Manual switch operating condition monitoring method based on noise reduction encoder

Abstract

The application provides a manual switch running state monitoring method based on a noise reduction encoder, comprising the steps of: collecting vibration signals of a manual switch for feature extraction sample data; outputting fused features through a multi-feature fusion network; inputting the fused features into an improved multi-time scale model to output multi-time scale fault features, performing noise reduction processing through a noise reduction encoder, and obtaining feature information; and then outputting a running state monitoring result of the manual switch based on the feature information. The method is efficient and simple, and effectively solves the deficiencies of traditional monitoring techniques in noise suppression, feature extraction and real-time performance through multi-dimensional feature extraction and improved algorithm processing, thereby providing technical support for intelligent operation and maintenance of switch cabinets, significantly reducing artificial inspection costs, and improving the reliability of power system operation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric data processing, and in particular to a manual switch operation state monitoring method based on a noise reduction encoder. BACKGROUND

[0002] As a core device of power system transmission and distribution, the switch cabinet bears the key functions of power distribution, circuit protection and device control, and the manual switch is the key equipment to ensure the safe and stable operation of the power grid. During the long-term operation of the manual switch, it needs to frequently perform opening and closing operations, and its internal mechanical structure, such as the contact, operating mechanism, transmission component, etc., is inevitably affected by multiple factors such as mechanical wear, arc ablation, environmental temperature and humidity fluctuations, and dust erosion. These factors will gradually lead to the occurrence of potential faults such as increased contact resistance, mechanism jamming, spring fatigue, and aging of insulation parts. If not discovered in time, it may cause local overheating, delayed operation response, or even regional power outage accidents, which poses a serious threat to industrial production, residents' life and power grid reliability. Therefore, real-time and accurate monitoring of the operation state of the manual switch is one of the core needs of intelligent operation and maintenance of the power system.

[0003] Traditional manual switch state monitoring mainly relies on manual inspection and periodic preventive testing. Manual inspection obtains state information through on-site observation and manual measurement by operation and maintenance personnel, but this method is limited by long inspection cycle, strong subjectivity, and high labor intensity, making it difficult to capture transient fault characteristics and prone to missed detection and misjudgment in complex distribution room environments. Although periodic preventive testing can evaluate switch performance through offline detection, the testing process needs to be performed with power off, which will affect power continuity and cannot reflect the dynamic state of the switch under actual operating conditions, making it difficult to meet the real-time requirements of modern power grid for equipment state perception.

[0004] With the development of sensing technology and intelligent algorithms, automated monitoring technology has been gradually applied to manual switch state evaluation. Among them, vibration signal monitoring has become a major direction because it can directly reflect the dynamic characteristics of mechanical structures. During the opening and closing operations and running of the manual switch, the movement and collision of its internal mechanical components will produce specific vibration signals, which contain a large amount of state information, such as abnormal vibration amplitude caused by poor contact, time domain characteristics of vibration waveform changed by mechanism jamming, and abnormal spectral peak values at specific frequency bands caused by component loosening. Therefore, by collecting vibration signals and extracting features, indirect evaluation of the switch state can be achieved.

[0005] However, existing vibration signal-based monitoring technologies still have many limitations. At the feature extraction level, traditional methods often employ single-domain feature analysis, such as peak value and kurtosis in the time domain, and power spectral density in the frequency domain. However, the vibration signals of manual switches are often non-stationary and non-linear, making it difficult for single-domain features to fully represent their complex states. Simultaneously, electromagnetic interference and noise from other equipment in the field environment can severely contaminate the signal, submerging effective features and affecting the accuracy of feature extraction. At the model processing level, existing methods lack the ability to capture features across multiple time scales. For example, the mechanical action of a manual switch includes processes at different time scales, such as instantaneous impact and continuous motion. Therefore, a single-time-scale model cannot adequately account for various features, resulting in limited ability to identify complex states. Furthermore, in scenarios involving parallel monitoring of multiple devices, existing algorithms have low computational efficiency, making real-time analysis and response difficult. Moreover, their model generalization ability is weak, exhibiting poor adaptability to different switch models and operating environments, failing to meet practical engineering needs.

