GIS equipment insulation defect identification method based on recurrent neural network

Abstract

The invention is suitable for the technical field of equipment defect identification, and provides a GIS equipment insulation defect identification method based on a recurrent neural network, and the method comprises the steps: collecting a multi-channel time sequence monitoring signal of GIS equipment; extracting time sequence features by using a recurrent neural network, and enhancing effective signal segments in the time sequence features by using an attention mechanism; processing through a pre-trained position classification model, and outputting defect position probability distribution vectors of the GIS electrical component units at preset candidate positions; constructing a graph structure, inputting the defect position probability distribution vector as a node feature into a graph neural network, and generating defect position probability distribution corrected based on spatial correlation through neighborhood feature aggregation; and comparing the defect probability value corresponding to each corrected candidate position with a preset dynamic threshold value to generate an insulation defect identification result. The false detection rate and the omission rate of equipment defect detection are effectively reduced, and the requirement of a power system for high-precision intelligent detection of GIS equipment is met.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment defect identification, and in particular to a GIS equipment insulation defect identification method based on a recurrent neural network. Background Art

[0002] In the field of intelligent detection of power equipment, insulation defect detection in gas-insulated switchgear (GIS) is crucial. As a key device in the power system, the insulation performance of GIS is directly related to the safe and stable operation of the power system. Traditional methods are mainly based on static features in the time or frequency domain and shallow machine learning models. However, time or frequency domain statistics only reflect the instantaneous or global distribution characteristics of the signal and cannot characterize the dynamic evolution of the signal on the time axis. For example, the ultrasonic signal caused by the motion of free metal particles has periodic time-varying characteristics, and the expansion of defects on the surface of insulators is accompanied by a gradual change in the amplitude of the partial discharge signal. Static statistics will lose such time series correlation. Shallow machine learning algorithms are essentially static classifiers whose input requires a fixed-dimensional feature vector and cannot directly model the long-term dependencies of the time series. Therefore, for the dynamic propagation process of defects, existing methods have difficulty establishing a mapping relationship between input features and dynamic defects, resulting in errors in equipment defect detection and difficulty in meeting high-precision detection requirements.

[0003] In view of this, a GIS equipment insulation defect identification method based on recurrent neural network is needed. Summary of the Invention

[0004] This application provides a GIS equipment insulation defect identification method based on a recurrent neural network, which is used to solve the problem that errors occur in equipment defect detection and it is difficult to meet high-precision detection requirements.

[0005] This application provides a GIS equipment insulation defect identification method based on a recurrent neural network, including:

[0006] Collecting multi-channel time series monitoring signals of a GIS device, wherein the GIS device includes a plurality of GIS electrical component units;

[0007] A recurrent neural network is used to extract the time series features corresponding to the multi-channel time series monitoring signal after removing noise interference, and an attention mechanism is used to enhance the effective signal segments in the time series features;

[0008] The enhanced time series features are processed by the pre-trained location classification model to output the defect location probability distribution vector of each GIS electrical component unit at the preset candidate location;

[0009] A graph structure is constructed with GIS electrical component units as nodes and the physical connection relationships between GIS electrical component units as edges. The defect location probability distribution vector is input into the graph neural network as the node feature. The defect location probability distribution corrected based on spatial correlation is generated by aggregating neighborhood features.

[0010] The corrected defect probability values ​​corresponding to the candidate positions are compared with the preset dynamic threshold, and the three-dimensional coordinates corresponding to the candidate positions whose probability values ​​exceed the threshold are extracted as the insulation defect recognition results.

[0011] Furthermore, the extraction of the time series features corresponding to the multi-channel time series monitoring signal after removing noise interference by using a recurrent neural network includes:

[0012] Performing normalization preprocessing on the multi-channel time series monitoring signal, using wavelet transform to remove Gaussian noise and periodic interference noise in the signal, and generating noise-reduced multi-channel time series data;

[0013] The denoised multi-channel time series data is converted into a three-dimensional tensor containing time dimension, channel dimension and feature dimension, and input into a recurrent neural network composed of a long short-term memory network;

[0014] The historical state of each channel time series and the current time step input are iteratively calculated through the hidden layer of the recurrent neural network to extract a time series feature vector containing time dependency and channel correlation.

