Gearbox fault diagnosis method and system

Abstract

The present invention discloses a gearbox fault diagnosis method and system. The method obtains horizontal vibration signals and vertical vibration signals of the gearbox shaft through an acceleration sensor set in a wind turbine, extracts characteristic data of the horizontal vibration signals and vertical vibration signals of the gearbox shaft, and inputs the extracted characteristic data into a deep convolutional neural network module to obtain a fault diagnosis result. The deep convolutional neural network module is a fault diagnosis model established by learning in advance the characteristic data of horizontal and vertical vibration signals of the gearbox shaft with different gearbox faults. The characteristic data extraction of the horizontal vibration signals and vertical vibration signals of the gearbox shaft includes: performing multi-scale time domain feature extraction and deep noise reduction feature extraction. The present invention improves the accuracy and automation level of wind turbine gearbox fault diagnosis, can monitor the working condition of the wind turbine gearbox in real time, and reduces the labor intensity of personnel.

Description

Technical Field

[0001] The present invention relates to a gearbox fault diagnosis method and system, which is a gearbox fault diagnosis method and system of a multi-feature fusion convolutional neural network. Background Art

[0002] Gearboxes play an important role in wind turbines to realize transmission, speed change, speed regulation and other functions. Especially with the development of wind power technology, in order to better realize the conversion of wind energy to mechanical energy and then to electrical energy, gearboxes are also developing in the direction of complexity and precision, resulting in complex vibration signals generated by fault points and increased difficulty in fault diagnosis. Traditional fault diagnosis methods generally first use feature extraction, such as time-frequency analysis to analyze the time-frequency characteristics of vibration signals, and extract features through Fourier transform, wavelet transform and other methods, and then classify and diagnose faults through expert systems, support vector machines and other models. Traditional methods have strong theoretical support, but in the process of processing nonlinear and non-stationary vibration signals, they often cannot deeply dig into deep information, resulting in misjudgment of faults and unnecessary economic losses. Summary of the invention

[0003] The purpose of the present invention is to provide a gearbox fault diagnosis method and system, which is a gearbox fault diagnosis method and system based on a multi-feature fusion convolutional neural network.

[0004] In order to achieve the above object, the technical solution of the present invention is:

[0005] A gearbox fault diagnosis method is used for wind turbine gearbox fault diagnosis. The method obtains horizontal vibration signals and vertical vibration signals of the gearbox shaft by an acceleration sensor arranged on the wind turbine, extracts characteristic data of the horizontal vibration signals and vertical vibration signals of the gearbox shaft, and inputs the extracted characteristic data into a deep convolutional neural network module to obtain a fault diagnosis result. The deep convolutional neural network module is a fault diagnosis model established by learning in advance the characteristic data of the horizontal and vertical vibration signals of the gearbox shaft for different gearbox faults. The characteristic data extraction of the horizontal and vertical vibration signals of the gearbox shaft includes: performing multi-scale time domain feature extraction and deep noise reduction feature extraction:

[0006] The multi-scale time domain feature extraction step is:

[0007] The first step is to decompose the vibration signal into stationary signal IMF components according to the set time length, and obtain the energy of each IMF component through formula (1).

[0008]

[0009] Where: N means the signal is divided into N equal parts, each part has a time length of △t, ci (tn) represents the amplitude of the i-th component at time tn, and h is the total number of IMF components;

[0010] The second step is to select the first S IMF components sorted from large to small in energy, and calculate the kurtosis index of the S IMF components using formula (2):

[0011]

[0012] The third step is to take the first P IMF components sorted from large to small in terms of kurtosis value. As a result of multi-scale time domain feature extraction, where Nc is the length of the signal, Indicates that the IMF component is 1xNc of real dimension, P is less than S;

[0013] The deep noise reduction feature extraction is: sending the vibration signal to the deep noise reduction autoencoder, and outputting the deep noise reduction feature through the deep noise reduction autoencoder.

[0014] The solution is further that: the horizontal vibration signal includes an axial horizontal vibration signal and a radial horizontal vibration signal.

