CN111539152B - Rolling bearing fault self-learning method based on two-stage twin convolutional neural network - Google Patents
Rolling bearing fault self-learning method based on two-stage twin convolutional neural network Download PDFInfo
- Publication number
- CN111539152B CN111539152B CN202010066718.6A CN202010066718A CN111539152B CN 111539152 B CN111539152 B CN 111539152B CN 202010066718 A CN202010066718 A CN 202010066718A CN 111539152 B CN111539152 B CN 111539152B
- Authority
- CN
- China
- Prior art keywords
- fault
- network
- training
- time
- cnn
- 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
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Acoustics & Sound (AREA)
- Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention discloses a rolling bearing fault self-learning method based on a two-stage twin convolutional neural network, which uses incomplete data to perform two-stage modeling: the first stage implements fault type identification. The second stage realizes fault damage degree classification, performs short-time Fourier transform on the signals of the identified fault types to obtain a sliding window time-frequency diagram, and extracts the slight differences of the faults with different damage degrees; and (3) taking the time-frequency diagram as the input of the second-stage S-CNN2, and designing a classifier based on the target space distance to realize classification and self-learning of the same fault damage degree. The method is applied to the rolling fault data acquired by the fault experimental platform, and the result shows that the method not only can finish accurate division of fault types and damage grades, but also can realize self-learning of faults and self-growth of a fault library under the condition of incomplete data modeling, and the intelligence of the classification process is enhanced.
Description
Technical Field
The invention relates to an intelligent diagnosis method for rolling bearing faults, in particular to an intelligent diagnosis method for the rolling bearing faults, which can realize accurate classification of the bearing faults and self-learning of models and enhance the intelligence of the classification process when the problem that an effective diagnosis model is difficult to establish by a traditional mode identification method aiming at a large amount of label-free data in the current industrial production is solved.
Background
With the rapid development of modern industry, in order to meet the production requirements of the current society, domestic and foreign mechanical equipment develops rapidly in the directions of intellectualization, complication, automation, precision and high speed, so that the connection among all components is tighter, and in addition, the mechanical equipment is very easy to break down due to the complex working environment, the operation and maintenance are difficult, and the faults of some key parts can cause the normal operation of the whole equipment, thereby causing serious property loss and casualties.
Rolling bearings, which are common elements of machinery, play an important role not only in industrial production but also in daily life, and are most likely to fail because they are used in operating environments of high-speed rotation, alternating load, and high temperature for a long time. According to statistics of relevant data, 40% of faults in the faults of the mechanical equipment are caused by the faults of the bearing, so that the rolling bearings are accurately identified and classified, major accidents can be effectively avoided, the maintenance cost is reduced, according to statistics, 25% -50% of maintenance cost and 75% of accident rate can be reduced by diagnosing and identifying the faults of the rolling bearings, the working potential of the bearing is furthest exerted, and the expenditure is saved. The traditional fault diagnosis method for the rolling bearing mainly comprises a wavelet analysis method, a support vector machine, a neural network and the like, and the bearing fault data obtained on site is usually incomplete and label-free. Along with the expansion of the application of the information acquisition system, the monitoring range and depth of the large-scale bearing state are continuously enhanced, and the generated data are in mass characteristics. However, the feature extraction process is a time-consuming and labor-intensive task and has a great impact on the final result. The traditional feature extraction and selection method has certain complexity and uncertainty, and the problems cause that a large amount of data in the field cannot be effectively utilized. And a deep network represented by a Convolutional Neural Network (CNN) provides an effective method capable of automatically extracting original data features, and can solve the problems of large data volume, difficult feature extraction and the like. The invention combines CNN and twin structure, provides a semi-supervised self-learning network of twin structure CNN (S-CNN) based on similarity measurement, for untrained new fault and new damage level of the rolling bearing, in a target space, the S-CNN can quickly separate the rolling bearing from the known level or fault and quickly aggregate with the same level or fault, and finally a new fault or new damage level is generated, thereby solving the problem of fault self-learning. The method has better practical significance for solving the classification of a large amount of label-free fault data in the current industrial production.
Disclosure of Invention
Aiming at the problem that the fault data acquired on site is often incomplete and label-free, so that an effective diagnosis model is difficult to establish by the traditional mode identification method, the invention provides a rolling bearing fault self-learning method based on a two-stage twin Convolutional Neural network (S-CNN), as shown in FIG. 1, the method is a general block diagram of the algorithm of the invention, and the general description of the method is as follows: firstly, combining a twin structure with similarity measurement with CNN to construct an S-CNN network structure; . Then, based on the constructed S-CNN network structure, establishing a first-level fault type identification network S-CNN 1: (1) removing part of noise by using the morphological enhancement pulse characteristics; (2) extracting the signal fault type common characteristics by applying S transformation, and constructing a time-frequency diagram; (3) inputting the time-frequency diagram sample into a first-stage S-CNN, training the time-frequency diagram sample to finally form a first-stage fault type identification network S-CNN1 (the trained network can realize identification and self-learning of the fault type of the rolling bearing by utilizing the characteristics of convergence and separation of the same type of faults and different types of faults of the network in a target space); and finally, establishing a second-stage fault damage degree classification network S-CNN 2: (1) amplifying differences of different damage degrees of faults of the same type by using short-time Fourier transform (STFT) with a sliding window to obtain a sliding time-frequency diagram sample; (2) and inputting the obtained time-frequency pattern into a second-stage S-CNN, training the time-frequency pattern, and finally forming a second-stage fault damage degree classification network S-CNN2 (the trained network is a classifier based on target space distance, and can realize self-learning and self-growth of fault damage degree).
