On-Load Tap Changer Fault Diagnosis Method Based on Bayesian Optimization Ladder Network

Abstract

The present invention discloses a method for diagnosing on-load tap-changer faults based on a Bayesian-optimized ladder network, belonging to the technical field of power equipment fault diagnosis. The method includes steps such as collecting vibration signals of the on-load tap-changer, dimensionality reduction of the vibration signals, optimizing the hyperparameters of the ladder network using the Bayesian optimization method, training the ladder network with the optimized hyperparameters, and performing fault diagnosis with the trained ladder network. The present invention utilizes labeled samples for supervised learning and unlabeled samples for unsupervised learning at the same time, improving the generalization performance of the neural network. Compared with traditional supervised learning methods, it reduces the dependence on labeled samples, can significantly improve the accuracy of on-load tap-changer fault diagnosis. At the same time, the present invention performs Bayesian optimization on two types of hyperparameters, namely the noise amplitude injected into the encoder of the ladder network and the reconstruction error weight of each layer of the decoder, further improving the performance of the ladder network.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment fault diagnosis, and particularly relates to a on-load tap-changer fault diagnosis method based on a Bayesian optimization ladder network. Background Art

[0002] On-load tap-changers are widely used in various power transformers and play a key role in on-load voltage regulation. Faults often lead to serious consequences. The on-load tap-changer is the only moving mechanism in the transformer, with complex mechanical structures and electrical characteristics. Its faults can be divided into mechanical faults and electrical faults. According to statistics, mechanical faults of on-load tap-changers account for more than 80%, and electrical faults usually evolve from mechanical faults. Therefore, it is of great significance to evaluate and monitor the mechanical state of on-load tap-changers.

[0003] The vibration signals of on-load tap-changers contain fault information of the driving mechanism, transmission mechanism, quick mechanism, switching switch, and tap selector, and can fully reflect the mechanical states of each component and the whole of the on-load tap-changer. The vibration signals of on-load tap-changers can be obtained through two methods: on-line monitoring and fault simulation experiments. Among them, the number of samples obtained by on-line monitoring is large, but effective labeling cannot be carried out; the fault simulation experiment can effectively label the samples, but the number of samples obtained is small. Therefore, on-load tap-changer fault diagnosis often faces the problem of lack of labeled samples.

[0004] At present, most on-load tap-changer fault diagnosis methods are based on supervised learning and have a large demand for labeled samples, but do not make full use of the unlabeled samples collected by on-line monitoring devices. Especially for deep learning methods, the lack of sufficient training samples will cause the neural network to quickly fall into overfitting, thereby reducing the generalization performance of the neural network. Summary of the Invention

[0005] In order to overcome the deficiencies in the prior art and make full use of the unlabeled samples collected by on-line monitoring devices, the present invention provides a on-load tap-changer fault diagnosis method based on a Bayesian optimization ladder network.

[0006] A on-load tap-changer fault diagnosis method based on a Bayesian optimization ladder network, the method comprising the following steps:

[0007] Step 1: Collect vibration signals of the on-load tap-changer;

[0008] Step 2: Reduce the dimension of the vibration signals;

[0009] Step 3: Optimize the hyperparameters of the ladder network by using the Bayesian optimization method;

[0010] Step 4: Train the ladder network with the optimized hyperparameters;

[0011] Step 5: Perform fault diagnosis using the trained ladder network.

[0012] Furthermore, in the above Step 1, the on-load tap-changer vibration signals are obtained through an on-line monitoring device and a fault simulation experiment. A large number of unlabeled samples are obtained by the on-line monitoring device, and a small number of labeled samples are obtained by the fault simulation experiment. The simulated faults include, but are not limited to, transmission jamming, insufficient core lubrication, and loose top cover.

[0013] Furthermore, in the above Step 2, the specific method for dimension reduction of the vibration signal is as follows: First, extract the main vibration wave during the switch core striking stage , then perform Fourier transform on to obtain the single-sided spectrum, and finally take the first 20% of the low-frequency part of the single-sided spectrum as the dimension-reduced data:

[0014]

[0015] Furthermore, in the above Step 3, the hyperparameters to be optimized in the ladder network are the noise amplitude injected into the encoder and the reconstruction error weight of each layer of the decoder .

[0016] Furthermore, in the above Step 3, when performing Bayesian optimization, the probabilistic surrogate model selects the Gaussian process, and the acquisition function selects the EI (Expected Improvement) acquisition function.

[0017] Furthermore, in the above Step 3, when performing Bayesian optimization, first generate 20 initial evaluation points in the hyperparameter space based on the Sobol sequence, and on this basis, perform 10 times of hyperparameter optimization based on the Gaussian process and the EI acquisition function.

[0018] Furthermore, in the above Step 4, the ladder network adopts a multi-layer perceptron ladder network. The encoder of the ladder network includes three steps: linear mapping, batch normalization, and non-linear activation. The decoder of the ladder network includes three steps: linear mapping, noise reduction, and normalization.

[0019] Furthermore, in the above Step 4, the encoder of the ladder network is supervised and trained with the labeled samples obtained by the fault simulation experiment, and the decoder of the ladder network is unsupervised and trained with the unlabeled samples obtained by the on-line monitoring device.

[0020] Furthermore, in the above Step 5, during diagnosis, the dimension-reduced data is input into the encoder of the ladder network without noise, and the output of the encoder is the classification result.

[0021] Advantages of the present invention:

[0022] (1) The present invention uses labeled samples for supervised learning and unlabeled samples for unsupervised learning, improving the generalization performance of the neural network and reducing the dependence on labeled samples compared with traditional supervised learning methods.

[0023] (2) The present invention performs Bayesian optimization on two types of hyperparameters, namely, the noise amplitude injected into the encoder of the ladder network and the reconstruction error weight of each layer decoder, further improving the performance of the ladder network. Description of the Drawings

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0025] Figure 1 is the flowchart of the present invention;

[0026] Figure 2 is the process of Bayesian optimization;

[0027] Figure 3 is the accuracy rate of the evaluation points;

[0028] Figure 4 is the structural diagram of the ladder network. Detailed Embodiment

[0029] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0030] A fault diagnosis method for on-load tap-changer based on a ladder network optimized by Bayesian optimization, as Figure 1 shown, the method includes the following steps:

[0031] Step 1: Collect the vibration signals of the on-load tap-changer; the collection of the vibration signals of the on-load tap-changer is obtained through an on-line monitoring device and a fault simulation experiment. The on-line monitoring device is used to obtain a large number of unlabeled samples, and the fault simulation experiment is used to obtain a small number of labeled samples. The faults that need to be simulated in the fault simulation experiment include transmission jamming, insufficient core lubrication, loose top cover, etc.

[0032] Step 2: Reduce the dimension of the vibration signals; the specific method is to first extract the main vibration wave during the switch core hitting stage , and then for Perform Fourier transform to obtain a single-sided spectrum, and finally take the first 20% of the low-frequency part of the single-sided spectrum as the dimensionality-reduced data: .

[0033] Step 3: Optimize the hyperparameters of the ladder network using the Bayesian optimization method; the hyperparameters to be optimized for the ladder network are the noise amplitude injected into the encoder and the reconstruction error weight of each layer of the decoder . When performing Bayesian optimization, the probabilistic surrogate model selects the Gaussian process, and the acquisition function selects the EI (Expected Improvement) acquisition function. At the same time, when performing Bayesian optimization, first generate 20 initial evaluation points in the hyperparameter space based on the Sobol sequence, and on this basis, perform 10 times of hyperparameter optimization based on the Gaussian process and the EI acquisition function.

[0034] Step 4: Train the ladder network with the optimized hyperparameters; the ladder network uses a multi-layer perceptron ladder network. The encoder of the ladder network includes three steps: linear mapping, batch normalization, and non-linear activation. The decoder of the ladder network includes three steps: linear mapping, noise reduction, and normalization. The encoder of the ladder network is supervised and trained with the labeled samples obtained from the fault simulation experiment, and the decoder of the ladder network is unsupervised and trained with the unlabeled samples obtained from the on-line monitoring device.

[0035] Step 5: Perform fault diagnosis using the trained ladder network; during diagnosis, input the dimensionality-reduced data into the encoder of the ladder network without noise, and the output of the encoder is the classification result.

[0036] In one embodiment, 4000 groups of unlabeled vibration signals of on-load tap changers are obtained through an on-line monitoring device, and 80 groups of labeled vibration signals of on-load tap changers are obtained through a fault simulation experiment, with 20 groups of signals for normal, transmission jamming, insufficient core lubrication, and loose top cover respectively.

[0037] Extract the main vibration wave during the stage when the switch core makes a clicking sound , and then perform Fourier transform on to obtain a single-sided spectrum, and finally take the first 20% of the low-frequency part of the single-sided spectrum as the dimensionality-reduced data:

[0038]

[0039] Perform Bayesian optimization on the hyperparameters of the ladder network, namely the noise amplitude injected into the encoder and the reconstruction error weight of each layer of the decoder . The process of Bayesian optimization is as shown in the appendix Figure 2As shown. 32 groups are extracted from 80 groups of labeled data as the validation set, and the remaining 48 groups of labeled data and 4000 groups of unlabeled data are used as the training set. The objective function of Bayesian optimization is the classification accuracy of the ladder network on the validation set.

[0040] The probability surrogate model of the Bayesian optimization algorithm selects the Gaussian process:

[0041]

[0042]

[0043]

[0044] When the Gaussian process regression updates the posterior distribution when some evaluation points (X, y) are known, the method is:

[0045]

[0046]

[0047]

[0048] Where and are the unknown evaluation point and its corresponding objective function value respectively, is the covariance function.

[0049] The acquisition function selects the EI (Expected Improvement) acquisition function:

[0050]

[0051] Where and are the cumulative distribution function and probability density function of the standard normal distribution respectively, is the maximum objective function value in the previous n evaluations.

[0052] The process of Bayesian optimization is as follows. First, 20 initial evaluation points are generated based on the Sobol sequence in the hyperparameter space, and on this basis, 10 hyperparameter optimizations based on the Gaussian process and the EI acquisition function are carried out. The accuracies of 30 evaluation points in the validation set during the Bayesian optimization process are as Figure 3 shown. In this embodiment, a 3-layer ladder network is used, and the optimal hyperparameters obtained through Bayesian optimization = 0.22, = 801.05, = 2.28, = 0.65.

[0053] Train the ladder network with the optimized hyperparameters. The structure of the ladder network is as Figure 4 shown. The ladder network adopts a multi-layer perceptron ladder network. The encoder of the ladder network includes three steps: linear mapping, batch normalization, and non-linear activation. The decoder of the ladder network includes three steps: linear mapping, noise reduction, and normalization.

[0054] The encoder of the ladder network is supervised and trained with the labeled samples obtained from the fault simulation experiment, and the decoder of the ladder network is unsupervised and trained with the unlabeled samples obtained from the online monitoring device.

[0055] Use the trained ladder network for fault diagnosis. Input the dimensionality-reduced data into the encoder of the ladder network without noise, and the output of the encoder is the classification result.

[0056] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification only illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed.

Claims

1. A on-load tap-changer fault diagnosis method based on a Bayesian optimization ladder network, characterized in that, The method includes the following steps: Step 1: Collect the vibration signals of the on-load tap-changer; Step 2: Reduce the dimension of the vibration signals; Step 3: Optimize the hyperparameters of the ladder network using the Bayesian optimization method; Step 4: Train the ladder network with the optimized hyperparameters; Step 5: Conduct fault diagnosis using the trained ladder network; In the above Step 1, the vibration signals of the on-load tap-changer are obtained through an on-line monitoring device and a fault simulation experiment. The on-line monitoring device is used to obtain unlabeled samples, and the fault simulation experiment is used to obtain labeled samples. The faults simulated in the fault simulation experiment include transmission jamming, insufficient core lubrication, and loose top cover; The specific method for reducing the dimension of the vibration signals in the above Step 2 includes the following steps: 1), extract the main vibration wave v in the stage when the switch core makes a clicking sound n , n = 1, ..., N; 2) Perform a Fourier transform on v n to obtain a single-sided spectrum; 3) Take the first 20% of the low-frequency part of the unilateral spectrum as the dimension-reduced data: The hyperparameters to be optimized in the ladder network in step 3 are the noise amplitude Anoise injected into the encoder and the reconstruction error weights of each layer of the decoder , l where \(l = 1,2,\cdots,L\); when performing Bayesian optimization, the probabilistic surrogate model is a Gaussian process, the acquisition function is the EI acquisition function, and the Bayesian optimization method is to generate 20 initial evaluation points based on the Sobol sequence in the hyperparameter space and perform 10 times of hyperparameter optimization based on the Gaussian process and the EI acquisition function on this basis.

2. The on-load tap-changer fault diagnosis method based on a Bayesian optimization ladder network according to claim 1, characterized in that, In the above Step 4, the ladder network adopts a multi-layer perceptron ladder network. The encoder of the ladder network includes three steps: linear mapping, batch normalization, and non-linear activation. The decoder of the ladder network includes three steps: linear mapping, noise reduction, and normalization.

3. The on-load tap-changer fault diagnosis method based on a Bayesian optimization ladder network according to claim 2, characterized in that, The encoder of the ladder network is supervised and trained with the labeled samples obtained from the fault simulation experiment, and the decoder of the ladder network is unsupervised and trained with the unlabeled samples obtained from the on-line monitoring device.

4. The on-load tap-changer fault diagnosis method based on a Bayesian optimization ladder network according to claim 1, characterized in that, In the above Step 5, during diagnosis, the dimension-reduced data is input into the encoder of the ladder network without noise, and the output of the encoder is the classification result.

Images