CN107643181B - A kind of rolling bearing variable working condition method for diagnosing faults based on image recognition - Google Patents
A kind of rolling bearing variable working condition method for diagnosing faults based on image recognition Download PDFInfo
- Publication number
- CN107643181B CN107643181B CN201610581598.7A CN201610581598A CN107643181B CN 107643181 B CN107643181 B CN 107643181B CN 201610581598 A CN201610581598 A CN 201610581598A CN 107643181 B CN107643181 B CN 107643181B
- Authority
- CN
- China
- Prior art keywords
- fault
- feature vector
- points
- point
- image
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 56
- 238000005096 rolling process Methods 0.000 title claims abstract description 56
- 239000013598 vector Substances 0.000 claims abstract description 55
- 238000013528 artificial neural network Methods 0.000 claims abstract description 37
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 16
- 238000003745 diagnosis Methods 0.000 claims description 39
- 239000011159 matrix material Substances 0.000 claims description 23
- 238000012545 processing Methods 0.000 claims description 17
- 238000000513 principal component analysis Methods 0.000 claims description 10
- 238000012216 screening Methods 0.000 claims description 10
- 238000012549 training Methods 0.000 claims description 10
- 230000009467 reduction Effects 0.000 claims description 9
- 230000004044 response Effects 0.000 claims description 5
- 238000001914 filtration Methods 0.000 claims description 3
- 238000012360 testing method Methods 0.000 abstract description 26
- 238000010586 diagram Methods 0.000 abstract description 8
- 238000005516 engineering process Methods 0.000 abstract description 8
- 238000000605 extraction Methods 0.000 description 16
- 238000006243 chemical reaction Methods 0.000 description 4
- 239000000284 extract Substances 0.000 description 4
- 238000005070 sampling Methods 0.000 description 4
- 241000282320 Panthera leo Species 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000001514 detection method Methods 0.000 description 3
- 239000000203 mixture Substances 0.000 description 3
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000009877 rendering Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 238000013519 translation Methods 0.000 description 2
- 101100491335 Caenorhabditis elegans mat-2 gene Proteins 0.000 description 1
- 101100495256 Caenorhabditis elegans mat-3 gene Proteins 0.000 description 1
- 102100040428 Chitobiosyldiphosphodolichol beta-mannosyltransferase Human genes 0.000 description 1
- 241001289753 Graphium sarpedon Species 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 230000000739 chaotic effect Effects 0.000 description 1
- 238000010835 comparative analysis Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000005260 corrosion Methods 0.000 description 1
- 230000007797 corrosion Effects 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000009499 grossing Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000036039 immunity Effects 0.000 description 1
- 238000003754 machining Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012567 pattern recognition method Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of rolling bearing variable working condition method for diagnosing faults based on image recognition comprising: bearing vibration signal is converted to by two dimensional image using recurrence diagram technology；By extracting feature from the two dimensional image with SIFT algorithm, multidimensional fault feature vector is obtained；By carrying out dimension-reduction treatment to the multidimensional fault feature vector with KPCA method, about subtracted rear eigenmatrix, and about subtract the final feature vector of singular value building of rear eigenmatrix by extracting；By carrying out failure modes to the final feature vector with trained PNN neural network.The present invention carries out case verifying using the bearing test data of Case Western Reserve University, and test result shows that method proposed by the invention is highly effective.
Description
Technical Field
The invention relates to a variable working condition fault diagnosis technology of a rolling bearing, in particular to a variable working condition fault diagnosis method of a rolling bearing based on image recognition.
Background
The rolling bearing is an important rotating part in the electromechanical equipment, and the fault of the rolling bearing often affects the normal operation of the electromechanical equipment, so that the failure of the electromechanical equipment can be seriously caused, and huge economic loss is brought. Therefore, it is of great significance to ensure the normal operation of the rolling bearing.
Since the bearing vibration signal contains a large amount of bearing state information, the bearing fault diagnosis method based on the vibration signal is receiving wide attention of researchers. The invention also extracts the fault characteristics from the bearing vibration signal so as to realize the fault diagnosis of the bearing. Currently, there are many fault feature extraction methods, such as Wigner-ville distribution (WVD), Wavelet Packet transformation (WPD), Empirical Mode Decomposition (EMD), and the like. However, these methods are based on the assumption that the rolling bearing operates under a fixed condition, and it is difficult to extract stable fault characteristics in the vibration signal of the bearing when the condition of the bearing changes. Because the rolling bearing has complex working environment and variable operation conditions, the development of a set of fault feature extraction and diagnosis method aiming at the rolling bearing under variable conditions has strong necessity.