[0015] Furthermore, the use of the attention mechanism to enhance the effective signal segments in the time series features includes:

[0016] Input the time series feature vector output by the recurrent neural network into the self-attention layer to calculate the attention weight of the signal segment at each time step;

[0017] Normalizing the attention weights to generate a weight vector reflecting the importance of the signal segment at each time step;

[0018] A weighted summation is performed on the time series feature vector based on the weight vector, thereby enhancing the amplitude of the effective signal segment containing the insulation defect feature and suppressing the interference of the irrelevant noise signal segment.

[0019] Furthermore, the pre-trained location classification model includes:

[0020] The input layer is used to receive a three-dimensional time series feature tensor containing time dimension, channel dimension, and feature dimension;

[0021] The convolution layer extracts the spatial correlation features between each monitoring channel through the convolution kernel of a preset size, and the preset size matches the physical layout dimensions of the GIS electrical component unit;

[0022] The fully connected layer is combined with a position encoding module, which generates a position embedding vector based on the three-dimensional spatial coordinate information of the GIS device, and then concatenates it with the output features of the convolutional layer and inputs it into the multi-layer perceptron.

[0023] The output layer generates a dimensional vector equal to the number of preset candidate positions through linear transformation. The preset candidate positions are determined based on the historical defect data and structural design drawings of the GIS equipment.

[0024] Furthermore, the processing of the enhanced time series features includes:

[0025] Transform the time series feature vector output by the attention mechanism from a three-dimensional tensor to a two-dimensional feature matrix, where the row dimension corresponds to the time step and the column dimension integrates the channel feature and the time feature;

[0026] Performing a sliding window pooling process on the two-dimensional feature matrix, where the window size is equal to the signal acquisition period of a single GIS electrical component unit, to generate a local feature vector reflecting the feature changes within the component unit time period;

[0027] The local feature vector is input into the input layer of the position classification model, and the three-dimensional spatial coordinate code of the corresponding GIS electrical component unit is simultaneously input as the position prior information.

[0028] Furthermore, the outputting of the defect location probability distribution vector of each GIS electrical component unit at the preset candidate location includes:

[0029] Generate an initial score vector with a dimension equal to the number of preset candidate locations through the output layer of the location classification model;

[0030] The initial score vector is normalized to generate a probability distribution vector; each element of the probability distribution vector corresponds to a preset candidate position, and the value of each element of the probability distribution vector represents the probability of an insulation defect occurring in the corresponding GIS electrical component unit at the preset candidate position.

[0031] Furthermore, the correcting the defect location probability distribution based on spatial correlation includes:

[0032] The defect probability distribution vectors of the current GIS electrical component unit and its adjacent units are aggregated through graph neural networks to generate probability values ​​with corrected spatial correlation.

[0033] Furthermore, the calculation formula for correlation correction includes:

[0034]

[0035] in: is the corrected defect probability distribution vector, p i and p j is the original defect probability distribution vector of unit j of GIS electrical component unit i, which represents the initial predicted probability of insulation defects occurring in each preset position of the GIS component unit, λ is the self-feature retention coefficient, is the set of adjacent units that are physically connected to the IS electrical component unit i, w ij is the spatial correlation weight between unit i and unit j, d i is the neighborhood weight sum of IS electrical component unit i, c i and c j are the spatial coordinates of unit i and unit j, and σ is the spatial attenuation coefficient.

[0036] Furthermore, the calculation formula for the graph neural network neighborhood aggregation includes:

[0037]

[0038] in: The defect features extracted by the neural network in the l+1 layer for the GIS electrical component unit i, the initial layer φ is the activation function, is the weight matrix of its own feature transformation, is the attention weight of unit j to unit i, reflecting the influence of the defect information of the neighboring unit on the current unit, and are the defect features extracted by unit i and unit J in the l-th layer of the neural network, a is the attention mechanism parameter vector, and || is the feature of the merged current unit and the neighboring unit.

