Rolling bearing fault identification method based on image information
By performing envelope analysis and characteristic frequency calculation of the vibration signals of rolling bearings, combined with deep learning models, the accuracy and automation problems of rolling bearing fault identification in the prior art are solved, and efficient fault diagnosis in noisy environments are achieved.
Patent Information
- Application Number
- CN202411480834.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-23
- Publication Date
- 2024-11-22
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing rolling bearing fault identification method is difficult to accurately capture fault characteristics when processing nonlinear and noise vibration signals. The deep learning method can only identify fault types but cannot extract feature frequency and harmonic information, which limits its practical application. effectiveness.
By enveloping the vibration signals of rolling bearings, the characteristic frequency and its harmonics are calculated, and the envelope spectrum pictures are marked using the annotation tool to build a rolling bearing information identification model, and feature fusion and prediction are performed through Backbone, Neck and Head networks to calculate the loss Function optimization model parameters to achieve fault identification.
It is more stable in noise environments, has higher automation and real-time performance, improves the accuracy of fault diagnosis, and reduces maintenance costs and downtime.
Smart Images

Figure CN119004220A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a rolling bearing fault identification method based on image information, belonging to the field of mechanical system fault diagnosis. Background Art
[0002] In modern industrial equipment, rolling bearings are key components in mechanical transmission systems and are often compared to joints in mechanical structures. They have the advantages of low friction, easy installation, high efficiency, and easy lubrication. They are perfectly suitable for rotational motion. Therefore, rolling bearings are widely used in various mechanical equipment, such as automobiles, aircraft, motors, and household appliances. Their operating conditions directly affect the stability and life of the equipment. Failure of rolling bearings will not only cause equipment downtime, but may also cause more serious mechanical damage or even safety accidents. Therefore, early identification and diagnosis of rolling bearing failures are particularly important.
[0003] Traditional fault identification methods mainly rely on vibration analysis technology and spectrum analysis technology. The former is to judge the operating status of the bearing by analyzing the vibration signal, and the latter is to detect the fault signal of the bearing through spectrum analysis technology to obtain the theoretical defect frequency and sideband frequency. However, these methods often require complex signal processing algorithms and a deep theoretical knowledge base. In addition, rolling bearing fault signals are usually nonlinear and non-stationary, and vibration signals often contain a lot of noise. Traditional spectrum analysis and vibration analysis methods may perform poorly when processing such signals, and it is difficult to accurately capture fault features, making it more difficult to determine the fault type. In addition, modern fault diagnosis methods based on deep learning use the features of extracted neural networks for fault diagnosis. They can often only identify the fault type but cannot extract the information contained in the fault data, such as fault characteristic frequency, harmonics of fault characteristic frequency, and other features. Therefore, when existing methods extract the features contained in the data while completing fault identification, it is usually difficult to obtain accurate results, which limits their effectiveness in practical applications. Summary of the invention
[0004] The present invention provides a rolling bearing fault recognition method based on image information, which obtains an envelope spectrum by performing envelope analysis on the vibration signal of a real rolling bearing; then, prior knowledge is introduced to calculate the characteristic frequencies and harmonics corresponding to different faults, and a labeling tool is used to label the fault characteristic frequencies and harmonics of the envelope spectrum image to participate in the training of a rolling bearing information recognition model; the unlabeled envelope feature map is used as the input of the trained rolling bearing information recognition model, and the actual frequency value corresponding to the recognition object can be calculated according to the model output result, thereby realizing the recognition and diagnosis of the fault.
[0005] The technical solution of the present invention is:
[0006] The present invention provides a rolling bearing fault identification method based on image information, comprising:
[0007] S1. Perform data segmentation processing on the real rolling bearing vibration signal to obtain the segmented rolling bearing vibration signal; perform bandpass filtering processing on each segment of the rolling bearing vibration signal to obtain the vibration signal after the bandpass filtering processing; apply Hilbert transform to extract the envelope signal of the vibration signal after the bandpass filtering processing, perform fast Fourier transform on the envelope signal to obtain the envelope spectrum at a preset frequency, save the envelope spectrum image in an image format to construct a sample set; divide the sample set into a training set, a validation set and a test set;
[0008] S2. Introduce prior knowledge, calculate the characteristic frequencies and harmonics corresponding to different rolling bearing faults, and use them as standard values; and use annotation tools to annotate the fault characteristic frequencies and harmonics of the images in the training set and the validation set to obtain the labeled envelope feature graph; wherein the labels are stored in text format;
[0009] S3. The labeled envelope feature map is used as the input of the rolling bearing information recognition model. The rolling bearing information recognition model is input into the Head network after feature fusion through the Backbone network and the Neck network to obtain the position and confidence of the prediction box. The prediction box is screened by non-maximum suppression, and the loss function is calculated to adjust the weight parameters. The model parameters are optimized through continuous training to obtain the optimal weight data; the optimal weight data is loaded to obtain the trained rolling bearing information recognition model;
[0010] S4. Input the envelope spectrum image of the test set into the trained rolling bearing information recognition model to obtain a prediction result; determine the fault type based on the prediction result.
