Wind turbine generator bearing fault diagnosis method based on semi-supervised time series learning

Abstract

A kind of wind turbine bearing fault diagnosis method based on semi-supervised time series learning, the method comprises the following steps: a. using wind turbine bearing original vibration signal to construct training set and test set;B. establish one-dimensional convolutional neural network;C. construct semi-supervised learning model;D. semi-supervised learning model is trained using MAML;E. the labeled data in test set is input to the semi-supervised learning model after training, and semi-supervised learning model is tested;F. with the semi-supervised learning model that has been tested qualified, the type of wind turbine bearing fault is judged.The type of wind turbine bearing fault is judged using semi-supervised learning model in the present application, can effectively utilize the limited labeled data and a large number of unlabeled data to train model, very good to solve the problem of insufficient fault data, reduce the difficulty of bearing fault intelligent diagnosis, improve the accuracy of bearing fault diagnosis result.

Description

Technical Field

[0001] This invention relates to a method for diagnosing bearing failures in wind turbine generator sets, belonging to the field of generator technology. Background Technology

[0002] Condition monitoring of wind turbine generators has always been a focus of wind turbine research. Due to their large size and harsh operating conditions, wind turbine generators are prone to failure, leading to significant maintenance and operating costs. Therefore, closely monitoring the condition of wind turbine generators and reducing their failure rate is crucial. Bearings are key components of wind turbine generators and are widely used in modern large-scale machinery. Statistics show that approximately 30% of mechanical failures in rotating machinery using rolling bearings are caused by bearings. In recent years, deep learning-based intelligent fault diagnosis methods for rolling bearings have emerged, providing powerful tools for monitoring the safe operation of mechanical equipment.

[0003] Chinese patent application CN103900816A discloses a method for diagnosing bearing faults in wind turbine generators. The method includes the following steps: acquiring the vibration signal of the bearing; performing three-level decomposition of the vibration signal using wavelet packet analysis, performing soft-threshold quantization on the high-frequency coefficients of each decomposition, and performing one-dimensional wavelet reconstruction based on the lowest-level low-frequency and high-frequency coefficients of the wavelet decomposition; performing wavelet packet decomposition on the reconstructed vibration signal, extracting the energy of each frequency band in the third layer, and using the energy of each frequency band to form a fault feature input vector as the fault diagnosis input vector for a BP neural network, thus constructing a three-level BP neural network; acquiring feature input vector samples from historical fault data and inputting them into the three-level BP neural network for training; acquiring fault diagnosis feature vectors from real-time bearing operation data and inputting them into the trained BP neural network; and achieving intelligent diagnosis of bearing fault types. This invention can accurately diagnose bearing fault types and precisely determine fault locations. However, its diagnostic method requires a large amount of historical fault data. With the continuous improvement of bearing processing precision and material properties, bearing service life is constantly extended, and bearings are replaced promptly when damage is detected, resulting in insufficient bearing fault data and only a large amount of unlabeled operation data.

[0004] In recent years, deep learning has made significant progress in the field of wind turbine bearing fault diagnosis. Its powerful feature learning and adaptive feature extraction capabilities have the potential to make fault diagnosis more intelligent and accurate. However, the operating environment of bearings is often variable, and fault conditions are complex, with different fault locations and types, making it extremely difficult to extract explicitly separable deep features, thus increasing the challenge of intelligent bearing fault diagnosis. Therefore, effectively utilizing limited labeled data and a large amount of unlabeled data to diagnose rolling bearing faults in wind turbines is of great significance. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a wind turbine generator bearing fault diagnosis method based on semi-supervised time series learning, thereby reducing the difficulty of intelligent bearing fault diagnosis and improving the accuracy of bearing fault diagnosis results.

[0006] The problem described in this invention is solved by the following technical solution:

[0007] A method for diagnosing bearing faults in wind turbine generators based on semi-supervised temporal learning, the method comprising the following steps:

[0008] a. Construct training and test sets using the original vibration signals from wind turbine generator bearings:

[0009] The original vibration signals of the wind turbine generator bearings are divided into training sets and test sets according to different operating conditions, different sampling frequencies, different fault degrees, and different fault types. The training set contains labeled data and unlabeled data, while the test set contains only labeled data.