Scale Invariant Feature Transform (SIFT) is a classic image invariant feature extraction method in the field of image processing at present. Because the generated 128-dimensional feature vector is not influenced by changes such as image translation, scaling and rotation, the SIFT has been successfully applied to various fields such as face recognition, identity recognition and target tracking. However, in the field of fault diagnosis, a fault feature extraction method based on SIFT is rarely reported. According to the invention, firstly, the vibration signal of the rolling bearing is converted into a two-dimensional image, and then stable fault feature extraction is carried out on the rolling bearing under the condition of variable working conditions by utilizing SIFT. The recursion diagram is a method for observing and analyzing a dynamics mechanism in a time sequence from a two-dimensional graph, can reflect phase space manifold in a dynamics system, well reveals the dynamics characteristics of the system, is suitable for carrying out characteristic analysis on a non-stable and non-linear time sequence, and is widely applied to the fields of electrocardiosignal analysis, fingerprint identification, noisy signal characteristic extraction, chaotic time sequence classification and the like at present. The invention adopts a recursion diagram to carry out graphical equivalent conversion on the vibration signal of the rolling bearing.
The 128-dimensional feature vector generated by the SIFT method has too high dimension, so that the problems of low identification precision, low running speed and the like are brought to the subsequent fault feature identification. Kernel Principal Component Analysis (KPCA) is a dimension reduction method developed on the basis of Principal Component Analysis (PCA), and KPCA is more suitable for the dimension reduction of nonlinear signals than PCA. Considering that the bearing vibration signal has strong nonlinearity, the invention adopts a KPCA method to reduce SIFT characteristics so as to reduce the redundancy of fault characteristics and improve the fault identification rate and the algorithm running speed.
After the feature extraction is completed, a pattern recognition method is needed to realize the fault classification of the rolling bearing. The Probabilistic Neural Network (PNN) is an artificial Neural Network with simple structure, simple training and wide application. The invention adopts the PNN neural network to realize the fault diagnosis of the rolling bearing.
Disclosure of Invention
The invention aims to introduce the method in the field of image recognition into the field of rolling bearing fault diagnosis and provide a set of new fault diagnosis method for stable fault feature extraction and fault diagnosis of the rolling bearing under variable working conditions.
The invention relates to a rolling bearing variable working condition fault diagnosis method based on image recognition, which comprises the following steps of:
converting the vibration signal of the rolling bearing into a two-dimensional image by adopting a recursive graph technology;
extracting features from the two-dimensional image by using an SIFT algorithm to obtain a multi-dimensional fault feature vector;
performing dimensionality reduction on the multi-dimensional fault feature vector by using a KPCA (kernel principal component analysis) method to obtain a reduced feature matrix, and extracting singular values of the reduced feature matrix to construct a final feature vector;
and carrying out fault classification on the final characteristic vector by using the trained PNN neural network, and diagnosing the variable working condition fault of the rolling bearing.
Preferably, the step of converting the rolling bearing vibration signal into the two-dimensional image by using the recursive graph technology comprises:
carrying out phase space reconstruction on the time sequence of the vibration signal of the rolling bearing to obtain a reconstructed phase space matrix;
calculating the distance between two phase points in the reconstructed phase space matrix by using the reconstructed phase space matrix;
calculating a recursion value in a recursion graph according to the distance between two phase points in the phase space;
by rendering a recursion map as the two-dimensional image with the recursion values.
Preferably, the extracting features from the two-dimensional image by using the SIFT algorithm to obtain a multi-dimensional fault feature vector includes:
constructing a Gaussian pyramid by performing Gaussian blur and downsampling on the two-dimensional images at different scales, and then subtracting adjacent upper and lower layers of images in each group of the Gaussian pyramid to construct a Gaussian difference scale space;
detecting local spatial extreme points in the constructed Gaussian difference scale space;
screening the local spatial extreme points, screening stable local extreme points from the local spatial extreme points, and taking the stable local extreme points as final key points;
and obtaining the multi-dimensional fault feature vector by performing direction distribution and description processing on each key point.
Preferably, the PNN neural network is trained by using a final feature vector of known fault data in advance to obtain the trained PNN neural network, specifically:
converting the known fault data into a two-dimensional image; performing dimensionality reduction processing on the multi-dimensional fault characteristic vector extracted from the two-dimensional image to obtain a final characteristic vector of known fault data; and then sending the final characteristic vector of the known fault data into the PNN neural network for training to obtain the PNN neural network capable of outputting fault labels corresponding to different faults.
Preferably, said fault classifying said final feature vector with a trained PNN neural network comprises: the feature vector is used as input and is sent to the trained PNN neural network for fault classification, and a corresponding fault label is obtained; and judging the fault type according to the fault label.
Preferably, the detecting the local spatial extreme point in the constructed gaussian difference image comprises: comparing the gray values of each pixel point in the Gaussian difference image with a plurality of adjacent pixel points in the image with the same scale and a plurality of adjacent pixel points in the image with the previous scale and the next scale; and if the comparison result is that the gray value of the pixel point is greater than or less than the gray values of all the adjacent pixel points, the pixel point is considered as a local spatial extreme point.
Preferably, the detecting the local spatial extreme point in the constructed gaussian difference image comprises: comparing the gray values of 8 adjacent pixel points in the image with the same scale and 9 adjacent pixel points in the images with the previous scale and the next scale; and if the comparison result is that the gray value of the pixel point is greater than or less than the gray values of all 26 adjacent pixel points, the pixel point is considered as a local spatial extreme point.
