Intelligent fault diagnosis method and system for rolling bearing based on enhanced event visual data

Abstract

The application discloses a rolling bearing intelligent fault diagnosis method and system based on enhanced event visual data, which adopts an event camera as a sensor to non-contact collect the health state of a rolling bearing; original event stream data are filtered according to a target region, are converted into event frame data through processing, and are subjected to data enhancement processing in combination with a diffusion model; then, the health state of the bearing is labeled as a corresponding label for the enhanced data, the data set is divided into a training data set and a test data set, and the test data set is subjected to noise adding processing; then, the training data set and the corresponding label are input into a convolutional neural network model for training, so as to build a mapping relationship between the event frame data and the corresponding label; finally, the health state of the rolling bearing is diagnosed according to the mapping relationship model of the event frame data and the corresponding label; the application can adapt to a non-contact sensor to realize fault diagnosis of the rolling bearing, and improve the precision and generalization performance of fault diagnosis using event data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical fault diagnosis, and in particular to a rolling bearing intelligent fault diagnosis method and system based on enhanced event visual data. BACKGROUND

[0002] With the rapid development of intelligent motor systems, the rotating machinery closely related thereto has also been widely applied, and the normal operation of the rotating machinery is closely related to the health state of the bearing, and the health state of the bearing directly reflects whether the entire rotating machinery system can operate normally. However, due to the fact that the bearing needs to work in a high-speed, high-temperature and high-pressure environment for a long time, the possibility of failure of the bearing during operation is very high, which will lead to damage of the mechanical equipment and even cause economic, environmental losses and personal safety accidents. Therefore, monitoring and diagnosing the health state of the rolling bearing is a very important task.

[0003] In engineering practice, vibration signals can effectively reflect the health state of the bearing and are widely used in various health state monitoring tasks. For the collection of rolling bearing vibration signals, the most common method is to use an acceleration sensor. However, since the acceleration sensor needs to be directly connected to the vibrating machinery, it requires a reserved installation position and installation space. In addition, since the sensor will vibrate at a high frequency with the machinery, the possibility of sensor failure is greatly increased, so the application of the acceleration sensor in engineering practice cannot achieve good results.

[0004] With the development of measurement technology, detection technology is also rapidly developing in the direction of non-contact, high precision and high speed, and some non-contact sensors have been used for signal collection of rolling bearings. At present, non-contact sensors for monitoring vibration mainly include eddy current sensors, laser vibration meters and high-speed cameras. Such sensors can complete the measurement task without direct contact with the vibrating machinery, but their defects are also obvious. The eddy current sensor has strict requirements on the shape and material of the measured object, and the test system is relatively expensive, and the installation method and position are also relatively strict. The laser vibration meter also has the problems of high cost and strict installation position. The high-speed camera is easily affected by environmental light and has a large amount of calculation, and the useless background data occupies too much computing resource, and the measured vibration is not accurate enough.

[0005] Event cameras, as a new type of non-contact visual sensor, have been used for vibration measurement due to their high time resolution and high dynamic range. However, there is no overall solution that can adapt to the collection, conversion, processing and diagnosis of the health state of rolling bearings, and the related fault diagnosis process is not perfect, and the engineering practical application value is not high.

[0006] In summary, the existing traditional contact and non-contact sensors have many defects, and the current rolling bearing fault diagnosis method based on event camera is not perfect, the overall fault effect is not good, and the generalization performance and precision of the fault diagnosis model have a lot of room for improvement. SUMMARY

[0007] In order to overcome the above-mentioned defects of the prior art, the purpose of the present application is to provide a rolling bearing intelligent fault diagnosis method and system based on enhanced event visual data, which can adapt to non-contact sensors to realize fault diagnosis of rolling bearings, and improve the precision and generalization performance of fault diagnosis using event data.

