CN116380466A - Rolling bearing intelligent fault diagnosis method and system based on enhanced event visual data - Google Patents
Rolling bearing intelligent fault diagnosis method and system based on enhanced event visual data Download PDFInfo
- Publication number
- CN116380466A CN116380466A CN202310473322.7A CN202310473322A CN116380466A CN 116380466 A CN116380466 A CN 116380466A CN 202310473322 A CN202310473322 A CN 202310473322A CN 116380466 A CN116380466 A CN 116380466A
- Authority
- CN
- China
- Prior art keywords
- event
- data
- model
- data set
- event stream
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000003745 diagnosis Methods 0.000 title claims abstract description 85
- 238000000034 method Methods 0.000 title claims abstract description 64
- 238000005096 rolling process Methods 0.000 title claims abstract description 47
- 230000000007 visual effect Effects 0.000 title claims abstract description 15
- 238000012549 training Methods 0.000 claims abstract description 46
- 230000036541 health Effects 0.000 claims abstract description 41
- 238000009792 diffusion process Methods 0.000 claims abstract description 31
- 238000012360 testing method Methods 0.000 claims abstract description 30
- 238000012216 screening Methods 0.000 claims abstract description 9
- 238000013507 mapping Methods 0.000 claims abstract description 8
- 238000006243 chemical reaction Methods 0.000 claims description 17
- 238000007781 pre-processing Methods 0.000 claims description 10
- 239000000284 extract Substances 0.000 claims description 9
- 238000010276 construction Methods 0.000 claims description 8
- 238000011176 pooling Methods 0.000 claims description 7
- 230000001502 supplementing effect Effects 0.000 claims description 7
- 238000012545 processing Methods 0.000 abstract description 9
- 238000013527 convolutional neural network Methods 0.000 abstract 1
- 230000008569 process Effects 0.000 description 10
- 238000013135 deep learning Methods 0.000 description 7
- 230000000694 effects Effects 0.000 description 7
- 238000002474 experimental method Methods 0.000 description 6
- 238000009434 installation Methods 0.000 description 4
- 230000001133 acceleration Effects 0.000 description 3
- 230000007547 defect Effects 0.000 description 3
- 238000005259 measurement Methods 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 230000003862 health status Effects 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 239000013589 supplement Substances 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M13/00—Testing of machine parts
- G01M13/04—Bearings
- G01M13/045—Acoustic or vibration analysis
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Landscapes
- Physics & Mathematics (AREA)
- Acoustics & Sound (AREA)
- General Physics & Mathematics (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The intelligent fault diagnosis method and system for the rolling bearing based on the enhanced event visual data comprises the steps of firstly adopting an event camera as a sensor to collect the health state of the rolling bearing in a non-contact manner; screening original event stream data according to a target area, converting the data into event frame data through processing, carrying out data enhancement processing by combining a diffusion model, marking enhancement data with corresponding labels according to the health state of a bearing, dividing a data set into a training data set and a test data set, and carrying out noise adding processing on the test data set; inputting the training data set and the corresponding label into a convolutional neural network model for training, and constructing a mapping relation between event frame data and the corresponding label; finally, diagnosing the health state of the rolling bearing according to the mapping relation model of the event frame data and the corresponding label; the invention can adapt to the non-contact sensor to realize the fault diagnosis of the rolling bearing, and improves the precision and generalization performance of fault diagnosis by using event data.
Description
Technical Field
The invention relates to the technical field of mechanical fault diagnosis, in particular to an intelligent fault diagnosis method and system for a rolling bearing based on enhanced event visual data.
Background
With the rapid development of intelligent motor systems, rotating machines closely related to the intelligent motor systems are widely applied, and the normal operation of the rotating machines is closely related to the health state of bearings, and the health state of the bearings directly reflects whether the whole rotating machine system can normally operate. However, since the bearing needs to operate in a high-speed, high-temperature and high-pressure environment for a long period of time, the possibility of failure of the bearing during operation is very high, which leads to damage to mechanical equipment and even economic, environmental and personal safety accidents. Monitoring and diagnosing the health of rolling bearings is therefore a very important task.
In engineering practice, the vibration signal can effectively reflect the health state of the bearing, and is widely applied to various health state monitoring tasks. For the collection of the vibration signals of the rolling bearing, an acceleration sensor is the most common method, however, because the acceleration sensor needs to be directly connected with a vibration machine, an installation position needs to be reserved and an installation space is needed, meanwhile, the probability of sensor failure is greatly increased because the sensor can vibrate at a high frequency along with the machine, and therefore, the application of the acceleration sensor in engineering practice cannot achieve a good effect.