[0039] Furthermore, the step of comparing the corrected defect probability values ​​corresponding to the candidate positions with a preset dynamic threshold value and extracting the three-dimensional coordinates corresponding to the candidate positions whose probability values ​​exceed the threshold value as insulation defect identification results includes:

[0040] Calculate preset dynamic thresholds under different operating conditions based on historical monitoring data and defect records of GIS equipment;

[0041] The defect probability value corresponding to each corrected candidate position is compared element by element with the preset dynamic threshold under the corresponding operating condition; if the defect probability value of a candidate position is greater than the corresponding threshold, the three-dimensional coordinates of the candidate position recorded in the three-dimensional layout model of the GIS equipment are extracted, and the three-dimensional coordinates of all candidate positions that exceed the threshold are combined into a set as the final insulation defect identification result.

[0042] It can be seen from the above technical solutions that this application has the following advantages:

[0043] This application uses a recurrent neural network to extract the dynamic timing features of multi-channel time series monitoring signals, and combines the attention mechanism to enhance the representation ability of effective signal segments, thereby solving the problem that traditional static features cannot capture the dynamic evolution law of insulation defects; using a graph neural network to aggregate the spatial correlation features between GIS component units, correct the defect probability distribution, and solve the defect that the shallow model does not adequately model the physical topology structure of the equipment; through dynamic threshold adaptive matching of defect judgment criteria under different working conditions, the positioning accuracy and reliability of insulation defect detection are effectively improved, which is especially suitable for complex dynamic defect scenarios such as free particle motion and gradual partial discharge. It can effectively reduce the false detection rate and missed detection rate, and meet the power system's demand for high-precision intelligent detection of GIS equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 The figure is a flow chart of an embodiment of a method for identifying insulation defects of GIS equipment based on a recurrent neural network in the present invention. DETAILED DESCRIPTION

[0045] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "corresponding to" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0046] Example 1

[0047] The implementation method in this embodiment can be implemented in the system, can be implemented in the server, and can also be implemented in the terminal, and the specific implementation is not clearly limited. The following will introduce the GIS equipment insulation defect identification method based on recurrent neural network in this application from the perspective of system implementation. Figure 1 , the method provided in the embodiment of the present application includes the following steps:

[0048] S11. Collecting multi-channel time series monitoring signals of a GIS device, where the GIS device includes several GIS electrical component units;

[0049] In this embodiment, vibration sensors, temperature sensors, ultra-high frequency (UHF) partial discharge sensors, and current / voltage monitoring modules are deployed at defect-prone locations of GIS equipment, such as circuit breaker contacts, insulator surfaces, busbar connections, and bushing sealing surfaces, to form a multi-dimensional monitoring network. Each sensor sets a sampling frequency based on signal characteristics, such as 1MHz for vibration signals and 300MHz-3GHz for partial discharge signals. A distributed data acquisition system is combined with the GPS time synchronization protocol to ensure strict time alignment of multi-channel signals. After anti-aliasing filtering and analog-to-digital conversion (A / D), the raw data is constructed into a three-dimensional data matrix based on timestamps, channel identifiers, and characteristic parameters. The time dimension records the dynamic evolution of the signal with millisecond-level accuracy. The channel dimension is mapped to specific component units and sensor types. The characteristic dimension contains the quantized value of the raw signal and the time domain statistics calculated in real time.

[0050] During the acquisition process, abnormal signal segments that trigger threshold alarms, such as excessive partial discharge amplitude and abnormal temperature rise rate, are labeled and marked together with normal operating data to form a multi-channel time series data set, providing an accurate monitoring data basis for subsequent dynamic defect feature extraction.

