Fan bearing diagnosis method and device based on gradient weighting type activation thermodynamic diagram

Abstract

The invention provides a fan bearing diagnosis method and device based on a gradient weighting type activation thermodynamic diagram, and relates to the technical field of wind driven generator bearing fault diagnosis, and the method comprises the steps: obtaining a bearing vibration signal of a wind turbine generator; carrying out noise reduction processing on the obtained bearing vibration signal; the bearing vibration signal after noise reduction is converted into a time-frequency image through short-time Fourier transform, the obtained time-frequency image serves as input, and a gradient weighting type activation thermodynamic diagram is obtained through a feature extraction unit of a fault diagnosis model; taking the obtained gradient weighting type activation thermodynamic diagram as input, performing fault diagnosis through a fault diagnosis unit of a fault diagnosis model according to the input gradient weighting type activation thermodynamic diagram, and obtaining a diagnosis result; and the fault diagnosis unit is obtained by training a preset machine learning algorithm according to historical gradient weighted class activation thermodynamic diagrams corresponding to different bearing faults of the wind turbine generator. The data processing calculation amount can be effectively reduced, and the fault diagnosis accuracy is improved.

Description

Technical Field

[0001] The present application relates to the technical field of wind turbine bearing fault diagnosis, and specifically to a wind turbine bearing diagnosis method based on a gradient-weighted class activation thermal map, a wind turbine bearing diagnosis device based on a gradient-weighted class activation thermal map, and a terminal device. Background Art

[0002] For wind turbines with large-sized flexible blades, the generator bearing is one of the key components in the wind turbine, which plays a vital role in the stable operation of the unit. However, due to factors such as harsh working environment and complex and variable loads, the generator bearing is prone to faults such as cracks, peeling, and wear. These faults not only affect the normal operation of the unit, but may also cause equipment damage, downtime for maintenance, and even safety accidents. Traditional fault diagnosis methods for wind turbine generator bearings mainly include vibration analysis and temperature monitoring. Although these methods can detect bearing faults to a certain extent, they often have problems such as low sensitivity, insufficient accuracy, and susceptibility to interference. In particular, for early faults and hidden faults, existing methods are often difficult to detect effectively. With the development of big data and artificial intelligence, data-driven fault diagnosis methods are widely used. Compared with traditional mathematical models, data-driven fault diagnosis methods do not need to obtain accurate mathematical models, but directly rely on feature data to infer faults. However, existing data-driven fault diagnosis methods have problems with low sensitivity and accuracy in fault detection of wind turbine bearings. Summary of the invention

[0003] The purpose of the embodiments of the present application is to provide a fan bearing diagnosis method based on a gradient-weighted class activation heat map, a fan bearing diagnosis device based on a gradient-weighted class activation heat map, and a terminal device to solve the above-mentioned problems.

[0004] In order to achieve the above-mentioned object, the first aspect of the present application provides a fan bearing diagnosis method based on a gradient-weighted class activation heat map, comprising:

[0005] Obtain bearing vibration signals of wind turbine generators;

[0006] Perform noise reduction processing on the acquired bearing vibration signal;

[0007] The denoised bearing vibration signal is converted into a time-frequency image by using short-time Fourier transform. The obtained time-frequency image is used as input. The input time-frequency image is forward propagated and backward propagated through the feature extraction unit of the fault diagnosis model to obtain a coarse gradient weighted class activation heat map and a back-propagation gradient map, respectively. The coarse gradient weighted class activation heat map is multiplied with the back-propagation gradient map to obtain the final gradient weighted class activation heat map.

[0008] Taking the obtained gradient weighted class activation heat map as input, the fault diagnosis unit of the fault diagnosis model performs fault diagnosis according to the input gradient weighted class activation heat map to obtain a diagnosis result;

[0009] The fault diagnosis unit is obtained by training a preset machine learning algorithm based on the historical gradient weighted class activation heat map corresponding to different bearing faults of the wind turbine generator.

[0010] Optionally, the bearing vibration signal of the wind turbine generator includes:

[0011] The wind turbine generator may include one or more of a front bearing radial vertical vibration signal, a front bearing radial horizontal vibration signal, a rear bearing radial vertical vibration signal, a rear bearing radial horizontal vibration signal, a generator transmission vertical envelope vibration signal, and a generator transmission horizontal envelope vibration signal.

[0012] Optionally, the acquired bearing vibration signal is subjected to noise reduction processing, including:

[0013] Perform wavelet decomposition of scale J on the acquired bearing vibration signal to obtain wavelet detail coefficients and approximate coefficients at each scale;

[0014] Determine the noise standard deviation of the current layer based on the wavelet detail coefficients of each layer, and determine the noise intensity of each layer based on the obtained noise standard deviation;

[0015] Determine the unified threshold of wavelet decomposition of each layer based on the standard deviation of noise of each layer;

[0016] The wavelet detail coefficients of each layer are sorted from small to large according to the absolute value, and the evaluation vectors of the wavelet detail coefficients of each layer are calculated based on the sorted wavelet detail coefficients. The obtained evaluation vectors are sorted from large to small, and the minimum value is used as the approximation error. The target wavelet detail coefficient corresponding to the approximation error is determined, and the wavelet decomposition adaptive threshold of the current layer is determined based on the obtained target wavelet detail coefficient;

