Power transformer early fault early warning method based on PMCAEN and SVDD

Abstract

The invention discloses a power transformer early fault early warning method based on PMCAEN and SVDD, and belongs to the technical field of power transformer fault early warning. Firstly, a parallel multi-scale convolution auto-encoder network is constructed, a parallel unsupervised learning architecture in the PMCAEN can extract related features from different angles, and feature complementation of double branches is achieved. And then, establishing a support vector data description model, and constructing a judgment domain of an abnormal state by taking a reconstruction error obtained by the PMCAEN under the normal operation condition of the transformer as new observation information. According to the method, the problem that early warning is difficult due to the fact that vibration signals do not change obviously when the transformer has a slight mechanical fault is fully considered, mining of local and global degradation information of the transformer can be achieved in a self-adaptive mode, and high-precision early fault early warning of the power transformer can be achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power transformer fault warning, and more specifically, relates to a power transformer early fault warning method based on PMCAEN and SVDD. Background Art

[0002] As one of the important equipment in the new power system, the reliability of power transformer operation directly affects the stable operation of the power grid. However, transformers will inevitably encounter various mechanical failures during long-term operation, such as winding deformation, loose core and loose fasteners. If these mechanical failures are not discovered and handled in time, they may cause serious equipment damage and even cause large-scale power outages. Therefore, early fault warning, as an emerging health status monitoring tool, plays a vital role in timely discovering and repairing early mechanical failures of power transformers and reducing transformer downtime.

[0003] At present, fault warning methods in various industrial fields are mainly divided into two categories, namely, process control-based methods and machine learning-based methods. In process control methods, control charts are a commonly used warning method, which determines whether the process has changed by detecting whether the process signal exceeds the control threshold. However, this type of method needs to set a suitable warning threshold according to the structure and characteristics of the data, resulting in poor applicability to different types of data and a high false alarm rate. Machine learning methods can automatically learn relevant knowledge from data, adaptively construct warning boundaries, avoid the selection of manual experience, and effectively overcome the defects of process control methods. However, most machine learning methods only use a single time window to extract features of time series, without considering local and global information, which will undoubtedly affect the accuracy of fault warning.

[0004] The vibration signal generated by the power transformer usually contains information about the mechanical state of the transformer changing over time. Therefore, vibration analysis is one of the most effective methods for early fault warning and fault diagnosis of power transformers. Early fault warning of power transformers is generally studied from two aspects: (1) extracting degradation information from the vibration signal; (2) establishing an abnormal judgment domain to effectively distinguish between the healthy state and the degraded state. It is worth noting that when a minor mechanical fault occurs in a power transformer, its vibration signal does not change significantly. At the same time, the vibration signal of the power transformer is a vector containing time information. If the time scale is not fully considered, it will affect the effective mining of power transformer degradation information. Summary of the invention

[0005] In view of the above defects or improvement needs of the prior art, the present invention proposes an early fault warning method for power transformer based on PMCAEN and SVDD to solve the problem that it is difficult to warn of early mechanical failures of power transformer.

[0006] To achieve the above object, according to one aspect of the present invention, a method for early fault warning of a power transformer based on PMCAEN and SVDD is provided, comprising:

[0007] Collecting a vibration signal of the power transformer, wherein the vibration signal is a historical normal sample composed of normal vibration signals;

[0008] Constructing a parallel multi-scale convolutional autoencoder network PMCAEN, and using the historical normal samples to train the parallel multi-scale convolutional autoencoder network PMCAEN, during the training process, adjusting the hyperparameters of the parallel multi-scale convolutional autoencoder network PMCAEN, and calculating the reconstruction error between the input and the output;

[0009] The reconstruction error output by the parallel multi-scale convolutional autoencoder network PMCAEN is used as observation information and input into the support vector data description SVDD model for training. The support vector data description SVDD model learns the intrinsic relationship from the structure of the observation information and constructs a hypersphere space. The radius of the hypersphere space is used as the warning threshold to construct a judgment domain for abnormal conditions.

[0010] In some optional embodiments, the parallel multi-scale convolutional autoencoder network PMCAEN uses two parallel multi-scale convolutional autoencoder paths to extract degradation characteristics of the power transformer, wherein the convolutional autoencoder with a first convolution kernel can perform feature extraction on the local degradation information of the power transformer, and the convolutional autoencoder with a second convolution kernel can consider the degradation trend of the entire sequence and perform feature extraction, and the size of the first convolution kernel is smaller than that of the second convolution kernel.