[0051] S12. Using a recurrent neural network to extract the time series features corresponding to the multi-channel time series monitoring signal after removing noise interference, and using an attention mechanism to enhance the effective signal segments in the time series features;

[0052] In this embodiment, multi-stage signal processing and deep learning models are used to extract dynamic time series features from multi-channel monitoring signals and enhance the effective signal segments related to insulation defects. A recurrent neural network is used to extract the time series features corresponding to the multi-channel time series monitoring signals after removing noise interference, including:

[0053] 1. Perform normalization preprocessing on the multi-channel time series monitoring signal, use wavelet transform to remove Gaussian noise and periodic interference noise in the signal, and generate noise-reduced multi-channel time series data;

[0054] Normalize the original monitoring signals (vibration, temperature, partial discharge, electrical parameters, etc.) to eliminate sensor range differences; select the db4 or sym8 wavelet basis to perform multi-layer decomposition of the signal, where the number of layers J = 3 to 5, and suppress noise through threshold processing; hard threshold: Soft Thresholding: in σ is the standard deviation of noise, N is the signal length. Signal reconstruction is the detail coefficient after noise reduction The inverse wavelet transform is performed on the approximate coefficients to generate the noise-reduced signal.

[0055] 2. Convert the denoised multi-channel time series data into a three-dimensional tensor consisting of time, channel, and feature dimensions, and input it into a recurrent neural network consisting of a long short-term memory network;

[0056] 3D tensor after denoising Where T is the time step, C is the number of monitoring channels, and F is the single channel feature dimension.

[0057] 3. Through the hidden layer of the recurrent neural network, the historical state of each channel time series and the current time step input are iteratively calculated to extract the time series feature vector containing time dependency and channel correlation.

[0058] The three-dimensional tensor is input into a recurrent neural network composed of a long short-term memory (LSTM) network to extract dynamic features across channels and time. The network structure: The hidden layer consists of 1 to 3 LSTM layers, each containing 128 to 512 memory units; the input dimension is F, the single-channel feature dimension; the output dimension is D, the hidden layer dimension, such as 256. The calculation formula of the LSTM gating mechanism is:

[0059] Forget gate: f t =σ(W f ·[h t―1 ,x t ]+b f ); Input gate: i t =σ(W i ·[h t―1 ,x t ]+b i ); Candidate status: Cell status update: Output gate: o t =σ(W o ·[h t―1 ,x t ]+b o ); hidden state: h t =o t ⊙tanh(C t );in, Input features for the current time step, is the hidden state at the previous time step, is the cell state at the previous time step, W * ,b * are trainable weights and biases, σ is the Sigmoid activation function, and ⊙ is element-by-element multiplication. Output time series feature matrix Characterize the dynamic evolution of signals, such as the periodicity and temperature rise trend of partial discharge pulses.

[0060] In this embodiment, the time series feature vector output by the LSTM is input into the self-attention layer, and the effective signal segment containing the insulation defect feature is enhanced through the following steps:

[0061] 1. Input the time series feature vector output by the recurrent neural network into the self-attention layer to calculate the attention weight of the signal segment at each time step;

[0062] 2. Normalize the attention weights to generate a weight vector that reflects the importance of the signal segment at each time step;

[0063] 3. Perform weighted summation of the time series feature vector based on the weight vector to enhance the amplitude of the effective signal segment containing the insulation defect feature and suppress the interference of irrelevant noise signal segments.

[0064] Map the time series feature vector into query vector, key vector and value vector respectively: Q = W Q H,K=W K The correlation score of the signal segment at each time step is calculated by dot product operation: in is the query, key, and value vector. Perform softmax normalization on the score matrix to generate the attention weight matrix where α i,j Indicates the degree of attention of the i-th time step to the j-th time step, reflecting the importance of the signal segment of this time step in defect recognition Based on the attention weight matrix, the value vector V is weighted and summed to obtain the enhanced temporal feature vector h attn : By amplifying the signal amplitude of high-weight time steps and suppressing low-weight noise segments, targeted enhancement of effective signal segments related to insulation defects is achieved.

[0065] S13. The enhanced time series features are processed by the pre-trained position classification model to output the defect location probability distribution vector of each GIS electrical component unit at the preset candidate location;

[0066] In this embodiment, the pre-trained location classification model is a hierarchical neural network architecture adapted to the GIS equipment insulation defect identification scenario, which includes the following core modules:

[0067] 1. The input layer is used to receive a three-dimensional time series feature tensor containing time dimension, channel dimension, and feature dimension;

[0068] 2. The convolution layer extracts the spatial correlation features between each monitoring channel using a convolution kernel of a preset size. The convolution kernel size matches the physical layout dimensions of the GIS electrical component units.

