A bearing fault diagnosis method, system and storage medium based on super wavelet

Abstract

The application discloses a bearing fault diagnosis method and system based on super wavelet and a storage medium. The method comprises the following steps: acquiring a bearing vibration signal; inputting the bearing vibration signal into a preset super wavelet convolutional neural network model, analyzing the bearing vibration signal through a super wavelet block, and obtaining bearing fault features of multiple scales; performing weighted fusion on the bearing fault features of the multiple scales to obtain fused fault features; and analyzing the fused fault features to obtain bearing fault diagnosis information. The bearing vibration signal of a fan gear box is processed through the SWCNN, the bearing vibration signal containing noise is processed through the super wavelet block, three different wavelet base functions are used for noise reduction processing and feature extraction, and the anti-noise capability of the SWCNN network is enhanced. In addition, the weight fusion layer of the super wavelet block can adaptively determine the fusion weight, so as to enhance the wavelet channel with obvious fault features, highlight the key features, and improve the diagnosis accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing and data transmission, and more particularly, to a fan gearbox bearing fault diagnosis method, system and storage medium based on hyperwave. BACKGROUND

[0002] Rolling bearings are widely used in fan gearboxes and are one of the key components that affect the normal operation of fan gearboxes. However, rolling bearings are in a state of high-speed rotation and heavy load for a long time, which causes 45% to 55% of rotating equipment failures to be caused by bearing damage. Rolling bearing failures will cause the equipment to be out of service for a long time, which will increase the production cost and even cause safety accidents. Monitoring and diagnosing the operating state of the rolling bearings of the fan gearbox can timely find problems and eliminate faults, thereby reducing safety accidents, which has important social significance and economic value. Therefore, it is crucial to research and develop an effective fan gearbox bearing fault diagnosis method.

[0003] With the development of artificial intelligence technology, bearing fault diagnosis based on deep learning has attracted widespread attention from experts and scholars. Among them, CNN is a typical deep learning method that has been successful in various image classification, object detection, natural language processing, speech recognition, etc. However, due to the harsh working environment of rolling bearings and the large background noise, the bearing vibration signal is disturbed by noise, which makes it difficult for the CNN model to extract effective fault features, resulting in a decrease in the fault diagnosis accuracy of the CNN model.

[0004] Therefore, the prior art has defects and needs to be improved. SUMMARY

[0005] In view of the above problems, the purpose of the present application is to provide a fan gearbox bearing fault diagnosis method, system and storage medium based on hyperwave, which can diagnose bearing faults in strong noise conditions.

[0006] The first aspect of the present application provides a bearing fault diagnosis method based on hyperwave, comprising:

[0007] Obtaining a fan gearbox bearing vibration signal;

[0008] Inputting the bearing vibration signal into a preset hyperwave convolutional neural network model, analyzing the bearing vibration signal through a hyperwave block to obtain bearing fault features of multiple scales;

[0009] Weighted fusion of the bearing fault features of multiple scales to obtain fused fault features;

[0010] Analyzing the fused fault features to obtain bearing fault diagnosis information.

[0011] The scheme further includes:

[0012] An initial overwave convolutional neural network model is established;

[0013] Obtain vibration signal sample data;

[0014] The vibration signal sample data is divided into a training set, a validation set, and a test set according to a preset ratio;

[0015] The training set, the validation set, and the test set are respectively input into the initial overwave convolutional neural network model for training diagnosis, and a preset overwave convolutional neural network model is obtained.

[0016] In the scheme, after the bearing vibration signal is input into the preset overwave convolutional neural network model, the scheme further includes:

[0017] Obtain a plurality of wavelet library data;

[0018] The wavelet library data is analyzed to obtain a plurality of adaptive wavelets;

[0019] The calculation method of the adaptive wavelet is represented by a formula as follows:

[0020]

[0021] wherein, is the nth wavelet basis function, a n is the scale factor of the wavelet basis function, b n is the translation factor of the wavelet basis function, w n is the combination coefficient of the overwave, and k represents the kth wavelet basis function.

[0022] In the scheme, the bearing vibration signal is analyzed by the overwave block to obtain a plurality of scale bearing fault features, including:

[0023] The overwave block performs wavelet transform on the bearing vibration signal through the plurality of adaptive wavelets, so that the overwave block decomposes and denoises the bearing vibration signal through wavelet transform to obtain a denoising vibration signal;

[0024] The denoising vibration signal is feature extracted through the plurality of adaptive wavelets to obtain a plurality of scale bearing fault features;

[0025] The wavelet transform is represented by a formula as follows:

[0026]

[0027] wherein, ψ represents a mother wavelet, t is time, a is a scale factor inversely proportional to frequency, and b is a translation factor.

[0028] In the scheme, the bearing fault features of the multiple scales are weighted and fused to obtain a fused fault feature, which comprises:

[0029] The bearing fault features of the multiple scales are multiplied by corresponding influence weights respectively to obtain weight scores of the multiple fault features;

[0030] A plurality of key fault features are selected according to the weight scores of the multiple fault features;

[0031] The multiple key fault features are fused to obtain a fused fault feature.

[0032] In the scheme, the bearing fault diagnosis information is obtained by analyzing the fused fault feature, which comprises:

[0033] The fused fault feature is input into a plurality of convolution layers;

[0034] The plurality of convolution layers process the fused fault feature according to a preset method, and send the processed fault feature to a fully connected layer;

[0035] The fully connected layer integrates and classifies the processed fault feature to obtain bearing fault diagnosis information.