[0011] In some optional implementations, for each multi-scale convolutional autoencoder, in the multi-scale convolutional layer, a set of kernels is convolved with the input feature vector, and the j-th output feature vector of the l-th convolutional layer is for, In the multi-scale deconvolution layer, the decoder reconstructs the original input data into Among them, * is the convolution operation, σ(·) is the activation function, is the dth eigenvector of the l-1 layer, is the convolution kernel, is the bias vector, is the feature vector in the deconvolution process, is the convolution kernel in the deconvolution process.

[0012] In some optional implementations, adjusting the hyperparameters of the parallel multi-scale convolutional autoencoder network PMCAEN and calculating the reconstruction error between the input and the output includes:

[0013] The internal weights and biases of the two parallel multi-scale convolutional autoencoders in the parallel multi-scale convolutional autoencoder network PMCAEN are updated by back propagation to minimize the reconstruction error. After obtaining the reconstruction errors of different scales of each multi-scale convolutional autoencoder, the reconstruction errors of each different scale are averaged to obtain the average error of the two branches. The average errors of the two branches are connected to obtain the final reconstruction error, where the reconstruction error δ of each branch is: N is the number of samples in the historical normal samples, is the input vector x i The reconstructed output of ||·|| 2 Represents the square of the norm.

[0014] In some optional embodiments, the hypersphere space is constructed by optimizing the objective function F: Among them, the training data set The length is N, y i is the historical normal sample x i , a and R are the center and radius of the hypersphere, ξ i is the slack variable and C is the penalty coefficient.

[0015] In some optional implementation schemes, by introducing Lagrange multipliers and kernel functions, the hypersphere space is optimized as follows: Among them, α is the Lagrangian operator, K(x i ,x j ) is the kernel function.

[0016] In some optional embodiments, Get the radius R of the hypersphere space, where x sv is the support vector.

[0017] In some optional implementation schemes, the construction of the judgment domain of the abnormal state includes:

[0018] The collected vibration signal to be measured is input into the trained parallel multi-scale convolutional autoencoder network PMCAEN to extract the reconstruction error vector;

[0019] Inputting the reconstructed error vector into the trained support vector data description (SVDD) model to evaluate whether the distance from the vibration signal to be measured to the center of the hypersphere space exceeds the warning threshold;

[0020] If the distance is less than or equal to the warning threshold, it is determined that the corresponding power transformer is in a healthy state;

[0021] If the distance is greater than the warning threshold, the corresponding power transformer is judged to be in a degraded state and a warning is issued.

[0022] In some optional implementation schemes, the distance d from the vibration signal z to be measured to the center of the hypersphere is: x i It is a historical normal sample.

[0023] According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of any one of the above methods are implemented.

[0024] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0025] The present invention combines PMCAEN and SVDD to establish an early fault warning model for power transformers using only normal samples. Among them, PMCAEN is committed to mining the global and local degradation information of the vibration signal of the power transformer and constructing the reconstruction error. SVDD focuses on establishing the judgment domain from normal samples and making early warning decisions. The present invention fully considers the problem that when a minor mechanical fault occurs in the transformer, the vibration signal does not change significantly, resulting in difficulty in early warning. It can adaptively realize the mining of local and global degradation information of the transformer, and can achieve high-precision early fault warning of the power transformer. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flow chart of a method for early fault warning of a power transformer based on PMCAEN and SVDD provided by an embodiment of the present invention;

[0027] Figure 2 It is a PMCAEN framework diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0029] In the examples of the present invention, "first", "second", etc. are used to distinguish different objects rather than to describe a specific order or sequence.

[0030] The present invention provides an early fault warning method for power transformers based on PMCAEN and SVDD. First, a parallel multi-scale convolutional autoencoder network (Parallel Multi-scale Convolutional Autoencoder Network, PMCAEN) is designed. Among them, the convolutional autoencoder with a small convolution kernel can extract features for the local degradation information of the power transformer. In contrast, the convolutional autoencoder with a large convolution kernel can consider the degradation trend of the entire sequence and perform feature extraction. At the same time, the parallel unsupervised learning architecture in PMCAEN can extract relevant features from different angles to achieve feature complementarity of dual branches. Subsequently, a support vector data description (Support Vector Data Description, SVDD) model is established, and the reconstruction error obtained by PMCAEN under normal operation of the transformer is used as new observation information to construct a judgment domain for abnormal conditions.

[0031] like Figure 1 As shown, the present invention proposes a power transformer early fault warning method based on PMCAEN and SVDD, comprising the following steps:

[0032] 1) Construct PMCAEN, which is an unsupervised deep learning network trained using normal samples of power transformer vibration signals.

[0033] 2) Among them, Figure 2 As shown in Figure 1, PMCAEN uses two parallel multi-scale convolutional autoencoder paths to extract the degradation features of the power transformer. It should be noted that the convolution kernel sizes of the two parallel paths are different, which can effectively mine the local and global degradation features of the power transformer vibration signal.