[0010] b. Establish a one-dimensional convolutional neural network:

[0011] A one-dimensional convolutional neural network includes convolutional layers, pooling layers, and fully connected layers. The convolutional layers consist of convolutional kernels, which perform convolution operations on the one-dimensional input data in the following manner:

[0012]

[0013] In the formula: σ represents the activation function of the convolutional layer; w is the weight of the convolutional kernel; x is the input; This is a one-dimensional convolution operation; b is the bias; c is the output of the convolution operation;

[0014] c. Constructing a semi-supervised learning model:

[0015] Using a one-dimensional convolutional neural network as the base network, a semi-supervised learning mechanism is introduced into the one-dimensional convolutional neural network to construct a semi-supervised learning model. This semi-supervised learning mechanism includes an encoder and a decoder. The encoder consists of four convolutional layers and four pooling layers. For each convolutional layer and max-pooling layer in the encoder, there is a corresponding deconvolutional layer and depooling layer in the decoder. The output of the deconvolution operation in the deconvolutional layer is defined as:

[0016]

[0017] In the formula: σ′ represents the activation function of deconvolution, is the deconvolution output of input x, and w′ is the filter of the deconvolution layer;

[0018] d. Use MAML to train a semi-supervised learning model:

[0019] The semi-supervised learning model is trained using MAML. Labeled and unlabeled data from the training set are input into the semi-supervised learning model to complete the fault identification and data reconstruction tasks, respectively. During the meta-training of MAML, the model parameters are updated through one or more gradient descent steps in the inner and outer loops.

[0020] e. Input the labeled data from the test set into the trained semi-supervised learning model and test the semi-supervised learning model;

[0021] f. Use a semi-supervised learning model that has been tested and approved to determine the type of bearing failure in the wind turbine generator set.

[0022] The above-mentioned wind turbine bearing fault diagnosis method based on semi-supervised temporal learning adds a temporal module to the parallel position of each layer of the one-dimensional convolutional neural network in the semi-supervised learning model. As the depth of the one-dimensional convolutional neural network extends, the temporal module also extends layer by layer. The temporal modules of different layers are combined together through dot product operation.

[0023] The above-mentioned wind turbine bearing fault diagnosis method based on semi-supervised temporal learning uses the following training process for training the semi-supervised learning model using MAML:

[0024] First, a certain number of samples are randomly sampled from each type of fault in the training set to form a task set. Then, samples are randomly sampled multiple times from the task set to form different tasks. Finally, the inner loop is updated. The inner loop parameter update process for a single task is defined as follows:

[0025]

[0026] In the formula: α is a hyperparameter, θ is the model parameter, and θ′ is the result of θ being updated once in the inner loop. Indicates the current task. It is the loss on the output f(θ). Perform gradient calculation;

[0027] The optimization objective of meta-learning is a process of summing up the updates of parameters for a single task, defined as follows:

[0028]

[0029] In the formula: Represents task set Follows the probability distribution P;

[0030] During the construction of the inner loop task set, one sample from each class is randomly selected from outside the inner loop task set of the training set to form the outer loop task set. The meta-optimization of the outer loop is described as follows:

[0031]

[0032] In the formula: β represents the gradient update learning rate of the outer loop.

[0033] Beneficial effects

[0034] This invention employs a semi-supervised learning model to determine the type of bearing fault in wind turbine generators. It effectively utilizes both limited labeled data and a large amount of unlabeled data to train the model, thus solving the problem of insufficient fault data, reducing the difficulty of intelligent bearing fault diagnosis, and improving the accuracy of bearing fault diagnosis results. The technical solution of this invention offers the following advantages:

[0035] a. By integrating semi-supervised learning and meta-learning mechanisms, typical 1D-CNNs can improve information mining and feature learning capabilities on limited datasets: the semi-supervised learning mechanism enables the model to extract features from a large amount of unlabeled data, thereby improving the accuracy of fault diagnosis; the meta-learning mechanism can enhance the feature learning ability of deep learning on limited datasets, improve the model's generalization ability, and is beneficial for diagnosing bearing faults.

[0036] b. The temporal module used in this invention can memorize useful features during training, preventing the disappearance of weak temporal features during deep learning and transmission. Attached Figure Description

[0037] The invention will now be described in further detail with reference to the accompanying drawings.

[0038] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0039] Figure 2 Time-domain plots of various fault conditions of rolling bearings under 100 rpm operating conditions;

[0040] Figure 3 a- Figure 3 b is the confusion matrix diagram for fault diagnosis under different operating conditions, where Figure 3 a represents the operating condition at 100 rpm; Figure 3 b represents the operating condition at 600 rpm; Figure 3 c represents the operating condition at 1800 rpm;

[0041] Figure 4 This is a visualization of the clustering results for fault diagnosis of various bearings under 100rpm operating conditions.

[0042] The symbols in the diagram have the following meanings: NC means normal, IF means inner ring fault, CF means cage fault, BF means rolling element fault, OF-3 means outer ring fault at the 3 o'clock position, OF-6 means outer ring fault at the 6 o'clock position, and OF-12 means outer ring fault at the 12 o'clock position. Detailed Implementation

[0043] In view of the problems existing in the prior art, the method provided by the present invention can effectively extract the deep element features of rolling bearing failure under variable operating conditions, thereby realizing the accurate identification of bearing failure status of wind turbine generator sets.

