An adaptive learning power grid equipment abnormality detection and diagnosis method and system

Abstract

The present invention relates to the technical field of fault diagnosis of power grid equipment, and in particular to an adaptive learning power grid equipment abnormality detection and diagnosis method and system, comprising the following steps: fusing the collected data of different modes, firstly encoding and decoding the data based on the informer network structure, then introducing a score-guided network into the structure, taking the encoded data as input, taking the data and the cluster center distance ratio of normal and abnormal data as labels, training to obtain scores, and then designing a loss function based on the scores, input data and decoded data, to achieve automatic updating and optimization of the network structure, and adapt to the changing equipment status and environmental conditions. The present invention adopts an unsupervised adaptive learning method, utilizes multimodal data, solves the difficult problem in the detection of abnormal working status of power equipment, and provides a new technical solution for power grid equipment status detection and diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid equipment fault diagnosis, and in particular to a method and system for detecting and diagnosing abnormalities of power grid equipment using adaptive learning. Background Art

[0002] The existing smart grid system realizes the data collection and integration of power grid equipment. At present, the analysis of the operating status of power grid equipment is increasingly using data-driven technical methods. However, the failure process of power grid equipment is usually very slow. When the equipment data changes after the abnormal state does not exceed the specified threshold, it will not be detected in time. Therefore, relying solely on equipment operation data for fault analysis and diagnosis has a certain lag. In addition, most of the current data-driven methods are supervised learning methods, which rely on accurate data labels and have certain requirements for data balance. In actual situations, abnormal data is relatively scarce, and accurate label data also requires resources to be verified. Therefore, supervised learning methods have certain defects.

[0003] Therefore, the present invention proposes an adaptive learning power grid equipment abnormality detection and diagnosis method and system to solve the above problems. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention develops an adaptive learning method and system for abnormality detection and diagnosis of power grid equipment to achieve abnormality detection and diagnosis of power grid equipment.

[0005] The technical solution to the technical problem of the present invention is as follows: On the one hand, the present invention provides an adaptive learning method for abnormal detection and diagnosis of power grid equipment, comprising the following steps:

[0006] S1. Data collection: Collect information about abnormalities in power grid equipment, including equipment operation sound data V, gas data F, power grid monitoring data B, and monitoring image data P, where the monitoring image data P is obtained by extracting monitoring video frame by frame;

[0007] S2. Data preprocessing: The sound data V is processed by Mel spectrum, the frequency domain features of the sound are extracted from the sound data V to form a logarithmic Mel spectrum, and then the logarithmic Mel spectrum is normalized by cepstral mean. The sound data obtained after processing is recorded as V m ;

[0008] S3. Constructing a dataset: Through the processed sound data V m , gas data F, power grid monitoring data B and monitoring image data P to construct sample data x, and arrange multiple sample data in chronological order to form a time series as a training set D;

[0009] S4. Multimodal data fusion: The image features P obtained after feature extraction of the monitoring image data P in the training set D feature Perform word segmentation operation to obtain image word segmentation P p ; The processed equipment operation sound data V m , gas data F, and power grid monitoring data B are spliced ​​to form numerical data O, and feature O is obtained after feature extraction of numerical data O feature Perform word segmentation operation to obtain word segmentation O p ; Segment the image into words P p and numerical data segmentation O p Combine and obtain the fused feature vector D f ;

[0010] S5. Construct Informer network: Design an Informer network, including an encoder module, a decoder module, and a score-guided module, to transform the feature vector D in S4 f As the input of the network constructed in S5, whether the data is abnormal is used as the output of the network;

[0011] S6. Design loss function: Use the distance ratio between the data and the cluster center of normal and abnormal data as labels, perform training to obtain scores, and then design a loss function based on the scores, input data, and decoded data;

[0012] S7. Train the network structure and evaluate it based on the training results: execute S6 in a loop. When the value of the loss function is less than the set threshold, the training of the network structure is completed.