Preferably, the screening the local spatial extreme point includes: and filtering the low-contrast point and the unstable edge response point contained in the local extreme point to obtain a stable local extreme point.
Preferably, the direction allocation and description processing for each key point includes: calculating the gradient direction distribution characteristic of each key point neighborhood pixel; and allocating the direction to each key point by using the calculated gradient direction distribution characteristic of the pixels in the neighborhood of each key point.
Preferably, the performing the direction distribution processing and description processing for each key point further includes: each keypoint is described by a gradient magnitude and direction, and a feature descriptor with rotation invariance is generated and serves as a multi-dimensional fault feature vector.
Compared with the prior art, the invention has the advantages that:
(1) the invention introduces the calculation method in the image field into the fault diagnosis field, and provides a new solution for the fault diagnosis of the rolling bearing;
(2) the invention provides a recursive graph-based graphical equivalent representation method for a vibration signal of a rolling bearing;
(3) the invention provides a stable fault feature extraction method under the condition of variable working conditions of a bearing based on SIFT;
(4) the invention adopts KPCA to reduce the dimension of SIFT feature vector, thereby greatly reducing the calculated amount and improving the operation speed.
The present invention will be described in detail with reference to the accompanying drawings.
Drawings
FIG. 1 is a schematic diagram of a rolling bearing variable working condition fault diagnosis method based on image recognition;
FIG. 2 is a Gaussian pyramid;
FIG. 3 is a diagram of a method of Gaussian difference scale space construction;
FIG. 4 is a rolling bearing data acquisition test bed;
FIG. 5 is a recursive diagram of four failure modes under different conditions of the bearing;
FIG. 6 is a DOG scale space constructed in a normal state of a bearing;
FIG. 7 is a DOG scale space constructed under a bearing inner ring failure mode;
FIG. 8 is a DOG scale space constructed under a fault mode of a bearing rolling element;
FIG. 9 is a DOG scale space constructed under a bearing outer ring failure mode;
FIG. 10 shows feature points detected by four failure modes under different working conditions of a bearing;
FIG. 11 is a fault signature scatter plot of four state modes under different operating conditions of a bearing;
FIG. 12 shows the cross-checking diagnosis result of the 1 st group for bearing variable-condition fault diagnosis;
FIG. 13 shows the cross-checking diagnosis result of the group 2 for bearing variable-condition fault diagnosis;
FIG. 14 shows the cross-checking diagnosis result of group 3 for bearing variable condition fault diagnosis;
FIG. 15 shows the cross-checking diagnosis result of the 4 th group for bearing variable condition fault diagnosis;
FIG. 16 is a schematic diagram of a rolling bearing variable working condition fault diagnosis method based on image recognition.
Detailed Description
Fig. 16 shows a rolling bearing variable condition fault diagnosis method based on image recognition, and as shown in fig. 16, the rolling bearing variable condition fault diagnosis method based on image recognition of the invention comprises the following steps:
converting the vibration signal of the rolling bearing into a two-dimensional image by adopting a recursive graph technology;
extracting features from the two-dimensional image by using an SIFT algorithm to obtain a multi-dimensional fault feature vector;
performing dimensionality reduction on the multi-dimensional fault feature vector by using a KPCA (kernel principal component analysis) method to obtain a reduced feature matrix, and extracting singular values of the reduced feature matrix to construct a final feature vector;
and carrying out fault classification on the final characteristic vector by using the trained PNN neural network, and diagnosing the variable working condition fault of the rolling bearing.
The above-mentioned converting the rolling bearing vibration signal into the two-dimensional image by using the recursive graph technique includes: carrying out phase space reconstruction on the time sequence of the vibration signal of the rolling bearing to obtain a reconstructed phase space matrix; calculating the distance between two phase points in the reconstructed phase space matrix by using the reconstructed phase space matrix; calculating a recursion value in a recursion graph according to the distance between two phase points in the phase space; by rendering a recursion map as the two-dimensional image with the recursion values.
The above-mentioned extracting features from the two-dimensional image by using the SIFT algorithm to obtain a multi-dimensional fault feature vector includes: constructing a Gaussian pyramid by performing Gaussian blur and downsampling on the two-dimensional images at different scales, and then subtracting adjacent upper and lower layers of images in each group of the Gaussian pyramid to construct a Gaussian difference scale space; detecting local spatial extreme points in the constructed Gaussian difference scale space; screening the local spatial extreme points, screening stable local extreme points from the local spatial extreme points, and taking the stable local extreme points as final key points; and obtaining the multi-dimensional fault feature vector by performing direction distribution and description processing on each key point.
The PNN neural network is trained by using the final feature vector of the known fault data in advance to obtain the trained PNN neural network, and the method specifically comprises the following steps: converting the known fault data into a two-dimensional image; performing dimensionality reduction processing on the multi-dimensional fault characteristic vector extracted from the two-dimensional image to obtain a final characteristic vector of known fault data; and then sending the final characteristic vector of the known fault data into the PNN neural network for training to obtain the PNN neural network capable of outputting fault labels corresponding to different faults.
