Power system fault waveform identification method based on self-supervised representation learning

Abstract

The invention relates to the technical field of deep learning, and provides a power system fault waveform recognition method based on self-supervised representation learning. Comprising the following steps: S1, constructing a positive and negative relation pair data set and an internal time relation data set based on sampling data; s2, performing self-supervised training on the network based on the positive and negative relation pair and the relation of the internal time sequence; s3, training an identification network by using a small amount of data with labels, and combining the identification network with the network trained in the step S2 to realize data classification; according to the method, a self-supervised learning mode is adopted during training, more underlying features of power system fault data can be learned, and real data instead of theoretical simulation data is adopted during training, so that the scheme has better recognition precision; in addition, the convenience brought by self-supervised learning that the fault data does not need to be marked in a large amount is also the huge advantage of the scheme provided by the invention.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and proposes a method for identifying fault waveforms in a power system based on self-supervised representation learning. Background Art

[0002] The transmission and distribution network is responsible for supplying power to users, and has a huge impact on power supply reliability. If a failure occurs, it will have a serious impact on production and life. Therefore, when a failure occurs in the transmission and distribution network, it is crucial to accurately and quickly locate the fault. At present, the recognition of power system fault waveforms is mainly based on machine learning methods, and there are few related studies on deep learning.

[0003] The paper "Study on the Identification of Lightning Waveforms of AC Transmission Lines Based on Time-Frequency Analysis and Deep Learning" proposes an identification scheme based on the combination of time-frequency analysis algorithm and deep learning. This scheme selects three time-frequency analysis algorithms, short-time Fourier transform, wavelet transform, and Hilbert-Huang transform, to perform time-frequency analysis on lightning simulation waveforms on matlab software to obtain the corresponding time-frequency spectrum data set. Two types of deep learning networks, Resnet18 and LSTM, were selected for classification experiments. The three types of time-frequency spectrum data sets were used as the input of the model. The loss function was improved based on the actual engineering characteristics of the lightning waveform, thereby optimizing the model. Finally, the degree of decrease in the loss function and the accuracy of the verification data set were used as criteria to analyze and compare the experimental results of different time-frequency algorithms combined with deep learning networks. Finally, the combination of wavelet transform and Resnet18 network was proposed as the main model for lightning waveform recognition, and the modified model was applied to the recognition of fault waveforms.

[0004] The above solution has the following advantages:

[0005] 1. Short-time Fourier transform, wavelet transform and Hilbert-Huang transform are introduced to perform time-frequency analysis on the simulated lightning waveform data, and different types of time-frequency data sets containing various features are obtained. Compared with the traditional Fourier transform, it has fewer limitations in analyzing non-stationary mutation signals.

[0006] 2. A method combining time-frequency analysis algorithm and Resnet model is proposed and applied to lightning waveform recognition, which has good recognition accuracy.

[0007] However, this solution also has some disadvantages and shortcomings:

[0008] 1. The research of this scheme is based on simulation. The simulated waveform has no complex environmental interference and the source is relatively clean. It is somewhat different from the real lightning waveform. Therefore, the performance of this scheme in actual scene application still needs to be investigated.

[0009] 2. This solution adopts the supervised learning training method. The training sample set must consist of labeled samples. In practical applications, the process of labeling all sample data may take a lot of time and effort. Summary of the invention

[0010] The purpose of the present invention is to provide a method for identifying power system fault waveforms based on self-supervised characterization learning in view of the deficiencies in the prior art.

[0011] The technical solution is as follows:

[0012] A method for identifying power system fault waveforms based on self-supervised representation learning includes data sampling and the following steps:

[0013] S1. Based on the sampled data, construct a positive and negative relationship pair dataset and an internal time relationship dataset;

[0014] S2, self-supervised training of the network based on the relationship between positive and negative relationship pairs and internal time series;

[0015] S3. Use a small amount of labeled data to train a recognition network, and combine the recognition network with the network trained in step S2 to achieve data classification.

[0016] Furthermore, the network includes a feature extraction network, an inter-sample relationship reasoning network, and an internal time relationship reasoning network.

[0017] Furthermore, step S3 is to combine the feature extraction network with the recognition network.

[0018] Furthermore, the positive and negative sample pair data set consists of positive relationship pairs and negative relationship pairs, the positive relationship pairs are two time series sampled from the same expansion set, and the negative relationship pairs are two time series sampled from different expansion sets, and the relationship between them is used as a label, the positive relationship is 1, and the negative relationship is 0.

[0019] Furthermore, the internal time relationship dataset is composed of different time series, and the time relationships are assigned according to their time distances. In the training task, three types of time relationships, short-term, medium-term and long-term, are specified, corresponding to different relationship labels respectively.

[0020] Furthermore, the feature extraction network is a convolutional neural network; the inter-sample relationship reasoning network and the internal time relationship reasoning network are fully connected neural networks.

[0021] Furthermore, the following time relationship sampling algorithm is also included:

[0022] S11. Data collection: Collect historical data, set t n is a time series of length T, pn,u and p n,v They are sampled from t n Two time series of length L, C is the number of time relationship categories;

[0023] S12. Define a tag set: in Yes n,u and p n,v Labels for temporal relationships;

[0024] S13. Define time distance d u,v And the distance threshold D: d u,v =|uv|, where u and v are p n,u and p n,v The starting point number;

[0025] S14, calculation label: if

[0026] like

[0027] …………

[0028] like

[0029] otherwise

[0030] After executing the above steps, the calculation of the label is completed.

[0031] Furthermore, the following self-supervised training algorithms are also included:

[0032] S21. Construct a convolutional neural network as a feature extraction network f θ , two fully connected neural networks, respectively serving as the sample relationship reasoning network γ μ and internal temporal relation reasoning network γ φ ;

[0033] S22, for t m and t n , by transforming the amplitude and time domain, two sets of enhanced sets A(t m ) and A(t n ), sample from two enhanced sets and construct positive relationship pairs and negative relationship

[0034] S23, taking the positive and negative relationship data set as input, using the network f θ Perform feature extraction:

[0035]

[0036]

[0037]

[0038] S24, using the extracted features as input, using the relationship between samples to infer the network γ μ Perform relational reasoning:

[0039]

[0040]

[0041] Binary cross entropy loss L inter The definition is as follows:

[0042]

[0043] in Indicates a positive relationship, Indicates a negative relationship.

[0044] S25, construct time segment relationship pairs according to the time relationship sampling algorithm (p n,u , p n,v )

[0045] S26, taking the internal time relationship dataset as input, using the network f θ Perform feature extraction:

[0046] z n,u =f θ (p n,u )

[0047] z n,v =f θ (p n,v )

[0048] S27, using the extracted features as input, using the internal temporal relationship reasoning network γ φ Perform relational reasoning:

[0049]

[0050] Cross entropy loss L intra The definition is as follows:

[0051]

[0052] Furthermore, the self-supervised training algorithm also includes:

[0053] S28, training loss is defined as follows:

[0054] L=Linter +L intra

[0055] With the goal of reducing the loss function L, continuously update f θ , γ μ , γ φ until the training is finished.

[0056] Furthermore, the step S3 specifically includes:

[0057] S31. After self-supervised training is completed, the relational reasoning network γ is discarded μ and γ φ , only retain the feature extraction network f θ ;

[0058] S32, build a fully connected neural network γ as a data recognition network, train the recognition network with a small number of labeled samples, and after the training is completed, the feature extraction network f θ Combined with recognition network γ, the classification of power data is realized.

[0059] The beneficial effects of the present invention are:

[0060] The present invention adopts a self-supervised learning method during training, which can learn more underlying features of power system fault data, and uses real data rather than theoretical simulation data during training, so the scheme has better recognition accuracy; in addition, the convenience brought by self-supervised learning that there is no need to mark a large amount of fault data is also a huge advantage of the scheme proposed by the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a flowchart of self-supervised representation learning;

[0062] Figure 2 It is the workflow diagram of the power data classification network. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.

[0064] Embodiment 1:

[0065] A method for identifying fault waveforms in a power system based on self-supervised representation learning, comprising:

[0066] S1. Sample the data to construct a positive-negative relationship pair dataset and an internal time relationship dataset. The positive-negative sample pair dataset consists of positive relationship pairs and negative relationship pairs. The positive relationship pairs are two time series sampled from the same expansion set, and the negative relationship pairs are two time series sampled from different expansion sets. The relationship between them is used as a label, with a positive relationship of 1 and a negative relationship of 0.

[0067] The internal temporal relation dataset consists of different time series. Temporal relations are assigned according to their temporal distance. In the training task, three types of temporal relations are specified: short-term, medium-term, and long-term, corresponding to different relation labels.

[0068] S2, based on the positive and negative relationship pairs and the relationship between the internal time series, the network is self-supervised and trained. The network includes a feature extraction network, an inter-sample relationship reasoning network, and an internal time relationship reasoning network. The goal is to reduce the multi-task loss function L, and the parameters of the three networks are continuously updated until the training is completed;

[0069] S3. Only retain the feature extraction network trained in the previous step. At this time, the feature extraction network has been trained and has learned some effective representations of the time series, which can be directly used for subsequent feature extraction tasks. Use a small amount of labeled data to train a fully connected neural network as a power data recognition network. After training, combine the feature extraction network with the recognition network to achieve classification of power data.

[0070] Embodiment 2:

[0071] Specifically, this embodiment also includes the following steps:

[0072] The network is made to learn some effective representations of time series by using self-supervised learning, so that the recognition network can still achieve high classification accuracy without the need for a large amount of labeled data for supervised training in the later stage. The algorithms used include time relationship sampling algorithm and self-supervised training algorithm.

[0073] (1) The temporal relationship sampling algorithm includes the following steps S11-S17:

[0074] S11. Data collection: Collect historical data, set t n is a time series of length T, p n,u and p n,v They are sampled from t n Two time series of length L, C is the number of time relationship categories;

[0075] S12. Define a tag set: in Yes n,u and p n,v Labels for temporal relationships;

[0076] S13. Define time distance d u,v And the distance threshold D: d u,v =|uv|, where u and v are p n,u and p n,v The starting point number;

[0077] S14, calculation label: if

[0078] like

[0079] …………

[0080] like

[0081] otherwise

[0082] After executing the above steps, the calculation of the label is completed.

[0083] (2) The self-supervised training algorithm includes the following steps S21-S27:

[0084] S21. Construct a 4-layer convolutional neural network as the feature extraction network f θ , two fully connected neural networks, respectively serving as the sample relationship reasoning network γ μ and the internal temporal relation reasoning network γφ;

[0085] S22, for t m and t n , by transforming the amplitude and time domain, two sets of enhanced sets A(t m ) and A(t n ), sample from two enhanced sets and construct positive relationship pairs and negative relationship

[0086] S23, taking the positive and negative relationship data set as input, using the network f θ Perform feature extraction:

[0087]

[0088]

[0089]

[0090] S24, using the extracted features as input, using the relationship between samples to infer the network γ μ Perform relational reasoning:

[0091]

[0092]

[0093] Binary cross entropy loss L inter The definition is as follows:

[0094]

[0095] in Indicates a positive relationship, Indicates a negative relationship.

[0096] S25, constructing a time segment relationship pair (p n,u , p n,v )

[0097] S26, taking the internal time relationship dataset as input, using the network f θ Perform feature extraction:

[0098] z n,u =f θ (p n,u )

[0099] z n,v =f θ (p n,v )

[0100] S27, using the extracted features as input, using the internal temporal relationship reasoning network γ φ Perform relational reasoning:

[0101]

[0102] Cross entropy loss L intra The definition is as follows:

[0103]

[0104] S28, training loss is defined as follows:

[0105] L=L inter +L intra

[0106] With the goal of reducing the loss function L, continuously update f θ , γ μ , γ φ until the training is finished.

[0107] Embodiment three:

[0108] Furthermore, this embodiment also includes the following steps:

[0109] Use neural network to identify fault data. In the feature extraction step, the trained network f can be directly used. θ ,The core idea is as follows,;

[0110] S31. After self-supervised training is completed, we discard the relational reasoning network γ μ and γ φ , only retain the feature extraction network f θ At this time, the feature extraction network has been trained and has learned some effective representations of the time series, which can be directly used in subsequent feature extraction tasks.

[0111] S32, construct a fully connected neural network γ as the power data recognition network, and train the recognition network with a small number of labeled samples. After the training is completed, the feature extraction network f θ Combined with the identification network γ, the classification of power data can be achieved.

[0112] The present invention proposes a general self-supervised time series representation learning framework, which can learn the underlying structural features of unlabeled time series by exploring the relationship between samples and internal time relationships of time series. After self-supervised training, the trained feature extraction network is directly used for subsequent feature extraction work. Only one recognition network γ needs to be trained, and the two networks can be combined to achieve the classification of power data. Since a self-supervised learning method is used during training, more underlying features of power system fault data can be learned, and real data rather than theoretical simulation data is used during training, so the scheme will have better recognition accuracy. In addition, the convenience brought by self-supervised learning that there is no need to mark a large amount of fault data is also a huge advantage of the scheme proposed by the present invention.

[0113] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying fault waveforms in a power system based on self-supervised representation learning, comprising data sampling, characterized in that: The following steps are also included: S1. Based on the sampled data, construct a positive and negative relationship pair dataset and an internal time relationship dataset; S2, self-supervised training of the network based on the relationship between positive and negative relationship pairs and internal time series; S3. Use a small amount of labeled data to train a recognition network, and combine the recognition network with the network trained in step S2 to achieve data classification.

2. A method for identifying power system fault waveforms based on self-supervised representation learning according to claim 1, characterized in that: The network includes a feature extraction network, an inter-sample relationship reasoning network and an internal time relationship reasoning network.

3. A method for identifying power system fault waveforms based on self-supervised representation learning according to claim 2, characterized in that: The step S3 is to combine the feature extraction network with the recognition network.

4. A method for identifying power system fault waveforms based on self-supervised representation learning according to claim 3, characterized in that: The positive and negative sample pair data set consists of positive relationship pairs and negative relationship pairs. The positive relationship pairs are two time series sampled from the same expansion set, and the negative relationship pairs are two time series sampled from different expansion sets. The relationship between them is used as a label, with a positive relationship of 1 and a negative relationship of 0.

5. The method for identifying power system fault waveforms based on self-supervised representation learning according to claim 3 is characterized in that: The internal time relationship dataset consists of different time series, and the time relationships are assigned according to their time distances. In the training task, three types of time relationships, short-term, medium-term and long-term, are specified, corresponding to different relationship labels.

6. A method for identifying power system fault waveforms based on self-supervised representation learning according to claim 4 or 5, characterized in that: The feature extraction network is a convolutional neural network; the inter-sample relationship reasoning network and the internal time relationship reasoning network are fully connected neural networks.

7. A method for identifying power system fault waveforms based on self-supervised representation learning according to claim 6, characterized in that: It also includes the following time relationship sampling algorithms: S11. Data collection: Collect historical data, set t n is a time series of length T, p n,u and p n,v They are sampled from t n Two time series of length L, C is the number of time relationship categories; S12. Define a tag set: in Yes n,u and p n,v Labels for temporal relationships; S13. Define time distance d u,v And the distance threshold D: d u,v =|uv|, where u and v are p n,u and p n,v The starting point number; S14, calculation label: if d u,v <D, If D <d u,v <2D, …………… If (C-2)D <d u,v <(C-1)D, otherwise After executing the above steps, the calculation of the label is completed.

8. A method for identifying fault waveforms in a power system based on self-supervised representation learning according to claim 7, characterized in that: Also included are the following self-supervised training algorithms: S21. Construct a convolutional neural network as a feature extraction network f θ , two fully connected neural networks, respectively serving as the sample relationship reasoning network γ μ and internal temporal relation reasoning network γ φ ; S22, for t m and t n , by transforming the amplitude and time domain, two sets of enhanced sets A(t m ) and A(t n ), sample from two enhanced sets and construct positive relationship pairs and negative relationship S23, taking the positive and negative relationship data set as input, using the network f θ Perform feature extraction: S24, using the extracted features as input, using the relationship between samples to infer the network γ μ Perform relational reasoning: Binary cross entropy loss L inter The definition is as follows: in Indicates a positive relationship, Indicates a negative relationship. S25, constructing a time segment relationship pair (p n,u , p n,v ) S26, taking the internal time relationship dataset as input, using the network f θ Perform feature extraction: z n,u =f θ (p n,u z n,v =f θ (p n,v ) S27, using the extracted features as input, using the internal temporal relationship reasoning network γ φ Perform relational reasoning: Cross entropy loss L intra The definition is as follows:

9. A method for identifying fault waveforms in a power system based on self-supervised representation learning according to claim 8, characterized in that: The self-supervised training algorithm also includes: S28, training loss is defined as follows: L=L inter +L intra With the goal of reducing the loss function L, continuously update f θ , γ μ , γ φ until the training is finished.

10. A method for identifying fault waveforms in a power system based on self-supervised representation learning according to claim 9, characterized in that: The step S3 specifically includes: S31. After self-supervised training is completed, the relational reasoning network γ is discarded μ and γ φ , only retain the feature extraction network f θ ; S32, build a fully connected neural network γ as a data recognition network, train the recognition network with a small number of labeled samples, and after the training is completed, the feature extraction network f θ Combined with recognition network γ, the classification of power data is realized.

Images