[0013] In a specific implementation manner, S2 is as follows:

[0014] The formula for Mel spectrum processing is:

[0015] Among them, f V represents the frequency of sound data, lg(·) represents the logarithmic function, and M(·) is the corresponding Mel frequency;

[0016] Set the Hamming window length and step size of the logarithmic Mel spectrum, and perform cepstrum mean normalization on the logarithmic Mel spectrum. The processing formula is: (MM mean ) / M std , where M mean is the average value of the Mel spectrum, M std is the variance of the Mel spectrum, and the processed data is recorded as V m .

[0017] In a specific implementation manner, S3 is as follows:

[0018] The sample data is recorded as x = [V m,F,B,P], the time series is recorded as D={x1,x2,…,x n}, where n represents the length of the time series, x n represents sample data with time series n, and training set D is the data required for network structure training.

[0019] In a specific implementation manner, S4 is specifically as follows:

[0020] The monitoring image data P in the sample data set x is subjected to feature extraction. The monitoring image data P first passes through the three-dimensional convolutional network Conv3D+BN+ReLU module, and then passes through the convolutional network Conv2D+BN+ReLU module to obtain the image feature P feature , the image feature P feature The intermediate variable P of the image is obtained through the Flatten layer flat , and then pass through the softmax layer to get the image segmentation P p , the calculation formula is:

[0021] P flat =T(Flatten(P feature )),

[0022] P p =softmax.T(P flat ·W ap ) / ·P flat ·W bp ,

[0023] Where T(·) represents the transposition function, W bp Represents the image segmentation weight parameter;

[0024] After processing, the sound data V m , gas data F and monitoring data M constitute numerical data O, and the numerical data O is passed through the two-dimensional convolutional network Conv2D+BN+ReLU module to obtain the feature O of the numerical data feature , the feature O of the numerical data feature Obtain the intermediate variable O of numerical data through the Flatten layer flat , and then pass through the softmax layer to obtain the numerical data segmentation O p , the calculation formula is:

[0025] O flat =T(Flatten(O feature )),

[0026] O p =softmax(T(O flat ·W ao))·O flat ·W bo ,

[0027] Where T(·) represents the transposition function, W bo Represents the weight parameter of numerical data segmentation;

[0028] Finally, the image segmentation P is directly concatenated. p and numerical data segmentation O p The fused feature vector is recorded as D f =[P p …O p ] T , T represents transpose.

[0029] In a specific implementation, the Informer network structure in S5 is as follows:

[0030] The encoder module is composed of a multi-head sparse self-attention layer, a normalization layer, a feedforward neural network layer and a normalization layer in sequence. The score-oriented module is composed of a four-layer fully connected network. The decoder module is composed of a multi-head sparse self-attention layer, a normalization layer, a multi-head sparse self-attention layer, a normalization layer, a feedforward neural network layer and a normalization layer in sequence. The feature vector D f Input is fed into the encoder, and the output is fed into both the decoder module and the score-guided module;

[0031] The output of the decoder module is passed through a fully connected neural network to obtain the output of the entire network structure.

[0032] In a specific implementation, S6 is as follows: the score-guided module performs supervised learning training, takes the data and the cluster center ratio of normal and abnormal data as labels, outputs the result as a score, and then designs and constructs a loss function based on the score, input data and decoded data:

[0033] The design of the loss function is:

[0034]

[0035]

[0036] Among them, L SG represents the loss function of the score-oriented module, N represents the number of samples, s represents the score result obtained by the score-oriented module, and x′ i represents the data reconstructed by the i-th decoder, ||·||2 represents the 2-norm, ∈ represents a threshold related to the ratio of normal and abnormal data, and λ SG , a represents the weight hyperparameter, a represents the ratio between the number of samples of normal and abnormal data, and xi represents the i-th data, s i Represents the scoring result given by the score-oriented module corresponding to the i-th data.