[0008] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:

[0009] A rolling bearing intelligent fault diagnosis method based on enhanced event visual data, comprising the following steps:

[0010] Step 1, data acquisition: using an event camera as a sensor to non-contact collect the health status of the rolling bearing, a total of seven types of original bearing vibration signals of different health states are collected as original event stream data, the seven types of different health states include normal state, outer ring fault, inner ring fault and rolling element fault, and three types of fault are collected under light load and heavy load respectively;

[0011] Step 2, data preprocessing: first, the original event stream data is screened according to the target area, then the screened target area event stream data is converted into event frame data through two-dimensional feature reconstruction, then the event frame data is divided into training data set and test data set, and the bearing health status is given a corresponding class label to construct the original data set; Then, additional samples are generated combined with the diffusion model to supplement the original data set; Finally, the test data set is subjected to noise processing, thereby constructing the final intelligent diagnosis data set;

[0012] Step 3, model construction: input the intelligent diagnosis data set into the initial intelligent diagnosis model for training, and construct the mapping relationship between the event frame data and the corresponding label, and finally train the intelligent fault diagnosis model;

[0013] Step 4, fault diagnosis: according to the intelligent fault diagnosis model, the health status of the rolling bearing is tested by the model, and the final diagnosis result is output.

[0014] In step 1, an event camera is used for data acquisition, wherein a single event is represented as e i :

[0015] e i = (t i , xi ,y i ,p i )

[0016] Where e i Let t represent the i-th event. i Let x represent the time when the i-th event occurs. i The x-coordinate of the location where the i-th event occurs is represented by y. i p represents the ordinate value of the location where the i-th event occurs. i p represents the polarity of the event. i =+1 represents an increase in pixel brightness, indicating a positive event has occurred; p i =-1 represents a decrease in pixel brightness, i.e., a negative event has occurred; when the event camera captures a series of events, it will continuously generate events, and the set of these events over a period of time is represented as the event stream E:

[0017]

[0018] Where N is the total number of events that occur during this period.

[0019] In step 2, the original event stream data is first processed by retaining only the events occurring in the target region and removing other background events; the original event stream is denoted as... The coordinate position where a single event occurs is represented as (x i y i If the event location is within the field of view of the event camera, then the initial range of the recorded event location is:

[0020]

[0021] Where x max The maximum event coordinate that the event camera can record in the x-axis direction is y. max This represents the maximum event coordinate that the event camera can record in the y-axis direction. A rectangular region is used to limit the range of event occurrences. After filtering by the rectangular region, an event stream containing only events within that region is generated. This filtered event stream is denoted as:

[0022]

[0023] Among them, e i Let N represent the i-th event. ROI E represents the total number of events that occurred within the defined rectangular area during this period. ROI The position coordinates of the event satisfy:

[0024]

[0025] Where xlower With x upper Let x be the lower and upper bounds of the rectangular region along the x-axis, and y be the upper and lower bounds of the region along the x-axis. lower With y upper This represents the lower and upper bounds of the rectangular region along the y-axis.

[0026] In step 2, the event stream data of the target area is converted into event frame data, specifically including:

[0027] The filtered target region event stream data is divided, event stream signals within a certain time period are obtained, and the number of events L required to generate a single sample and the step size D between two samples are set. Then, the event stream data required for the first sample is represented as E1 = [e1, e2, ..., e...]. L The second sample is represented as E2 = [e] 1+D e 2+D , ..., e L+D The event stream data required for the k-th sample is represented as E. k =[e 1+(k-1)D e 2+(k-1)D , ..., e L+(k-1)D ], and so on, to obtain the event stream data needed to generate all samples.

[0028] Based on the event stream conversion model, the event stream of the corresponding target area captured by the event camera is converted into event frames. When a positive event occurs in the event stream, and no positive event has occurred at the location where the event occurred before this moment, the event stream conversion model is expressed as follows:

[0029]

[0030] Where I(x, y) represents the pixel value at (x, y) in the event frame, and the subscripts R, G, and B of I indicate that a single channel of the three channels in the event frame needs to be operated on. V base Set the base pixel value to half of the highest pixel value in the event frame;

[0031] When a positive event occurs in the event stream, and a positive event has occurred at the position where the event occurred before this moment, the event stream transformation model is represented as:

[0032]

[0033] Where M P With M N These represent the number of positive events and the number of negative events that occur most frequently at the same position in the event stream, respectively.