[0006] In summary, existing manual switch operation status monitoring technologies have significant shortcomings in terms of the comprehensiveness of feature extraction, noise suppression capability, multi-timescale feature capture and real-time performance, and generalization ability, making it difficult to achieve accurate perception and reliable assessment of switch operation status. Summary of the Invention

[0007] This application addresses the problems existing in the prior art by providing a monitoring method that can integrate multi-domain features, adapt to dynamic changes at multiple time scales, effectively suppress noise, and has efficient real-time processing capabilities.

[0008] To achieve the above objectives, the technical solution adopted in this application is as follows:

[0009] This application provides a method for monitoring the operating status of a manual switch based on a noise-reducing encoder, which mainly includes the following steps:

[0010] Collect vibration signals from manual switches;

[0011] The vibration signal is preprocessed to generate time-domain sample data and frequency-domain sample data for feature extraction;

[0012] Based on the time-domain sample data and frequency-domain sample data, the time-domain features and frequency-domain features are encoded and fused in parallel through a multi-feature fusion network, and the fused features are output.

[0013] The fused features are input into the improved multi-timescale model to output multi-timescale fault features.

[0014] input the multi-time scale fault features into a denoising encoder; in the denoising encoder, the multi-time scale fault features are respectively extracted into global features and local features through a pooling operation; global threshold coefficients and local threshold coefficients are respectively generated based on the global features and the local features; the global threshold coefficients and the local threshold coefficients are modulated with the multi-time scale fault features, and denoising thresholds are obtained; the global features and the local features are denoised by using a soft threshold function based on the denoising thresholds; the denoised global features and the denoised local features are subjected to feature splicing and feature regularization operations, and feature information is obtained;

[0015] The operation state monitoring result of the manual switch is output based on the feature information.

[0016] Optionally, when collecting the vibration signal of the manual switch, an electric signal of mechanical vibration of the manual switch is collected by a vibration sensor;

[0017] The collection process includes vibration signals of the manual switch under multiple operation states, including opening operation, closing operation, normal operation state and simulated fault state.

[0018] The vibration signal contains complete mechanical action timing information.

[0019] Optionally, the vibration signal is preprocessed, including the following steps:

[0020] The continuously collected vibration signal is segmented according to a preset fixed time length, and a plurality of signal samples are obtained;

[0021] Time domain feature extraction is performed on each signal sample to generate the time domain sample data;

[0022] Time-frequency conversion processing is performed on each signal sample to generate the frequency domain sample data.

[0023] Optionally, the multi-feature fusion network includes a first feature processing module, a second feature processing module and a fusion module;

[0024] The first feature processing module is provided with a waveform encoder and a state regressor; the waveform encoder includes a plurality of cascaded residual convolution blocks, each residual convolution block is provided with a convolution layer, a normalization layer and a ReLU activation layer; the state regressor has the same network structure as the waveform encoder, and a full connection layer is arranged at the end of the state regressor;

[0025] The second feature processing module is provided with a spectrum graph encoder and a spectrum graph decoder; the spectrum graph encoder includes a Swin-Transformer block and a patch synthesis layer; the spectrum graph decoder includes an inverse patch synthesis layer and a Swin-Transformer block;

[0026] The fusion module comprises a multi-layer perception, a convolution layer and a Transformer decoder; the multi-layer perception is used to process the output features of the first feature processing module; and the convolution layer is used to process the output features of the second feature processing module.

[0027] Optionally, in the training process of the first feature processing module, the working voltage of the manual switch is taken as a target parameter, the working voltage prediction value is output through the state regressor, and the network parameters are adjusted by using a mean square error loss function;

[0028] In the training process of the second feature processing module, the reconstructed input spectrum is taken as a target, the reconstructed spectrum is output through the spectrum graph decoder, and the network parameters are adjusted by using a hybrid loss function comprising a mean square error loss and a structural similarity loss;

[0029] The fusion module performs multi-level decoding on the waveform features output by the first feature processing module and the spectrum features output by the second feature processing module through the Transformer decoder to generate the fusion features.