[0017] Determine the average value of the absolute value of the wavelet detail coefficient of the current layer and the minimum energy level value of the wavelet detail coefficient;

[0018] The wavelet decomposition threshold corresponding to each layer is determined based on the wavelet decomposition unified threshold of each layer, the wavelet decomposition adaptive threshold, the average value of the absolute value of the wavelet detail coefficient, the minimum energy level value of the wavelet detail coefficient and the noise standard deviation;

[0019] The wavelet detail coefficients of each layer are adjusted based on the obtained wavelet decomposition threshold, and the wavelet detail coefficients of each layer after the adjustment are reconstructed to obtain the bearing vibration signal after noise reduction.

[0020] Optionally, determining the noise standard deviation of each layer based on the wavelet detail coefficient of each layer includes:

[0021] The noise standard deviation of each layer is calculated by the following formula:

[0022]

[0023] Among them, σ j represents the noise standard deviation of the jth layer, Wa j,k represents the kth wavelet detail coefficient of the jth layer, N represents the total number of wavelet detail coefficients at the jth scale, and k is the number of wavelet detail coefficients at the current scale;

[0024] The noise intensity of the current layer is determined based on the obtained noise standard deviation, including:

[0025]

[0026] Among them, Γ(σ j ) represents the noise intensity of the j-th layer wavelet detail coefficient, A j Represents the amplitude of the j-th layer wavelet detail coefficient.

[0027] Optionally, determining a unified threshold of wavelet decomposition of each layer based on a noise standard deviation of each layer includes:

[0028] The unified threshold of wavelet decomposition of each layer is calculated by the following formula:

[0029]

[0030] Among them, λ 1,j represents the uniform threshold of wavelet decomposition at the jth layer;

[0031] The evaluation vector of the wavelet detail coefficients of each layer is calculated based on the sorted wavelet detail coefficients, including:

[0032] The evaluation vector of the wavelet detail coefficients of each layer is calculated by the following formula:

[0033] R=[r n ],1≤n≤N

[0034] in, Wa j,n Represents the nth wavelet detail coefficient of the jth layer.

[0035] Optionally, determining a wavelet decomposition adaptive threshold of the current layer based on the obtained target wavelet detail coefficient includes:

[0036] The wavelet decomposition adaptive threshold of the current layer is calculated by the following formula:

[0037]

[0038] Among them, Wα min is the target wavelet detail coefficient, λ α,j is the adaptive threshold of wavelet decomposition of the jth layer;

[0039] Determine the average of the absolute values ​​of the wavelet detail coefficients of the current layer, including:

[0040] The average of the absolute values ​​of the wavelet detail coefficients of the current layer is calculated by the following formula:

[0041]

[0042] Among them, P a,j Represents the average of the absolute values ​​of the wavelet detail coefficients of the jth layer.

[0043] Optionally, the wavelet decomposition threshold corresponding to each layer is determined based on the wavelet decomposition unified threshold of each layer, the wavelet decomposition adaptive threshold, the average value of the absolute value of the wavelet detail coefficient, the minimum energy level value of the wavelet detail coefficient and the noise standard deviation, including:

[0044] The wavelet decomposition threshold corresponding to each layer is calculated by the following formula:

[0045]

[0046] Among them, ρ N,j It is the minimum energy level value of the wavelet detail coefficient vector of the jth layer.

[0047] Optionally, the wavelet detail coefficients of each layer are adjusted based on the obtained wavelet decomposition threshold, including:

[0048] Calculating adjustment values ​​of wavelet detail coefficients of each layer based on wavelet decomposition thresholds of each layer, and adjusting the wavelet detail coefficients of each layer based on the obtained adjustment values ​​of the wavelet detail coefficients of each layer;

[0049] Among them, the adjustment value of the wavelet detail coefficient of each layer is calculated by the following formula:

[0050]

[0051] in, Represents the adjustment value of the k-th wavelet detail coefficient of the j-th layer.

[0052] In a second aspect of the present application, a fan bearing diagnosis device based on a gradient-weighted class activation heat map is provided, comprising:

[0053] A signal acquisition module is configured to obtain a bearing vibration signal of a wind turbine generator;

[0054] A signal noise reduction module is configured to perform noise reduction processing on the acquired bearing vibration signal;

[0055] The heat map generation module is configured to use short-time Fourier transform to convert the denoised bearing vibration signal into a time-frequency image, take the obtained time-frequency image as input, perform forward propagation and reverse propagation on the input time-frequency image through the feature extraction unit of the fault diagnosis model, and obtain a coarse gradient weighted class activation heat map and a reverse propagation gradient map respectively, and perform image multiplication of the obtained coarse gradient weighted class activation heat map and the reverse propagation gradient map to obtain a final gradient weighted class activation heat map;

[0056] The fault diagnosis module is configured to take the obtained gradient-weighted class activation heat map as input, and the fault diagnosis unit of the fault diagnosis model performs fault diagnosis according to the input gradient-weighted class activation heat map to obtain a diagnosis result, wherein the fault diagnosis unit is obtained by training a preset machine learning algorithm with historical gradient-weighted class activation heat maps corresponding to different bearing faults of the wind turbine generator.

