosis represents the tensor singularity kurtosis of the k-th order reconstructed signal. This index is the average of the kurtosis of each channel and is used to characterize the overall prominence of the multi-channel impact characteristics by the reconstruction result of this order. The relative change rate of TSVK between adjacent orders is further defined as follows: (10). In the formula, This represents the relative rate of change of TSVK between the k-th order and the (k+1)-th order.

[0064] By calculating all orders By determining the position corresponding to the largest absolute value, the optimal reconstruction order can be determined. Let: (11). In the formula, I represents the candidate position corresponding to the optimal reconstruction order; This indicates the index at which the objective function reaches its maximum value. When When, it indicates that the first-order reconstructed signal F I It exhibits more prominent impact characteristics, and in this case, the optimal reconstruction order is I; when When the optimal reconstruction order is found to be I+1, it indicates that the reconstruction result of order I+1 is better. Finally, the characteristic signal determined by the optimal reconstruction order is used as the fault characteristic signal output by this invention.

[0065] S6. Generation of time-domain feature samples and time-frequency feature tensors.

[0066] The optimal-order reconstructed fault feature signal obtained in step S5 is used as the input sample for the diagnostic model, and a short-time Fourier transform is performed on it to construct the corresponding time-frequency feature tensor. This forms two complementary inputs: time-domain reconstructed features and time-frequency distribution features, providing a data foundation for subsequent bi-branch feature learning.

[0067] To further extend the application from fault feature extraction to mechanical condition recognition, after determining the optimal reconstruction order I, the first-order extracted signal F... I As input to the fault diagnosis model, let the multi-channel sample of a single action after truncation, alignment, and normalization be denoted as: (12). In the formula, D represents the multi-channel time-domain feature sample matrix of the input diagnostic model; n3 is the number of vibration measurement channels; m t The length of a single action sample; This indicates a sequence of n3 rows and m... t A matrix space composed of real number elements. To simultaneously preserve the transient evolution information of the fault impact and the energy distribution information of the fault frequency band, a short-time Fourier transform is performed on D to construct the corresponding time-frequency feature tensor: (13). In the formula, G represents the time-frequency feature tensor obtained by transforming the time-domain feature sample D; Indicates the short-time Fourier transform operation; m s For time frames; This represents the third-order real-valued tensor space.

[0068] S7. Two-branch multi-scale attention feature learning.

[0069] A dual-branch, multi-scale attention fusion fault diagnosis model is constructed to deeply mine the reconstructed time-domain and time-frequency features, respectively, to achieve intelligent identification of the mechanical state of on-load tap changers. Specifically, the time-domain branch employs multi-scale one-dimensional convolution and bidirectional gated recurrent units to model the fault impact pulse, envelope fluctuations, and temporal evolution characteristics; the time-frequency branch uses two-dimensional convolution and a channel-space joint attention mechanism to deeply mine the fault frequency band, harmonic structure, and local energy accumulation regions. This step enables parallel learning and enhanced discrimination of time-domain and time-frequency information.

[0070] In the temporal branch, multi-scale one-dimensional convolutions are used to extract features from D in parallel. Let the set of convolution kernel scales be... Then the response at the u-th scale is: (14). In the formula, This represents the temporal feature response extracted at the u-th convolutional scale; u represents the convolutional scale index. This represents the length of the u-th one-dimensional convolution kernel; Indicates the kernel length is One-dimensional convolution operation; For batch normalization; The activation function is nonlinear. The outputs from each scale are concatenated to obtain a multi-scale representation in the time domain: (15). In the formula, A represents the temporal multi-scale feature formed by concatenating the outputs of multi-scale convolution; Indicates feature concatenation operation; This represents the output features corresponding to each convolution scale. This process can simultaneously capture impact pulses, envelope fluctuations, and local decay patterns at different durations, thereby enhancing the model's ability to perceive subtle differences in the early stages of a fault.