The above fault classification of the final feature vector by using the trained PNN neural network includes: the feature vector is used as input and is sent to the trained PNN neural network for fault classification, and a corresponding fault label is obtained; and judging the fault type according to the fault label.
The above detecting the local spatial extreme point in the constructed gaussian difference image includes: comparing the gray values of each pixel point in the Gaussian difference image with a plurality of adjacent pixel points in the image with the same scale and a plurality of adjacent pixel points in the image with the previous scale and the next scale; and if the comparison result is that the gray value of the pixel point is greater than or less than the gray values of all the adjacent pixel points, the pixel point is considered as a local spatial extreme point.
In a specific embodiment, the detecting the local spatial extreme point in the constructed gaussian difference image includes: comparing the gray values of 8 adjacent pixel points in the image with the same scale and 9 adjacent pixel points in the images with the previous scale and the next scale; and if the comparison result is that the gray value of the pixel point is greater than or less than the gray values of all 26 adjacent pixel points, the pixel point is considered as a local spatial extreme point.
The above-mentioned screening process of the local spatial extreme point includes: and filtering the low-contrast point and the unstable edge response point contained in the local extreme point to obtain a stable local extreme point.
The above direction allocation and description processing for each key point includes: calculating the gradient direction distribution characteristic of each key point neighborhood pixel; and allocating the direction to each key point by using the calculated gradient direction distribution characteristic of the pixels in the neighborhood of each key point.
The above-mentioned direction allocation processing and description processing for each key point further includes: each keypoint is described by a gradient magnitude and direction, and a feature descriptor with rotation invariance is generated and serves as a multi-dimensional fault feature vector.
The present invention is further illustrated by the following specific examples, which are intended to be merely illustrative of, and not limiting of, the above-described methods of the present invention.
FIG. 1 is a general flow chart of a rolling bearing variable working condition fault diagnosis method based on image recognition. Firstly, a recursive graph technology is adopted to convert rolling bearing vibration signals under different working conditions into a two-dimensional image. And then, extracting stable fault feature vectors by applying an SIFT algorithm to the two-dimensional image to obtain a 128-dimensional feature matrix, reducing the dimensions of the feature matrix by using a KPCA algorithm, and extracting singular values of the reduced dimension matrix to construct final feature vectors. Under different working conditions, part of feature vectors are respectively selected as training data to train the PNN neural network, and the rest feature vectors are used as test data to be sent to the trained PNN neural network to realize fault classification. The fault diagnosis method mainly comprises the following three steps: the method comprises the steps of image conversion based on a recursive graph, feature extraction based on SIFT and KPCA and fault classification based on a PNN neural network.
1. The specific embodiment is as follows:
1.1 recursive graph-based image conversion
In the scheme, the recursive graph technology is adopted to convert the rolling bearing vibration signal into a two-dimensional image, and an image basis is provided for the subsequent SIFT-based fault feature extraction. The recursive graph technology is realized by the following steps:
(1) for a time sequence u with a sampling interval Δ tk(k is 1,2, …, N), selecting a suitable embedding dimension m and delay time τ by using a Cao method and a mutual information method to perform phase space reconstruction on the time sequence, and obtaining a matrix with N rows and m columns as follows after reconstruction:
xi＝(ui,ui+τ,…,ui+(m-1)τ)
wherein i is 1,2, …, N- (m-1) tau
(2) Calculating two phase points x in reconstructed phase spaceiAnd xjA distance S betweenij＝||xi-xj||，i＝1,2,…,N-(m-1)τ,j＝1,2,…,N-(m-1)τ
(3) Calculating a recursion value in the recursion graph:
R(i,j)＝Θ(εi-Sij)
wherein epsiloniFor the threshold, it can be fixed or changed with i, and its size only affects the recursion graphThe density of the points does not change the structure of the recursive graph, and is generally selected to be n times of the standard deviation of the original time series or selected according to experience. Θ is the Heaviside unit function:
(4) and drawing a recursive graph. The recursion map is obtained by plotting R (i, j) on a coordinate axis with i as the abscissa and j as the ordinate. The value of R (i, j) is 0 or 1, representing the white and black points in the figure, respectively. As can be seen from R (i, j) ═ R (j, i) and R (i, j) ═ 1, (i ═ j) there is a main diagonal in the recursive graph, and the recursive graph is symmetric about the main diagonal.
1.2 feature extraction based on SIFT and KPCA
After image transformation, features are extracted from the two-dimensional image by using a SIFT algorithm. The SIFT algorithm was first proposed by Lowe in 1999 and has good performance in the aspect of gray level image feature detection. The algorithm mainly comprises the following four steps:
step 1: and establishing a scale space. The scale space is constructed using a gaussian pyramid as shown in fig. 2. The Gaussian pyramid is obtained by performing Gaussian blur and downsampling on the image at different scales. Assuming that I (x, y) is a two-dimensional image, its scale space L (x, y, σ) can be defined as follows:
L(x,y,σ)＝G(x,y,σ)*I(x,y)
wherein,is a variable scale gaussian function, represents a convolution operation, (x, y) is a spatial coordinate, and σ represents a scale factor. As the σ value increases, the degree of smoothing also increases, and the image is blurred. Thus, the large scale and the small scale reflect the overview features and the detail features of the image, respectively.