With the development of measurement technology, detection technology is rapidly developing in the directions of non-contact, high precision and high speed, and part of non-contact sensors have also been used for signal acquisition of rolling bearings, and currently, non-contact sensors for monitoring vibration mainly include: the eddy current sensor, the laser vibration meter and the high-speed camera can finish the measurement task without directly contacting with a vibrating machine, but the defects are obvious; the eddy current sensor has strict requirements on the shape and the material of the object to be tested, the test system is more expensive, and the installation mode and the position are also more strict; the laser vibration meter has the problems of higher cost and strict installation position; the high-speed camera is easily affected by ambient light, the calculated amount is large, useless background data occupies excessive calculation resources, and the measurement vibration is not accurate enough.
The event camera is used as a novel non-contact visual sensor, and has the characteristics of high time resolution and high dynamic range, so that the event camera is used for measuring vibration, but at present, the processing of event data is not well adapted to the overall scheme of rolling bearing health state acquisition, conversion, processing and diagnosis, and the related fault diagnosis flow is not perfect enough, so that the engineering practical application value is not high.
In summary, the conventional contact and non-contact sensors have many defects, and the current rolling bearing fault diagnosis method based on the event camera is not complete in flow, the overall fault effect is not good enough, and the generalization performance and the precision of the fault diagnosis model have great improvement space.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention aims to provide an intelligent fault diagnosis method and system for a rolling bearing based on enhanced event visual data, which can be adapted to a non-contact sensor to realize fault diagnosis of the rolling bearing and improve the precision and generalization performance of fault diagnosis by using the event data.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:
an intelligent fault diagnosis method for a rolling bearing based on enhanced event visual data comprises the following steps:
step 1, data acquisition: the method comprises the steps that an event camera is used as a sensor to collect health states of the rolling bearing in a non-contact mode, original bearing vibration signals of seven different health states are collected as original event stream data, the seven different health states comprise a normal state, an outer ring fault, an inner ring fault and a rolling body fault, and three fault types are collected respectively in a light load state and a heavy load state;
step 2, data preprocessing: firstly screening original event stream data according to a target area, converting the screened target area event stream data into event frame data through two-dimensional feature reconstruction, dividing the event frame data into a training data set and a test data set, and endowing the event frame data with labels of corresponding categories according to the health state of a bearing so as to construct the original data set; then generating additional samples for supplementing the original data set by combining the diffusion model; finally, the test data set is subjected to noise adding treatment, so that a final intelligent diagnosis data set is constructed;
step 3, model construction: inputting the intelligent diagnosis data set into an initial intelligent diagnosis model for training, constructing a mapping relation between event frame data and corresponding labels, and finally training to obtain an intelligent fault diagnosis model;
In step 1, data acquisition is performed using an event camera, wherein a single event is denoted as e i ：
e i ＝(t i ,x i ,y i ,p i )
Wherein e i Represents the ith event, t i Indicating the time at which the ith event occurred, x i An abscissa value, y, representing an i-th event occurrence position i Ordinate value, p, representing the position of occurrence of the ith event i Indicating the polarity of the occurrence of an event, p i = +1 represents an increase in pixel brightness, i.e. a positive event occurs; p is p i -1 represents a decrease in pixel brightness, i.e. a negative event has occurred; after capturing a series of events by the event camera, the events are continuously generated, and the set of the events is expressed as an event stream E in a period of time:
where N is the total number of events that occur during this period of time.
In step 2, firstly, for original event stream data, only the events occurring in a target area are reserved, and other background events are removed; the original event stream is recorded as
Wherein the coordinate position at which a single event occurs is expressed as (x i ，y i ) The initial range of recorded event positions within the event camera field of view is:
wherein x is max Maximum value of event coordinates, y, which can be recorded by the event camera in the x direction of the abscissa max The maximum value of the event coordinates which can be recorded by the event camera in the y direction of the ordinate is obtained; the method comprises the steps that a rectangular area is adopted to limit the occurrence range of an event, a section of event stream only comprising events in the area is generated after the event stream is screened by the rectangular area, and the screened event stream is recorded as:
wherein e is i Represents the ith event, N ROI E, for the total number of events occurring in the set rectangular area within the period of time ROI The position coordinates of the event in (a) satisfy:
wherein x is lower And x upper Is the lower bound and the upper bound of the rectangular frame area in the x-axis direction, y lower And y is upper Is the lower and upper bounds of the rectangular box area in the y-axis direction.
