Hydro-generator fault diagnosis method based on polynomial kernel induced maximum mean difference transfer learning

Abstract

The invention discloses a hydro-generator fault diagnosis method based on polynomial kernel-induced maximum mean difference transfer learning, and the method employs a polynomial kernel-induced maximum mean difference evaluation index to measure the difference between the distribution of a source domain data set and a target domain data set, and improves the robustness of a domain adaptive process. An unbiased estimation technology is adopted to improve the calculation process of the maximum mean difference evaluation index based on polynomial kernel induction, the calculation efficiency is improved, and a large amount of calculation cost is saved; a modulation factor changing based on the iteration process is adopted to balance the importance degree of the maximum mean difference loss and the classification cross entropy loss induced by the polynomial kernel, an overall loss function is constructed, and the migratable features between the source domain and the target domain can be accurately extracted through an iteratively trained model; according to the method, the learning effect of the target domain is improved, optimal convergence is effectively realized, the diagnosis precision is improved, and a reliable fault migration diagnosis task is completed.

Description

Technical Field

[0001] The invention relates to the technical field of hydraulic generator fault diagnosis, in particular to a hydraulic generator fault diagnosis method based on maximum mean difference transfer learning induced by polynomial kernel. Background Art

[0002] In recent years, with the rapid development of modern industry, the reliability and safety requirements for the integrated machine systems of hydropower stations have been greatly improved. After long-term operation, various faults will occur in rotating parts such as bearings and impellers of hydropower generator sets, which may cause damage to machinery and equipment, or even cause major accidents. The condition monitoring and fault diagnosis of hydropower generator sets are important links in the operation and maintenance management of power stations. By using real-time online monitoring systems and online swing monitoring systems, and based on massive available data, advanced data-driven fault diagnosis and identification technologies are developed to accurately diagnose the operating status of hydropower generator sets, providing strong guidance and basis for condition maintenance. By timely discovering and handling faults, major operating accidents can be effectively avoided, ensuring the safe and stable operation of power stations.

[0003] Existing research methods mainly focus on feature extraction-based methods for fault diagnosis of hydro-turbine generator sets. The accuracy of fault prediction is improved by extracting feature information from monitoring signals that can fully characterize the operating status of the unit, or constructing intelligent fault diagnosis models based on methods such as convolutional neural networks. However, there are still some problems that need to be improved: (1) Feature extraction-based methods require prior knowledge of signal processing technology and diagnostic expertise, and it is time-consuming and labor-intensive to produce these features; (2) Data-driven intelligent fault diagnosis models require a large number of monitoring signals and iteratively train the model to achieve high accuracy. However, in actual industrial scenarios, since mechanical equipment is in a healthy state for a long time, the number of fault state samples is very limited, resulting in high information repetition in the monitoring data and missing typical fault information, which makes it difficult to meet the assumption that deep learning requires a large number of labeled identically distributed fault samples.

[0004] Therefore, new technologies need to be introduced to accurately diagnose the status of the hydro-turbine generator units, to achieve precise monitoring of the operating status of the hydro-turbine generator units and fault warning, and to provide strong guarantees for the safe and stable operation of the hydropower station. Summary of the invention

[0005] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a hydro-generator fault diagnosis method based on maximum mean difference transfer learning induced by polynomial kernel, so as to realize accurate monitoring of the operating status of the hydro-generator unit and fault warning, and provide strong guarantee for the safe and stable operation of the hydropower station.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for fault diagnosis of a hydro-generator based on maximum mean difference transfer learning induced by a polynomial kernel, which comprises the following steps:

[0007] Step S1, collecting vibration data samples of the upper guide bearing of a hydro-turbine generator in normal and abnormal states that are similar to the actual equipment structure and parameters of the project as a source domain data set, and dividing the source domain into training samples and test samples;

[0008] Step S2, collecting measured upper guide bearing vibration data samples of the hydro-turbine generator in normal and abnormal states as target domain data sets, and dividing the target domain into training samples and test samples;