[0034] When a negative event occurs in the event stream, and no negative event has occurred at the position where the event occurred before this moment, the event stream transformation model is represented as:

[0035]

[0036] When a negative event occurs in the event stream, and at this moment before the event occurs at the location of the negative event, the event stream conversion model is represented as:

[0037]

[0038] In step 2, a diffusion model is trained using the converted event frame data described above. First, a training sample is randomly selected from the converted event frame data described above, and then a value t is randomly selected from the noise sequence 1-T. After that, the sample is processed by adding noise, that is, the forward noise adding process of the diffusion model, which is represented by the following formula:

[0039]

[0040] Where s0 represents the initial input sample, s t represents the sample after adding noise t times, q(s t |s0) represents the conditional probability distribution of the sample s t after adding noise t times derived from the sample s0, I is the unit matrix, is a hyperparameter related to the mean and variance of the normal distribution; after obtaining the data of the sample after adding noise from the above formula, the image is restored through the reverse denoising process of the diffusion model, and the distribution of this process is learned through the deep learning network. The reverse process is represented by the following formula:

[0041]

[0042] Where p θ (s t-1 |s t ) represents the conditional probability distribution of the sample s t after denoising once derived from the sample s t-1 , μ θ (s t ,t) is the mean of the normal distribution in the reverse diffusion; after denoising the noisy sample, it is sent into the deep learning training network together with the noise number t for iteration, and the predicted noise is compared with the real noise and the loss is calculated. Finally, the gradient is calculated and the deep learning network used for denoising the sample is updated, and the process is repeated until the network is finally trained; after obtaining the final diffusion model, random noise data is input into it, that is, additional samples are generated, and these samples generated by the diffusion model are used as a supplement to the original training set.

[0043] The transformation of the event stream data into samples is completed, and then it is divided into a training data set and a test data set, and noise is added to the test data set, and the distribution of the added noise is subject to:

[0044]

[0045] where σ is the variance of the added noise normal distribution, μ is the mean of the added noise normal distribution, and is the test data set after adding noise. The data set after the above preprocessing can be used for the training and testing of the initial intelligent diagnosis model.

[0046] In step 3, an initial intelligent diagnosis model for learning fault features is constructed, which specifically consists of a convolution block, a max pooling layer, a fully connected layer, and a Drop-out layer.

[0047] The initial intelligent diagnosis model first extracts the shallow features of the event frame by a convolution block containing 32 3x3 size convolution kernels, then compresses the features by a 2x2 pooling layer, and then further extracts the features by a convolution block containing 16 3x3 size convolution kernels and compresses the features by a 2x2 pooling layer. After that, the flattening operation is performed, and finally the Drop-out layer with a loss rate of 0.5 is used and two fully connected layers are used to output the fault category.

[0048] An intelligent fault diagnosis system for rolling bearings based on enhanced event visual data is implemented by the above method, comprising:

[0049] The data acquisition module: an event camera is used as a sensor to non-contact collect the health status of the rolling bearing, and seven types of original bearing vibration signals of different health statuses are collected as original event stream data, including normal state, outer ring fault, inner ring fault and rolling element fault, and three types of fault are collected under light load and heavy load respectively.

[0050] The data preprocessing module: data preprocessing: first, the original event stream data is screened according to the target area, then the screened target area event stream data is converted into event frame data through two-dimensional feature reconstruction, and then the event frame data is divided into a training data set and a test data set, and the bearing health status is assigned a corresponding category label to construct the original data set; then, additional samples are generated by combining the diffusion model to supplement the original data set; finally, the test data set is added with noise, thereby constructing the final intelligent diagnosis data set;

[0051] The model construction module: the intelligent diagnosis data set is input into the initial intelligent diagnosis model for training, and the mapping relationship between the event frame data and the corresponding label is constructed, and finally the intelligent fault diagnosis model is trained.

[0052] Fault diagnosis module: Based on the intelligent fault diagnosis model, the module performs model tests on the health status of the rolling bearing and finally outputs the diagnostic results.

[0053] The beneficial effects of this invention are as follows:

[0054] This invention proposes an intelligent fault diagnosis method and system for rolling bearings based on enhanced event visual data. After acquiring event data using an event camera, the obtained event stream data is filtered for target regions. Then, the event stream is converted into event frames using the proposed two-dimensional feature construction method. Finally, data augmentation processing is performed on the event frame data using a diffusion model, thus successfully completing the intelligent fault diagnosis of rolling bearings using event data. This invention not only completes the fault diagnosis task for rolling bearings using event data but also improves the related processes of event data generation and conversion. Because this invention uses event stream filtering, it effectively reduces the model's computational load and saves computing resources. Furthermore, because this invention uses a special two-dimensional feature construction method, it effectively improves the accuracy and training stability of the intelligent fault diagnosis model. Finally, because this invention uses a data augmentation algorithm based on a diffusion model, it further improves the accuracy and generalization performance of the intelligent fault diagnosis model. Attached Figure Description

[0055] Figure 1 This is a flowchart of the method of the present invention.