[0011] The S3 is specifically as follows: taking the labeled envelope feature map as input, and then preprocessing the labeled envelope feature map, the preprocessing includes Mosaic data enhancement, adaptive anchor frame calculation and adaptive image scaling; the preprocessed image is used as the input of the Backbone network, and three feature maps P1, P2, and P3 of different sizes are obtained after feature extraction; the three feature maps of different sizes obtained by feature extraction are used as the input of the Neck network, and three feature maps M1, M2, and M3 of different scales are obtained after feature fusion; the feature maps of three scales obtained by feature fusion are input into the Head network, and three feature tensors of different sizes are obtained after the detector, i.e., three feature tensors of different sizes are obtained. The feature map is automatically assigned to the corresponding anchor frame and feature map according to the size of the annotated object frame, and then prediction is performed on the feature map to obtain intermediate prediction information; the product of the confidence and the probability of each fault type is calculated through the intermediate prediction information, and the scores of all types in each prediction frame are calculated. The prediction frame with the largest score in each prediction frame determines the corresponding fault type; by setting a threshold, the prediction frame above the threshold is retained; for the prediction frame that determines the fault type, the non-maximum suppression method is used to continuously screen the overlapping prediction frames, leaving the prediction frame with the highest score as the prediction frame of the fault type; the weight parameters are adjusted by calculating the loss function, and the model parameters are continuously optimized by calculating the accuracy and average precision.
[0012] The method of determining the fault type based on the prediction result is specifically as follows: according to the pixel coordinates of the upper left corner and the lower right corner of the prediction box output by the model, the pixel coordinates of the center point of the prediction box are calculated; according to the horizontal pixel size of the blank space and the pixel coordinates of the center point of the prediction box, the pixel horizontal coordinates of the center point of the prediction box relative to the coordinate system are obtained; by calculating the ratio of the horizontal interval length of the coordinate system to the horizontal pixel size of the coordinate system, and according to the pixel horizontal coordinates of the center point of the prediction box relative to the coordinate system, the actual frequency value corresponding to the horizontal coordinate of the center point of the prediction box is obtained. According to this step, the actual frequency value corresponding to the center point of each prediction box can be calculated in turn, and the actual frequency value corresponding to the center point of each prediction box is compared with the standard value under different fault types, so as to determine the fault type.
[0013] The beneficial effects of the present invention are as follows: the present invention obtains an envelope spectrum by performing envelope analysis on the vibration signal of the rolling bearing, introduces prior knowledge such as the rolling bearing model and the motor speed, calculates the frequency value corresponding to the identification object, and uses a labeling tool to label the fault characteristic frequency and its harmonics in the envelope characteristic diagram, and then uses the labeled envelope characteristic diagram as the input of the rolling bearing information recognition model, optimizes the model parameters through training, obtains the best weight data, selects the test set data to test the trained rolling bearing information recognition model, calculates the actual frequency value corresponding to the identification object according to the model output result, and then determines the fault category. Compared with the traditional vibration analysis method, the method of the present invention is more stable in a noisy environment, and has a higher degree of automation and real-time performance, which can not only improve the accuracy of fault diagnosis, but also effectively reduce maintenance costs and downtime. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 It is a flowchart of the present invention;
[0015] Figure 2 A schematic diagram of a rolling bearing information identification model provided according to an embodiment of the present invention;
[0016] Figure 3 A confusion matrix diagram of a rolling bearing information recognition model training according to an embodiment of the present invention;
[0017] Figure 4 , Figure 5 A first set of test schematic diagrams provided according to an embodiment of the present invention;
[0018] Figure 6 , Figure 7 A second set of test schematic diagrams provided according to an embodiment of the present invention;
[0019] Figure 8 , Fig. 9 A third set of test schematic diagrams provided according to an embodiment of the present invention. DETAILED DESCRIPTION
[0020] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should be noted that the embodiments in this application and the features in the embodiments can be combined with each other arbitrarily without conflict.