[0009] Step S3: linearly transform the source domain and target domain data to the range of [0, 1] using the minimum-maximum normalization method;

[0010] Step S4, using the maximum mean difference evaluation index induced by the polynomial kernel to measure the difference between the distribution of the source domain data set and the target domain data set, and using unbiased estimation technology to improve calculation efficiency;

[0011] Step S5, using a modulation factor that varies based on an iterative process to weigh the importance of the maximum mean difference loss and the classification cross entropy loss induced by the polynomial kernel, and constructing an overall loss function;

[0012] Step S6, constructing an unsupervised domain adaptive network based on the maximum mean difference evaluation index induced by the polynomial kernel and the modulation factor that changes based on the iterative process, inputting the training samples in the source domain data set and the target domain data set into the unsupervised domain adaptive network at the same time to extract transferable features, training to obtain a hydro-turbine generator transfer fault diagnosis model, and using the test samples in the target domain data set to verify the fault classification effect of the transfer fault diagnosis model;

[0013] Step S7: using the trained hydro-generator migration fault diagnosis model to identify the operating status of the hydro-generator set.

[0014] Furthermore, the step S4 specifically includes the following steps:

[0015] Step S41, using the polynomial kernel as the nonlinear mapping function kernel function of the maximum mean difference, mapping the source domain data set and the target domain data set to the reproducing kernel Hilbert high-dimensional inner product space to measure the difference in distribution between the two domains:

[0016]

[0017] in, is the source domain dataset, containing n1 source domain samples is the target domain dataset, containing n2 target domain samples It is an evaluation index based on the maximum mean difference induced by the polynomial kernel. is the reproducing kernel Hilbert high-dimensional inner product space, a, b, c represent the slope, intercept and order respectively;

[0018] Step S42: Under the premise that the number of samples of the source domain dataset and the target domain dataset is equal, in order to quickly and accurately use the maximum mean difference evaluation index induced by the polynomial kernel to evaluate the distribution difference between the source domain and the target domain in the domain adaptation process, an unbiased estimation method is used to reduce the computational complexity of the formula to linear and perform calculations, and a four-tuple is defined: It includes two source domain samples and two target domain samples. For each quadruple z i Use nonlinear function φ(·) for mapping:

[0019]

[0020] Then, the polynomial kernel function is selected as the mapping function and z i As units, the X calculated based on the maximum mean difference evaluation index induced by the unbiased estimate polynomial kernel is s With X t The distance is:

[0021]

[0022] Furthermore, the step S5 specifically includes the following steps:

[0023] Step S51: adopt a modulation factor that changes based on an iterative process, whose value changes dynamically as the number of iterations increases:

[0024]

[0025] Among them, λ C is a modulation factor that changes based on the iterative process, epoch is the current iteration number; max-epoch is the preset maximum iteration number; mid-epoch is the preset hyperparameter;

[0026] Step S52, introducing a modulation factor that varies based on the iterative process into the overall model loss function to weigh the importance of the maximum mean difference loss and the classification cross entropy loss induced by the polynomial kernel:

[0027]

[0028] Among them, L represents the overall loss function, represents the output value of the i-th source domain sample after the Softmax classification layer, Represents the label of the i-th sample.

[0029] Further, in step S6, the classification layer activation function of the unsupervised domain adaptation network is Softmax, and the pooling layer activation function is ReLu.

[0030] Beneficial effects of the present invention:

[0031] 1. The present invention adopts the maximum mean difference evaluation index induced by polynomial kernel to measure the difference between the distribution of source domain data set and target domain data set, which improves the robustness of the domain adaptation process, and adopts unbiased estimation technology to improve the calculation process of the maximum mean difference evaluation index induced by polynomial kernel, which improves the calculation efficiency and saves a lot of calculation cost; adopts a modulation factor that changes based on the iterative process to weigh the importance of the maximum mean difference loss induced by polynomial kernel and the classification cross entropy loss, and constructs an overall loss function. The model trained by iterative training can accurately extract the transferable features between the source domain and the target domain;

[0032] 2. The present invention solves the problem that the number of fault state samples in actual industrial scenarios is limited and difficult to use for deep learning. It can transfer the knowledge of the source domain to the target domain that has a certain correlation with it, improve the learning effect of the target domain, achieve optimal convergence and improve diagnostic accuracy, and complete reliable fault migration diagnosis tasks.