[0036] The second aspect of the application provides a fan gearbox bearing fault diagnosis system based on hyperwave, comprising a memory and a processor, the memory comprises a fan gearbox bearing fault diagnosis method program based on hyperwave, and the fan gearbox bearing fault diagnosis method program based on hyperwave realizes the following steps when the processor is executed:

[0037] Obtain the vibration signal of the fan gearbox bearing;

[0038] The bearing vibration signal is input into a preset hyperwave convolutional neural network model, and the bearing vibration signal is analyzed by a hyperwave block to obtain bearing fault features of multiple scales;

[0039] The bearing fault features of the multiple scales are weighted and fused to obtain a fused fault feature;

[0040] The bearing fault diagnosis information is obtained by analyzing the fused fault feature.

[0041] In the scheme, it also comprises:

[0042] An initial hyperwave convolutional neural network model is established;

[0043] Obtain the vibration signal sample data of the fan gearbox;

[0044] The vibration signal sample data is divided into a training set, a validation set and a test set according to a preset ratio;

[0045] The training set, the validation set and the test set are respectively input into the initial ultra wavelet convolutional neural network model for training diagnosis, and a preset ultra wavelet convolutional neural network model is obtained.

[0046] In the scheme, after the bearing vibration signal is input into the preset ultra wavelet convolutional neural network model, the following steps are further included:

[0047] Obtaining a plurality of wavelet library data;

[0048] Analyzing the wavelet library data to obtain a plurality of adaptive wavelets;

[0049] The calculation method of the adaptive wavelet is expressed by a formula as follows:

[0050]

[0051] wherein, is the nth wavelet base function, a n is the scale factor of the wavelet base function, b n is the translation factor of the wavelet base function, w n is the combination coefficient of the ultra wavelet, and k represents the kth wavelet base function.

[0052] In the scheme, the bearing vibration signal is analyzed by the ultra wavelet block to obtain a plurality of scale bearing fault features, including:

[0053] The ultra wavelet block performs wavelet transform on the bearing vibration signal through the plurality of adaptive wavelets, so that the ultra wavelet block decomposes and denoises the bearing vibration signal through wavelet transform to obtain a denoising vibration signal;

[0054] The denoising vibration signal is feature extracted through the plurality of adaptive wavelets to obtain a plurality of scale bearing fault features;

[0055] The wavelet transform is expressed by a formula as follows:

[0056]

[0057] wherein, ψ represents a mother wavelet, t is time, a is a scale factor inversely proportional to frequency, and b is a translation factor.

[0058] In the scheme, the plurality of scale bearing fault features are weighted and fused to obtain a fused fault feature, including:

[0059] The plurality of scale bearing fault features are respectively multiplied by corresponding influence weights to obtain weight scores of the plurality of fault features.

[0060] select a plurality of key fault features according to the weight scores of the plurality of fault features;

[0061] perform feature fusion on the plurality of key fault features to obtain a fused fault feature.

[0062] In this scheme, the analysis according to the fused fault feature to obtain bearing fault diagnosis information comprises:

[0063] input the fused fault feature to a plurality of convolutional layers;

[0064] the plurality of convolutional layers process the fused fault feature according to a preset method, and send the processed fault feature to a fully connected layer;

[0065] the fully connected layer integrates and classifies the processed fault feature to obtain bearing fault diagnosis information.

[0066] The third aspect of the present application provides a computer readable storage medium, the computer readable storage medium comprises a kind of based on super wave fan gear box bearing fault diagnosis method program of fan gear box bearing fault diagnosis method program based on super wave of the processor is executed when, realize the steps of any one of the kind based on super wave fan gear box bearing fault diagnosis method as described above.

[0067] The present application discloses a kind of based on super wave fan gear box bearing fault diagnosis method, system and storage medium, the method comprises: obtaining bearing vibration signal;The bearing vibration signal is input to preset super wave convolutional neural network model, the bearing vibration signal is analyzed by super wave block, obtains a plurality of scale bearing fault features;The bearing fault features of the plurality of scales are weighted and fused, to obtain fused fault feature;According to the analysis of the fused fault feature, obtain bearing fault diagnosis information.The present application is processed by SWCNN to bearing vibration signal, the bearing vibration signal containing noise is handled by super wave block, is respectively reduced noise processing and feature extraction using 3 different wavelet base functions, enhances the anti-noise ability of SWCNN network.In addition, the weight fusion layer of super wave block can adaptively determine fusion weight, to enhance the wavelet channel of obvious fault feature, to highlight key features, improve the diagnosis accuracy of SWCNN. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 The flow chart of the present application based on super wave fan gear box bearing fault diagnosis method is shown;

[0069] Figure 2 The flow chart of the present application is shown in a preset super wave convolutional neural network model training method.

[0070] Figure 3 A flow chart of a bearing fault diagnosis information acquisition method of the application is shown;

[0071] Figure 4 A block diagram of a bearing fault diagnosis system based on super wavelet of the application is shown;

[0072] Figure 5 A display diagram of a bearing data acquisition system of the application is shown;

[0073] Figure 6 A training flow chart of a super wavelet convolutional neural network of the application is shown. DETAILED DESCRIPTION

[0074] In order to enable the above-mentioned purposes, features and advantages of the application to be more clearly understood, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the embodiments of the application and the features in the embodiments can be combined with each other without conflict.

[0075] In the following description, many specific details are set forth in order to provide a thorough understanding of the application, but the application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the application is not limited by the specific embodiments disclosed below.

[0076] Figure 1 A flow chart of a fan gear box bearing fault diagnosis method based on super wavelet of the application is shown.

