Rolling bearing fault diagnosis method

Abstract

The invention relates to a rolling bearing fault diagnosis method. The invention relates to the technical field of rolling bearing fault diagnosis, and the method comprises the steps: decomposing a bearing signal through peak envelope spectrum Fourier decomposition to obtain a component signal, carrying out the envelope spectrum analysis of the component signal, and converting a component envelope spectrum into a gray-scale map as an input feature of a convolutional neural network for fault diagnosis. According to the method, the original bearing signal is decomposed by using peak envelope spectrum Fourier decomposition, a frequency band with rich fault information is effectively separated from a noise frequency band, the component signal is converted from a time domain into an envelope spectrum and further converted into a feature enhanced grey-scale map, the obtained envelope spectrum grey-scale map has obvious fault features, and the fault detection accuracy is improved. And the problem that a traditional fault diagnosis method needs complex feature extraction is avoided. Diagnosis experiments are carried out on bearing vibration signals of different fault types, and the result shows that the method is higher in diagnosis precision and higher in anti-noise capacity.

Description

Technical Field

[0001] The invention relates to the technical field of rolling bearing fault diagnosis, and relates to a rolling bearing fault diagnosis method. Background Art

[0002] With the development of science and technology, industrial equipment is gradually developing in the direction of intelligence and precision, and the condition monitoring and maintenance of machinery and equipment have received widespread attention. Rotating machinery is one of the most common types of mechanical equipment and plays an important role in industrial equipment. As a key component of rotating machinery and equipment, rolling bearings will inevitably suffer from wear, pitting, cracks and other faults due to long-term operation under harsh working conditions such as high speed and heavy load. According to statistics, 30% of the failures of rotating machinery and equipment are caused by rolling bearing failures. If the bearing failure cannot be effectively identified, it may lead to accidents such as reduced processing quality and unexpected shutdown of machinery and equipment, which will cause economic losses and even safety accidents. Therefore, it is of great significance to study the fault diagnosis method of rolling bearings and realize intelligent monitoring of the health status of rolling bearings during operation, so as to reasonably arrange the maintenance time and service life of the equipment and ensure the normal operation of machinery and equipment.

[0003] The fault diagnosis method based on traditional machine learning requires complex fault feature vector extraction, and the fault feature extraction method directly affects the accuracy of the diagnosis result. Deep learning is a new research direction in the field of machine learning. In the past decade, with the rapid development of computer technology, the model complexity and data capacity of deep learning have continued to increase, and its classification accuracy has continued to improve. Deep learning does not require manual feature extraction, and has excellent ability to extract complex nonlinear features, which is suitable for the complex characteristics of mechanical equipment failure. The two-dimensional convolutional neural network (2D-CNN) in deep learning has strong image feature extraction capabilities, powerful prediction and classification functions, and is a widely used method in the field of fault diagnosis. The input images of the fault diagnosis method based on 2D-CNN include time-frequency diagrams, polar coordinate symmetric images, time domain grayscale images, etc. Among them, the time-frequency diagrams generated by the time-frequency analysis method are particularly widely used. Commonly used time-frequency analysis tools include: short-time Fourier transform, continuous wavelet transform, synchronous compression wavelet transform, etc. However, the generation speed of the time-frequency diagram is not fast enough, which is undoubtedly very time-consuming for the large number of samples to be prepared for CNN training and testing. In addition, the existing feature extraction methods cannot make good use of fault diagnosis knowledge to obtain enhanced fault feature representation, which limits the improvement of the final diagnosis accuracy and has weak anti-noise interference ability. If a good two-dimensional fault feature input image is obtained, it will be beneficial to improve the fault diagnosis accuracy of 2D-CNN and reduce the complexity of the network model. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a rolling bearing fault diagnosis method based on peak envelope spectrum Fourier decomposition envelope spectrum grayscale image and convolutional neural network. This method overcomes the problem that traditional machine learning fault diagnosis methods require complex fault feature vector extraction, and has higher fault diagnosis accuracy and stronger anti-noise interference ability, which is superior to traditional rolling bearing fault diagnosis methods based on machine learning and 2D-CNN.

[0005] The present invention provides a rolling bearing fault diagnosis method, and the present invention provides the following technical solutions:

[0006] A rolling bearing fault diagnosis method, based on peak envelope spectrum Fourier decomposition envelope spectrum grayscale image and convolutional neural network, comprises the following steps:

[0007] Step 1: Collect vibration signals of rolling bearings in normal state, different fault types and different damage degrees;

[0008] Step 2: Segment the collected vibration signals of the rolling bearing to form labeled data samples, and divide the data samples into a training set and a test set;

[0009] Step 3: Decompose the data sample x(k) using the peak envelope spectrum Fourier decomposition to obtain a series of component signals {x l (k)};

[0010] Step 4: Perform envelope spectrum analysis on all component signals to obtain the component envelope spectrum ES l (n);

[0011] Step 5: Convert the envelope spectra of all components into data-enhanced grayscale images;

[0012] Step 6: Design a 2D-CNN for rolling bearing fault feature learning;

[0013] Step 7: Input the training set into the 2D-CNN to train the network, save the trained network model, and input the test set data into the trained 2D-CNN to complete fault diagnosis.

[0014] Preferably, in step 3, the signal is decomposed using PESFDM, and the number of PESFDM components is selected in the range of 6-10.

[0015] Preferably, the specific steps of step 4 are:

[0016] For signal x l (k) Perform Hilbert transform to get x l '(k),x l The analytical signal of (k) is q l (k), as shown in formula (1), xl (k) Envelope signal |q l (k)|As shown in formula (2):

[0017] q l (k) = x l (k)+i·x l '(k) k=0,2,3…N-1 (1)

[0018]

[0019] Yes | l (k) | Perform FFT to obtain the complex Q of the envelope signal l (n), as shown in equation (3), where the FFT transform length is the signal length, |Q l (n)| is twice the first half, which is x l The envelope spectrum of (k) is expressed as ES l (n) is expressed as shown in formula (4):

[0020]

[0021] Preferably, the specific steps of step 5 are:

[0022] Step 5.1: The one-dimensional envelope spectrum amplitude sequence of the selected component is converted into a grayscale image in the following manner: in order to enhance the fault characteristics in the grayscale image, the amplitude sequence of the first 1 / 10 part of each component signal is first formed into a matrix of 3 rows and N / 20 columns, that is, the selected envelope spectrum amplitude sequence is repeatedly arranged into three rows;

[0023] Step 5.2: Combine the selected envelope spectra matrices of all components into a matrix of 3M rows and N / 20 columns in the order of component numbers. This matrix is ​​the grayscale image matrix used by the grayscale image. The individual elements in the grayscale image matrix are numbered from left to right and from top to bottom. There are 3×M×N / 20 elements in total. Then, the grayscale image matrix is ​​normalized to 0-255 using formula (5):

[0024]

[0025] Where A represents all elements of the matrix, A i is the i-th element in the matrix, min(A) and max(A) represent the minimum and maximum values ​​in A respectively, A′ i is the normalized element value. The value of each element after normalization is the grayscale value of a color block in the grayscale image. Finally, according to the rule that grayscale from 0 to 255 corresponds to grayscale from dark to bright, the matrix is ​​converted into a grayscale image.

[0026] Preferably, a 2D-CNN including two convolution-pooling pairs and two fully connected layers is designed. To speed up network training, a batch normalization layer is added between the convolution layer and the activation function and between the fully connected layer and the activation function. The random inactivation ratio adopts the recommended value: 0.5.

[0027] Preferably, the moving step of the convolution kernel is fixed to 1, the number of convolution kernels of the first convolution layer is 16, and the number of convolution kernels of the second layer is 32, and the number of convolution kernels is increased by multiples to improve the feature extraction capability;

[0028] Both convolutional layers use a small receptive field of size 3×3 to capture detail information, and the padding method is “same”. When the convolution kernel movement step is 1, the side length of the output feature map after the convolution operation remains unchanged, which helps to retain information about the input.

[0029] Preferably, the pooling layer uses a 2×2 pooling kernel used by most models, with a moving step of 2 and no padding. After the pooling operation, the length and width of the feature map are reduced by half, thereby improving the calculation speed of the model.

[0030] After the convolution and pooling pairs, there are two fully connected layers. The number of neurons in the first fully connected layer is 128, and the number of neurons in the second fully connected layer is the number of classification patterns.

[0031] A rolling bearing fault diagnosis system, the system comprising:

[0032] A collection module, which collects vibration signals of the normal state of the rolling bearing, different fault types and different damage degrees;

[0033] A segmentation module, wherein the segmentation module segments the collected vibration signals of the rolling bearing to form labeled data samples, and divides the data samples into a training set and a test set;

[0034] A decomposition module is used to decompose the data sample x(k) using the peak envelope spectrum Fourier decomposition to obtain a series of component signals {x l (k)};

[0035] An analysis module performs envelope spectrum analysis on all component signals to obtain the component envelope spectrum ES l (n);

[0036] A conversion module, wherein the conversion module converts the envelope spectra of all components into a data enhanced grayscale image;

[0037] A design module, wherein the design module designs a 2D-CNN for rolling bearing fault feature learning;

[0038] The fault diagnosis module inputs the training set into the 2D-CNN to train the network, saves the trained network model, and inputs the test set data into the trained 2D-CNN to complete the fault diagnosis.

[0039] A computer-readable storage medium stores a computer program, which is executed by a processor to implement a rolling bearing fault diagnosis method.

[0040] A computer device comprises a memory and a processor, wherein the memory stores a computer program and the processor implements a rolling bearing fault diagnosis method when executing the computer program.

[0041] The present invention has the following beneficial effects:

[0042] Compared with the prior art, the present invention has the following advantages:

[0043] The signal preprocessing method of the peak envelope spectrum Fourier decomposition envelope spectrum grayscale image proposed in the present invention has strong ability to extract rolling bearing fault feature information, strong anti-noise ability, and can accurately reflect the health status information of the bearing. Using it as an input parameter can improve the accuracy of CNN fault mode recognition.

[0044] The present invention uses bearing data of different fault types to carry out fault diagnosis experiments. Compared with the traditional fault diagnosis method based on machine learning and 2D-CNN, the proposed method has stronger anti-noise interference ability.

[0045] The PESFDM-ESGI signal preprocessing method proposed in the present invention does not require complex mathematical operations. Compared with the method of obtaining time-frequency feature images through complex time-frequency analysis, it greatly shortens the preparation time of CNN input feature maps.

[0046] The envelope grayscale image proposed in the present invention has obvious features and a small image size, and has low requirements on the calculation amount of the CNN model. The fault diagnosis method based on the envelope grayscale image and 2D-CNN requires shorter training time than other methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0048] Figure 1 A rolling bearing fault diagnosis method based on peak envelope spectrum Fourier decomposition envelope spectrum grayscale image and convolutional neural network;

[0049] Figure 2 It is the envelope spectrum of the components of normal signal, inner race fault signal, outer race fault signal and rolling element fault signal after being decomposed by PESFDM;

[0050] Figure 3 The grayscale image of the envelope spectrum of PESFDM for normal signal, inner race fault signal, outer race fault signal and rolling element fault signal;

[0051] Figure 4 CWT time-frequency diagram of four signals;

[0052] Figure 5 It is the training process based on PESFDM-CNN and CWT-CNN;

[0053] Figure 6 It is the confusion matrix of the test set based on PESFDM-CNN and CWT-CNN. DETAILED DESCRIPTION

[0054] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of 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 scope of protection of the present invention.

[0055] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0056] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0057] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0058] The present invention is described in detail below in conjunction with specific embodiments. Specific embodiment one:

[0060] according to Figures 1 to 6 As shown, the specific optimization technical solution adopted by the present invention to solve the above technical problems is: the present invention relates to a rolling bearing fault diagnosis method.

[0061] A rolling bearing fault diagnosis method, based on peak envelope spectrum Fourier decomposition envelope spectrum grayscale image and convolutional neural network, comprises the following steps:

[0062] Step 1: Collect vibration signals of rolling bearings in normal state, different fault types and different damage degrees;

[0063] Step 2: Segment the collected vibration signals of the rolling bearing to form labeled data samples, and divide the data samples into a training set and a test set;

[0064] Step 3: Decompose the data sample x(k) using the peak envelope spectrum Fourier decomposition to obtain a series of component signals {x l (k)};

[0065] Step 4: Perform envelope spectrum analysis on all component signals to obtain the component envelope spectrum ES l (n);

[0066] Step 5: Convert the envelope spectra of all components into data-enhanced grayscale images;

[0067] Step 6: Design a 2D-CNN for rolling bearing fault feature learning;

[0068] Step 7: Input the training set into the 2D-CNN to train the network, save the trained network model, and input the test set data into the trained 2D-CNN to complete fault diagnosis. Specific embodiment 2:

[0070] The difference between the second embodiment of the present application and the first embodiment is that:

[0071] Step 3: Decompose the signal using PESFDM, where the number of PESFDM components is selected within the range of 6-10. Specific embodiment three:

[0073] The difference between the third embodiment of the present application and the second embodiment is that:

[0074] The specific steps of step 4 are:

[0075] For signal x l (k) Perform Hilbert transform to get xl '(k),x l The analytical signal of (k) is q l (k), as shown in formula (1), x l (k) Envelope signal |q l (k)|As shown in formula (2):

[0076] q l (k) = x l (k)+i·x l '(k) k=0,2,3…N-1 (1)

[0077]

