Audio recognition and fault diagnosis method for mechanical faults of wind turbine generator sets

Abstract

The present invention belongs to the technical field of wind turbine generator fault diagnosis, specifically to a method for audio recognition and fault diagnosis of mechanical faults of wind turbine generator sets, including fault signal detection and convolutional neural network, characterized in that the fault signal detection includes vibration signal detection, acoustic emission signal detection, strain signal detection, temperature signal detection, oil parameter detection and electrical signal detection. The present invention proposes a set of method flows based on vibration fault signal monitoring and convolutional neural network model for intelligent bearing fault diagnosis. The vibration signal is obtained by an acceleration sensor. After reasonable sampling and 1D-2D signal processing transformation of historical data, the bearing signal samples under various fault states are reasonably divided into training sets and test sets. The training set is sent to the established deep convolutional neural network for model learning. After the model learning is completed, the generalization ability of the model is verified with the test set, that is, the test accuracy.

Description

Technical Field

[0001] The invention relates to the technical field of wind turbine generator fault diagnosis, and in particular to a method for audio recognition and fault diagnosis of mechanical faults of a wind turbine generator set. Background Art

[0002] With the continuous progress of science and technology and the sustained growth of economic development, people's living standards have been unprecedentedly improved. In addition, with the rapid development of industrial manufacturing and product production, energy supply is facing unprecedented challenges. Against the background of energy shortage, the development of clean and renewable energy has received more and more attention. Among them, wind energy, as a safe, clean and renewable energy, has great development potential and is an important alternative energy to traditional fossil fuels. With the continuous development and maturity of wind power technology, wind power generation has entered a period of rapid development. Wind energy has been highly favored in my country and even around the world.

[0003] As the scale of wind power generation continues to expand, the operation and maintenance of wind turbines has become increasingly prominent, especially due to the high operation and maintenance costs caused by failures and repairs. The maintenance cost even occupies as much as 10-25% of the total investment. Due to the geographical location of wind energy resources, the existing large-scale wind farms are generally built in remote areas such as northwest my country, plateaus and coastal areas. The operating environment of wind turbines is relatively harsh, and the transportation conditions are inconvenient, which brings many difficulties to the operation and maintenance of wind turbines. Wind turbines may be affected by extreme weather such as heavy rain and snow, storms and lightning. With the increase of operating time, wind turbine blades, gearboxes, generators and other components are prone to various abnormalities and failures, resulting in abnormal operation or even shutdown of the unit. Among them, the gearbox, as an important transmission component in the wind turbine, shows a high failure rate in actual operation. According to statistics, about 20% of the downtime is caused by gearbox failures, and most of the failures occur in bearings. Therefore, bearing failure is the main factor causing the failure of wind turbine gearboxes.

[0004] Health status monitoring is indispensable to ensure the safe and stable operation of wind turbines. Effective fault diagnosis technology can detect wind turbine faults in a timely manner at an early stage. Due to the continuous expansion of wind power installed capacity, traditional regular maintenance methods have become increasingly unable to meet the actual operation and maintenance needs of wind farms. Later, condition monitoring technology emerged, which uses sensors to obtain real-time operation data of wind turbines to achieve the purpose of online monitoring and fault diagnosis. However, this relies on expert knowledge and work experience, and requires a lot of manpower, material resources and time costs. In addition, with the increase in system complexity, the increase in fault diversity, the increase in data volume and the need for online diagnosis, it has become increasingly difficult to rely on traditional methods for effective fault diagnosis. More effective intelligent diagnosis technology is urgently needed to solve these problems. Therefore, we propose an audio recognition and fault diagnosis method for mechanical faults of wind turbines. Summary of the invention

[0005] 1. Purpose of the invention

[0006] In order to solve the technical problems existing in the background technology, the present invention proposes an audio recognition and fault diagnosis method for mechanical faults of a wind turbine generator set, which has the characteristics of intelligent detection.

[0007] (II) Technical solution

[0008] To solve the above technical problems, the present invention provides an audio recognition and fault diagnosis method for mechanical faults of wind turbine generator sets, including fault signal detection and convolutional neural networks, characterized in that the fault signal detection includes vibration signal detection, acoustic emission signal detection, strain signal detection, temperature signal detection, oil parameter detection and electrical signal detection, and the intelligent diagnosis of bearing faults based on convolutional neural networks includes data acquisition, feature extraction and feature classification. With the help of the function of deep learning models in automatic feature extraction, the fault diagnosis problem is converted into a similar image recognition and classification task for processing, and signal data preprocessing is performed. After that, a model structure is established for training and optimization, and overlapping sampling is performed. The two-dimensional representation of vibration data is used to diagnose whether the bearing is healthy.

[0009] The model used to establish the model structure is designed as a network structure of 3 convolution-pooling pairs and 2 fully connected layers. After obtaining the samples and preprocessing them, the training optimization uses a two-dimensional image and the corresponding health status label as a sample pair. A part of the entire sample set is randomly selected as a training set according to a given ratio, and the remaining samples are used as a test set. The training set is used for model training, and supervised learning is performed with health status category labels. After the model training is completed, the unlabeled test set data is sent to the model for testing to obtain the classification prediction results. After comparing with the actual results, the generalization ability of the model can be evaluated to reduce the randomness of the experimental results.

[0010] The bearing faults include inner ring fault, outer ring fault, rolling element fault and cage fault. The damage point caused by the fault repeatedly collides with the surface of other components to produce periodic fault impact, which is a low-frequency vibration signal. The greater the fault degree, the greater the impact amplitude. The calculation formula of the fault characteristic frequency of each part of the rolling bearing is as follows:

[0011] Inner race fault frequency:

[0012] Outer race fault frequency:

[0013] Rolling element failure frequency:

[0014] Cage failure frequency:

[0015] In the above formulas, N is the number of rolling elements, D is the bearing diameter, is the rolling element diameter, is the contact angle, is the rotation frequency of the shaft.

[0016] The convolutional neural network learning module consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The back propagation of the convolutional neural network learning module is a key step in optimizing the neural network parameters. Each weight parameter and bias parameter is updated according to the optimization algorithm after partial derivatives of the objective function. After back propagation, the weight optimization is updated using the optimization algorithm.

[0017] The fault signal detection also includes comprehensive composite fault diagnosis. Composite faults are regarded as belonging to multiple single faults, that is, they need to be identified as multiple fault modes at the same time. At this time, the model becomes a single-input multi-output model. It is necessary to identify the given composite faults as corresponding single fault modes respectively. In the pattern recognition problem, this problem can be called multi-label classification.

[0018] Preferably, the pooling layer is the key image recognition capability, with sparse connections and weight sharing, the input layer is the image input, including single-channel grayscale image and three-channel color image input forms, the output layer is a Softmax classifier, which is used to output image classification and recognition results. There is a special output layer structure for target detection or image segmentation. The convolution layer is used for feature learning, the pooling layer is used for feature selection, and the classifier uses the learned deep features for classification output, and the network parameters of each layer are optimized simultaneously during training.