[0021] Example 1: Figure 1-Figure 9 As shown, according to a first aspect of an embodiment of the present invention, a rolling bearing fault identification method based on image information is provided, comprising:
[0022] S1. Perform data segmentation processing on the real rolling bearing vibration signal to obtain the segmented rolling bearing vibration signal; perform bandpass filtering processing on each segment of the rolling bearing vibration signal to obtain the vibration signal after the bandpass filtering processing; apply Hilbert transform to extract the envelope signal of the vibration signal after the bandpass filtering processing, perform fast Fourier transform on the envelope signal to obtain the envelope spectrum at a preset frequency, save the envelope spectrum image in an image format to construct a sample set; divide the sample set into a training set, a verification set and a test set; in the embodiment of the present invention, the preset frequency is 0-1000Hz;
[0023] S2. Introduce prior knowledge, calculate the characteristic frequency and harmonics corresponding to different rolling bearing faults, and use this as the standard value; and use the annotation tool to annotate the fault characteristic frequency and harmonics of the images in the training set and the validation set to obtain a labeled envelope characteristic diagram; wherein the label is stored in text format; illustratively, the tool used to annotate the samples is LabelImg, and there are a total of 5 labels, namely characteristic freq, harmonic 2x, harmonic 3x, harmonic4x, and harmonic 5x; in this embodiment, characteristic freq corresponds to the characteristic frequency, harmonic 2x corresponds to the 2nd harmonic, harmonic 3x corresponds to the 3rd harmonic, harmonic 4x corresponds to the 4th harmonic, and harmonic 5x corresponds to the 5th harmonic. When annotating specifically, import all the training set and validation set envelope spectrum images, annotate the targets with rectangular boxes, define them as corresponding labels, and use txt as the suffix of the annotated files, and the txt file name is consistent with the name of the envelope spectrum image;
[0024] S3. The labeled envelope feature map is used as the input of the rolling bearing information recognition model. The rolling bearing information recognition model is input into the Head network after feature fusion through the Backbone network and the Neck network to obtain the position and confidence of the prediction box. The prediction box is screened by non-maximum suppression, and the loss function is calculated to adjust the weight parameters. The model parameters are optimized through continuous training to obtain the optimal weight data; the optimal weight data is loaded to obtain the trained rolling bearing information recognition model;
[0025] S4. Input the envelope spectrum image of the test set into the trained rolling bearing information recognition model to obtain the prediction result (i.e., the image including the prediction box, label, confidence level and pixel coordinates); determine the fault type based on the prediction result.
[0026] Furthermore, the S3 is specifically:
[0027] The labeled envelope feature map is taken as input, and then the labeled envelope feature map is preprocessed, and the preprocessing includes Mosaic data enhancement, adaptive anchor frame calculation and adaptive image scaling; in an embodiment of the present invention, the size of the image after preprocessing is 640×640×3, and the Mosaic data enhancement includes random scaling, random cropping, and random arrangement. The initial anchor frame of the rolling bearing information recognition model is set to [116, 90, 156, 198, 373, 326], [30, 61, 62, 45, 59, 119], and [10, 13, 16, 30, 33, 23]. The model outputs a predicted frame based on the initial anchor frame, compares it with the real frame, calculates the difference between the two, and then reversely updates it, and iterates the model parameters;
[0028] The preprocessed image is used as the input of the Backbone network. After feature extraction, three feature maps of different sizes, P1, P2, and P3, are obtained.
[0029] The three feature maps of different sizes obtained by feature extraction are used as the input of the Neck network. After feature fusion, three feature maps of different sizes M1, M2, and M3 are obtained.
[0030] The feature maps of three scales obtained by feature fusion are input into the Head network, and three feature tensors of different sizes are obtained after the detector, that is, they contain three feature maps of different sizes. According to the different sizes of the labeled object boxes, they are automatically assigned to the corresponding anchor boxes and feature maps, and then predictions are made on the feature maps to obtain intermediate prediction information; the product of the confidence and the probability of each fault type is calculated through the intermediate prediction information, and the scores of all fault types in each prediction box are calculated. The prediction box with the largest score in each prediction box determines the corresponding fault type; by setting a threshold, the prediction boxes above the threshold are retained; for the prediction boxes that determine the fault type, the non-maximum suppression method is used to continuously screen the overlapping prediction boxes, leaving the prediction box with the highest score as the prediction box of the fault type; the weight parameters are adjusted by calculating the loss function, and the model parameters are continuously optimized by calculating the accuracy and average precision.
[0031] Furthermore, the intermediate prediction information is ; is the confidence information, is the horizontal coordinate of the center point of the prediction box relative to the image after scaling, is the vertical coordinate of the center point of the prediction box relative to the image after scaling, is the width of the predicted box relative to the image after scaling, is the height of the predicted box relative to the image after scaling, is the probability that the predicted box belongs to a certain fault type.
[0032] Furthermore, when determining the fault type of the prediction box, it is necessary to first calculate the product of the confidence and the probability of each fault type, calculate the scores of all fault types in each prediction box, and the one with the largest score in each prediction box determines the corresponding fault type.
[0033] The score calculation formula is:
[0034] Scores = p x × [ c 1 , c 2 , c 3 , c 4 , c 5 ] T ;
[0035] in, The confidence information output by the Detect detector indicates the probability of whether an object exists in the prediction box; is the probability that the predicted box belongs to a certain fault type;
[0036] Assumptions The score of is the largest, then the fault type of this prediction box is The corresponding fault type.
[0037] Furthermore, the threshold is set to 0.25, and a mask is created according to the threshold. Prediction boxes above 0.25 are retained, and their corresponding masks are True or 1. Prediction boxes below 0.25 are discarded, and their corresponding masks are False or 0. The dimension of the mask is the number of prediction boxes.