[0077] As Figure 1 shown, the application discloses a fan gear box bearing fault diagnosis method based on super wavelet, comprising:

[0078] S102, acquiring a fan gear box bearing vibration signal;

[0079] S104, inputting the bearing vibration signal to a preset super wavelet convolutional neural network model, analyzing the bearing vibration signal through a super wavelet block, and obtaining bearing fault features of multiple scales;

[0080] S106, weighting and fusing the bearing fault features of multiple scales to obtain fused fault features;

[0081] S108, analyzing according to the fused fault features to obtain bearing fault diagnosis information.

[0082] According to the embodiment of the present application, the scheme establishes a preset super wavelet convolutional neural network (SWCNN) according to the wavelet transform principle. The SWCNN takes the super wavelet convolutional block as the first layer of the CNN network, denoises the bearing vibration signal and enhances the key fault features. The scheme selects the Morlet wavelet, the Laplace wavelet and the Mexhat wavelet to constitute the wavelet kernel of the super wavelet block. The super wavelet block adaptively selects and constructs the wavelet basis that can effectively represent the bearing fault signal from the original bearing vibration signal, and simultaneously adopts multiple wavelet basis functions for wavelet transform, highlights the bearing fault features on different scales, and provides the CNN model with multi-scale fault features. In addition, the super wavelet block can decompose the key feature information from the noise through the multi-resolution analysis characteristic of the wavelet transform, thereby improving the noise immunity of the CNN model.

[0083] The SWCNN adopts the adaptive wavelet transform (AWT) to adopt the Laplace wavelet, the Morlet wavelet and the Mexh wavelet to constitute the adaptive wavelet, which can fuse the advantages of different wavelet bases. Moreover, the SWCNN does not need to manually select the wavelet basis, and it optimizes the adaptive wavelet through the CNN for the wavelet parameters and the combination coefficients, thereby learning a set of adaptive wavelet filters with good noise reduction performance. The scheme analyzes the SWCNN through two different bearing data sets. As can be known from the results, the SWCNN can not only well preserve the original information, but also fully consider the internal relationship of the sequence data, and realize the fault diagnosis in the strong noise condition.

[0084] Figure 2 A flowchart of a preset super wavelet convolutional neural network model training method is shown.

[0085] As Figure 2 shown, according to the embodiment of the present application, further comprising:

[0086] S202, an initial super wavelet convolutional neural network model is established;

[0087] S204, vibration signal sample data is acquired;

[0088] S206, the vibration signal sample data is divided into a training set, a verification set and a test set according to a preset proportion;

[0089] S208, the training set, the verification set and the test set are respectively input to the initial super wavelet convolutional neural network model for training diagnosis, and a preset super wavelet convolutional neural network model is obtained.

[0090] It should be noted that the bearing data set of the experiment of the scheme is from a fan gear box simulation bearing fault simulation comprehensive test bench. The experimental device is as Figure 5As shown, the experimental platform mainly consists of a driving motor, a transmission part, a support part, a controller, etc. In the rolling bearing fault experiment, signals in four states are collected, and normal bearings, inner ring fault bearings, outer ring fault bearings and rolling element fault bearings are obtained. Among them, the sampling frequency is 8192Hz, the collection time is 120s, and the sampling signal sequence length of each fault working condition is 983040.

[0091] In order to simulate the bearing crack, pitting and other faults in the actual working condition, the inner ring, outer ring and rolling element of the rolling bearing are implanted with different depth faults by means of wire cutting processing groove. The depth is 0.4mm, 0.8mm and 1.2mm respectively. The healthy bearing can be regarded as a special fault mode, so there are ten kinds of bearing fault types in the experimental data set.

[0092] The training process of the SWCNN model is as shown in Figure 6 , which includes three steps: data set division, model training and model diagnosis. The specific steps are as follows:

[0093] (1) The original data, that is, the vibration signal sample data, is divided into three data sets: training set, validation set and test set according to the preset ratio of 8:1:1. The data of the training set is the original noise-free data, so that the SWCNN can learn the appropriate wavelet basis. The validation set and the test set are added with noise of a certain signal-to-noise ratio to judge whether the trained SWCNN model can resist noise interference and diagnose bearing faults.

[0094] (2) When training the model, first, initialize the network parameters of SWCNN. Then input the training set to train the parameters of each layer of SWCNN, so that SWCNN can fully learn the characteristics of the original data. Finally, use the validation set containing different noises to test the trained SWCNN model to judge whether the accuracy meets the requirements. Repeat the above steps several times, and select the optimal model as the diagnosis model.

[0095] (3) When diagnosing the model, first, take the optimal model in step (2) as the diagnosis model. Then, input the test set into the diagnosis model. Finally, the result of the diagnosis model is taken as the judgment standard for evaluating the performance of SWCNN.

[0096] According to the embodiment of the application, after the bearing vibration signal is input into the preset superwave convolutional neural network model, the method further comprises:

[0097] obtaining wavelet library data;

[0098] analyzing the wavelet library data to obtain a plurality of adaptive wavelets;

[0099] The calculation method of the adaptive wavelet is represented by a formula as follows:

[0100]

[0101] wherein, is the nth wavelet basis function, a n is the scale factor of the wavelet basis function, b n is the translation factor of the wavelet basis function, w n is the combination coefficient of the super wavelet, k represents the kth wavelet basis function.