[0019] Preferably, the data preprocessing rearranges the sampling values ​​of the one-dimensional signal into a two-dimensional matrix in order, which is regarded as a vibration grayscale map, and performs overlapping sampling at a certain length when constructing samples to increase the number of samples.

[0020] Preferably, the fault signal detection is performed by setting a sensor on the part to be detected to perform data collection and monitoring control.

[0021] Preferably, the convolutional neural network is verified for effectiveness through simulation signals and test bench signals.

[0022] Preferably, the fault signal detection is subjected to simulation training based on a convolutional neural network under noise interference to increase the model's anti-noise capability.

[0023] Preferably, after the model structure is constructed, objective function design and model training are performed.

[0024] Preferably, the vibration signal monitoring and convolutional neural network model are used to identify different fault types and different fault degrees, detect fault signals under noise interference, and effectively identify complex faults.

[0025] The above technical solution of the present invention has the following beneficial technical effects:

[0026] 1. The present invention proposes a method flow based on vibration fault signal monitoring and convolutional neural network model for intelligent bearing fault diagnosis. The vibration signal is obtained by an acceleration sensor. After reasonable sampling and 1D-2D signal processing transformation of historical data, the bearing signal samples under various fault conditions are reasonably divided into training sets and test sets. The training set is sent to the established deep convolutional neural network for model learning. After the model learning is completed, the generalization ability of the model is verified by the test set, that is, the test accuracy. The convolutional neural network model that passes the test can be further deployed in the wind turbine monitoring and control system to realize online real-time status monitoring and fault diagnosis.

[0027] 2. The present invention verifies the effectiveness of the fault diagnosis method based on convolutional neural network through simulation signals and test bench signals. Simulation is a simplified mathematical modeling of bearing vibration. It simplifies the problem to a certain extent, which helps to quickly verify the effectiveness of the method. The test bench data set is data collected during the operation of a real faulty bearing. Testing is performed using data on a variety of fault types and fault severities, which helps to comprehensively analyze and evaluate the effectiveness and superiority of the intelligent diagnosis algorithm.

[0028] 3. The present invention has experimentally proved that under the premise that composite fault samples are available, whether the composite fault is treated as a separate category or the model is trained using a multi-label method, a satisfactory fault identification accuracy rate can be achieved. Next, different from the previous method of adding composite fault samples to model training, this paper studies the method of constructing pseudo-composite fault signal samples using the intrinsic mode function obtained by the empirical mode decomposition of single fault signals and adding them to model training under the premise of lack of real composite fault data, so that the multi-label classification convolutional neural network model has the initial ability to diagnose and identify real composite faults. Compared with the model trained only with single fault samples, the accuracy of simultaneously identifying composite faults as containing multiple single fault components is greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the bearing structure of the present invention;

[0030] Figure 2 A schematic diagram of the convolution operation after adjusting the moving step size in the present invention;

[0031] Figure 3 A schematic diagram of the convolution operation after padding is added to the present invention;

[0032] Figure 4 Schematic diagram of the three-dimensional convolution operation of the present invention;

[0033] Figure 5 It is a commonly used activation function curve diagram of the present invention;

[0034] Figure 6 Schematic diagram of maximum pooling and average pooling operation of the present invention;

[0035] Figure 7 Schematic diagram of the fully connected layer and output layer of the present invention;

[0036] Figure 8 This is a flow chart of bearing fault diagnosis based on CNN of the present invention;

[0037] Fig. 9 This is a schematic diagram of overlapping sampling for signal preprocessing of the present invention;

[0038] Fig.10 It is a schematic diagram of a two-dimensional representation of the vibration data of the signal preprocessing of the present invention;

[0039] Fig.11 Schematic diagram of the convolutional neural network structure of the present invention;

[0040] Fig.12 It is the time domain waveform diagram of two fault signals of the present invention;

[0041] Fig.13 It is a simulation signal diagram of normal and fault states of the present invention;

[0042] Fig.14 This is a t-SNE feature visualization diagram of the simulation test results of the present invention;

[0043] Fig.15 A schematic diagram of a confusion matrix of simulation test results of the present invention;

[0044] Fig.16 The waveform diagram of the vibration signal containing noise interference of the present invention;

[0045] Fig.17 This is a diagram of the diagnostic test result of the present invention containing a noise interference signal;

[0046] Fig.18 This is a schematic diagram of a single label for fault classification according to the present invention;

[0047] Fig.19 It is a multi-label graph of fault classification of the present invention;

[0048] Fig. 20 This is a diagram of the pseudo composite fault signal construction process of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.

[0050] like Figure 1 As shown, the method for audio recognition and fault diagnosis of mechanical faults of wind turbine generator sets proposed in the present invention includes fault signal detection and convolutional neural network. The fault signal detection includes vibration signal detection, acoustic emission signal detection, strain signal detection, temperature signal detection, oil parameter detection and electrical signal detection. The intelligent diagnosis of bearing faults based on convolutional neural network includes data acquisition, feature extraction and feature classification. With the help of the function of deep learning model in automatic feature extraction, the fault diagnosis problem is converted into a similar image recognition and classification task for processing, and signal data preprocessing is performed. After that, the model structure is established for training optimization, overlapping sampling is performed, and the two-dimensional representation of vibration data is used to diagnose whether the bearing is healthy.

[0051] The bearing faults include inner ring fault, outer ring fault, rolling element fault and cage fault. The damage point caused by the fault repeatedly collides with the surface of other components to produce periodic fault impact, which is a low-frequency vibration signal. The greater the fault degree, the greater the impact amplitude. The calculation formula of the fault characteristic frequency of each part of the rolling bearing is as follows:

[0052] Inner race fault frequency:

[0053] Outer race fault frequency:

[0054] Rolling element failure frequency:

[0055] Cage failure frequency:

[0056] In the above formulas, N is the number of rolling elements, D is the bearing diameter, is the rolling element diameter, is the contact angle, is the rotation frequency of the shaft.

[0057] When a certain type of fault occurs, the spectral peak of the corresponding fault frequency will appear in the bearing vibration spectrum. Taking into account factors such as errors in production, processing and installation, elastic deformation caused by load, and interference from signal noise, the actual observed fault frequency components may not be completely consistent with those calculated by the formula, but are basically close to the theoretical calculated values. Ideally, this can be used to make a preliminary judgment on the type of fault.