[0030] Optionally, the improved multi-time scale model is constructed based on the SE-DSCNN model, and a data input layer is additionally arranged at the input end;

[0031] The improved multi-time scale model has a four-branch parallel convolution structure, the first branch processes the input features through a one-dimensional convolution layer; the second branch splices the output features of the first branch with the input features and processes them through a one-dimensional convolution layer; the third branch splices the output features of the second branch with the input features and processes them through a one-dimensional convolution layer; and the fourth branch comprises a maximum pooling layer and a convolution layer.

[0032] After the output features of each branch are spliced, the feature weights are calculated through a self-attention mechanism, and the multi-time scale fault features are generated.

[0033] Optionally, in the pooling operation, global average pooling and maximum pooling are included.

[0034] The global average pooling extracts global features of the distribution by calculating the average value of the multi-time scale fault features in the spatial dimension.

[0035] The maximum pooling extracts local features in the form of pulses by selecting the maximum value of the multi-time scale fault features in the spatial dimension.

[0036] Optionally, when the global threshold coefficient is generated, the global features are input into a first type of attention module.

[0037] The first type of attention module comprises a batch normalization layer, a ReLU activation layer, a full connection layer and a Sigmoid activation layer; after the global feature is processed by the batch normalization layer and the ReLU activation layer, the dimension is adjusted through the full connection layer, and then the global threshold coefficient matching the dimension of the global feature is output through the Sigmoid activation layer;

[0038] When the local threshold coefficient is generated, the local feature is input into a second type of attention module, the network structure of the second type of attention module is the same as that of the first type of attention module, the local feature is input into the second type of attention module, and the local threshold coefficient matching the dimension of the local feature is output.

[0039] Optionally, the soft threshold function is an improved adaptive soft threshold function, and a control term is additionally added.

[0040] When the global feature is denoised, the adaptive soft threshold function corresponding to the global threshold coefficient is used.

[0041] When the local feature is denoised, the adaptive soft threshold function corresponding to the local threshold coefficient is used.

[0042] Optionally, in the feature splicing and feature regularization operation, the following steps are included:

[0043] The feature splicing is performed by connecting the denoised global feature and the local feature in the channel dimension.

[0044] The feature regularization comprises a channel attenuation convolution and an identity mapping, the channel attenuation convolution is used to adjust the channel number of the spliced feature through a convolution layer, so that the channel number of the spliced feature is consistent with that of the input feature; and the identity mapping is used to perform element-wise summation on the output feature of the channel attenuation convolution and the feature after splicing and before convolution processing.

[0045] Compared with the prior art, the present application has the following beneficial effects:

[0046] The method of the present application is efficient and simple, and the deficiencies of traditional monitoring technology in noise suppression, feature extraction and real-time performance are effectively solved through multi-dimensional feature extraction and improved algorithm processing, thereby providing technical support for intelligent operation and maintenance of switch cabinets, significantly reducing the cost of artificial inspection, and improving the reliability of power system operation. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0048] Figure 1 Method flow chart for specific embodiments of the present application;

[0049] Figure 2 Internal structure diagram of Swin Transformer in specific embodiments of the present application;

[0050] Figure 3 Internal structure diagram of Transformer in specific embodiments of the present application;

[0051] Figure 4 Principle diagram of multi-time scale module in specific embodiments of the present application;

[0052] Figure 5 Principle diagram of noise reduction encoder in specific embodiments of the present application. DETAILED DESCRIPTION

[0053] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in connection with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0054] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0055] It is worth noting that the methods used in the present application are conventional methods unless otherwise specified; the raw materials and devices used are conventional commercially available products unless otherwise specified, and their sources are not specifically limited.

[0056] The present embodiment provides a manual switch running state monitoring method based on a noise reduction encoder, as shown in Figure 1 The present embodiment provides a manual switch running state monitoring method based on a noise reduction encoder, as shown in

[0057] When collecting the vibration signal of the manual switch, a piezoelectric acceleration sensor and a data acquisition instrument are used to form a collection system, which converts mechanical vibration into an electrical signal through piezoelectric effect. During the collection process, the vibration signal of the manual switch under multiple operating states needs to be obtained, including the opening operation, the closing operation, the normal operating state and the simulated fault state, and the vibration signal needs to contain complete mechanical action timing information.