[0102] It should be noted that the wavelet transform can decompose the bearing vibration signal to reduce the interference of noise on the bearing vibration signal, and more information can be obtained by using the multi-resolution analysis characteristic. In order to use the multi-resolution analysis characteristic of the wavelet transform, the super wavelet block (SWB) is proposed in the scheme. The SWB is composed of two convolution layers. The first convolution layer is a wavelet transform layer, which is composed of multiple different wavelet channels in parallel, and each wavelet channel represents a wavelet basis. The second convolution layer is a weighted fusion layer, which is a standard 1x1 convolution layer, used for weighted fusion of the features output by different wavelet channels.

[0103] The principle of SWB is to extract fault features from the bearing vibration signal by using adaptive wavelets. The adaptive wavelets use a multi-wavelet basis complementary mechanism to obtain complementary features using different wavelet bases, and assign different weights to the features to fully extract bearing fault features. In addition, the adaptive wavelets satisfy the admissibility condition of wavelets, that is, they are considered as a wavelet. Therefore, SWB can fuse the advantages of different wavelet bases, obtain more fault feature information, and learn new wavelet bases at the same time.

[0104] SWB has two main advantages. The first advantage is that SWB extracts features through multiple wavelet basis functions to match different fault features in the vibration signal, obtain more useful information, and thus increase the noise resistance of the CNN model. The second advantage is that SWB can directly input the original vibration signal, which well preserves the original information and reduces the loss of original vibration signal information caused by denoising preprocessing.

[0105] According to an embodiment of the present application, the bearing vibration signal is analyzed by the super wavelet block to obtain bearing fault features at multiple scales, comprising:

[0106] The super wavelet block performs wavelet transform on the bearing vibration signal through the multiple adaptive wavelets, so that the super wavelet block decomposes and denoises the bearing vibration signal by wavelet transform to obtain a denoised vibration signal;

[0107] The denoised vibration signal is feature-extracted by the multiple adaptive wavelets to obtain bearing fault features at multiple scales;

[0108] The wavelet transform is expressed by the formula:

[0109]

[0110] Where ψ represents the mother wavelet, t is time, a is a scale factor inversely proportional to frequency, and b is a translation factor.

[0111] It should be noted that the wavelet transform layer can extract bearing fault features from different angles using multiple wavelet bases. When the SWB is working, the bearing vibration signal is first input to the wavelet transform layer. The input signal is decomposed by wavelet transform in the wavelet channel to reduce noise interference and extract bearing fault features.

[0112] The wavelet transform provides a time-frequency window that changes with the signal frequency, overcoming the shortcoming of the short-time Fourier transform window size not changing with the frequency. It can be used for time-frequency analysis and processing of signals. Therefore, the wavelet transform has the characteristic of multi-resolution analysis, and can simultaneously perform localized analysis of bearing vibration signals in time domain and frequency domain. The signal can be decomposed into subbands with multiple scales and frequencies. The signal has high time resolution and low frequency resolution at high frequencies, and has high frequency resolution and low time resolution at low frequencies. The wavelet transform can represent time and frequency information simultaneously. The multi-resolution characteristic of the wavelet transform makes it excellent in signal decomposition, noise reduction, and analysis of non-stationary signals.

[0113] A wavelet is a function that has localization in both time and frequency domains. Each wavelet is a variation of a single mother wavelet ψ, called a wavelet member. A wavelet library {ψ a,b (t)} is composed of multiple wavelet members, which can be constructed by controlling the scale factor a and the translation factor b of the mother wavelet ψ(·). The scale factor a can make the signal contract or stretch. When the scale factor is low, the signal contracts. When the scale factor is high, the signal stretches. The wavelet transform includes three steps:

[0114] (1) Select a wavelet basis function, then calculate the similarity coefficient with the partial signal.

[0115] (2) Shift the wavelet to the right by b, continue to calculate the similarity coefficient, until all parts of the signal are calculated.

[0116] (3) Change the scale parameter a of the wavelet, repeat steps (1) and (2) until all scale analyses are completed.

[0117] According to an embodiment of the present application, the weighted fusion of the bearing fault features of the multiple scales obtains a fused fault feature, including:

[0118] The bearing fault features of the plurality of scales are respectively multiplied by corresponding influence weights to obtain weight scores of the plurality of fault features;

[0119] A plurality of key fault features are selected according to the weight scores of the plurality of fault features;

[0120] The plurality of key fault features are fused to obtain a fused fault feature.

[0121] It should be noted that the weighted fusion layer is a 1x1 standard convolutional layer, which is used for weighted fusion of the features output by different wavelet channels. Because the wavelet basis functions in the wavelet channels are different, the bearing fault features output from different wavelet channels differ in importance. Subsequently, the weighted fusion layer assigns different weights to the output bearing fault features through a set of 1x1 standard convolutional layers, highlights key fault features, determines the best bearing fault features, and fuses the features. Finally, the SWB outputs a set of weighted and fused fault features, and takes the set of features as the input of the CNN network, thereby enhancing the noise resistance of the CNN network.

[0122] Figure 3 A flowchart of a bearing fault diagnosis information acquisition method according to an embodiment of the present application is shown.

[0123] As shown in Figure 3 According to an embodiment of the present application, the analysis according to the fused fault feature to obtain bearing fault diagnosis information comprises:

[0124] S302, inputting the fused fault feature to a plurality of convolutional layers;

[0125] S304, the plurality of convolutional layers processing the fused fault feature according to a preset method, and sending the processed fault feature to a fully connected layer;

[0126] S306, the fully connected layer integrating and classifying the processed fault feature to obtain bearing fault diagnosis information.