The specific technical scheme and the implementation steps of the invention are as follows:
1. combining twin structure with CNN to construct S-CNN network structure
The first step of the present invention is to construct a network structure (S-CNN) in which a twin structure is combined with a Convolutional Neural Network (CNN), the structure of which is shown in fig. 10. The constructed S-CNN structure is a similarity measurement method in nature. As can be seen from FIG. 10, the input is two time-frequency graphs x 1 And x 2 At the time of model training, they have a known label Y (where Y is a binary label (0 or 1)). When time-frequency diagram x is input 1 And x 2 When belonging to the same class, Y is 0; otherwise, Y is 1, and the label is used for network training. Then, Gw (x) 1 ) And Gw (x) 2 ) Are respectively x 1 And x 2 Corresponding outputs over two CNN networks, which will be x 1 And x 2 Mapping to two points in a low dimensional space. Where w is the shared parameter to be learned by both CNN networks. Then, the outputs Gw (x) of the two networks are combined 1 ) And Gw (x) 2 ) And (4) performing operation to obtain a similarity measurement index Ew, wherein the index is called an energy function. The energy function Ew (x) 1 ,x 2 ) The following can be defined:
Ew(x 1 ,x 2 )=||Gw(x 1 )-Gw(x 2 )|| (1)
the above is the constructed S-CNN, basic structure, and then different S-CNN networks can be trained by inputting different training samples, so that the S-CNN has different classification or identification functions. Specifically, in the second step, the input part of rolling bearing different fault type samples train the S-CNN1 network to become a network for identifying the fault type of the bearing and have a self-learning function; in the third step, the S-CNN2 network is trained by inputting samples of the same fault type and different fault damage levels of partial rolling bearings, so that the network becomes a network for identifying the fault damage levels of the bearings and has a self-learning function. And finally, in the training process of the S-CNN network, the output of the whole S-CNN network is expanded into a target space mainly according to a method for training network parameters layer by Back Propagation (BP), and the final output value of the network is the position in the target space.
2. Establishing a first-level S-CNN1 fault type self-learning network
(a) Data preprocessing: assuming that there is one unknown fault type X and several known fault types X i (i is more than or equal to 2), performing mathematical morphology filtering on all original vibration signals, removing part of background noise, and dividing data according to each group of Q sampling points to be used as an input sample; and then, carrying out S transformation on each group to obtain a time-frequency graph corresponding to each group of vibration data and containing time domain and frequency domain information, and converting the time-frequency graph into an n multiplied by n matrix. 80% of all signal data were taken as training set data and 20% as test set data.
(b) Parameter selection: selecting network parameters, wherein the two same CNN network parameters adopted by the first-stage S-CNN1 network are selected as follows: selecting 5 layers of convolution networks, a first layer of 16 convolution kernels, a second layer of 32 convolution kernels, a third layer of 64 convolution kernels, a fourth layer of 128 convolution kernels and a fifth layer of 256 convolution kernels, wherein the sizes of the convolution kernels are 5 multiplied by 5, selecting Adam by an optimization algorithm, selecting the Adam to have good robustness on super parameters (learning rate, regularization coefficients and the like), and setting the maximum training times to be 4000 times.
(c) Training process: after the network parameters are determined, taking the time-frequency diagram (n multiplied by n matrix) of the training set in the step (a) as the input of the S-CNN1, and calculating the corresponding Gw (x) 1 )、Gw(x 2 ) And an energy function Ew (x) 1 ,x 2 ) And training the S-CNN network structure by using the known label Y of the training set.The training process mainly trains parameters of each layer of the network layer by layer according to Back Propagation (BP), and finally the trained network structure S-CNN1 is obtained.
The detailed training process is as follows: first, a loss function l is defined, whose magnitude is dependent only on the inputs and parameters of the network via an "energy function Ew", which is defined as:
L(W,(Y,x 1 ,x 2 ) i )=(1-Y)L G (E W (x 1 ,x 2 ) i )+YL I (E W (x 1 ,x 2 ) i ) (3)
L G (E W (x 1 ,x 2 ) i )=||Gw(x 1 )-Gw(x 2 )|| 2 (4)
wherein (Y, x) 1 ,x 2 ) i Represents the ith sample pair, (Y, x) 1 ,x 2 ) i Comprising a pair of images and a known label Y, L G Represents the part of the loss function when Y is 0 (true pair), L I The representative loss function is a portion when Y is 1 (false pair), and P represents the number of training samples. Design such L G And L I In order to reduce the energy of a true pair and increase the energy of a false pair, training layer by layer according to Back Propagation (BP) is carried out to minimize the energy of l, and when the energy of l is minimum, the training is finished, and the parameters of each layer of the network are trained.
(d) The testing and self-learning process comprises the following steps: after the network is trained, the time-frequency diagram (n multiplied by n matrix) of the test set in the step (a) is used as the input of the S-CNN1 trained in the step (c), the time-frequency diagram of each fault vibration signal obtains a coordinate value (x, y) mapped to a target space through the last layer of the S-CNN1, and the time-frequency diagram of each fault vibration signal is formed by twin structuresThe principle shows that the same type of fault is closer to the target space, and different types of faults are further and further. After 4000 times of training, finally, the two types of faults are converged into two clusters which are not overlapped in the target space, and the learning process of the network on different characteristics is completed. Calculating the mass center of each type of fault in the training set in the target space set, and recording the mass center of the ith type in the target space as T _ rainC i 。
(e) Minimum clustering radius (T _ M) of each class in target space min ) Is a very important parameter, and directly influences the classification and the quality of the fault self-growth result. Determining a radius T _ M from a noisy Density-Based Clustering of Applications with Noise (DBSCAN) scatter plot min The value of (c).