[0037] On the other hand, the present invention also provides an adaptive learning power grid equipment anomaly detection and diagnosis system, which implements an adaptive learning power grid equipment anomaly detection and diagnosis method, including the following modules:

[0038] Data collection module: collects information about abnormalities in power grid equipment, including equipment operation sound data V, gas data F, monitoring data itself B, and monitoring image data P;

[0039] Data preprocessing module: Process the sound data V through the Mel spectrum, extract the frequency domain features of the sound from the sound data V to form a logarithmic Mel spectrum, and then perform Doppler mean normalization on the logarithmic Mel spectrum;

[0040] Dataset building module: processed sound data V m , gas data F, monitoring data itself M and monitoring image data P to construct a sample data set x, and arrange multiple sample data sets in chronological order to form a time series D;

[0041] Multimodal data fusion module: extract the features of the monitoring image data P in the data set to obtain the image features P feature Perform word segmentation operation to obtain image word segmentation P p ; Construct numerical data O, extract the features of numerical data O to obtain the features of numerical data O feature Perform word segmentation operation to obtain numerical data word segmentation O p ; Segment the image into words P p and numerical data segmentation O p Combine and obtain the fused feature vector D f ;

[0042] Construct network modules: Design a network including encoder module, decoder module and score-guided module;

[0043] Network structure training and result evaluation: The feature vector D f As the input of the network structure, the output of the model is used as the input of the loss function to train the network.

[0044] The effects provided in the content of the invention are only the effects of the embodiments, rather than all the effects of the invention. The above technical solution has the following advantages or beneficial effects:

[0045] The present invention proposes an adaptive learning method for abnormal detection and diagnosis of power grid equipment. By utilizing multimodal data, early acquisition of abnormal working status of power equipment is achieved, and the lag problem in fault detection is solved. Through an unsupervised adaptive learning method, the operation data of the equipment is continuously analyzed, and the detection model is automatically updated and optimized, thereby achieving abnormal detection and diagnosis of power grid equipment and avoiding resource consumption in data acquisition and data balance of supervised learning methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0047] Figure 1 It is a schematic diagram of the method flow of the present invention.

[0048] Figure 2 The figure is a schematic diagram of the structure of the invented method. DETAILED DESCRIPTION

[0049] In order to clearly illustrate the technical features of the present invention, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described below.

[0050] Example 1

[0051] like Figure 1 As shown, a method for detecting and diagnosing abnormalities in power grid equipment based on adaptive learning includes the following steps:

[0052] S1. Data collection: Collect information about abnormalities in power grid equipment, including equipment operation sound data V, gas data F, power grid monitoring data B, and monitoring image data P, where the monitoring image data P is obtained by extracting monitoring video frame by frame;

[0053] S2. Data preprocessing: The sound data V is processed by Mel spectrum, the frequency domain features of the sound are extracted from the sound data V to form a logarithmic Mel spectrum, and then the logarithmic Mel spectrum is normalized by cepstral mean. The sound data obtained after processing is recorded as V m ;

[0054] S3. Constructing a dataset: Through the processed sound data V m , gas data F, power grid monitoring data B and monitoring image data P to construct sample data x, and arrange multiple sample data in chronological order to form a time series as a training set D;

[0055] S4. Multimodal data fusion: The image features P obtained after feature extraction of the monitoring image data P in the training set D feature Perform word segmentation operation to obtain image word segmentation P p ; The processed equipment operation sound data V m , gas data F, and power grid monitoring data B are spliced ​​to form numerical data O, and feature O is obtained after feature extraction of numerical data O feature Perform word segmentation operation to obtain word segmentation O p ; Segment the image into words P p and numerical data segmentation O p Combine and obtain the fused feature vector D f ;

[0056] S5. Construct Informer network: Design an Informer network, including an encoder module, a decoder module, and a score-guided module, to transform the feature vector D in S4 f As the input of the network constructed in S5, whether the data is abnormal is used as the output of the network;

[0057] S6. Design loss function: Use the distance ratio between the data and the cluster center of normal and abnormal data as labels, perform training to obtain scores, and then design a loss function based on the scores, input data, and decoded data;

[0058] S7. Train the network structure and evaluate it based on the training results: execute S6 in a loop. When the value of the loss function is less than the set threshold, the training of the network structure is completed.