[0069] 3. The fully connected layer is combined with the position encoding module. The position encoding module generates a position embedding vector based on the 3D spatial coordinate information of the GIS device, concatenates it with the output features of the convolutional layer, and then inputs it into the multi-layer perceptron.

[0070] 4. The output layer generates a dimensional vector equal to the number of preset candidate locations through linear transformation. The preset candidate locations are determined based on the historical defect data and structural design drawings of the GIS equipment.

[0071] Specifically, the model input is a three-dimensional time series feature tensor enhanced with an attention mechanism, with dimensions: time steps × number of channels × single-channel features. After normalization at the input layer, it is fed into a convolutional layer to extract spatially correlated features. The convolution kernel size is customized based on the physical layout of the GIS components. For example, if the sensors at a busbar connection are spatially arranged in 2 rows and 3 columns, a 2×3 convolution kernel is used to capture the propagation patterns of adjacent sensor signals, such as the diffusion path of partial discharges along the busbar. The fully connected layer, combined with the position encoding module, encodes the three-dimensional coordinates (x, y, z) of the component unit into a position embedding vector. This vector is then concatenated with the convolutional features and fed into a multilayer perceptron to fuse the spatiotemporal information. The output layer uses a linear transformation to generate a raw score vector equal to the number of preset candidate locations. The candidate locations are determined based on historically high-incidence areas of defects (circuit breaker contacts, insulator surfaces) and equipment structural design drawings (busbar connections), providing a spatial prior for subsequent probability distribution.

[0072] In this embodiment, processing the enhanced time series features includes the following steps:

[0073] 1. Transform the time series feature vector output by the attention mechanism from a three-dimensional tensor to a two-dimensional feature matrix, where the row dimension corresponds to the time step and the column dimension integrates the channel features and time features;

[0074] 2. Perform sliding window pooling on the two-dimensional feature matrix, with a window size equal to the signal acquisition period of a single GIS electrical component unit, to generate a local feature vector that reflects the feature changes within the component unit time period;

[0075] 3. Input the local feature vector into the input layer of the location classification model, and simultaneously input the three-dimensional spatial coordinate code of the corresponding GIS electrical component unit as the location prior information.

[0076] The three-dimensional time series feature tensor output by the attention mechanism is first transformed in dimension, merging the channel dimension and feature dimension into a two-dimensional matrix (rows correspond to time steps, and columns fuse multi-channel features). A sliding window pooling operation is then used, with the window size consistent with the signal acquisition period of a single GIS component unit (e.g., every 100 time steps). Mean pooling is used to extract the trend of feature changes within the period, such as the periodic signal of free metal particle motion. The pooled local feature vector is concatenated with the three-dimensional coordinate encoding of the component unit (normalized to the interval [0,1]) to form an input feature that integrates spatial position information. For example, after pooling, the periodic amplitude fluctuations of the vibration signal at the busbar connection are combined with the coordinate encoding to enhance the model's perception of the propagation of spatially adjacent defects.

[0077] After processing the enhanced time series features through the pre-trained location classification model, the defect probability distribution vector of each GIS electrical component unit at the preset candidate location is finally output. The specific process is as follows:

[0078] 1. Generate an initial score vector with a dimension equal to the number of preset candidate locations through the output layer of the location classification model;

[0079] The output layer of the location classification model maps the high-dimensional representation of the fused spatiotemporal features into a raw score vector with a dimension equal to the number of preset candidate locations through linear transformation. The candidate locations are pre-defined based on the historical defect data of GIS equipment (such as 30% of circuit breaker contact defects and 40% of insulator surface defects) and structural design drawings (with vulnerable connection nodes marked), covering typical high-incidence areas of defects such as busbar connections and bushing interfaces.