[0071] Considering that different measurement points have varying sensitivities to fault states, a channel attention allocation mechanism is further introduced in the time-domain branch to adaptively weight the contributions of multiple channels. The weight vector is defined as follows: (16). In the formula, Represents the channel attention weight vector; This represents the normalization exponential function, used to map the weights of each channel to a probability distribution that sums to 1; Indicates global average pooling; For activation functions; and Let be the trainable parameter matrix. The weighted time-domain features are represented as: (17). In the formula, This represents the time-domain characteristics after channel weighting; This indicates an element-wise weighted operation based on the channel. Subsequently, a bidirectional gated loop unit is used to... By modeling the temporal correlation, a temporal discriminant vector is obtained: (18). In the formula, e d Characterize the temporal evolution of fault impacts during the switching process; This represents a bidirectional gated recurrent unit network, used to characterize temporal dependencies; This represents pooling operations, used for compressing and aggregating temporal features. In the time-frequency branch, the time-frequency feature tensor G is used as input, and two-dimensional convolution is employed to extract fault frequency bands and local energy accumulation regions. (19). In the formula, B represents the initial time-frequency features extracted by two-dimensional convolution; This represents a two-dimensional convolution operation.

[0072] To further highlight key frequency bands and significant time-frequency regions, a channel-space joint attention mechanism is introduced in this branch. First, the channel attention coefficients are constructed: (20). In the formula, This represents the channel attention coefficient in the time-frequency branch; This represents the Sigmoid activation function, used to compress the output to the range of 0 to 1; Represents a multilayer perceptron mapping; This represents the result after performing global average pooling on B; This represents the result after performing global max pooling on B. After channel weighting, we get: (21). In the formula, This represents the time-frequency features after channel attention weighting. Further, the spatial attention mapping is defined as: (22). In the formula, Represents spatial attention mapping; [;] indicates a 7×7 convolution operation; [;] indicates feature concatenation; This represents the average pooling operation; This represents the max pooling operation. The enhanced time-frequency features are: (23). In the formula, This represents the time-frequency features after further enhancement through spatial attention. The time-frequency discriminant vector is then obtained through pooling. (24). In the formula, This represents the discriminative feature vector output by the time-frequency branch. This branch is mainly used to enhance the expressive power of fault frequency, harmonic structure, and their time-varying distribution patterns, thereby improving the separability between similar mechanical faults at the frequency domain level.

[0073] S8. Cross-domain feature fusion and mechanical state recognition.

[0074] Gated fusion of discriminative features extracted from the time-domain and time-frequency branches is performed to construct a fused deep representation, which is then input into the classification layer. The output is the posterior probability of the mechanical state corresponding to the on-load tap changer and the final diagnostic category. To further improve the inter-class separation ability and intra-class compactness under small sample conditions, cross-entropy loss and center constraint loss are introduced for joint optimization during model training, thereby achieving accurate diagnosis of the mechanical state of the on-load tap changer.

[0075] To achieve coordinated discrimination of time-domain impact information and time-frequency energy information, a gating fusion mechanism is further constructed to adaptively couple the dual-branch features. The fusion gate vector is defined as: (25). In the formula, This represents the gate weight vector during the fusion of two-branch features; Represents the trainable parameter moments of the fusion gate layer; Represents the bias vector of the fusion gate layer; This indicates that the time-domain discriminant vector With time-frequency discriminant vector The layers are then spliced ​​together. The resulting deep representation is: (26). In the formula, h represents the fused deep feature vector; This represents the trainable parameter matrix of the fusion mapping layer; Represents the bias vector of the fusion mapping layer; This represents the weighting coefficient that complements the gating weights; This represents the element-wise interaction term between the time-domain discriminant vector and the time-frequency discriminant vector. This fusion method not only preserves the complementary information of the time-domain and time-frequency-domain features but also explicitly introduces cross-domain interaction terms, thereby enhancing the model's ability to represent complex coupled mechanical faults. This is one of the important improvements that distinguishes it from traditional single-branch fault classification methods. In the output layer, the fused feature h is input into the classifier to obtain the posterior probability distribution of each mechanical state: (27). In the formula, This represents the posterior probability vector corresponding to each mechanical state category; This represents the trainable parameter matrix of the classification layer; This represents the bias vector of the classification layer; This represents the classification normalization function. The corresponding fault identification result is: (28). In the formula, This represents the probability that the sample belongs to the j-th type of mechanical state. Indicates taking The category index that reaches the maximum value; This represents the final diagnostic category. To improve intra-class compactness and inter-class separability under small sample conditions, cross-entropy loss and center constraint loss are further introduced to form a joint optimization objective function: (29). In the formula, J represents the joint optimization objective function; Represents the cross-entropy loss term; Indicates the central constraint loss term; This represents the weighting coefficient used to balance the contributions of the two types of losses; (30). (31). In the formula, v represents the batch sample size; v represents the batch sample index. Here, j represents the number of mechanical state categories; j represents the category index. Let be the label value of the v-th sample for the j-th class; c represents the center vector of the class to which the v-th sample belongs in the fused feature space; v This represents the true class label of the v-th sample; This represents the squared L2 norm. This joint constraint effectively compresses the dispersion of similar samples in the feature space and increases the discrimination interval between dissimilar samples, thereby improving the accuracy of identifying early weak faults and similar fault states, and achieving accurate diagnosis of the mechanical state of on-load tap changers.