[0033] 3. The present invention realizes accurate monitoring of the operating status of the hydro-generator unit and fault warning, providing a strong guarantee for the safe and stable operation of the hydropower station. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A schematic diagram of a flow chart of a hydro-generator fault diagnosis method based on maximum mean difference transfer learning induced by polynomial kernel;

[0035] Figure 2 This is a structural flow chart of the unsupervised domain adaptation network based on convolutional neural network. DETAILED DESCRIPTION

[0036] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0037] Embodiment 1: A method for fault diagnosis of a hydro-generator based on maximum mean difference transfer learning induced by a polynomial kernel, comprising the following steps:

[0038] Step S1, collecting vibration data samples of the upper guide bearing of a hydro-turbine generator in normal and abnormal states that are similar to the actual equipment structure and parameters of the project as a source domain data set, and dividing the source domain into training samples and test samples;

[0039] Step S2, collecting measured upper guide bearing vibration data samples of the hydro-turbine generator in normal and abnormal states as target domain data sets, and dividing the target domain into training samples and test samples;

[0040] Step S3: linearly transform the source domain and target domain data to the range of [0, 1] using the minimum-maximum normalization method;

[0041] Step S4, using the maximum mean difference evaluation index induced by the polynomial kernel to measure the difference between the distribution of the source domain data set and the target domain data set, and using unbiased estimation technology to improve calculation efficiency;

[0042] Step S41, using the polynomial kernel as the nonlinear mapping function kernel function of the maximum mean difference, mapping the source domain data set and the target domain data set to the reproducing kernel Hilbert high-dimensional inner product space to measure the difference in distribution between the two domains:

[0043]

[0044] in, is the source domain dataset, containing n1 source domain samples is the target domain dataset, containing n2 target domain samples It is an evaluation index based on the maximum mean difference induced by the polynomial kernel. is the reproducing kernel Hilbert high-dimensional inner product space, a, b, c represent the slope, intercept and order respectively;

[0045] Step S42: Under the premise that the number of samples of the source domain dataset and the target domain dataset is equal, in order to quickly and accurately use the maximum mean difference evaluation index induced by the polynomial kernel to evaluate the distribution difference between the source domain and the target domain in the domain adaptation process, an unbiased estimation method is used to reduce the computational complexity of the formula to linear and perform calculations, and a four-tuple is defined: It includes two source domain samples and two target domain samples. For each quadruple z i Use nonlinear function φ(·) for mapping:

[0046]

[0047] Then, the polynomial kernel function is selected as the mapping function and z i As units, the X calculated based on the maximum mean difference evaluation index induced by the unbiased estimate polynomial kernel is s With X t The distance is:

[0048]

[0049] Step S5, using a modulation factor that varies based on an iterative process to weigh the importance of the maximum mean difference loss and the classification cross entropy loss induced by the polynomial kernel, and constructing an overall loss function;

[0050] Step S51: adopt a modulation factor that changes based on an iterative process, whose value changes dynamically as the number of iterations increases:

[0051]

[0052] Among them, λ C is a modulation factor that changes based on the iterative process, epoch is the current iteration number; max-epoch is the preset maximum iteration number; mid-epoch is the preset hyperparameter;

[0053] Step S52, introducing a modulation factor that varies based on the iterative process into the overall model loss function to weigh the importance of the maximum mean difference loss and the classification cross entropy loss induced by the polynomial kernel:

[0054]

[0055] Among them, L represents the overall loss function, represents the output value of the i-th source domain sample after the Softmax classification layer, Represents the label of the i-th sample.

[0056] Step S6, constructing an unsupervised domain adaptive network based on the maximum mean difference evaluation index induced by the polynomial kernel and the modulation factor that changes based on the iterative process, inputting the training samples in the source domain data set and the target domain data set into the unsupervised domain adaptive network at the same time to extract transferable features, training to obtain a hydro-turbine generator transfer fault diagnosis model, and using the test samples in the target domain data set to verify the fault classification effect of the transfer fault diagnosis model;

[0057] In step S6, the activation function of the classification layer of the unsupervised domain adaptation network is Softmax, and the activation function of the pooling layer is ReLu.

[0058] Step S7: using the trained hydro-generator migration fault diagnosis model to identify the operating status of the hydro-generator set.

[0059] Example 2: In this example, the vibration data samples of the upper guide bearing of the hydro-turbine generator with similar structure and parameters to the actual engineering equipment are used as the source domain data set, and the measured vibration data samples of the upper guide bearing of the hydro-turbine generator are used as the target domain data set, which are divided into a training set and a test set in a ratio of 7:3. As shown in Table 1.

[0060] Table 1 Source domain and target domain datasets

[0061]

[0062] Construct a hydropower generator unit migration fault diagnosis model based on an unsupervised domain adaptive network induced by a polynomial kernel; in this example, the structural flow chart of the unsupervised domain adaptive network based on a convolutional neural network is as follows: Figure 2 As shown in the figure, it contains two convolutional layers, two pooling layers, two fully connected layers and one classification layer. The model can extract features from source domain and target domain samples simultaneously; the specific network structure and parameters are shown in Table 2. As can be seen from the table, the proposed network contains two convolutional layers, two pooling layers, two fully connected layers and one classification layer. The model can extract features from source domain and target domain samples simultaneously. The hyperparameters are set as follows: in the polynomial kernel parameters, a and b are fixed to 1, c is searched from {1,2,3,4,5,6,7,8,9,10}, the batch size is 32, the weight decay coefficient λ=0.001, the learning rate η=0.01, the number of iterations t=120, the maximum number of iterations is max-epoch=2000, and the preset hyperparameter mid-epoch=30. Table 3 shows the comparison results of the method of the present invention with transfer component analysis, deep convolutional neural network, domain adaptive neural network and joint distribution adaptation method after ten independent runs. The input of the deep convolutional neural network, domain adaptive neural network and joint distribution adaptation is the same as the method of the present invention, which is the original time domain vibration signal. The input of the migration component analyzer is the frequency domain amplitude spectrum signal converted from the time domain vibration signal. Specifically analyzing the recognition result of one of the 10 runs of the invented method, the overall classification accuracy of the method of the present invention is 99.13%. The method proposed by the present invention has the highest average recognition accuracy and can effectively distinguish between normal and abnormal categories.

[0063] Table 2 Unsupervised domain adaptation network structure and parameters

[0064]

[0065] Table 3 Comparison of prediction results

[0066]

[0067] Reference Figure 1As shown, the content of the present invention can be mainly divided into three parts. The first part is to collect the upper guide bearing vibration data samples of the hydro-turbine generator in normal state and abnormal state which are similar to the actual engineering equipment structure and parameters as the source domain data set and divide the source domain training samples and test samples, collect the upper guide bearing vibration data samples of the hydro-turbine generator in normal state and abnormal state as the target domain data set and divide the target domain training samples and test samples, and divide them into the training samples and test samples of the source domain and the target domain respectively after removing the mean and normalization processing; the second part is to design the maximum mean difference metric index induced by polynomial kernel, use the unbiased estimation technology to reduce the calculation complexity of the designed metric index, and finally design the maximum mean difference metric index induced by polynomial kernel based on unbiased estimation; the third part is to design the modulation factor which changes based on the iterative process, and finally construct the overall loss function; the fourth part is to test the trained hydro-turbine generator group migration fault diagnosis model based on the unsupervised domain adaptive network induced by polynomial kernel by the target domain test sample.

[0068] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A method for fault diagnosis of hydro-generator based on maximum mean difference transfer learning induced by polynomial kernel, characterized in that: It includes the following steps: Step S1, collecting vibration data samples of the upper guide bearing of a hydro-turbine generator in normal and abnormal states that are similar to the actual equipment structure and parameters of the project as a source domain data set, and dividing the source domain into training samples and test samples; Step S2, collecting measured upper guide bearing vibration data samples of the hydro-turbine generator in normal and abnormal states as target domain data sets, and dividing the target domain into training samples and test samples; Step S3: linearly transform the source domain and target domain data to the range of [0, 1] using the minimum-maximum normalization method; Step S4, using the maximum mean difference evaluation index induced by the polynomial kernel to measure the difference between the distribution of the source domain data set and the target domain data set, and using unbiased estimation technology to improve calculation efficiency; Step S5, using a modulation factor that varies based on an iterative process to weigh the importance of the maximum mean difference loss and the classification cross entropy loss induced by the polynomial kernel, and constructing an overall loss function; Step S6, constructing an unsupervised domain adaptive network based on the maximum mean difference evaluation index induced by the polynomial kernel and the modulation factor that changes based on the iterative process, inputting the training samples in the source domain data set and the target domain data set into the unsupervised domain adaptive network at the same time to extract transferable features, training to obtain a hydro-turbine generator transfer fault diagnosis model, and using the test samples in the target domain data set to verify the fault classification effect of the transfer fault diagnosis model; Step S7: using the trained hydro-generator migration fault diagnosis model to identify the operating status of the hydro-generator set.

2. A method for fault diagnosis of a hydro-turbine generator based on maximum mean difference transfer learning induced by a polynomial kernel according to claim 1, characterized in that: The step S4 specifically comprises the following steps: Step S41, using the polynomial kernel as the nonlinear mapping function kernel function of the maximum mean difference, mapping the source domain data set and the target domain data set to the reproducing kernel Hilbert high-dimensional inner product space to measure the difference in distribution between the two domains: in, is the source domain dataset, containing n1 source domain samples is the target domain dataset, containing n2 target domain samples It is an evaluation index based on the maximum mean difference induced by the polynomial kernel. is the reproducing kernel Hilbert high-dimensional inner product space, a, b, c represent the slope, intercept and order respectively; Step S42: Under the premise that the number of samples of the source domain dataset and the target domain dataset is equal, in order to quickly and accurately use the maximum mean difference evaluation index induced by the polynomial kernel to evaluate the distribution difference between the source domain and the target domain in the domain adaptation process, an unbiased estimation method is used to reduce the computational complexity of the formula to linear and perform calculations, and a four-tuple is defined: It includes two source domain samples and two target domain samples. For each quadruple z i Use nonlinear function φ(·) for mapping: Then, the polynomial kernel function is selected as the mapping function and z i As units, the X calculated based on the maximum mean difference evaluation index induced by the unbiased estimate polynomial kernel is s With X t The distance is:

3. A method for fault diagnosis of a hydro-generator based on maximum mean difference transfer learning induced by a polynomial kernel according to claim 2, characterized in that: The step S5 specifically includes the following steps: Step S51: adopt a modulation factor that changes based on an iterative process, whose value changes dynamically as the number of iterations increases: Among them, λ C is a modulation factor that changes based on the iterative process, epoch is the current iteration number; max-epoch is the preset maximum iteration number; mid-epoch is the preset hyperparameter; Step S52, introducing a modulation factor that varies based on the iterative process into the overall model loss function to weigh the importance of the maximum mean difference loss and the classification cross entropy loss induced by the polynomial kernel: Among them, L represents the overall loss function, represents the output value of the i-th source domain sample after the Softmax classification layer, Represents the label of the i-th sample.

4. The method for fault diagnosis of a hydro-generator based on maximum mean difference transfer learning induced by a polynomial kernel according to claim 1, characterized in that: In step S6, the activation function of the classification layer of the unsupervised domain adaptation network is Softmax, and the activation function of the pooling layer is ReLu.

Images