[0056] Figure 2 This is a structural diagram of the initial intelligent diagnostic model of the present invention. Detailed Implementation

[0057] The present invention will now be described in detail with reference to the embodiments and accompanying drawings.

[0058] Reference Figure 1 A method for intelligent fault diagnosis of rolling bearings based on enhanced event visual data includes the following steps:

[0059] Step 1, Data Acquisition: An event camera was used as a sensor to collect the health status of the rolling bearing in a non-contact manner. A total of seven different health statuses were collected as raw bearing vibration signals as raw event stream data. The seven different health statuses include normal status, outer ring fault, inner ring fault and rolling element fault. The three types of faults were collected under light load and heavy load conditions respectively.

[0060] In step 1, data on the health status of the rolling bearing is collected using a non-contact vision sensor such as an event camera. A single event can be represented as e. i :

[0061] e i = (t i , x i , y i , p i )

[0062] where e i represents the i-th event, t i represents the time when the i-th event occurs, x i represents the horizontal coordinate value of the i-th event occurrence position, y i represents the vertical coordinate value of the i-th event occurrence position, and p i represents the polarity of the event occurrence, p i = +1 represents that the pixel brightness increases, i.e., a positive event occurs; p i = -1 represents that the pixel brightness decreases, i.e., a negative event occurs. When the event camera captures a series of event occurrences, events are continuously generated, and the set of events in a period of time can be represented as an event stream E:

[0063]

[0064] where e i represents the i-th event, and N is the total number of events occurring in this period of time;

[0065] Step 2, data preprocessing: first, the original event stream data is filtered according to the target area, then the filtered target area event stream data is converted into event frame data through two-dimensional feature reconstruction, then the event frame data is divided into a training data set and a test data set, and the bearing is given a label of the corresponding category according to the health state, so as to construct the original data set; then, additional samples are generated in combination with the diffusion model to supplement the original data set; finally, the test data set is subjected to noise processing, thereby constructing the final intelligent diagnosis data set;

[0066] In step 2, first, the original event stream data is filtered to retain only the events occurring in the target area, and other background events are removed. The original event stream is denoted as where the coordinate position of a single event can be represented as (x i , y i ), and the initial range of the event position recorded in the field of view of the event camera is:

[0067]

[0068] where x max is the maximum value of the event coordinate that can be recorded by the event camera in the horizontal coordinate x direction, and y maxThe maximum event coordinate that can be recorded by the event camera in the y direction is the longitudinal coordinate. In order to retain events occurring on the rolling bearing in the target area and eliminate meaningless background events, a rectangular area is used to limit the range of event occurrence. After filtering through the rectangular area, an event stream containing only events in the area is generated. The filtered event stream is denoted as:

[0069]

[0070] wherein, e i represents the i-th event, N ROI is the total number of events occurring in the set rectangular area within this period of time, and E ROI is the event stream data of the target area.

[0071]

[0072] wherein, x lower and x upper are the lower and upper bounds of the rectangular area in the x-axis direction, y lower and y upper are the lower and upper bounds of the rectangular area in the y-axis direction.

[0073] The event stream data of the target area is converted into event frame data, which specifically includes:

[0074] The filtered target area event stream data is divided to obtain event stream signals within a certain period of time, and the number of events L required to generate a single sample and the step size D between two samples are set. The event stream data required for the first sample is represented as E1=[e1, e2,..., e L ], the second sample is represented as E2=[e 1+D , e 2+D ,..., e L+D ], the event stream data required for the k-th sample is represented as E k =[e 1+(k-1)D , e 2+(k-1)D ,..., e L+(k-1)D ], and so on to obtain the event stream data required to generate all samples.