[0076] To verify the effectiveness of the proposed method in extracting weak fault features and identifying faults in the mechanical vibration signals of on-load tap changers, online monitoring data of mechanical vibration signals from a KM-type oil-immersed on-load tap changer were selected as an example for verification. Experimental analysis was conducted on the multi-channel vibration response during the switch switching process under different mechanical conditions.

[0077] In terms of test system setup, a YD-38D piezoelectric accelerometer was selected as the vibration sensor and installed on the top flange of the on-load tap changer. Three vibration measurement channels were arranged to synchronously acquire multi-channel vibration signals during the switch switching process. The main parameters of the YD-38D vibration sensor are shown in Table 1. The signal acquisition unit used was the NI9231 vibration signal acquisition module, which has 8-channel analog input capability and a maximum single-channel sampling rate of 51.2 kS / s, meeting the high-precision synchronous acquisition requirements of transient vibration signals during on-load tap changer switching. The KM-type oil-immersed on-load tap changer, the YD-38D vibration sensor, the NI9231 acquisition module, and the host computer online monitoring software together constitute the online monitoring hardware system used in the verification of this invention, providing an experimental basis for subsequent multi-channel fault feature extraction and fault diagnosis.

[0078] Table 1 Sensor Parameters

[0079] YD-38D 0.0005 1000 1-12000 40k 0.04 14 14*14*28

[0080] To verify the effectiveness of the overall strategy of this invention, a multi-mechanical-state vibration sample dataset was constructed based on the KM-type oil-immersed on-load tap changer online monitoring platform. Specifically, taking the normal operating condition of the on-load tap changer as the benchmark, and combining it with the three typical mechanical defect conditions mentioned above—namely, loose contact and arc plate conditions, loose drive shaft screw conditions, and jammed drive gear conditions—multi-channel vibration signals during the switch switching process were repeatedly collected. The signal sampling frequency was set to 51.2Hz, and the sampling mode was trigger sampling. When the switch contact actuates, the signal reaches the trigger threshold, and the system automatically saves 10 seconds of data before and after the trigger. In the formula, the vibration signal generated by the contact closing at the moment of switch switching is extracted, with a duration of approximately 150ms. The effective vibration segment extracted during each switching process is taken as an independent sample. Ultimately, a total of 800 sets of data were constructed, including 200 sets of normal state samples, 200 sets of loose contact and arc plate samples, 200 sets of loose drive shaft screw samples, and 200 sets of stuck drive gear samples. The final diagnostic results are shown in Table 2.

[0081] Regarding dataset partitioning, to ensure the objectivity and repeatability of training and testing results, a stratified random sampling method was used to divide all samples into a training set and a test set in a 4:1 ratio, with 640 groups in the training set and 160 groups in the test set. The training set is used to complete the parameter learning and model training of the overall strategy proposed in this invention. Specifically, fault features are first extracted through phase space reconstruction, tensor singular value decomposition, amplitude-preserving non-convex tensor low-rank separation, and tensor kurtosis adaptive order selection. Then, the optimal reconstructed features are input into a two-branch multi-scale attention fusion fault diagnosis model for classification learning. The test set is used to evaluate the recognition performance of the overall strategy for different mechanical states of on-load tap changers.