(f) And mapping the data of the test set to a target space through S-CNN1 to obtain the coordinates of the test set in the target space, and calculating the centroid of the test set as T _ testC. Solving T _ testC in test set and T _ trainC in training set i Euclidean distance between T _ testC and I-th fault T _ trainC i Is less than T _ M min If so, judging that the unknown fault in the test set belongs to the ith fault type in the training set; when T _ testC and all T _ trainCs in the training set i Are not less than T _ M min And judging that the unknown fault in the test set belongs to a new fault type, thereby realizing the classification and self-learning process of the fault type.
3, establishing a second-stage S-CNN2 fault damage level self-learning network
(a) Data preprocessing: and restoring the samples of the identified fault types (namely the data identified by the S-CNN1 network) into original vibration signal data, and then carrying out fault damage level judgment and self-learning. Firstly, carrying out overlapping segmentation pretreatment on the data, and dividing the data into input samples according to Q sampling points of each segment; and then, carrying out STFT (space time Fourier transform) on each section to obtain a time-frequency graph containing time domain and frequency domain information corresponding to each group of vibration data, and then compressing the time-frequency graph to convert the time-frequency graph into an n multiplied by n matrix.
(b) Parameter selection: selecting network parameters, wherein the two same CNN network parameters adopted by the second-stage S-CNN2 network are selected as follows: the method comprises the steps of selecting 3 layers of convolutional networks from the CNN, selecting 64 convolutional kernels on the first layer, 128 convolutional kernels on the second layer and 256 convolutional kernels on the third layer, wherein the convolutional kernels are 3 x 3, selecting Adam Optimizer by an optimization algorithm, and setting the maximum training frequency to 20000 times.
(c) Training: after the network parameters are determined, taking the time-frequency diagram (n multiplied by n matrix) of the training set in the step (a) as the input of the S-CNN2, and calculating the corresponding Gw (x matrix) 1 )、Gw(x 2 ) And an energy function Ew (x) 1 ,x 2 ) And training the S-CNN network structure by using the known label Y of the training set. The training process mainly trains parameters of each layer of the network layer by layer according to Back Propagation (BP), and finally the trained network structure S-CNN2 is obtained. The whole training process is similar to the S-CNN1 network training process and is not described in detail.
(d) Testing and self-learning processes: after the network structure is determined, the matrix of the test set n multiplied by n in the step (1) is used as the input of the S-CNN2, and a coordinate value (x, y) mapped to a target space is obtained by a time-frequency diagram of each fault vibration signal passing through the last layer of the S-CNN 2. After 20000 times of training, calculating the centroid of each fault degree in the target space set in the training set, and recording the centroid of the ith class in the target space as L _ rainC i 。
(e) Similar to the previous network, determining the radius second-level network L _ M according to the scatter diagram of the DBSCAN min The value of (c).
(f) And mapping the data of the test set to a target space through S-CNN2 to obtain the coordinates of the test set in the target space, and calculating the centroid of the test set as L _ testC. Solving L _ testC in test set and L _ train C in training set i Euclidean distance between the two classes, when L _ testC and i-th class damage degree L _ trainC i Is less than L _ M min Judging that the unknown fault in the test set belongs to the i-th damage degree; when the distance between the L _ testC and all the damage degrees in the training set is not less than the L _ M min And judging that the unknown faults in the test set belong to the new type of damage degree, thereby realizing the classification and self-learning process of the fault damage degree.
Advantageous effects
The general algorithm flow chart of the invention is shown in fig. 2. The invention provides a rolling bearing fault intelligent learning method based on two-stage S-CNN, which provides a better solution for solving the data self-learning problem in the current industrial production and has certain practical value.
Drawings
FIG. 1 is a general block diagram of the algorithm.
Figure 2 is a general algorithm flow diagram.
Fig. 3 morphological filtering process.
FIG. 4 is a S-transform time-frequency spectrum of an original signal.
Fig. 5 is a time-frequency diagram obtained by performing S transform after morphological filtering of the original signal.
Fig. 6S-CNN 1 shows the partial results of the results of each failure mapped to the target space.
FIG. 7 is a time-frequency diagram obtained by STFT of the original data.
8S-CNN 2 each fault maps to a partial result of the target space.
Fig. 9S-CNN 2 illustrates partial results of mapping of various failures to the target space.
FIG. 10 is a schematic diagram of a twin structure of the similarity metric.
Detailed Description
The invention is described in detail below with reference to the figures and examples.
In order to show the effectiveness of the method, a bearing fault experiment platform is adopted to collect data for algorithm verification, and the experiment aims at a two-horsepower motor bearing, and the model is 6205-2RS JEM SKF. Through the electric spark machining technology, single-point faults are respectively arranged on a rolling body, an inner ring and an outer ring of the bearing, the fault damage grades comprise 0.007inch, 0.014inch and 0.021inch, a vibration acceleration signal of the fault bearing is collected through an acceleration sensor arranged on a bearing seat at the driving end of the motor, the sampling frequency is 12kHz, and experimental data are shown in Table 1.