[0058] The original vibration data collected completely records the mechanical state characteristics in the manual switch operation process, and the data has the following characteristics:

[0059] 1. Contains rich mechanical action feature information, which can accurately reflect the mechanism operation state during opening and closing operation;

[0060] 2. Accurately captures the impact vibration characteristics of the contact contact moment, providing key data for contact state evaluation;

[0061] 3. Real-time record of the vibration response of the operating mechanism, which can effectively identify abnormal conditions such as mechanical jamming and component loosening;

[0062] 4. Provides a high-fidelity data basis for subsequent real-time state monitoring, fault feature extraction and mechanical performance evaluation.

[0063] Therefore, the vibration signal collection system has high sensitivity, wide frequency response range and other characteristics, and provides reliable data support for real-time operation state detection.

[0064] The vibration signal is preprocessed, specifically including the following steps:

[0065] First, the continuously collected vibration signal is segmented according to a pre-set fixed time length to obtain a plurality of signal samples;

[0066] Then, time domain feature extraction is performed on each signal sample to generate time domain sample data, and time-frequency conversion processing is performed on each signal sample to generate frequency domain sample data.

[0067] The above preprocessing method can not only retain transient impact information, but also reflect the overall operation state of the mechanism, providing optimal feature expression for subsequent real-time state evaluation.

[0068] Based on the above time domain sample data and frequency domain sample data, the time domain features and frequency domain features are encoded and fused in parallel through a multi-feature fusion network, and the fused features are output. The multi-feature fusion network encodes the time domain and frequency domain features of the acquired signal in parallel. The preprocessed vibration signal is input into the designed multi-feature fusion network for multi-feature parallel coding. The core of the multi-feature fusion network includes a first feature processing module, a second feature processing module and a fusion module.

[0069] The first feature processing module is a waveform restoration module (WR), which is provided with a waveform encoder (WE) and a state regressor (SG). The WE includes a plurality of cascaded residual convolution blocks (RCDs). Each RCD block includes two convolution layers with a kernel size of 3, a batch normalization layer and a ReLU activation layer, and realizes down-sampling through a 4x4 convolution layer with a step size of 2. The sound wave signal is processed layer by layer through the cascaded residual convolution blocks. The convolution layer is used to extract different features of the signal, the batch normalization can accelerate the network convergence and reduce the internal covariate shift, and the ReLU activation layer increases the nonlinear expression ability of the network. The down-sampling operation reduces the data dimension while preserving the key features. Finally, the WE extracts the working voltage features from the waveform signal, providing a data basis for the state regressor to output the working voltage prediction value. The training of the WE uses a mean square error loss function, and the optimal target is that the predicted voltage is equal to the actual voltage. Further, the SG has the same network structure as the WE, and a fully connected layer is arranged at the end of the SG. Finally, the working voltage prediction value is output through the fully connected layer.

[0070] In the training process of the first feature processing module, the working voltage of the manual switch is taken as the target parameter, the working voltage prediction value is output through the state regressor, and the network parameters are adjusted using the mean square error loss function. The first feature processing module training uses a mean square error loss function As follows:

[0071] ;

[0072] In the formula, represents the actual voltage, represents the predicted voltage, and the optimal state is reached when the predicted voltage is equal to the actual voltage.

