An ultra-lightweight device bearing fault diagnosis method and system

Abstract

The application discloses a kind of ultra-light equipment bearing fault diagnosis method and system, belong to the field of fault diagnosis.The method comprises: the vibration signal in corresponding direction of equipment is collected by at least three vibration sensors installed in different positions of equipment but in the same direction;The vibration signal is visualized as gray scale image;Three vibration sensors are a group, and the gray scale image of each group of vibration sensors is fused into three-channel RGB image;Improved ultra-light network StarNet model is used to carry out bearing fault diagnosis to the RGB image.The correlation and specificity between the multi-sensor information are fully considered, the effective fusion of multi-sensor information is realized, and through the improved ultra-light network StarNet model, the feature mapping between levels is more comprehensive in the case of small parameter quantity, so as to improve the performance of the model, and then improve the bearing fault diagnosis accuracy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of fault diagnosis, and in particular, the present application relates to an ultra-lightweight device bearing fault diagnosis method and system. BACKGROUND

[0002] Bearing fault diagnosis is a very important task in the industrial field, as bearings are critical components in many mechanical devices. In existing technologies, vibration analysis is often used for bearing fault diagnosis, that is, by monitoring the vibration characteristics of the bearing, the faults of the bearing can be detected, such as internal damage, wear and looseness of the bearing. Vibration analysis can be monitored by acceleration sensors or vibration sensors, and signal processing techniques are used to analyze the frequency spectrum and amplitude of the vibration signal.

[0003] For example, the comparative document (CN118329450A) discloses a bearing fault diagnosis method, system, device and storage medium, relating to the technical field of bearing fault diagnosis, the method flow is: obtaining original bearing vibration data to construct a bearing vibration sample set; constructing a bearing fault diagnosis model, using unlabeled data and labeled data in the bearing vibration sample set to pre-train the bearing fault diagnosis model unsupervisedly and fine-tune the bearing fault diagnosis model supervisedly; obtaining bearing vibration data to be classified, using the bearing fault diagnosis model fine-tuned supervisedly to perform end-to-end bearing fault diagnosis on the bearing vibration data to be classified to obtain bearing fault classification results. The present application uses time series modeling technology and masking auto-encoding technology to pre-train and fine-tune the unlabeled data and labeled data unsupervisedly and supervisedly, so that the bearing fault diagnosis model trained can perform end-to-end high-precision bearing fault classification on the bearing vibration data.

[0004] However, the above-mentioned comparative document, like most existing technologies, has the following shortcomings:

[0005] 1. The reliability and stability of fault diagnosis based on a single signal are poor, and the fault tolerance is not good. Once the sensor fails, the diagnosis will stop. When the fault types are multiple and complex, the diagnosis or prediction effect obtained from a single information or a few information is not reliable, and different faults may also manifest in the form of the same symptom.

[0006] 2. Multi-sensor information fusion can provide more abundant information, but there is a lack of effective and reliable fusion method in existing technologies, and effective information cannot be obtained from multi-sensor information fusion.

[0007] 3. In order to improve the performance of the network, the method of widening or deepening the network is often used, which increases the parameter amount of the whole network and increases the difficulty of model deployment. SUMMARY

[0008] The present application aims to overcome the deficiencies of the prior art, and provides an ultra-light equipment bearing fault diagnosis method and system, so as to fully consider the correlation and specificity between multi-sensor information, realize effective fusion of multi-sensor information, and improve the performance of the model by using an improved ultra-light network StarNet model, so that the feature mapping between levels is more comprehensive under the condition of small parameter quantity, thereby improving the bearing fault diagnosis accuracy.

[0009] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows: an ultra-light equipment bearing fault diagnosis method, the method comprising the following steps:

[0010] S1, collecting vibration signals in the corresponding direction of the equipment through at least three vibration sensors installed at different positions of the equipment but in the same direction;

[0011] S2, visualizing the vibration signals as a gray-scale image;

[0012] S3, three vibration sensors installed at different positions of the equipment but in the same direction form a group, and the gray-scale image of each group of vibration sensors is fused into a three-channel RGB image;

[0013] S4, using an improved ultra-light network StarNet model to perform bearing fault diagnosis on the RGB image.