[0015] The scheme further includes: the steps of establishing the fault diagnosis model are:

[0016] The first step is to perform the multi-scale time domain feature extraction and deep noise reduction feature extraction according to different faults;

[0017] The second step is to fuse the multi-scale time domain features and the deep noise reduction features: each IMF component in the multi-scale time domain features is concatenated with the deep noise reduction features, and these IMF components are stacked into a multi-channel training sample in time sequence, and a deep convolutional neural network is constructed and trained. The training of the convolutional neural network is completed through the input of the training samples.

[0018] Each convolutional layer outputs the feature map of formula (3) to complete feature fusion:

[0019]

[0020] Where * represents convolution, M i , i=1,2,...,P,P+1 represents the convolution kernel of P+1 channels;

[0021] The third step is to output the fault diagnosis result based on the probability of occurrence of each fault type formed by feature fusion.

[0022] The solution is further that: the feature data extraction is feature data extraction of multiple sample signals of the vibration signal, and the multiple sample signals are obtained by dividing the length of the vibration signal into windows of equal length.

[0023] The solution is further: the method further includes marking the device name, number of sampling points, sensitivity, and acquisition time parameters during the acquisition process of each group of multiple sample signals, and checking whether the corresponding fields of each group of multiple sample signals have missing values ​​and whether the fields are complete.

[0024] The solution is further that: the method of decomposing the vibration signal into the IMF components of the stationary signal according to the set time length is:

[0025] The first step is to define the total number of EMDs to be performed MAX_ITER;

[0026] In the second step, normally distributed white noise is added to the original signal, and the signal to be processed is decomposed by empirical mode decomposition (EMD) to obtain the IMF component C ij , C ij represents the i IMF components decomposed for the jth time, and the above operation is repeated until the number of empirical mode decompositions reaches MAX_ITER;

[0027] The third step is to calculate the IMF component corresponding to each time point and find its average value according to formula (4).

[0028]

[0029] Finally, they are spliced ​​in chronological order. As the final obtained stationary signal IMF component.

[0030] A gearbox fault diagnosis system for implementing the method comprises: a gear transmission box of a wind turbine, an input shaft of the gear transmission box connected to a wind turbine rotating blade hub, an output shaft of the gear transmission box connected to a generator, characterized in that an acceleration sensor is respectively arranged on a bearing seat where the input shaft is docked with the wind turbine rotating blade hub, the gear transmission box and a generator housing to sense a horizontal vibration signal and a vertical vibration signal of the gear shaft, the acceleration sensor is connected to a fault diagnosis and analysis computer via a wired or wireless connection, a signal receiving interface card is arranged in the computer, the signal receiving interface card is connected to the acceleration sensor to send the received horizontal vibration signal and vertical vibration signal of the gear to a central processing unit, and a vibration signal preprocessing module, a feature data extraction module and a deep convolutional neural network module are arranged in the central processing unit;

[0031] The vibration signal preprocessing module divides the length of the acquired horizontal vibration signal and vertical vibration signal of the gear into multiple sample signals using windows of equal length, records the device name, number of sampling points, sensitivity, and acquisition time parameters of the corresponding vibration signal, and checks whether the corresponding fields of each group of multiple sample signals have vacant values ​​and whether the fields are complete;

[0032] The feature data extraction module extracts features from the sample signal provided by the preprocessing module;

[0033] The deep convolutional neural network module performs fault diagnosis on the feature data extracted by the feature data extraction module and outputs the diagnosis results.

[0034] The scheme is further: the acceleration sensor includes: an input shaft axial horizontal vibration sensor arranged on the axial end face of the bearing seat where the input shaft is connected to the wind turbine rotating blade hub, an input shaft radial horizontal vibration sensor arranged on the horizontal radial end face of the bearing seat, a box vertical vibration sensor and a box radial horizontal vibration sensor arranged on the gear transmission box, and a generator shaft radial horizontal vibration sensor respectively arranged on the input and output shafts of the generator.

[0035] A further solution is: the box vertical vibration sensor is arranged in the middle of the upper side of the box, and there are two box radial horizontal vibration sensors, which are arranged on both sides of the box vertical vibration sensor on the side of the box.

[0036] A further solution is that the acceleration sensor is fixed at a set position by embedding.