[0058] In addition, after being impacted by a fault, the impact pulse will also cause high-frequency vibration of the bearing. Its vibration frequency is the natural vibration frequency of the bearing. Usually, the natural frequency of the inner and outer rings of the bearing can reach several thousand hertz, while the natural frequency of the rolling element is even higher, reaching hundreds of kilohertz. The vibration of the inner and outer rings is the most obvious in the natural vibration. The natural frequencies of the inner and outer rings are as follows:

[0059]

[0060] Where n is the vibration order (deformation coefficient), E is the elastic modulus of the material, the unit is , I is the moment of inertia of the ring cross section, D is the diameter of the neutral axis, and M is the mass per unit length. For steel balls, their natural frequency can be expressed by formula (2-6), where d represents the ball diameter, is the material density.

[0061]

[0062] The natural frequency vibration caused by the short-term pulse impact emitted by the impact of a local bearing fault can be measured by the sensor and manifested as an exponentially decaying sinusoidal vibration. The larger the natural frequency and the decay index, the shorter the vibration exists. Each impact caused by the fault will cause such high-frequency oscillation. If the mutual modulation effect of different frequency components and their harmonics is not considered, the bearing fault is reflected in the time domain vibration signal as a continuous low-frequency pulse appearing at the fault characteristic frequency, and each pulse decays in the form of a high-frequency oscillation.

[0063] The convolutional neural network learning module consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The back propagation of the convolutional neural network learning module is a key step in optimizing the neural network parameters. Each weight parameter and bias parameter is updated according to the optimization algorithm after taking the partial derivative of the objective function. The weight optimization is updated by the optimization algorithm after back propagation. Among them, the convolutional layer, for a computer, a picture is essentially a matrix composed of pixel values ​​(pixel values ​​are between 0-255), and maximum and minimum normalization processing is often performed before inputting into the neural network. The convolution layer performs convolution operations on the input image pixel value matrix in order to extract features from the image and output feature maps. For a network with multiple convolution layers, subsequent convolution layers continue to perform convolutions on the feature maps obtained in the previous layer to extract deeper features. The convolution layer consists of a series of convolution kernels (or filters). Typical convolution kernel sizes are 3×3, 5×5, etc. In addition to length and width, the convolution kernel size also has depth or the number of channels, which is equal to the number of input channels. The larger the length and width of the convolution kernel, the larger the image area covered by a single convolution operation. The convolution operation retains the spatial relationship between pixels, and the size of each convolution kernel in a convolution layer is consistent. The number of convolution kernels determines the number of channels of the output feature map.

[0064] There are two more hyperparameters that need to be set in the convolution layer: the moving stride and the padding method. The moving stride refers to the length of each convolution kernel movement. For example, a step of 1 means that the convolution kernel slides pixel by pixel in one direction, while a step of 2 means that each slide translates 2 pixels. As shown in the figure below, the larger the moving stride, the less overlap in the area involved in each convolution operation. This can reduce redundant calculations but may also lead to insufficient feature extraction. In practical applications, Stride = 1 or Stride = 2 is often used, and a larger moving stride is rarely used. represents the input side length, represents the side length of the convolution kernel, Represents the moving step size, then the output size is given by formula (2-1):

[0065] (2-1)

[0066] In the above formula, floor represents the integer function. Since the values ​​of f and s are at least 1, the output size is always less than or equal to the input size. If you want to maintain the same dimensions of the input and output, you need to add padding operations. In addition, since the edges of the original output participate in the convolution calculation less times than the middle area, in order to reduce the impact of insufficient edge feature extraction, it is also necessary to perform padding operations around the input matrix. Common terms for Padding include: (1) Valid, that is, no padding, the convolution kernel is calculated only when it completely covers the image pixels; (2) Same, the convolution kernel center is filled with spaces around it so that the calculation is performed when it coincides with the image pixels. When the moving step size is 1, the side length of the feature map after convolution can be kept unchanged, as shown in formula (2-2). If used Represents the padding size, then the output size after Stride and Padding operation is given by formula (2-3):

[0067] (2-2)

[0068] (2-3)

[0069] By coordinating the step size s and the padding p, the dimension size of the feature map after the convolution operation can be kept unchanged. The above examples are all based on square matrix inputs with the same length and width. The principle is similar for inputs with different lengths and widths. The schematic diagram of the convolution operation after adjusting the moving step size is shown in the figure below. Figure 2 As shown, the convolution operation after adding padding is as follows Figure 3 shown.

[0070] When the input image has multiple channels, for example, a color image has three color channels, RGB, the convolution kernel needs to have the same number of channels. The channel of each convolution kernel is convolved with the channel corresponding to the input layer. The convolution operation is done in the same way as above. Finally, the convolution results of each channel are added bit by bit to obtain the final feature map, such as Figure 4 The figure shows a schematic diagram of a three-dimensional convolution operation with three-channel input.

[0071] In addition, after the convolution operation, the convolution layer often needs to perform nonlinear mapping, that is, pass through an activation layer. Since the problems that the actual application of convolutional neural networks hopes to solve are often nonlinear problems, and the convolution operation is a linear operation, the purpose of introducing the activation layer is to introduce nonlinear factors. The so-called activation layer refers to an activation function operation on the result of the convolution operation. Commonly used activation functions include Sigmoid function, hyperbolic tangent function (Tanh) and linear rectified unit (Relu). These three activation functions perform very well in practical applications and are the most commonly used. Among them, the linear rectified unit is the default choice for most models. The expressions and function images of the three activation functions are as follows Figure 5 shown.

[0072] (2-4)

[0073] (2-5)

[0074] (2-6)

[0075] By the above Figure 5 It can be seen that the Sigmoid function and the Tanh function have an obvious disadvantage that they are easy to saturate. When the input value is relatively large or relatively small, the function value does not change significantly. When the neural network is deep, it is difficult to train and it is easy to cause the gradient to disappear. However, the derivative of the ReLU activation function is always 1 when the input is greater than 0, and is always 0 when the input is less than 0. It can achieve better results when training using the back propagation algorithm.

[0076] In summary, the forward propagation of the convolutional layer can be expressed by the following formula, where the superscript represents the number of layers, * represents the cross-correlation convolution operation, W represents the weight, b represents the bias, and σ represents the activation function.

[0077] (2-7)

[0078] In addition, after a single or multi-layer convolutional layer, a pooling layer is usually added to reduce the dimension in order to reduce network parameters and computational complexity. Therefore, it is also called a downsampling layer. The pooling layer can not only reduce the model size and shorten the training time, but also enhance the robustness of the extracted features, which helps prevent model overfitting. Pooling methods are generally divided into maximum pooling and average pooling. The calculation method is given by equations (2-8) and (2-9). Maximum pooling takes the maximum value in the window as the output, and average pooling takes the average of all values ​​in the window as the output. Figure 6 These are the effects of two pooling operations. In practical applications, maximum pooling is the main one.