[0127] It should be noted that the SWCNN includes a super wavelet block, four standard convolutional layers, some operations such as batch normalization, ReLu and global maximum pooling, and a fully connected layer. The preset method is specifically that the output of the super wavelet block will pass through a plurality of convolutional layers to enhance feature expression. Then, some operations such as batch normalization, ReLu and global maximum pooling are sequentially performed on the output of each convolutional layer. Finally, a classification result is output through the fully connected layer.

[0128] In addition, the second layer convolution of the SWCNN adopts a wide convolution kernel to increase the receptive field and improve the diagnostic accuracy. The subsequent convolution layers all adopt a small convolution kernel to deeply mine the hidden deep information between samples and have a small amount of network parameters. In addition, a batch normalization and a pooling layer are added after all the convolution layers to speed up the training and reduce the amount of parameters.

[0129] According to the embodiment of the application, the method further comprises:

[0130] The vibration signal sample data is expanded by the method of overlapping sampling.

[0131] It should be noted that, due to the multifunctional bearing fault simulation test bench, each fault data set has only 983040 sampling points, and the training samples are limited. When the training samples are too few, the model will be under-fitted, and the accuracy of the training set and the test set will be reduced. In order to avoid the problem of under-fitting, the method of overlapping sampling is used to expand the data set. The size of the sliding window is 2048, and the sliding step (offset) is 100. Starting from the starting position, the sliding window collects 2048 sampling points to successively build data samples by sliding 100 sampling points backward each time, until there are 600 samples.

[0132] According to the embodiment of the application, the method further comprises:

[0133] In the preset hyperwave convolutional neural network model training process, the training data is processed through a cross-entropy function and an Adam algorithm.

[0134] It should be noted that, in order to improve the performance of the model and make the experimental results optimal, the model parameters are selected in the present scheme. In the training strategy of the present scheme, the cross-entropy function and Adam are selected as the loss function and the optimizer, so as to speed up the convergence speed of the model and obtain the optimal result. The cross-entropy is a loss function used to evaluate the difference between the predicted probability distribution and the actual distribution. The smaller the cross-entropy, the closer the two probability distributions, and the higher the accuracy of the model. Adam is a learning rate adaptive optimization algorithm, which can improve the robustness of the learning rate by dynamically adjusting the learning rate parameter, and help the model converge. The learning rate needs to be set in an appropriate range when training the CNN model. If the learning rate is too high, it may cause instability, and if the learning rate is too small, it will increase the training time. Under the premise of ensuring the stability of the training, the training time is reduced, and the learning rate is set to 0.001 in the present scheme. At the same time, the training batch size is set to 128 in the present scheme.

[0135] Figure 4 A block diagram of a bearing fault diagnosis system based on hyperwave according to the present application is shown.

[0136] As Figure 4As shown, the second aspect of the present application provides a hyperlet-based bearing fault diagnosis system 4, comprising a memory 41 and a processor 42, the memory comprising a hyperlet-based bearing fault diagnosis method program, the hyperlet-based bearing fault diagnosis method program being executed by the processor to implement the following steps:

[0137] Obtaining a bearing vibration signal;

[0138] Inputting the bearing vibration signal into a preset hyperlet convolutional neural network model, analyzing the bearing vibration signal through a hyperlet block to obtain bearing fault features of multiple scales;

[0139] Weighted fusion of the bearing fault features of multiple scales to obtain a fused fault feature;

[0140] Analyzing the fused fault feature to obtain bearing fault diagnosis information.

[0141] According to the embodiment of the present application, the preset hyperlet convolutional neural network (SWCNN) is established according to the wavelet transform principle. The SWCNN takes the hyperlet convolution block as the first layer of the CNN network, denoises the bearing vibration signal and enhances the key fault features. The present application selects Morlet wavelet, Laplace wavelet and Mexhat wavelet to form the wavelet kernel of the hyperlet block. The hyperlet block adaptively selects and constructs the wavelet basis that can effectively represent the bearing fault signal from the original bearing vibration signal, and simultaneously performs wavelet transform using multiple wavelet basis functions to highlight the bearing fault features at different scales and provide the CNN model with multi-scale fault features. In addition, the hyperlet block can decompose the key feature information from the noise through the multi-resolution analysis characteristic of wavelet transform, thereby improving the noise immunity of the CNN model.

[0142] The SWCNN adopts adaptive wavelet transform (AWT) to adopt Laplace wavelet, Morlet wavelet and Mexh wavelet to form adaptive wavelets, which can integrate the advantages of different wavelet bases. Moreover, the SWCNN does not need to manually select the wavelet basis, but optimizes the adaptive wavelet through the CNN of the wavelet parameters and combination coefficients, thereby learning a set of adaptive wavelet filters with good noise reduction performance. The present application analyzes the SWCNN through two different bearing data sets. As can be seen from the results, the SWCNN not only can well preserve the original information, but also fully considers the internal relationship of the sequence data, and realizes the fault diagnosis in the case of strong noise.

[0143] According to the embodiment of the present application, further comprising:

[0144] Establishing an initial hyperlet convolutional neural network model;

[0145] Obtaining vibration signal sample data;

[0146] The vibration signal sample data is divided into a training set, a validation set and a test set according to a preset ratio;

[0147] The training set, the validation set and the test set are respectively input into the initial superwavelet convolutional neural network model for training diagnosis, and a preset superwavelet convolutional neural network model is obtained.