[0014] Preferably, the equipment includes a fan end bearing, a motor, a drive end bearing, a speed reducer bearing, a load equipment, a base, and a coupling, and the vibration signals collected by the vibration sensors according to the different installation positions thereof include fan end bearing vibration signals, motor vibration signals, drive end bearing vibration signals, speed reducer bearing vibration signals, load equipment vibration signals, base vibration signals, and coupling vibration signals.

[0015] Preferably, in the step S2, each collected vibration signal is subjected to M 2 length cutting to generate N sample numbers, and M and N are both preset values; each sample is subjected to gray-scale image conversion so as to be converted into a gray-scale image with a size of 1×M×M, wherein the gray-scale value of each pixel point in the gray-scale image with a size of 1×M×M is represented as follows:

[0016]

[0017] wherein the values of i and j are both 1 to M; L i,j represents the amplitude of the pixel point (i, j) in the cut vibration signal; maxL i,j represents the maximum amplitude that the pixel point (i, j) can reach in the cut vibration signal; and minL i,jrepresents the minimum amplitude that pixel point (i, j) can reach in the intercepted vibration signal; L max represents the maximum pixel value supported by the grayscale image, which ranges between 255 and 0; L min represents the minimum pixel value supported by the grayscale image, which ranges between 0 and 255.

[0018] Preferably, in the step S3, the three-channel RGB image is obtained by channel fusion of the three vibration sensor grayscale images, wherein the grayscale images participating in the fusion have the same time domain.

[0019] Preferably, in the step S3, the channel fusion is performed by principal component analysis technology.

[0020] Preferably, the step S3 includes the following steps:

[0021] S31, respectively converting the three vibration sensor grayscale images into corresponding vector forms;

[0022] S32, merging the vectors representing the three grayscale images into a matrix Xraw, wherein the matrix Xraw is n rows and m columns, and the values of n and m are determined by the vectors representing the three grayscale images;

[0023] S33, transforming the matrix Xraw by principal component analysis technology, which is represented by the following formula:

[0024] X pca =W T X raw ={x ij};i=1,2,3;j=1,2,3…m;

[0025] Wherein, X pca ∈R 3×m represents the principal component matrix obtained by transformation by principal component analysis technology, R 3×m represents that the matrix R is 3 rows and m columns; W T ∈R 3×n represents the transformation matrix based on principal component analysis technology, T represents transposition, R 3×n represents that the matrix R is 3 rows and n columns.

[0026] S34, obtaining a three-channel RGB image according to the X pca matrix.

[0027] Preferably, the improved ultra-lightweight network StarNet model includes a convolutional layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, and a global average pooling layer / full connection layer connected in sequence, wherein the output of the convolutional layer is input to the first feature extraction layer after batch normalization and ReLU activation.

[0028] Preferably, all the feature extraction layers comprise a convolution layer and a StarBlocks layer connected in sequence, wherein the convolution layer is used for down-sampling, and the StarBlocks layer is used for feature extraction; the StarBlocks layer comprises a first point-wise convolution layer, a second point-wise convolution layer, a depth convolution layer and a full connection layer, wherein the initially input features are processed in sequence through the first point-wise convolution layer and the depth convolution layer; then the output of the depth convolution layer is respectively multiplied element by element after passing through two full connection layers, and the result of the element-by-element multiplication is output to the second point-wise convolution layer after passing through a full connection layer; finally, the features processed by the second point-wise convolution layer are output after being connected with the initially input features in a skip connection manner.

[0029] Preferably, in the step S4, the T-SNE algorithm is used to visualize the fault diagnosis result of the improved ultra-lightweight network StarNet model.

[0030] Meanwhile, the application also provides an ultra-lightweight equipment bearing fault diagnosis system according to the above method, which comprises at least three vibration sensors and a controller, wherein the vibration sensors are connected with the controller, the vibration sensors are installed at different positions of the equipment but in the same direction, and are used for collecting vibration signals of the equipment in the corresponding direction; and the controller is used for bearing fault diagnosis.

[0031] The technical effects of the application are as follows:

[0032] (1) The multiple sensors of one equipment are used simultaneously, different sensor information can be selected for fusion according to the different installation positions of the sensors when the sensor information is fused, and the correlation and specificity between the information are fully considered.

[0033] (2) The principal component analysis technology is used to realize effective fusion of the multiple sensor information, reduce the redundant information, improve the reliability of the fused image data, and thus improve the fault diagnosis accuracy.

[0034] (3) The improved ultra-lightweight network StarNet model is used, the element multiplication is introduced, the input is mapped to a high-dimensional nonlinear feature space without widening the network, the feature mapping between the levels is more comprehensive under the condition of small parameter quantity, and thus the performance of the model is improved. Further, based on the improvement of the ultra-lightweight network StarNet model, the inverted bottleneck structure, i.e. the point-wise convolution (PW) + depth convolution (DW) + point-wise convolution (PW) structure, is introduced in the ultra-lightweight network StarNet, the feature extraction capability of the light structure is deepened without increasing the network parameters. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 A flow chart of an ultra-lightweight equipment bearing fault diagnosis method is provided for an embodiment of the present application.

[0036] Figure 2 A device structure diagram is provided for an embodiment of the present application.

[0037] Figure 3 An improved ultra-lightweight network StarNet model structure diagram is provided for an embodiment of the present application.

[0038] Figure 4 A model training set and validation set accuracy comparison diagram is provided for an embodiment of the present application.

[0039] Figure 5 A model training set and validation set loss rate comparison diagram is provided for an embodiment of the present application.

[0040] Figure 6 A fault classification result diagram after T-SNE visualization is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0041] The specific embodiments of the present application will be further described below with reference to the accompanying drawings, and the description of the embodiments will help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solutions of the present application, and will be helpful for their implementation. It should be noted that the terms "first", "second" and the like in the present application are only for the convenience of describing the technical solutions to distinguish different components, and do not limit the present application. In order to make the technical solutions of the present application more clear, the present application is explained and described by the following embodiments.

[0042] An ultra-lightweight equipment bearing fault diagnosis method, as shown in Figure 1 the method comprises the following steps:

[0043] S1, collecting vibration signals in the corresponding direction of the equipment through at least three vibration sensors installed at different positions of the equipment but in the same direction;

[0044] S2, visualizing the vibration signals into a grayscale image;

[0045] S3, three vibration sensors installed at different positions of the equipment but in the same direction form a group, and the grayscale image of each group of vibration sensors is fused into a three-channel RGB image;

[0046] S4, using an improved ultra-lightweight network StarNet model to perform bearing fault diagnosis on the RGB image.

[0047] In this embodiment, the above method is applied to various equipment in the cement production process. As shown in Figure 2The device structure diagram provided by the embodiment is shown, wherein the device comprises a fan end bearing, a motor, a driving end bearing, a speed reducer bearing, a load device, a base, and a shaft coupling. In order to monitor the safety of various devices in the cement production process and simultaneously perform predictive maintenance on the devices, vibration sensors are installed at different positions on the devices to collect vibration data of the devices in real time, and the data is analyzed to predict the state of the devices. In the embodiment, according to the different installation positions of the vibration sensors, the vibration signals collected by the vibration sensors comprise fan end bearing vibration signals, motor vibration signals, driving end bearing vibration signals, speed reducer bearing vibration signals, load device vibration signals, base vibration signals, and shaft coupling vibration signals. The vibration sensors of the embodiment can be accelerometers, and the sampling frequency is 51.2 kHz. In specific implementation, the selection of the vibration sensors and the setting of the corresponding sampling frequency can be flexibly selected according to actual conditions.

[0048] Specifically, in step S1, in order to effectively utilize the vibration sensors, the vibration sensors are arranged in the same direction. For example, in order to obtain vibration data in the horizontal (vertical) direction of the device, different vibration sensors are arranged in the horizontal (vertical) direction, so that fault judgment is performed through the vibration data in the horizontal (vertical) direction. Meanwhile, in order to solve the problem of unreliable signals and poor stability of a single sensor, the application can fuse the signals of other sensors to comprehensively judge when performing bearing fault diagnosis. For example, when performing fault diagnosis on the driving end bearing, the fan end bearing vibration signal and the base vibration signal can be combined with the driving end bearing vibration signal to comprehensively perform driving end bearing fault diagnosis through the fusion of the three vibration signals, thereby improving the judgment accuracy. In summary, according to the different installation positions of the sensors, different sensor information can be flexibly selected for fusion when performing bearing fault diagnosis, and the correlation and specificity between the information are fully considered.

[0049] Since the vibration signal of the sensor is one-dimensional data, it is not suitable for being directly sent into an algorithm model for fault diagnosis. Therefore, in step S2 of the embodiment, the vibration signal of the sensor is converted into a two-dimensional gray image to realize visualization, thereby making it more convenient to extract and analyze the data. Specifically, in step S2 of the embodiment, the collected vibration signal is converted into a two-dimensional gray image through M 2The length is intercepted to generate N samples, and M and N are preset values. In this embodiment, M is 64, and N can be optional. It should be noted that, for all sensor signals participating in fusion, they need to be intercepted on the same time domain, so as to ensure that the time domain of the samples obtained by interception is consistent and the reliability of the data is ensured. Then, a gray scale image transformation is performed on each sample, so that each sample is transformed into a gray scale image with a size of 1*M*M (the channel number is 1, the length has M pixels, and the width has M pixels). In the gray scale image with a size of 1*M*M, the gray scale value of each pixel point is represented as follows:

[0050]

[0051] wherein the values of i and j are both 1 to M; L i,j represents the amplitude of pixel point (i, j) in the intercepted vibration signal. For example, the amplitude of point i in the vibration signal is 1, and the amplitude of point j is 2, so L i,j = (1, 2); maxL i,j represents the maximum amplitude that pixel point (i, j) can reach in the intercepted vibration signal; minL i,j represents the minimum amplitude that pixel point (i, j) can reach in the intercepted vibration signal; L max represents the maximum pixel value supported by the gray scale image, which ranges between 255 and 0. In this embodiment, L max may be 255; L min represents the minimum pixel value supported by the gray scale image, which ranges between 0 and 255. In this embodiment, L min may be 0.

[0052] The gray scale images obtained by processing the signals of the plurality of sensors in step S2 cannot be directly sent into the algorithm model, and need to be fused before use. The present application selects to fuse the gray scale images into a colorful RGB image. The colorful RGB image usually corresponds to three-channel data, and the gray scale image is single-channel. Therefore, in step S3, three sensors are grouped into one group, and the channel fusion of the gray scale images of the three vibration sensors can obtain a three-channel RGB image. It should be noted that the gray scale images participating in fusion should have the same time domain to ensure the reliability of the data after fusion.

[0053] To further improve the reliability of the data, reduce the redundancy in the data, step S3 of the embodiment adopts principal component analysis technology (PCA) for channel fusion, fusing three single-channel grayscale images into one three-channel RGN image. Principal component analysis technology (PCA) is a common technique for those skilled in the art, and simply speaking, the work of PCA is to sequentially find a set of mutually orthogonal coordinate axes in the original space. The selection of the new coordinate axes is closely related to the data itself. Among them, the first new coordinate axis selection is the direction with the maximum variance in the original data, the second new coordinate axis selection is the maximum variance in the orthogonal plane with the first coordinate axis, the third axis is the maximum variance in the orthogonal plane with the first and second new coordinate axes, and so on. Through this way, we find that most of the variances are contained in the first k coordinate axes, and the variances contained in the remaining coordinate axes are almost 0. Therefore, we can ignore the remaining coordinate axes and only keep the first k coordinate axes containing most of the variances. In fact, this is equivalent to retaining the dimension characteristics containing most of the variances and ignoring the characteristic dimensions containing almost 0 variance, thereby realizing the dimension reduction processing of the data characteristics. The above principle is mainly reflected in the transformation matrix of PCA in actual operation.

[0054] Specifically, the step S3 comprises the following steps:

[0055] S31, respectively converting the grayscale images of the three vibration sensors into corresponding vector forms;

[0056] S32, combining the vectors representing the three grayscale images into a matrix Xraw, wherein the matrix Xraw is n rows and m columns, and the values of n and m are determined by the vectors representing the three grayscale images;

[0057] S33, transforming the matrix Xraw through principal component analysis technology, wherein, because three principal components need to be retained to obtain a three-channel RGB image finally, the transformation is specifically represented by the following formula:

[0058] X pca =W T X raw ={x ij};i=1,2,3;j=1,2,3…m;

[0059] Wherein, X pca ∈R 3×m represents the principal component matrix obtained after transformation through principal component analysis technology, R 3×m represents that the matrix R is 3 rows and m columns; W T ∈R 3×n represents the transformation matrix based on principal component analysis technology, T represents transposition, R 3×n represents that the matrix R is 3 rows and n columns.

[0060] S34、according to the X pca The matrix obtains a three-channel RGB image. Through the X pca The matrix obtains three principal components, and then a three-channel RGB image can be recovered from the three principal components. The three-channel RGB image obtained in this way has the advantages of low data redundancy and high reliability, and can be directly sent to a subsequent algorithm model for fault diagnosis.

[0061] The traditional algorithm model uses the method of widening or deepening the network when performing fault diagnosis, which increases the parameter quantity of the network as a whole and increases the difficulty of model deployment. However, with the proposal of a super-lightweight network (StarNet), the above problems have been solved. The present application applies the super-lightweight network (StarNet) to bearing fault diagnosis, and through the element multiplication introduced by the super-lightweight network (StarNet), the input is mapped to a high-dimensional nonlinear feature space without widening the network, so that the feature mapping between levels is more comprehensive under the condition of smaller parameter quantity, thereby improving the performance of the model. At the same time, the present application also improves the super-lightweight network (StarNet), that is, in the super-lightweight network, the structure of inverted bottleneck is introduced, the inverted bottleneck is a point-wise convolution (PW) + depthwise convolution (DW) + point-wise convolution (PW) structure, which can deepen the feature extraction capability of the lightweight structure without increasing the network parameters, thereby improving the accuracy of fault diagnosis of the present application.

[0062] Specifically, as shown in Figure 3 The improved super-lightweight network StarNet model of the present embodiment includes convolution layers (Conv), a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, a global average pooling layer (GAP) / a fully connected layer (FC) connected in sequence, wherein the output of the convolution layer is input to the first feature extraction layer after batch normalization (BN) and ReLU activation. In the present embodiment, the related parameters of the improved super-lightweight network StarNet model, such as the convolution kernel size, the channel expansion factor, etc., are consistent with the existing super-lightweight network StarNet.

[0063] In the present embodiment, the improvement of introducing inverted bottleneck is mainly aimed at the feature extraction layer of the super-lightweight network StarNet model, as shown in Figure 3As shown, all the feature extraction layers of the embodiment include a convolution layer and a Star Blocks layer connected in sequence, wherein the convolution layer is used for down-sampling, and the Star Blocks layer is used for feature extraction; the Star Blocks layer includes a first point-wise convolution layer, a second point-wise convolution layer, a depth convolution layer, and a full connection layer, wherein the initially input features are processed by the first point-wise convolution layer and the depth convolution layer in sequence; then, the outputs of the depth convolution layer are respectively multiplied by two full connection layers, and the result of the element-wise multiplication is output to the second point-wise convolution layer after passing through a full connection layer; finally, the features processed by the second point-wise convolution layer are output after being connected with the initially input features in a skip connection mode. As can be found, the Star Blocks layer includes the first point-wise convolution layer, the depth convolution layer, and the second point-wise convolution layer in sequence, and constitutes an inverted bottleneck structure, thereby improving the feature extraction capability of the light-weight structure without increasing the network parameters, and further improving the accuracy of the fault diagnosis of the embodiment.

[0064] Before use, the improved ultra-lightweight network StarNet model also needs to be trained (train) and verified (val) to stabilize the model. In the embodiment, the model training parameters are set as shown in Table 1.

[0065] Parameter name Value Number of training epochs 100 Input image size (224,224) Learning rate 0.1 Optimizer SGD Loss function Cross-entropy loss function

[0066] Table 1

[0067] As shown in the experiment, Figure 4 and Figure 5 after 100 training epochs, the accuracy (acc) and loss rate (loss) of the model reach a stable state, and the average accuracy reaches 98.9%, which indicates that the performance of the model reaches the best stable state.

[0068] Finally, the trained model can be used for bearing fault diagnosis, and different categories of fault data can be obtained as the fault diagnosis result after the sensor data is input into the trained model. Further, to improve the user experience, in the step S4, the improved ultra-lightweight network StarNet model is visualized by using the T-SNE algorithm, different categories of fault data are classified, and the classification effect is obvious.

[0069] Meanwhile, the application also provides an ultra-light equipment bearing fault diagnosis system according to the above method, which comprises at least three vibration sensors and a controller, wherein the vibration sensors are connected with the controller, the vibration sensors are installed at different positions of the equipment but in the same direction, and are used for collecting vibration signals of the equipment in the corresponding direction; and the controller is used for bearing fault diagnosis. The controller can be a CPU, a computer or other intelligent equipment with data processing capacity or a control chip, and can be flexibly selected according to the needs in the specific implementation.

[0070] To verify the effectiveness of the method, the application also uses the rolling bearing data set of Case Western Reserve University (CWRU) to conduct experiments. The experiments use the rolling bearing vibration signals of the CWRU data set collected under a sampling frequency of 12 kHz in three different working conditions of the driving end, and the bearing type is SKF6205. The three different working conditions are inner ring fault (Inner Race), rolling element fault (Ball) and outer ring fault (Outer Race), and each working condition contains bearing states with fault diameters (used to represent the fault degree) of 0.007 inch, 0.014 inch and 0.021 inch, so a total of 9 bearing states can be obtained, i.e. BO07, BO14, BO21, IR07, IR14, IR21, OR07, OR14 and OR21, wherein BO07 represents rolling element fault of the driving end bearing with a fault diameter of 0.007 inch, BO14 represents rolling element fault of the driving end bearing with a fault diameter of 0.014 inch, and BO21 represents rolling element fault of the driving end bearing with a fault diameter of 0.021 inch; IR07 represents inner ring fault of the driving end bearing with a fault diameter of 0.007 inch, IR14 represents inner ring fault of the driving end bearing with a fault diameter of 0.014 inch, and IR21 represents inner ring fault of the driving end bearing with a fault diameter of 0.021 inch; and OR07 represents outer ring fault of the driving end bearing with a fault diameter of 0.007 inch, OR14 represents outer ring fault of the driving end bearing with a fault diameter of 0.014 inch, and OR21 represents outer ring fault of the driving end bearing with a fault diameter of 0.021 inch.

[0071] Specific experimental data related parameters are shown in Table 2.

[0072]

[0073] Table 2

[0074] In this experiment, the vibration signals of the drive end bearing, the base vibration signals, and the fan end bearing vibration signals are converted into gray-scale images and fused into a group of fault data, and so on, to obtain multiple groups of fault data in combination with the CWRU data set.

[0075] All the above fault data are input into the improved ultra-light network StarNet model trained to perform fault classification prediction, and are visualized by the T-SNE algorithm, and the results are as shown in Figure 6 It can be seen from Figure 6 that the classification effect of the drive end bearing under different working conditions and different fault degrees is obvious, indicating that the performance of the model can achieve the specified classification effect.

[0076] In order to further show the lightness of the model, the parameter amounts of different models are compared, wherein: Vgg16 represents a conventional model, the detection accuracy of this model is 98.87%, but the parameter amount is dozens of times of the light model; similarly, MobileNet is a representative of a light network, the parameter amount is greatly reduced, but the detection accuracy on the data is only 78.5%, which does not reach the detection standard. The model described in the present application maintains an excellent level in the parameter amount and the detection accuracy, fully illustrating the effectiveness of the method described in the present application.

[0077] The present application is described above in conjunction with the drawings. Obviously, the specific implementation of the present application is not limited by the above manner. As long as various non-essential improvements are made by adopting the method concept and technical solution of the present application; or without improvement, the above concept and technical solution of the present application are directly applied to other occasions, which are all within the protection scope of the present application.

Claims

1. A method for diagnosing bearing faults in ultra-lightweight equipment, characterized in that: The method includes the following steps: S1. Vibration signals in the corresponding direction of the equipment are collected by at least three vibration sensors installed at different locations on the equipment but in the same direction; S2. Visualize the vibration signal as a grayscale image; S3. Three vibration sensors installed at different locations on the equipment but in the same direction are grouped together, and the grayscale images of each group of vibration sensors are fused into a three-channel RGB image. S4. An improved ultra-lightweight StarNet model is used to diagnose bearing faults in the RGB image. The improved ultra-lightweight StarNet model includes a convolutional layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, and a global average pooling layer / fully connected layer connected in sequence. The output of the convolutional layer is sequentially processed by batch normalization and ReLU activation before being input into the first feature extraction layer. All feature extraction layers include a convolutional layer and a Star Blocks layer connected in sequence. The convolutional layer is used for downsampling, and the Star Blocks layer is used for feature extraction. The Star Blocks layer includes a first pointwise convolutional layer, a second pointwise convolutional layer, a depthwise convolutional layer, and a fully connected layer. The initially input features are processed sequentially by the first pointwise convolutional layer and the depthwise convolutional layer. Then, the output of the depthwise convolutional layer is multiplied element-wise by two fully connected layers. The result of the element-wise multiplication is then processed by a fully connected layer before being output to the second pointwise convolutional layer. Finally, the features processed by the second pointwise convolutional layer are connected to the initially input features and then output.

2. The method for diagnosing bearing faults in ultra-lightweight equipment according to claim 1, characterized in that: The device includes a fan-end bearing, a motor, a drive-end bearing, a reducer bearing, a load device, a base, and a coupling. Correspondingly, depending on the installation position of the vibration sensor, the vibration signals it collects include fan-end bearing vibration signals, motor vibration signals, drive-end bearing vibration signals, reducer bearing vibration signals, load device vibration signals, base vibration signals, and coupling vibration signals.

3. The method for diagnosing bearing faults in ultra-lightweight equipment according to claim 1, characterized in that: In step S2, each acquired vibration signal is processed. The length is truncated to generate N samples, where M and N are preset values; Each sample is transformed into a grayscale image of size 1×M×M. In this 1×M×M grayscale image, the grayscale value of each pixel is represented as follows: ； Wherein, the values ​​of i and j are both from 1 to M; This represents the amplitude of pixel (i,j) in the extracted vibration signal; This represents the maximum amplitude that pixel (i,j) can achieve in the captured vibration signal; This represents the minimum amplitude that pixel (i,j) can achieve in the captured vibration signal; This indicates the maximum pixel value supported by the grayscale image, which ranges from 255 to 0. This represents the minimum pixel value supported by the grayscale image, ranging from 0 to 255.

4. The method for diagnosing bearing faults in ultra-lightweight equipment according to claim 1, characterized in that: In step S3, the grayscale images of the three vibration sensors are fused to obtain a three-channel RGB image, wherein the grayscale images involved in the fusion have the same time domain.

5. The method for diagnosing bearing faults in ultra-lightweight equipment according to claim 4, characterized in that: In step S3, principal component analysis is used to perform channel fusion.

6. The method for diagnosing bearing faults in ultra-lightweight equipment according to claim 5, characterized in that: Step S3 includes the following steps: S31. Convert the grayscale images of the three vibration sensors into their corresponding vector forms respectively; S32. Merge the vectors representing the three grayscale images into a matrix Xraw, where the matrix Xraw is n rows and m columns, and the values ​​of n and m are determined by the vectors representing the three grayscale images. S33. The matrix Xraw is transformed using principal component analysis, as expressed by the following formula: ； in, This represents the principal component matrix obtained after transformation using principal component analysis techniques. This indicates that matrix R is 3 rows and m columns; This represents the transformation matrix based on principal component analysis, where T denotes transpose. This indicates that matrix R is 3 rows and n columns; S34, according to the above The matrix produces a three-channel RGB image.

7. The method for diagnosing bearing faults in ultra-lightweight equipment according to claim 1, characterized in that: In step S4, the fault diagnosis results of the improved ultralightweight network StarNet model are visualized using the T-SNE algorithm.

8. An ultra-lightweight equipment bearing fault diagnosis system, characterized in that: The system includes at least three vibration sensors and a controller. The vibration sensors are all connected to the controller. The vibration sensors are installed at different locations on the equipment but in the same direction to collect vibration signals in the corresponding direction of the equipment. The controller is used to diagnose bearing faults in an ultra-lightweight equipment bearing fault diagnosis method according to any one of claims 1-7.

Images