[0073] The second feature processing module is a spectrum restoration module (SR), which is provided with a spectrum encoder (SE) and a spectrum decoder (SD). The SE includes a Swin-Transformer block and a patch synthesis layer, which gradually compresses the feature size and captures global spectral features through the Swin-Transformer block and the patch synthesis layer. Specifically, as shown in Figure 2As shown, the Swin Transformer block is the core component of SE, and its internal structure includes a hierarchical processing procedure of Layer Normalization (LN), Window-based Multi-Head Self-Attention (W-MSA), Shifted Window-based Multi-Head Self-Attention (SW-MSA), Multi-Layer Perceptron (MLP), etc. The input of the Swin Transformer block is y1, and the processed results are intermediate features y2, y3, y4, etc. The final output is T. The Swin Transformer block mainly realizes feature extraction and compression through hierarchical Transformer blocks and patch merging operations. Specifically, the Swin Transformer divides the input spectral graph data y1 into non-overlapping patches, and each patch is converted into a feature vector through linear embedding. Then, the feature vector is processed through multiple consecutive Swin-Transformer blocks containing LN, W-MSA / SW-MSA, and MLP. In the processing procedure, intermediate features y2, y3, y4, etc. are generated in sequence. Each block contains a self-attention mechanism based on a moving window, i.e., the window mechanism corresponding to W-MSA and SW-MSA, and an MLP. The moving window mechanism can reduce the computational complexity while maintaining the relevance of local features. In the feature processing procedure, the features of adjacent patches are gradually merged through the patch merging layer to compress the feature map size and increase the channel number. Finally, the output feature T is obtained, thereby effectively capturing the global spectral features of the spectral graph and providing a high-quality feature basis for the subsequent spectral graph decoder (SD) to reconstruct the spectral graph.

[0074] The SD contains an inverse patch merging layer and a Swin-Transformer block, and uses the inverse patch merging layer to cooperate with the Swin-Transformer block to reconstruct the spectral graph.

[0075] In the training process of the second feature processing module, the input spectral graph is reconstructed as the target, the spectral graph decoder is used to output the reconstructed spectral graph, and a hybrid loss function containing a mean square error loss and a structural similarity loss is used to adjust the network parameters. The objective function of the second feature processing module is

[0076] ;

[0077] In the formula,​ is the Mean Squared Error loss, is the Structural Similarity Index loss, is a constant between 0 and 1, representing the proportion of the Mean Squared Error loss in training. This objective function is used to capture some good representations in the spectrogram.

[0078] Further, the Structural Similarity loss is represented by the following formula:

[0079] ;

[0080] In the formula, represents the mean, represents the variance; and are constants to maintain stability, whose formula is as follows:

[0081] ;

[0082] In the formula, L represents the dynamic range of pixel values, and are equal to 0.01 and 0.03 respectively to maintain computational stability.

[0083] The Feature Fusion Module (FM) includes a multi-layer perceptron (MLP), a convolutional layer, and a Transformer decoder. Among them, the MLP is used to process the output features of the first feature processing module; the convolutional layer is used to process the output features of the second feature processing module. Specifically, the FM first processes the waveform and spectrogram features through the MLP and the convolutional layer respectively, the MLP is used to capture global coarse-grained features, and the convolutional layer is used to extract local detailed features. The processed waveform features are used as guide information, and the spectrogram features are used as main features. The fusion process adopts the Transformer decoder structure as shown in Figure 3 , generates joint feature representation through multiple decoding layers, and finally generates fusion features.

[0084] The fusion features are input into the improved multi-time scale model, and multi-time scale fault features are output. Among them, as Figure 4As shown, the improved multi-time scale model is constructed based on a SE-DSCNN (Squeeze-and-Excitation Deeply Supervised Convolutional Neural Network) model. The multi-time scale model is constructed to extract multi-time scale fault features of the signal and enhance the discriminability of the features under noise interference. The improved multi-time scale model adds a data input layer at the input end, specifically using a wide convolutional layer 1x32, for receiving and processing the fused feature signal , and obtaining features ; wherein C and L represent the channel number and the feature map length, respectively.

[0085] Further, the internal structure of the improved multi-time scale model is a four-branch parallel convolution structure, which realizes multi-time scale local fault feature extraction; the input features of the shallow branch are aggregated as supplementary input features into the next deep branch, thereby enriching the different time scale features. Specifically, the first branch processes the input features through a one-dimensional convolutional layer; the second branch concatenates the output features of the first branch with the input features and processes them through a one-dimensional convolutional layer; the third branch concatenates the output features of the second branch with the input features and processes them through a one-dimensional convolutional layer; and the fourth branch includes a max pooling layer, specifically a max average pooling layer (MAP layer) and a convolutional layer; wherein the convolutional layer of the fourth branch uses the MAP layer to retain the key parts of the multi-time scale elements. Thus, to avoid an increase in computational burden caused by a large kernel size, a 1x3 convolution is used in the MAP layer to minimize the number of parameters while maintaining effective feature extraction. In addition, a 1x1 convolutional layer is added to adjust the dimensions and reduce the computational cost. After processing the parallel convolution, the fault features of the signal are effectively mined.

[0086] After the output features of each branch are concatenated, the feature weights are calculated through a self-attention mechanism (SA, Self-Attention), and multi-time scale fault features are generated. Specifically, the signal processed by the feature fusion module is input into the model, and first, a wide convolutional layer is used to process it to obtain features ; then, four branches are constructed in a parallel convolution manner to calculate as follows:

[0087] First branch: ; wherein Conv represents a convolution operation, and the branch output is features Figure 1 .

[0088] Second branch: ; wherein Cat represents a concatenation operation, and the branch output is features Figure 2 .

[0089] Third branch: , and the branch output is features Figure 3 .

[0090] Fourth branch: Branch output is a feature Figure 4 .

[0091] Output characteristics of each branch Features are obtained after splicing Then, feature weights are calculated using a self-attention mechanism to obtain multi-timescale fault features. .

[0092] Input multi-timescale fault features into the noise reduction encoder; such as Figure 5 As shown, in the denoising encoder, multi-timescale fault features are extracted into global and local features through pooling operations. These pooling operations include Global Average Pooling (GAP) and Maximum Average Pooling (MAP). Global Average Pooling extracts multi-timescale fault features by calculating the average value of the multi-timescale fault features across the spatial dimension. Global characteristics of distribution Max pooling extracts pulse-like local features by selecting the maximum value of fault features across multiple time scales in the spatial dimension. .

[0093] The global threshold coefficient and local threshold coefficient are generated based on global features and local features, respectively. The steps are as follows:

[0094] When generating the global threshold coefficient, the global features are input into the first type of attention module. The first type of attention module includes a batch normalization layer, a ReLU activation layer, a fully connected layer, and a Sigmoid activation layer. After the global features are processed by the batch normalization and ReLU activation layers, their dimensions are adjusted by the fully connected layer, and then the global threshold coefficient, which matches the dimensions of the global features, is output by the Sigmoid activation layer. ;in, Figure 5 The intermediate quantities shown are represented by C×W×1 / C×1×1, which are used to describe the dimensional information of the feature map. The first "C" represents the number of channels, the second "W" or "1" represents the size of the feature map in the height direction, and the third "1" represents the size of the feature map in the width direction.

[0095] When generating local threshold coefficients, local features are input into a second type of attention module. The second type of attention module has the same network structure as the first type of attention module. Local features are input into the second type of attention module, and the output is a local threshold coefficient that matches the dimension of the local features. .

[0096] The aforementioned global and local threshold coefficients are modulated element-wise with the multi-timescale fault features, that is, the coefficients are combined with the input features by element-wise multiplication to obtain different noise reduction thresholds, namely the soft smoothing thresholds corresponding to the global and local thresholds.

[0097] Based on the noise reduction threshold, a soft thresholding function (STF) is used to denoise both global and local features. However, traditional STFs incorrectly eliminate valid fault features that are close to zero. Therefore, this embodiment employs an adaptive soft thresholding function (ASSTF) for noise reduction, which adds a control term compared to the traditional soft thresholding function. Its formula is as follows:

[0098] ;

[0099] In the formula, h Input features can be global or local features; τ The noise reduction threshold is generated by modulating the global or local threshold coefficients with the fault features at multiple time scales. That is, it is the global noise reduction threshold when denoising global features and the local noise reduction threshold when denoising local features. λ It is a control item used to adjust threshold characteristics. λ When =0, it is equivalent to the traditional soft threshold function. F The output feature represents either the global or local denoised features after noise reduction. Therefore, when denoising global features, an adaptive soft thresholding function corresponding to the global threshold coefficient is used; when denoising local features, an adaptive soft thresholding function corresponding to the local threshold coefficient is used.