[0037] The beneficial effects of the present invention are:

[0038] The system of the present invention can diagnose the fault of the wind turbine gearbox, improve the accuracy and automation level of the wind turbine gearbox fault diagnosis, monitor the working condition of the wind turbine gearbox in real time, and reduce the labor intensity of personnel.

[0039] The present invention uses different methods to extract fault features in vibration signals. It uses the method of ensemble empirical mode decomposition to decompose the original signal and extract the IMF component with obvious fault feature information. In addition, it uses the deep denoising autoencoder to extract the deep denoising features in the original signal, which makes up for the insufficient feature extraction ability of a single method.

[0040] The method uses a deep convolutional neural network for feature fusion and fault diagnosis. Since the components after feature extraction are difficult to model uniformly, the deep convolutional neural network can mine the correlation between them and establish a complex mapping with the fault type. By training and updating the model parameters, the accuracy and robustness of the model are improved, which improves the accuracy of wind turbine fault diagnosis to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a schematic diagram of the distribution of acceleration sensors in the system of the present invention;

[0042] Figure 2 It is a block diagram of the system module of the present invention;

[0043] Figure 3 This is a multi-scale feature extraction flow chart of the present invention;

[0044] Figure 4 Output 5 IMF components for the multi-scale time domain feature extraction module;

[0045] Figure 5 It is the first layer of convolution for fusion features. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0047] A gearbox fault diagnosis method technical solution is used for wind turbine gearbox fault diagnosis. The method is implemented based on a gearbox fault diagnosis system. Figure 1 and Figure 2 As shown, the fault diagnosis system includes: a gear transmission box 1 of a wind turbine generator, an input shaft 101 of the gear transmission box 1 is connected to a wind turbine rotating blade hub 2, and an output shaft 102 of the gear transmission box is connected to a generator 3, wherein an acceleration sensor is respectively arranged on a bearing seat 8 where the input shaft is connected to the wind turbine rotating blade hub, the gear transmission box 1 and a housing of the generator 3 to sense a horizontal vibration signal and a vertical vibration signal of the gear shaft, the acceleration sensor is connected to a fault diagnosis and analysis computer (not shown in the figure) via a wired or wireless connection, a signal receiving interface card is arranged in the computer, and the signal receiving interface card is connected to the acceleration sensor to send the received horizontal vibration signal and vertical vibration signal of the gear to a central processing unit, such as Figure 2 As shown, a vibration data signal acquisition and preprocessing module 5, a feature data extraction module 6 and a deep convolutional neural network module 7, i.e., a multi-feature fusion convolutional neural network module, are provided in the central processing unit;

[0048] The vibration data signal acquisition and preprocessing module acquires vibration signals, and divides the lengths of the acquired horizontal vibration signals and vertical vibration signals of the gear shaft into multiple sample signals using windows of equal length, records the device name, number of sampling points, sensitivity, and acquisition time parameters of the corresponding vibration signals, and checks whether the corresponding fields of each group of multiple sample signals have vacant values ​​and whether the fields are complete;

[0049] The feature data extraction module performs multi-scale time-domain IMF component feature extraction and deep noise reduction feature extraction on the sample signal provided by the preprocessing module;

[0050] The deep convolutional neural network module performs fault diagnosis on the feature data extracted by the feature data extraction module and outputs the diagnosis results.

[0051] The acceleration sensor includes: an input shaft axial horizontal vibration sensor 9 arranged on the axial end face of a bearing seat 8 where the input shaft is connected to the hub of a wind turbine rotating blade, an input shaft radial horizontal vibration sensor 10 arranged on the horizontal radial end face of the bearing seat, a housing vertical vibration sensor 11 and a housing radial horizontal vibration sensor 12 arranged on the gear transmission housing, and a generator shaft radial horizontal vibration sensor 13 respectively arranged on the input and output shafts of the generator.

[0052] Among them: the input shaft axial horizontal vibration sensor 9 set on the axial end face of the bearing seat 8 and the input shaft radial horizontal vibration sensor 10 set on the horizontal radial end face of the bearing seat are set at an angle of 90 degrees, the box vertical vibration sensor 11 is set in the middle of the upper side of the box, and there are two box radial horizontal vibration sensors 12, which are set on both sides of the box vertical vibration sensor on the side of the box. In order to fix it firmly, the acceleration sensor is fixed at the set position by inlaying, that is, it is fixed by inlaying through the processing groove.