[0044] like Figure 1 As shown, the present invention includes the following steps:

[0045] a. First, in order to meet the requirements of the meta-learning task, the original vibration signals of the wind turbine bearings are constructed into training and testing sets according to different working conditions, different sampling frequencies, different fault degrees, and different fault types. The training set contains a limited amount of labeled data and a relatively large amount of unlabeled data.

[0046] b. Establish a one-dimensional convolutional neural network (1D-CNN) as the base network to construct a prototype for semi-supervised learning. As the foundational network for SSTML, 1D-CNN consists of three main parts: convolutional layers, pooling layers, and fully connected layers. The convolutional layers consist of convolutional kernels, which perform convolution operations on the input data. Due to the one-dimensional input, convolution is designed as a one-dimensional operation. Given the convolutional kernel and the input samples, the one-dimensional convolution operation can be determined as follows:

[0047]

[0048] In the formula: σ represents the activation function of the convolutional layer; w is the weight of the convolutional kernel; x is the input; is a one-dimensional convolution operation; b is the bias; c is the output of the convolution operation.

[0049] c. Supervised 1D-CNNs cannot directly utilize unlabeled data, which wastes a significant amount of unlabeled data in practical industrial applications. Therefore, a semi-supervised learning mechanism (consisting of an encoder and a decoder) is designed into 1D-CNNs. Semi-supervised learning is required to perform fault identification (using limited labeled data) and data reconstruction tasks (using a relatively large amount of unlabeled data). In this design, the encoder consists of four convolutional layers and four pooling layers, and the decoder's structure is the reverse of the encoder's for each layer.

[0050] According to formula (1), the output of the deconvolution operation is defined as:

[0051]

[0052] In the formula: σ′ represents the activation function of deconvolution, is the deconvolution output of input x, and w′ is the filter of the deconvolution layer.

[0053] d. MAML is a model-independent meta-learning method that accelerates the training of deep learning models and improves their performance with limited training samples. During the meta-training of MAML, the model parameters are updated through one or more gradient descent steps in the inner and outer loops. MAML is used to train a semi-supervised learning model. According to the MAML training flow, a certain number of samples are first randomly sampled from each fault class in the training set to form a task set. Then, samples are randomly sampled multiple times from the task set to form different tasks, and the inner loop update is then completed. The parameter update process of the inner loop for a single task can be defined as follows:

[0054]

[0055] In the formula: α is a hyperparameter, θ is the model parameter, and θ′ is the result of θ being updated once in the inner loop. Indicates the current task. It is the loss on the output f(θ). Perform gradient calculation.

[0056] In practice, the model parameters are computed across all tasks, so the optimization objective of meta-learning is a process of summing up the parameter updates for individual tasks, defined as follows:

[0057]

[0058] In the formula: Represents task set It follows a probability distribution P.

[0059] During the construction of the inner loop task set, one sample from each class is randomly selected from outside the task set of the training set to form the outer loop task set. The meta-optimization description of the outer loop is as follows:

[0060]

[0061] In the formula: β represents the gradient update learning rate of the outer loop.

[0062] e. The temporal module is proposed to memorize temporal features in deep networks. As the model depth increases, the output relationships between different layers become weaker. The temporal module is designed at the parallel position of each layer of the basic 1D-CNN network in step b. As the 1D-CNN extends in depth, the temporal module also extends layer by layer. Temporal modules from different layers are combined through dot product operations, thus enabling the memorization of features from each layer. These features reflect the sequential relationship of model parameters during transmission, hence the term "temporal features." The temporal module is an optimization of 1D-CNN, and together with 1D-CNN, they form the network skeleton of the fault diagnosis model.

[0063] f. Input the labeled data from the test set into the trained semi-supervised learning model and test the semi-supervised learning model;

[0064] g. Use a semi-supervised learning model that has been tested and approved to determine the type of bearing failure in wind turbine generator sets.

[0065] The following is an analysis of bearing fault signals.

[0066] The invention was validated using bearing failure data, which consisted of bearing failure signals collected on a wind turbine drive system test bench. Detailed data usage is shown in Table 1, where the laboratory bearing data sampling frequency was f = 12.8 kHz.

[0067] Table 1. Bearing data for laboratory wind turbine generators

[0068]

[0069] The time-domain waveform of the bearing multi-fault test signal at 100 rpm is as follows: Figure 2 As shown, from Figure 2 As can be seen, the time-domain signal alone is insufficient to distinguish effective fault characteristics, thus making it impossible to determine the presence of a fault in the bearing.

[0070] The specific implementation process of analyzing bearing signals using this invention is as follows:

[0071] First, in order to meet the requirements of the meta-learning task, the original vibration signals were constructed into a task set according to different fault types, and a one-dimensional convolutional neural network (1D-CNN) was established as the basic network to build a prototype of semi-supervised learning, which serves as the basic network of SSTML.

[0072] Then, a temporal module is designed to enhance the performance of the supervised learning part in semi-supervised learning. This module can memorize useful features during training to prevent the disappearance of weak temporal features during deep learning transmission.

[0073] Finally, the semi-supervised model trained using MAML yielded the following classification results: Figure 3 a- Figure 3 As shown in c: In Figure 3 In category a, only BF had a classification accuracy of 90%, while the others had 100%; considering multiple working conditions... Figure 3 a- Figure 3 c. The classification accuracy of each component exceeds 95%. Therefore, this invention analyzes the experimental signals of rolling bearings, and the results show that it can accurately identify the fault conditions of rolling bearings in a limited labeled dataset, effectively mine deep meta-feature information, and achieve accurate diagnosis of bearing fault conditions.

Claims

1. A method for diagnosing bearing faults in wind turbine generators based on semi-supervised temporal learning, characterized in that, The method includes the following steps: a. Construct training and test sets using the original vibration signals from wind turbine generator bearings: The original vibration signals of the wind turbine generator bearings are divided into training sets and test sets according to different operating conditions, different sampling frequencies, different fault degrees, and different fault types. The training set contains labeled data and unlabeled data, while the test set contains only labeled data. b. Establish a one-dimensional convolutional neural network: A one-dimensional convolutional neural network includes convolutional layers, pooling layers, and fully connected layers. The convolutional layers consist of convolutional kernels, which perform convolution operations on the one-dimensional input data in the following manner: In the formula: σ represents the activation function of the convolutional layer; w is the weight of the convolutional kernel; x is the input; This is a one-dimensional convolution operation; b is the bias; c is the output of the convolution operation; c. Constructing a semi-supervised learning model: Using a one-dimensional convolutional neural network as the base network, a semi-supervised learning mechanism is introduced into the one-dimensional convolutional neural network to construct a semi-supervised learning model. This semi-supervised learning mechanism includes an encoder and a decoder. The encoder consists of four convolutional layers and four pooling layers. For each convolutional layer and max-pooling layer in the encoder, there is a corresponding deconvolutional layer and depooling layer in the decoder. The output of the deconvolution operation in the deconvolutional layer is defined as: In the formula: σ′ represents the activation function of deconvolution, is the deconvolution output of input x, and w′ is the filter of the deconvolution layer; d. Use MAML to train a semi-supervised learning model: The semi-supervised learning model is trained using MAML. Labeled and unlabeled data from the training set are input into the semi-supervised learning model to complete the fault identification and data reconstruction tasks, respectively. During the meta-training of MAML, the model parameters are updated through one or more gradient descent steps in the inner and outer loops. e. Input the labeled data from the test set into the trained semi-supervised learning model and test the semi-supervised learning model; f. Use a semi-supervised learning model that has been tested and approved to determine the type of bearing failure in the wind turbine generator set.

2. The method for diagnosing wind turbine bearing faults based on semi-supervised temporal learning according to claim 1, characterized in that, Temporal modules are added to the parallel positions of each layer of the one-dimensional convolutional neural network in the semi-supervised learning model. As the depth of the one-dimensional convolutional neural network increases, the temporal modules also extend layer by layer. Temporal modules from different layers are combined together through dot product operations.

3. A method for diagnosing wind turbine generator bearing faults based on semi-supervised time series learning according to claim 1 or 2, characterized in that, The training process for training a semi-supervised learning model using MAML is as follows: First, a certain number of samples are randomly sampled from each type of fault in the training set to form a task set. Then, samples are randomly sampled multiple times from the task set to form different tasks. Finally, the inner loop is updated. The inner loop parameter update process for a single task is defined as follows: In the formula: α is a hyperparameter, θ is the model parameter, and θ′ is the result of θ being updated once in the inner loop. Indicates the current task. It is the loss on the output f(θ). Perform gradient calculation; The optimization objective of meta-learning is a process of summing up the updates of parameters for a single task, defined as follows: In the formula: Represents task set Follows the probability distribution P; During the construction of the inner loop task set, one sample from each class is randomly selected from outside the inner loop task set of the training set to form the outer loop task set. The meta-optimization of the outer loop is described as follows: In the formula: β represents the gradient update learning rate of the outer loop.

Images