[0100] The denoised global and local features are then concatenated and normalized to obtain feature information. The feature concatenation and normalization operations include the following steps:

[0101] Feature concatenation connects the denoised global and local features along the channel dimension;

[0102] Feature normalization includes channel decay convolution and identity mapping. Channel decay convolution uses convolutional layers to adjust the number of channels in the concatenated features so that the number of channels in the concatenated features is consistent with the number of channels in the input features.

[0103] The identity mapping sums the output features of the channel decay convolution with the concatenated features that have not undergone convolution processing element-wise.

[0104] Finally, output features The formula is:

[0105] ;

[0106] In the formula, represents element-wise summation.

[0107] The running state monitoring result of the manual switch is output based on the feature information. A more optimal method is that, assuming that the vibration signal in the opening and closing operation process of the manual switch is monitored, the pre-processed signal collected by the sensor in one opening operation is input into the trained feature fusion multi-time scale self-help noise reduction model, and the final output is a feature vector F. Then, the feature vector F is input into a classifier composed of a fully connected layer + Softmax, which has been pre-trained and can identify four states: normal, mechanism jam, contact ablation, and other faults. The feature vector F is input into the classifier, the classifier performs calculation, and a probability distribution vector is output. For example: P(normal) = 0.02, P(mechanism jam) = 0.91, P(contact ablation) = 0.05, and P(other fault) = 0.02. The probability of mechanism jam (0.91) is much higher than that of other states. Therefore, the system finally outputs the monitoring result: "10kV manual switch running state is abnormal: suspected mechanism jam fault, confidence 91%".

[0108] The method of the embodiment aims to overcome the shortcomings of traditional real-time monitoring technology in noise suppression, feature extraction, and real-time performance, and realizes more efficient and accurate running state detection through improved technical means. The application of the method can provide technical support for intelligent operation and maintenance of switch cabinets, significantly reduce the cost of manual inspection, and improve the reliability of power system operation.

[0109] Finally, it should be noted that the above content is only used to illustrate the technical solutions of the present application, and is not a limitation on the protection scope of the present application. Simple modifications or equivalent replacements of the technical solutions of the present application made by those skilled in the art do not deviate from the essence and scope of the technical solutions of the present application.

Claims

1. A method for monitoring the operating status of a manual switch based on a noise-reducing encoder, characterized in that, Includes the following steps: Collect vibration signals from manual switches; The vibration signal is preprocessed to generate time-domain sample data and frequency-domain sample data for feature extraction; Based on the time-domain sample data and frequency-domain sample data, the time-domain features and frequency-domain features are encoded and fused in parallel through a multi-feature fusion network, and the fused features are output. The fused features are input into the improved multi-timescale model to output multi-timescale fault features. The improved multi-timescale model is constructed based on the SE-DSCNN model, and a data input layer is added at the input end. The internal structure of the improved multi-timescale model is a four-branch parallel convolutional structure, and the first branch processes the input features through a one-dimensional convolutional layer. The second branch concatenates the output features of the first branch with the input features and processes them through a one-dimensional convolutional layer; the third branch concatenates the output features of the second branch with the input features and processes them through a one-dimensional convolutional layer; the fourth branch includes a max pooling layer and a convolutional layer; after the output features of each branch are concatenated, feature weights are calculated through a self-attention mechanism, and the multi-timescale fault features are generated. The multi-timescale fault features are input into the noise reduction encoder; in the noise reduction encoder, the multi-timescale fault features are used to extract global and local features through pooling operations. Global threshold coefficients and local threshold coefficients are generated based on the global and local features, respectively; the global and local threshold coefficients are modulated unequally with the multi-timescale fault features to obtain a noise reduction threshold; based on the noise reduction threshold, a soft thresholding function is used to denoise the global and local features; the denoised global and local features are then concatenated and regularized to obtain feature information. The operating status monitoring results of the manual switch are output based on the aforementioned feature information.

2. The method for monitoring the operating status of a manual switch based on a noise-reducing encoder according to claim 1, characterized in that: When collecting vibration signals from a manual switch, an electrical signal of the mechanical vibration of the manual switch is collected using a vibration sensor. The data acquisition process includes vibration signals from manual switches under various operating conditions, including opening operation, closing operation, normal operation, and simulated fault condition. The vibration signal contains complete timing information of the mechanical action.