[0059] In a specific implementation manner, S2 is as follows:

[0060] The formula for Mel spectrum processing is:

[0061] Among them, f V represents the frequency of sound data, lg(·) represents the logarithmic function, and M(·) is the corresponding Mel frequency;

[0062] Set the Hamming window length of the logarithmic Mel spectrum to 2048, the step length to 128, and perform cepstrum mean normalization on the logarithmic Mel spectrum. The processing formula is: (MM mean ) / M std , where M mean is the average value of the Mel spectrum, M std is the variance of the Mel spectrum, and the processed data is recorded as V m .

[0063] In a specific implementation manner, S3 is as follows:

[0064] The sample data is recorded as x = [Vm ,F,B,P], the time series is recorded as D={x1,x2,…,x n}, where n represents the length of the time series, x n represents sample data with time series n, and training set D is the data required for network structure training.

[0065] In a specific implementation manner, S4 is specifically as follows:

[0066] The monitoring image data P in the sample data set x is subjected to feature extraction. The monitoring image data P first passes through the three-dimensional convolutional network Conv3D+BN+ReLU module, and then passes through the convolutional network Conv2D+BN+ReLU module to obtain the image feature P feature , the image feature P feature The intermediate variable P of the image is obtained through the Flatten layer flat , and then pass through the softmax layer to get the image segmentation P p , the calculation formula is:

[0067] P flat =T(Flatten(P feature )),

[0068] P p =softmax.T(P flat ·W ap ) / ·P flat ·W bp ,

[0069] Where T(·) represents the transposition function, W bp Represents the image segmentation weight parameter;

[0070] After processing, the sound data V m , gas data F and monitoring data M constitute numerical data O, and the numerical data O is passed through the two-dimensional convolutional network Conv2D+BN+ReLU module to obtain the feature O of the numerical data feature , the feature O of the numerical data feature Obtain the intermediate variable O of numerical data through the Flatten layer flat , and then pass through the softmax layer to obtain the numerical data segmentation O p , the calculation formula is:

[0071] O flat =T(Flatten(O feature )),

[0072] O p =softmax(T(O flat ·Wao ))·O flat ·W bo ,

[0073] Where T(·) represents the transposition function, W bo Represents the weight parameter of numerical data segmentation;

[0074] Finally, the image segmentation P is directly concatenated. p and numerical data segmentation O p The fused feature vector is recorded as D f =[P p …O p ] T , T represents transpose.

[0075] In a specific implementation, the Informer network structure in S5 is as follows:

[0076] The encoder module is composed of a multi-head sparse self-attention layer, a normalization layer, a feedforward neural network layer and a normalization layer in sequence. The score-oriented module is composed of a four-layer fully connected network. The decoder module is composed of a multi-head sparse self-attention layer, a normalization layer, a multi-head sparse self-attention layer, a normalization layer, a feedforward neural network layer and a normalization layer in sequence. The feature vector D f Input is fed into the encoder, and the output is fed into both the decoder module and the score-guided module;

[0077] The output of the decoder module is passed through a fully connected neural network to obtain the output of the entire network structure.

[0078] In a specific implementation, S6 is as follows: the score-guided module performs supervised learning training, takes the data and the cluster center ratio of normal and abnormal data as labels, outputs the result as a score, and then designs and constructs a loss function based on the score, input data and decoded data:

[0079] The design of the loss function is:

[0080]

[0081]

[0082] Among them, L SG represents the loss function of the score-oriented module, N represents the number of samples, s represents the score result obtained by the score-oriented module, and x′ i represents the data reconstructed by the i-th decoder, ||·||2 represents the 2-norm, ∈ represents a threshold related to the ratio of normal and abnormal data, and λ SG , arepresents the weight hyperparameter, a represents the ratio between the number of samples of normal and abnormal data, and x i represents the i-th data, s i Represents the scoring result given by the score-oriented module corresponding to the i-th data.

[0083] like Figure 2 As shown above, it can be seen that in the figure, x represents sample data, L represents the loss function, s represents the evaluation result obtained by the score-guided module, and x′ represents the data reconstructed by the decoder.