[0082] To quantitatively evaluate the recognition performance of the method of this invention for different mechanical states of on-load tap changers, classification accuracy, precision, recall, and F1 score are selected as evaluation indicators. Let TP be the number of samples correctly identified as belonging to a certain mechanical state category, FP be the number of samples misclassified as belonging to that category, FN be the number of samples misclassified as belonging to other categories, and TN be the number of samples correctly identified as belonging to other categories. Then, classification accuracy is defined as: (32). In the formula, Acc represents the classification accuracy, which reflects the proportion of correctly identified samples in the entire sample. Precision is defined as: Precision = (33). In the formula, precision measures the proportion of samples judged as belonging to a certain mechanical state that actually do not. Recall is defined as: Recall = (34). In the formula, recall measures the proportion of a sample of a given mechanical state that is correctly identified. The F1 score is defined as: (35). In the formula, F1 represents the harmonic mean of precision and recall, which is used to comprehensively evaluate the overall recognition performance of the classification model.

[0083] To further verify the effectiveness of the overall strategy proposed in this invention, four sets of comparative algorithms were set up for experimental verification under the same dataset, the same training and test set partitioning method, and the same evaluation index, and the method of this invention was compared with them in a unified manner. The four sets of comparative algorithms are denoted as A1, A2, A3, and A4, respectively. In the formula, A1 is a conventional fault diagnosis method, which directly extracts the time-domain statistical features and frequency-domain features of multi-channel vibration signals and combines them with support vector machines to complete mechanical state classification, thereby characterizing the performance level of traditional shallow feature diagnosis methods. A2 is a direct diagnosis method without feature extraction enhancement, which directly inputs the preprocessed multi-channel original vibration signal into a dual-branch multi-scale attention fusion diagnosis model, thereby illustrating the recognition effect of the model without front-end fault feature extraction. A3 is a comparative method without amplitude-preserving non-convex tensor low-rank separation, which uses phase space reconstruction and tensor singular value decomposition to complete feature modeling, but does not perform amplitude-preserving noise reduction processing, and directly combines adaptive order selection with subsequent diagnosis models for state recognition, thereby verifying the role of amplitude-preserving noise reduction in enhancing weak fault features. A4 is a comparison method without adaptive order selection. It uses a fixed reconstruction order for feature reconstruction based on phase space reconstruction, tensor singular value decomposition, and amplitude-preserving non-convex tensor low-rank separation. The feature is then input into a diagnostic model for classification, which is used to illustrate the necessity of the adaptive order selection mechanism in the feature reconstruction process. Figures 2-6 The confusion matrices of four comparison algorithms (A1, A2, A3, and A4) and the method of this invention on the test set are presented respectively. The proposed method fully includes phase space reconstruction, tensor singular value decomposition, amplitude-preserving non-convex tensor low-rank separation, tensor kurtosis adaptive order selection, and a two-branch multi-scale attention fusion fault diagnosis model. All methods were trained and tested on a unified dataset, and the final diagnostic results are shown in Table 2.

[0084] Table 2 Comparison of specific indicators for each method

[0085] A1 Time-frequency statistical features + SVM 71.86 70.94 69.87 70.4 A2 Original signal + dual-branch attention diagnostic model 81.73 80.96 80.18 80.57 A3 Phase space reconstruction + TSVD + TSVK + diagnostic model 88.64 87.95 87.21 87.58 A4 Phase space reconstruction + TSVD + amplitude-preserving low-rank separation + fixed order selection + diagnostic model 92.47 91.88 91.36 91.62 Method of the present invention Feature extraction + intelligent diagnosis overall strategy 96.83 96.37 95.98 96.17