[0038] Set the maximum number of training iterations to 150, the initial training weight data to yolov5s.pt, the number of input images for each model training to 32, the initial learning rate of the model to 0.01, and the width_multiple parameter to 0.5, which is used to control the model width and mainly affects the number of output channels in the parameters.
[0039] Furthermore, if Figure 2 As shown, the Backbone network includes Focus structure, CBS structure, CSP structure and SPPF structure, specifically:
[0040] When the original image of size 640×640×3 passes through the Focus structure, a slice operation is performed to generate a feature map of size 320×320×32;
[0041] The 320×320×32 feature map is subjected to CBS convolution with a convolution kernel of 3×3 and 128 channels and a CSP structure to generate a feature map of size 160×160×64;
[0042] After the 160×160×64 feature map is subjected to CBS convolution with a convolution kernel of 3×3 and a channel number of 256 and a CSP structure, a feature map P1 of size 80×80×128 is generated;
[0043] After the feature map P1 undergoes CBS convolution with a convolution kernel of 3×3 and a channel number of 512 and a CSP structure, a feature map P2 with a size of 40×40×256 is generated;
[0044] After the feature map P2 undergoes CBS convolution with a convolution kernel of 3×3 and 1024 channels, as well as CSP structure and SPPF structure, a feature map P3 with a size of 20×20×512 is generated.
[0045] The CBS structure consists of a convolution layer, a normalization layer and an activation function layer, and its function is to perform network convolution.
[0046] The CSP structure divides the input feature map into two parts, one part is passed through a series of convolution operations, and the other part is directly connected to the output. This structure helps to fuse features between different layers, improve feature expression capabilities, and effectively extract features.
[0047] The SPPF structure aggregates information of different scales through multi-scale pooling operations, which helps to enhance the expressiveness of feature maps.
[0048] Furthermore, if Figure 2 As shown, the Neck network adopts the structure of FPN+PAN. Specifically, the operation of FPN of the Neck network includes:
[0049] The 20×20×512 feature map P3 output by the Backbone network is convolved with a CBS convolution kernel of 1×1 and 512 channels to generate a feature map N1 of size 20×20×256.
[0050] After upsampling, feature map N1 generates a feature map of size 40×40×256, which is then used for feature fusion with feature map P2 of size 40×40×256 to generate feature map N2 of size 40×40×512;
[0051] After passing through the CSP structure, feature map N2 generates a feature map of size 40×40×512; then after passing through the CBS convolution with a convolution kernel of 1×1 and 256 channels, a feature map of size 40×40×128 is generated; then after upsampling, a feature map of size 80×80×128 is generated, which is used for feature fusion with the feature map P1 of size 80×80×128 to generate a feature map N3 of size 80×80×256.
[0052] Furthermore, if Figure 2 As shown, the operations of the PAN of the Neck network include:
[0053] The 80×80×256 feature map N3 output by the FPN network is converted into a feature map M1 of size 80×80×128 after the CSP structure.
[0054] After the feature map M1 is convolved by CBS with a convolution kernel of 3×3 and a channel number of 256, a feature map of size 40×40×128 is generated. Then, after feature fusion with the feature map of size 40×40×128 output from FPN, a feature map of size 40×40×256 is generated.
[0055] The 40×40×256 feature map is converted into a 40×40×256 feature map M2 after the CSP structure.
[0056] After the feature map M2 is convolved with a convolution kernel of 3×3 and a channel number of 512, a feature map of size 20×20×256 is generated;
[0057] After the feature map of 20×20×256 is fused with the feature map N1 of size 20×20×256, a feature map of size 20×20×512 is generated;
[0058] After the 20×20×512 feature map passes through the CSP structure, a feature map M3 of size 20×20×512 is generated.
[0059] The Neck network mentioned above is a technology for processing multi-scale target detection, which can be achieved by adding feature layers of different scales to the backbone network. FPN uses upsampling from top to bottom to fuse feature maps of different levels together, and the path aggregation network PAN uses a convolutional layer from bottom to top to fuse feature maps from different levels; the upsampling multiple is 2.
[0060] In the embodiment of the present invention, the CSP structure in the Backbone network uses residual links, and the CSP structure in the Neck network does not use residual links. Figure 2 In the figure, the CSP structure in the Backbone network is represented by CSP1_x, and the CSP structure in the Neck network is represented by CSP2_x.
[0061] Furthermore, the Head network contains three Detect detector structures, including some convolutional layers, pooling layers and fully connected layers, which are mainly responsible for multi-scale target detection on the feature maps extracted by the backbone network;
[0062] The operation of the Head network is that the feature maps M1, M2, and M3 are used as the input of the Head network. After passing through the Detect detector structure, the output is three feature tensors of different sizes, which are 20×20×A×(5+B), 40×40×A×(5+B), and 80×80×A×(5+B), respectively. A is the number of anchor frames, A=3, and B is the number of fault types, B=5. 80×80 represents a shallow feature map, which is suitable for detecting small targets. 20×20 represents a deep feature map, which is suitable for detecting large targets. The 40×40 feature map uses an anchor frame between these two scales to detect medium-sized targets.
[0063] Furthermore, the fault type is determined based on the prediction result, specifically:
[0064] According to the pixel coordinates of the upper left and lower right corners of the prediction box output by the model, the pixel coordinates of the center point of the prediction box are calculated; according to the horizontal pixel size of the blank space and the pixel coordinates of the center point of the prediction box, the pixel horizontal coordinates of the center point of the prediction box relative to the coordinate system are obtained; by calculating the ratio of the horizontal interval length of the coordinate system to the horizontal pixel size of the coordinate system, and according to the pixel horizontal coordinates of the center point of the prediction box relative to the coordinate system, the actual frequency value corresponding to the horizontal coordinate of the center point of the prediction box is obtained. According to this step, the actual frequency value corresponding to the center point of each prediction box can be calculated in turn, and the actual frequency value corresponding to the center point of each prediction box can be compared with the standard value under different fault types to determine the fault type.
[0065] Assume that the coordinates of the upper left corner of the prediction box output by the model are , the coordinates of the lower right corner are , then the calculation formula for the pixel coordinates of the center point of the prediction box is:
[0066] ;
[0067] ;
[0068] The calculation formula for the pixel horizontal coordinate of the center point of the prediction box relative to the coordinate system is:
[0069] ;
[0070] Wherein, a is the horizontal pixel size of the blank space at the origin of the coordinate system relative to the left side of the envelope spectrum image, and in the embodiment of the present invention, a=114 is set;
[0071] The calculation formula for the actual frequency value corresponding to the horizontal coordinate of the center point of the prediction box is:
[0072] ;
[0073] Wherein, z is the difference between the maximum value and the minimum value of the preset frequency, that is, the horizontal interval length of the coordinate system, and in the embodiment of the present invention, z=1000; e is the horizontal pixel size of the coordinate system, and e=676.
[0074] Furthermore, the actual frequency value corresponding to the center point of each prediction box is compared with the standard value under different fault types to determine the fault type: assuming that the number of actual frequency values corresponding to the center point of the prediction box is M; under the standard value of different fault types, the absolute value of the difference between the actual frequency value corresponding to the center point of each prediction box and the standard value under the selected fault type is calculated, and the actual frequency value corresponding to the center point of the prediction box that satisfies the absolute value of the difference between the two is less than the preset value is statistically counted; the fault type selected when the count is the largest is used as the prediction result to determine the fault type (it should be noted that if the maximum count value is less than M / 2, it is classified as an unknown fault). For example, for a sample, 5 prediction frames are obtained through prediction by the model of the present invention, and the actual frequency value corresponding to the horizontal coordinate of the center point of the prediction frame can be obtained for each of the 5 prediction frames; under three fault types: inner ring, outer ring, and rolling element fault; when the standard value of the inner ring fault is selected, the absolute value of the actual frequency value corresponding to the center point of the 5 prediction frames and the standard value of the inner ring fault are both less than the preset value, the count is 5, and the other two counts are 0; therefore, the fault type corresponding to the standard value is determined to be the fault type of the current sample, that is, the inner ring fault. For ease of understanding, further explanation is as follows: For example, Figure 4 The sample is used to illustrate that 5 prediction frames are obtained after model testing. The absolute value of the difference between the actual frequency values of the 5 prediction frames and the standard value under the outer ring fault type is less than the preset value of 5Hz, so it is judged to be an outer ring fault.
[0075] The non-maximum suppression method NMS is used to remove redundant detected frames, retain only the frames that are most likely to contain the target object, and retain the best detection results.
[0076] The most important function used by the non-maximum suppression method is called the intersection over union (IoU), which is calculated as follows:
[0077] ;
[0078] Among them, A is the first prediction box, and B is the second prediction box;
[0079] Assume that the coordinates of the upper left corner of the first predicted box are , the coordinates of the lower right corner are , the coordinates of the upper left corner of the second prediction box are , the coordinates of the lower right corner are , then the calculation formula of IoU is:
[0080] ;
[0081] IoU=0 means that the two prediction boxes have no intersection, and IoU=1 means that the two prediction boxes completely overlap;
[0082] The IoU threshold is set to 0.45, and prediction boxes below the threshold are retained, which can ensure that the most accurate prediction box is retained from multiple overlapping boxes.
[0083] The loss function includes a positioning loss function , category loss function And the confidence loss function ,The calculation formula of the loss function is:
[0084] ;
[0085] in, The loss function and The loss function used is binary cross entropy loss; The loss function is calculated as:
[0086] ;
[0087] ;
[0088] ;
[0089] ;
[0090] in, is the intersection-over-union ratio of the predicted box and the true box; B is the predicted box, is the real frame, is the distance between the center points of the two boxes; is the diagonal length of the minimum enclosing rectangle of the two boxes, is the similarity of the aspect ratio of the two boxes, for Impact factor; , Represent the width of the predicted box and the real box respectively, , Represent the heights of the predicted box and the true box respectively.
[0091] According to a second aspect of an embodiment of the present invention, a rolling bearing fault identification system based on image information is provided, comprising a module of any one of the above methods.
[0092] According to a third aspect of an embodiment of the present invention, a terminal device is provided, comprising a memory and a processor, wherein the memory is used to store instructions executed by the processor, and when the instructions are executed by the processor, the terminal device is used to implement a rolling bearing fault identification method based on image information as described in any one of the above.
[0093] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium includes stored instructions, wherein when the instructions are executed, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned rolling bearing fault identification methods based on image information.
[0094] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.
[0095] Taking the 6205-2RSL JME SKF model deep groove ball bearing as an example, the optional specific embodiments of the present invention are described as follows:
[0096] The bearing vibration signal data disclosed by Case Western Reserve University in the United States was selected to test the model. The motor speed was about 1797r / min, and the sampling frequency was 12kHz under no-load conditions. The vibration data of the drive end (DE) was collected. The data set contains vibration signals in three states: inner ring fault, outer ring fault, and rolling element fault. For the vibration signals of each state of the rolling bearing, the data is segmented with 2048 data points, and each vibration signal is bandpass filtered. The envelope of the signal is extracted by Hilbert transform, and the envelope signal is fast Fourier transformed to obtain the envelope spectrum, and the envelope spectrum image is saved in .png format; finally, the present invention selects 78 samples of each state, with a total of 234 samples. The obtained data set is shown in the following table:
[0097] Table 1 Experimental dataset
[0098]
[0099] In an embodiment of the present invention, prior knowledge such as rolling bearing model and motor speed is introduced, the frequency values corresponding to different faults of the identification object are calculated, and the fault feature frequencies and harmonics of the images in the training set and the validation set are labeled in turn using the LabelImg labeling tool to obtain a labeled envelope feature diagram; wherein the label information is stored in text format, specifically in .txt format, to facilitate subsequent model training.
[0100] For 6205-2RSL JME SKF model deep groove ball bearings, it involves inner ring fault, outer ring fault, and rolling element fault.
[0101] The inner race fault characteristic frequency The calculation formula is:
[0102]
[0103] The outer race fault characteristic frequency The calculation formula is:
[0104]
[0105] The rolling element failure characteristic frequency The calculation formula is:
[0106] F BSF = Df r 2 d × 60 [ 1 − ( d D cos α ) 2 ]
[0107] Where n is the number of rolling elements, n=9; is the motor speed, ; d is the ball diameter, d=7.94mm; D is the bearing pitch diameter, D=39.04mm; is the contact angle, ;
[0108] Let the characteristic frequency of the inner ring, outer ring and rolling element fault be F, and the calculation formula of its integer multiple harmonics is:
[0109] ;
[0110] Wherein, m is an integer multiple, F is the fault characteristic frequency (i.e., taking the embodiment of the present invention as an example, it refers to the above , , );
[0111] After calculation, the characteristic frequency of the inner race fault is 162.18Hz, the 2nd harmonic is 324.36Hz, the 3rd harmonic is 486.54Hz, the 4th harmonic is 648.72Hz, and the 5th harmonic is 810.9Hz;
[0112] The characteristic frequency of the outer race fault is 107.36Hz, its 2nd harmonic is 214.72Hz, 3rd harmonic is 322.08Hz, 4th harmonic is 429.44Hz, and 5th harmonic is 536.8Hz;
[0113] The characteristic frequency of rolling element failure is 141.17 Hz, its second harmonic is 282.34 Hz, the third harmonic is 423.51 Hz, the fourth harmonic is 564.68 Hz, and the fifth harmonic is 705.85 Hz.
[0114] In the embodiment of the present application, after training based on the training set and the validation set, the model classification results are as follows: Figure 3 As shown, it can be seen that the accuracy of all feature recognitions reached 100%.
[0115] Furthermore, the rolling bearing information recognition model weight data adopts the optimal weight data to obtain the trained rolling bearing information recognition model, and the unlabeled envelope feature map of the test set is used as the input of the trained rolling bearing information recognition model to output the target recognition result.
[0116] It should be noted that the function of the inner ring of the rolling bearing is to cooperate with the shaft and rotate with the shaft, the function of the outer ring is to cooperate with the bearing seat and play a supporting role, and the rolling elements are evenly distributed between the inner ring and the outer ring with the help of the retaining cage to bear axial and radial loads; due to the interference of environmental noise and vibration at the signal collection site and the uncontrollable error factors when the marking tool is marked, there is a certain error in the extraction of the fault characteristic frequency. Therefore, it is normal for the error of the characteristic frequency and its harmonics to be within 5Hz.
[0117] For example, the three unlabeled envelope feature maps of the test set are used as the input of the trained rolling bearing information recognition model to obtain the output results as follows: Figure 4-Figure 9 As shown, Figure 4 , Figure 6 , Figure 8 is the prediction box of the obtained characteristic frequency and its harmonics, Figure 5 , Figure 7 , Fig. 9 is the corresponding prediction box coordinate information. The prediction box coordinate information is output in order from low to high confidence. The four values of each output represent the upper left corner horizontal coordinate, upper left corner vertical coordinate, lower right corner horizontal coordinate, and lower right corner vertical coordinate of the prediction box. According to the calculation formula described above, calculate Figure 4 , Figure 6 , Figure 8 The actual frequency value of the horizontal coordinate of the center point of each prediction box.
[0118] Depend on Figure 5 As shown in the figure, the coordinate information of 5 prediction boxes is generated. Figure 4 The confidence of the prediction box ranges from low to high. Figure 5The first output is the coordinate information corresponding to the characteristic frequency prediction frame, the second is the coordinate information corresponding to the 5th harmonic prediction frame, the third is the coordinate information corresponding to the 2nd harmonic prediction frame, the fourth is the coordinate information corresponding to the 4th harmonic prediction frame, and the second is the coordinate information corresponding to the 3rd harmonic prediction frame. Figure 4 The actual frequency value of the horizontal coordinate of the center point of the prediction box shown is calculated as follows:
[0119] First prediction box: ;
[0120] Second prediction box: ;
[0121] Third prediction box: ;
[0122] Fourth prediction box: ;
[0123] Fifth prediction box: ;
[0124] It can be obtained that the actual frequency value of the horizontal coordinate of the center point of the first prediction frame corresponding to the characteristic frequency is 106.51 Hz, the actual frequency value of the horizontal coordinate of the center point of the second prediction frame corresponding to the 2nd harmonic is 218.93 Hz, the actual frequency value of the horizontal coordinate of the center point of the third prediction frame corresponding to the 3rd harmonic is 322.49 Hz, the actual frequency value of the horizontal coordinate of the center point of the fourth prediction frame corresponding to the 4th harmonic is 433.43 Hz, and the actual frequency value of the horizontal coordinate of the center point of the fifth prediction frame corresponding to the 5th harmonic is 540.68 Hz; the above actual frequency values are compared with the theoretical characteristic frequency values of the outer ring fault and their harmonic standard values. The absolute values of the frequency value differences are all within 5 Hz, so it can be judged that Figure 4 In the subsequent examples, the first prediction box corresponds to the characteristic frequency, the second prediction box corresponds to the 2nd harmonic, the third prediction box corresponds to the 3rd harmonic, the fourth prediction box corresponds to the 4th harmonic, and the fifth prediction box corresponds to the 5th harmonic.
[0125] Figure 7 As shown, the coordinate information of 4 prediction boxes is generated by Figure 6 The confidence of the prediction box ranges from low to high. Figure 7 The first output is the coordinate information corresponding to the 5th harmonic prediction frame, the second is the coordinate information corresponding to the characteristic frequency prediction frame, the third is the coordinate information corresponding to the 3rd harmonic prediction frame, and the fourth is the coordinate information corresponding to the 2nd harmonic prediction frame. Figure 6 The actual frequency value of the horizontal coordinate of the center point of the prediction box shown is calculated as follows:
[0126] First prediction box: ;
[0127] Second prediction box: ;
[0128] Third prediction box: ;
[0129] Fifth prediction box: ;
[0130] According to the above calculation results, the absolute value of the difference between each actual frequency value and the theoretical characteristic frequency value of the inner ring fault and its harmonic standard value is within 5Hz, so it can be judged that Figure 6 It is an inner race fault.
[0131] Fig. 9 As shown, the coordinate information of two prediction boxes is generated by Figure 8 The confidence of the prediction box ranges from low to high. Fig. 9 The first output is the coordinate information corresponding to the third harmonic prediction frame, and the second output is the coordinate information corresponding to the characteristic frequency prediction frame. Figure 8 The actual frequency value of the horizontal coordinate of the center point of the prediction box shown is calculated as follows:
[0132] First prediction box: ;
[0133] Third prediction box: ;
[0134] According to the above calculation results, the absolute value of the difference between the actual frequency value and the theoretical characteristic frequency value of the rolling element fault and its harmonic standard value is within 5Hz, so it can be judged that Figure 8 Rolling element failure.
[0135] In summary, the present invention obtains an envelope spectrum by performing envelope analysis on the vibration signal of the rolling bearing, introduces prior knowledge such as the rolling bearing model and the motor speed, calculates the characteristic frequencies and harmonics corresponding to different faults, and uses a labeling tool to label the fault characteristic frequencies and harmonics of the envelope spectrum image, and uses the unlabeled envelope spectrum image as the input of the trained rolling bearing information recognition model, and then the actual frequency value corresponding to the recognition object can be calculated according to the model output result, so as to realize the recognition and diagnosis of the fault. By applying the above technical solution, the problem that the identification and diagnosis of rolling bearing faults in engineering applications requires a certain professional background and diagnostic experience is solved. At the same time, the rolling bearing fault recognition method based on image information is simulated and verified on real data, which shows that the method of the present invention can adapt to the recognition of vibration signals in a noisy environment, and has a higher degree of automation and real-time performance, which can not only improve the accuracy of fault diagnosis, but also effectively reduce maintenance costs and downtime.
[0136] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.
Claims
1. A rolling bearing fault identification method based on image information, characterized in that: include: S1. Performing data segmentation processing on the real rolling bearing vibration signal to obtain the segmented rolling bearing vibration signal; Perform bandpass filtering on each section of rolling bearing vibration signal to obtain a vibration signal after bandpass filtering; apply Hilbert transform to extract the envelope signal of the vibration signal after bandpass filtering, perform fast Fourier transform on the envelope signal to obtain the envelope spectrum at a preset frequency, save the envelope spectrum image in image format to construct a sample set; divide the sample set into a training set, a validation set and a test set; S2. Introduce prior knowledge, calculate the characteristic frequencies and harmonics corresponding to different rolling bearing faults, and use them as standard values; and use annotation tools to annotate the fault characteristic frequencies and harmonics of the images in the training set and the validation set to obtain the labeled envelope feature graph; wherein the labels are stored in text format; S3. The labeled envelope feature map is used as the input of the rolling bearing information recognition model. The rolling bearing information recognition model is input into the Head network after feature fusion through the Backbone network and the Neck network to obtain the position and confidence of the prediction box. The prediction box is screened by non-maximum suppression, and the loss function is calculated to adjust the weight parameters. The model parameters are optimized through continuous training to obtain the optimal weight data; the optimal weight data is loaded to obtain the trained rolling bearing information recognition model; S4. Input the envelope spectrum image of the test set into the trained rolling bearing information recognition model to obtain a prediction result; determine the fault type based on the prediction result.
2. The rolling bearing fault identification method based on image information according to claim 1 is characterized in that: The S3 is specifically: The labeled envelope feature map is taken as input, and then the labeled envelope feature map is preprocessed, including Mosaic data enhancement, adaptive anchor box calculation and adaptive image scaling; The preprocessed image is used as the input of the Backbone network. After feature extraction, three feature maps of different sizes, P1, P2, and P3, are obtained. The three feature maps of different sizes obtained by feature extraction are used as the input of the Neck network. After feature fusion, three feature maps of different sizes M1, M2, and M3 are obtained. The feature maps of three scales obtained by feature fusion are input into the Head network, and three feature tensors of different sizes are obtained after the detector, that is, they contain three feature maps of different sizes. According to the different sizes of the labeled object boxes, they are automatically assigned to the corresponding anchor boxes and feature maps, and then predictions are made on the feature maps to obtain intermediate prediction information; the product of the confidence and the probability of each fault type is calculated through the intermediate prediction information, and the scores of all types in each prediction box are calculated. The prediction box with the largest score in each prediction box determines the corresponding fault type; by setting a threshold, the prediction boxes above the threshold are retained; for the prediction box that determines the fault type, the non-maximum suppression method is used to continuously screen the overlapping prediction boxes, leaving the prediction box with the highest score as the prediction box of the fault type; the weight parameters are adjusted by calculating the loss function, and the model parameters are continuously optimized by calculating the accuracy and average precision.
3. The rolling bearing fault identification method based on image information according to claim 1 is characterized in that: The fault type is determined according to the prediction result, specifically: According to the pixel coordinates of the upper left and lower right corners of the prediction box output by the model, the pixel coordinates of the center point of the prediction box are calculated; according to the horizontal pixel size of the blank space and the pixel coordinates of the center point of the prediction box, the pixel horizontal coordinates of the center point of the prediction box relative to the coordinate system are obtained; by calculating the ratio of the horizontal interval length of the coordinate system to the horizontal pixel size of the coordinate system, and according to the pixel horizontal coordinates of the center point of the prediction box relative to the coordinate system, the actual frequency value corresponding to the horizontal coordinate of the center point of the prediction box is obtained. According to this step, the actual frequency value corresponding to the center point of each prediction box can be calculated in turn, and the actual frequency value corresponding to the center point of each prediction box can be compared with the standard value under different fault types to determine the fault type.
Citation Information
Patent Citations
Fish feeding behavior identification method based on YOLOv5
CN113537106A
Knowledge-driven industrial robot intelligent migration fault diagnosis method and system
CN113962254A
Bearing diagnosis method for enhancing fault sensitive feature extraction through deep attention mechanism
CN117387947A
Rolling bearing fault intelligent identification method based on envelope form analysis
CN117648669A
System and method for estimating remaining useful life of a bearing
WO2023052390A1
Cited By
Fault identification method and device for speed reducer gear in scraper conveyor
CN119249117A