In step 2, converting event stream data of a target area into event frame data, specifically including:
dividing the screened event stream data of the target area, obtaining event stream signals within a certain time, setting the number L of events required for generating a single sample and the step length D between two samples, and expressing the event stream data required for the first sample as E 1 ＝[e 1 ，e 2 ，...，e L ]The second sample is denoted as E 2 ＝[e 1+D ，e 2+D ，...，e L+D ]The event stream data required for the kth sample is denoted as E k ＝[e 1+(k-1)D ，e 2+(k-1)D ，...，e L+(k-1)D ]The event stream data required to generate all samples is obtained by such a push.
Based on an event stream conversion model, converting the event stream of the corresponding target area acquired by the event camera into an event frame; when a positive event occurs in the event stream and the positive event has not occurred at a position before the moment, the event stream conversion model is expressed as:
where I (x, y) represents the pixel value at the event frame (x, y), the subscript R, G, B of I represents the need to operate on one of the three channels of the event frame alone, V base Setting the pixel value to be half of the highest pixel value of the event frame as the base pixel value;
when a positive event occurs in the event stream and the position of the event occurs before the moment, the event stream conversion model is expressed as:
wherein M is P And M is as follows N Respectively representing the number of positive events and negative events which occur most times at the same position in the event stream;
when a negative event occurs in the event stream and the position of the event occurs before the moment is not negative, the event stream conversion model is expressed as:
when a negative event occurs in the event stream and the position of the event occurs before the moment is over-negative, the event stream conversion model is expressed as:
in step 2, training a diffusion model by using the converted event frame data, randomly selecting a training sample from the converted event frame data, randomly selecting a value T from the noise adding sequence 1-T, and then performing noise adding processing on the sample, namely performing a forward noise adding process of the diffusion model, wherein the forward noise adding process is represented by the following formula:
wherein s is 0 Representing an initial sample of the input s t Represents the samples after the addition of the noise t times, q(s) t |s 0 ) Representing the result of the sample s 0 Sample s after adding noise t times t I is the identity matrix,
obtaining a super parameter for the mean value and variance related to normal distribution; after the data of the sample after the noise superposition is obtained by the method, the image is restored by the reverse denoising process of the diffusion model, and the distribution of the process is learned by the deep learning network, wherein the reverse process is expressed as the following formula:
wherein p is θ (s t-1 |s t ) Representing the result of the sample s t Sample s after denoising 1 time is deduced t-1 Conditional probability distribution, mu θ (s t T) is the mean value of the normal distribution in the reverse diffusion; after denoising the denoised sample, sending the denoised sample and the denoised times t into a deep learning training network for iteration, comparing the predicted noise with the real noise, calculating the loss, calculating the gradient, updating the deep learning network for denoising the sample, and repeating until the final training of the network is completed; after the final diffusion model is obtained, noise data is randomly input into the diffusion model, namely additional samples are generated, and the samples generated by the diffusion model are used as supplements of the original training set.
The operation of converting event stream data into samples is completed, then the event stream data 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 obeys to:
wherein sigma is the variance of the normal distribution of the added noise, mu is the mean value of the normal distribution of the added noise, and after adding noise to the test data set, the data set after the pretreatment can be used for training and testing the initial intelligent diagnosis model.
In step 3, an initial intelligent diagnosis model for learning fault characteristics is constructed, and specifically comprises a convolution block, a maximum pooling layer, a full connection layer and a Drop-out layer.
The initial intelligent diagnosis model firstly extracts shallow layer characteristics of event frames from a convolution block containing convolution kernels with the size of 32 pieces of 3 multiplied by 3, then extracts characteristics further from a 2 multiplied by 2 pooling layer laminated characteristics, then extracts characteristics further from a convolution block containing convolution kernels with the size of 16 pieces of 3 multiplied by 3, and then passes through the 2 multiplied by 2 pooling layer laminated characteristics, then performs flattening operation, and finally outputs fault categories through a Drop-out layer with the loss rate of 0.5 and through two fully connected layers.