In order to simplify the calculation, Lowe uses a gaussian difference scale-space (DOG scale-space) to realize the detection of the extreme points in the scale space. The Gaussian difference scale space is obtained by subtracting adjacent upper and lower layers of images in each group of the Gaussian pyramid:
in this formula, D (x, y, σ) is the gaussian difference of the image, L (x, y, σ) is the scale representation of the scale space, and k is a constant multiple factor to change the scale. Potential scale and rotation invariant points of interest may be identified by searching for local extrema points. An efficient method for constructing D (x, y, σ) is shown in fig. 3.
Step 2: and detecting local spatial extreme points. In order to detect local extreme points of a Gaussian difference image in a constructed DoG, each pixel point is compared with 8 adjacent points in the image with the same scale and 9 adjacent points in the images with the previous scale and the next scale. If the value is greater than or less than all 26 adjacent points, the pixel point is considered as a local spatial extreme point.
And step 3: and (4) positioning key points. In order to improve the noise immunity and matching stability of the extracted features, the low contrast point (sensitive to noise) and the unstable edge response point (caused by the edge response of the gaussian difference function) included in the selected local extreme points are filtered. After the two kinds of screening, the obtained stable local extreme points are called key points, and the key points are obtained by utilizing scale invariance, so that the key points have scaling invariance.
And 4, step 4: and (4) direction distribution. Neighborhood pixels of a keypoint have a gradient direction distribution characteristic, and therefore, the characteristic can be used to assign directions to keypoints to generate a feature descriptor with rotation invariance. The detection of all key points of the image is completed through the steps, and each key point comprises position, direction and scale information and has invariance to translation, rotation and scaling.
And 5: and (5) key point description. The key point descriptors are finally obtained by using the gradient magnitude and direction as basic elements. The Lowe proposed descriptor uses 8 directions of gradient information computed in a 4 x 4 window within the keypoint scale space. Thus, each keypoint descriptor includes a feature vector of dimensions 4 × 4 × 8 ═ 128. The descriptor vector remains invariant to rotation, scaling and illumination variations.
The SIFT algorithm usually extracts a huge number of features from the image, so the huge computational consumption limits its application in fault diagnosis. To solve this problem, a dimension reduction method is required to reduce the amount of computation. While PCA can extract essential structures from high dimensional datasets, PCA as a linear method does not extract the nonlinear structures of the datasets. Kernel PCA is an extended nonlinear form of PCA that can compute the principal components of a data set that is nonlinearly mapped into a high-dimensional feature space. Therefore, the present invention utilizes KPCA to reduce the dimensionality of SIFT feature vectors. And after the dimension reduction is finished, extracting singular values in the matrix and constructing a final eigenvector. At this point, the feature extraction process is complete.
1.3 PNN neural network-based Fault Classification
And after the feature extraction is finished, training a PNN neural network classifier to realize fault classification. Input samples of the PNN neural network are fault feature vectors of four modes, namely normal, inner ring fault, outer ring fault and rolling element fault, extracted by the rolling bearing under each working condition, and the feature vectors are respectively marked as 1,2, 3 and 4. Then, a 4-class PNN neural network classifier is trained for fault classification.
For test vibration signals under different working conditions, the test vibration signals are converted into a recursion graph, then the SIFT algorithm is adopted to extract fault features, and the fault features are reduced based on KPCA to construct final feature vectors. And finally, the feature vectors are sent to a trained PNN neural network classifier for classification, and the classification accuracy is calculated by comparing the prediction label with the training label.
2. The experimental results are as follows:
2.1 Experimental data
In order to verify the effectiveness of the method, the method is proved by using a test bed data set of a bearing data center of Kaiser university. The bearing test apparatus is shown in fig. 4. The test platform consists of a 2 horsepower motor (left side) (1hp 746W), a torque sensor (center), a power meter (right side) and electronic control equipment. Bearing faults were injected using an electro-discharge machining technique with injected fault diameters of 0.007, 0.014, 0.021, 0.028, 0.040 inches, respectively. The acceleration sensor was mounted on the motor housing using a magnetic base, and the vibration signals it generated were collected by a 16-channel DAT recorder and later processed in the MATLAB environment. The sampling frequency of the digital signal is 12000Hz, and the sampling frequency of the fault data of the bearing at the driving end is 48000 Hz. Bearing outer race faults are arranged in the 3 o ' clock, 6 o ' clock and 12 o ' clock directions.
The method selects the SKF bearing at the driving end as a research object, the diameter of the pitting corrosion fault is 0.021 inch, and the sampling frequency of the vibration data of the bearing at the driving end is 48000 Hz. Keeping the load and the rotating speed of the motor unchanged, and obtaining the data of the normal state of the bearing at the driving end, the fault of the inner ring, the fault of the rolling body and the fault of the outer ring under different working conditions. The invention selects the bearing test data under 4 working conditions for analysis, and the data composition is shown in table 1. And verifying the feasibility of the variable working condition fault diagnosis method based on the accelerated robust features and the equidistant mapping by using the following test data.