[0080] 2. Normalize the initial score vector to generate a probability distribution vector; each element of the probability distribution vector corresponds to a preset candidate location, and the value of each element of the probability distribution vector represents the probability of an insulation defect occurring in the corresponding GIS electrical component unit at the preset candidate location.

[0081] The initial score vector is converted into a defect probability distribution vector through softmax normalization, where each element represents the probability of an insulation defect occurring at the corresponding candidate location. For example, after normalization, the probability value for the contact surface of a circuit breaker unit is significantly higher than that for other locations, reflecting the high incidence of contact defects in historical data. This probability distribution vector serves as the spatially corrected input of the graph neural network, providing a quantitative basis for defect localization while ensuring the interpretability of the model output and compatibility with dynamic threshold determination.

[0082] The above steps realize spatial position-sensitive processing of enhanced time series features, and generate defect probability distribution based on the physical layout information of GIS equipment, providing high-precision initial probability input for subsequent spatial correlation correction based on graph neural networks, effectively improving the positioning accuracy and reliability of insulation defect identification.

[0083] S14. Construct a graph structure with GIS electrical component units as nodes and the physical connections between GIS electrical component units as edges. Input the defect location probability distribution vector as the node feature into the graph neural network. Aggregate the neighborhood features to generate a defect location probability distribution corrected for spatial correlation.

[0084] In this embodiment, correcting the defect location probability distribution based on spatial correlation involves aggregating the defect probability distribution vectors of the current GIS electrical component unit and its adjacent units through a graph neural network to generate probability values ​​corrected for spatial correlation. The graph structure represents a node for each GIS electrical component unit; if two components have a direct physical connection, an undirected edge is established; and the initial node feature is the defect probability distribution vector output by the location classification model, representing the defect probability of the component at several candidate locations.

[0085] GNN generates a modified probability distribution by aggregating the features of adjacent nodes:

[0086]

[0087] in: is the corrected defect probability distribution vector, p i and p j is the original defect probability distribution vector of unit j of GIS electrical component unit i, which represents the initial predicted probability of insulation defects occurring in each preset position of the GIS component unit, λ is the self-feature retention coefficient, is the set of adjacent units that are physically connected to the IS electrical component unit i, w ij is the spatial correlation weight between unit i and unit j, d i is the neighborhood weight sum of IS electrical component unit i, c i and c j are the spatial coordinates of unit i and unit j, and σ is the spatial attenuation coefficient.

[0088] The calculation formula of graph neural network neighborhood aggregation is:

[0089]

[0090] in: The defect features extracted by the neural network in the l+1 layer for the GIS electrical component unit i, the initial layer φ is the activation function, is the weight matrix of its own feature transformation, is the attention weight of unit j to unit i, reflecting the influence of the defect information of the neighboring unit on the current unit, and are the defect features extracted by unit i and unit J in the l-th layer of the neural network, a is the attention mechanism parameter vector, and || is the feature of the merged current unit and the neighboring unit.

[0091] The above steps deeply integrate the physical topology of GIS equipment with the defect propagation law through spatial correlation correction of graph neural networks, solving the problem of insufficient modeling of multi-component correlation in traditional methods and significantly improving the detection robustness in complex defect scenarios.

[0092] S15. Compare the corrected defect probability values ​​corresponding to the candidate positions with a preset dynamic threshold, and extract the three-dimensional coordinates corresponding to the candidate positions whose probability values ​​exceed the threshold as insulation defect recognition results.

[0093] 1. Calculate preset dynamic thresholds under different operating conditions based on historical monitoring data and defect records of GIS equipment;

[0094] 2. The corrected defect probability values ​​corresponding to each candidate location are compared element by element with the preset dynamic threshold value under the corresponding operating condition. If the defect probability value of a candidate location is greater than the corresponding threshold, the three-dimensional coordinates of the candidate location recorded in the three-dimensional layout model of the GIS equipment are extracted, and the three-dimensional coordinates of all candidate locations that exceed the threshold are combined into a set as the final insulation defect identification result.