TABLE 1 bearing failure data
1. Fault type classification model self-learning
First, level _1, level _2 and level _3 are defined to represent the failure of the rolling bearing with damage sizes of 0.007in, 0.014in and 0.021in, respectively. Capital letters I, O and B are used to represent fault types of the inner ring, the outer ring and the balls respectively, for example, I _ level _1 represents the fault of the bearing with the damage degree of 0.007inch of the inner ring, and N represents the normal state of the bearing, and subsequent expressions are similar to the fault types and are not repeated. When the fault type self-learning is verified, a data set is constructed as follows: and (3) training by using incomplete data, wherein two faults are arbitrarily selected from each group of experiments as known types, different experiments are designed in each group, one fault is optionally selected from each experiment as an unknown fault, and the fault is determined to belong to a known fault or a new type of unknown fault according to a judgment result. Table 2 shows the selected combinations of fault types for each group and each experiment.
Table 2 fault type classification self-growth experimental grouping
As shown in fig. 3, the filtering result of the morphological opening operation is performed on the vibration signal, the dotted line at the upper half is the original signal, and the solid line at the lower half is the filtered signal. The lower envelope of the signal is obtained by the shape open operation, the pulse characteristic of the signal can be effectively enhanced by using the shape filter, and part of noise is suppressed. As shown in fig. 4, a time-frequency diagram obtained by performing S transform on 3 kinds of original fault data is shown. As can be seen from fig. 4, there is similarity between the spectrograms of the same fault type signals after S-conversion, but it can be found from the graphs that there is similarity between the pulse waveforms in the time-frequency graphs of the inner circle (fig. 4(d)) and the outer circle (fig. 4(f)), and it is obvious that certain misjudgment is caused for the following classification identification. To solve the problem, morphological filtering is added before S transformation, and interference of partial noise is removed. As can be seen from fig. 5, after morphological filtering, S transformation is performed to obtain a spectrogram, which has strong similarity to the same fault and great difference between different faults. Therefore, it is necessary to add morphological filtering preprocessing before S transform.
The analysis is performed by taking the group 2 experiment as an example, the training sets in the group 2 are B _ level _1 and O _ level _1, and the classification result in the target space is shown in fig. 6. FIG. 6a shows the output of the last layer of network with test sets B _ level _1 (experiment 7) and O _ level _1 (experiment 8), where the upper right corner in the figure is the position of O _ level _ 1in the target space, and the lower left corner is the position of B _ level _1, and it can be seen that S-CNN1 has the characteristic of converging similar faults and separating different types of faults. In fig. 6a, all the frequency spectrum graphs obtained by applying the above-mentioned preprocessing method to the fault data are frequency spectrum graphs, and for easier identification, the time frequency graph of each fault is replaced by a number, where the number "2" represents B _ level _1, the number "4" represents O _ level _1, and fig. 6B is obtained by replacing the number in fig. 6 a. FIG. 6c shows the positions of the centroids of the training sets B _ level _1 and O _ level _1 at the target spatial coordinates.
In order to enable each fault to have a reference object in the target space, the converging and separating characteristics in the target space are better shown. Meanwhile, the test sets B _ level _1, O _ level _1, and B _ level _3 (i.e., experiment 7, experiment 8, and experiment 9) are mapped to the target space, and the result is shown in fig. 6d, where the numeral "5" represents B _ level _ 3. In the figure, it can be obviously seen that "2" and "5" converge in the target space, which indicates that B _ level _1 and B _ level _3 are the same type of fault, and have stronger similarity in the time-frequency diagram, and the two samples are fused to the lower left corner of the target space. Meanwhile, the test sets B _ level _1, O _ level _1, and O _ level _3 (i.e., experiment 7, experiment 8, and experiment 12) are mapped to the target space, and the result is shown in fig. 8e, where the numeral "5" represents O _ level _ 3. Similar to FIG. 8d, "4" and "5" merge into the upper right corner of the target space, since O _ level _1 and O _ level _3 are the same type of failure. By the above example, it can be shown that the S-CNN1 has strong identification capability for the same type of failure with different damage degrees which are not learned.
Meanwhile, the test sets B _ level _1, O _ level _1, and I _ level _3 (i.e., experiment 7, experiment 8, and experiment 11) are output in the target space, and the result is shown in fig. 6f, where the numeral "5" represents I _ level _ 3. It can be seen that the graphs collectively form 3 clusters of sets, that is, since B _ level _1, O _ level _1 and I _ level _3 have no similarity in the time-frequency graph, they each form a cluster after being mapped to the target space. It can be seen from the figure that the S-CNN1 has good identification capability for different types of unknown faults which are not learned.
Analyzing the mapping result of the training set in the target space, and determining the minimum clustering radius (T _ M) according to the scatter diagram of the DBSCAN min ) The value of (c). By calculating T _ M min =1.5。
TABLE 3 distance of centroid coordinates in target space and classification results
(a) The test set is the classification result of B _ level _1 and I _ level _2
(b) The test set is the classification result of B _ level _1 and O _ level _1
(c) The test set is the classification result of I _ level _3 and O _ level _1
And calculating Euclidean distances (T _ D) between the training set and each centroid in the test set. The coordinates and distances of each centroid in the training set and the test set are shown in table 3. According to the discriminant rule (if the distance T _ D between the unknown fault in the test set and the i-th type fault of the known fault type in the training set it <T_M min Unknown fault belongs to the known fault) to identify the fault type for each test sample. The results of all experiments are listed in table 3, where "NEW" represents a NEW type of failure. As can be seen from Table 3, the classification accuracy of each experiment is very high, which shows that the method can realize self-learning of fault types and self-growth of fault libraries under the condition of incomplete data modeling, and the effectiveness of the proposed algorithm is verified.
2. Fault damage degree classification model self-learning
The second-level network is used for searching the differences of the same type of faults with different damage degrees, and the S transformation can find the similarity of the same type of faults with different damage degrees in the time-frequency domain more easily according to the work, so that different preprocessing strategies are adopted in the first-level network. STFT observes signals through a sliding window of fixed size, which is more sensitive to small changes in signals and makes it easier to find differences in different signals.
Table 4 fault damage degree classification self-growth experimental grouping
The vibration data is subjected to overlapping segmentation preprocessing, and then each segment of data is subjected to STFT to obtain a time-frequency diagram as shown in FIG. 7. As can be seen from the graph, the time-frequency graphs of the same type of faults with different damage degrees have obvious differences. Therefore, it is reasonable for the second level network to employ this preprocessing method. Next, 6 sets of experiments were constructed as shown in table 4. Groups 4-6 are classification and learning processes for the case where normal signals and a fault damage level signal are known, and groups 7-9 are further classification and learning processes for the case where two fault damage level signals are known.
(1) A process of identifying the extent of damage is known: the training sets in group 6 are N and O _ level _1, and their spatial distribution in the target is shown in fig. 8. FIG. 8a shows the output of the last layer network of the training set N and O _ level _1, where the number "1" represents N, the number "4" represents O _ level _1, N converges at the lower left corner of the graph at the target spatial coordinate value, and O _ level _1 converges at the upper right corner of the graph.
In order to observe the distribution of each fault in the target space, the test sets N and O _ level _1 (i.e. experiments 27 and 28) are mapped to the target space, and the result is shown in fig. 8 b. The test set samples are also more centrally distributed in two regions of the target space, almost coinciding with the locations where the samples of the training set are located. It is shown that S-CNN2 can be used to identify known damage levels. Meanwhile, mapping the test sets N, O _ level _1 and O _ level _2 (i.e. experiment 27, experiment 28 and experiment 29) to the target space results are shown in FIG. 8 c. Because O _ level _1 and O _ level _2 are the same fault type and have similarity in time frequency spectrum, the fault damage level signal O _ level _2 in the test set is closer to O _ level _1, but the two types of damage levels can still be clearly found or separated into two clusters. FIG. 8d is the output result obtained by mapping the test sets N, O _ level _1 and O _ level _2 (i.e. experiment 27, experiment 28, and experiment 30) to the target space at the same time, and O _ level _3 (indicated by "5" in the figure), because O _ level _1 and O _ level _3 are the same fault type signals, the test set O _ level _3 will be closer to O _ level _ 1in the target space, but still effectively separated into two clusters. Analyzing the training set in a target space, and determining the minimum clustering radius (L _ M) according to the scatter diagram of the DBSCAN min ) The value of (c). By calculating L _ M min 0.1. The Euclidean distance (L _ D) between each centroid of the training set and the test set is calculated. According to the discriminant rule (if the distance L _ D between the unknown fault degree in the test set and the ith fault degree of the known fault degree in the training set it <L_M min1 Unknown fault level belongs to the known fault level) to identify the fault damage level of each test sample. All experimental results are listed in table 5, where "NEW" represents a NEW type of failure; "L-1" is represented as a first type of lesion grade. As can be seen from table 5, the classification accuracy of each experiment was high.
TABLE 5 distance of centroid coordinates in target space and classification results
(a) Training set N and B _ level _1 classification result
(b) Training set N and I _ level _1 classification result
(c) Training set N and O _ level _1 classification result
(2) And when two fault damage grades are known, realizing self-learning and grade self-increasing of a third fault damage degree: the training sets in group 7 are B _ level _1 and B _ level _2, and their spatial distribution in the target is shown in FIG. 9. FIG. 9a shows the outputs of the last layer networks of training sets B _ level _1 and B _ level _2, where B _ level _1 (indicated by "2" in the figure) converges at the upper left corner of the figure at the target spatial coordinate value and B _ level _2 (indicated by "3" in the figure) converges at the lower right corner of the figure.
Meanwhile, the test sets B _ level _1 and B _ level _2 (i.e., experiment 31 and experiment 32) are mapped to the target space, and the result is shown in fig. 9B. Fig. 9c shows that B _ level _3 is represented by "5" in the target space mapping result graph by using the test sets B _ level _1, B _ level _2, and B _ level _3 (experiment 31, experiment 32, and experiment 33) at the same time, and it can be clearly seen that three damage levels are distributed in the target space into three cluster sets, which are easily divided into three types.
Analyzing the training set in a target space, and determining a minimum clustering radius (L _ M) according to a scatter diagram of DBSCAN min2 ) The value of (c). By calculating L _ M min2 1.5. According to similar discrimination rules, classification and self-learning of 3 fault damage degrees can be realized. The results of all experiments are shown in Table 6, where "L-2" represents the second type of injury rating. As can be seen from table 6, the classification accuracy of each experiment was high.
The above experiments are combined to show that the method can realize self-learning of fault damage levels, and indirectly show that the method can continue classified self-growth when the damage degree is further increased, so that certain learning intelligence is embodied.
TABLE 6 distance of centroid coordinates in target space and classification results
(a) Training the classification result of B _ level _1 and B _ level _2
(b) Classification results of training sets I _ level _1 and I _ level _2
(c) The classification results of the training sets O _ level _1 and O _ level _2
Claims (2)
1. A rolling bearing fault self-learning method based on a two-stage twin convolution neural network is characterized by comprising the following steps: firstly, combining a twin structure with similarity measurement with CNN to construct an S-CNN network structure; then, based on the constructed S-CNN network structure, establishing a first-level fault type identification network S-CNN 1: (1) removing part of noise by using the morphological enhancement pulse characteristics; (2) extracting the signal fault type common characteristics by applying S transformation, and constructing a time-frequency diagram; (3) inputting the time-frequency diagram sample into a first-stage S-CNN, training the time-frequency diagram sample to finally form a first-stage fault type identification network S-CNN1, wherein the trained network utilizes the characteristics of the network in a target space for converging and separating faults of the same type and different types to realize the identification and self-learning of the fault type of the rolling bearing;
and finally, establishing a second-stage fault damage degree classification network S-CNN 2: (1) amplifying differences of different damage degrees of the same type of faults by using short-time Fourier transform (STFT) with a sliding window to obtain a sliding time-frequency diagram sample; (2) inputting the obtained time-frequency pattern and the time-frequency pattern into a second-stage S-CNN, training the time-frequency pattern and the time-frequency pattern to finally form a second-stage fault damage degree classification network S-CNN2, wherein the trained network is a classifier based on a target space distance, and self-learning and self-growth of fault damage degrees are realized;
the specific process of establishing the first stage S-CNN1 fault type self-learning network is as follows,
(a) data preprocessing: assuming that there is one unknown fault type X and several known fault types X i If i is more than or equal to 2, performing mathematical morphology filtering on all original vibration signals, removing part of background noise, and dividing data according to each group of Q sampling points to be used as input samples; then, each group is subjected to S transformation processing to obtain a time-frequency graph containing time domain and frequency domain information corresponding to each group of vibration data, and then the time-frequency graph is converted into an n multiplied by n matrix; taking 80% of all signal data as training set data and 20% as test set data;
(b) parameter selection: selecting network parameters, wherein the two same CNN network parameters adopted by the first-stage S-CNN1 network are selected as follows: selecting 5 layers of convolution networks, namely a first layer of 16 convolution kernels, a second layer of 32 convolution kernels, a third layer of 64 convolution kernels, a fourth layer of 128 convolution kernels and a fifth layer of 256 convolution kernels, wherein the sizes of the convolution kernels are 5 multiplied by 5, selecting Adam by an optimization algorithm, selecting the hyper-parameters by the Adam to have good robustness, and training the maximum times to be 4000 times;
(c) training process: after the network parameters are determined, taking the time-frequency diagram of the training set in the step (a), namely n multiplied by n matrix, as the input of the S-CNN1, and calculating the corresponding Gw (x) 1 )、Gw(x 2 ) And an energy function Ew (x) 1 ,x 2 ) Training the S-CNN network structure by using the known label Y of the training set; the training process mainly comprises the steps of training parameters of each layer of the network layer by layer according to back propagation, and finally obtaining a trained network structure S-CNN 1;
(d) the testing and self-learning process comprises the following steps: after the network is trained, the test set time-frequency diagram (n multiplied by n matrix) in the step (a) is used as the one trained in the step (c)When the S-CNN1 is input, a time-frequency graph of each fault vibration signal can obtain a coordinate value (x, y) mapped to a target space through the last layer of the S-CNN1, and the fact that faults of the same type are closer to the target space and faults of different types are farther from the target space can be known through the twin structure principle; after 4000 times of training, finally, the two types of faults can be converged into two clusters which are not overlapped in the target space, and the learning process of the network on different characteristics is completed; calculating the mass center of each type of fault in the training set in the target space set, and recording the mass center of the ith type in the target space as T _ rainC i ;
(e) Determining radius T _ M from a scatter plot with noise based on a density clustering method min A value of (d);
(f) mapping the test set data to a target space through S-CNN1 to obtain the coordinates of the test set in the target space, and calculating the centroid of the test set as T _ testC; solving T _ testC of test set and T _ trainC in training set i Euclidean distance between T _ testC and I-th failed T _ trainC i Is less than T _ M min Judging that the unknown fault in the test set belongs to the ith fault type in the training set; when T _ testC and all T _ trainCs in the training set i Is not less than T _ M min Judging that the unknown fault in the test set belongs to a new fault type, thereby realizing the classification and self-learning process of the fault type;
the specific process of establishing the second stage S-CNN2 fault damage level self-learning network is as follows,
(a) data preprocessing: restoring the identified fault type, namely the sample of the data identified by the S-CNN1 network into original vibration signal data, and then carrying out fault damage grade discrimination and self-learning; firstly, carrying out overlapping segmentation pretreatment on the data, and dividing the data into input samples according to Q sampling points of each segment; then, each section is subjected to STFT (space time Fourier transform) conversion to obtain a time-frequency graph which corresponds to each group of vibration data and contains time domain and frequency domain information, and then the time-frequency graph is compressed and converted into an n multiplied by n matrix;
(b) parameter selection: selecting network parameters, wherein the two same CNN network parameters adopted by the second-level S-CNN2 network are selected as follows: selecting 3 layers of convolution networks from the CNN, wherein the first layer comprises 64 convolution kernels, the second layer comprises 128 convolution kernels, the third layer comprises 256 convolution kernels, the convolution kernels are 3 x 3, the optimization algorithm selects an Adam Optimizer, and the maximum training frequency is 20000 times;
(c) training process: after the network parameters are determined, taking the time-frequency diagram n multiplied by n matrix of the training set in the step (a) as the input of the S-CNN2, and calculating the corresponding Gw (x) 1 )、Gw(x 2 ) And an energy function Ew (x) 1 ,x 2 ) Training the S-CNN network structure by using the known label Y of the training set; in the training process, parameters of each layer of the network are trained layer by layer according to back propagation, and a trained network structure S-CNN2 is finally obtained;
(d) testing and self-learning processes: after the network structure is determined, taking a matrix of the test set n multiplied by n as the input of the S-CNN2, and obtaining a coordinate value (x, y) mapped to a target space through the last layer of the S-CNN2 of a time-frequency graph of each fault vibration signal; after 20000 times of training, calculating the centroid of each fault degree in the target space set in the training set, and recording the centroid of the ith class in the target space as L _ rainC i ;
(e) Determining radius second-level network L _ M according to scatter diagram of DBSCAN min A value of (d);
(f) mapping the test set data to a target space through S-CNN2 to obtain the coordinates of the test set in the target space, and calculating the centroid of the test set as L _ testC; solving L _ testC in test set and L _ train C in training set i Euclidean distance between the two classes, when L _ testC and i-th class damage degree L _ trainC i Is less than L _ M min Judging that the unknown fault in the test set belongs to the i-th damage degree; when the distance between the L _ testC and all the damage degrees in the training set is not less than the L _ M min And judging that the unknown faults in the test set belong to the new type of damage degree, thereby realizing the classification and self-learning process of the fault damage degree.
2. The self-learning method for the faults of the rolling bearing based on the two-stage twin convolutional neural network as claimed in claim 1, wherein the self-learning method comprises the following steps: the S-CNN structure is essentially a similarity measurement method and inputIs two time-frequency graphs x 1 And x 2 At the time of model training, they have a known label Y, where Y is a binary label; when time-frequency diagram x is input 1 And x 2 When belonging to the same class, Y is 0; otherwise, Y is 1, and the label is used for network training; then, Gw (x) 1 ) And Gw (x) 2 ) Are each x 1 And x 2 Through the corresponding output of the two CNN networks, the two CNN networks will x 1 And x 2 Mapping to two points in a low dimensional space; wherein, w is a shared parameter to be learned by two CNN networks; then, the outputs Gw (x) of the two networks are combined 1 ) And Gw (x) 2 ) Calculating to obtain a similarity measurement index Ew, and calling the index as an energy function; the energy function Ew (x) 1 ,x 2 ) The definition is as follows:
Ew(x 1 ,x 2 )=||Gw(x 1 )-Gw(x 2 )|| (1)
the above is the constructed S-CNN basic structure, and then different S-CNN networks are trained by inputting different training samples.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010066718.6A CN111539152B (en) | 2020-01-20 | 2020-01-20 | Rolling bearing fault self-learning method based on two-stage twin convolutional neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010066718.6A CN111539152B (en) | 2020-01-20 | 2020-01-20 | Rolling bearing fault self-learning method based on two-stage twin convolutional neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111539152A CN111539152A (en) | 2020-08-14 |
CN111539152B true CN111539152B (en) | 2022-08-26 |
Family
ID=71976663
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010066718.6A Active CN111539152B (en) | 2020-01-20 | 2020-01-20 | Rolling bearing fault self-learning method based on two-stage twin convolutional neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111539152B (en) |
Families Citing this family (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112115999B (en) * | 2020-09-15 | 2022-07-01 | 燕山大学 | Wind turbine generator fault diagnosis method of space-time multi-scale neural network |
CN112881017A (en) * | 2021-01-07 | 2021-06-01 | 西北工业大学 | Intelligent fault diagnosis method for aeroengine control system sensor based on mode gradient spectral entropy |
CN112881018A (en) * | 2021-01-07 | 2021-06-01 | 西北工业大学 | Intelligent fault diagnosis method of aeroengine control system sensor based on improved mode gradient spectral entropy |
CN113050602B (en) * | 2021-03-26 | 2022-08-09 | 杭州电子科技大学 | Industrial process fault classification method based on robust semi-supervised discriminant analysis |
CN113138080B (en) * | 2021-04-22 | 2022-06-03 | 东北大学 | Rolling bearing intelligent fault diagnosis method based on vibration twinning and extreme learning |
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 |
CN114157552B (en) * | 2021-10-29 | 2024-04-05 | 国网河南省电力公司漯河供电公司 | Distribution network fault detection method based on twin time sequence diagram network |
CN115753103A (en) * | 2022-08-30 | 2023-03-07 | 山东建筑大学 | Fault diagnosis method and system based on standard self-learning data enhancement |
CN117997690B (en) * | 2024-04-03 | 2024-06-18 | 厦门大学 | Class comparison learning type underwater sound signal modulation recognition system and recognition method |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106874957A (en) * | 2017-02-27 | 2017-06-20 | 苏州大学 | A kind of Fault Diagnosis of Roller Bearings |
CN107421741A (en) * | 2017-08-25 | 2017-12-01 | 南京信息工程大学 | A kind of Fault Diagnosis of Roller Bearings based on convolutional neural networks |
CN107957551A (en) * | 2017-12-12 | 2018-04-24 | 南京信息工程大学 | Stacking noise reduction own coding Method of Motor Fault Diagnosis based on vibration and current signal |
CN109632309A (en) * | 2019-01-17 | 2019-04-16 | 燕山大学 | Based on the rolling bearing fault intelligent diagnosing method for improving S-transformation and deep learning |
CN109765053A (en) * | 2019-01-22 | 2019-05-17 | 中国人民解放军海军工程大学 | Utilize the Fault Diagnosis of Roller Bearings of convolutional neural networks and kurtosis index |
CN110132598A (en) * | 2019-05-13 | 2019-08-16 | 中国矿业大学 | Slewing rolling bearing fault noise diagnostics algorithm |
CN110490218A (en) * | 2019-06-10 | 2019-11-22 | 内蒙古工业大学 | A kind of rolling bearing fault self-learning method based on two-stage DBN |
-
2020
- 2020-01-20 CN CN202010066718.6A patent/CN111539152B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106874957A (en) * | 2017-02-27 | 2017-06-20 | 苏州大学 | A kind of Fault Diagnosis of Roller Bearings |
CN107421741A (en) * | 2017-08-25 | 2017-12-01 | 南京信息工程大学 | A kind of Fault Diagnosis of Roller Bearings based on convolutional neural networks |
CN107957551A (en) * | 2017-12-12 | 2018-04-24 | 南京信息工程大学 | Stacking noise reduction own coding Method of Motor Fault Diagnosis based on vibration and current signal |
CN109632309A (en) * | 2019-01-17 | 2019-04-16 | 燕山大学 | Based on the rolling bearing fault intelligent diagnosing method for improving S-transformation and deep learning |
CN109765053A (en) * | 2019-01-22 | 2019-05-17 | 中国人民解放军海军工程大学 | Utilize the Fault Diagnosis of Roller Bearings of convolutional neural networks and kurtosis index |
CN110132598A (en) * | 2019-05-13 | 2019-08-16 | 中国矿业大学 | Slewing rolling bearing fault noise diagnostics algorithm |
CN110490218A (en) * | 2019-06-10 | 2019-11-22 | 内蒙古工业大学 | A kind of rolling bearing fault self-learning method based on two-stage DBN |
Non-Patent Citations (3)
Title |
---|
Feature extraction for rolling element bearing fault diagnosis utilizing generalized S transform and two-dimensional non-negative matrix factorization;Bing Li等;《Journal of Sound and Vibration》;20101222;第330卷;第2388–2399页 * |
Intelligent fault diagnosis of rolling bearing using hierarchical convolutional network based health state classification;Chen Lu等;《Advanced Engineering Informatics》;20170227;第32卷;第139-151页 * |
基于两级DBN的滚动轴承故障自学习方法;师芳;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;20200115;C029-237 * |
Also Published As
Publication number | Publication date |
---|---|
CN111539152A (en) | 2020-08-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111539152B (en) | Rolling bearing fault self-learning method based on two-stage twin convolutional neural network | |
AU2020103923A4 (en) | Fault diagnosis method and system for gear bearing based on multi-source information fusion | |
CN110849626B (en) | Self-adaptive sparse compression self-coding rolling bearing fault diagnosis system | |
Chen et al. | Dual-path mixed-domain residual threshold networks for bearing fault diagnosis | |
CN107784325B (en) | Spiral fault diagnosis method based on data-driven incremental fusion | |
CN109827777B (en) | Rolling bearing fault prediction method based on partial least square method extreme learning machine | |
CN110070060B (en) | Fault diagnosis method for bearing equipment | |
CN111562108A (en) | Rolling bearing intelligent fault diagnosis method based on CNN and FCMC | |
CN112257530B (en) | Rolling bearing fault diagnosis method based on blind signal separation and support vector machine | |
CN110647830B (en) | Bearing fault diagnosis method based on convolutional neural network and Gaussian mixture model | |
CN114358123B (en) | Generalized open set fault diagnosis method based on deep countermeasure migration network | |
CN111753891B (en) | Rolling bearing fault diagnosis method based on unsupervised feature learning | |
Zhang et al. | Multi-model ensemble deep learning method for intelligent fault diagnosis with high-dimensional samples | |
CN113375941A (en) | Open set fault diagnosis method for high-speed motor train unit bearing | |
CN110991471B (en) | Fault diagnosis method for high-speed train traction system | |
CN114169377A (en) | G-MSCNN-based fault diagnosis method for rolling bearing in noisy environment | |
Liao et al. | Research on a rolling bearing fault detection method with wavelet convolution deep transfer learning | |
CN114118138A (en) | Bearing composite fault diagnosis method based on multi-label field self-adaptive model | |
CN115824519B (en) | Comprehensive diagnosis method for valve leakage faults based on multi-sensor information fusion | |
CN114429152A (en) | Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption | |
Chen et al. | A visualized classification method via t-distributed stochastic neighbor embedding and various diagnostic parameters for planetary gearbox fault identification from raw mechanical data | |
CN114462480A (en) | Multi-source sensor rolling mill fault diagnosis method based on non-equilibrium data set | |
CN115392333A (en) | Equipment fault diagnosis method based on improved end-to-end ResNet-BilSTM dual-channel model | |
CN115587290A (en) | Aero-engine fault diagnosis method based on variational self-coding generation countermeasure network | |
CN117030263A (en) | Bearing fault diagnosis method based on improved residual error network under multi-sensor signal fusion |
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 |