An intelligent fault diagnosis system for a rolling bearing based on enhanced event visual data, which realizes the method, comprises the following steps:
and a data acquisition module: the method comprises the steps that an event camera is used as a sensor to collect health states of the rolling bearing in a non-contact mode, original bearing vibration signals of seven different health states are collected as original event stream data, the seven different health states comprise a normal state, an outer ring fault, an inner ring fault and a rolling body fault, and three fault types are collected respectively in a light load state and a heavy load state;
and a data preprocessing module: data preprocessing: firstly screening original event stream data according to a target area, converting the screened target area event stream data into event frame data through two-dimensional feature reconstruction, dividing the event frame data into a training data set and a test data set, and endowing the event frame data with labels of corresponding categories according to the health state of a bearing so as to construct the original data set; then generating additional samples for supplementing the original data set by combining the diffusion model; finally, the test data set is subjected to noise adding treatment, so that a final intelligent diagnosis data set is constructed;
model construction module: inputting the intelligent diagnosis data set into an initial intelligent diagnosis model for training, constructing a mapping relation between event frame data and corresponding labels, and finally training to obtain an intelligent fault diagnosis model;
and a fault diagnosis module: and performing model test on the health state of the rolling bearing according to the intelligent fault diagnosis model, and finally outputting a diagnosis result.
The beneficial effects of the invention are as follows:
the invention provides an intelligent fault diagnosis method and system for a rolling bearing based on enhanced event visual data, which are characterized in that event data are acquired through an event camera, then target area screening is carried out on the obtained event stream data, then the event stream is converted into event frames through the two-dimensional feature construction method, and finally data enhancement processing is carried out on the event frame data through a diffusion model, so that intelligent fault diagnosis of the rolling bearing is successfully completed through the event data. The method not only utilizes the event data to complete the fault diagnosis task for the rolling bearing, but also improves the related flow of event data generation and conversion, and as the method adopts the event flow screening operation, the method can effectively reduce the model operand and save the calculation resources; in addition, the invention adopts a special two-dimensional feature construction method, so that the accuracy and the training stability of the intelligent fault diagnosis model can be effectively improved; finally, as the data enhancement algorithm based on the diffusion model is adopted, the accuracy and generalization performance of the intelligent fault diagnosis model are further improved.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
FIG. 2 is a block diagram of an initial intelligent diagnostic model of the present invention.
Detailed Description
The present invention will be described in detail with reference to the following examples and the accompanying drawings.
Referring to fig. 1, a rolling bearing intelligent fault diagnosis method based on enhanced event visual data includes the steps of:
step 1, data acquisition: the method comprises the steps that an event camera is used as a sensor to collect health states of the rolling bearing in a non-contact mode, original bearing vibration signals of seven different health states are collected as original event stream data, the seven different health states comprise a normal state, an outer ring fault, an inner ring fault and a rolling body fault, and three fault types are collected respectively in a light load state and a heavy load state;
in step 1, for the health status information of the rolling bearing, a non-contact visual sensor such as an event camera is used for data acquisition, and for the health status information of the rolling bearing, a non-contact visual sensor such as an event camera is used for data acquisition, wherein a single event can be expressed as e i ：
e i ＝(t i ,x i ,y i ,p i )
Wherein e i Represents the ith event, t i Represent the firstTime of occurrence of i events, x i An abscissa value, y, representing an i-th event occurrence position i Ordinate value, p, representing the position of occurrence of the ith event i Indicating the polarity of the occurrence of an event, p i = +1 represents an increase in pixel brightness, i.e. a positive event occurs; p is 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, the events are generated continuously, and the set of the events can be expressed as an event stream E:
wherein e is i Representing the ith event, N being the total number of events occurring during the period of time;
step 2, data preprocessing: firstly screening original event stream data according to a target area, converting the screened target area event stream data into event frame data through two-dimensional feature reconstruction, dividing the event frame data into a training data set and a test data set, and endowing the event frame data with labels of corresponding categories according to the health state of a bearing so as to construct the original data set; then generating additional samples for supplementing the original data set by combining the diffusion model; finally, the test data set is subjected to noise adding treatment, so that a final intelligent diagnosis data set is constructed;
in step 2, firstly, for original event stream data, only the events occurring in a target area are reserved, and other meaningless background events are removed; the original event stream is recorded as
Wherein the coordinate locations where the individual events occur can be expressed as (x) i ,y i ) The initial range of recorded event positions within the event camera field of view is:
wherein x is max Maximum value of event coordinates, y, which can be recorded by the event camera in the x direction of the abscissa max For the maximum value of the event coordinates which can be recorded by the event camera in the y direction of the ordinate, in order to keep the event occurring on the rolling bearing of the target area and remove nonsensical background events, a rectangular area is adopted to limit the occurrence range of the event, a section of event stream only comprising the event in the area is generated after the event stream is screened by the rectangular area, and the screened event stream is marked as follows:
wherein e is i Represents the ith event, N ROI E, for the total number of events occurring in the set rectangular area within the period of time ROI The position coordinates of the event in (a) satisfy:
wherein x is lower And x upper Is the lower bound and the upper bound of the rectangular frame area in the x-axis direction, y lower And y is upper A lower boundary and an upper boundary of the rectangular frame area in the y-axis direction;
converting event stream data of a target area into event frame data, specifically comprising:
dividing the screened event stream data of the target area, obtaining event stream signals within a certain time, setting the number L of events required for generating a single sample and the step length D between two samples, and expressing the event stream data required for the first sample as E 1 ＝[e 1 ，e 2 ，...，e L ]The second sample is denoted as E 2 ＝[e 1+D ，e 2+D ，...，e L+D ]The event stream data required for the kth sample is denoted as E k ＝[e 1+(k-1)D ，e 2+(k-1)D ，...，e L+(k-1)D ]Obtaining all samples required for generating by such a pushIs provided.