[0086] Combined with Table 2 and Figures 2-6 The experimental results show that, under the same dataset, unified training and testing partitioning, and identical experimental conditions, the method of this invention outperforms the comparison algorithms in all four evaluation metrics: accuracy, precision, recall, and F1 score. Furthermore, its confusion matrix exhibits a more concentrated distribution along the main diagonal and fewer off-diagonal errors, indicating that the method of this invention has stronger discriminative ability and higher classification stability for different mechanical states. Where, Figure 2 The corresponding A1 method has a relatively weak overall recognition effect. There are obvious cross-class misclassifications in the confusion matrix, indicating that the traditional diagnostic method based on time-frequency statistical features and shallow classifiers is difficult to fully characterize the nonlinear and non-stationary characteristics of multi-channel vibration signals under the complex mechanical state of on-load tap changers. Figure 3 The corresponding A2 method improves classification performance after introducing a dual-branch deep diagnostic model. However, due to the lack of front-end fault feature extraction and enhancement, noise and interference components in the original signal still weaken the model's ability to identify weak fault information, thus still resulting in a certain degree of class confusion. Figure 4 The corresponding A3 method, after further introducing phase space reconstruction and tensor singular value decomposition, shows a significant increase in the number of diagonal elements in the confusion matrix, indicating that the feature representation method based on phase space tensor modeling can effectively improve the separability of fault samples. However, due to the lack of amplitude preservation and noise reduction processing, some weak fault features are still easily affected by noise during the separation process. Figure 5 The corresponding A4 method, after incorporating amplitude-preserving low-rank separation, further improves diagnostic performance, significantly reducing misclassifications in the confusion matrix. This indicates that amplitude-preserving denoising has a positive effect on enhancing fault features. However, due to the use of a fixed reconstruction order, it is difficult to achieve optimal reconstruction for the feature distribution of different samples, thus the overall effect is still not optimal. In contrast, Figure 6 The method of the present invention shown exhibits more ideal recognition results under various mechanical conditions. In the confusion matrix, samples of various types are mainly concentrated on the main diagonal position, and the number of misclassified samples is the fewest. This fully demonstrates that the present invention achieves effective separation of fault main features and noise interference through amplitude-preserving non-convex tensor low-rank separation, and accurately determines the optimal reconstruction order by means of tensor kurtosis adaptive order selection, providing more discriminative input features for the subsequent bi-branch multi-scale attention fusion diagnostic model, thereby achieving the best comprehensive recognition performance.

[0087] The present invention discloses a multi-channel feature extraction and fault diagnosis system for on-load tap changers, used to execute a multi-channel feature extraction and fault diagnosis method for on-load tap changers. It includes a multi-channel vibration sensing module, a signal conditioning module, a synchronization triggering module, a multi-channel data acquisition module, a data processing and system control module, a data storage module, and a communication module. The multi-channel vibration sensing module includes multiple vibration sensors, which are installed on the operating mechanism housing, transmission mechanism connection, drive motor mounting base, transmission bearing housing, and switching switch support structure of the on-load tap changer, used to acquire vibration signals from multiple measurement points during a single switching process of the on-load tap changer. The input terminal of the signal conditioning module is connected to the output terminal of the multi-channel vibration sensing module, used to perform charge conversion, amplification, filtering, anti-aliasing processing, and impedance matching on the vibration signals of each channel. The synchronization triggering module is connected to the action control circuit, auxiliary contacts, drive motor current detection unit, or position switch of the on-load tap changer, used to acquire the vibration signals from the single switching process of the on-load tap changer. The system initiates the switching action and outputs a synchronous trigger signal to the multi-channel data acquisition module. The multi-channel data acquisition module, connected to the signal conditioning module and the synchronous trigger module, synchronously acquires and converts the vibration signals of each channel to digital under the action of the synchronous trigger signal, forming multi-channel raw vibration data. The data processing and system control module, connected to the multi-channel data acquisition module, receives and processes the multi-channel raw vibration data, performing signal interception, alignment, normalization, phase space reconstruction, tensor feature extraction, amplitude-preserving noise reduction, multi-order feature reconstruction, time-frequency feature generation, cross-domain feature fusion, and mechanical condition identification, outputting the mechanical condition diagnosis results of the on-load tap changer. The data storage module, connected to the data processing and system control module, stores the raw vibration data, feature data, diagnostic model parameters, historical diagnostic results, and condition assessment records. The communication module, connected to the data processing and system control module, uploads the mechanical condition diagnosis results to the condition monitoring platform, substation backend system, or maintenance terminal.

[0088] In summary, the on-load tap changer multi-channel fault feature extraction and fault diagnosis method and system proposed in this invention can effectively enhance weak fault features and improve the separability between samples of different mechanical states under complex background noise conditions, and significantly improve the accuracy and stability of mechanical state identification. This verifies the effectiveness, superiority and engineering application value of the method in on-load tap changer mechanical state monitoring and fault diagnosis.

