A Fault Diagnosis Method for Wind Turbines

Abstract

The present invention provides a method for fault diagnosis of a wind turbine, specifically a method for fault diagnosis of a wind turbine using an early stopping optimized set expansion convolutional neural network (ESO-EDCNN). The method includes the following steps: First, vibration signals of the wind turbine in different operating states are collected using an acceleration sensor, and the collected signals are randomly shuffled and divided into a training set and a test set. Then, the training set is divided into a training subset and a validation subset, which are used to train and validate the ESO-EDCNN model respectively. Finally, the test set is input into the trained model to achieve intelligent fault diagnosis of the wind turbine. The present invention overcomes the problems of low reliability and low accuracy of a single diagnosis model, and solves the problems of long training time and easy overfitting of existing deep learning models. The experimental results of wind turbine fault diagnosis show that the proposed method can accurately identify various single / composite fault types and has great practical value.

Description

Technical Field

[0001] The present invention relates to the technical field of rotating machinery fault diagnosis and artificial intelligence, and particularly to a fault diagnosis method for wind turbines. Background Technique

[0002] Wind energy has a short construction period and flexible investment scale, and is considered the clean energy with the most commercial potential and development prospects. However, harsh environmental conditions make wind turbines prone to failures. A minor fault in a wind turbine may lead to the overall failure of the entire machine, resulting in economic and efficiency losses. Therefore, early fault diagnosis of wind turbines is of great significance.

[0003] Currently, traditional fault diagnosis methods for wind turbines are signal processing-based fault diagnosis methods, machine learning-based fault diagnosis methods, and deep neural network-based fault diagnosis methods. Signal processing-based fault diagnosis methods usually use techniques such as wavelet transform, empirical mode decomposition, and local mean decomposition to perform time-frequency analysis on the collected fault signals to find the fault characteristic frequency components, thereby realizing fault diagnosis. However, this type of method has weak feature extraction ability and requires technicians to have strong expert knowledge. Machine learning-based fault diagnosis methods manually extract time-domain features, frequency-domain features, or time-frequency domain features from the collected signals, and then use machine learning methods such as random forest, support vector machine, or KNN to realize fault diagnosis. However, this type of method requires manual feature extraction and has weak adaptability, which is not conducive to practical applications in wind turbine fault diagnosis. Deep neural network-based fault diagnosis methods can automatically extract features and achieve end-to-end fault diagnosis, with strong adaptability. However, the diagnosis results relying on a single diagnosis model lack reliability and are difficult to make accurate diagnosis decisions under complex working conditions. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a fault diagnosis method for wind turbines, which gives full play to the advantages of the deep neural network algorithm in adaptive fault diagnosis and greatly increases the feature extraction ability of the model.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions: A fault diagnosis method for wind turbines, comprising the following steps:

[0006] Step 1: Use an acceleration sensor to collect the vibration acceleration signals of the wind turbine under normal, single-fault, and compound-fault conditions;

[0007] Step 2: Randomly shuffle the signal sample sets collected in each state and divide them into a training set and a test set;

[0008] Step 3: Divide the training set into a training subset and a validation subset for training and validating the proposed ESO-EDCNN model;

[0009] Step 4: Input the test set into the trained ESO-EDCNN model to output the fault diagnosis result.

[0010] In a preferred embodiment, the vibration acceleration signal in Step 1 is the radial vibration acceleration signal along the transmission shaft.

[0011] In a preferred embodiment, the ESO-EDCNN algorithm in Step 3 is as follows

[0012] The EDCNN algorithm consists of multiple sub-classifiers; each sub-classifier has a different dilated convolutional kernel; each sub-classifier is composed of three stacked feature extractors and a classifier;

[0013] The three-layer fully connected neural network serves as the classifier of the diagnostic model to convert the feature map into a fault decision; and each feature extractor includes convolution calculation, activation calculation, normalization calculation, and pooling calculation; there is no need to preprocess the original data, and the collected vibration signal is directly used as the model output; the principles of convolution calculation, activation calculation, normalization calculation, and pooling calculation are as follows:

[0014]

[0015] In the formula, is the j-th element of the n-th convolutional layer, M j represents the convolution region of the input signal; i represents the position within the convolutional kernel and weight matrix, is M j the input of the previous layer, and are the weight matrix and bias of the convolutional kernel; the activation function selects the rectified linear unit ReLU or the exponential linear unit ELU non-linear function;

[0016] The pooling calculation is expressed as:

[0017]

[0018] In the formula, is the pooling calculation value of the k-th element at (i,j) in the l-th layer, is the pooling region within (i,j), is the input node within the pooling region (m,n); the pooling form selects average pooling or max pooling.