Based on an event stream conversion model, converting the event stream of the corresponding target area acquired by the event camera into an event frame; when a positive event occurs in the event stream and the positive event has not occurred at a position before the moment, the event stream conversion model is expressed as:
where I (x, y) represents the pixel value at the event frame (x, y), the subscript R, G, B of I represents the need to operate on one of the three channels of the event frame alone, V base Setting the pixel value to be half of the highest pixel value of the event frame as the base pixel value;
when a positive event occurs in the event stream and the position of the event occurs before the moment, the event stream conversion model is expressed as:
wherein M is P And M is as follows N Respectively representing the number of positive events and negative events which occur most times at the same position in the event stream;
when a negative event occurs in the event stream and the position of the event occurs before the moment is not negative, the event stream conversion model is expressed as:
when a negative event occurs in the event stream and the position of the event occurs before the moment is over-negative, the event stream conversion model is expressed as:
training a diffusion model by adopting the converted event frame data, randomly selecting a training sample from the converted event frame data, randomly selecting a value T from a noise adding sequence 1-T, and then carrying out noise adding processing on the sample, namely carrying out a forward noise adding process of the diffusion model, wherein the forward noise adding process is represented by the following formula:
wherein s is 0 Representing an initial sample of the input s t Represents the samples after the addition of the noise t times, q(s) t |s 0 ) Representing the result of the sample s 0 Sample s after adding noise t times t I is the identity matrix,
obtaining a super parameter for the mean value and variance related to normal distribution; after the data of the sample after the noise superposition is obtained by the method, the image is restored by the reverse denoising process of the diffusion model, and the distribution of the process is learned by the deep learning network, wherein the reverse process is expressed as the following formula:
wherein p is θ (s t-1 |s t ) Representing the result of the sample s t Sample s after denoising 1 time is deduced t-1 Conditional probability distribution, mu θ (s t T) is the mean value of the normal distribution in the reverse diffusion; after denoising the denoised sample, sending the denoised sample and the denoised times t into a deep learning training network for iteration, comparing the predicted noise with the real noise, calculating the loss, calculating the gradient, updating the deep learning network for denoising the sample, and repeating until the final training of the network is completed; after the final diffusion model is obtained, noise data is randomly input into the diffusion model, namely additional samples are generated, and the diffusion model is used for generating the noise dataThe generated samples supplement the original training set.
The operation of converting event stream data into samples is completed, then the event stream data 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 obeys to:
wherein sigma is the variance of the normal distribution of the added noise, mu is the mean value of the normal distribution of the added noise, and after adding noise to the test data set, the data set after the pretreatment can be used for training and testing the initial intelligent diagnosis model.
Step 3, model construction: inputting the intelligent diagnosis data set into an initial intelligent diagnosis model for training, constructing a mapping relation between event frame data and corresponding labels, and finally training to obtain an intelligent fault diagnosis model;
in step 3, an initial intelligent diagnosis model for learning fault characteristics is constructed, and the network specifically comprises a convolution block, a maximum pooling layer, a full connection layer and a Drop-out layer; as shown in fig. 2, the initial smart diagnostic model first extracts shallow features of event frames from a convolution block containing 32 convolution kernels of 3×3 size, then extracts features from a 2×2 pooled layer, then further extracts features from a convolution block containing 16 convolution kernels of 3×3 size, and then passes through a 2×2 pooled layer, then performs a flattening operation, and finally outputs fault categories through Drop-out layer with a loss rate of 0.5 and through two fully connected layers.