[0084] Example 2

[0085] An adaptive learning power grid equipment anomaly detection and diagnosis system implements an adaptive learning power grid equipment anomaly detection and diagnosis method, including the following modules:

[0086] Data collection module: collects information about abnormalities in power grid equipment, including equipment operation sound data V, gas data F, monitoring data itself B, and monitoring image data P;

[0087] Data preprocessing module: Process the sound data V through the Mel spectrum, extract the frequency domain features of the sound from the sound data V to form a logarithmic Mel spectrum, and then perform Doppler mean normalization on the logarithmic Mel spectrum;

[0088] Dataset building module: processed sound data V m , gas data F, monitoring data itself M and monitoring image data P to construct a sample data set x, and arrange multiple sample data sets in chronological order to form a time series D;

[0089] Multimodal data fusion module: extract the features of the monitoring image data P in the data set to obtain the image features P feature Perform word segmentation operation to obtain image word segmentation P p ; Construct numerical data O, extract the features of numerical data O to obtain the features of numerical data O feature Perform word segmentation operation to obtain numerical data word segmentation O p ; Segment the image into words P p and numerical data segmentation O p Combine and obtain the fused feature vector D f ;

[0090] Construct network modules: Design a network including encoder module, decoder module and score-guided module;

[0091] Network structure training and result evaluation: The feature vector D f As the input of the network structure, the output of the model is used as the input of the loss function to train the network.

[0092] Example 3

[0093] From Table 1, it can be seen that the comparison between the method in the present invention and other methods in terms of accuracy and recall rate shows that the accuracy and recall rate of the present invention are higher than those of other methods. It can be seen that the method in the present invention can better detect and diagnose abnormalities of power grid equipment. The method in the present invention uses an unsupervised adaptive learning method to continuously analyze the operation data of the equipment, automatically update and optimize the detection model, thereby realizing abnormal detection and diagnosis of power grid equipment and avoiding resource consumption of supervised learning methods in data acquisition and data balance.

[0094] Table 1. Comparison of the detection results of the method in the present invention and other methods

[0095] Technology Name Accuracy Recall Data labeling based method 94％ 90％ Graph Neural Networks 68％ 79％ Statistical Hierarchical Clustering Algorithm 85％ 78％ Enhanced nonparametric sequence learning algorithm based on ensemble clustering technology 78％ 65％ This method 96％ 90％

[0096] Although the above describes the specific implementation mode of the invention in conjunction with the drawings, it is not intended to limit the scope of protection of the invention. Based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims

1. An adaptive learning method for abnormality detection and diagnosis of power grid equipment, characterized by: The following steps are involved: S1. Data collection: Collect information about abnormalities in power grid equipment, including equipment operation sound data V, gas data F, and monitoring image data P, where the monitoring image data P is obtained by extracting monitoring video frame by frame; S2. Data preprocessing: The sound data V is processed by Mel spectrum, the frequency domain features of the sound are extracted from the sound data V to form a logarithmic Mel spectrum, and then the logarithmic Mel spectrum is normalized by cepstral mean. The sound data obtained after processing is recorded as ; S3. Constructing the dataset: through the processed sound data , gas data F and monitoring image data P construct sample data x, the multiple sample data are arranged in chronological order to form a time series as a training set D; S4. Multimodal data fusion: Image features obtained after feature extraction of monitoring image data P in training set D Perform word segmentation operation to obtain image segmentation ; The processed equipment operation sound data V m The features are obtained by combining the gas data F and numerical data O. Perform word segmentation operation to obtain word segmentation ; Segment the image into words and numeric data segmentation Combine to obtain the fused feature vector ; S4 is as follows: The monitoring image data P in the sample data set x is subjected to feature extraction. The monitoring image data P first passes through the three-dimensional convolutional network Conv3D+BN+ReLU module, and then passes through the two-dimensional convolutional network Conv2D+BN+ReLU module to obtain image features. , the image features Get the intermediate variable of the image through the Flatten layer , after calculation, the image segmentation is obtained through the softmax layer , the calculation formula is: ， ， in, represents the transpose function, and Represents two different weight parameters in image segmentation; After processing the sound data The gas data F and the numerical data O are combined to obtain the features of the numerical data through the two-dimensional convolutional network Conv2D+BN+ReLU module. , the features of numerical data Get the intermediate variable of numerical data through the Flatten layer , after calculation, the softmax layer is used to obtain the numerical data segmentation , the calculation formula is: ， ， in, represents the transpose function, and Represents two different weight parameters in the segmentation of numerical data; Finally, the image is segmented by direct concatenation and numeric data segmentation The fused feature vector is recorded as , T represents transpose; S5. Construct Informer network: Design an Informer network, including encoder module, decoder module and score-guided module, to transform the feature vector in S4 As the input of the network constructed in S5, whether the data is abnormal is used as the output of the network; The Informer network structure is as follows: The encoder module is composed of a multi-head sparse self-attention layer, a normalization layer, a feedforward neural network layer and a normalization layer in sequence. The score-guided module is composed of a four-layer fully connected network. The decoder module is composed of a multi-head sparse self-attention layer, a normalization layer, a multi-head sparse self-attention layer, a normalization layer, a feedforward neural network layer and a normalization layer in sequence. The feature vector Input is fed into the encoder, and the output is fed into both the decoder module and the score-guided module; The output of the decoder module is passed through a fully connected neural network to obtain the output of the entire network structure; S6. Design loss function: Use the distance ratio between the data and the cluster center of normal and abnormal data as labels, perform training to obtain scores, and then design a loss function based on the scores, input data, and decoded data; S6 is as follows: The score-oriented module performs supervised learning training, takes the data and the cluster center ratio of normal and abnormal data as labels, outputs the score, and then designs and constructs the loss function based on the score, input data and decoded data: The design of the loss function is: ， ， in, represents the loss function of the score-oriented module, N represents the number of samples, s represents the scoring result obtained by the score-oriented module, represents the data reconstructed by the i-th decoder, represents the 2-norm, Represents a threshold value related to the ratio of normal and abnormal data. represents the weight hyperparameter, a represents the ratio between the number of samples of normal and abnormal data, represents the i-th data, represents the scoring result given by the score-oriented module corresponding to the i-th data; S7. Train the network structure and evaluate it based on the training results: execute S6 in a loop. When the value of the loss function is less than the set threshold, the training of the network structure is completed.

2. The method for detecting and diagnosing abnormalities of power grid equipment using adaptive learning according to claim 1, characterized in that: S2 is as follows: The formula for Mel spectrum processing is: , in, Indicates the frequency of sound data, represents the logarithmic function, is the corresponding Mel frequency; Set the Hamming window length and step size of the logarithmic Mel spectrogram, and perform cepstrum mean normalization on the logarithmic Mel spectrogram. The processing formula is: , where is the average value of the Mel spectrum, is the variance of the Mel spectrum, and the processed data is recorded as .

3. The method for detecting and diagnosing abnormalities of power grid equipment using adaptive learning according to claim 2, wherein S3 The details are as follows: The sample data is recorded as , the time series is recorded as , where n represents the length of the time series, represents sample data with time series n, and training set D is the data required for network structure training.

4. An adaptive learning power grid equipment anomaly detection and diagnosis system, which executes an adaptive learning power grid equipment anomaly detection and diagnosis method according to any one of claims 1 to 3, characterized in that: Includes the following modules: Data acquisition module: collects information about abnormalities in power grid equipment, including equipment operation sound data V, gas data F, and monitoring image data P; Data preprocessing module: Process the sound data V through the Mel spectrum, extract the frequency domain features of the sound from the sound data V to form a logarithmic Mel spectrum, and then perform Doppler mean normalization on the logarithmic Mel spectrum; Dataset building blocks: processed sound data , gas data F and monitoring image data P construct a sample data set x, the multiple sample data sets are arranged in chronological order to form a time series D; Multimodal data fusion module: image features obtained after feature extraction of monitoring image data P in the data set Perform word segmentation operation to obtain image segmentation ; Construct numerical data O, and extract the features of the numerical data obtained after extracting the features of the numerical data O Perform word segmentation to obtain numeric data segmentation ; Segment the image into words and numeric data segmentation Combine to obtain the fused feature vector ; Construct network modules: Design a network including encoder module, decoder module and score-guided module; Network structure training and result evaluation: The feature vector As the input of the network structure, the output of the model is used as the input of the loss function to train the network.

Images