[0034] 3) Further, in the multi-scale convolutional layer, a set of kernels are convolved with the input feature vector. The j-th output feature vector of the l-th convolutional layer is It can be expressed as:

[0035]

[0036] Where * is the convolution operation; σ(·) is the activation function; is the dth eigenvector of the l-1 layer; is the convolution kernel; is the bias vector.

[0037] 4) Further, in the multi-scale deconvolution layer, the decoder reconstructs the original input data, which can be expressed as:

[0038]

[0039] in, is the feature vector in the deconvolution process, is the convolution kernel in the deconvolution process.

[0040] 5) PMCAEN updates the internal weights and biases through back-propagation to minimize the reconstruction error.

[0041] The reconstruction error can be expressed as:

[0042]

[0043] Where N is the number of samples; is the input vector x i The reconstructed output of ||·|| 2 Represents the square of the norm. After obtaining the reconstruction errors of different scales, an average operation is performed, and the average errors of the two branches are concatenated.

[0044] 6) Furthermore, in PMCAEN, the Markov Transition Field (MTF) method is used to convert the one-dimensional vibration signal into a two-dimensional image. For a given time series signal X = {x1, x2, ..., x n}, which consists of a sampled signal x with n timestamps i The MTF method first divides the time series signal into Z regions (bins) according to the amplitude, and each bin can be expressed as q i (i=1,2,…,Z), each sampling signal x i Mapping to region q i Then, calculate the sampling signal x before and after i-1 and x i From area q i Convert to q j The probability of calculating P is taken as the element w ij , we get the Z×Z Markov state transfer matrix W:

[0045]

[0046] 7) From formula (4), it can be seen that the Markov state transfer matrix only calculates the transition probability of the previous and next moments, without considering the dynamic probability transfer of time series data. MTF improves this defect and generates a dynamic probability transfer matrix M across time scales based on the Markov state transfer matrix:

[0047]

[0048] 8) The calculation steps of MTF include determining x1, ..., x1 by comparing the Markov probability transfer matrix W in equation (4). n The transition probability between them is finally obtained, and the MTF with dimension n×n is obtained. Each element w in MTF ij represents the transition probability between points with a time interval of |ij|.

[0049] 9) Furthermore, the reconstruction error output by PMCAEN is used as observation information, and the support vector data description (SVDD) method is used to determine whether the mechanical performance of the power transformer has entered the degradation stage. Given a training data set The length is N, where y i is the historical normal sample x i The hypersphere is constructed by optimizing the objective function F:

[0050]

[0051] Where a and R are the center and radius of the hypersphere respectively; ξ i is the slack variable; C is the penalty coefficient.

[0052] 10) Furthermore, by introducing Lagrange multipliers and kernel functions, the original problem is constructed as follows:

[0053]

[0054]

[0055] Where α is the Lagrangian operator; K(x i ,x j ) is the kernel function.

[0056] 11) Furthermore, the radius R of the hypersphere is calculated by equation (9):

[0057]

[0058] Where x sv is the support vector.

[0059] 12) For the test sample z, its distance to the center of the hypersphere can be expressed as:

[0060]

[0061] When d≤R, it means that the test sample is on or inside the hypersphere and is judged as a normal sample; otherwise, it is an abnormal sample.

[0062] 13) The present invention combines PMCAEN and SVDD to establish an early fault warning model for power transformers using only normal samples. The invention mainly includes two parts: error reconstruction and early fault warning. Among them, PMCAEN is committed to mining the global and local degradation information of power transformer vibration signals and constructing reconstruction errors. SVDD focuses on establishing a judgment domain from normal samples and making early warning decisions. The specific steps of the present invention are summarized as follows:

[0063] a) Vibration signal acquisition. The vibration signal is obtained by sensors installed at different locations on the outer wall of the power transformer oil tank. The collected vibration signals are divided into offline historical normal samples and online samples. The offline samples are used as training data to construct the judgment domain, and the online samples are used as test data to evaluate the operating status of the power transformer in real time.

[0064] b) Error reconstruction: The PMCAEN model proposed in the present invention is trained. During the training process, the hyperparameters of PMCAEN are adjusted to make the output as close to the input as possible, and the reconstruction error between the input and the output is calculated.

[0065] c) Construct the judgment domain. The reconstruction error of the normal sample is used as new observation information, input into the SVDD model for training, and k-fold cross validation is performed. Through the training process, SVDD learns the intrinsic relationship from the structure of the input data and constructs a hypersphere space. The radius of the hypersphere space is used as the warning threshold, which is adaptively learned based on the input data rather than set by human experience.

[0066] d) Early fault warning. The newly acquired vibration signal is input into the trained PMCAEN to extract the reconstructed error vector. Then, the reconstructed error vector is input into the trained SVDD model to evaluate whether its distance exceeds the threshold. If the distance is less than or equal to the threshold, the sample is judged to be in a healthy state. Otherwise, it will be judged as a degraded state and a warning will be given.

[0067] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, and two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0068] It will be easily understood by those skilled in the art that 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 protection scope of the present invention.

Claims

1. A power transformer early fault warning method based on PMCAEN and SVDD, It is characterized in that include: Collecting a vibration signal of the power transformer, wherein the vibration signal is a historical normal sample composed of normal vibration signals; Constructing a parallel multi-scale convolutional autoencoder network PMCAEN, and using the historical normal samples to train the parallel multi-scale convolutional autoencoder network PMCAEN, during the training process, adjusting the hyperparameters of the parallel multi-scale convolutional autoencoder network PMCAEN, and calculating the reconstruction error between the input and the output; The reconstruction error output by the parallel multi-scale convolutional autoencoder network PMCAEN is used as observation information and input into the support vector data description SVDD model for training. The support vector data description SVDD model learns the intrinsic relationship from the structure of the observation information and constructs a hypersphere space. The radius of the hypersphere space is used as the warning threshold to construct a judgment domain for abnormal conditions.

2. The method according to claim 1, It is characterized in that The parallel multi-scale convolutional autoencoder network PMCAEN uses two parallel multi-scale convolutional autoencoder paths to extract the degradation characteristics of the power transformer, wherein the convolutional autoencoder with the first convolution kernel can perform feature extraction on the local degradation information of the power transformer, and the convolutional autoencoder with the second convolution kernel can consider the degradation trend of the entire sequence and perform feature extraction, and the size of the first convolution kernel is smaller than that of the second convolution kernel.

3. The method according to claim 2, It is characterized in that For each multi-scale convolutional autoencoder, in the multi-scale convolutional layer, a set of kernels is convolved with the input feature vector, and the j-th output feature vector of the l-th convolutional layer is for, In the multi-scale deconvolution layer, the decoder reconstructs the original input data into Among them, * is the convolution operation, σ(·) is the activation function, is the dth eigenvector of the l-1 layer, is the convolution kernel, is the bias vector, is the feature vector in the deconvolution process, is the convolution kernel in the deconvolution process.

4. The method according to claim 3, It is characterized in that The step of adjusting the hyperparameters of the parallel multi-scale convolutional autoencoder network PMCAEN and calculating the reconstruction error between the input and the output includes: The internal weights and biases of the two parallel multi-scale convolutional autoencoders in the parallel multi-scale convolutional autoencoder network PMCAEN are updated by back propagation to minimize the reconstruction error. After obtaining the reconstruction errors of different scales of each multi-scale convolutional autoencoder, the reconstruction errors of each different scale are averaged to obtain the average error of the two branches. The average errors of the two branches are connected to obtain the final reconstruction error, where the reconstruction error δ of each branch is: N is the number of samples in the historical normal samples, is the input vector x i The reconstructed output of ||·|| 2 Represents the square of the norm.

5. The method according to claim 4, It is characterized in that The hypersphere space is constructed by optimizing the objective function F: Among them, the training data set The length is N, y i is the historical normal sample x i , a and R are the center and radius of the hypersphere, ξ i is the slack variable and C is the penalty coefficient.

6. The method according to claim 5, It is characterized in that By introducing Lagrange multipliers and kernel functions, the hypersphere space is optimized as follows: 0≤α i ≤C, where α is the Lagrangian operator, K(x i ,x j ) is the kernel function.

7. The method according to claim 6, It is characterized in that Depend on Get the radius R of the hypersphere space, where x sv is the support vector.

8. The method according to any one of claims 1 to 7, It is characterized in that The construction of the judgment domain of the abnormal state includes: The collected vibration signal to be measured is input into the trained parallel multi-scale convolutional autoencoder network PMCAEN to extract the reconstruction error vector; Inputting the reconstructed error vector into the trained support vector data description (SVDD) model to evaluate whether the distance from the vibration signal to be measured to the center of the hypersphere space exceeds the warning threshold; If the distance is less than or equal to the warning threshold, it is determined that the corresponding power transformer is in a healthy state; If the distance is greater than the warning threshold, the corresponding power transformer is judged to be in a degraded state and a warning is issued.

9. The method according to claim 8, It is characterized in that The distance d from the vibration signal z to the center of the hypersphere is: x i It is a historical normal sample.

10. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

Images