An intelligent fault diagnosis system for a rolling bearing based on enhanced event visual data, which realizes the method, comprises the following steps:
and a data acquisition module: the method comprises the steps that an event camera is used as a sensor to collect health states of the rolling bearing in a non-contact mode, original bearing vibration signals of seven different health states are collected as original event stream data, the seven different health states comprise a normal state, an outer ring fault, an inner ring fault and a rolling body fault, and three fault types are collected respectively in a light load state and a heavy load state;
and a data preprocessing module: data preprocessing: firstly screening original event stream data according to a target area, converting the screened target area event stream data into event frame data through two-dimensional feature reconstruction, dividing the event frame data into a training data set and a test data set, and endowing the event frame data with labels of corresponding categories according to the health state of a bearing so as to construct the original data set; then generating additional samples for supplementing the original data set by combining the diffusion model; finally, the test data set is subjected to noise adding treatment, so that a final intelligent diagnosis data set is constructed;
model construction module: inputting the intelligent diagnosis data set into an initial intelligent diagnosis model for training, constructing a mapping relation between event frame data and corresponding labels, and finally training to obtain an intelligent fault diagnosis model;
and a fault diagnosis module: and performing model test on the health state of the rolling bearing according to the intelligent fault diagnosis model, and finally outputting a diagnosis result.
Experimental example: the method provided by the invention is used for verifying the effectiveness of the method based on the experimental data of the rolling bearing by taking the rolling bearing in mechanical equipment as a case.
The event camera model used was a pu-femee version 3.1 event camera. The method mainly aims at the health states of the rolling bearing under different working conditions, the experiment records seven motor health states including normal states, fault types comprise ball faults, bearing outer ring faults and bearing inner ring faults, each fault type is divided into a light load state and a heavy load state, each working condition is subjected to the experiment under four different rotating speeds, and the rotating speeds are respectively 1000rpm, 1500rpm, 2000rpm and 2500 rpm. Preprocessing data, and training relevant parameters in a diffusion model to be: training deviceThe basic parameters of the diffusion model are as follows: training small batch is 32, training learning rate is 8×10 -5 The total training times are 30000 times, the gradient accumulation step value is 2, the exponential moving average attenuation value is 0.995, after the data set is divided, the data set is input into an initial intelligent diagnosis model for training, the initial intelligent diagnosis model learns the characteristics of each type of faults, the training related parameters are set as follows, and the learning rate is set to be 5 multiplied by 10 -3 The small batch size is 32, the loss function is a cross entropy function, the optimization method adopts a random gradient descent method, the redundant iteration number of training termination is 100, after training is completed, the intelligent fault diagnosis model can be used for distinguishing fault event frames, and each group of experiments is carried out five times so as to eliminate the influence of the deep learning randomness as much as possible.
In addition, two other groups of experiments are selected to compare the diagnosis results of the method, the method 1 is an event frame representation method which is not provided by the invention, and the method 2 is a method for carrying out data enhancement processing without using a diffusion model. As can be seen from the comparison of the diagnostic results of the different methods shown in Table 1, the accuracy of the intelligent fault diagnosis can reach more than 96% after the event frame data representation method of the invention is adopted and the data enhancement is carried out by adopting the diffusion model, and if the event frame representation method of the invention is not adopted, the accuracy can be greatly reduced, and the average accuracy of 5 experiments without adopting the event frame representation method of the invention in the method 1 can only reach 70.76%, which is far lower than the method of the invention. If the diffusion model provided by the invention is not adopted for data enhancement, the accuracy of the model is also reduced, the training is unstable, the average accuracy of 5 experiments can only reach 78.30% as can be seen from the method 2, and the test result has larger fluctuation.
Table 1 comparison of diagnostic results for different methods
By comparing the diagnosis effects of the invention with the diagnosis effects of the method 1 and the diagnosis effects of the method 2, the invention can further improve the effect of fault diagnosis by adopting the event camera, the event frame representation method provided by the invention can fully absorb the characteristics of event data, thereby realizing better fault diagnosis effect, improving the overall accuracy of fault diagnosis after using the diffusion model, and having higher generalization performance of the model.
Claims (9)
1. An intelligent fault diagnosis method for a rolling bearing based on enhanced event visual data is characterized by comprising the following steps:
step 1, data acquisition: the method comprises the steps that an event camera is used as a sensor to collect health states of the rolling bearing in a non-contact mode, original bearing vibration signals of seven different health states are collected as original event stream data, the seven different health states comprise a normal state, an outer ring fault, an inner ring fault and a rolling body fault, and three fault types are collected respectively in a light load state and a heavy load state;
step 2, data preprocessing: firstly screening original event stream data according to a target area, converting the screened target area event stream data into event frame data through two-dimensional feature reconstruction, dividing the event frame data into a training data set and a test data set, and endowing the event frame data with labels of corresponding categories according to the health state of a bearing so as to construct the original data set; then generating additional samples for supplementing the original data set by combining the diffusion model; finally, the test data set is subjected to noise adding treatment, so that a final intelligent diagnosis data set is constructed;
step 3, model construction: inputting the intelligent diagnosis data set into an initial intelligent diagnosis model for training, constructing a mapping relation between event frame data and corresponding labels, and finally training to obtain an intelligent fault diagnosis model;
step 4, fault diagnosis: and performing model test on the health state of the rolling bearing according to the intelligent fault diagnosis model, and finally outputting a diagnosis result.
2. The method according to claim 1, characterized in that: in step 1, an event camera is used for enteringData acquisition, wherein a single event is denoted as e i ：
e i ＝(t i ,x i ,y i ,p i )
Wherein e i Represents the ith event, t i Indicating the time at which the ith event occurred, x i An abscissa value, y, representing an i-th event occurrence position i Ordinate value, p, representing the position of occurrence of the ith event i Indicating the polarity of the occurrence of an event, p i = +1 represents an increase in pixel brightness, i.e. a positive event occurs; p is 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, the events are generated continuously, and the set of the events can be expressed as an event stream E:
where N is the total number of events that occur during this period of time.
3. The method according to claim 1, characterized in that: in step 2, only the events occurring in the target area are reserved for the original event stream data, and other background events are removed.
4. The method according to claim 1, characterized in that: in step 2, the step of converting the event stream data of the target area into an event frame specifically includes:
dividing the screened event stream data of the target area, obtaining event stream signals within a certain time, setting the number L of events required for generating a single sample and the step length D between two samples, and expressing the event stream data required for the first sample as E 1 ＝[e 1 ,e 2 ,…,e L ]The second sample is denoted as E 2 ＝[e 1+D ,e 2+D ,…,e L+D ]The event stream data required for the kth sample is denoted as E k ＝[e 1+(k-1)D ,e 2+(k-1)D ,…,e L+(k-1)D ]Obtaining event stream data required for generating all samples by such a push;
based on an event stream conversion model, converting the event stream of the corresponding target area acquired by the event camera into an event frame; when a positive event occurs in the event stream and the positive event has not occurred at a position before the moment, the event stream conversion model is expressed as:
where I (x, y) represents the pixel value at the event frame (x, y), the subscript R, G, B of I represents the need to operate on one of the three channels of the event frame alone, V base Setting the pixel value to be half of the highest pixel value of the event frame as the base pixel value;
when a positive event occurs in the event stream and the position of the event occurs before the moment, the event stream conversion model is expressed as:
wherein M is P And M is as follows N Respectively representing the number of positive events and negative events which occur most times at the same position in the event stream;
when a negative event occurs in the event stream and the position of the event occurs before the moment is not negative, the event stream conversion model is expressed as:
when a negative event occurs in the event stream and the position of the event occurs before the moment is over-negative, the event stream conversion model is expressed as:
5. the method according to claim 1, characterized in that: in step 2, a diffusion model is trained using the converted event frame data.
6. The method according to claim 5, wherein: using the diffusion model, after random input of any noise data, additional samples are generated for supplementing the original data set.
7. The method according to claim 1, characterized in that: in step 3, an initial intelligent diagnosis model for learning fault characteristics is constructed, and specifically comprises a convolution block, a maximum pooling layer, a full connection layer and a Drop-out layer.
8. The method according to claim 7, wherein: the initial intelligent diagnosis model firstly extracts shallow layer characteristics of event frames from a convolution block containing convolution kernels with the size of 32 pieces of 3 multiplied by 3, then extracts characteristics further from a 2 multiplied by 2 pooling layer laminated characteristics, then extracts characteristics further from a convolution block containing convolution kernels with the size of 16 pieces of 3 multiplied by 3, and then passes through the 2 multiplied by 2 pooling layer laminated characteristics, then performs flattening operation, and finally outputs fault categories through a Drop-out layer with the loss rate of 0.5 and through two fully connected layers.
9. A rolling bearing intelligent fault diagnosis system based on enhanced event visual data implementing the method of any one of claims 1-8, comprising:
and a data acquisition module: the method comprises the steps that an event camera is used as a sensor to collect health states of the rolling bearing in a non-contact mode, original bearing vibration signals of seven different health states are collected as original event stream data, the seven different health states comprise a normal state, an outer ring fault, an inner ring fault and a rolling body fault, and three fault types are collected respectively in a light load state and a heavy load state;
and a data preprocessing module: data preprocessing: firstly screening original event stream data according to a target area, converting the screened target area event stream data into event frame data through two-dimensional feature reconstruction, dividing the event frame data into a training data set and a test data set, and endowing the event frame data with labels of corresponding categories according to the health state of a bearing so as to construct the original data set; then generating additional samples for supplementing the original data set by combining the diffusion model; finally, the test data set is subjected to noise adding treatment, so that a final intelligent diagnosis data set is constructed;
model construction module: inputting the intelligent diagnosis data set into an initial intelligent diagnosis model for training, constructing a mapping relation between event frame data and corresponding labels, and finally training to obtain an intelligent fault diagnosis model;
and a fault diagnosis module: and performing model test on the health state of the rolling bearing according to the intelligent fault diagnosis model, and finally outputting a diagnosis result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310473322.7A CN116380466A (en) | 2023-04-27 | 2023-04-27 | Rolling bearing intelligent fault diagnosis method and system based on enhanced event visual data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310473322.7A CN116380466A (en) | 2023-04-27 | 2023-04-27 | Rolling bearing intelligent fault diagnosis method and system based on enhanced event visual data |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116380466A true CN116380466A (en) | 2023-07-04 |
Family
ID=86963468
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310473322.7A Pending CN116380466A (en) | 2023-04-27 | 2023-04-27 | Rolling bearing intelligent fault diagnosis method and system based on enhanced event visual data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116380466A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117473397A (en) * | 2023-12-25 | 2024-01-30 | 清华大学 | Diffusion model data enhancement-based emotion recognition method and system |
-
2023
- 2023-04-27 CN CN202310473322.7A patent/CN116380466A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117473397A (en) * | 2023-12-25 | 2024-01-30 | 清华大学 | Diffusion model data enhancement-based emotion recognition method and system |
| CN117473397B (en) * | 2023-12-25 | 2024-03-19 | 清华大学 | Diffusion model data enhancement-based emotion recognition method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Hu et al. | Data-driven fault diagnosis method based on compressed sensing and improved multiscale network | |
| Meng et al. | Data segmentation and augmentation methods based on raw data using deep neural networks approach for rotating machinery fault diagnosis | |
| Chengyang et al. | Spindle thermal error prediction approach based on thermal infrared images: A deep learning method | |
| CN109858352B (en) | Fault diagnosis method based on compressed sensing and improved multi-scale network | |
| Wang et al. | Self-supervised signal representation learning for machinery fault diagnosis under limited annotation data | |
| CN112084974A (en) | Multi-label rolling bearing fault diagnosis method based on meta-learning | |
| CN109782603A (en) | The detection method and monitoring system of rotating machinery coupling fault | |
| Li et al. | Joint attention feature transfer network for gearbox fault diagnosis with imbalanced data | |
| Liang et al. | Unsupervised fault diagnosis of wind turbine bearing via a deep residual deformable convolution network based on subdomain adaptation under time-varying speeds | |
| CN112763214A (en) | Rolling bearing fault diagnosis method based on multi-label zero-sample learning | |
| CN116416497B (en) | Bearing fault diagnosis system and method | |
| Pu et al. | A one-class generative adversarial detection framework for multifunctional fault diagnoses | |
| CN112487890B (en) | Bearing acoustic signal fault diagnosis method based on parallel sparse filtering | |
| CN114993677B (en) | Rolling bearing fault diagnosis method and system for unbalanced small sample data | |
| CN116793682A (en) | Bearing fault diagnosis method based on iCORAL-MMD and anti-migration learning | |
| CN116380466A (en) | Rolling bearing intelligent fault diagnosis method and system based on enhanced event visual data | |
| CN115563853A (en) | Rolling bearing fault diagnosis method based on digital twinning | |
| CN113916535B (en) | Bearing diagnosis method, system, equipment and medium based on time frequency and CNN | |
| CN116952583A (en) | Rolling bearing fault diagnosis method of digital twin driving deep transfer learning model | |
| CN116256174A (en) | Noise robust bearing fault diagnosis method based on interpretable multilayer wavelet drive | |
| CN115099264A (en) | Ship part fault diagnosis method and device, computer and computer storage medium | |
| CN113723592A (en) | Fault diagnosis method based on wind power gear box monitoring system | |
| CN114332531B (en) | Mechanical equipment rapid fault diagnosis method and device based on 2D-CNN and t-SNE | |
| Wang et al. | Fault Diagnosis Method of Aeroengine Bearing Based on Convolution Self-Coded Neural Network | |
| CN117516925A (en) | MSDC-Swin-T-based gearbox fault diagnosis method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |