Rolling bearing fault diagnosis method based on multi-channel feature fusion image

Abstract

The invention discloses a rolling bearing fault diagnosis method based on a multi-channel feature fusion image, and the method comprises a fault diagnosis model training step and a fault diagnosis step, and the fault diagnosis model training step comprises the steps: collecting a multi-channel vibration signal of a rolling bearing; performing feature mode decomposition on the vibration signals of each channel, and selecting a part of feature mode components as target feature mode components; converting the target feature modal component to be displayed and output under the same polar coordinate system to obtain a multi-channel feature fusion image; constructing a fault diagnosis model; the fault diagnosis step comprises the step of inputting the multi-channel feature fusion image of the rolling bearing into the fault diagnosis model. According to the rolling bearing fault diagnosis method, the data processing amount can be reduced. A plurality of one-dimensional time sequences are converted into a single two-dimensional feature image to be presented, the complex relation between the overall image structure and key features is learned, and automation of rolling bearing fault diagnosis is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of rolling bearing fault diagnosis, and in particular relates to a rolling bearing fault diagnosis method based on multi-channel feature fusion images. Background Art

[0002] Rotating equipment is a critical component of industrial production. Failure can halt production and even lead to serious accidents. According to incomplete statistics, approximately 30% of rotating machinery failures are caused by rolling bearings. Research on condition monitoring and fault diagnosis of rolling bearings can effectively reduce the operating costs of rotating machinery and ensure the long-term safety and stability of rotating machinery.

[0003] Because rolling bearings operate in harsh and volatile environments for long periods of time, and under conditions of dynamic changes in speed and load, the vibration signals of rolling bearings in industrial applications often contain interference from vibrations and noise from other components, making it difficult to directly diagnose rolling bearing faults. Signal decomposition methods can decompose signals into multiple sub-signal components with clear physical meanings. These sub-signal components contain rolling bearing fault characteristics that are easier to extract. Variational mode decomposition establishes a constrained optimization problem using component narrowband conditions to estimate the center frequency of the signal components and reconstruct the components. Eigenmode decomposition considers the impact and periodicity of the equipment fault response, adaptively adjusting the filter center frequency and bandwidth to obtain a series of eigenmode components. However, each eigenmode component only contains the characteristics of a specific rolling bearing fault. Therefore, fault diagnosis requires checking each eigenmode component one by one, which is time-consuming and labor-intensive.

[0004] In addition, due to the influence of working conditions and noise, the fault information of one-dimensional signals may be reduced, blurred, and shifted, resulting in reduced fault diagnosis accuracy.

[0005] With the development of computer technology, shallow machine learning algorithms such as support vector machines and artificial neural networks have been gradually applied to fault diagnosis, but the extracted features have weak generalization capabilities. Deep learning, with its powerful feature extraction capabilities, can learn efficient feature representations from input. Recurrent neural networks (RNNs) are used to extract time series features, but their accuracy is easily affected by noise. Using the two-dimensional feature image of the original signal as the neural network input, a convolutional neural network (CNN) extracts fault information from the image, which can extract more complex features. However, CNNs lack global information and their convolution kernel weights cannot change dynamically. Therefore, a fault diagnosis model with strong global information capture and generalization capabilities is needed. Summary of the Invention

[0006] To solve the technical problems that when diagnosing the faults of rolling bearings of existing rotating equipment, it is necessary to check one by one according to each characteristic mode component, which is time-consuming and laborious and the accuracy of single-channel signal fault diagnosis is low, the present invention proposes a rolling bearing fault diagnosis method based on multi-channel feature fusion images, which can solve the above problems.

[0007] To solve the above technical problems, the present invention is implemented by adopting the following technical solutions:

[0008] A rolling bearing fault diagnosis method based on multi-channel feature fusion images includes a fault diagnosis model training step and a fault diagnosis step. The fault diagnosis model training step includes:

[0009] Step 1, collect multi-channel vibration signals of the rolling bearing, preprocess the multi-channel vibration signals to obtain multi-channel vibration signals x(n), where n represents the number of sampling points;

[0010] Step 2, set the parameters of the empirical mode decomposition, perform empirical mode decomposition on the vibration signals x(n) of each channel respectively to obtain a plurality of intrinsic mode functions, and select a part of the intrinsic mode functions as the target intrinsic mode functions;

[0011] Step 3, convert the target intrinsic mode functions to be displayed and output in the same polar coordinate system to obtain multi-channel feature fusion images, gather the multi-channel feature fusion images of rolling bearings of different fault types together to construct a rolling bearing fault data set, preprocess the multi-channel feature fusion images, and divide the rolling bearing fault data set into a training set and a test set according to a ratio;

[0012] Step 4, construct a fault diagnosis model, and use the training set to train the model. When the loss function converges, obtain the fault diagnosis model;

[0013] The fault diagnosis step includes:

[0014] Input the multi-channel feature fusion image of the rolling bearing into the fault diagnosis model, and output the rolling bearing fault diagnosis result.

[0015] In some embodiments, in Step 1, the method for preprocessing the multi-channel vibration signals includes:

[0016] Perform high-frequency sampling on the multi-channel vibration signals, and the sampling frequency of the high-frequency sampling is at least 2 times the highest frequency of the rolling bearing vibration signals;

[0017] Align the multi-channel vibration signals after high-frequency sampling on the time axis, and intercept the signals of a set duration to obtain multi-channel vibration signals x(n).

[0018] In some embodiments, the characteristic mode decomposition parameters set in step two include the number of modes N, the number of filters K, the maximum number of iterations M, and the filter length L.

[0019] In some embodiments, in step two, the method for obtaining the target characteristic mode components includes:

[0020] Perform characteristic mode decomposition on the vibration signals of each channel respectively, and multiple characteristic mode components are obtained by decomposing the vibration signal of each channel;

[0021] Calculate the correlation kurtosis of each characteristic mode component of each channel, and sort the correlation kurtosis in descending order, and select the characteristic mode components corresponding to the first A correlation kurtosis as the target characteristic mode components.

[0022] In some embodiments, the calculation method of the correlation kurtosis is:

[0023]

[0024] where y n represents the n-point vibration value of the mode component y, and y n-T represents the n-T point vibration value of the mode component y, T represents the fault period, and N represents the length of the mode component y.

[0025] In some embodiments, the method for obtaining the fault period T includes:

[0026] Establish the autocorrelation function R x (τ) of the mode component y:

[0027]

[0028] where τ represents the time delay value for collecting the multi-channel vibration signals of the rolling bearing, and y n+τ represents the n+τ point vibration value of the mode component y;

[0029] Calculate the local maximum value of R x (τ) as the fault period T.

[0030] In some embodiments,

[0031] The calculation methods of the polar radius r(ij) and the polar angles θ(ij), β(ij), α(ij) of the multi-channel feature fusion image are:

[0032]

[0033] In the formula, i represents the i-th characteristic mode component; represents the j-point vibration value of the i-th mode component; x max represents the maximum value of the vibration mode component; xmin represents the minimum value of the vibration mode component; a, b, and c represent offset coefficients; η represents the amplification coefficient; m represents the number of target characteristic mode components; represents the vibration value at the j + a point of the i-th mode component.

[0034] In some embodiments, in step three, the method for preprocessing the multi-channel feature fusion image includes:

[0035] Set the resolution of the multi-channel feature fusion image to m 2 ×m 2 ;

[0036] Convert the multi-channel feature fusion image into a grayscale image.

[0037] In some embodiments, in step four, when constructing the fault diagnosis model, the grayscale image of the multi-channel feature fusion image is divided into m 2 image blocks of size m×m, and each image block is mapped to a vector of m 2 .

[0038] In some embodiments, the fault diagnosis model is constructed based on the self-attention mechanism.

[0039] Compared with the prior art, the advantages and positive effects of the present invention are as follows: The rolling bearing fault diagnosis method based on multi-channel feature fusion images of the present invention can reduce the amount of data processing by selecting a part of the characteristic mode components as the target characteristic mode components. By converting the target characteristic mode components to be displayed and output in the same polar coordinate system, a multi-channel feature fusion image is obtained, realizing the conversion of multiple one-dimensional time series into a single two-dimensional feature image for presentation, reflecting the global information of the rolling bearing fault state from the visual information of the feature image, learning the complex relationship between the overall image structure and the key features, and training the fault diagnosis model according to the patterns presented by the multi-channel feature fusion images and the corresponding faults, realizing the automation of rolling bearing fault diagnosis.

[0040] After reading the detailed description of the embodiments of the present invention in conjunction with the accompanying drawings, other features and advantages of the present invention will become clearer. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a flowchart of the fault diagnosis model training steps in an embodiment of the rolling bearing fault diagnosis method based on multi-channel feature fusion images proposed by the present invention;

[0042] Figure 2 is a schematic structural diagram of a mechanical performance and fault comprehensive simulation test bench in an embodiment of the rolling bearing fault diagnosis method based on multi-channel feature fusion images proposed by the present invention;

[0043] Figure 3a It is the vibration signal in the x - direction collected in an embodiment of the rolling bearing fault diagnosis method based on multi - channel feature - fused images proposed by the present invention;

[0044] Figure 3b It is the vibration signal in the y - direction collected in an embodiment of the rolling bearing fault diagnosis method based on multi - channel feature - fused images proposed by the present invention;

[0045] Figure 3c It is the vibration signal in the z - direction collected in an embodiment of the rolling bearing fault diagnosis method based on multi - channel feature - fused images proposed by the present invention;

[0046] Figure 4a It is one of the decomposed characteristic mode component diagrams in an embodiment of the rolling bearing fault diagnosis method based on multi - channel feature - fused images proposed by the present invention;

[0047] Figure 4b It is another decomposed characteristic mode component diagram in an embodiment of the rolling bearing fault diagnosis method based on multi - channel feature - fused images proposed by the present invention;

[0048] Figure 4c It is yet another decomposed characteristic mode component diagram in an embodiment of the rolling bearing fault diagnosis method based on multi - channel feature - fused images proposed by the present invention;

[0049] Figure 5 It is the grayscale image of the multi - channel feature - fused image in an embodiment of the rolling bearing fault diagnosis method based on multi - channel feature - fused images proposed by the present invention;

[0050] Figure 6 It is the comparison diagram between the multi - channel feature - fused images of different rolling bearing fault types and the multi - channel feature - fused image without fault in an embodiment of the rolling bearing fault diagnosis method based on multi - channel feature - fused images proposed by the present invention;

[0051] Figure 7 It is the structural schematic diagram of the fault diagnosis model in an embodiment of the rolling bearing fault diagnosis method based on multi - channel feature - fused images proposed by the present invention;

[0052] Figure 8 It is the iteration - loss diagram of the fault diagnosis model in an embodiment of the rolling bearing fault diagnosis method based on multi - channel feature - fused images proposed by the present invention;

[0053] Figure 9 It is the iteration - accuracy diagram of the fault diagnosis model in an embodiment of the rolling bearing fault diagnosis method based on multi - channel feature - fused images proposed by the present invention. Detailed Embodiments

[0054] The following further describes in detail the specific embodiments of the present invention in conjunction with the accompanying drawings.

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0056] It should be noted that in the description of the present invention, the terms indicating directions or positional relationships such as "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be construed as a limitation of the present invention. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance. In the description of the present invention, "a plurality" means two or more unless otherwise specifically defined.

[0057] In the present invention, unless otherwise clearly specified and limited, the terms "installation", "connection", "connection", "fixation", etc. should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two elements or the interaction relationship between two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0058] Embodiment 1, as shown in Figure 1 A rolling bearing fault diagnosis method based on multi-channel feature fusion images is proposed in this embodiment, including a fault diagnosis model training step and a fault diagnosis step. The fault diagnosis model training step includes:

[0059] Step 1: Collect multi-channel vibration signals of rolling bearings with different fault types through sensors, preprocess the multi-channel vibration signals, and obtain multi-channel vibration signals x(n), where n represents the number of sampling points.

[0060] Step 2: Set the parameters of the empirical mode decomposition, perform empirical mode decomposition on the vibration signals x(n) of each channel respectively, obtain a plurality of intrinsic mode functions, and select a part of the intrinsic mode functions as the target intrinsic mode functions.

[0061] Step 3: Convert the target feature modal components to the same polar coordinate system for display output to obtain a multi-channel feature fusion image. Combine the multi-channel feature fusion images of rolling bearings with different fault types to construct a rolling bearing fault dataset. Preprocess the multi-channel feature fusion images, and divide the rolling bearing fault dataset into a training set and a test set according to a ratio.

[0062] Step 4: Construct a fault diagnosis model and use the training set to train the model. After the loss function converges, obtain the fault diagnosis model.

[0063] The fault diagnosis steps include:

[0064] Input the multi-channel feature fusion image of the rolling bearing into the fault diagnosis model and output the rolling bearing fault diagnosis result.

[0065] In the fault diagnosis steps, it can be to preprocess the multi-channel vibration signals of the rolling bearing faults collected in real time, perform feature modal decomposition in step 2, obtain the multi-channel feature fusion image according to step 3, and then input the multi-channel feature fusion image into the fault diagnosis model for diagnosis and output the diagnosis result. It can also be the historical acquisition data of the multi-channel vibration signals of the rolling bearing faults. After being processed by the above steps, the multi-channel feature fusion image is obtained and input into the fault diagnosis model for diagnosis and output the diagnosis result.

[0066] For the rolling bearing fault diagnosis method based on multi-channel feature fusion images in this embodiment, by selecting a part of the feature modal components as the target feature modal components, the data processing amount can be reduced. By converting the target feature modal components to the same polar coordinate system for display output, a multi-channel feature fusion image is obtained, realizing the conversion of multiple one-dimensional time series into a single two-dimensional feature image for presentation, reflecting the global information of the rolling bearing fault state from the visual information of the feature image, learning the complex relationship between the overall image structure and key features, and training the fault diagnosis model according to the patterns presented by the multi-channel feature fusion images and the corresponding faults, realizing the automation of rolling bearing fault diagnosis.

[0067] By constructing a polar coordinate mapping, convert the multi-channel one-dimensional vibration signals into two-dimensional feature fusion images, establish cross-channel feature collaborative analysis while retaining the independent features of each channel, and at the same time the differences in the images can accurately reflect the state of the rolling bearing, enhancing the overall expression of different fault states of the rolling bearing.

[0068] In some embodiments, in step 1, the method for preprocessing the multi-channel vibration signals includes:

[0069] Perform high-frequency sampling on the multi-channel vibration signals. The sampling frequency of the high-frequency sampling is at least twice the highest frequency of the rolling bearing vibration signals. High-frequency sampling can completely record the high-frequency transient impact signals generated by the rolling bearing faults, avoiding signal aliasing or loss caused by insufficient sampling rate. In this embodiment, by setting the high-frequency sampling rate to be at least twice the highest frequency of the rolling bearing vibration signals, the multi-channel vibration signals are aligned and intercepted based on the sampling frequency to ensure the consistency of the fault-related features of the vibration signals in different channels in the time dimension.

[0070] Align the multi-channel vibration signals after high-frequency sampling on the time axis, and intercept the signals with a set duration to obtain the multi-channel vibration signals x(n).

[0071] In some embodiments, the method for preprocessing the multi-channel vibration signals further includes, but is not limited to, denoising processing.

[0072] In some embodiments, the parameters of the empirical mode decomposition set in step two include the number of modes N, the number of filters K, the maximum number of iterations M, and the filter length L.

[0073] In some embodiments, in step two, the method for obtaining the target characteristic mode components includes:

[0074] Perform empirical mode decomposition on the vibration signals of each channel respectively. The vibration signals of each channel are decomposed into multiple characteristic mode components.

[0075] Calculate the correlation kurtosis of each characteristic mode component of each channel, and sort the correlation kurtosis in descending order. Select the first A characteristic mode components corresponding to the correlation kurtosis as the target characteristic mode components. It can be understood that A should be a positive integer and A is not greater than the total number of all characteristic mode components.

[0076] The correlation kurtosis can reflect the periodicity and impulsiveness of the vibration signals, and the rolling bearing faults also have periodicity and impulsiveness. In this solution, by selecting the characteristic mode components with larger correlation kurtosis as the target characteristic mode components, the characteristic mode components only related to the faults can be screened out for processing, which can greatly reduce the calculation amount, improve the fault diagnosis speed, and realize real-time fault diagnosis.

[0077] In some embodiments, the calculation method of the correlation kurtosis CK is:

[0078]

[0079] where y n represents the n-point vibration value of the mode component y, and y n-T represents the n-T point vibration value of the mode component y, T represents the fault period, and N represents the length of the mode component y.

[0080] In some embodiments, the method for obtaining the fault cycle T includes:

[0081] Establish the autocorrelation function R x (τ) of the modal component y:

[0082]

[0083] where τ represents the time delay value for collecting the multi-channel vibration signals of the rolling bearing, and y n+τ represents the vibration value at the n + τ point of the modal component y.

[0084] Calculate the local maximum value of R x (τ) as the fault cycle T.

[0085] In some embodiments, before the step of calculating the correlation kurtosis of each characteristic modal component of each channel, it further includes a step of processing the characteristic modal components, including:

[0086] Align all the characteristic modal components in time.

[0087] Stitch the characteristic modal components of each channel according to the interception order of the multi-channel vibration signal x(n).

[0088] In some embodiments, in step three, the method for obtaining the multi-channel feature fusion image includes:

[0089] Convert the time series coordinates of the characteristic modal components to polar coordinates, with the center of the plane as the pole, draw a ray from the pole as the polar axis, the unit of the radial coordinate of the polar coordinates is 1, the unit of the angular coordinate of the polar angle is 1°, and the counterclockwise direction is the positive direction of the angle. Calculate the radial radius r(ij), polar angles θ(ij), β(ij), α(ij), and use (r(ij), θ(ij)), (r(ij), β(ij)), (r(ij), α(ij)) as polar coordinate points to form a multi-channel feature fusion image. The image can effectively characterize the state differences generated by different fault types through the radial texture features and the angular pole distribution features, realizing the mapping between the fault mode and the visual features of the image. r(ij), θ(ij), β(ij), α(ij) are specifically as follows:

[0090]

[0091]

[0092] ]>

[0093] In the formula, i represents the i-th characteristic modal component; represents the vibration value at the j-th point of the i-th modal component; x maxrepresents the maximum value of the vibration mode component; x min represents the minimum value of the vibration mode component; a, b, and c represent offset coefficients; η represents the amplification coefficient; m represents the number of target characteristic mode components; represents the vibration value at the j + a point of the i-th mode component.

[0094] During the calculation of the polar radius, data normalization is adopted, which reduces the influence of working conditions, avoids the excessive influence of certain characteristics on the results, and ignores the contributions of other characteristics

[0095] In some embodiments, in step three, the method for preprocessing the multi-channel feature fusion image includes:

[0096] Set the resolution of the multi-channel feature fusion image to m 2 ×m 2 .

[0097] Convert the multi-channel feature fusion image into a grayscale image to reduce redundant color information, which can effectively reduce the input data volume, reduce the computational cost and memory consumption of training, and at the same time make subsequent diagnosis more focused on the structure and texture features of the image, improving the running efficiency of the model.

[0098] In some embodiments, in step four, in constructing the fault diagnosis model, the grayscale image of the multi-channel feature fusion image is divided into m×m sizes to obtain m 2 image blocks, and each image block is mapped to a vector of m 2 . In this embodiment, by setting the resolution of the multi-channel feature fusion image to m 2 ×m 2 , it is convenient to segment the image.

[0099] To facilitate the operation of subsequent methods, the image is preprocessed. By dividing the grayscale image of the multi-channel feature fusion image into several small image blocks, the resolution of each image block becomes smaller and more compact compared to the original image, which can reduce the computational complexity and improve the computational speed in the training of the fault diagnosis model.

[0100] In some embodiments, the fault diagnosis model is constructed based on the self-attention mechanism. Construct a fault diagnosis model, which includes multiple encoders. The encoders capture the global information of the image through the self-attention mechanism, which is achieved by defining three learnable weight matrices query Q, key K, and value V. The output calculation of the self-attention layer is shown in the following formula:

[0101]

[0102] where Attention(Q, K, V) represents the output of the self-attention layer; softmax represents the function that converts to a probability distribution; QKT represents the dot product between the query and all keys; is the normalization factor.

[0103] The multi-channel feature fusion image is subjected to feature extraction through a fault diagnosis model based on the self-attention mechanism for fault diagnosis of rolling bearings; the self-attention mechanism is used to capture the global features in the image, learn the complex relationship between the overall image structure and key features, and realize the automation of rolling bearing fault diagnosis. In industrial applications, due to the existence of a large amount of monitoring data, after pre-training and fine-tuning, the model has good generalization ability and accuracy.

[0104] Example 2. In a specific embodiment, as Figure 2 shown, taking a certain fluid machinery performance and fault comprehensive simulation test bench as an example, the test bench includes a motor 1, an optoelectronic sensor 2, a coupling 3, a vibration acceleration sensor 4, a bearing housing 5, and a gearbox 6. The optoelectronic sensor 2 is installed on the front side of the motor 1 to measure the movement of the inner ring of the bearing. The motor 1 drives the movement of the inner ring of the bearing through the coupling 3 to drive the rotor. The outer ring of the bearing is fixed on the bearing housing 5, and the vibration acceleration sensor 4 is installed at the bearing housing 5 to measure the vibration signal of the bearing. The gearbox 6 is installed at the tail of the rotor and is connected through a coupling. The experiment mainly includes various conditions such as faults of the outer ring of the rolling bearing, faults of the inner ring, and health. In a specific example, the collected fault vibration signals are subjected to high-frequency sampling, truncation, and denoising processing, where the sampling frequency is 51200 Hz, and the data with 51200 sampling points in the entire signal segment is intercepted. As Figure 3a - Figure 3c shown, they are the vibration signals collected in the x, y, and z directions in space respectively.

[0105] The original fault signal is a multi-channel one-dimensional time-domain signal. Due to the interference of noise and other components, the various fault characteristics of the rolling bearing are hidden in the time-domain signal, making it difficult to directly perform rolling bearing fault diagnosis. Combining the impulsiveness and periodicity of rolling bearing faults, the multi-channel one-dimensional time-domain signal is decomposed into multi-channel feature mode components through the method of feature mode decomposition, and the various characteristics of the faults are separated to form multiple feature mode components, thereby judging the bearing fault type.

[0106] In step S2, the set parameters include the number of modes N, the number of filters K, the filter length L, and the maximum number of iterations. In a specific example, the number of modes is 1, the number of filters is 10, the filter length is 30, and the maximum number of iterations is 30.

[0107] According to the method described in Step 2 to Step 3 of Embodiment 1, the multi-channel vibration signal is decomposed into multi-channel characteristic mode components through characteristic mode decomposition, and Figure 4a - Figure 4c the multi-channel characteristic mode components shown in Figure 4a - Figure 4c are transformed into multi-channel feature fusion images, and the multi-channel feature fusion images are converted into grayscale images, as shown in Figure 5 . The image can effectively characterize the state differences generated by different fault types through the radial texture feature and the angular pole distribution feature, realizing the mapping between the fault mode and the visual features of the image. The multi-channel feature fusion image fully contains the global features and overall information of the vibration signal, enhancing the visualization of the fault features.

[0108] As shown in Figure 6 , the figure shows that the multi-channel feature fusion images formed by different fault types have large differences, can accurately reflect the states of different rolling bearing fault types, and are convenient for subsequent fault diagnosis, indicating the effectiveness of the multi-channel feature fusion image for rolling bearing fault diagnosis.

[0109] The self-attention mechanism can focus on and capture the global features of the image without a convolutional layer. At the same time, in industrial applications, due to the complexity of the equipment environment and the massive monitoring data, the deep learning fault diagnosis model has better generalization ability and accuracy. The structure of the fault diagnosis model is shown in Figure 7 . The image is segmented into a series of image blocks, and each image block is converted into a low-dimensional vector to meet the requirements of the input data.

[0110] Set the hyperparameters of the fault diagnosis model, as shown in Table 1.

[0111] [[ID=**********]]Table 1 Model Hyperparameter Settings

[0112]

[0113] Train the initial model with the rolling bearing training set. When it is judged whether the number of iterations reaches the set value, if satisfied, output the model. In a specific embodiment provided by the present invention, after processing the rolling bearing vibration data, input it into the rolling bearing fault diagnosis model to classify the rolling bearing and give the corresponding diagnosis results. The iterative loss and accuracy of the model are shown in Figure 8 and Figure 9 . The accuracy of the rolling bearing of the proposed method reaches 97%. The results show that the method proposed in this example can efficiently and accurately complete the rolling bearing fault diagnosis function and can effectively ensure the efficient operation of the rolling bearing.

[0114] A rolling bearing fault diagnosis method based on a multi-channel feature fusion image proposed by the present invention decomposes the rolling bearing vibration signal by using a feature mode decomposition method. The obtained multi-channel feature mode components contain different rolling bearing fault information. All the mode components are fused through a multi-channel feature fusion image. At the same time, due to normalization, the influence of working conditions is reduced, and the impact and periodicity of rolling bearing faults are more reflected. A model is used to extract the global features and local features of the multi-channel feature fusion image to realize the rolling bearing fault diagnosis function.

[0115] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by those of ordinary skill in the art within the scope of the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A rolling bearing fault diagnosis method based on multi-channel feature fusion images, characterized in that It includes a fault diagnosis model training step and a fault diagnosis step. The fault diagnosis model training step includes: Step 1: Collect multi-channel vibration signals of a rolling bearing, preprocess the multi-channel vibration signals to obtain multi-channel vibration signals x(n), where n represents the number of sampling points; Step 2: Set the parameters of empirical mode decomposition, perform empirical mode decomposition on the vibration signals x(n) of each channel respectively to obtain multiple intrinsic mode functions, and select a part of the intrinsic mode functions as the target intrinsic mode functions; Step 3: Convert the target intrinsic mode functions to be displayed and output in the same polar coordinate system to obtain a multi-channel feature fusion image. Assemble the multi-channel feature fusion images of rolling bearings with different fault types together to construct a rolling bearing fault data set. Preprocess the multi-channel feature fusion image, and divide the rolling bearing fault data set into a training set and a test set according to a ratio; Step 4: Construct a fault diagnosis model, and use the training set to train the model. After the loss function converges, obtain the fault diagnosis model; The fault diagnosis step includes: Input the multi-channel feature fusion image of the rolling bearing into the fault diagnosis model, and output the rolling bearing fault diagnosis result.

2. The rolling bearing fault diagnosis method according to claim 1, characterized in that In Step 1, the method for preprocessing the multi-channel vibration signals includes: Perform high-frequency sampling on the multi-channel vibration signals, and the sampling frequency of the high-frequency sampling is at least 2 times the highest frequency of the rolling bearing vibration signals; Align the multi-channel vibration signals after high-frequency sampling on the time axis, and intercept a signal with a set duration to obtain multi-channel vibration signals x(n).

3. The rolling bearing fault diagnosis method according to claim 1, wherein, The parameters of empirical mode decomposition set in Step 2 include the number of modes N, the number of filters K, the maximum number of iterations M, and the filter length L.

4. The rolling bearing fault diagnosis method according to claim 2, wherein, In Step 2, the method for obtaining the target intrinsic mode functions includes: Perform empirical mode decomposition on the vibration signals of each channel respectively, and each channel's vibration signal is decomposed into multiple intrinsic mode functions; Calculate the correlation kurtosis of each intrinsic mode function of each channel, sort the correlation kurtosis in descending order, and select the intrinsic mode functions corresponding to the first A correlation kurtosis as the target intrinsic mode functions.

5. The rolling bearing fault diagnosis method according to claim 4, wherein, The calculation method of the correlation kurtosis is: where y n represents the vibration value of the n points of the modal component y, and y n-T represents the vibration value of the n-T points of the modal component y, T represents the fault period, and N represents the length of the modal component y.

6. The rolling bearing fault diagnosis method according to claim 5, characterized in that, The method for obtaining the fault period T includes: Establish the autocorrelation function \(R\) of the modal component \(y\) x (\(\tau\)): where τ represents the time delay value for collecting multi-channel vibration signals of the rolling bearing, and y n+τ represents the vibration value at the n+τ point of the modal component y; Calculate R x The local maximum of (τ) is used as the fault period T.

7. The rolling bearing fault diagnosis method according to claim 1, characterized in that In Step 3, the calculation methods of the polar radius r(ij) and the polar angles θ(ij), β(ij), α(ij) of the multi-channel feature fusion image are: In the formula, i represents the i-th characteristic mode component; represents the vibration value at the j-th point of the i-th mode component; x max represents the maximum value of the vibration mode component; x min represents the minimum value of the vibration mode component; a, b, and c represent offset coefficients; η represents a magnification factor; m represents the number of target characteristic mode components; represents the vibration value at the j + a-th point of the i-th mode component.

8. The rolling bearing fault diagnosis method according to claim 1, characterized in that, In Step 3, the method for preprocessing the multi-channel feature fusion image includes: Set the resolution of the multi-channel feature fusion image to m 2 ×m 2 ; Convert the multi-channel feature fusion image into a grayscale image.

9. The rolling bearing fault diagnosis method according to claim 8, characterized in that In step four, when constructing the fault diagnosis model, the grayscale image of the multi-channel feature fusion image is divided into blocks of size m×m to obtain m 2 image blocks, and each image block is mapped to a vector of m 2 .

10. The rolling bearing fault diagnosis method according to any one of claims 1-9, characterized in that, The fault diagnosis model is constructed based on the self-attention mechanism.

Images