[0053] The method obtains horizontal vibration signals and vertical vibration signals of the gearbox shaft by an acceleration sensor arranged on a wind turbine, extracts characteristic data of the horizontal vibration signals and vertical vibration signals of the gearbox shaft, and inputs the extracted characteristic data into a deep convolutional neural network module to obtain a fault diagnosis result. The deep convolutional neural network module is a fault diagnosis model established by learning in advance the characteristic data of the horizontal vibration signals and vertical vibration signals of the gearbox shaft for different gearbox faults. The characteristic data extraction of the horizontal vibration signals and vertical vibration signals of the gearbox shaft includes: performing multi-scale time domain feature extraction and deep noise reduction feature extraction:

[0054] The multi-scale time domain feature extraction step is:

[0055] The first step is to decompose the vibration signal into stationary signal IMF components according to the set time length, and obtain the energy of each IMF component through formula (1).

[0056]

[0057] Where: N means the signal is divided into N equal parts, each part has a time length of △t, c i (tn) represents the amplitude of the i-th component at time tn, and h is the total number of IMF components;

[0058] The second step is to select the first S IMF components sorted from large to small in energy, and calculate the kurtosis index of the S IMF components by formula (2). The value of S is less than 10, usually 6 to 8.

[0059]

[0060] The third step is to take the first P IMF components sorted from large to small in terms of kurtosis value. As a result of multi-scale time domain feature extraction, where Nc is the length of the signal, Indicates that the IMF component is 1xNc of real dimension, P is less than S, usually P is less than 2 to 3 values ​​of S, for example, if S is 8, P is 6 or 5;

[0061] The deep noise reduction feature extraction is: sending the vibration signal to the deep noise reduction autoencoder, and outputting the deep noise reduction feature through the deep noise reduction autoencoder

[0062] The horizontal vibration signal includes an axial horizontal vibration signal and a radial horizontal vibration signal.

[0063] Wherein: the steps of establishing the fault diagnosis model are:

[0064] The first step is to perform the multi-scale time domain feature extraction and deep noise reduction feature extraction according to different faults;

[0065] The second step is to fuse the multi-scale time domain features and the deep noise reduction features: each IMF component in the multi-scale time domain features is concatenated with the deep noise reduction features, and these IMF components are stacked into a multi-channel training sample in time sequence, and a deep convolutional neural network is constructed and trained. The training of the convolutional neural network is completed through the input of the training samples.

[0066] Each convolutional layer outputs the feature map of formula (3) to complete feature fusion:

[0067]

[0068] Where * represents convolution, M i , i=1,2,...,P,P+1 represents the convolution kernel of P+1 channels;

[0069] The third step is to output the fault diagnosis result based on the probability of each fault type occurring formed by feature fusion.

[0070] Wherein: the feature data extraction is the feature data extraction of multiple sample signals of the vibration signal, and the multiple sample signals are obtained by dividing the length of the vibration signal into windows of equal length.

[0071] The method further includes marking the device name, sampling point number, sensitivity, and acquisition time parameters during the acquisition process of each group of multiple sample signals, and checking whether the corresponding fields of each group of multiple sample signals have missing values ​​and whether the fields are complete.

[0072] Wherein: the method of decomposing the vibration signal into the IMF components of the stationary signal according to the set time length is:

[0073] The first step is to define the total number of EMDs to be performed MAX_ITER;

[0074] In the second step, normally distributed white noise is added to the original signal, and the signal to be processed is decomposed by empirical mode decomposition (EMD) to obtain the IMF component C ij , C ij represents the i IMF components decomposed for the jth time, and the above operation is repeated until the number of empirical mode decompositions reaches MAX_ITER;

[0075] The third step is to calculate the IMF component corresponding to each time point and find its average value according to formula (4).

[0076]

[0077] Finally, they are spliced ​​in chronological order. As the final obtained stationary signal IMF component.

[0078] The following is combined Figures 2 to 5 The above method is further described in detail.