[0148] It should be noted that the bearing data set of the present scheme experiment comes from a multifunctional bearing fault simulation test bench, and the experimental device is as shown in Figure 5 The experimental bench mainly consists of a driving motor, a transmission part, a support part, a controller and the like. In the rolling bearing fault experiment, signals in four states are collected, and normal bearing, inner ring fault bearing, outer ring fault bearing and rolling element fault bearing are obtained. Among them, the sampling frequency is 8192Hz, the collection time is 120s, and the sampling signal sequence length of each fault working condition is 983040.

[0149] In order to simulate the bearing crack, pitting and other faults in the actual working condition, different depths of faults are implanted on the inner ring, outer ring and rolling element of the rolling bearing by the way of wire cutting processing groove, and the depths are 0.4mm, 0.8mm and 1.2mm respectively. The healthy bearing can be regarded as a special fault mode, so there are ten types of bearing fault in the experimental data set.

[0150] The training process of the SWCNN model is as shown in Figure 6 , which includes three steps: data set division, model training and model diagnosis, and the specific steps are as follows:

[0151] (1) The original data, that is, the vibration signal sample data, is divided into three data sets: training set, validation set and test set according to a preset ratio, and the preset ratio is 8:1:1. Among them, the data of the training set is the original noise-free data, so that the SWCNN can learn the appropriate wavelet basis. The validation set and the test set are added with noise of a certain signal-to-noise ratio, so as to judge whether the trained SWCNN model can resist noise interference and diagnose bearing faults.

[0152] (2) When training the model, first, the network parameters of SWCNN are initialized. Then the training set is input, and the parameters of each layer of SWCNN are trained, so that SWCNN can fully learn the characteristics of the original data. Finally, the trained SWCNN model is tested using the validation set containing different noises to determine whether the accuracy meets the requirements. The above steps are repeated several times, and the optimal model is selected as the diagnosis model.

[0153] (3) In the model diagnosis, firstly, the optimal model in step (2) is taken as the diagnosis model. Then, the test set is input into the diagnosis model. Finally, the result of the diagnosis model is taken as the judgment standard for evaluating the performance of the SWCNN.

[0154] According to the embodiment of the application, after the bearing vibration signal is input into the preset super wavelet convolution neural network model, the method further comprises:

[0155] obtaining wavelet library data;

[0156] analyzing the wavelet library data to obtain a plurality of adaptive wavelets;

[0157] The calculation method of the adaptive wavelet is expressed by a formula as follows:

[0158]

[0159] wherein, is the nth wavelet base function, a n is a scale factor of the wavelet base function, b n is a translation factor of the wavelet base function, w n is a combination coefficient of the super wavelet, and k represents the kth wavelet base function.

[0160] It should be noted that the wavelet transform can decompose the bearing vibration signal to reduce the interference of noise on the bearing vibration signal, and can obtain more information by using the multi-resolution analysis characteristic. In order to use the multi-resolution analysis characteristic of the wavelet transform, the present scheme proposes a super wavelet block (SWB). The SWB is composed of two convolution layers. The first convolution layer is a wavelet transform layer, which is composed of a plurality of different wavelet channels in parallel, and each wavelet channel represents a wavelet base. The second convolution layer is a weighted fusion layer, which is a 1x1 standard convolution layer, and is used for weighted fusion of the features output by different wavelet channels.

[0161] The principle of the SWB is to extract fault features from the bearing vibration signal by using adaptive wavelets. The adaptive wavelets use a multi-wavelet base complementary mechanism to obtain complementary features by using different wavelet bases, and give different weights to the features to fully extract bearing fault features. In addition, the adaptive wavelets satisfy the admissibility condition of the wavelets, that is, they are regarded as a wavelet. Therefore, the SWB can fuse the advantages of different wavelet bases, obtain more fault feature information, and learn new wavelet bases.

[0162] The SWB has two main advantages, the first advantage is that the SWB extracts features through multiple wavelet basis functions to match different fault characteristics in the vibration signal, to obtain more useful information, and to increase the noise resistance of the CNN model. The second advantage is that the SWB can directly input the original vibration signal, well preserving the original information and reducing the loss of original vibration signal information caused by denoising preprocessing.

[0163] According to the embodiment of the present application, the bearing vibration signal is analyzed by the super wavelet block to obtain bearing fault features of multiple scales, comprising:

[0164] The super wavelet block performs wavelet transform on the bearing vibration signal through the multiple adaptive wavelets, so that the super wavelet block decomposes and denoises the bearing vibration signal through wavelet transform to obtain a denoised vibration signal;

[0165] The denoised vibration signal is feature-extracted by the multiple adaptive wavelets to obtain bearing fault features of multiple scales;

[0166] The wavelet transform is expressed by a formula as follows:

[0167]

[0168] Wherein, ψ represents a mother wavelet, t is time, a is a scale factor inversely proportional to frequency, and b is a translation factor.

[0169] It should be noted that the wavelet transform layer can extract bearing fault features from different angles using multiple wavelet bases. When the SWB is working, the bearing vibration signal is first input to the wavelet transform layer, the input signal is subjected to wavelet transform in the wavelet channel to decompose the bearing vibration signal, reduce noise interference, and extract bearing fault features.

[0170] The wavelet transform provides a time-frequency window that changes with the signal frequency, overcoming the shortcoming that the window size of the short-time Fourier transform does not change with the frequency. It can be used for time-frequency analysis and processing of signals. Therefore, the wavelet transform has the characteristic of multi-resolution analysis, and can simultaneously perform localized analysis of the bearing vibration signal in the time domain and the frequency domain. The signal can be decomposed into subbands with multiple scales and multiple frequencies. The signal has high time resolution and low frequency resolution at high frequencies, and has high frequency resolution and low time resolution at low frequencies. The wavelet transform can represent time and frequency information simultaneously. The multi-resolution characteristic of the wavelet transform makes it excellent in signal decomposition, denoising, and analysis of non-stationary signals.