[0078] Yes | l (k) | Perform FFT to obtain the complex Q of the envelope signal l (n), as shown in equation (3), where the FFT transform length is the signal length, |Q l (n)| is twice the first half, which is x l The envelope spectrum of (k) is expressed as ES l (n) is expressed as shown in formula (4):

[0079] Specific embodiment four:

[0081] The difference between the fourth embodiment of the present application and the third embodiment is that:

[0082] The specific steps of step 5 are:

[0083] Step 5.1: The one-dimensional envelope spectrum amplitude sequence of the selected component is converted into a grayscale image in the following manner: in order to enhance the fault characteristics in the grayscale image, the amplitude sequence of the first 1 / 10 part of each component signal is first formed into a matrix of 3 rows and N / 20 columns, that is, the selected envelope spectrum amplitude sequence is repeatedly arranged into three rows;

[0084] Step 5.2: Combine the selected envelope spectra matrices of all components into a matrix of 3M rows and N / 20 columns in the order of component numbers. This matrix is ​​the grayscale image matrix used by the grayscale image. The individual elements in the grayscale image matrix are numbered from left to right and from top to bottom. There are 3×M×N / 20 elements in total. Then, the grayscale image matrix is ​​normalized to 0-255 using formula (5):

[0085]

[0086] Where A represents all elements of the matrix, A i is the i-th element in the matrix, min(A) and max(A) represent the minimum and maximum values ​​in A respectively, A′i is the normalized element value. The value of each element after normalization is the grayscale value of a color block in the grayscale image. Finally, according to the rule that grayscale from 0 to 255 corresponds to grayscale from dark to bright, the matrix is ​​converted into a grayscale image. Specific embodiment five:

[0088] The difference between the fifth embodiment of the present invention and the fourth embodiment is that:

[0089] A 2D-CNN with two convolution-pooling pairs and two fully connected layers is designed. To speed up network training, batch normalization layers are added between the convolution layer and the activation function and between the fully connected layer and the activation function. The random dropout ratio adopts the recommended value: 0.5. Specific embodiment six:

[0091] The difference between the sixth embodiment of the present invention and the fifth embodiment is that:

[0092] The moving step of the convolution kernel is fixed to 1. The number of convolution kernels in the first convolution layer is 16, and the second layer is 32. The number of convolution kernels increases exponentially to improve the feature extraction capability.

[0093] Both convolutional layers use a small receptive field of size 3×3 to capture detail information, and the padding method is “same”. When the convolution kernel movement step is 1, the side length of the output feature map after the convolution operation remains unchanged, which helps to retain information about the input. Specific embodiment seven:

[0095] The difference between the seventh embodiment of the present invention and the sixth embodiment is that:

[0096] The pooling layer uses a 2×2 pooling kernel used by most models, with a moving step of 2 and no padding. After the pooling operation, the length and width of the feature map are reduced by half, which improves the calculation speed of the model.

[0097] After the convolution and pooling pairs, there are two fully connected layers. The number of neurons in the first fully connected layer is 128, and the number of neurons in the second fully connected layer is the number of classification patterns. Specific embodiment eight:

[0099] The difference between the eighth embodiment of the present invention and the seventh embodiment is that:

[0100] The present invention provides a rolling bearing fault diagnosis system, the system comprising:

[0101] A collection module, which collects vibration signals of the normal state of the rolling bearing, different fault types and different damage degrees;

[0102] A segmentation module, wherein the segmentation module segments the collected vibration signals of the rolling bearing to form labeled data samples, and divides the data samples into a training set and a test set;

[0103] A decomposition module is used to decompose the data sample x(k) using the peak envelope spectrum Fourier decomposition to obtain a series of component signals {x l (k)};

[0104] An analysis module performs envelope spectrum analysis on all component signals to obtain the component envelope spectrum ES l (n);

[0105] A conversion module, wherein the conversion module converts the envelope spectra of all components into a data enhanced grayscale image;

[0106] A design module, wherein the design module designs a 2D-CNN for rolling bearing fault feature learning;

[0107] The fault diagnosis module inputs the training set into the 2D-CNN to train the network, saves the trained network model, and inputs the test set data into the trained 2D-CNN to complete the fault diagnosis. Specific embodiment nine:

[0109] The difference between the ninth embodiment of the present invention and the eighth embodiment is that:

[0110] The present invention provides a computer-readable storage medium on which a computer program is stored. The program is executed by a processor to implement a rolling bearing fault diagnosis method.