[0079] (2-8)

[0080] (2-9)

[0081] The pooling layer is similar to the convolutional layer. There are several hyperparameters that need to be set, such as sampling window size, moving step size, and padding. Let the size of the input feature map be , respectively represent the width, height and number of channels. The window size and moving step size have the same symbol convention as the convolutional layer. The output size of the pooling layer is given by formula (2-10). The expression here omits the downward value function and Padding parameters because most models in practical applications use a maximum pooling sampling window of 2×2 and a moving step of Stride = 2. Padding is generally not used, and the side length of the feature map is generally an even number. After such a pooling operation, the length and width of the feature map can be reduced by half, effectively reducing the model size and increasing the calculation speed. It is worth noting that there are no parameters in the pooling layer that need to be learned during model training.

[0082] (2-10)

[0083] Among them, the fully connected layer, the above convolutional layer (including the activation part) and the pooling layer can be regarded as the feature extraction layer of the convolutional neural network. The extracted features are finally used for image classification or target detection. The fully connected layer is consistent with the traditional feedforward neural network layer and can be regarded as the feature classification layer. It will not be described here. All the extracted features are flattened and connected to one or more fully connected layers for model learning. The fully connected layer can be regarded as a buffer between the feature map and the output layer, such as Figure 7 As shown, it is equivalent to the hidden layer of the traditional neural network, which can enhance the model's ability to recognize the extracted image features. The expression of the fully connected layer is as follows:

[0084] (2-11)

[0085] Among them, the output layer, the output layer of the convolutional neural network also varies according to the different uses of the model. In the image classification task, the output layer is generally a Softmax classifier. Assuming that there are k categories in the classification task, the output of the Softmax function can be expressed as follows:

[0086] (2-12)

[0087] In the formula, zi represents the output value of each neuron in the last layer of the neural network without activation. It is worth noting that the premise for the above formula to be valid is that the number of neurons in the last layer of the neural network is consistent with the total number of classification categories. Some reference materials will classify the last layer of the neural network as a Softmax classifier, which is essentially a fully connected layer, but the activation function of this layer is the Softmax function. This article only defines the output result after the Softmax function operation as the output layer. A careful observation of the above formula shows that the sum of each item is 1. Therefore, the Softmax function can be used to map multiple scalars output by the fully connected layer into a normalized probability distribution, and output the confidence of the classification result. Each output can be regarded as the probability of belonging to each category, that is, the larger the output of the neuron, the greater the possibility that the category corresponding to the neuron is the true category, and the item with the largest probability is taken as the final classification prediction value.

[0088] (2-13)

[0089] Among them, model training, after the convolutional neural network model is built, the network depth and the size and number of convolution kernels have been determined, and the classification object is also known, but the specific network parameters (including weights and biases) need to be trained and optimized. At the same time, the setting of training parameters such as learning rate is also very important. Improper setting of training parameters may lead to overfitting or underfitting of model training, resulting in insufficient generalization ability of the model. This section mainly introduces the basic objective function design and model training methods.

[0090] Among them, the objective function, the training of convolutional neural networks can be regarded as an optimization problem. It is necessary to design an objective function to maximize or minimize it as the optimization goal. In deep learning tasks, the cost function is often used with necessary regularization terms as the objective function. The cost function is a measure of the error between the classification prediction result and the actual true value. The commonly used cost function in classification tasks is the average value of the cross entropy loss function of each sample. The cross entropy is a means of measuring the similarity of two probability distributions. The target distribution is expressed as p(x), and the distribution obtained by the prediction estimate is expressed as q(x). The cross entropy between them is defined as follows:

[0091] (2-14)

[0092] If the label value is represented by a one-hot vector, that is, the label value for the k-category classification is represented as a target vector of length k: [ p 1 ,..., p j ,..., p k ] , if the target category is , then let , and all other items are 0, then the final objective function can be expressed as formula (2-15), where the expression 1{c = yi} means that if the condition in the brackets is met, it is 1, otherwise it is 0.

[0093] (2-15)

[0094] Among them, back propagation is a key step in neural network parameter optimization. All weight parameters and bias parameters are updated according to the optimization algorithm after partial derivatives of the objective function. According to the previous article, the objective function is a measure of error. Therefore, back propagation actually transmits the error between the output value and the target value, or more accurately, the sensitivity of the error to the network parameters. Unlike the back propagation of ordinary deep neural networks, the error propagation of the pooling layer and the convolution layer requires special treatment, and the pooling layer has no parameters that need to be learned. The back propagation of the fully connected layer is basically the same as the traditional BP algorithm. For convenience, the subsequent expressions in this section are expressed as symbols using a single neuron or a single channel of a single filter.

[0095] In order to facilitate the calculation of error propagation, we define is the objective function about The sensitivity, also known as the error, is given by the partial derivative form shown in equation (2-16). Depending on whether the current layer is a fully connected layer, a convolutional layer, or a pooling layer, the expression is different, as shown in formula (2-17). All symbols with superscripts represent the lth layer of the neural network.

[0096] (2-16)

[0097] (2-17)

[0098] When the bias of the fully connected layer is known to be When reviewing the BP algorithm, it is easy to know that the previous hidden layer The sensitivity of the objective function to the weights and biases is shown in Equations (2-18)-(2-20).

[0099] (2-18)

[0100] (2-19)

[0101] (2-20)

[0102] When the bias of the pooling layer is known to be Derivation of the previous hidden layer Since the pooling layer generally uses maximum pooling or average pooling for downsampling during forward propagation, it is necessary to perform an upsampling to restore a larger area during backward propagation. First, all The sub-matrix size is restored to the size before pooling. For the maximum pooling, Each value is placed at the position of the maximum value in the corresponding pooling area during forward propagation. For average pooling, the average is taken and placed in the corresponding pooling area after restoration. Upsample is used to represent the process of pooling error matrix amplification and error redistribution. Since the pooling layer has no parameters, there is no need to consider the errors of weights and biases.

[0103] (2-21)

[0104] if is the bias of the convolutional layer, then we derive The expression of is different. Combined with the above, the deviation of the previous hidden layer at this time will be given by (2-22), where rot180(·) means that the matrix is ​​flipped 180 degrees, that is, flipped once up, down, left and right. This is the result obtained in the back-propagation derivation. In addition, the convolution operation in the back-propagation process needs to be padded to align the dimensions.

[0105] (2-22)

[0106] (2-23)

[0107] (2-24)

[0108] Among them, after calculating the back propagation, the optimization algorithm needs to be used to optimize and update the weights. The most commonly used optimization algorithm in neural network optimization is the gradient descent method. The basic idea of ​​the gradient descent method is to reduce the function value along the negative gradient direction of the variable when optimizing the objective function, so as to achieve the optimization goal. According to the number of samples calculated in a single time, it can be divided into stochastic gradient descent method, batch gradient descent method and small batch gradient descent method. The difference lies in the number of samples used for parameter update in each training iteration. Stochastic gradient descent uses one sample to update the parameters in each iteration. The advantage is fast update speed, but it is not conducive to parallel processing and easy to converge to the local optimal solution. Batch gradient descent uses all training set samples, which can represent the overall sample distribution and can realize parallel operation. The disadvantage is that when the training set is relatively large, all samples need to be operated in each iteration, so the training process will be very slow. Small batch is a compromise between the two. It is necessary to set the number of samples used in each iteration, that is, the batch size, to achieve better operation efficiency and parameter update speed. For neural network models, the BP algorithm can be used to efficiently calculate gradients, which is very conducive to using the gradient descent algorithm to train the neural network. The following table shows the algorithm flow of network optimization using the small batch gradient descent method.