[0075] Based on the event stream conversion model, the event stream corresponding to the target area collected by the event camera is converted into an event frame. When a positive event occurs in the event stream, and there is no positive event at the location where the event occurred before this time, the event stream conversion model is represented as:

[0076]

[0077] where I(x, y) represents the pixel value at event frame (x, y), the subscript R, G, B of I represents that the operation needs to be performed on a certain channel of the three channels of the event frame, V base The pixel value is set to half of the highest pixel value of the event frame as the basic pixel value;

[0078] When there is a positive event in the event stream, and the location where the event occurs at this moment has occurred a positive event before this moment, the event stream conversion model is represented as:

[0079]

[0080] where M P and M N respectively represent the number of the most positive events and the number of negative events occurring at the same location in the event stream;

[0081] When there is a negative event in the event stream, and the location where the event occurs at this moment has not occurred a negative event before this moment, the event stream conversion model is represented as:

[0082]

[0083] When there is a negative event in the event stream, and the location where the event occurs at this moment has occurred a negative event before this moment, the event stream conversion model is represented as:

[0084]

[0085] The diffusion model is trained using the converted event frame data. First, a training sample is randomly selected from the converted event frame data, and then a value t is randomly selected from the noise sequence 1-T. After that, the sample is processed by adding noise, that is, the forward noise adding process of the diffusion model, which is represented by the following formula:

[0086]

[0087] where s0represents the initial sample input, s t represents the sample after adding noise t times, q(s t |s0) represents the conditional probability distribution of the sample s t after adding noise t times derived from the sample s0, I is the unit matrix, and is a hyperparameter related to the mean and variance of the normal distribution. After obtaining the data of the sample after adding noise from the above formula, the image is restored through the reverse denoising process of the diffusion model, and the distribution of this process is learned through the deep learning network. The reverse process is represented by the following formula:

[0088]

[0089] where p θ(s t-1 |s t ) indicates that the sample s t After denoising once, the sample s is derived. t-1 The conditional probability distribution, μ θ (s t ,t) is the mean of the normal distribution in the reverse diffusion; after denoising the noisy samples, they are fed into the deep learning training network along with the number of times t is added for iterative processing, and the predicted noise is compared with the real noise and the loss is calculated. Finally, the gradient is calculated and the deep learning network used for denoising the samples is updated. This process is repeated until the network is finally trained; after obtaining the final diffusion model, noisy data is randomly input into it, that is, additional samples are generated. These samples generated by the diffusion model are used as a supplement to the original training set.

[0090] The process involves transforming the event stream data into samples, then dividing it into training and test datasets. Noise is added to the test dataset, and the distribution of the added noise follows the following pattern:

[0091]

[0092] Where σ is the variance of the normal distribution of the added noise, and μ is the mean of the normal distribution of the added noise. After adding noise to the test dataset, the dataset after the above preprocessing can be used for the training and testing of the initial intelligent diagnostic model.

[0093] Step 3, Model Building: Input the intelligent diagnostic dataset into the initial intelligent diagnostic model for training, build the mapping relationship between event frame data and corresponding labels, and finally train to obtain the intelligent fault diagnosis model;

[0094] In step 3, an initial intelligent diagnostic model for learning fault features is constructed. The network specifically consists of convolutional blocks, max-pooling layers, fully connected layers, and drop-out layers; as follows: Figure 2 As shown, the initial intelligent diagnostic model first extracts shallow features of event frames from a convolutional block containing 32 3×3 kernels, then compresses the features with a 2×2 pooling layer, further extracts features from a convolutional block containing 16 3×3 kernels and compresses the features again with a 2×2 pooling layer, then performs a flattening operation, and finally outputs the fault category through a drop-out layer with a dropout rate of 0.5 and two fully connected layers.

[0095] Step 4, Fault Diagnosis: Based on the intelligent fault diagnosis model, the health status of the rolling bearing is tested using the model, and the final diagnosis results are output.

[0096] A rolling bearing intelligent fault diagnosis system based on enhanced event visual data for realizing the above method comprises:

[0097] A data acquisition module: an event camera is used as a sensor to non-contact collect the health state of the rolling bearing, and seven types of original bearing vibration signals of different health states are collected as original event stream data, including normal state, outer ring fault, inner ring fault and rolling element fault, and three types of faults are collected under light load and heavy load conditions respectively.

[0098] A data preprocessing module: data preprocessing: first, the original event stream data is screened according to the target area, then the screened target area event stream data is converted into event frame data through two-dimensional feature reconstruction, then the event frame data is divided into training data set and test data set, and the bearing health state is labeled with corresponding category label to construct the original data set; then, additional samples are generated by combining the diffusion model to supplement the original data set; finally, the test data set is subjected to noise processing, thereby constructing the final intelligent diagnosis data set.

[0099] A model construction module: the intelligent diagnosis data set is input into the initial intelligent diagnosis model for training to construct the mapping relationship between the event frame data and the corresponding label, and finally the intelligent fault diagnosis model is trained.

[0100] A fault diagnosis module: according to the intelligent fault diagnosis model, the health state of the rolling bearing is tested, and finally the diagnosis result is output.

[0101] Experimental example: taking the rolling bearing in the mechanical equipment as an example, the effectiveness of the method is verified based on the rolling bearing experimental data.

[0102] The event camera used is Prophesee 3.1 version event camera. The health state of the rolling bearing under different working conditions is mainly tested, and seven types of motor health states including normal state are recorded in this experiment, the fault types include rolling ball fault, bearing outer ring fault and bearing inner ring fault, and each fault type is divided into light load and heavy load two states, and each working condition is tested at four different speeds, the speeds are 1000rpm, 1500rpm, 2000rpm and 2500rpm. The related parameters for training the diffusion model are as follows: the basic parameters for training the diffusion model are as follows: the training small batch is 32, the training learning rate is 8x10 -5, the total number of training is 30000 times, the gradient accumulation step value is 2, the exponential moving average attenuation value is 0.995, after the data set is divided, it is input into the initial intelligent diagnosis model for training, so that the initial intelligent diagnosis model learns the characteristics of each type of fault, and the related parameter settings are as follows, the learning rate is set to 5x10 -3 , the small batch size is 32, the loss function is cross entropy function, the optimization method adopts the stochastic gradient descent method, the redundant iteration number of training termination is 100, after the training is completed, the intelligent fault diagnosis model can be used to distinguish the fault event frame, each group of experiments is carried out five times, so as to exclude the influence of deep learning randomness as much as possible.

[0103] In addition, another two groups of experiments are selected to compare the diagnosis results of the method of the application, method 1 is not using the event frame representation method proposed by the application, and method 2 is not using the diffusion model for data enhancement processing. The comparison of the diagnosis results of different methods is shown in Table 1, it can be seen that the precision of intelligent fault diagnosis can reach more than 96% by using the event frame data representation method of the application and using the diffusion model for data enhancement, and if the event frame representation method of the application is not used, the precision will decrease a lot, the average precision of 5 experiments of method 1 without using the event frame representation method of the application can only reach 70.76%, which is much lower than the method of the application. If the diffusion model proposed by the application is not used for data enhancement, the precision of the model will also decrease, and the training has non-stationarity, as can be seen from method 2, the average precision of 5 experiments can only reach 78.30%, and the test result has great volatility.

[0104] Table 1 Comparison of diagnosis results of different methods

[0105]

[0106] By comparing the diagnosis effect of the application with method 1 and method 2, it is shown that the application can further improve the effect of fault diagnosis using event camera, the event frame representation method proposed by the application can fully absorb the characteristics of event data, so as to realize better fault diagnosis effect, and after using the diffusion model, not only the overall precision of fault diagnosis is improved, but also the generalization performance of the model is higher.

Claims

1. A method for intelligent fault diagnosis of rolling bearings based on enhanced event visual data, characterized in that, The method comprises the following steps: Step 1, data acquisition: using an event camera as a sensor to non-contact collect the health state of a rolling bearing, and collecting original bearing vibration signals of seven different health states as original event stream data, the seven different health states including a normal state, an outer ring fault, an inner ring fault and a rolling element fault, and the three types of faults are collected under light load and heavy load conditions respectively; Step 2, data preprocessing: first, the original event stream data is screened according to the target area, then the screened target area event stream data is converted into event frame data through two-dimensional feature reconstruction, then the event frame data is divided into a training data set and a test data set, and the bearing is assigned a corresponding category label according to the health state, so as to construct an original data set; Then, additional samples are generated by combining a diffusion model to supplement the original data set; finally, the test data set is subjected to noise processing, thereby constructing a final intelligent diagnosis data set; The step of converting the target area event stream data into event frame data specifically comprises: The target region event stream data after screening is divided to obtain event stream signals in a certain time, and the number of events L required for generating a single sample and the step D between two samples are set, and the event stream data required for the first sample is represented as , the second sample is represented as , and the event stream data required for the kth sample is represented as , and so on to obtain the event stream data required for generating all samples; Based on an event stream conversion model, the corresponding target area event stream collected by the event camera is converted into an event frame; when a positive event occurs in the event stream, and no positive event has occurred at the position of the event before this moment, the event stream conversion model is represented as: wherein, represents a pixel value at an event frame (x, y), the subscript R, G, B of the pixel value represents that the operation needs to be performed on a certain channel of the three channels of the event frame separately, the pixel value is set to half of the highest pixel value of the event frame as a basic pixel value; When a positive event occurs in the event stream, and a positive event has occurred at the position of the event before this moment, the event stream conversion model is represented as: wherein with respectively represent the number of positive events and negative events that occur most frequently at the same location in the event stream; When a negative event occurs in the event stream, and no negative event has occurred at the position of the event before this moment, the event stream conversion model is represented as: When a negative event occurs in the event stream, and a negative event has occurred at the position of the event before this moment, the event stream conversion model is represented as: ； Step 3, model construction: inputting the intelligent diagnosis data set into an initial intelligent diagnosis model for training to construct a mapping relationship between the event frame data and the corresponding label, and finally training an intelligent fault diagnosis model; Step 4, fault diagnosis: according to the intelligent fault diagnosis model, the health state of the rolling bearing is tested by the model, and finally the diagnosis result is output.

2. The method of claim 1, wherein: In step 1, data acquisition is performed with an event camera, where a single event represents : wherein represents the ith event, represents the time at which the ith event occurs, represents the horizontal coordinate value of the location at which the ith event occurs, represents the vertical coordinate value of the location at which the ith event occurs, represents the polarity of the event, = +1 represents an increase in the brightness of the pixel, i.e. a positive event has occurred; = -1 represents a decrease in the brightness of the pixel, i.e. a negative event has occurred; when an event camera captures a series of events, it continuously generates events, and the set of events in a period of time can be represented as an event stream : wherein, is the total number of events that occurred during this time period.

3. The method of claim 1, wherein: In step 2, only the events occurring in the target area are retained, and other background events are removed.

4. The method of claim 1, wherein: In step 2, a diffusion model is trained using the converted event frame data.

5. The method of claim 4, wherein: Using the diffusion model, after randomly inputting any noise data, additional samples are generated to supplement the original data set.

6. The method of claim 1, wherein: In step 3, an initial intelligent diagnosis model for learning fault features is constructed, which specifically comprises a convolution block, a maximum pooling layer, a full connection layer and a Drop-out layer.

7. The method of claim 6, wherein: The initial intelligent diagnosis model first extracts the shallow features of the event frame by a convolution block containing 32 3x3 size convolution kernels, then compresses the features by a 2x2 pooling layer, further extracts the features by a convolution block containing 16 3x3 size convolution kernels, and then compresses the features by a 2x2 pooling layer, then performs a flattening operation, and finally outputs the fault category through a Drop-out layer with a loss rate of 0.5 and two full connection layers.

8. An intelligent fault diagnosis system for rolling bearings based on enhanced visual data of events, implementing the method according to any one of claims 1 to 7, characterized in that, The application relates to a rolling bearing intelligent fault diagnosis method and device. The data acquisition module: an event camera is used as a sensor to non-contact collect the health state of a rolling bearing, and original bearing vibration signals of seven different health states are collected as original event stream data, the seven different health states including a normal state, outer ring fault, inner ring fault and rolling element fault, and three types of faults are collected under light load and heavy load conditions respectively; The data preprocessing module: data preprocessing: firstly, the original event stream data is screened according to a target region, then the screened target region event stream data is converted into event frame data through two-dimensional feature reconstruction, then the event frame data is divided into a training data set and a test data set, and the bearing is given a corresponding category label according to the health state, so as to construct an original data set; Then, additional samples are generated in combination with a diffusion model to supplement the original data set; finally, the test data set is subjected to noise processing, so as to construct a final intelligent diagnosis data set; The model construction module: the intelligent diagnosis data set is input into an initial intelligent diagnosis model for training, a mapping relationship between event frame data and corresponding labels is constructed, and finally an intelligent fault diagnosis model is trained; The fault diagnosis module: according to the intelligent fault diagnosis model, the health state of the rolling bearing is subjected to model testing, and finally a diagnosis result is output.

Images