Claims

1. A method for multi-channel feature extraction and fault diagnosis of on-load tap changers, characterized in that, Includes the following steps: S1. Multi-channel vibration signal acquisition and sample construction The original vibration signals from multiple channels during a single switching process of the on-load tap changer were acquired to obtain the time-domain sequence corresponding to each vibration measurement channel; the acquired signals were then truncated, aligned, and normalized. S2. Phase Space Reconstruction and Multichannel Attractor Tensor Construction The original vibration signal is reconstructed in phase space to obtain the corresponding attractor matrix; then the attractor matrices are stacked along the channel dimension to construct a multi-channel attractor tensor. S3. Tensor Singular Value Decomposition and Singular Subspace Expansion Tensor singular value decomposition is performed on the multichannel attractor tensor to decompose the original vibration signal into different singular subspaces, and the corresponding tensor singular value tensor and characteristic components of each order are obtained. S4. Amplitude-preserving non-convex tensor low-rank separation noise reduction A low-rank separation strategy for amplitude-preserving non-convex tensors is introduced to perform low-rank feature separation and sparse interference suppression on attractor tensors, thereby obtaining denoised low-rank feature tensors. S5. Multi-order feature reconstruction and determination of the optimal reconstruction order The denoised low-rank feature tensor is expanded according to the singular value order, and the first k order components are used to reconstruct the reconstructed feature signals corresponding to each order. Furthermore, the singular value kurtosis index of the tensor and its relative rate of change are constructed. By searching for the candidate position with the largest absolute value of the relative rate of change, the optimal reconstruction order is adaptively determined, thereby outputting a pure fault feature signal. S6. Generation of Temporal Feature Samples and Time-Frequency Feature Tensors The fault feature signal is used as the input sample of the diagnostic model, and a short-time Fourier transform is performed on it to construct the corresponding time-frequency feature tensor. S7. Two-branch multi-scale attention feature learning A dual-branch, multi-scale attention fusion fault diagnosis model is constructed, and the reconstructed time-domain features and time-frequency features are deeply mined to achieve intelligent identification of the mechanical state of on-load tap changers. S8. Cross-domain feature fusion and machine condition recognition Gated fusion is performed on the discriminative features extracted from the time-domain branch and the time-frequency branch to construct a fused deep representation, which is then input into the classification layer. The output is the posterior probability of the mechanical state corresponding to the on-load tap changer and the final diagnostic category.

2. The method for multi-channel feature extraction and fault diagnosis of on-load tap changers according to claim 1, characterized in that, In step S1, the original vibration signal sequence of the c-th channel is acquired during one switching process of the on-load tap changer. In the formula, C s This indicates the total number of vibration measurement channels; N represents the number of sampling points for a single channel signal. Indicates the transpose operation; This represents the amplitude of the c-th channel at the Nth sampling time.

3. The method for multi-channel feature extraction and fault diagnosis of on-load tap changers according to claim 2, characterized in that, In step S2, phase space reconstruction is performed on the signals of each channel to obtain the attractor matrix of the c-th channel. In the formula, Indicates a time delay; Indicates the embedding dimension; Let represent the effective sampling length after reconstruction, and satisfy: .

4. The method for multi-channel feature extraction and fault diagnosis of on-load tap changers according to claim 3, characterized in that, In step S2, the multichannel attractor tensor In the formula, This represents the tensor construction operation that stacks along the channel dimension; This represents the third-order real-valued tensor space.

5. The method for multi-channel feature extraction and fault diagnosis of on-load tap changers according to claim 4, characterized in that, In step S3, the formula is used. Perform tensor singular value decomposition on T; where U represents a left orthogonal tensor and V represents a right orthogonal tensor; Represents a singular value tensor; Represents t-product; This indicates the transpose of a tensor.

6. The method for multi-channel feature extraction and fault diagnosis of on-load tap changers according to claim 5, characterized in that, In step S4, amplitude-preserving denoising is performed on the attractor tensor to establish an optimization model. In the formula, This represents the low-rank feature tensor after denoising; Represents the sparse disturbance tensor; Tensor The qth singular value; Indicates the total number of singular values; Represents the generalized nonconvex penalty function; Indicates the sparse constraint weights; This represents the L1 norm.

7. The method for multi-channel feature extraction and fault diagnosis of on-load tap changers according to claim 6, characterized in that, In step S5, Expanding by singular value order and reconstructing using the first k components, we obtain the k-th order feature tensor: In the formula, This represents the feature tensor reconstructed from the first k singular components; The tensor component corresponding to the q-th singular value is represented by k; k represents the reconstruction order; then... By performing inverse tensor transformation and inverse phase space mapping, the corresponding k-th order reconstructed feature signal is obtained. .