[0095] A dynamic threshold determination strategy matches the corrected defect probability distribution with the actual operating conditions of GIS equipment, outputting accurate insulation defect location results. In specific implementation, based on historical monitoring data (such as the distribution of partial discharge amplitudes under normal and high-load conditions) and defect records (such as the temperature rise rate when insulator surface defects occur), a sliding window statistical method or exponential moving average (EMA) is used to calculate dynamic thresholds for different operating conditions. For example, the threshold is increased under high load to avoid false positives. The corrected probability values ​​of each candidate location are compared element-by-element with the threshold for the corresponding operating condition. If the probability of a location exceeds the threshold (such as the probability at a busbar connection > 0.6), its coordinates (x, y, z) are extracted from the GIS 3D layout model to generate a set of defect location coordinates. This set can be directly mapped to the physical structure of the equipment, guiding operation and maintenance personnel to quickly locate the defect point (such as the coordinates of the contact surface of a circuit breaker). Furthermore, the dynamic threshold is adapted to different operating conditions (such as low-temperature startup and full-load operation), significantly reducing the risk of missed and false positives.

[0096] It is understandable that those skilled in the art can, under the guidance of the above embodiments, combine various implementation methods in the above embodiments to obtain technical solutions of multiple implementation methods.

[0097] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A GIS equipment insulation defect identification method based on recurrent neural network, characterized in that: include: Collecting multi-channel time series monitoring signals of a GIS device, wherein the GIS device includes a plurality of GIS electrical component units; A recurrent neural network is used to extract the time series features corresponding to the multi-channel time series monitoring signal after removing noise interference, and an attention mechanism is used to enhance the effective signal segments in the time series features; The enhanced time series features are processed by the pre-trained location classification model to output the defect location probability distribution vector of each GIS electrical component unit at the preset candidate location; A graph structure is constructed with GIS electrical component units as nodes and the physical connection relationships between GIS electrical component units as edges. The defect location probability distribution vector is input into the graph neural network as the node feature. The defect location probability distribution corrected based on spatial correlation is generated by aggregating neighborhood features. The corrected defect probability values ​​corresponding to the candidate positions are compared with the preset dynamic threshold, and the three-dimensional coordinates corresponding to the candidate positions whose probability values ​​exceed the threshold are extracted as the insulation defect recognition results.

2. The GIS equipment insulation defect identification method based on recurrent neural network according to claim 1 is characterized in that: The method of extracting the time series features corresponding to the multi-channel time series monitoring signal after removing noise interference by using a recurrent neural network includes: Performing normalization preprocessing on the multi-channel time series monitoring signal, using wavelet transform to remove Gaussian noise and periodic interference noise in the signal, and generating noise-reduced multi-channel time series data; The denoised multi-channel time series data is converted into a three-dimensional tensor containing time dimension, channel dimension and feature dimension, and input into a recurrent neural network composed of a long short-term memory network; The historical state of each channel time series and the current time step input are iteratively calculated through the hidden layer of the recurrent neural network to extract a time series feature vector containing time dependency and channel correlation.

3. The GIS equipment insulation defect identification method based on recurrent neural network according to claim 2 is characterized in that: The use of the attention mechanism to enhance the effective signal segment in the time series feature includes: Input the time series feature vector output by the recurrent neural network into the self-attention layer to calculate the attention weight of the signal segment at each time step; Normalizing the attention weights to generate a weight vector reflecting the importance of the signal segment at each time step; A weighted summation is performed on the time series feature vector based on the weight vector, thereby enhancing the amplitude of the effective signal segment containing the insulation defect feature and suppressing the interference of the irrelevant noise signal segment.

4. The GIS equipment insulation defect identification method based on recurrent neural network according to claim 1 is characterized in that: The pre-trained location classification model includes: The input layer is used to receive a three-dimensional time series feature tensor containing time dimension, channel dimension, and feature dimension; The convolution layer extracts the spatial correlation features between each monitoring channel through the convolution kernel of a preset size, and the preset size matches the physical layout dimensions of the GIS electrical component unit; The fully connected layer is combined with a position encoding module, which generates a position embedding vector based on the three-dimensional spatial coordinate information of the GIS device, and then concatenates it with the output features of the convolutional layer and inputs it into the multi-layer perceptron. The output layer generates a dimensional vector equal to the number of preset candidate positions through linear transformation. The preset candidate positions are determined based on the historical defect data and structural design drawings of the GIS equipment.