[0111] The following steps are involved:

[0112] Step 1: Collect vibration signals of rolling bearings in normal state, different fault types and different damage degrees;

[0113] Step 2: Segment the collected vibration signals of the rolling bearing to form labeled data samples, and divide the data samples into a training set and a test set;

[0114] Step 3: Decompose the data sample x(k) using the peak envelope spectrum Fourier decomposition to obtain a series of component signals {x l (k)};

[0115] Step 4: Perform envelope spectrum analysis on all component signals to obtain the component envelope spectrum ES l (n);

[0116] Step 5: Convert the envelope spectra of all components into data-enhanced grayscale images;

[0117] Step 6: Design a 2D-CNN for rolling bearing fault feature learning;

[0118] Step 7: Input the training set into the 2D-CNN to train the network, save the trained network model, and input the test set data into the trained 2D-CNN to complete fault diagnosis. Specific embodiment ten:

[0120] The difference between the tenth embodiment of the present invention and the ninth embodiment is that:

[0121] The present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements a rolling bearing fault diagnosis method when executing the computer program.

[0122] Step 1: Collect vibration signals of rolling bearings in normal state, different fault types and different damage degrees;

[0123] Step 2: Segment the collected vibration signals of the rolling bearing to form labeled data samples, and divide the data samples into a training set and a test set;

[0124] Step 3: Decompose the data sample x(k) using the peak envelope spectrum Fourier decomposition to obtain a series of component signals {x l (k)};

[0125] Step 4: Perform envelope spectrum analysis on all component signals to obtain the component envelope spectrum ES l (n);

[0126] Step 5: Convert the envelope spectra of all components into data-enhanced grayscale images;

[0127] Step 6: Design a 2D-CNN for rolling bearing fault feature learning;

[0128] Step 7: Input the training set into the 2D-CNN to train the network, save the trained network model, and input the test set data into the trained 2D-CNN to complete fault diagnosis. Specific embodiment eleven:

[0130] The difference between the eleventh embodiment of the present invention and the tenth embodiment is that:

[0131] Using vibration signals to identify whether a rolling bearing is faulty is currently the main diagnostic method, but the many noise interferences in the vibration signal will seriously reduce the accuracy of fault diagnosis. PESFDM is an effective signal decomposition and denoising method that decomposes the fault information band submerged in the noise information. Therefore, the present invention combines the characteristics of PESFDM for signal decomposition and denoising with the advantages of 2D-CNN in extracting two-dimensional image features, and proposes a fault diagnosis method based on PESFDM envelope grayscale image and 2D-CNN.

[0132] like Figure 1 As shown in FIG. 1 , a rolling bearing fault diagnosis method based on peak envelope spectrum Fourier decomposition envelope spectrum grayscale image and convolutional neural network is described in detail in the following steps:

[0133] Step 1: Collect vibration signals of rolling bearings in normal state, different fault types and different damage degrees;

[0134] Step 2: Segment the collected vibration signal of the rolling bearing to form labeled data samples;

[0135] Step 3: Decompose the data sample x(k) using the peak envelope spectrum Fourier decomposition to obtain a series of component signals {x l (k)};

[0136] Step 4: Perform envelope spectrum analysis on all component signals to obtain the component envelope spectrum ES l (n);

[0137] Step 5: Convert the envelope spectra of all components into grayscale images for data enhancement, and divide the data samples into training set and test set.

[0138] Step 6: Design a 2D-CNN for rolling bearing fault feature learning;

[0139] Step 7: Input the training set into the 2D-CNN to train the network, save the trained network model, and input the test set data into the trained 2D-CNN to complete fault diagnosis.

[0140] In the above technical solution, the different fault types in step 1 include rolling element fault, rolling inner ring fault, and rolling outer ring fault, and each fault type can be divided into a weak fault, a medium fault, and a severe fault.

[0141] In step 2, the feature learning ability of 2D-CNN is improved by using the sliding overlapping sampling method to increase the number of samples, and the number of samples is adjusted according to the size of the overlap rate of sample data points.

[0142] In step 3, the number of PESFDM components can be selected in the range of 6-10, which can be adjusted according to the width of the signal spectrum.

[0143] In step 4, the component signal x l (k) The specific steps of envelope spectrum analysis are: l (k) Perform Hilbert transform to get x l '(k),x l The analytical signal of (k) is q l (k), as shown in formula (1), x l (k) Envelope signal |q l (k)|As shown in formula (2).

[0144] q l (k) = x l (k)+i·x l '(k) k=0,2,3…N-1 (1)

[0145]

[0146] Yes | l (k) | Perform FFT to obtain the complex Q of the envelope signal l (n), as shown in equation (3), where the FFT transform length is the signal length. l (n)| is twice the first half, which is x l The envelope spectrum of (k) is expressed as ES l (n) is represented as shown in formula (4).

[0147]

[0148] In step 5, the one-dimensional envelope spectrum amplitude sequence of the selected component is converted into a grayscale image in the following manner: to enhance the fault features in the grayscale image, the amplitude sequence of the first 1 / 10 part of each component signal is first formed into a matrix of 3 rows and N / 20 columns, that is, the selected envelope spectrum amplitude sequence is repeatedly arranged into three rows; then, the matrices of the selected envelope spectra of all components are combined into a matrix of 3M rows and N / 20 columns in the order of component numbers, which is the grayscale image matrix used for the grayscale image. The individual elements in the grayscale image matrix are numbered from left to right and from top to bottom, with a total of 3×M×N / 20 elements, and then the grayscale image matrix is ​​normalized to 0-255 using formula (5).

[0149]

[0150] Where A represents all elements of the matrix, A i is the i-th element in the matrix, min(A) and max(A) represent the minimum and maximum values ​​in A respectively, A′ iis the normalized element value. The value of each element after normalization is the grayscale value of a color block in the grayscale image. Finally, according to the rule that grayscale from 0 to 255 corresponds to grayscale from dark to bright, the matrix is ​​converted into a grayscale image.

[0151] Furthermore, a 2D-CNN is designed, which mainly consists of two convolution-pooling pairs and two fully connected layers. To speed up the network training, a batch normalization layer is added between the convolution layer and the activation function, and between the fully connected layer and the activation function. In addition, random dropout is added between the two fully connected layers as a simple and effective regularization method to reduce the overfitting of the model and improve the generalization ability of the model. The random dropout ratio in this paper adopts the recommended value: 0.5.

[0152] The moving step of the convolution kernel is fixed to 1. The number of convolution kernels in the first convolution layer is 16, and the second layer is 32. The number of convolution kernels increases exponentially to improve the feature extraction capability. Both convolution layers use a small receptive field of size 3×3 to capture detail information. The padding mode is "same". When the moving step of the convolution kernel is 1, the side length of the output feature map after the convolution operation can be kept unchanged, which helps to retain information about the input. The pooling layer uses a 2×2 pooling kernel of size used by most models, with a moving step of 2 and no padding. After the pooling operation, the length and width of the feature map are reduced by half, which improves the calculation speed of the model. The two pairs of convolution and pooling pairs are followed by two fully connected layers. The number of neurons in the first fully connected layer is 128, and the number of neurons in the second fully connected layer is the number of classification patterns.

[0153] This embodiment uses real bearing data for analysis, and the experimental data comes from the widely used bearing data of Case Western Reserve University. Electrospark machining is used to pre-set faults on the inner ring, outer ring and rolling element of the motor bearing. The three types of faults include faults with diameters of 0.18mm, 0.36mm and 0.53mm. The signal from the drive end with a working condition of 1hp and a sampling frequency of 12kHz is used for fault diagnosis. Including normal signals and fault signals of different damage degrees of the three fault types, there are a total of ten signals.

[0154] Combined with the signal data length used for a single sample and the length of the original signal, if the non-overlapping sampling method is used for data segmentation, the number of samples of a single type of signal is too small, and CNN learning is not sufficient, which may limit the improvement of diagnostic accuracy. Therefore, in order to increase the number of samples, a sliding sampling method is used to obtain samples. When the signal length of a single sample is 4096 and the number of samples is 118, the data overlap rate of PESFDM-ESGI is 75.83%, and the slip length is 990. For 10 types of bearing signals, the number of components M of PESFDM is set to 6, and the size of the PESFDM envelope spectrum grayscale image is 18×205. Among them, the envelope spectra of normal signals and three minor fault signals are as follows Figure 3 , its grayscale image is as follows Figure 4 .

[0155] exist Figure 3 In the envelope spectrum of the normal signal, the fault characteristic frequency cannot be observed, but the IMF2 can observe f r and 2f r For the inner race and outer race fault signals, the periodic fault impact in their time domain signals is obvious and less affected by noise. After PESFDM decomposition, the envelope spectra of their component signals almost all contain the peaks of the fault characteristic frequency. It is difficult to observe the peak of the characteristic frequency of the rolling element fault in the envelope spectrum of the component signal of the rolling element fault signal, and only the peak of the rotation frequency can be observed.

[0156] exist Figure 4 In the PESFDM characteristic diagram of the normal signal, the two white bright spots corresponding to the frequency conversion in the envelope spectrum of IMF2 are more obvious; the periodic strip-shaped white bright spots of the frequency conversion and characteristic frequency peaks in the PESFDM characteristic diagram of the inner race fault signal and the outer race fault signal are more significant; in the PESFDM characteristic diagram of the rolling element fault signal, the grayscale value of the grayscale image corresponding to IMF4 and IMF5 is large, and the white bright spots are concentrated in this area. The PESFDM envelope grayscale diagrams of the four signals each have characteristics that represent different health information of the bearing, which is conducive to CNN pattern recognition.

[0157] Parameter settings during network training: learning rate is 0.001, minimum batch size is set to 32, a total of 20 cycles, 440 iterations, verification is performed every 10 iterations, and the optimizer uses Adam.

[0158] In order to verify the effectiveness of the proposed method, the proposed method is compared with the traditional fault diagnosis method which uses the continuous wavelet transform (CWT) time-frequency diagram as the 2D-CNN input.

[0159] The signal length used by CWT is 1024. The size of the time-frequency diagram obtained by CWT with the default setting is too large. Directly inputting it into CNN will greatly increase the computational complexity of the network model. The resolution of the time-frequency diagram of CWT is set to the commonly used 224×224 pixels. In addition, CWT uses the default "morse" basis function. The signal length used for a single data sample is 1024, and the number of samples is consistent with the PESFDM-ESGI method.

[0160] The abscissa of the CWT time-frequency diagram is time, and the ordinate is frequency. The time-frequency diagrams input into CNN have all removed information that is irrelevant to fault diagnosis, such as the abscissa and ordinate. The time-frequency diagrams of normal signals and three minor fault signals are shown in the figure. The CWT time-frequency diagram reflects the time-frequency distribution of various signals in a more detailed manner, with more obvious characteristics, which will be beneficial to fault diagnosis.

[0161] The average calculation time for generating one feature map by the two methods is 81.26ms and 376.21ms respectively. The calculation time of PESFDM-ESGI proposed in the present invention is only about one fifth of that of CWT, which greatly saves the time for producing feature maps.

[0162] In order to simulate the bearing signals with different levels of noise contained in actual industrial production, Gaussian white noise was added to the data set, and a total of four data sets were obtained with vibration signals with signal-to-noise ratios of -4db, 0db, and 4db, plus the original signal. Five experiments were conducted for each method, and the average accuracy and standard deviation of the test set are shown in Table 2.

[0163] Table 2 Average accuracy of the two methods under different signal-to-noise ratios

[0164]

[0165] Without adding noise, the average accuracy of the two methods is 100%. For data with three different noise backgrounds, the method proposed in the present invention has strong noise resistance and still achieved an accuracy of 96.5% at a low signal-to-noise ratio of -4db, which is significantly better than the comparison method. For the -4db bearing signal data set, the confusion matrix of the test sets of the two methods in each test result is as follows: Figure 6 As shown in Figure 3, only a few samples are misclassified by the proposed method, while a large number of samples are misclassified by the CWT-based method.

[0166] Next, we analyze the training time of the two methods. The training process of the two methods is as follows: Figure 5 ,The training time of the network model is shown in Table 3.

[0167] Table 3 Training time of two methods

[0168]

[0169] The proposed method is beneficial to feature learning of CNN because of its obvious fault characteristics. Therefore, the two methods require fewer iterations to achieve the highest verification accuracy. The number of iterations required for CWT network verification accuracy to no longer decrease is more than that of PESFDM-ESGI. Due to the large image size, the time required for a single cycle is also more than that of PESFDM-ESGI. Combining the time required for the two methods to complete an iteration and the number of iterations required when the verification accuracy no longer decreases, the PESFDM-ESGI method proposed in the present invention has a fast training speed and requires a short training time, which has significant advantages over the comparative methods.

[0170] The above is only a preferred implementation of a rolling bearing fault diagnosis method. The protection scope of a rolling bearing fault diagnosis method is not limited to the above embodiment. All technical solutions under this idea belong to the protection scope of the present invention. It should be pointed out that for those skilled in the art, several improvements and changes without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A rolling bearing fault diagnosis method, characterized in that: The method is based on peak envelope spectrum Fourier decomposition envelope spectrum grayscale image and convolutional neural network. The steps include: Step 1: Collect vibration signals of rolling bearings in normal state, different fault types and different damage degrees; Step 2: Segment the collected vibration signals of the rolling bearing to form labeled data samples, and divide the data samples into a training set and a test set; Step 3: Decompose the data sample x(k) using the peak envelope spectrum Fourier decomposition to obtain a series of component signals {x l (k)}; Step 4: Perform envelope spectrum analysis on all component signals to obtain the component envelope spectrum ES l (n); Step 5: Convert the envelope spectra of all components into data-enhanced grayscale images; Step 6: Design a 2D-CNN for rolling bearing fault feature learning; Step 7: Input the training set into the 2D-CNN to train the network, save the trained network model, and input the test set data into the trained 2D-CNN to complete fault diagnosis.

2. The method according to claim 1, characterized in that: Step 3: Decompose the signal using PESFDM, where the number of PESFDM components is selected within the range of 6-10.

3. The method according to claim 2, characterized in that: The specific steps of step 4 are: For signal x l (k) Perform Hilbert transform to get x l '(k),x l The analytical signal of (k) is q l (k), as shown in formula (1), x l (k) Envelope signal |q l (k)|As shown in formula (2): q l (k)=x l (k)+i·x l '(k) k=0,2,3···N-1 (1) Yes | l (k) | Perform FFT to obtain the complex Q of the envelope signal l (n), as shown in equation (3), where the FFT transform length is the signal length, |Q l (n)| is twice the first half, which is x l The envelope spectrum of (k) is expressed as ES l (n) is expressed as shown in formula (4):

4. The method according to claim 3, characterized in that: The specific steps of step 5 are: Step 5.1: The one-dimensional envelope spectrum amplitude sequence of the selected component is converted into a grayscale image in the following manner: in order to enhance the fault characteristics in the grayscale image, the amplitude sequence of the first 1 / 10 part of each component signal is first formed into a matrix of 3 rows and N / 20 columns, that is, the selected envelope spectrum amplitude sequence is repeatedly arranged into three rows; Step 5.2: Combine the selected envelope spectra matrices of all components into a matrix of 3M rows and N / 20 columns in the order of component numbers. This matrix is ​​the grayscale image matrix used by the grayscale image. The individual elements in the grayscale image matrix are numbered from left to right and from top to bottom. There are 3×M×N / 20 elements in total. Then, the grayscale image matrix is ​​normalized to 0-255 using formula (5): Where A represents all elements of the matrix, A i is the i-th element in the matrix, min(A) and max(A) represent the minimum and maximum values ​​in A respectively, and A i ' is the normalized element value. The value of each element after normalization is the grayscale value of a color block in the grayscale image. Finally, according to the rule that grayscale from 0 to 255 corresponds to grayscale from dark to bright, the matrix is ​​converted into a grayscale image.

5. The method according to claim 4, characterized in that: A 2D-CNN with two convolution-pooling pairs and two fully connected layers is designed. To speed up network training, batch normalization layers are added between the convolution layer and the activation function and between the fully connected layer and the activation function. The random dropout ratio adopts the recommended value: 0.

5.

6. The method according to claim 5, characterized in that: The moving step of the convolution kernel is fixed to 1. The number of convolution kernels in the first convolution layer is 16, and the second layer is 32. The number of convolution kernels increases exponentially to improve the feature extraction capability. Both convolutional layers use a small receptive field of size 3×3 to capture detail information, and the padding method is "same". When the convolution kernel movement step is 1, the side length of the output feature map after the convolution operation remains unchanged, which helps to retain information about the input.

7. The method according to claim 6, characterized in that: The pooling layer uses a 2×2 pooling kernel used by most models, with a moving step of 2 and no padding. After the pooling operation, the length and width of the feature map are reduced by half, which improves the calculation speed of the model. After the convolution and pooling pairs, there are two fully connected layers. The number of neurons in the first fully connected layer is 128, and the number of neurons in the second fully connected layer is the number of classification patterns.

8. A rolling bearing fault diagnosis system, characterized by: The system comprises: A collection module, which collects vibration signals of the rolling bearing in normal state, different fault types and different damage degrees; A segmentation module, wherein the segmentation module segments the collected vibration signals of the rolling bearing to form labeled data samples, and divides the data samples into a training set and a test set; A decomposition module is used to decompose the data sample x(k) using the peak envelope spectrum Fourier decomposition to obtain a series of component signals {x l (k)}; An analysis module performs envelope spectrum analysis on all component signals to obtain the component envelope spectrum ES l (n); A conversion module, wherein the conversion module converts the envelope spectra of all components into a data enhanced grayscale image; A design module, wherein the design module designs a 2D-CNN for rolling bearing fault feature learning; The fault diagnosis module inputs the training set into the 2D-CNN to train the network, saves the trained network model, and inputs the test set data into the trained 2D-CNN to complete the fault diagnosis.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the method according to claims 1-7.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method of claims 1-7 is implemented.

Images