[0079] Figure 2 In the data collection and preprocessing module, data collection and preprocessing are responsible for data collection and data preprocessing, which are specifically implemented as follows:

[0080] Divide the data into windows of equal length and perform data enhancement processing by overlapping slices to obtain as many sample signals as possible. That is, set the sliding window step size to l window Less than the length of a single sample signal l input , the length is l signal The original signal can be divided into Nums = (l signal -l input ) / l window Sample signal;

[0081] Mark the equipment failure type, equipment name, number of sampling points, sensitivity, acquisition time and other related parameters during the acquisition process of each set of sample signals, check whether the corresponding fields of each set of sample signals have missing values, whether the fields are complete, and complete data preprocessing.

[0082] The feature data extraction module includes a multi-scale time domain feature extraction module and a deep noise reduction feature extraction module, where:

[0083] The multi-scale time domain feature extraction module uses the method of ensemble empirical mode decomposition to decompose the collected vibration signal into a series of stable IMF components with different characteristic scales and obvious fault characteristics. Figure 3 Indicates the specific process of this module, which is implemented as follows:

[0084] 1. Define the total number of EMD to be performed MAX_ITER;

[0085] 2. Add normally distributed white noise to the original signal, decompose the signal to be processed by empirical mode decomposition (EMD), and obtain the IMF component C ij , represents the i IMF components decomposed for the jth time, and the above operation is repeated until the number of empirical mode decompositions reaches MAX_ITER;

[0086] 3. Calculate the IMF components corresponding to each stage and find their average values As the final obtained stationary IMF component;

[0087] 4. Solve the energy of each IMF component In the formula, N means that the signal is divided into N equal parts, and the length of each part is △t, c i (tn) represents the amplitude of the i-th component at time tn, h is the total number of IMF components, the first 8 IMF components with larger energy are selected, and then the kurtosis index of these 8 IMF components is calculated. Take the five IMF components with the largest kurtosis values, As a result of multi-scale time domain feature extraction, where Nc is the length of the signal, Indicates that the IMF component is 1xNc of real dimension; the output result is as follows Figure 4 shown.

[0088] The deep noise reduction feature extraction module trains the deep noise reduction autoencoder by adding noise. The obtained model can effectively perform noise reduction filtering and feature extraction on the original vibration signal. The specific implementation is as follows:

[0089] 1. Add random noise to train the deep denoising autoencoder, that is, take the vibration signal with added noise as input and the vibration signal without noise as output, and train multiple stacked autoencoders by changing the reconstruction error term of the loss function, so that the autoencoder learns to remove this noise to obtain the real input that is not contaminated by noise. The input of one autoencoder is the output of the next autoencoder. Stacking L autoencoders constitutes L hidden layers. The output of each layer of the autoencoder network is in represents the hidden layer activation function, W l a , represents the parameters learned by the autoencoder during training, X l-1 Represents the output of the previous layer of autoencoder;

[0090] 2. After the training is completed, the original network parameters are retained, the vibration signal to be processed is input, and the deep noise reduction features can be output through the deep noise reduction autoencoder.

[0091] Figure 2 The multi-feature fusion convolutional neural network module in the CNN combines the IMF components and deep noise reduction features extracted by the first two modules, inputs them into the deep convolutional neural network for training, and finally outputs the probability value of each fault to complete the fault diagnosis task. The specific implementation is as follows:

[0092] 1. The characteristics of the above two modules, that is, each IMF component and deep noise reduction features Splicing is performed to stack these components into a multi-channel sample in the order of time series, and the data dimension increases by one dimension, that is, the data dimension changes from 1×Nc to 1×Nc×6;

[0093] 2. Construct and train a deep convolutional neural network, where the convolution kernel width of the first layer is 1 and the number of channels is 6. Extract the convolution information of the corresponding moments of the 5 IMF components and 1 deep noise reduction feature, and output the feature map:

[0094]

[0095] Where * represents convolution, M i (i=1,2,...,6) represents the convolution kernel of these 6 channels. The process of the first layer of convolution is as follows Figure 5 As shown in the figure, the first layer has a set of convolution kernel dimensions of 1×q×6, and the resulting first layer convolution graph has a dimension of 1×(Nc-q+1)×1. Subsequently, multiple convolution layers, batch normalization layers, activation layers, and pooling layers are established, and finally, the Dropout layer, fully connected layer, and classifier are established to complete the construction of the neural network. Through the input of training samples, the convolutional neural network is trained in a supervised manner;

[0096] 3. Apply deep convolutional neural network, input combined features, and output diagnosis results, that is, the probability of each type of fault occurring, to complete the fault diagnosis task.

[0097] The following is an example of gearbox fault diagnosis:

[0098] There are four types of gear fault conditions collected, including slight gear peeling, severe gear peeling, local wear, and comprehensive wear. Through the data collection and preprocessing module, the fault samples were expanded from the original 120 to 600, enriching the fault samples and laying a good foundation for training deep networks. Each sample contains 1024 points of vibration data.

[0099] Through the multi-scale time domain feature extraction module, the original signal is decomposed into multiple IMF components, and the energy value and kurtosis value of the IMF components are calculated. The calculation results for the sample signal of severe gear spalling fault are shown in Table 1. According to the energy value method and the kurtosis value method, IMF components 1, 2, 3, 4, and 6 are selected as the multi-scale time domain features to characterize the sample signal. The waveforms of the five components are shown in Figure 4 shown.

[0100] Table 1 Calculation results of energy value and kurtosis value

[0101]

[0102] In the deep denoising feature extraction module, a deep denoising autoencoder is used for training, where the noise parameter is set to 0.5, the number of neurons in the hidden layer are 512, 256, 128, 128, 256, 512 respectively, the Dropout is 0.3, the activation function is ReLu, and the gradient descent method is used for training. The trained deep denoising autoencoder can effectively filter out noise and extract features from the original signal.

[0103] In the multi-feature fusion convolutional neural network module, the parameters of each layer of the neural network are designed as shown in Table 2. In addition, the Dropout parameter is 0.3, and ReLu is selected as the activation function. By building the network model shown in Table 2, a deep convolutional neural network is established, which improves the expression ability of the network and diagnoses the probability of fault occurrence.

[0104] Table 2 Convolutional neural network parameters

[0105]

[0106] Through the construction of the above modules, the test fault diagnosis accuracy is 95.32%. Compared with the diagnosis accuracy of 89.43% of the model established by the more traditional method SVM support vector machine, the fault diagnosis rate of the model established by this method has been greatly improved, indicating that by combining multi-scale time domain features and deep noise reduction features, the network shows high robustness and achieves good classification effect.

[0107] In summary, the present invention solves the problem of single fault features and low diagnostic efficiency to a certain extent. By combining multi-scale time domain features with deep noise reduction features, a deep convolutional neural network is used to fuse the above two features to improve the fault diagnosis accuracy.

[0108] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A gearbox fault diagnosis method for wind turbine gearbox fault diagnosis, which obtains horizontal vibration signals and vertical vibration signals of the gearbox shaft by an acceleration sensor installed in the wind turbine, extracts characteristic data of the horizontal vibration signals and vertical vibration signals of the gearbox shaft, and inputs the extracted characteristic data into a deep convolutional neural network module to obtain a fault diagnosis result. The deep convolutional neural network module is a fault diagnosis model established by pre-learning characteristic data of horizontal and vertical vibration signals of the gearbox shaft for different gearbox faults, and is characterized in that: The feature data extraction of the horizontal vibration signal and the vertical vibration signal of the gearbox shaft includes: performing multi-scale time domain feature extraction and deep noise reduction feature extraction: The multi-scale time domain feature extraction step is: The first step is to decompose the vibration signal into stationary signal IMF components according to the set time length, and obtain the energy of each IMF component through formula (1). Where: N means the signal is divided into N equal parts, each part has a time length of Δt, c i (tn) represents the amplitude of the i-th component at time tn, and h is the total number of IMF components; The second step is to select the first S IMF components sorted from large to small in energy, and calculate the kurtosis index of the S IMF components using formula (2): The third step is to take the first P IMF components sorted from large to small in terms of kurtosis value. As a result of multi-scale time domain feature extraction, where Nc is the length of the signal, Indicates that the IMF component is 1xNc of real dimension, P is less than S; The deep noise reduction feature extraction is: sending the vibration signal to the deep noise reduction autoencoder, and outputting the deep noise reduction feature through the deep noise reduction autoencoder.

2. The method according to claim 1, characterized in that: The horizontal vibration signal includes an axial horizontal vibration signal and a radial horizontal vibration signal.

3. The method according to claim 1 or 2, characterized in that: The steps of establishing the fault diagnosis model are: The first step is to perform the multi-scale time domain feature extraction and deep noise reduction feature extraction according to different faults; The second step is to fuse the multi-scale time domain features and the deep noise reduction features: each IMF component in the multi-scale time domain features is concatenated with the deep noise reduction features, and these IMF components are stacked into a multi-channel training sample in time sequence, and a deep convolutional neural network is constructed and trained. The training of the convolutional neural network is completed through the input of the training samples. Each convolutional layer outputs the feature map of formula (3) to complete feature fusion: Where * represents convolution, M i , i=1,2,...,P,P+1 represents the convolution kernel of P+1 channels; The third step is to output the fault diagnosis result based on the probability of occurrence of each fault type formed by feature fusion.

4. The method according to claim 1 or 2, characterized in that: The feature data extraction is feature data extraction of multiple sample signals of the vibration signal, and the multiple sample signals are obtained by dividing the length of the vibration signal into windows of equal length.

5. The method according to claim 4, characterized in that The method further includes marking the device name, sampling point number, sensitivity, and acquisition time parameters during the acquisition process of each group of multiple sample signals, and checking whether the corresponding fields of each group of multiple sample signals have missing values ​​and whether the fields are complete.

6. The method according to claim 1 or 2, characterized in that: The method of decomposing the vibration signal into the IMF components of the stationary signal according to the set time length is: The first step is to define the total number of EMDs to be performed MAX_ITER; In the second step, normally distributed white noise is added to the original signal, and the signal to be processed is decomposed by empirical mode decomposition (EMD) to obtain the IMF component C ij , C ij represents the i IMF components decomposed for the jth time, and the above operation is repeated until the number of empirical mode decompositions reaches MAX_ITER; The third step is to calculate the IMF component corresponding to each time point and find its average value according to formula (4). Finally, they are spliced ​​in chronological order. As the final obtained stationary signal IMF component.

7. A gearbox fault diagnosis system for implementing the method described in any one of claims 1 to 6, comprising: A gear transmission box of a wind turbine generator, wherein the input shaft of the gear transmission box is connected to the hub of a wind-driven rotating blade, and the output shaft of the gear transmission box is connected to a generator, wherein acceleration sensors are respectively arranged on a bearing seat where the input shaft is docked with the hub of a wind-driven rotating blade, the gear transmission box, and a generator housing to sense horizontal vibration signals and vertical vibration signals of the gear shaft, the acceleration sensor is connected to a fault diagnosis and analysis computer via a wired or wireless connection, a signal receiving interface card is arranged in the computer, the signal receiving interface card is connected to the acceleration sensor to send the received horizontal vibration signals and vertical vibration signals of the gear to a central processing unit, and a vibration signal preprocessing module, a feature data extraction module, and a deep convolutional neural network module are arranged in the central processing unit; The vibration signal preprocessing module divides the length of the acquired horizontal vibration signal and vertical vibration signal of the gear into multiple sample signals using windows of equal length, records the device name, number of sampling points, sensitivity, and acquisition time parameters of the corresponding vibration signal, and checks whether the corresponding fields of each group of multiple sample signals have vacant values ​​and whether the fields are complete; The feature data extraction module extracts features from the sample signal provided by the preprocessing module; The deep convolutional neural network module performs fault diagnosis on the feature data extracted by the feature data extraction module and outputs the diagnosis result; The acceleration sensor comprises: an input shaft axial horizontal vibration sensor arranged on the axial end face of the bearing seat where the input shaft is connected to the hub of the wind turbine rotating blade, an input shaft radial horizontal vibration sensor arranged on the horizontal radial end face of the bearing seat, a housing vertical vibration sensor and a housing radial horizontal vibration sensor arranged on the gear transmission housing, and a generator shaft radial horizontal vibration sensor respectively arranged on the input and output shafts of the generator; The box vertical vibration sensor is arranged in the middle of the upper side of the box, and there are two box radial horizontal vibration sensors, which are arranged on both sides of the box vertical vibration sensor on the side of the box; The acceleration sensor is fixed at a set position by means of embedding.

Images