8. The method for multi-channel feature extraction and fault diagnosis of on-load tap changers according to claim 7, characterized in that, In step S5, the optimal reconstruction order is adaptively determined using the tensor singular value kurtosis (TSVK) as the order selection index. Let the kurtosis of the k-th order reconstructed signal on the c-th channel be... In the formula, This is the time-domain sequence of the k-th order reconstructed signal on the c-th channel; for The mean; for Standard deviation; This represents the expectation operation; further, the tensor singular value kurtosis of the k-th order reconstructed signal is defined. ; Redefine the relative rate of change of adjacent orders TSVK By calculating all orders And determine the position corresponding to the largest absolute value, and determine the optimal reconstruction order; set up: In the formula, I represents the candidate position corresponding to the optimal reconstruction order; This represents the index that maximizes the objective function; when When, it indicates that the first-order reconstructed signal F I It exhibits more prominent impact characteristics, and in this case, the optimal reconstruction order is I; when When the optimal reconstruction order is I+1, it indicates that the reconstruction result of order I+1 is better. Finally, the characteristic signal determined by the optimal reconstruction order is taken as the fault characteristic signal.

9. The method for multi-channel feature extraction and fault diagnosis of on-load tap changers according to claim 8, characterized in that, In step S6, after determining the optimal reconstruction order I, the first-order extracted signal F is... I As input to the fault diagnosis model; let Single-action multi-channel temporal feature sample matrix after truncation, alignment and normalization In the formula, n3 is the number of vibration measurement channels; m t The length of a single action sample; This indicates a sequence of n3 rows and m... t D is a matrix space composed of real elements; a short-time Fourier transform is performed on D to construct the corresponding time-frequency feature tensor. In the formula, Indicates the short-time Fourier transform operation; m s For time frames; This represents the third-order real-valued tensor space.

10. The method for multi-channel feature extraction and fault diagnosis of on-load tap changers according to claim 9, characterized in that, In step S7, the time-domain branch uses multi-scale one-dimensional convolution and bidirectional gated recurrent units to model the fault impact pulse, envelope fluctuations and temporal evolution characteristics. The time-frequency branch employs two-dimensional convolution and channel-space joint attention mechanism to deeply mine fault frequency bands, harmonic structures, and local energy accumulation regions.

11. The method for multi-channel feature extraction and fault diagnosis of on-load tap changers according to claim 10, characterized in that, In step S7, in the temporal branch, multi-scale one-dimensional convolution is used to extract features from D in parallel; let the set of convolution kernel scales be... The temporal feature response extracted at the u-th convolution scale In the formula, u represents the convolution scale index; This represents the length of the u-th one-dimensional convolution kernel; Indicates the kernel length is One-dimensional convolution operation; For batch normalization; The activation function is nonlinear; the outputs at each scale are concatenated to obtain a multi-scale representation in the time domain. In the formula, Indicates feature concatenation operation; This represents the output features corresponding to each convolution scale.

12. The method for multi-channel feature extraction and fault diagnosis of on-load tap changers according to claim 11, characterized in that, In step S7, a channel attention allocation mechanism is further introduced in the time-domain branch to adaptively weight the contributions of multiple channels, resulting in a channel attention weight vector. In the formula, This represents the normalized exponential function; Indicates global average pooling; For activation functions; and The trainable parameter matrix; the weighted time-domain features are represented as: In the formula, This indicates an element-wise weighted operation based on the channel. Subsequently, a bidirectional gated loop unit was used to... The temporal correlation is modeled to obtain the temporal discriminant vector. In the formula, This represents a bidirectional gated cyclic cell network; Indicates pooling operation; In the time-frequency branch, the time-frequency feature tensor G is used as input, and two-dimensional convolution is used to extract the fault frequency band and local energy accumulation region: In the formula, B represents the initial time-frequency features extracted by two-dimensional convolution; This represents a two-dimensional convolution operation.

13. The method for multi-channel feature extraction and fault diagnosis of on-load tap changers according to claim 12, characterized in that, In step S7, a channel-space joint attention mechanism is introduced into the time-frequency branch. First, the channel attention coefficients are constructed. In the formula, This represents the Sigmoid activation function; Represents a multilayer perceptron mapping; This represents the result after performing global average pooling on B; This represents the result after global max pooling of B; after channel weighting, we get: In the formula, This represents the time-frequency features after channel attention weighting; further, the spatial attention mapping is defined. In the formula, [;] indicates a 7×7 convolution operation; [;] indicates feature concatenation; This represents the average pooling operation; Represents max pooling operation; enhanced time-frequency features ; and obtain the time-frequency discrimination vector through pooling. .