TABLE 1 test bearing data information
| Working conditions | Motor speed/rpm | Is normal | Inner ring | Rolling body | Outer ring |
| 0 | 1797 | 97.mat | 213.mat | 226.mat | 238.mat |
| 1 | 1772 | 98.mat | 214.mat | 227.mat | 239.mat |
| 2 | 1750 | 99.mat | 215.mat | 228.mat | 240.mat |
| 3 | 1730 | 100.mat | 217.mat | 229.mat | 241.mat |
2.2 recursive graph-based image conversion
And (3) performing graphical equivalent representation on normal, inner ring fault, outer ring fault and rolling body fault vibration data of the bearing under 4 different working conditions by adopting a recursion diagram. And 20 groups of vibration data are respectively selected under each working condition and each fault mode, and each group of data comprises 1000 points. For vibration data under each working condition, a Cao method and a mutual information method are adopted to select a proper embedding dimension m and a proper delay time tau to carry out phase space reconstruction on a vibration signal time sequence, parameters m and tau obtained by calculation under each working condition are shown in a table 2, and a threshold value epsilon is selected to be 1. And analyzing the recursive behavior of the vibration signal in the reconstruction phase space to generate a recursive graph, thereby realizing the graphical equivalent representation of the bearing vibration signal. In order to analyze the influence of the working condition change on the recursion graph, a group of test data is randomly selected to generate the recursion graph for comparative analysis for each fault mode under 4 working conditions, as shown in fig. 5.
TABLE 2 Experimental parameters for each failure mode under different conditions
2.3 feature extraction based on SIFT and KPCA
In the part, firstly, a SIFT algorithm is adopted to perform feature extraction on recursive graphs of the rolling bearing under different fault modes, and FIGS. 6-9 are scale spaces constructed under the conditions of a normal state, an inner ring fault, a rolling element fault and an outer ring fault of the rolling bearing under a working condition 1, wherein the DOG differential pyramid comprises 7 groups (o is 7), each group comprises 5 layers (s is 5), and the fuzzy is performed between different layers through a scale factor sigma. The corresponding detected feature points are shown in fig. 10.
Considering that the problem that the SIFT feature dimension is too high, which causes low accuracy of subsequent fault classification and large consumption of computing resources, the invention adopts a KPCA method to reduce the dimension of the feature matrix. In the invention, each SIFT feature is reduced into a 3-dimensional feature vector by using KPCA, and singular values of the SIFT features are extracted to construct a final feature matrix. FIG. 11 is a scatter plot of four fault signatures extracted under four different operating conditions.
2.4 fault classification based on PNN neural networks
The invention adopts the PNN neural network as a fault classifier to realize fault classification of the rolling bearing. In order to verify the accuracy of the method, the fault diagnosis adopts a cross-check mode, 1 kind of working condition data is sequentially selected from test data acquired under 4 kinds of operating working conditions as training data, and the other 3 kinds of working condition data are used as test data for fault identification. The data composition is shown in table 3.
TABLE 3 bearing variable working condition fault diagnosis cross-checking data composition
In each set of cross-tests, the training data and the test data contained the following number of data sets:
training data: 20 groups of data are selected from 4 state modes (normal, inner ring fault, rolling element fault and outer ring fault) respectively, and the total number of the data is 80;
test data: under each working condition, 20 groups of data are respectively selected from 4 state modes (normal, inner ring fault, rolling element fault and outer ring fault), wherein 1-80 groups are test data under the 1 st working condition, 81-160 groups are test data under the 2 nd working condition, and 161-240 groups are test data under the 3 rd working condition;
the results of the diagnosis using the Probabilistic Neural Network (PNN) classifier are shown below. Wherein fig. 12-15 are 4 sets of cross-validation results. The red circle is the actual failure mode category, and the blue triangle is the classifier decision result. The vertical axes 1-4 represent 4 failure modes of normal, inner ring failure, rolling element failure, outer ring failure, respectively. Tables 4 and 5 summarize the results of the cross-testing of the rolling bearing 4 sets.
TABLE 4 bearing variable working condition diagnosis 4 groups cross-checking each fault mode identification error rate
TABLE 5 summary of bearing variable condition fault diagnosis cross-examination results
| Number of cross-test sets | 1 | 2 | 3 | 4 | Average/total number |
| Average classification accuracy | 97.08％ | 97.08％ | 97.08％ | 97.5％ | 97.19％ |
| Total number of erroneous samples | 7 | 7 | 7 | 6 | 27 |
| Total number of test samples | 240 | 240 | 240 | 240 | 960 |
The PNN neural network diagnosis result can show that the classification accuracy of each group of cross tests is higher than 97%, and the average classification accuracy of 4 groups of cross test results reaches 97.19%. The diagnosis result shows that the rolling bearing variable working condition fault diagnosis method based on image recognition is very effective.