[0019] In a preferred embodiment, the process of early stopping optimization ESO is as follows:

[0020] During each iteration, when the loss function of the validation subset decreases, the model is saved; when the evaluation metric of the model no longer improves and the number of iterations is within the early stopping optimization range, the training process is stopped; the early stopping optimization evaluation criterion expression is:

[0021]

[0022] where t is the number of iterations, L obt (t) is the loss function of the obtained validation subset, and L va (t') is the loss function of the validation subset corresponding to the time t'. The loss function can be selected as the cross-entropy loss function or the mean squared error loss function.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] (1) The present invention gives full play to the advantages of the deep neural network algorithm in adaptive fault diagnosis.

[0025] (2) The present invention introduces an ensemble learning algorithm for comprehensive diagnostic decision-making, making the fault diagnosis results more reliable and accurate.

[0026] (3) The present invention introduces a dilated convolutional neural network into the ensemble model, greatly increasing the feature extraction ability of the model.

[0027] (4) The present invention introduces an early stopping mechanism to save the training time of the model and prevent the degradation of the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flowchart of the diagnostic method of the preferred embodiment of the present invention.

[0029] Figure 2 It is the ensemble learning mechanism of the preferred embodiment of the present invention.

[0030] Figure 3 It is the convolution kernel of different sub-classifiers of the preferred embodiment of the present invention.

[0031] Figure 4 It is a flowchart of the ESO method of the preferred embodiment of the present invention.

[0032] Figure 5 It is the fault recognition result of the EDCNN of the preferred embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0034] It should be noted that the following detailed description is illustrative and aims to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present application pertains.

[0035] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should also be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0036] A fault diagnosis method for a wind turbine, referring to Figures 1 to 5 , includes the following steps:

[0037] Step 1: Use an acceleration sensor to collect the vibration acceleration signals of the wind turbine under normal, single-fault, and compound-fault conditions;

[0038] Step 2: Randomly shuffle the signal sample sets collected in each state and divide them into a training set and a test set;

[0039] Step 3: Divide the training set into a training subset and a validation subset for training and validating the proposed ESO-EDCNN model;

[0040] Step 4: Input the test set into the trained ESO-EDCNN model to output the fault diagnosis result.

[0041] The vibration acceleration signal described in Step 1 is the radial vibration acceleration signal along the transmission shaft.

[0042] The process of the ESO-EDCNN algorithm described in Step 3 is as follows:

[0043] The EDCNN algorithm is a deep neural network method that extracts features through a multi-layer convolutional neural network to achieve a fault diagnosis method with strong adaptability and end-to-end. The EDCNN algorithm consists of multiple sub-classifiers. Each sub-classifier has different dilated convolutional kernels to extract different feature elements of the vibration signal. Each sub-classifier is composed of three stacked feature extractors and a classifier.

[0044] A three-layer fully connected neural network is used as the classifier of the diagnostic model to convert the feature map into a fault decision. And each feature extractor includes convolutional calculation, activation calculation, normalization calculation, and pooling calculation. It is not necessary to preprocess the original data, and the collected vibration signal is directly used as the model output. The principles of convolutional calculation, activation calculation, normalization calculation, and pooling calculation are as follows:

[0045]

[0046] Wherein, is the j-th element of the n-th convolutional layer, and M j represents the convolution region of the input signal. i represents the position within the convolution kernel and weight matrix. is M j the input of the previous layer, and are the weight matrix and bias of the convolution kernel. The activation function generally selects a non-linear function such as the rectified linear unit (ReLU) or the exponential linear unit (ELU).

[0047] The pooling calculation is expressed as:

[0048]

[0049] Wherein, is the pooling calculation value of the k-th element at (i, j) in the l-th layer, is the pooling region within (i, j), is the input node within the pooling region (m, n). The pooling form generally selects average pooling or max pooling.

[0050] The early stopping optimization (ES0) process is as follows

[0051] The early stopping method is used to avoid overfitting of the model and save the model calculation and training time. This method is an early stopping optimization based on the loss function of the validation subset. In each iteration process, when the loss function of the validation subset decreases, the model will be saved. When the evaluation index of the model no longer improves and the number of iterations is within the early stopping optimization range, the training process is stopped. The early stopping optimization can greatly reduce the calculation cost and effectively avoid the degradation of the model generalization ability. The evaluation criterion expression of this early stopping optimization is:

[0052]

[0053] where t is the number of iterations, and L obt (t) is the loss function of the obtained validation subset, and L va (t′) is the loss function of the validation subset corresponding to the time t′. The loss function can be selected as the cross-entropy loss function or the mean squared error loss function.

[0054] To verify the effectiveness of the proposed ESO-EDCNN fault diagnosis method, it is applied to the signal analysis process of the wind turbine power generation simulation test bench. The vibration acceleration signals of 9 fault conditions such as normal, single faults (inner race fault of bearing, outer race fault of bearing, rolling element fault of bearing, gear wear fault, and gear tooth breakage fault), and compound faults (inner and outer race faults of bearing, outer race and gear wear faults, outer race and gear tooth breakage faults) are collected by using acceleration sensors. The sampling frequency is f s = 12,800 Hz, and the number of sampling points is set to 1024.

[0055] The specific process of the wind turbine fault diagnosis method based on the proposed early stopping optimized set expansion convolutional neural network (ESO-EDCNN) of the present invention is as Figure 1 shown, and the specific steps are as follows.

[0056] Step 1: Data acquisition

[0057] In this step, the acceleration sensor is used to collect the experimental data of the wind turbine equipment to obtain its vibration acceleration signal. It is impossible to effectively distinguish each fault type only from the collected time-domain signal, and there is a certain background noise interference in the experimentally collected signal.

[0058] Step 2: Sample division

[0059] In this step, each of the 800 groups of samples of the above 9 conditions is labeled (0 - 8) and the dataset is divided. The ratio of the training set to the test set is set to 4:1, and the ratio of the training subset to the validation subset is set to 4:1.

[0060] Step 3: Model construction

[0061] In this step, the proposed ESO-EDCNN method has high self-adaptability and does not require preprocessing of the dataset.

[0062] The proposed EDCNN is set to have four sub-classifiers, as Figure 2 shown. Each sub-classifier has three feature extractors and one classifier. Each feature extraction block contains a convolutional layer, an activation layer, a batch normalization layer, and a pooling layer. The classifier is a three-layer fully connected neural network.

[0063] The hyperparameters of the diagnostic model are set as follows: the effective length of the convolutional kernel of each sub-classifier is set to 5, the stride is 1, but the convolutional regions are inconsistent, as Figure 3 shown. ELU is selected as the activation function and max pooling is used as the pooling function. The length of the pooling kernel is 2 and the stride is 2. The dimension of the hidden layer in the classifier is 128. The early stopping optimization threshold is set to 5.

[0064] The neural network model is trained on the training subset, and the model parameters are updated with each iteration. When the model update threshold exceeds the set early stopping optimization threshold, the training is immediately stopped to obtain the trained model.

[0065] Step 4: Pattern recognition

[0066] In this step, the data of the test set is input into the trained model to obtain the fault diagnosis result. As Figure 5 shown in the confusion matrix, all fault types are identified with 100% accuracy, indicating the reliability and robustness of the proposed deep learning model in the fault diagnosis of wind turbines.

Claims

1. A fault diagnosis method for a wind turbine, characterized in that It includes the following steps: Step 1: Use an acceleration sensor to collect the vibration acceleration signals of the wind turbine under normal, single-fault, and compound-fault conditions; Step 2: Randomly shuffle the signal sample sets collected in each state and divide them into a training set and a test set; Step 3: Divide the training set into a training subset and a validation subset for training and validating the proposed ESO-EDCNN model; Step 4: Input the test set into the trained ESO-EDCNN model and output the fault diagnosis result; The ESO-EDCNN algorithm process in Step 3 is as follows The EDCNN algorithm consists of multiple sub-classifiers; each sub-classifier has a different dilated convolutional kernel; each sub-classifier is composed of three stacked feature extractors and a classifier; A three-layer fully connected neural network is used as the classifier of the diagnostic model to convert the feature map into a fault decision; and each feature extractor includes convolutional calculation, activation calculation, normalization calculation, and pooling calculation; there is no need to preprocess the original data, and the collected vibration signals are directly used as the model output; the principles of convolutional calculation, activation calculation, normalization calculation, and pooling calculation are as follows: In the formula, is the j-th element of the n-th convolutional layer, and M j represents the convolutional region of the input signal; i represents the position within the convolutional kernel and the weight matrix, is the input of the previous layer of M j , and are the weight matrix and bias of the convolutional kernel; the activation function uses the rectified linear unit ReLU or the exponential linear unit ELU non-linear function; The pooling calculation is expressed as: Wherein, is the pooling calculation value of the k-th element at (i, j) in the l-th layer, is the pooling area within (i, j), is the input node within the pooling area (m, n); the pooling form is selected as average pooling or max pooling; The process of early stopping to optimize ESO is as follows: During each iteration, when the loss function of the validation subset decreases, the model is saved; when the evaluation index of the model no longer improves and the number of iterations is within the range of early stopping optimization, the training process is stopped; the expression of the early stopping optimization evaluation criterion is: where t is the number of iterations, and L obt (t) is the loss function of the obtained validation subset, and L va (t') is the loss function of the validation subset corresponding to the time t'.

2. The method for diagnosing faults of a wind turbine according to claim 1, wherein The vibration acceleration signal described in Step 1 is the vibration acceleration signal along the radial direction of the transmission shaft.

Images