5. The GIS equipment insulation defect identification method based on recurrent neural network according to claim 4 is characterized in that: The processing of the enhanced time series features includes: Transform the time series feature vector output by the attention mechanism from a three-dimensional tensor to a two-dimensional feature matrix, where the row dimension corresponds to the time step and the column dimension integrates the channel feature and the time feature; Performing a sliding window pooling process on the two-dimensional feature matrix, where the window size is equal to the signal acquisition period of a single GIS electrical component unit, to generate a local feature vector reflecting the feature changes within the component unit time period; The local feature vector is input into the input layer of the position classification model, and the three-dimensional spatial coordinate code of the corresponding GIS electrical component unit is simultaneously input as the position prior information.

6. The GIS equipment insulation defect identification method based on recurrent neural network according to claim 5 is characterized in that: Outputting the defect location probability distribution vector of each GIS electrical component unit at a preset candidate location includes: Generate an initial score vector with a dimension equal to the number of preset candidate locations through the output layer of the location classification model; The initial score vector is normalized to generate a probability distribution vector; each element of the probability distribution vector corresponds to a preset candidate position, and the value of each element of the probability distribution vector represents the probability of an insulation defect occurring in the corresponding GIS electrical component unit at the preset candidate position.

7. The GIS equipment insulation defect identification method based on recurrent neural network according to claim 1 is characterized in that: The method of correcting the probability distribution of defect positions based on spatial correlation includes: The defect probability distribution vectors of the current GIS electrical component unit and its adjacent units are aggregated through graph neural networks to generate probability values ​​with corrected spatial correlation.

8. The GIS equipment insulation defect identification method based on recurrent neural network according to claim 7 is characterized in that: The calculation formula for the correlation correction includes: in: is the corrected defect probability distribution vector, p i and p j is the original defect probability distribution vector of unit j of GIS electrical component unit i, which represents the initial predicted probability of insulation defects occurring in each preset position of the GIS component unit, λ is the self-feature retention coefficient, is the set of adjacent units that are physically connected to the IS electrical component unit i, w ij is the spatial correlation weight between unit i and unit j, d i is the neighborhood weight sum of IS electrical component unit i, c i and c j are the spatial coordinates of unit i and unit j, and σ is the spatial attenuation coefficient.

9. The GIS equipment insulation defect identification method based on recurrent neural network according to claim 8 is characterized in that: The calculation formula for the graph neural network neighborhood aggregation includes: in: The defect features extracted by the neural network in the l+1 layer for the GIS electrical component unit i, the initial layer φ is the activation function, is the weight matrix of its own feature transformation, is the attention weight of unit j to unit i, reflecting the influence of the defect information of the neighboring unit on the current unit, and are the defect features extracted by unit i and unit J in the l-th layer of the neural network, a is the attention mechanism parameter vector, and || is the feature of the merged current unit and the neighboring unit.

10. The GIS equipment insulation defect identification method based on recurrent neural network according to claim 1 is characterized in that: The step of comparing the corrected defect probability values ​​corresponding to the candidate positions with a preset dynamic threshold value and extracting the three-dimensional coordinates corresponding to the candidate positions whose probability values ​​exceed the threshold value as insulation defect identification results includes: Calculate preset dynamic thresholds under different operating conditions based on historical monitoring data and defect records of GIS equipment; The defect probability value corresponding to each corrected candidate position is compared element by element with the preset dynamic threshold under the corresponding operating condition; if the defect probability value of a candidate position is greater than the corresponding threshold, the three-dimensional coordinates of the candidate position recorded in the three-dimensional layout model of the GIS equipment are extracted, and the three-dimensional coordinates of all candidate positions that exceed the threshold are combined into a set as the final insulation defect identification result.

Images