3. The method for monitoring the operating status of a manual switch based on a noise-reducing encoder according to claim 1, characterized in that: The vibration signal is preprocessed, including the following steps: The continuously collected vibration signals are segmented according to a preset fixed duration to obtain multiple signal samples; Time-domain features are extracted from each signal sample to generate the time-domain sample data; Each signal sample undergoes time-frequency conversion processing to generate the frequency domain sample data.

4. The method for monitoring the operating status of a manual switch based on a noise-reducing encoder according to claim 1, characterized in that: The multi-feature fusion network includes a first feature processing module, a second feature processing module, and a fusion module; The first feature processing module includes a waveform encoder and a state regressor; the waveform encoder contains multiple cascaded residual convolutional blocks, each residual convolutional block having a convolutional layer, a normalization layer and a ReLU activation layer; the state regressor has the same network structure as the waveform encoder, and the end of the state regressor is provided with a fully connected layer; The second feature processing module includes a spectrogram encoder and a spectrogram decoder; the spectrogram encoder includes a Swing-Transformer block and a patch synthesis layer; the spectrogram decoder includes a reverse patch synthesis layer and a Swing-Transformer block. The fusion module includes a multilayer perceptron, a convolutional layer, and a Transformer decoder; the multilayer perceptron is used to process the output features of the first feature processing module; the convolutional layer is used to process the output features of the second feature processing module.

5. The method for monitoring the operating status of a manual switch based on a noise-reducing encoder according to claim 4, characterized in that: During the training process of the first feature processing module, the operating voltage of the manual switch is used as the target parameter. The predicted value of the operating voltage is output by the state regressor, and the network parameters are adjusted by the mean square error loss function. During the training process of the second feature processing module, with the goal of reconstructing the input spectrogram, the reconstructed spectrogram is output through the spectrogram decoder, and the network parameters are adjusted using a hybrid loss function that includes mean square error loss and structural similarity loss. The fusion module performs multi-level decoding on the waveform features output by the first feature processing module and the spectral features output by the second feature processing module through a Transformer decoder to generate the fused features.

6. The method for monitoring the operating status of a manual switch based on a noise-reducing encoder according to claim 1, characterized in that: The pooling operations include global average pooling and max pooling; The global average pooling extracts the distributed global features by calculating the average value of the multi-timescale fault features in the spatial dimension. The max pooling method extracts local features in the form of pulses by selecting the maximum value of the multi-timescale fault features in the spatial dimension.

7. The method for monitoring the operating status of a manual switch based on a noise-reducing encoder according to claim 1, characterized in that: When generating the global threshold coefficient, the global features are input into the first type of attention module; The first type of attention module includes a batch normalization layer, a ReLU activation layer, a fully connected layer, and a Sigmoid activation layer; after the global features are processed by the batch normalization and ReLU activation layers, the dimensions are adjusted by the fully connected layer, and then the global threshold coefficient that matches the dimensions of the global features is output by the Sigmoid activation layer. When generating local threshold coefficients, the local features are input into a second type of attention module. The second type of attention module has the same network structure as the first type of attention module. The local features are input into the second type of attention module, and the output is a local threshold coefficient that matches the dimension of the local features.

8. The method for monitoring the operating status of a manual switch based on a noise-reducing encoder according to claim 1, characterized in that: The soft threshold function is an improved adaptive soft threshold function, with an added control term; When denoising global features, an adaptive soft thresholding function corresponding to the global threshold coefficient is used; When denoising local features, an adaptive soft thresholding function corresponding to the local threshold coefficient is used.

9. The method for monitoring the operating status of a manual switch based on a noise-reducing encoder according to claim 1, characterized in that: The feature splicing and feature regularization operations include the following steps: Feature concatenation connects the denoised global and local features along the channel dimension; Feature normalization includes channel decay convolution and identity mapping. The channel decay convolution uses convolutional layers to adjust the number of channels in the concatenated features so that the number of channels in the concatenated features is consistent with the number of channels in the input features. The identity mapping sums the output features of the channel decay convolution with the concatenated features that have not undergone convolution processing element by element.

Images