[0171] A wavelet is a function that has localization in both time and frequency domains. Each wavelet is a variation of a single mother wavelet ψ, called a wavelet member. A wavelet library {ψ a,b(t)} is composed of multiple wavelet members, which can be constructed by controlling the scale factor a and the translation factor b of the mother wavelet ψ(·). The scale factor a can make the signal contract or stretch. When the scale factor is low, the signal contracts. When the scale factor is high, the signal stretches. The wavelet transform contains three steps:

[0172] (1) A wavelet base function is selected, and then a similarity coefficient with a partial signal is calculated.

[0173] (2) The wavelet is shifted to the right by b, and the similarity coefficient is continuously calculated until all parts of the signal are calculated.

[0174] (3) The scale parameter a of the wavelet is changed, steps (1) and (2) are repeated, and all scale analyses are completed.

[0175] According to an embodiment of the present application, the bearing fault features of the multiple scales are weighted and fused to obtain a fused fault feature, including:

[0176] The bearing fault features of the multiple scales are multiplied by corresponding influence weights respectively to obtain weight scores of the multiple fault features;

[0177] A plurality of key fault features are selected according to the weight scores of the multiple fault features;

[0178] The multiple key fault features are fused to obtain a fused fault feature.

[0179] It should be noted that the weighted fusion layer is a 1x1 standard convolution layer, which is used to weight and fuse the features output by different wavelet channels. Since the wavelet base functions in the wavelet channels are different, the importance of the bearing fault features output from different wavelet channels is different. Then, the weighted fusion layer gives different weights to the output bearing fault features through a group of 1x1 standard convolution layers, highlights the key fault features, determines the best bearing fault features, and fuses the features. Finally, the SWB outputs a set of weighted and fused fault features, and takes the set of features as the input of the CNN network, thereby enhancing the anti-noise ability of the CNN network.

[0180] According to an embodiment of the present application, the bearing fault diagnosis information is obtained by analyzing the fused fault feature, including:

[0181] The fused fault feature is input to a plurality of convolution layers;

[0182] The plurality of convolution layers process the fused fault feature according to a preset method, and send the processed fault feature to a fully connected layer;

[0183] The fully connected layer integrates and classifies the processed fault feature to obtain bearing fault diagnosis information.

[0184] It should be noted that the SWCNN includes a super wavelet block, 4 standard convolution layers, some operations such as batch normalization, ReLu and global maximum pooling, and a fully connected layer. The preset method is specifically that the output of the super wavelet block will pass through multiple convolution layers for enhancing feature expression. Then, some operations such as batch normalization, ReLu and global maximum pooling are sequentially performed on the output of each convolution layer. Finally, a classification result is output through the fully connected layer.

[0185] In addition, the second layer convolution of the SWCNN adopts a wide convolution kernel to increase a receptive field and improve diagnostic accuracy. The subsequent convolution layers all adopt small convolution kernels to deeply mine hidden deep information among samples and reduce network parameter quantity. In addition, batch normalization and pooling layers are added after all the convolution layers to accelerate the training speed and reduce the parameter quantity.

[0186] According to the embodiment of the present application, further comprising:

[0187] The vibration signal sample data is expanded by the method of overlapping sampling.

[0188] It should be noted that, in the CWRU data set, each fault data set has only 120,000 sampling points, and the training sample is limited. When the training sample is too small, the model will be under-fitted, and the accuracy of the training set and the test set will be reduced. In order to avoid the problem of under-fitting, the method of overlapping sampling is adopted to expand the data set. The size of the sliding window is 2048, and the sliding step (offset) is 100. Starting from the starting position, the sliding window collects 2048 sampling points to sequentially construct data samples by sliding 100 sampling points backward each time, until there are 600 samples.

[0189] According to the embodiment of the present application, further comprising:

[0190] In the preset super wavelet convolution neural network model training process, the training data is processed through a cross-entropy function and an Adam algorithm.

[0191] It should be noted that in order to improve the performance of the model and make the experimental results optimal, the model parameters are selected in the scheme. In the training strategy of the scheme, the cross-entropy function and Adam are selected as the loss function and the optimizer to speed up the convergence speed of the model and obtain the optimal result. Cross-entropy is a loss function used to evaluate the difference between the predicted probability distribution and the actual distribution. The smaller the cross-entropy, the closer the two probability distributions, and the higher the accuracy of the model. Adam is a learning rate adaptive optimization algorithm that can improve the robustness of the learning rate by dynamically adjusting the learning rate parameter, and help the model converge. The learning rate needs to be set in an appropriate range when training the CNN model. Too high learning rate may cause instability, and too small learning rate will increase the training time. Under the premise of ensuring the stability of training, reduce the training time, the learning rate of the scheme is set to 0.001. At the same time, the training batch size of the scheme is 128.

[0192] The third aspect of the application provides a computer readable storage medium, the computer readable storage medium comprises a hyperwave-based bearing fault diagnosis method program, and the hyperwave-based bearing fault diagnosis method program is executed by a processor to realize the steps of the hyperwave-based bearing fault diagnosis method according to any one of the preceding aspects.

[0193] The application discloses a hyperwave-based fan gear box bearing fault diagnosis method and system and a storage medium. The method comprises the following steps: acquiring a bearing vibration signal; inputting the bearing vibration signal into a preset hyperwave convolutional neural network model; analyzing the bearing vibration signal through a hyperwave block to obtain bearing fault features of multiple scales; performing weighted fusion on the bearing fault features of the multiple scales to obtain fused fault features; and analyzing the fused fault features to obtain bearing fault diagnosis information. The application processes the bearing vibration signal through an SWCNN, processes the bearing vibration signal containing noise through a hyperwave block, performs noise reduction processing and feature extraction by using three different wavelet basis functions, and enhances the noise resistance of the SWCNN network. In addition, the weight fusion layer of the hyperwave block can adaptively determine the fusion weight, thereby enhancing the wavelet channel with obvious fault features, highlighting the key features, and improving the diagnosis accuracy of the SWCNN.

[0194] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other manners. The described device embodiments are merely illustrative. For example, the division of the units is only a logical function division. There can be another division manner for the actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling, or direct coupling or communication connection between the components can be indirect coupling or communication connection through some interfaces, devices, or units, and can be electrical, mechanical, or in other forms.

[0195] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units; they can be located in one place, or distributed on multiple network units; and some or all of the units can be selected as needed to achieve the purposes of the embodiments.

[0196] In addition, each functional unit in each embodiment of the present application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; and the integrated unit can be implemented in the form of hardware or hardware plus software functional units.

[0197] Those of ordinary skill in the art can understand that all or part of the steps of the above-described method embodiments can be completed by a program instructing related hardware, and the foregoing program can be stored in a computer readable storage medium, and when the program is executed, the steps of the method embodiments are executed; and the foregoing storage medium includes mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks or optical disks, and various media that can store program codes.

[0198] Alternatively, the integrated units of the present application, if implemented in the form of software functional modules and sold or used as independent products, can also be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present application can be embodied in the form of a software product, and the computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the embodiments of the present application. The foregoing storage medium includes mobile storage devices, ROMs, RAMs, magnetic disks or optical disks, and various media that can store program codes.

Claims

1. A method for diagnosing bearing faults of a fan gearbox based on super wavelets, characterized in that, The method comprises the following steps: obtaining a bearing vibration signal; inputting the bearing vibration signal into a preset hyperwave convolutional neural network model, analyzing the bearing vibration signal through a hyperwave block to obtain bearing fault features of multiple scales; performing weighted fusion on the bearing fault features of multiple scales to obtain fused fault features; analyzing the fused fault features to obtain bearing fault diagnosis information; establishing an initial hyperwave convolutional neural network model; obtaining a wind turbine gearbox bearing vibration signal sample data; dividing the vibration signal sample data into a training set, a validation set and a test set according to a preset proportion; inputting the training set, the validation set and the test set into the initial hyperwave convolutional neural network model for training diagnosis to obtain a preset hyperwave convolutional neural network model; after inputting the bearing vibration signal into the preset hyperwave convolutional neural network model, further comprising the following steps: obtaining a plurality of wavelet library data; analyzing the wavelet library data to obtain a plurality of adaptive wavelets; the calculation method of the adaptive wavelet is expressed by a formula as follows: ； wherein is a first wavelet basis function, is a scale factor of the wavelet basis function, is a translation factor of the wavelet basis function, is a combination coefficient of the hyperwavelet, denotes a first class wavelet basis function; analyzing the bearing vibration signal through the hyperwave block to obtain bearing fault features of multiple scales, comprising the following steps: the hyperwave block performs wavelet transform on the bearing vibration signal through the plurality of adaptive wavelet bases, so that the hyperwave block decomposes and denoises the bearing vibration signal through wavelet transform to obtain a denoised vibration signal; extracting features from the denoised vibration signal through the plurality of adaptive wavelets to obtain bearing fault features of multiple scales; the wavelet transform is expressed by a formula as follows: ； wherein, denotes a mother wavelet, t is time, is a scale factor inversely proportional to the frequency, is a translation factor; analyzing the fused fault features to obtain bearing fault diagnosis information, comprising the following steps: inputting the fused fault features into a plurality of convolutional layers; the plurality of convolutional layers process the fused fault features according to a preset method, and send the processed fault features to a fully connected layer; the fully connected layer integrates and classifies the processed fault features to obtain bearing fault diagnosis information.

2. The ultra-wavelet-based bearing fault diagnosis method for a fan gearbox according to claim 1, characterized in that, the weighted fusion of the bearing fault features of multiple scales to obtain fused fault features, comprising the following steps: multiplying the bearing fault features of multiple scales by corresponding influence weights to obtain weight scores of the plurality of fault features; selecting a plurality of key fault features according to the weight scores of the plurality of fault features; performing feature fusion on the plurality of key fault features to obtain fused fault features.

3. A system for ultra wavelet based bearing fault diagnosis of a fan gearbox, for implementing the method of claim 1, wherein, The device comprises a memory and a processor, and the memory comprises a hyperwave-based bearing fault diagnosis method program, which is executed by the processor to implement the following steps: obtaining a bearing vibration signal of a wind turbine gearbox; inputting the bearing vibration signal into a preset hyperwave convolutional neural network model, analyzing the bearing vibration signal through a hyperwave block to obtain bearing fault features of multiple scales; performing weighted fusion on the bearing fault features of multiple scales to obtain fused fault features; analyzing the fused fault features to obtain bearing fault diagnosis information.

4. A computer-readable storage medium, characterized in that, The computer readable storage medium comprises an ultra-wavelet-based bearing fault diagnosis method program, and the processor executes the ultra-wavelet-based bearing fault diagnosis method program to realize the steps of the ultra-wavelet-based bearing fault diagnosis method in any one of claims 1 to 2.

Images