[0109] Mini-batch stochastic gradient descent optimization process

[0110]

[0111] In the actual deep learning training process, multiple techniques are often used in combination to achieve the fastest and best results, but this process often requires repeated attempts to finally find the best strategy. The general rule of thumb is to perform batch normalization before sending it to each convolutional layer or fully connected layer, and then use random dropout after the fully connected layer. This can greatly reduce the possibility of overfitting of the model. In terms of learning rate setting, it often starts from 0.001 or even smaller and gradually increases by ten or five times to find the most appropriate learning rate hyperparameter.

[0112] Figure 8 The figure shows the diagnostic process from sample acquisition to data preprocessing and then to model training. Once the model is trained and optimized, it can be used to classify and identify samples of unknown status to diagnose whether the bearing is healthy.

[0113] like Figure 9-10 The fault signal detection also includes comprehensive composite fault diagnosis. Composite faults are regarded as belonging to multiple single faults, that is, they need to be identified as multiple fault modes at the same time. At this time, the model becomes a single-input multi-output model. It is necessary to identify the given composite faults as corresponding single fault modes respectively. In the pattern recognition problem, this problem can be called multi-label classification.

[0114] It should be noted that the pooling layer is the key image recognition capability, with sparse connections and weight sharing. The input layer is the image input, including single-channel grayscale images and three-channel color image input forms. The output layer is a Softmax classifier, which is used to output image classification and recognition results. There is a special output layer structure for target detection or image segmentation. The convolution layer is used for feature learning, the pooling layer is used for feature selection, and the classifier uses the learned deep features for classification output. The network parameters of each layer are optimized simultaneously during training.

[0115] Furthermore, the data preprocessing rearranges the sampling values ​​of the one-dimensional signal into a two-dimensional matrix in order, which is regarded as a vibration grayscale image, and overlaps the samples at a certain length when constructing the samples to increase the number of samples. In this system, this method is used by default for signal sampling and sample construction. In fact, the length of a single signal sample (that is, the number of sampling points contained in a single sample) can be adjusted according to the sampling frequency and the system rotation frequency. In order to facilitate the processing of the convolutional neural network, the side length of the input image is often an even number, especially when the side length is an exponential power with a base of 2, it helps to simplify the network structure parameter design. Commonly used sample lengths are 400 = 20×20, 1024 = 32×32, 4096 = 64×64, etc. Many literatures have adopted this signal processing method to achieve certain research results. Therefore, this method will also be directly used as one of the signal preprocessing means, and the relevant diagnostic test results can be used as a comparison benchmark value. In addition, the present invention also compares short-time Fourier transform (STFT) and continuous wavelet transform (CWT) as signal processing methods to comprehensively evaluate the feasibility and efficiency of the algorithm system.

[0116] It should be noted that the fault signal detection is carried out by setting sensors on the part to be detected to collect data and monitor and control. Different signals are suitable for monitoring different components. For example, vibration signals are particularly suitable for gear boxes, blades, drive shafts, etc.; the temperature can reflect whether the equipment has deteriorated or whether the operation is abnormal, and is suitable for monitoring gears, generators, converters, etc.; the amplitude and harmonic components of the electrical signal directly reflect electrical faults, and are particularly suitable for generators, sensors, etc. In addition, the wind turbine is a strong electromechanical coupling system, and mechanical faults can also cause electrical signal abnormalities; monitoring of oil component content is suitable for analyzing whether mechanical components lubricated by lubricating oil are damaged; it should be noted that in the actual system, multiple sensors will be installed at different locations at the same time and even multiple signals will be monitored at the same time. Data fusion technology is used to jointly perform fault diagnosis, and the processing methods and analysis difficulties of different signals are also different; among them, vibration signals are the most widely used monitoring signals, and play an important role in the fault diagnosis of components such as bearings, gears, blades, and drive shafts. Since vibration signals often contain inherent fault characteristics and reflect the operating status of the fan, a large number of practices have shown that a large amount of characteristic information can be extracted from vibration signals through a variety of signal processing technologies, and the type of fault, fault location, and fault severity can be accurately and effectively identified based on these characteristic information; it is worth mentioning that a variety of signal sensors, including vibration signals, are invasive and need to be embedded in the equipment, which increases the complexity of the equipment and the possibility of failure. Non-invasive sound signal detection also has great application prospects in fault diagnosis.

[0117] Furthermore, the effectiveness of the convolutional neural network is verified through simulation signals and test bench signals. Simulation is a simplified mathematical modeling of bearing vibration, which simplifies the problem to a certain extent and helps to quickly verify the effectiveness of the method. The test bench data set is data collected during the operation of a real faulty bearing. Testing is performed using data on a variety of fault types and fault severities, which helps to comprehensively analyze and evaluate the effectiveness and superiority of the intelligent diagnosis algorithm.

[0118] It should be noted that after the model structure is built, the objective function design and model training are carried out. A network structure with 3 convolution-pooling pairs and 2 fully connected layers is as follows: Fig.11 shown.

[0119] It is worth mentioning that when the fault diagnosis task changes, for example, when the fault types increase, the number of neurons in the output layer can be directly increased while keeping the structure of the previous feature extraction layer unchanged, so that the network structure has a certain applicability. In this paper, the convolution layer uses the Same filling method to keep the input and output sizes unchanged, and the pooling layer uses the Valid filling method, that is, no filling, to achieve downsampling. Table 3.1 shows the structural details of the designed convolutional neural network model, where 32×32@16 means that there are 16 channels and the size of each channel is 32×32, and the other items have similar meanings. In addition to the table, there are some details and techniques used for training in the experiment, mainly the batch normalization (BatchNormalization) and random dropout (Dropout) mentioned in the previous article. Batch normalization is used in the convolution layer and the fully connected layer to accelerate training, while the random dropout regularization method is used in the fully connected layer to prevent the model from overfitting and improve the generalization ability of the model. The random dropout ratio uses the recommended value of 0.5 as the choice in the experiment of this paper.

[0120] The convolutional neural network structure is designed as follows:

[0121] Layer Name Core Size Number Output Size Activation Function

[0122] 1 Input layer - 1 32×32@1 -

[0123] 2 Convolutional layer 1 5×5@1 16 32×32@16 ReLU

[0124] 3 Pooling layer 1 2×2@16 - 16×16@16 -

[0125] 4 Convolutional layer 2 3×3@16 32 16×16@32 ReLU

[0126] 5 Pooling layer 2 2×2@32 - 8×8@32 -

[0127] 6 Convolutional layer 3 3×3@32 64 8×8@64 ReLU

[0128] 7 Pooling layer 3 2×2@64 - 4×4@64 -

[0129] 8 Fully connected layer 1 512 1 512×1 ReLU

[0130] 9 Fully connected layer 2 128 1 128×1 ReLU

[0131] 10 Output layer 10 1 10×1 Softmax

[0132] The effectiveness of the CNN-based fault diagnosis method is verified by simulation signals and test bench signals. Simulation is a simplified mathematical modeling of bearing vibration. It simplifies the problem to a certain extent and helps to quickly verify the effectiveness of the method. The test bench data set is the data collected during the operation of the real faulty bearing. The test is carried out using data of various fault types and fault severity, which helps to comprehensively analyze and evaluate the effectiveness and superiority of the intelligent diagnosis algorithm.

[0133] Through numerical simulation, vibration signal simulation is carried out. The fault signal of rolling bearing is mainly composed of three parts: pulse impact caused by the fault defect point, vibration harmonics caused by bearing imbalance or gear meshing in the gearbox, and Gaussian white noise introduced in the measurement, as shown in the following formula.

[0134] (3-1)

[0135] in: is the amplitude of the ith pulse excited by the fault, is the time when the pulse occurs; and is the amplitude and initial phase of the kth harmonic caused by bearing imbalance or gear meshing in the gearbox; is Gaussian white noise.

[0136] Here we consider the vibration signals in two cases, namely normal case and fault case. In the normal case, two vibration harmonics are generated, and the simplified mathematical expressions are given in (3-2) and (3-3):

[0137] (3-2)

[0138] (3-3)

[0139] in and represents two vibration harmonic frequencies, Represents time and satisfies , is the sampling point, is the sampling frequency. Considering that the vibration is very small under normal circumstances, the amplitude of the above simulation signal is relatively low. In addition, the simulation expressions of the two fault signals are as follows. Formula (3-4) represents the vibration caused by the outer ring fault, and formula (3-5) represents the vibration caused by the inner ring fault. There is an additional signal modulation component. In the following formula, Represent the resonance frequency and fault characteristic frequency of the outer ring and inner ring respectively. It represents the rotation frequency of the neutral axis. Fig.12 The time domain waveforms of two fault signals are shown.

[0140] (3-4)

[0141] (3-5)

[0142] Depend on Fig.12 It can be seen that under rational circumstances, the outer ring fault causes an impact signal, the maximum amplitude of each pulse is equal, and the interval between them is the inverse of the fault characteristic frequency. Each pulse oscillates at a decaying frequency at the resonance frequency. Figure (b) shows a typical inner ring fault signal. Since the relative position of the vibration acceleration sensor and the inner ring fault point changes with the rotation of the shaft, the impact amplitude generated by each impact defect is modulated by the rotation speed frequency, causing the signal envelope to show a "wave-like" change.

[0143] Based on the signal generated by the above vibration source, after superimposing Gaussian white noise with a signal-to-noise ratio of 20dB, various simulation signals of bearing health status can be constructed according to formula 3-1, namely healthy state, outer ring fault, inner ring fault, and composite fault of inner and outer ring faults at the same time, such as Fig.13 shown.

[0144] In the experiment, samples were collected with a sampling length of 1024 points and an overlap rate of 30%. 250 samples were obtained in each state, resulting in a total of 1000 samples, of which 60% were randomly divided as training sets and the remaining 40% were used as test sets. First, the data of the training set was normalized and rearranged, and then sent to CNN for training as a 32×32 two-dimensional grayscale image. When the training accuracy is no longer significantly improved, it means that the training has been completed. Sending the test set data to the trained model in the same way can not only obtain the test accuracy, but also be used to evaluate the recognition accuracy and generalization ability of the model. Since the simulation data is more idealized and the fault characteristics are more obvious, the experiment very smoothly achieved a test accuracy of 100%, verifying the feasibility and effectiveness of this chapter's model based on two-dimensional grayscale images and CNN algorithm. Since the neural network model is a black box model with poor interpretability, feature visualization is used here to further help understand why the model can distinguish and identify various fault types and health states. Fig.14 This is a visualization of the distribution of high-dimensional features extracted from each sample using the t-SNE method in the feature space after dimensionality reduction. It can be seen from the figure that before being sent to the classifier, the signals of each category have been separated from each other in the feature space, and the farther the clusters are from each other, the more significant the difference in features between the categories is, which is more conducive to classification and recognition. This shows that the convolutional neural network has a very good feature learning ability and can learn discriminative features from the original input signal for classification. Fig.15 is the confusion matrix of the test results. The confusion matrix can be used to view the relationship between the predicted results and the actual results of each category, so as to determine whether the model is still insufficient in the identification of certain categories or which categories are easily confused. From the figure, it can be seen that the prediction accuracy of each category in the diagnosis results of this simulation signal has reached 100%. It should be pointed out that the above ideal results are also obtained under the condition that there are relatively few fault categories and the signal data quality is relatively ideal. The bearing test data in the real environment often cannot achieve such a perfect diagnosis result. Therefore, it is necessary to use the actual test data for further analysis and research.

[0145] Furthermore, the vibration signal monitoring and convolutional neural network model are used to identify different fault types and fault degrees, detect fault signals under noise interference, and effectively identify complex faults.

[0146] The signal form of the bearing vibration signal after the noise is superimposed can be simulated by additive Gaussian white noise. The signal-to-noise ratio is defined as formula (4-1), which represents the logarithm of the ratio of signal energy to noise energy in decibels (dB). For example, when the signal-to-noise ratio is 10dB, it means that the signal energy is 10 times the noise energy, and when the signal-to-noise ratio is 0dB, it means that the signal energy is equivalent to the noise energy. At this time, the influence of noise is already quite significant. Fig.16The figure shows the time domain waveform of a 0.007-inch rolling element failure before and after adding Gaussian white noise with a signal-to-noise ratio of 0dB. It can be seen from the figure that it is difficult to distinguish the fault impact component from the signal waveform after superimposing the noise.

[0147] (4-1)

[0148] In order to improve the generalization ability of the model, the improved cost function is used as the optimization target for model training. At the same time, in order to improve the model's ability to learn the features of samples, some training samples are subjected to noise interference (Sample Interrupt, SI) during the training process, and randomly mixed with the original signal samples and sent to the model for training. Since some samples are interfered by noise, the feature learning process of these samples also strengthens the model's ability to resist noise interference to a certain extent. The specific experimental implementation details are shown in the following table, where the training set adds signal samples under three types of signal-to-noise ratio interference.

[0149] Training set construction and parameter setting

[0150]

[0151] In addition, the model was trained by improving the cost function without adding interference samples and by adding interference samples without improving the cost function. The experimental parameters were kept unchanged, and the test results of the four models were compared and analyzed in combination with the conventional convolutional neural network model mentioned above. In all experiments, the vibration signals were not processed by complex time-frequency analysis, but were rearranged into a two-dimensional vibration grayscale image according to the order of sampling points to verify the improvement in test accuracy under noise interference brought about by the improved cost function. In the experiment, after the model training was completed, Gaussian white noise was added to the original signal at a signal-to-noise ratio of -4dB to 6dB, which was consistent with the test range of the previous experiment, and then sent to the trained model for testing. The accuracy results under various noise interference conditions were obtained and plotted into relevant curves as shown below. Fig.17 shown.

[0152] It can be seen intuitively from the above curve that after the cost function is improved, the convolutional neural network model can achieve better fault recognition accuracy in a noisy interference environment, especially after adding sample interference, which further enhances the generalization ability of the model. Under the condition of signal-to-noise ratio SNR = -4dB, the original CNN model can only achieve a fault recognition accuracy of about 81%, while the LDCNN+SI model improved in this chapter can achieve an accuracy of about 94%, which is a significant improvement, mainly due to the improvement of the cost function and the expansion of interference samples. At the same time, by comparing the model test results before and after adding sample interference, it can be seen that algorithm improvement and data sample expansion are very important. For data-driven artificial intelligence algorithms, the more data, the more comprehensive the model's feature learning will be, and the stronger the model's generalization ability will be. Sample expansion is relatively easy to implement compared to algorithm improvement, so this link should be taken very seriously in practical applications.

[0153] In addition, when the signal-to-noise ratio is high, both the original model and the improved model can achieve satisfactory classification results. It is worth mentioning that although the recognition rate of the improved model is still inferior to that after the wavelet transform signal processing, its value lies in that it can still obtain a significant improvement in model accuracy by simply improving the cost function without complex signal processing steps. Therefore, when the signal-to-noise ratio SNR = 0dB, the improved model has achieved a satisfactory fault diagnosis accuracy of 99.71%, which can be considered as a result of practical and effective improvement.

[0154] For example, for bearing fault diagnosis, in addition to the healthy state, there are three known single faults: inner ring fault, outer ring fault and rolling element fault. Compound fault is the simultaneous existence of inner and outer ring faults. For single-label classification, there are 5 independent categories, and for multi-label classification, there are 3 fault categories. Compound faults belong to both inner ring and outer ring faults, and the normal state does not belong to any of them. The relevant fault diagnosis process is as follows: Fig.18 and 19As shown. For multi-label classification, special attention should be paid to the label setting method when constructing samples. That is, different from ordinary one-hot encoding, for the labels of multi-label samples, there are multiple elements in the label vector that are 1, and all elements of the normal state label are 0, that is, no single fault exists. In addition, the setting of the loss function and the evaluation method of the accuracy will also be different. In this paper, multiple binary classifiers are used to implement multi-label classification, which is similar to the one-vs-rest classification strategy in the multi-classification support vector machine. For convenience, a logistic regression binary classifier with a Sigmoid activation function is used for each label category. In theory, other machine learning classifiers such as support vector machines can also be used. The features extracted by the CNN model can be mapped to a probability output of 0-1 after being activated by the Sigmoid function of the logistic regression classifier. The threshold is often used in practice. As the criterion for determining the positive and negative categories of the output, the corresponding cross entropy cost function is adjusted as shown in formula (5-1), where is the activation function output value of the i-th sample in the c-th classifier, representing the probability that the predicted category of the label category is positive, and k is the total number of binary classifiers, that is, the total number of single fault categories. The accuracy rate can be defined as the average accuracy rate of each label, that is, a single fault can be identified as a composite fault and it can also be considered to contribute to the overall accuracy rate, or it can be defined as the average accuracy rate of each category, that is, only when each label category of the composite fault is completely and accurately identified can it be considered to contribute to the overall accuracy rate. This article uses the latter as an evaluation indicator, as the absolute accuracy rate.

[0155] (5-1)

[0156] From the previous description, it can be seen that compound faults are closely related to single faults and can be considered as nonlinear coupling of multiple single faults. Even if the generation mechanism is complex and difficult to explain, it can be considered that the fault frequency components that constitute the single fault still exist in the compound fault. The traditional diagnostic method based on signal decoupling and fault frequency component analysis is also based on this principle. Therefore, it can be considered that constructing "pseudo-composite fault" data containing multiple single fault frequency components and adding them to the training samples will help the model to obtain more powerful signal decoupling analysis capabilities. This article mainly attempts to superimpose time domain signals. The superposition method of time domain signals is to perform empirical mode decomposition (EMD) on multiple single fault signals and linearly superimpose the selected intrinsic mode functions (IMF). In addition, considering that the occurrence order and fault impact interval of different single faults in the time domain are uncertain, a variety of sampling points are linearly superimposed with staggered ratios according to the sample length. Table 5.1 is the EMD algorithm flow. Fig. 20 The “pseudo-composite fault” signal construction process is shown.

[0157] Table 5.1 Empirical mode decomposition algorithm flow

[0158]

[0159] The above EMD decomposition method can be used to decompose the signal of a single fault to obtain the intrinsic mode function that can characterize the fault information. By superimposing the intrinsic mode functions of signals of different fault types, the signal characteristics of the composite fault can be approximately described. Therefore, the approximate signal obtained in this way is called a "pseudo-composite fault" signal in this article, which is used to assist the deep learning model in the feature learning and pattern recognition capabilities of single faults and composite faults. This method does not require setting parameters, and can not only be used for the analysis of the coexistence of two faults, but can also be extended to the situation where more faults exist at the same time, and has a certain degree of versatility.

[0160] By proposing a method to construct pseudo-composite fault samples using the EMD decomposition results of single fault data, and using the multi-label classification method to train the convolutional neural network feature extractor and multiple binary classifiers, the correct diagnosis of real composite fault samples was initially achieved, proving that the method proposed in this chapter is practically feasible. Without increasing the complexity and difficulty of signal processing and model construction, the convolutional neural network model can achieve effective diagnosis when multiple single faults coexist in the absence of composite fault samples.

[0161] In general, the present invention first proposes a method flow based on vibration signal monitoring and convolutional neural network model for intelligent bearing fault diagnosis. The vibration signal is obtained by an acceleration sensor. After reasonable sampling and 1D-2D signal processing transformation of historical data, the bearing signal samples under various fault conditions are reasonably divided into training sets and test sets. The training set is sent to the established deep convolutional neural network for model learning. After the model learning is completed, the test set is used to verify the generalization ability of the model, that is, the test accuracy. The convolutional neural network model that passes the test can be further deployed in the wind turbine monitoring and control system to realize online real-time status monitoring and fault diagnosis.

[0162] Then, due to the use of a two-dimensional input convolutional neural network, the one-dimensional time domain vibration signal needs to be specially processed and converted into a two-dimensional form. With the help of the powerful image recognition ability of the deep convolutional neural network, sampling any two-dimensional image form can achieve a relatively ideal fault diagnosis accuracy. Among them, the vibration grayscale image based on the rearrangement of the signal sampling points has a unique advantage because it is simple to implement and does not require complex transformation.

[0163] Then, based on the principle of keeping the structure as simple as possible, the training as efficient as possible, and the prediction as accurate as possible, a comparative analysis was conducted on the number of convolutional pooling layers, the number of filters, etc., and a deep convolutional neural network model that is most suitable for the task studied in this article was obtained, which achieved a good balance between training cost and prediction accuracy.

[0164] Subsequently, considering the universality of noise interference in actual industrial production environments, the present invention studies the recognition ability of convolutional neural network models for fault signals under strong noise interference. Experiments show that the recognition method based on vibration grayscale images still has very good fault recognition accuracy when the signal-to-noise ratio is not too low, but when the signal-to-noise ratio drops below 0dB, the recognition accuracy drops significantly, while the time-frequency image recognition based on continuous wavelet transform still has a very high fault recognition accuracy due to the noise reduction function of wavelet transform. Based on this, an improved strategy based on clustering cost function and training sample interference is proposed, which improves the fault signal recognition accuracy of the method based on vibration grayscale images under low signal-to-noise ratio conditions.

[0165] Finally, the present invention first proves through experiments that under the premise that composite fault samples are available, whether the composite fault is treated as a separate category or the model is trained using a multi-label method, a satisfactory fault identification accuracy rate can be achieved. Next, different from the previous method of adding composite fault samples to model training, this paper studies the method of constructing pseudo composite fault signal samples using the intrinsic mode function obtained by the empirical mode decomposition of single fault signals and adding them to model training under the premise of lack of real composite fault data, so that the multi-label classification convolutional neural network model has the initial ability to diagnose and identify real composite faults. Compared with the model trained only with single fault samples, the accuracy of simultaneously identifying composite faults as containing multiple single fault components is greatly improved.

[0166] It should be understood that the above specific embodiments of the present invention are only used to illustrate or explain the principles of the present invention, and do not constitute a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc. made without departing from the spirit and scope of the present invention should be included in the protection scope of the present invention. In addition, the appended claims of the present invention are intended to cover all changes and modifications that fall within the scope and boundaries of the appended claims, or the equivalent forms of such scope and boundaries.

Claims

1. Audio recognition and fault diagnosis methods for mechanical faults of wind turbine generator sets, including fault signal detection and convolutional neural networks, It is characterized in that The fault signal detection includes vibration signal detection, acoustic emission signal detection, strain signal detection, temperature signal detection, oil parameter detection and electrical signal detection. The intelligent diagnosis of bearing fault detection based on convolutional neural network includes data acquisition, feature extraction and feature classification. With the help of the function of deep learning model in automatic feature extraction, the fault diagnosis problem is converted into a similar image recognition and classification task for processing, and signal data preprocessing is performed. After that, the model structure is established for training optimization, and overlapping sampling is performed. Through the two-dimensional representation of vibration data, the health of the bearing is diagnosed; The model used to establish the model structure is designed as a network structure of 3 convolution-pooling pairs and 2 fully connected layers. After obtaining and preprocessing the samples, the two-dimensional image and the corresponding health status label are used as a sample pair. A part of the entire sample set is randomly selected as the training set according to a given ratio, and the remaining samples are used as the test set. The training set is used for model training, and supervised learning is performed with health status category labels. After the model training is completed, the unlabeled test set data is sent to the model for testing to obtain the classification prediction results. After comparing with the actual results, the generalization ability of the model can be evaluated to reduce the randomness of the experimental results. The bearing faults include inner ring fault, outer ring fault, rolling element fault and cage fault. The damage point caused by the fault repeatedly collides with the surface of other components to produce periodic fault impact, which is a low-frequency vibration signal. The greater the fault degree, the greater the impact amplitude. The calculation formula of the fault characteristic frequency of each part of the rolling bearing is as follows: Inner race fault frequency: Outer race fault frequency: ; Rolling element failure frequency: Cage failure frequency: ; In the above formulas, N is the number of rolling elements, D is the bearing diameter, is the rolling element diameter, is the contact angle, is the rotation frequency of the shaft; The convolutional neural network learning module consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The back propagation of the convolutional neural network learning module is a key step in optimizing the neural network parameters. Each weight parameter and bias parameter is updated according to the optimization algorithm after partial derivative of the objective function. After back propagation, the weight optimization is updated using the optimization algorithm. The fault signal detection also includes comprehensive composite fault diagnosis. Composite faults are regarded as belonging to multiple single faults, that is, they need to be identified as multiple fault modes at the same time. At this time, the model becomes a single-input multi-output model. It is necessary to identify the given composite faults as corresponding single fault modes respectively. In the pattern recognition problem, this problem is divided into multi-label classification.

2. The method for audio recognition and fault diagnosis of mechanical faults of wind turbine generator sets according to claim 1, It is characterized in that The pooling layer is the key image recognition capability, with sparse connections and weight sharing. The input layer is the image input, including single-channel grayscale images and three-channel color image input forms. The output layer is a Softmax classifier, which is used to output image classification and recognition results. There is a special output layer structure for target detection or image segmentation. The convolution layer is used for feature learning, the pooling layer is used for feature selection, and the classifier uses the learned deep features for classification output. The network parameters of each layer are optimized simultaneously during training.

3. The method for wind turbine generator mechanical fault audio recognition and fault diagnosis according to claim 1, It is characterized in that The data preprocessing rearranges the sampling values ​​of the one-dimensional signal into a two-dimensional matrix in order, which is regarded as a vibration grayscale image, and performs overlapping sampling according to a certain length when constructing samples to increase the number of samples.

4. The method for wind turbine generator mechanical fault audio recognition and fault diagnosis according to claim 1, It is characterized in that The fault signal detection is performed by setting a sensor on the part to be detected to perform data collection and monitoring control.

5. The method for wind turbine generator mechanical fault audio recognition and fault diagnosis according to claim 1, It is characterized in that The effectiveness of the convolutional neural network is verified by simulation signals and test bench signals.

6. The method for audio recognition and fault diagnosis of mechanical faults of wind turbine generator sets according to claim 1, It is characterized in that The fault signal detection is subjected to simulation training based on a convolutional neural network under noise interference, thereby increasing the anti-noise capability of the model.

7. The method for wind turbine generator mechanical fault audio recognition and fault diagnosis according to claim 1, It is characterized in that After the model structure is constructed, objective function design and model training are performed.

8. The method for audio recognition and fault diagnosis of mechanical faults of wind turbine generator sets according to claim 1, It is characterized in that The vibration signal detection and convolutional neural network model are used for identifying different fault types and different fault degrees, fault signal detection under noise interference, and effective identification of compound faults.

Images