14. The method for multi-channel feature extraction and fault diagnosis of on-load tap changers according to claim 13, characterized in that, In step S8, a gating fusion mechanism is constructed to adaptively couple the dual-branch features, and the gating weight vector for dual-branch feature fusion is defined. In the formula, Represents the trainable parameter moments of the fusion gate layer; Represents the bias vector of the fusion gate layer; This indicates that the time-domain discriminant vector With time-frequency discriminant vector The data is spliced ​​together; the resulting deep representation is: In the formula, h represents the fused deep feature vector; This represents the trainable parameter matrix of the fusion mapping layer; Represents the bias vector of the fusion mapping layer; This represents the weighting coefficient that complements the gating weights; This represents the element-wise interaction term between the time-domain discriminant vector and the time-frequency discriminant vector; in the output layer, h is input into the classifier to obtain the posterior probability vector corresponding to each mechanical state category. In the formula, This represents the trainable parameter matrix of the classification layer; This represents the bias vector of the classification layer; the corresponding fault identification result is: In the formula, This represents the probability that the sample belongs to the j-th type of mechanical state; Indicates taking The category index that reaches the maximum value; This is the final diagnostic category.

15. The method for multi-channel feature extraction and fault diagnosis of on-load tap changers according to claim 14, characterized in that, In step S8, cross-entropy loss and center constraint loss are further introduced to form a joint optimization objective function. In the formula, Represents the cross-entropy loss term; Indicates the central constraint loss term; This represents the weighting coefficient used to balance the contributions of the two types of losses; ; ; v represents the batch sample size; v represents the batch sample index. Here, j represents the number of mechanical state categories; j represents the category index. Let be the label value of the v-th sample for the j-th class; c represents the center vector of the class to which the v-th sample belongs in the fused feature space; v This represents the true class label of the v-th sample; This represents the square of the L2 norm.

16. A multi-channel feature extraction and fault diagnosis system for on-load tap changers, used to execute the multi-channel feature extraction and fault diagnosis method for on-load tap changers as described in claim 15, characterized in that, The system includes a multi-channel vibration sensing module, a signal conditioning module, a synchronization triggering module, a multi-channel data acquisition module, a data processing and system control module, a data storage module, and a communication module. The multi-channel vibration sensing module comprises multiple vibration sensors mounted on the operating mechanism housing, transmission mechanism connection points, drive motor mounting base, transmission bearing housing, and switching switch support structure of the on-load tap changer. These sensors are used to acquire vibration signals from multiple measurement points during a single switching operation of the on-load tap changer. The input terminal of the signal conditioning module is connected to the output terminal of the multi-channel vibration sensing module, and it performs charge conversion, amplification, filtering, anti-aliasing processing, and impedance matching on the vibration signals from each channel. The synchronization triggering module is connected to the on-load tap changer's action control circuit, auxiliary contacts, drive motor current detection unit, or position switch, and it acquires the start time of a single switching operation of the on-load tap changer and outputs a synchronization trigger signal to the multi-channel data acquisition module. The system transmits signals; the multi-channel data acquisition module is connected to the signal conditioning module and the synchronization triggering module, and is used to synchronously acquire and convert the vibration signals of each channel under the action of the synchronization triggering signal to form multi-channel raw vibration data; the data processing and system control module is connected to the multi-channel data acquisition module, and is used to receive and process the multi-channel raw vibration data, perform signal interception, alignment, normalization, phase space reconstruction, tensor feature extraction, amplitude preservation and noise reduction, multi-order feature reconstruction, time-frequency feature generation, cross-domain feature fusion and mechanical condition identification, and output the mechanical condition diagnosis results of the on-load tap changer; the data storage module is connected to the data processing and system control module, and is used to store raw vibration data, feature data, diagnostic model parameters, historical diagnostic results and condition assessment records; the communication module is connected to the data processing and system control module, and is used to upload the mechanical condition diagnosis results to the condition monitoring platform, substation backend system or operation and maintenance terminal.

Images