Although the present invention has been described in detail hereinabove, the present invention is not limited thereto, and various modifications can be made by those skilled in the art in light of the principle of the present invention. Thus, modifications made in accordance with the principles of the present invention should be understood to fall within the scope of the present invention.
Claims (9)
1. A rolling bearing variable working condition fault diagnosis method based on image recognition comprises the following steps:
carrying out phase space reconstruction on the time sequence of the vibration signal of the rolling bearing to obtain a reconstructed phase space matrix;
calculating the distance between two phase points in the reconstructed phase space matrix by using the reconstructed phase space matrix;
calculating a recursion value in a recursion graph according to the distance between two phase points in the phase space;
drawing a recursion graph by using the recursion value, and converting the vibration signal of the rolling bearing into a two-dimensional image by using the drawn recursion graph;
extracting features from the two-dimensional image by using an SIFT algorithm to obtain a multi-dimensional fault feature vector, performing dimension reduction processing on the multi-dimensional fault feature vector by using a KPCA (kernel principal component analysis) method to obtain a reduced feature matrix, and extracting singular values of the reduced feature matrix to construct a final feature vector;
and carrying out fault classification on the final eigenvector constructed by extracting the singular value of the reduced eigenvector matrix by using the trained PNN neural network, and diagnosing the variable working condition fault of the rolling bearing.
2. The method of claim 1, wherein obtaining a multi-dimensional fault feature vector by extracting features from the two-dimensional image with a SIFT algorithm comprises:
constructing a Gaussian pyramid by performing Gaussian blur and downsampling on the two-dimensional images at different scales, and then subtracting adjacent upper and lower layers of images in each group of the Gaussian pyramid to construct a Gaussian difference scale space;
detecting local spatial extreme points in the constructed Gaussian difference scale space;
screening the local spatial extreme points, screening stable local extreme points from the local spatial extreme points, and taking the stable local extreme points as final key points;
and obtaining the multi-dimensional fault feature vector by performing direction distribution and description processing on each key point.
3. The method according to claim 1 or 2, wherein the PNN neural network is trained in advance using a final eigenvector of known fault data to obtain the trained PNN neural network, specifically:
obtaining a final feature vector of the known fault data by processing the known fault data according to claim 1;
and sending the final characteristic vector of the known fault data into the PNN neural network for training to obtain the PNN neural network capable of outputting fault labels corresponding to different faults.
4. The method of claim 3, wherein said fault classifying said final feature vector with a trained PNN neural network comprises:
the feature vector is used as input and is sent to the trained PNN neural network for fault classification, and a corresponding fault label is obtained;
and judging the fault type according to the fault label.
5. The method of claim 2, wherein the detecting the local spatial extreme point in the constructed gaussian difference image comprises:
comparing the gray values of each pixel point in the Gaussian difference image with a plurality of adjacent pixel points in the image with the same scale and a plurality of adjacent pixel points in the image with the previous scale and the next scale;
and if the comparison result is that the gray value of the pixel point is greater than or less than the gray values of all the adjacent pixel points, the pixel point is considered as a local spatial extreme point.
6. The method of claim 2, wherein the detecting the local spatial extreme point in the constructed gaussian difference image comprises:
comparing the gray values of 8 adjacent pixel points in the image with the same scale and 9 adjacent pixel points in the images with the previous scale and the next scale;
and if the comparison result is that the gray value of the pixel point is greater than or less than the gray values of all 26 adjacent pixel points, the pixel point is considered as a local spatial extreme point.
7. The method of claim 2, wherein the screening the local spatial extreme point comprises:
and filtering the low-contrast point and the unstable edge response point contained in the local extreme point to obtain a stable local extreme point.
8. The method of claim 2, wherein the assigning direction and description processing for each keypoint comprises:
calculating the gradient direction distribution characteristic of each key point neighborhood pixel;
and allocating the direction to each key point by using the calculated gradient direction distribution characteristic of the pixels in the neighborhood of each key point.