[0057] In a third aspect of the present application, a terminal device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned fan bearing diagnosis method based on the gradient-weighted class activation thermal map when executing the computer program.

[0058] The embodiments provided in this application have the following beneficial effects:

[0059] The present application utilizes vibration data to generate a gradient-weighted class activation heat map based on the time-frequency domain, establishes the connection between the time-frequency region and the fault mode of the wind turbine generator bearing fault, and constructs a convolutional neural network model based on wavelet time-frequency images for generator bearing fault diagnosis to perform fault diagnosis on the collected wind turbine generator bearing vibration signals. It can more accurately extract the characteristic information of the generator bearing fault and realize intelligent diagnosis of the fault. At the same time, the present application uses the wind turbine generator bearing fault data under different working conditions for testing, and analyzes the results based on the time-frequency region that the model pays the most attention to, which can effectively reduce the dimension of the input parameters, reduce the computational complexity of data processing, improve the accuracy of fault diagnosis, and have better robustness.

[0060] Other features and advantages of the embodiments or implementations of the present application will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the following specific implementations, they are used to explain the embodiments of the present application, but do not constitute a limitation on the embodiments of the present application. In the accompanying drawings:

[0062] Figure 1 A method flow chart of a fan bearing diagnosis method based on a gradient-weighted class activation heat map according to an embodiment of the present application is schematically shown;

[0063] Figure 2 The schematic diagram schematically shows the effect of the improved semi-soft threshold denoising method of the embodiment of the present application on the denoising of the collected high-frequency data;

[0064] Figure 3 The schematic diagram of the structure of the fault diagnosis model of the embodiment of the present application is shown schematically;

[0065] Figure 4 A schematic diagram of network structure parameters of an embodiment of the present application is schematically shown;

[0066] Figure 5 The schematic diagram of the implementation principle of the gradient weighted class activation heat map of the embodiment of the present application is schematically shown;

[0067] Figure 6 The following is a schematic diagram of a fault diagnosis process of an embodiment of the present application;

[0068] Figure 7 Schematically shows a schematic block diagram of a fan bearing diagnosis device based on a gradient-weighted class activation heat map according to an embodiment of the present application;

[0069] Figure 8 A schematic diagram of the structure of a terminal device according to an embodiment of the present application is shown schematically.

[0070] Description of Reference Numerals

[0071] 10-terminal device, 100-processor, 101-memory, 102-computer program. DETAILED DESCRIPTION

[0072] The specific implementation of the embodiment of the present application is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the embodiment of the present application, and is not used to limit the embodiment of the present application.

[0073] In order to solve the above problems, Figure 1 As shown, the first aspect of the present application provides a fan bearing diagnosis method based on a gradient weighted class activation heat map, comprising:

[0074] S100, obtaining a bearing vibration signal of a wind turbine generator;

[0075] S200, performing noise reduction processing on the acquired bearing vibration signal;

[0076] S300, using short-time Fourier transform, converting the denoised bearing vibration signal into a time-frequency image, taking the obtained time-frequency image as input, forward propagating and backward propagating the input time-frequency image through the feature extraction unit of the fault diagnosis model, respectively obtaining a coarse gradient weighted class activation heat map and a backward propagation gradient map, and performing image multiplication of the obtained coarse gradient weighted class activation heat map and the backward propagation gradient map to obtain a final gradient weighted class activation heat map;

[0077] S400. Taking the obtained gradient-weighted class activation heat map as input, the fault diagnosis unit of the fault diagnosis model performs fault diagnosis according to the input gradient-weighted class activation heat map to obtain a diagnosis result; wherein the fault diagnosis unit is obtained by training a preset machine learning algorithm with historical gradient-weighted class activation heat maps corresponding to different bearing faults of the wind turbine generator.

[0078] In this way, the present application uses vibration data to generate a gradient-weighted class activation heat map based on the time-frequency domain, establishes the connection between the time-frequency area and the fault mode of the wind turbine generator bearing fault, and constructs a convolutional neural network model based on wavelet time-frequency images for generator bearing fault diagnosis. The model performs fault diagnosis on the collected wind turbine generator bearing vibration signals, which can more accurately extract the characteristic information of the generator bearing fault and realize intelligent diagnosis of the fault. At the same time, the present application uses the wind turbine generator bearing fault data under different working conditions for testing, and analyzes the results based on the time-frequency area that the model pays the most attention to, which can effectively reduce the dimension of the input parameters, reduce the computational complexity of data processing, improve the accuracy of fault diagnosis, and have better robustness.

[0079] In step S100, the bearing vibration signal of the wind turbine generator includes: one or more of the radial vertical vibration signal of the front bearing of the wind turbine generator, the radial horizontal vibration signal of the front bearing, the radial vertical vibration signal of the rear bearing, the radial horizontal vibration signal of the rear bearing, the vertical envelope vibration signal of the generator transmission, and the horizontal envelope vibration signal of the generator transmission. Among them, the vertical envelope vibration signal of the generator transmission and the horizontal envelope vibration signal of the generator transmission refer to the signal envelope data extracted by performing envelope detection on the vibration signal in the vertical direction of the generator transmission chain, and the signal envelope data extracted by performing envelope detection on the vibration signal in the horizontal direction of the generator transmission chain. It can be understood that in a wind turbine, the vibration of the generator transmission chain in the vertical direction is often related to faults such as rotor imbalance, bearing damage, and gear meshing problems, while the vibration of the generator transmission chain in the horizontal direction is usually related to faults such as misalignment, loose foundation, and shaft bending. It can be understood that in this application, the collected data may also include the input shaft speed and generator power of the wind turbine generator.

[0080] In step S200, the acquired bearing vibration signal is subjected to noise reduction processing, including:

[0081] S210, performing wavelet decomposition of a scale J on the acquired bearing vibration signal to obtain wavelet detail coefficients and approximate coefficients at each scale;

[0082] S220, determining the noise standard deviation of the current layer based on the wavelet detail coefficients of each layer, and determining the noise intensity of each layer according to the obtained noise standard deviation;

[0083] S230, determining a unified threshold of wavelet decomposition of each layer based on the noise standard deviation of each layer;

[0084] S240, sorting the wavelet detail coefficients of each layer from small to large according to the absolute value, calculating the evaluation vector of the wavelet detail coefficient of each layer based on the sorted wavelet detail coefficients, sorting the obtained evaluation vectors from large to small, taking the minimum value as the approximation error, determining the target wavelet detail coefficient corresponding to the approximation error, and determining the wavelet decomposition adaptive threshold of the current layer based on the obtained target wavelet detail coefficient;

[0085] S250, determining the average value of the absolute value of the wavelet detail coefficient of the current layer and the minimum energy level value of the wavelet detail coefficient;

[0086] S260, determining the wavelet decomposition threshold corresponding to each layer based on the wavelet decomposition unified threshold of each layer, the wavelet decomposition adaptive threshold, the average value of the absolute value of the wavelet detail coefficient, the minimum energy level value of the wavelet detail coefficient and the noise standard deviation;

[0087] S270, adjusting the wavelet detail coefficients of each layer based on the obtained wavelet decomposition threshold, performing wavelet reconstruction on the adjusted wavelet detail coefficients of each layer, and obtaining a bearing vibration signal after noise reduction.

[0088] In step S210, the collected original signal x(t) is discretized and J-layer wavelet decomposition is performed by setting a high-pass filter h and a low-pass filter g. The process of performing wavelet decomposition on the signal x(t) is the process of performing multi-resolution analysis on the signal, that is, the process of gradually decomposing the original signal from scale j to j+1, that is, the process from low resolution to high resolution, and finally decomposing the original signal into a high-frequency signal A j x (wavelet detail coefficient or detail signal) and a low-frequency signal D j x (approximate coefficient or approximate signal). By performing J-layer decomposition on the original signal, the wavelet detail coefficients and the amplitude A of the wavelet detail signal at each layer of decomposition can be obtained. j It can be understood that in wavelet decomposition, the decomposition scale and the number of layers are one-to-one corresponding. Each time a layer of decomposition is performed, the low-frequency signal obtained by the previous decomposition is decomposed into two parts: high-frequency component and low-frequency component.

[0089] In this application, the process of wavelet decomposition of the original can be expressed as the following formula:

[0090]

[0091] Where h(n) is the impulse response function, g(n) = (-1) 1-n h(1-n).

[0092] In step S220, the noise standard deviation σ of each layer is calculated according to the detail coefficient of the signal of each layer. j and the noise intensity coefficient Γ(σ j ), wherein the noise standard deviation of each layer is determined based on the wavelet detail coefficient of each layer, including:

[0093] The noise standard deviation of each layer is calculated by the following formula:

[0094]

[0095] Among them, σ j represents the noise standard deviation of the jth layer, Wa j,k represents the kth wavelet detail coefficient of the jth layer, N represents the total number of wavelet detail coefficients at the jth scale, and k is the number of wavelet detail coefficients at the current scale.

[0096] The noise intensity of the current layer is determined according to the obtained noise standard deviation, including:

[0097]

[0098] Among them, Γ(σ j ) represents the noise intensity of the j-th layer wavelet detail coefficient, A j Represents the amplitude of the j-th layer wavelet detail coefficient.

[0099] In step S230, the wavelet decomposition unified threshold of each layer is determined based on the noise standard deviation of each layer, including:

[0100] The unified threshold of wavelet decomposition of each layer is calculated by the following formula:

[0101]

[0102] Among them, λ 1,j It represents the uniform threshold of the wavelet decomposition of the jth layer. It can be understood that the setting of the threshold can also be changed according to the actual situation according to the different signal-to-noise ratios of the original signal.

[0103] In step S240, after the original signal is decomposed by wavelet with a scale of J, J groups of high-frequency signal coefficients, i.e., J groups of wavelet detail coefficients, are obtained. Each group of obtained wavelet detail coefficients is arranged from small to large according to the absolute value, and a vector P=[Wa j,n ], 1≤n≤N, and use the obtained vector to calculate the evaluation vector of the wavelet detail coefficient of the jth layer. Specifically, the evaluation vector of the wavelet detail coefficient of each layer is calculated based on the sorted wavelet detail coefficient, including:

[0104] The evaluation vector of the wavelet detail coefficients of each layer is calculated by the following formula:

[0105] R=[r n ],1≤n≤N

[0106] in, Wa j,n Represents the nth wavelet detail coefficient of the jth layer.

[0107] After calculating the evaluation vector of the wavelet detail coefficient of the j-th layer, the obtained evaluation vectors are sorted from large to small, and the minimum value of the evaluation vector is used as the approximation error. The wavelet detail coefficient corresponding to the minimum value is determined as the target wavelet detail coefficient, and the wavelet decomposition adaptive threshold of the j-th layer is calculated based on the target wavelet detail coefficient. The calculation formula of the wavelet decomposition adaptive threshold is: Among them, Wα min is the target wavelet detail coefficient, λ α,j is the adaptive threshold of wavelet decomposition of the jth layer.

[0108] In step S250, determining the average value of the absolute values ​​of the wavelet detail coefficients of the current layer includes: calculating the average value of the absolute values ​​of the wavelet detail coefficients of the current layer by the following formula:

[0109]

[0110] Among them, P a,j Represents the average of the absolute values ​​of the wavelet detail coefficients of the jth layer.

[0111] Among them, the minimum energy level value ρ of the wavelet detail coefficient N,j It can be determined according to actual conditions, for example, directly determined based on experience or statistical analysis.

[0112] In step S260, the wavelet decomposition threshold corresponding to each layer is determined based on the wavelet decomposition unified threshold of each layer, the wavelet decomposition adaptive threshold, the average value of the absolute value of the wavelet detail coefficient, the minimum energy level value of the wavelet detail coefficient and the noise standard deviation, including:

[0113] The wavelet decomposition threshold corresponding to each layer is calculated by the following threshold selection function:

[0114]

[0115] Among them, ρ N, j is the minimum energy level value of the wavelet detail coefficient vector of the jth layer.

[0116] In step S270, the wavelet detail coefficients of each layer are adjusted based on the obtained wavelet decomposition threshold, including: calculating the adjustment value of the wavelet detail coefficients of each layer based on the wavelet decomposition threshold of each layer, and adjusting the wavelet detail coefficients of each layer based on the obtained adjustment value of the wavelet detail coefficients of each layer.

[0117] Among them, the adjustment value of the wavelet detail coefficient of each layer is calculated by the following formula:

[0118]

[0119] in, Represents the adjustment value of the k-th wavelet detail coefficient of the j-th layer.

[0120] After calculating the wavelet decomposition threshold of each layer through the above steps, compare the wavelet detail coefficient obtained by decomposing the layer with the obtained wavelet decomposition threshold, determine the adjustment value of the wavelet detail coefficient according to the comparison result, and replace the wavelet detail coefficient obtained by decomposing the current layer with the determined adjustment value of the wavelet detail coefficient to complete the adjustment of the wavelet detail coefficient of the current layer. Finally, after adjusting the wavelet detail coefficients of each layer, perform wavelet reconstruction to obtain the bearing vibration signal after noise reduction. Convert the bearing vibration signal after noise reduction into a time-frequency image, extract features through a preset fault diagnosis model, and perform fault diagnosis. Among them, a schematic diagram of the effect of denoising the collected high-frequency data by the improved semi-soft threshold denoising method of the present application is shown in the figure. Figure 2 shown.

[0121] In this application, the fault diagnosis model includes a feature extraction unit and a fault diagnosis unit. In this application, the feature extraction unit is a CNN model, and the fault diagnosis unit is a BiLSTM model. This application forms a new feature vector by connecting the spatial features extracted by the CNN layer and the temporal features extracted by the BiLSTM layer, and further processes the fused feature vector using a fully connected layer, and outputs it to the final classification layer for fault diagnosis. Figure 3As shown, the feature extraction unit of the present application includes multiple convolutional layers and maximum pooling layers, and the fault diagnosis unit is composed of a BiLSTM layer and a Dropout layer connected in sequence, and the Dropout layer is sequentially connected with a fully connected layer and a Softmax output layer. For example, the feature extraction unit can be composed of a convolutional layer 1, a maximum pooling layer, a convolutional layer 2, a maximum pooling layer, a convolutional layer 3, a convolutional layer 4, a convolutional layer 5, a convolutional layer 6, a maximum pooling layer, a fully connected layer 1, a Dropout layer, a fully connected layer 2, a Dropout layer, a fully connected layer 3 and a Softmax output layer connected in sequence, and its network structure parameters are as follows: Figure 4 As shown. It can be understood that the specific structure and network structure parameters of the CNN model can be adjusted as needed.

[0122] In step S300, Figure 5 and Figure 6 As shown in the figure, after obtaining the de-noised bearing vibration signal, first use the short-time Fourier transform to plot the de-noised high-frequency data as a 2D time-frequency diagram. For example, the Morlet basis function can be used to perform wavelet transform on the collected high-frequency data, the frequency range can be set to half the sampling frequency, 512 frequency points are taken at equal intervals, the obtained wavelet coefficient matrix is ​​converted into a 512×512 grayscale image, and the input channel of the feature extraction unit of the fault diagnosis model is adjusted to 1 to adapt to the data sample. Secondly, the generated grayscale image of size 512×512 is converted into a vector of shape 1×512, and the vector is input into the feature extraction unit of the fault diagnosis model for forward propagation, and the coarse gradient weighted class activation heat map (blue heat map) is calculated. The original category score S is generated after Softmax normalization. It can be understood that the gradients of all categories are set to 0 except for the required categories such as "fault features" which are set to 1. The signal is then back-propagated to obtain the corrected convolution feature map. Finally, the obtained coarse gradient weighted class activation heat map heat map is dot-multiplied with the back-propagated gradient map to obtain the visualization result of the gradient weighted class activation heat map with high resolution and category distinction.

[0123] Specifically, the implementation principle of the gradient weighted class activation heat map is as follows:

[0124] For a given 2D time-frequency diagram, where f k (x, y) represents the feature map of the kth layer in the last convolution layer at the spatial position (x, y), and global average pooling is performed on it to obtain To calculate the gradient weighted class activation heat map of fault category c, we first need to calculate the softmax classification score f c , for a feature map A of a convolutional layer k Gradient These reflowing gradients are globally averaged pooled in the width and height dimensions (indexed by i and j, respectively) to obtain the importance weights of neurons, and then weighted and processed, and then passed through ReLU to obtain the gradient weighted class activation heat map. By superimposing the class activation map on the original image in the form of a heat map to intuitively display the key areas in the image, it can help understand the classification basis of the network model. The redder the color, the more attention the model pays to the area, so that when the sample is intuitive or has a clear physical meaning, it can be used to judge whether the network has learned the correct information for image classification.

[0125] Among them, the process of global average pooling can be expressed as follows:

[0126]

[0127] in, is the average gradient of the k-th channel feature map in the c-th output back propagation, Z is the total number of pixels in the feature map, and y c is the output value of the largest class of output, is the feature map corresponding to the kth channel.

[0128] The calculation process of the gradient weighted class activation heat map can be expressed as follows:

[0129]

[0130] in, It is the class activation map obtained by gradient weighted class activation heat map, and ReLU is the nonlinear activation function used to remove negative values.

[0131] In step S400, the gradient-weighted class activation heat map generated in step S300 has extracted local features of the data through convolution operations, and reduced the dimension of the data through pooling operations, while retaining important features; on this basis, the present application inputs the one-dimensional time-frequency features extracted by the CNN layer into the BiLSTM layer, based on the forward and backward information in the sequence, so as to more accurately capture the dependencies in the sequence. The present application forms a new feature vector by connecting the spatial features extracted by the CNN layer and the temporal features extracted by the BiLSTM layer, further processes the fused features using a fully connected layer, and outputs them to the final classification layer for fault diagnosis output. Among them, the calculation of the fully connected layer can be expressed as: y=Wx+b, y is the output, W is the weight matrix, x is the input feature, and b is the bias term. The output layer uses a softmax classifier to classify the fault state of the generator bearing, where the formula for softmax is: Where p(y=jx) is the probability of outputting category j given input x, and c is the total number of categories. It can be understood that the implementation principle of the CNN layer and the BiLSTM layer is the prior art, which will not be repeated here.

[0132] In the present application, the fault diagnosis unit is obtained by training the BiLSTM model with the historical gradient weighted class activation heat map corresponding to different bearing faults of the wind turbine generator. The training process is a prior art and is not limited here.

[0133] It is understandable that the method of the present application can be called and executed by online or offline data analysis platforms and servers, by calling the fault diagnosis model to identify fault anomalies and output detection and warning results for the historical operation data of the unit before the execution time, such as within 24 hours. Among them, the model calling code script can be an executable data analysis script such as C, C++, Matlab, R, Scala, etc., and the script model timing operation cycle can be determined according to the application requirements and the amount of data to be processed. For example, the script operation cycle can be half a day, one day, one week, one month, etc. The script model can be called and executed online or offline by a variety of big data analysis platforms, data mining systems, fault diagnosis servers, early warning monitoring systems, etc.

[0134] like Figure 7 As shown, in a second aspect of the present application, a fan bearing diagnosis device based on a gradient weighted class activation heat map is provided, comprising:

[0135] A signal acquisition module is configured to obtain a bearing vibration signal of a wind turbine generator;

[0136] A signal noise reduction module is configured to perform noise reduction processing on the acquired bearing vibration signal;

[0137] The heat map generation module is configured to use short-time Fourier transform to convert the denoised bearing vibration signal into a time-frequency image, take the obtained time-frequency image as input, perform forward propagation and reverse propagation on the input time-frequency image through the feature extraction unit of the fault diagnosis model, and obtain a coarse gradient weighted class activation heat map and a reverse propagation gradient map respectively, and perform image multiplication of the obtained coarse gradient weighted class activation heat map and the reverse propagation gradient map to obtain a final gradient weighted class activation heat map;

[0138] The fault diagnosis module is configured to take the obtained gradient-weighted class activation heat map as input, and the fault diagnosis unit of the fault diagnosis model performs fault diagnosis based on the input gradient-weighted class activation heat map to obtain a diagnosis result, wherein the fault diagnosis unit is obtained by training a preset machine learning algorithm based on the historical gradient-weighted class activation heat map corresponding to different bearing faults of the wind turbine generator.

[0139] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0140] In a third aspect of the present application, a terminal device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned fan bearing diagnosis method based on the gradient-weighted class activation thermal map when executing the computer program.

[0141] like Figure 8 FIG. 1 is a schematic diagram of a terminal device provided in an embodiment of the present application. Figure 8 As shown, the terminal device 10 of this embodiment includes: a processor 100, a memory 101, and a computer program 102 stored in the memory 101 and executable on the processor 100. When the processor 100 executes the computer program 102, the steps in the above method embodiment are implemented. Alternatively, when the processor 100 executes the computer program 102, the functions of each module / unit in the above device embodiments are implemented.

[0142] Exemplarily, the computer program 102 may be divided into one or more modules / units, one or more modules / units are stored in the memory 101, and are executed by the processor 100 to complete the present application. One or more modules / units may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program 102 in the terminal device 10.

[0143] The terminal device 10 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The terminal device 10 may include, but is not limited to, a processor 100 and a memory 101. Those skilled in the art will appreciate that Figure 8It is only an example of the terminal device 10 and does not constitute a limitation of the terminal device 10. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the terminal device may also include input and output devices, network access devices, buses, etc.

[0144] The processor 100 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.

[0145] The memory 101 may be an internal storage unit of the terminal device 10, such as a hard disk or memory of the terminal device 10. The memory 101 may also be an external storage device of the terminal device 10, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal device 10. Further, the memory 101 may also include both an internal storage unit of the terminal device 10 and an external storage device. The memory 101 is used to store computer programs and other programs and data required by the terminal device 10. The memory 101 may also be used to temporarily store data that has been output or is to be output.

[0146] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0147] In summary, this application collects the vibration data of the generator bearings of each wind turbine set, and denoises the collected high-frequency data through an improved semi-soft threshold denoising method to generate a 2D time-frequency diagram of the high-frequency data and a corresponding gradient-weighted class activation heat map. The reconstructed image is extracted from the image features of the wind turbine generator bearings under different fault operating states through a fault diagnosis model built based on the CNN-BiLSTM classifier, which can effectively reduce the dimension of the input parameters, reduce the computational complexity of data processing, and improve the accuracy of fault diagnosis.

[0148] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0149] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A fan bearing diagnosis method based on gradient weighted class activation heat map, characterized in that: include: Obtain bearing vibration signals of wind turbine generators; Perform noise reduction processing on the acquired bearing vibration signal; The denoised bearing vibration signal is converted into a time-frequency image by using short-time Fourier transform. The obtained time-frequency image is used as input. The input time-frequency image is forward propagated and backward propagated through the feature extraction unit of the fault diagnosis model to obtain a coarse gradient weighted class activation heat map and a back-propagation gradient map, respectively. The coarse gradient weighted class activation heat map is multiplied with the back-propagation gradient map to obtain the final gradient weighted class activation heat map. Taking the obtained gradient weighted class activation heat map as input, the fault diagnosis unit of the fault diagnosis model performs fault diagnosis according to the input gradient weighted class activation heat map to obtain a diagnosis result; The fault diagnosis unit is obtained by training a preset machine learning algorithm based on the historical gradient weighted class activation heat map corresponding to different bearing faults of the wind turbine generator.

2. The fan bearing diagnosis method based on gradient weighted class activation heat map according to claim 1 is characterized in that: The bearing vibration signal of the wind turbine generator includes: The wind turbine generator may include one or more of a front bearing radial vertical vibration signal, a front bearing radial horizontal vibration signal, a rear bearing radial vertical vibration signal, a rear bearing radial horizontal vibration signal, a generator transmission vertical envelope vibration signal, and a generator transmission horizontal envelope vibration signal.

3. The wind turbine bearing diagnosis method based on gradient weighted class activation heat map according to claim 1 is characterized in that: The acquired bearing vibration signal is subjected to noise reduction processing, including: Perform wavelet decomposition of scale J on the acquired bearing vibration signal to obtain wavelet detail coefficients and approximate coefficients at each scale; Determine the noise standard deviation of the current layer based on the wavelet detail coefficients of each layer, and determine the noise intensity of each layer based on the obtained noise standard deviation; Determine the unified threshold of wavelet decomposition of each layer based on the standard deviation of noise of each layer; The wavelet detail coefficients of each layer are sorted from small to large according to the absolute value, and the evaluation vectors of the wavelet detail coefficients of each layer are calculated based on the sorted wavelet detail coefficients. The obtained evaluation vectors are sorted from large to small, and the minimum value is used as the approximation error. The target wavelet detail coefficient corresponding to the approximation error is determined, and the wavelet decomposition adaptive threshold of the current layer is determined based on the obtained target wavelet detail coefficient; Determine the average value of the absolute value of the wavelet detail coefficient of the current layer and the minimum energy level value of the wavelet detail coefficient; The wavelet decomposition threshold corresponding to each layer is determined based on the wavelet decomposition unified threshold of each layer, the wavelet decomposition adaptive threshold, the average value of the absolute value of the wavelet detail coefficient, the minimum energy level value of the wavelet detail coefficient and the noise standard deviation; The wavelet detail coefficients of each layer are adjusted based on the obtained wavelet decomposition threshold, and the wavelet detail coefficients of each layer after the adjustment are reconstructed to obtain the bearing vibration signal after noise reduction.

4. The wind turbine bearing diagnosis method based on gradient weighted class activation heat map according to claim 3 is characterized in that: The noise standard deviation of each layer is determined based on the wavelet detail coefficient of each layer, including: The noise standard deviation of each layer is calculated by the following formula: Among them, σ j represents the noise standard deviation of the jth layer, Wa j,k represents the kth wavelet detail coefficient of the jth layer, N represents the total number of wavelet detail coefficients at the jth scale, and k is the number of wavelet detail coefficients at the current scale; The noise intensity of the current layer is determined based on the obtained noise standard deviation, including: Among them, Γ(σ j ) represents the noise intensity of the j-th layer wavelet detail coefficient, A j Represents the amplitude of the j-th layer wavelet detail coefficient.

5. The wind turbine bearing diagnosis method based on gradient weighted class activation heat map according to claim 4 is characterized in that: The unified threshold of wavelet decomposition of each layer is determined based on the noise standard deviation of each layer, including: The unified threshold of wavelet decomposition of each layer is calculated by the following formula: Among them, λ 1,j represents the uniform threshold of wavelet decomposition at the jth layer; The evaluation vector of the wavelet detail coefficients of each layer is calculated based on the sorted wavelet detail coefficients, including: The evaluation vector of the wavelet detail coefficients of each layer is calculated by the following formula: R=[r n ],1≤n≤N in, Wa j,n Represents the nth wavelet detail coefficient of the jth layer.

6. The wind turbine bearing diagnosis method based on gradient weighted class activation heat map according to claim 5 is characterized in that: The adaptive threshold of wavelet decomposition of the current layer is determined based on the obtained target wavelet detail coefficient, including: The wavelet decomposition adaptive threshold of the current layer is calculated by the following formula: Among them, Wα min is the target wavelet detail coefficient, λ α,j is the adaptive threshold of wavelet decomposition of the jth layer; Determine the average of the absolute values ​​of the wavelet detail coefficients of the current layer, including: The average of the absolute values ​​of the wavelet detail coefficients of the current layer is calculated by the following formula: Among them, P a,j Represents the average of the absolute values ​​of the wavelet detail coefficients of the jth layer.

7. The wind turbine bearing diagnosis method based on gradient weighted class activation heat map according to claim 6 is characterized in that: The wavelet decomposition threshold corresponding to each layer is determined based on the wavelet decomposition unified threshold of each layer, the wavelet decomposition adaptive threshold, the average value of the absolute value of the wavelet detail coefficient, the minimum energy level value of the wavelet detail coefficient and the noise standard deviation, including: The wavelet decomposition threshold corresponding to each layer is calculated by the following formula: Among them, ρ N,j It is the minimum energy level value of the wavelet detail coefficient vector of the jth layer.

8. The wind turbine bearing diagnosis method based on gradient weighted class activation heat map according to claim 7 is characterized in that: The wavelet detail coefficients of each layer are adjusted based on the obtained wavelet decomposition threshold, including: Calculating adjustment values ​​of wavelet detail coefficients of each layer based on wavelet decomposition thresholds of each layer, and adjusting the wavelet detail coefficients of each layer based on the obtained adjustment values ​​of the wavelet detail coefficients of each layer; Among them, the adjustment value of the wavelet detail coefficient of each layer is calculated by the following formula: in, Represents the adjustment value of the k-th wavelet detail coefficient of the j-th layer.

9. A fan bearing diagnosis device based on gradient weighted class activation heat map, characterized in that: include: A signal acquisition module is configured to obtain a bearing vibration signal of a wind turbine generator; A signal noise reduction module is configured to perform noise reduction processing on the acquired bearing vibration signal; The heat map generation module is configured to use short-time Fourier transform to convert the denoised bearing vibration signal into a time-frequency image, take the obtained time-frequency image as input, perform forward propagation and reverse propagation on the input time-frequency image through the feature extraction unit of the fault diagnosis model, and obtain a coarse gradient weighted class activation heat map and a reverse propagation gradient map respectively, and perform image multiplication of the obtained coarse gradient weighted class activation heat map and the reverse propagation gradient map to obtain a final gradient weighted class activation heat map; The fault diagnosis module is configured to take the obtained gradient-weighted class activation heat map as input, and the fault diagnosis unit of the fault diagnosis model performs fault diagnosis according to the input gradient-weighted class activation heat map to obtain a diagnosis result, wherein the fault diagnosis unit is obtained by training a preset machine learning algorithm with historical gradient-weighted class activation heat maps corresponding to different bearing faults of the wind turbine generator.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the fan bearing diagnosis method based on the gradient weighted class activation thermal map described in any one of claims 1 to 8 is implemented.

Images