Substation equipment fault diagnosis method based on integrated deep generative model

Abstract

The present invention discloses a substation equipment fault diagnosis method based on an integrated deep generative model, the method comprising: collecting substation equipment operating data, performing data preprocessing, constructing a sample training data set and a random potential data set; constructing an AMBI-GAN integrated deep generative model based on a bidirectional long short-term memory network and an attention mechanism; training the AMBI-GAN integrated deep generative model; inputting test data into the trained model, and calculating the total loss L of the model. test ; by 1‑L test An identification score is obtained, and if the identification score exceeds a preset value, it is judged to be abnormal; the advantage of the present invention is that: a small amount of labeled data is fully utilized to further improve the fault diagnosis performance.

Description

Technical Field

[0001] The present invention relates to the field of substation equipment fault diagnosis, and more specifically to a substation equipment fault diagnosis method based on an integrated deep generative model. Background Art

[0002] The development and improvement of substation construction has led to higher requirements for data processing. The collection of various data with sensing and monitoring capabilities and the detection of faults are key research issues. Substation equipment, as a crucial component of substations, generates massive amounts of heterogeneous power time-series data during operation. Real-time fault detection and fault cause analysis of this data will help improve UHV design, construction, operation, and maintenance, ensuring substation safety and stability.

[0003] Research is currently underway within the industry to use various deep learning models to detect anomalies in power time series data. These models rely on multidimensional network structures to exploit spatiotemporal distribution characteristics. Considering the spatiotemporal characteristics of time series, some current models, such as Hidden Markov Model (HMM) and Local Anomaly Factor (LAF), have achieved high detection accuracy.

[0004] Most existing time series anomaly detection research is limited by the randomness and complexity of power data. These issues can be summarized as follows: 1. Existing methods struggle to effectively extract temporal information from power data, limiting their ability to represent data features. The lack of integration between temporal and feature relationships limits the accuracy of the model. 2. Existing symbolic detection methods, such as Chinese Patent Publication No. CN114218998A, which discloses a method for analyzing abnormal behavior in power systems based on a hidden Markov model, do not extract features before symbolization, thus neglecting valid data information. Long symbol sequences can be noisy and affect computational efficiency. Data noise caused by voltage and current fluctuations can easily be detected as abnormal power operation samples, resulting in reduced anomaly detection performance. Existing similarity-based detection methods, such as the local anomaly factor method, are suitable for cases where data points are highly clustered and outliers are rare. Furthermore, because similarity algorithms typically require separate calculations for each data point, they are often computationally intensive and unsuitable for large, high-dimensional power data. 3. Accurately labeling power data requires significant labor, and in practice, insufficient labeled data is often a problem. Therefore, how to make full use of a small amount of labeled data to further improve model performance is a key issue in the anomaly detection task. Summary of the Invention

[0005] The technical problem to be solved by the present invention is how to make full use of a small amount of labeled data to further improve fault diagnosis performance.

[0006] The present invention solves the above technical problems through the following technical means: a substation equipment fault diagnosis method based on an integrated deep generative model, the method comprising:

[0007] Step 1: Collect substation equipment operation data, perform data preprocessing, and construct a sample training dataset and a random potential dataset;

[0008] Step 2: Build an AMBI-GAN integrated deep generative model based on bidirectional long short-term memory network and attention mechanism;

[0009] Step 3: Train the AMBI-GAN integrated deep generative model;

[0010] Step 4: Input test data into the trained model and calculate the total loss L of the model test ;

[0011] Step 5: Pass 1-L test Get the identification score, and if the identification score exceeds the preset value, it is judged to be abnormal.

[0012] The present invention has made corresponding improvements to the GAN model and proposed a generative adversarial network integrated deep generative model AMBI-GAN based on a bidirectional long short-term memory network and an attention mechanism. The model is constructed based on a bidirectional long short-term memory network, which can capture the dependency of time series, effectively extract the temporal characteristics of power data, and improve the accuracy of the model. Moreover, based on the inherent performance of the GAN model, it can generate a large amount of simulated data based on limited real power data, thereby making full use of a small amount of labeled data, with small computational complexity, small amount of original data, and low labor cost. In addition, through training, the influence of noise on data can be reduced, thereby realizing the anomaly detection of power time series, improving the detection accuracy of multidimensional time series anomalies, and further improving the fault diagnosis performance.

[0013] Furthermore, the data preprocessing in step 1 includes:

[0014] The input data is divided into a sample training data set and a random potential data set. All data in the sample training data set are normal, and there are abnormal data in the random potential data set. A sliding time window of length n is used to divide all data into several subsequences of equal length.

[0015] Furthermore, the step 2 includes:

[0016] Step 2.1: Build the Bi-LSTM structure.

[0017] Step 2.2: Add the attention mechanism to the Bi-LSTM structure to obtain the AMBi-LSTM structure.

[0018] Step 2.3: Use the AMBi-LSTM structure to obtain the generator and discriminator respectively. The generator and discriminator together constitute the AMBI-GAN integrated deep generation model.

[0019] Furthermore, the step 2.1 includes:

[0020] Construct a bidirectional LSTM network as a Bi-LSTM structure, where the input of the bidirectional LSTM network is X = {x1, ...x n}, x n Represents the input at the nth moment;

[0021] The hidden layer state of the Bi-LSTM structure is

[0022]

[0023]

[0024] Among them, LSTM backward For backward LSTM, LSTM forward For the forward LSTM, x t represents the input at the tth moment, represents the hidden state of the backward LSTM at time t+1, Represents the hidden state of the forward LSTM at time t-1;

[0025] The hidden layer state output at n moments is

[0026]

[0027] Furthermore, the step 2.2 includes:

[0028] By formula Set the input weight of the Bi-LSTM structure, thereby adding the attention mechanism to the Bi-LSTM structure to obtain the AMBi-LSTM structure; where w i represents the weight parameter of the Bi-LSTM structure at the i-th moment, softmax is the normalization function, and Bi-LSTM(X,H) represents the output of the Bi-LSTM structure;

[0029] The output of the AMBi-LSTM structure is

[0030]

[0031] represents the transpose of the input matrix at the i-th moment, represents the transpose of the hidden state matrix at the i-th moment.

[0032] Furthermore, the step three includes:

[0033] Step 3.1: The input of the generator is a random potential data set, and the mean of the data distribution of the generator is calculated;

[0034] Step 3.2: Design the discriminator’s identification model.

[0035] Step 3.3: Set the objective function according to the identification model, and continuously adjust the network weight parameters and data distribution mean so that the mean of the generator's output samples is closer to the mean of the sample training data set. When the objective function iteration reaches Nash equilibrium, stop training.

[0036] Furthermore, the step 3.1 includes:

[0037] Generator input sample z={z1,…z n} is the data sequence in the random potential data set, and the mean of the output sample of the generator is

[0038]

[0039] Among them, z i is the i-th data input to the generator, p(·) is the probability density function of the output, and log(·) is the logarithmic function;

[0040] During the training process, the network weight parameter W is continuously optimized, and the output sample E(z) is continuously brought close to the mean of the sample training data set.

[0041] Furthermore, the step 3.2 includes:

[0042] When the input of the discriminator is a sample training dataset, the discriminator's identification model is

[0043] E x ~p data(x) log(D(x))

[0044] Among them, p data(x) represents the probability distribution of input data x, E x ~p data(x) Indicates that p data(x) The mean of the variable x of the distribution, D(x) represents the output of the discriminator for x;

[0045] When the input of the discriminator is the output of the generator G(z), the discriminator's identification model is

[0046] E x ~p z(z) log(1-D(G(z)))

[0047] Among them, p z(z)represents the probability distribution of random latent data z, E x ~p z(z) Indicates that p z(z) The mean of the distribution z, D(G(z)) represents the output of the discriminator for G(z).

[0048] Furthermore, the step 3.3 includes:

[0049] By formula

[0050] min G max D V(D,G)=E x ~p data(x) log(D(x))+E x ~p z(z) log(1-D(G(z)))

[0051] Construct the objective function;

[0052] The output G(z) of the generator is used as the input of the discriminator, which continuously adjusts the network's weight parameters and the mean of the data distribution so that the data generated by the generator can be identified as real data by the discriminator. That is, when the objective function iteration reaches Nash equilibrium, the training stops.

[0053] Furthermore, the step 4 includes:

[0054] The original sample test data is directly input into the trained AMBI-GAN integrated deep generation model, through the formula Calculate the identification loss value, where Represents the model subsequence N={N i ,i=1,2,…,n},x test,i Represents the i-th data in the original test data, Represents the output of the generator;

[0055] By formula Calculate the total loss, where λ represents an adjustable parameter, represents the reconstruction loss.

[0056] The advantages of the present invention are:

[0057] (1) The present invention makes corresponding improvements to the GAN model and proposes a generative adversarial network integrated deep generative model AMBI-GAN based on a bidirectional long short-term memory network and an attention mechanism. The model is constructed based on a bidirectional long short-term memory network, which can capture the dependency of time series, effectively extract the temporal characteristics of power data, and improve the accuracy of the model. Moreover, based on the inherent performance of the GAN model, a large amount of simulated data can be generated based on limited real power data, thereby making full use of a small amount of labeled data, with small computational complexity, small amount of original data, and low labor cost. In addition, the influence of noise on data can be reduced through training, thereby realizing the anomaly detection of power time series, improving the detection accuracy of multidimensional time series anomalies, and further improving the fault diagnosis performance.

[0058] (2) The present invention constructs an objective function, and the output G(z) of the generator is used as the input of the discriminator. The generator hopes to reduce the objective function value so that the data distribution it generates cannot be distinguished, and the output D(x) of the discriminator hopes to increase the objective function value so that it can efficiently distinguish the true and false categories of the data. By playing a game between the generator and the discriminator, the weight parameters of the network and the mean value of the data distribution are continuously adjusted, so that the data generated by the generator can be distinguished as real data by the discriminator, that is, the objective function iteration reaches Nash equilibrium, thereby realizing accurate identification of abnormal data.

[0059] (3) The present invention calculates the total loss L by calculating the identification loss and reconstruction loss test , total loss L test It reflects the gap between the actual value and the predicted value. test It is defined as calculating the identification score. When the current score exceeds the preset value, it is judged that a fault exists. This can achieve the purpose of time series anomaly detection. The detection method does not require a complex calculation process and is simple and convenient. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a logical process framework diagram of the substation equipment fault diagnosis method based on the integrated deep generative model. DETAILED DESCRIPTION

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0062] like Figure 1 As shown, a substation equipment fault diagnosis method based on an integrated deep generative model includes:

[0063] S1: Data preprocessing, constructing sample training datasets and random potential datasets;

[0064] The operating data of substation equipment is collected and divided into several subsequences of equal length using a sliding time window of length n. Each subsequence is taken as a sample and each subsequence is tested for abnormality. All normal samples are selected to construct a sample training dataset, and the data sequences doped with abnormal samples are used to construct a random latent dataset. This ensures that all data in the sample training dataset are normal and there are abnormal data in the random latent dataset.

[0065] S2: Constructing an AMBI-GAN integrated deep generative model based on bidirectional long short-term memory network and attention mechanism;

[0066] Step 2.1: Build the Bi-LSTM structure.

[0067] A bidirectional LSTM network is constructed as a Bi-LSTM structure. The bidirectional LSTM network belongs to the prior art. The present invention does not improve its network architecture. It only adds an attention mechanism to the network structure by increasing the weight parameter in step 2.2. The input of the bidirectional LSTM network is X = {x1, ...x n}, x n Represents the input at the nth moment;

[0068] The hidden layer state of the Bi-LSTM structure is

[0069]

[0070]

[0071] Among them, LSTM backward For backward LSTM, LSTM forward For the forward LSTM, x t represents the input at the tth moment, represents the hidden state of the backward LSTM at time t+1, Represents the hidden state of the forward LSTM at time t-1.

[0072] The hidden layer state output at n moments is

[0073]

[0074] Step 2.2: Add the attention mechanism to the Bi-LSTM structure to obtain the AMBi-LSTM structure.

[0075] By formula Set the input weight of the Bi-LSTM structure, thereby adding the attention mechanism to the Bi-LSTM structure to obtain the AMBi-LSTM structure; where w i represents the weight parameter of the Bi-LSTM structure at the i-th moment, softmax is the normalization function, and Bi-LSTM(X,H) represents the output of the Bi-LSTM structure;

[0076] The output of the AMBi-LSTM structure is

[0077] represents the transpose of the input matrix at the i-th moment, represents the transpose of the hidden state matrix at the i-th moment.

[0078] In step 2.3, the generator and discriminator are obtained using the AMBi-LSTM structure respectively. The generator and discriminator together constitute the AMBI-GAN integrated deep generative model. It should be noted that the generator and discriminator have the same network structure, both of which are AMBi-LSTM structures. The only difference is the input data. The generator inputs a random potential data set, while the discriminator inputs a sample training data set and the output of the generator.

[0079] S3: Train the AMBI-GAN ensemble deep generative model. During the training process, the objective function reaches a Nash equilibrium by updating the weight parameters and the mean of the data distribution. The weight parameters have been explained in step 2.2 above. The following details the calculation of the mean of the data distribution and how to train the model.

[0080] Step 3.1: The input of the generator is a random potential data set, and the mean of the data distribution of the generator is calculated;

[0081] Generator input sample z={z1,…z n} is the data sequence in the random potential data set, and the mean of the output sample of the generator is

[0082]

[0083] Among them, z i is the i-th data input to the generator, p(·) is the probability density function of the output, and log(·) is the logarithmic function;

[0084] During the training process, the network weight parameter W is continuously optimized, and the output sample E(z) is continuously brought close to the mean of the sample training data set.

[0085] Step 3.2: Design the discriminator’s identification model.

[0086] When the input of the discriminator is a sample training dataset, the discriminator's identification model is

[0087] E x ~p data(x) log(D(x))

[0088] Among them, p data(x) represents the probability distribution of input data x, E x ~p data(x) Indicates that p data(x) The mean of the variable x of the distribution, D(x) represents the output of the discriminator for x;

[0089] When the input of the discriminator is the output of the generator G(z), the discriminator's identification model is

[0090] E x ~p z(z) log(1-D(G(z)))

[0091] Among them, p z(z) represents the probability distribution of random latent data z, E x ~p z(z) Indicates that p z(z) The mean of the distribution z, D(G(z)) represents the output of the discriminator for G(z).

[0092] Step 3.3, through the formula

[0093] min G max D V(D, G)=E x ~p data(x) log(D(x))+E x ~p z(z) log(1-D(G(z)))

[0094] Construct the objective function;

[0095] The output G(z) of the generator G serves as the input to the discriminator D. The goal is to reduce V so that the generated data distribution cannot be distinguished. The goal is to increase V for the discriminator's output D(x) so that it can effectively distinguish true from false data. By playing a game between the generator and the discriminator, the network's weight parameters and the mean of the data distribution are continuously adjusted so that the data produced by the generator can be distinguished as real data by the discriminator. In other words, training stops when the objective function reaches a Nash equilibrium.

[0096] S4: After the above model is trained, use the trained model to detect abnormal data, input test data into the trained model, and calculate the total loss L of the model. test , through the total loss L test Further judge the abnormality. The main process is as follows:

[0097] Step 4.1, calculation of identification loss. Input the original sample test data directly into the discriminator and calculate the test space residual R. The function is expressed as:

[0098]

[0099] Among them, N k represents the result of k iterations of the model subsequence N = {Ni, i = 1, 2, ..., n} during the model training process, X test represents the original test data, Representative generator AMB i -Output of LSTM network. By calculating the sum of residuals at i moments, the anomaly detection model obtains the identification loss value R(X test ):

[0100]

[0101] in, Represents the model subsequence N={N i ,i=1,2,…,n},x test,i Represents the i-th data in the original test data, Represents the output of the generator;

[0102] Step 4.2: Calculate the total loss. In the process of time series anomaly detection, the detection result consists of two parts: identification loss and reconstruction loss. The total loss is obtained by calculating the sum of identification loss and reconstruction loss. The result is as follows:

[0103]

[0104] Among them, λ represents an adjustable parameter, Represents the reconstruction loss, which is the direct output of the discriminator after inputting the test data into the discriminator.

[0105] S5: Loss L test Indicates the gap between the true value and the predicted value, 1-L test It is defined as calculating the identification score. If the identification score exceeds the preset value, it is judged to be abnormal, so as to achieve the purpose of time series anomaly detection. test A value greater than or equal to 0.5 indicates normal, and a value less than 0.5 indicates abnormal.

[0106] Through the above technical scheme, the present invention has made corresponding improvements to the GAN model and proposed a generative adversarial network integrated deep generative model AMBI-GAN based on a bidirectional long short-term memory network and an attention mechanism. The model is constructed based on a bidirectional long short-term memory network, which can capture the dependency of time series, effectively extract the temporal characteristics of power data, and improve the accuracy of the model. Moreover, based on the inherent performance of the GAN model, a large amount of simulated data can be generated based on limited real power data, thereby making full use of a small amount of labeled data, with small computational complexity, small amount of original data, and low labor cost. In addition, the influence of noise on the data can be reduced through training, thereby realizing the anomaly detection of power time series, improving the detection accuracy of multidimensional time series anomalies, and further improving the fault diagnosis performance.

[0107] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A substation equipment fault diagnosis method based on an integrated deep generative model, characterized in that: The method comprises: Step 1: Collect substation equipment operation data, perform data preprocessing, and construct a sample training dataset and a random potential dataset; Step 2: Build a bidirectional long short-term memory network and attention mechanism Integrating a deep generative model; the second step includes: Step 2.1, Build structure; Step 2.2, The structure is added with attention mechanism to obtain structure; Step 2.3, respectively use The structure obtains the generator and discriminator, the generator and discriminator as a whole Integrate deep generative models; Step 3: Training Integrate deep generative models; Step 4: Input test data into the trained model and calculate the total loss of the model ; The original sample test data is directly input into the trained model, and the formula Calculate the identification loss value, where Represents a model subsequence carried out The result of the iterations, represents the i-th data in the original test data, Represents the output of the generator; By formula Calculate the total loss, where represents an adjustable parameter, represents reconstruction losses; Step 5: Pass Get the identification score, and if the identification score exceeds the preset value, it is judged to be abnormal.

2. The substation equipment fault diagnosis method based on integrated deep generative model according to claim 1 is characterized in that: The data preprocessing in step 1 includes: The input data is divided into a sample training data set and a random potential data set. All data in the sample training data set are normal, and there are abnormal data in the random potential data set. The sliding time window divides all data into several subsequences of equal length.

3. The substation equipment fault diagnosis method based on integrated deep generative model according to claim 1 is characterized in that: The step 2.1 includes: Build a two-way Network as Structure, in which two-way The input of the network is , Represents the input at the nth moment, The hidden layer state of the structure is in, Backward , Forward , represents the input at the tth moment, Indicates the Backward The hidden state of Indicates the Always looking forward The hidden state of The hidden layer state output at the moment is 。 4. The method for substation equipment fault diagnosis based on integrated deep generative model according to claim 3 is characterized in that: The step 2.2 includes: By formula set up The input weight of the structure, thus The structure is added with attention mechanism to obtain structure; among them, represents the i-th moment The weight parameter of the structure, is the normalization function, express Output of the structure; The output of the structure is , in, represents the transpose of the input matrix at the i-th moment, represents the transpose of the hidden state matrix at the i-th moment.

5. The method for fault diagnosis of substation equipment based on integrated deep generative model according to claim 4 is characterized in that: The step three includes: Step 3.1: The input of the generator is a random potential data set, and the mean of the data distribution of the generator is calculated; Step 3.2: Design the discriminator’s identification model. Step 3.3: Set the objective function according to the identification model, and continuously adjust the network weight parameters and data distribution mean so that the mean of the generator's output samples is closer to the mean of the sample training data set. When the objective function iteration reaches Nash equilibrium, stop training.

6. The method for fault diagnosis of substation equipment based on integrated deep generative model according to claim 5, characterized in that: The step 3.1 includes: Generator input samples is the data sequence in the random latent dataset, and the mean of the output sample of the generator is in, is the i-th data input to the generator, is the probability density function of the output, is a logarithmic function.

7. The method for fault diagnosis of substation equipment based on integrated deep generative model according to claim 6, characterized in that: The step 3.2 includes: When the input of the discriminator is a sample training dataset, the discriminator's identification model is in, Represents input data The probability distribution of Express satisfaction Distributed variables The mean of Denotes the discriminator pair Output; When the input of the discriminator is the output of the generator , then the identification model of the discriminator is in, Represents random latent data The probability distribution of Express satisfaction distributed The mean of Denotes the discriminator pair Output.

8. The method for fault diagnosis of substation equipment based on integrated deep generative model according to claim 7, characterized in that: The step 3.3 includes: Construct the following objective function The output of the generator As the input of the discriminator, the network weight parameters and data distribution mean are continuously adjusted so that the data generated by the generator can be identified as real data by the discriminator. That is, when the objective function iteration reaches Nash equilibrium, the training is stopped.

Images