9. The method of claim 8, wherein the performing direction assignment processing and description processing for each keypoint further comprises:
each keypoint is described by a gradient magnitude and direction, and a feature descriptor with rotation invariance is generated and serves as a multi-dimensional fault feature vector.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610581598.7A CN107643181B (en) | 2016-07-21 | 2016-07-21 | A kind of rolling bearing variable working condition method for diagnosing faults based on image recognition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610581598.7A CN107643181B (en) | 2016-07-21 | 2016-07-21 | A kind of rolling bearing variable working condition method for diagnosing faults based on image recognition |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107643181A CN107643181A (en) | 2018-01-30 |
| CN107643181B true CN107643181B (en) | 2019-11-12 |
Family
ID=61109845
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610581598.7A Active CN107643181B (en) | 2016-07-21 | 2016-07-21 | A kind of rolling bearing variable working condition method for diagnosing faults based on image recognition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107643181B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108664936B (en) * | 2018-05-14 | 2020-09-01 | 浙江师范大学 | Diagnosis method and system based on machine fault |
| CN108763377B (en) * | 2018-05-18 | 2021-08-13 | 郑州轻工业学院 | Multi-source telemetering big data feature extraction preprocessing method based on satellite fault diagnosis |
| CN110598768B (en) * | 2019-08-30 | 2022-05-17 | 武汉科技大学 | Gear fault classification method, classification device and readable storage medium |
| CN111308985B (en) * | 2020-02-18 | 2021-03-26 | 北京航空航天大学 | Performance degradation evaluation method for control assembly of airplane environmental control system based on NSCT and DM |
| KR102309559B1 (en) * | 2020-06-17 | 2021-10-07 | 광주과학기술원 | The method for detecting fault of motor |
| CN112488179A (en) * | 2020-11-26 | 2021-03-12 | 中国舰船研究设计中心 | Rotary machine fault diagnosis method based on GRU |
| CN113405799B (en) * | 2021-05-20 | 2022-06-28 | 新疆大学 | Bearing early fault detection method based on health state index construction and fault early warning limit self-learning |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6629058B2 (en) * | 2000-04-20 | 2003-09-30 | Rion Co., Ltd. | Fault diagnosis method and apparatus |
| CN202793793U (en) * | 2012-08-30 | 2013-03-13 | 桂林电子科技大学 | Large wind generation set bearing fault diagnosis system |
| CN103048137A (en) * | 2012-12-20 | 2013-04-17 | 北京航空航天大学 | Fault diagnosis method of rolling bearing under variable working conditions |
| CN105181110A (en) * | 2015-09-13 | 2015-12-23 | 北京航空航天大学 | Rolling bearing fault diagnosis method based on SIFT-KPCA and SVM |
| CN105758645A (en) * | 2014-12-20 | 2016-07-13 | 哈尔滨智晟天诚科技开发有限公司 | Probabilistic neural network based engine fault diagnosis system |
-
2016
- 2016-07-21 CN CN201610581598.7A patent/CN107643181B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6629058B2 (en) * | 2000-04-20 | 2003-09-30 | Rion Co., Ltd. | Fault diagnosis method and apparatus |
| CN202793793U (en) * | 2012-08-30 | 2013-03-13 | 桂林电子科技大学 | Large wind generation set bearing fault diagnosis system |
| CN103048137A (en) * | 2012-12-20 | 2013-04-17 | 北京航空航天大学 | Fault diagnosis method of rolling bearing under variable working conditions |
| CN105758645A (en) * | 2014-12-20 | 2016-07-13 | 哈尔滨智晟天诚科技开发有限公司 | Probabilistic neural network based engine fault diagnosis system |
| CN105181110A (en) * | 2015-09-13 | 2015-12-23 | 北京航空航天大学 | Rolling bearing fault diagnosis method based on SIFT-KPCA and SVM |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107643181A (en) | 2018-01-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107643181B (en) | A kind of rolling bearing variable working condition method for diagnosing faults based on image recognition | |
| Babu et al. | Statistical features based optimized technique for copy move forgery detection | |
| Tao et al. | An unsupervised fault diagnosis method for rolling bearing using STFT and generative neural networks | |
| Gao et al. | Automatic change detection in synthetic aperture radar images based on PCANet | |
| CN107657279B (en) | Remote sensing target detection method based on small amount of samples | |
| Nogueira et al. | Evaluating software-based fingerprint liveness detection using convolutional networks and local binary patterns | |
| CN111103139A (en) | Rolling bearing fault diagnosis method based on GRCMSE and manifold learning | |
| CN108345827B (en) | Method, system and neural network for identifying document direction | |
| CN109858352B (en) | Fault diagnosis method based on compressed sensing and improved multi-scale network | |
| CN107145829B (en) | Palm vein identification method integrating textural features and scale invariant features | |
| CN102147858B (en) | License plate character identification method | |
| CN108520215B (en) | Single-sample face recognition method based on multi-scale joint feature encoder | |
| Tang et al. | Distinctive image features from illumination and scale invariant keypoints | |
| CN101140216A (en) | Gas-liquid two-phase flow type recognition method based on digital graphic processing technique | |
| Hussain et al. | Robust pre-processing technique based on saliency detection for content based image retrieval systems | |
| CN111400528B (en) | Image compression method, device, server and storage medium | |
| Sun et al. | Curvature enhanced bearing fault diagnosis method using 2D vibration signal | |
| Zhou et al. | Fault diagnosis for rolling bearing under variable conditions based on image recognition | |
| CN110942473A (en) | Moving target tracking detection method based on characteristic point gridding matching | |
| CN112784754A (en) | Vehicle re-identification method, device, equipment and storage medium | |
| CN116910752A (en) | Malicious code detection method based on big data | |
| Lee et al. | Fast object localization using a CNN feature map based multi-scale search | |
| CN109871825B (en) | Portrait identification method based on improved local two-dimensional mode | |
| Amiri et al. | RASIM: a novel rotation and scale invariant matching of local image interest points | |
| CN111127407B (en) | Fourier transform-based style migration forged image detection device and 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |