CN114330096B - Rail transit motor bearing fault degree classification model construction method based on deep learning - Google Patents
Rail transit motor bearing fault degree classification model construction method based on deep learning Download PDFInfo
- Publication number
- CN114330096B CN114330096B CN202111409465.9A CN202111409465A CN114330096B CN 114330096 B CN114330096 B CN 114330096B CN 202111409465 A CN202111409465 A CN 202111409465A CN 114330096 B CN114330096 B CN 114330096B
- Authority
- CN
- China
- Prior art keywords
- network
- fault
- layer
- classification
- data
- 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
- 238000013135 deep learning Methods 0.000 title claims abstract description 30
- 238000013145 classification model Methods 0.000 title claims abstract description 15
- 238000010276 construction Methods 0.000 title description 4
- 238000012549 training Methods 0.000 claims abstract description 55
- 238000000034 method Methods 0.000 claims abstract description 51
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 34
- 238000005070 sampling Methods 0.000 claims abstract description 34
- 230000001133 acceleration Effects 0.000 claims abstract description 29
- 230000015654 memory Effects 0.000 claims abstract description 19
- 238000012795 verification Methods 0.000 claims description 29
- 238000012360 testing method Methods 0.000 claims description 15
- 238000012545 processing Methods 0.000 claims description 12
- 238000011176 pooling Methods 0.000 claims description 11
- 230000006870 function Effects 0.000 claims description 10
- 230000008569 process Effects 0.000 claims description 10
- 238000004364 calculation method Methods 0.000 claims description 9
- 238000000605 extraction Methods 0.000 claims description 7
- 230000007787 long-term memory Effects 0.000 claims description 7
- 230000006403 short-term memory Effects 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 4
- 238000009432 framing Methods 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- 230000005540 biological transmission Effects 0.000 claims description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 2
- 230000001419 dependent effect Effects 0.000 claims description 2
- 230000007774 longterm Effects 0.000 claims description 2
- 230000000737 periodic effect Effects 0.000 claims description 2
- 230000001174 ascending effect Effects 0.000 claims 1
- 238000003745 diagnosis Methods 0.000 abstract description 15
- 238000012544 monitoring process Methods 0.000 abstract description 4
- 238000001514 detection method Methods 0.000 abstract description 3
- 230000036541 health Effects 0.000 abstract description 3
- 230000008901 benefit Effects 0.000 abstract description 2
- 238000005457 optimization Methods 0.000 abstract description 2
- 238000012423 maintenance Methods 0.000 description 13
- 238000010586 diagram Methods 0.000 description 8
- 238000004458 analytical method Methods 0.000 description 7
- 238000011161 development Methods 0.000 description 3
- 230000009466 transformation Effects 0.000 description 3
- 230000007547 defect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000003137 locomotive effect Effects 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000001228 spectrum Methods 0.000 description 2
- 238000010200 validation analysis Methods 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 238000010183 spectrum analysis Methods 0.000 description 1
Classifications
-
- 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
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention discloses a method for constructing a classification model of the fault degree of a rail transit motor bearing based on deep learning, and relates to the field of motor bearing fault diagnosis in rail transit. The method specifically uses the rail transit motor bearing acceleration sensor signal to carry out deep learning network training, and then the deep learning network adopts a convolutional neural network or a long-short-term memory network, so that a novel efficient and accurate solution model can be provided for rail transit motor bearing health monitoring. And an acceleration sensor is arranged on the traction motor to collect and intelligently analyze motor vibration signals in real time, so that real-time normal state monitoring and fault degree detection of the motor bearing are realized. According to the invention, the network model is trained by using a deep learning method, so that the classification accuracy of fault types and fault degrees can be effectively improved; the data sampling is a single-axis acceleration sensor, so that the hardware cost of data acquisition can be effectively reduced, and economic benefits are brought; the type and the depth of the network can be selected to perform model optimization by combining the characteristics of the data.
Description
Technical Field
The invention relates to the field of motor bearing fault diagnosis in rail transit, in particular to a method for constructing a rail transit motor bearing fault degree classification model based on deep learning.
Background
In recent years, with the development and construction of national rail transit, the total mileage of rail transit operation and the total number of vehicles are continuously increased, and the complexity and integration of train vehicles are continuously improved, which puts higher demands on the safety and maintenance efficiency of rail transit equipment. However, the current common train maintenance and repair mode cannot completely keep pace with the development of rail transit, and the current maintenance mode not only consumes more manpower resources, but also causes equipment faults due to untimely maintenance and other reasons, thereby causing economic loss and even endangering the safety of passengers. Current rail transit vehicle "planned maintenance" and "post-maintenance" maintenance modes have failed to meet their maintenance needs, and fault Prediction and Health Management (PHM) technologies with "on-demand maintenance" capability are receiving increasing attention.
The large and complex size of rail transit vehicles results in high failure rates. The motor of the rail transit vehicle is the most critical part of the transmission chain part, and the analysis of the motor bearing faults and severity can directly influence whether the vehicle runs normally. Predictive maintenance is one of the key innovation points proposed by "industry 4.0". Based on the continuous measurements and analysis, predictive maintenance can predict institutional metrics such as the remaining useful life of machine parts. The key operation parameter data can assist in decision making, judging the operation state of the machine and optimizing the maintenance time of the machine. One notable direction of development is the application of deep learning in predictive maintenance. The method is characterized in that the signals of the acceleration sensor of the rail transit motor bearing are monitored, and the diagnosis and prediction of faults are realized through analysis, statistics and classification of the signals. The method is particularly important for processing sensor data and constructing a fault diagnosis model of a deep learning algorithm.
In the prior art, there are two directions related to this, as follows:
One is to extract features from time-domain or frequency-domain signals and then analyze the features using a machine learning algorithm in the processing of acceleration data for rail transit motor bearings. When the bearing breaks down, the vibration amplitude of the bearing can be greatly increased, and a corresponding impact signal can be generated. The peak index is generally used for detecting impact vibration caused by bearing flaking, cracks and the like, and the kurtosis index is used for diagnosing the earliest faults of the bearing. Waveform indicators are often used to detect mechanical failure of bearing components due to discrete imperfections such as pits, nicks, flaking, and scratches. In the bearing fault diagnosis process, the fault characteristics of the rolling bearing are usually modulation phenomenon, the time domain waveform of the vibration signal is complex, and the fault signal characteristics cannot be intuitively expressed. The vibration frequency signal has objectivity and can better reflect the basic characteristics of the vibration signal. The time domain vibration signal is converted into the frequency domain vibration signal through Fourier transformation, and the frequency spectrum can intuitively reflect the energy size, the frequency composition and the phase of the vibration signal. However, this method has the following drawbacks: when training the time domain or frequency domain feature model of the motor bearing acceleration sensor data, the 3-axis signal of the acceleration sensor is often required to extract the features. Moreover, when the bearing professional knowledge of algorithm engineers is deficient, the bearing fault degree is different, noise interference and the like, the classification effect of the model is poorer, the classification precision is not high, hidden hazards of misjudgment and missed judgment exist, the generalization capability of the model is poorer, and the model is difficult to simultaneously respond to signal characteristics of different rotating speed conditions.
Other directions: the time domain fault diagnosis method and the frequency domain fault diagnosis method have certain limitations, and the comprehensiveness and the locality of the vibration signal can not be well reflected, so that a new diagnosis method is generated, namely the time-frequency fault diagnosis method. The time-frequency fault diagnosis method organically combines the time domain and the frequency domain into a time-frequency phase plane to obtain vibration signal spectrograms at different moments. The time-frequency fault diagnosis methods widely used at present mainly comprise Hilbert-Huang transform (HHT) and an envelope modulation method. The envelope demodulation method utilizes envelope detection and spectrum analysis to the envelope signal, and then diagnoses and identifies faults according to the demodulated spectrum peaks. The envelope demodulation method is particularly suitable for high-frequency impact vibration, and until now, the envelope demodulation method is the only effective important analysis means for high-frequency impact vibration. The envelope demodulation method is mainly used for high-precision fault diagnosis of the bearing, and not only can the fault part be diagnosed, but also the fault type can be judged. However, this method has the following drawbacks: the envelope analysis, wavelet transformation and the like used still rely on manual feature extraction and expertise accumulation to a great extent, and higher requirements are put on modeling of a fault algorithm model. Moreover, when the fault classification is too many and the training data amount is small, the model is easy to be subjected to the fitting phenomenon, so that the accuracy of the test set is reduced, and the problem that the training model cannot be generalized is caused.
Therefore, based on the defects of the two methods, a new fault diagnosis method needs to be constructed.
Disclosure of Invention
The invention provides a method for constructing a classification model of the fault degree of a rail transit motor bearing based on deep learning, which aims to solve the defect problem of the rail transit motor bearing fault diagnosis in the method.
The invention aims to use the signal of the acceleration sensor of the rail transit motor bearing to carry out deep learning network training, thereby forming a diagnosis method of the fault type and the fault severity of the rail transit motor bearing. The deep learning network adopts a convolutional neural network or a long-short-time memory network, and can provide a new efficient and accurate solution paradigm for the rail transit motor bearing health monitoring. And an acceleration sensor is arranged on the traction motor to collect and intelligently analyze motor vibration signals in real time, so that real-time normal state monitoring and fault degree detection of the motor bearing are realized. The invention is realized by the following technical scheme:
a method for constructing a rail transit motor bearing fault degree classification model based on deep learning comprises the following steps:
Step 1, presetting bearing fault types and quantity: the number of the types of the preset bearing faults and the severity is A;
step 2, original signal acquisition and framing processing:
a. The track traffic motor is arranged on the joint debugging test bed and runs in a no-load mode; collecting Z-axis direction acceleration signals of a bearing at a transmission end by using a single-axis acceleration sensor, respectively collecting continuous Z-axis direction acceleration signals within a T time range of 1000-4500 rpm under each bearing fault type to obtain A original signals, dividing each original signal into B parts, adding fault type labels to each part, and enabling each original signal to form an original signal data set composed of B time sequence data with fault labels; the signal sampling frequency of the single-axis acceleration sensor is Fs, so that each time series data with the fault tag comprises N sampling points, and N=T/B xFs;
b. N sampling points in each time series data with fault labels of each original signal data set are subjected to a fast algorithm FFT (Fast Fourier Transform) of discrete Fourier transform to obtain FFT results of the N sampling points, so that a time domain value of each sampling point of each time series data with fault labels corresponds to a frequency domain value, each time domain value and the corresponding frequency domain value are combined, each time series data with fault labels is changed from time domain one-dimensional 1 XN to time domain frequency domain two-dimensional 2 XN, before FFT transformation, each time series data with fault labels belongs to a one-dimensional time series, and analysis of signals with only one time domain dimension is difficult, a deep network structure is often needed to be compared, and more calculation resources are consumed to complete identification requirements. Some signals are very difficult to see what features are in the time domain, but if transformed into the frequency domain, the features are easily seen, and FFT has once been regarded as an epoch-making advance in signal analysis technology. Each time sequence data with fault labels is changed from time domain one-dimensional 1 XN to time domain frequency domain two-dimensional 2 XN, and the original N point sequences are sequentially decomposed into a series of short sequences; the symmetrical property and the periodic property of the exponential factor in the discrete Fourier transform calculation are fully utilized, and further, the discrete Fourier transform values corresponding to the short sequences are obtained and combined, so that the purposes of deleting repeated calculation, reducing multiplication operation and simplifying the structure are achieved; in order to realize fast algorithm FFT of discrete Fourier transform, the signal sampling frequency Fs of the single-axis acceleration sensor needs to be more than 2 times of the maximum frequency of the signal, namely the maximum rotating speed (rps) of the motor, so as to meet the shannon sampling theorem;
c. Randomly distributing data according to the proportion of the training set, the verification set and the test set in each original signal data set of 8:1:1, and finally obtaining AXnXB x 80% data as training samples, wherein AXnXB x 10% data are used as verification samples, and the rest AXnXB x 10% data are used as test samples;
Step 3, establishing a deep learning network:
Training a deep learning network model for processing classification tasks, wherein the model is a convolutional neural network CNN or a long and short memory network LSTM; CNNs typically include several key levels, convolution layers, reLU activation layer and pooling layer. Long and short term memory networks (LSTM) are the most classical structures in Recurrent Neural Networks (RNN), and are also well suited for classification tasks involving sequential data:
a. The processing procedure of the convolutional neural network CNN network comprises the following steps: as shown in fig. 2, the two-dimensional sequence data is up-scaled to three-dimensional image data to be input to the image input layer; then, extracting image features by using a convolution layer, wherein the convolution layer, a batch normalization layer, a ReLU activation layer and a pooling layer are used as a complete feature extraction process; when needed, the method can be used as a whole for expansion to form a plurality of convolution characteristic extraction processes; after the feature extraction is completed, a discarding layer is used for preventing the model from being over fitted; then using the fully connected layer for combining the features found before; the classification function layer Softmax performs weighting and classification calculation on the output fully-connected characteristics; the classification layer outputs the result of the network training as the fault type;
b. Long and short term memory network LSTM network processing procedure: as shown in fig. 3, two-dimensional sequence data is input to the sequence input layer; the long-term dependency relationship between time steps of the sequence data is learned by the long-term and short-term memory layer; discarding the layer, and randomly neglecting hidden layer nodes with a certain probability; when needed, the long and short time memory layer and the discarding layer can be used as a whole to be expanded to form a multi-layer memory dependent process; randomly ignoring some hidden nodes when each network iterates to prevent the network from being over-fitted; the full connection layer is used for combining the features found before; the classification function layer Softmax performs weighting and classification calculation on the output fully-connected characteristics; the classification layer outputs the result of the network training as the fault type;
c. Inputting the data in the training sample in the step 2 into a CNN or LSTM network for training, and designating training parameters of the network; verifying the identification accuracy of the current training network by the data in the verification sample in the step 2, achieving the required accuracy, and storing the CNN or LSTM network model at the moment;
Step 4, verifying a model classification result:
The verification sample in the step 2 is led into the CNN model or the LSTM model formed in the step 3 for classification, so that a classification result of fault type and severity degree is obtained, and the overall accuracy of the classification result is obtained; if the total accuracy of the classification does not meet the user index requirement, returning to the step 3 to adjust the depth or the layer number of the CNN or LSTM network, and modifying the training parameters until the total accuracy of the classification meets the requirement.
Compared with the prior art, the invention has the following beneficial effects: according to the track traffic motor bearing fault degree classification model construction method based on deep learning, which is provided by the invention, (1) the network model is trained by using the deep learning method, so that the classification accuracy of fault types and fault degrees can be effectively improved; (2) The data sampling is a single-axis acceleration sensor, so that the hardware cost of data acquisition can be effectively reduced, and considerable economic benefits are brought. (3) The data characteristics can be combined, and the type and the depth of the network can be selected for model optimization.
Drawings
FIG. 1 is a diagram of a fault identification process based on deep learning.
Fig. 2 is a network structure diagram of a convolutional neural network.
Fig. 3 is a network structure of a long and short memory network.
Fig. 4 is a time domain waveform diagram for each type of fault type in two embodiments.
Fig. 5 is a schematic diagram of a convolutional neural network used in the first embodiment.
Fig. 6 is a training process of the convolutional neural network in the first embodiment.
FIG. 7 is a diagram of a validation sample classification confusion matrix in accordance with the first embodiment.
Fig. 8 is a schematic diagram of a long-short-time memory network used in the second embodiment.
Fig. 9 is a diagram of training process of long-short-time memory network in the second embodiment.
FIG. 10 is a diagram of a validation sample classification confusion matrix in embodiment two.
Detailed Description
The invention is further illustrated below with reference to specific examples.
Example 1
A method for constructing a classification model of the fault degree of a rail transit motor bearing based on deep learning, which is designed as shown in figure 1, comprises the following steps:
step1, presetting bearing fault types of a certain model and 8 types of bearing faults:
fault type one: inner race 1.1 mm deep (inner race failure degree 1 grade)
Fault type two: inner ring 1.5 mm deep (inner ring failure degree 2 level)
Fault type three: inner race 1.7 mm deep (inner race failure degree 3 level)
Fault type four: inner race 2.0mm deep (inner race failure degree 4 level)
Fault type five: roller 0.6 mm deep (roller failure degree 1 grade)
Fault type six: roller 1.2 mm deep (roller failure degree 2 level)
Fault type seven: roller 1.5 mm deep (roller failure degree 3 level)
Fault type eight: roller 1.7 mm deep (roller failure degree 4 grade)
Step 2, original signal acquisition and framing processing: the time domain waveforms for each type of fault type are shown in fig. 4:
a. The motor of a locomotive of a certain model is arranged on the joint debugging test bed and runs in a no-load mode; collecting continuous Z-axis direction acceleration signals of each fault type bearing at the motor driving end within 240 seconds at the motor rotating speed of 1000-4500 rpm to obtain 8 original signals, dividing each original signal into 2400 parts, adding fault type labels to each part, and enabling each original signal to form an original signal data set composed of 2400 time series data with fault labels; the signal sampling frequency of the single-axis acceleration sensor is 25.6 kHz, so that each time series data with fault labels comprises 2560 sampling points;
b. 2560 sampling points in each time series data with fault labels of each original signal data set are subjected to fast algorithm FFT of discrete Fourier transform to obtain FFT results of 2560 sampling points, so that a time domain value of each sampling point of each time series data with fault labels corresponds to a frequency domain value, each time domain value and the corresponding frequency domain value are combined, and each time series data with fault labels is changed from one dimension (1X 2560) of the time domain to two dimensions (2X 2560) of the time domain frequency domain; in order to realize fast algorithm FFT of discrete Fourier transform, the signal sampling frequency Fs of the single-axis acceleration sensor is more than 2 times of the maximum frequency of the signal, namely the maximum rotating speed of the motor;
c. Data are randomly distributed according to the training set, the verification set and the test set proportion of 8:1:1, and 15360 (8×2400×80%) data are obtained as training samples, 1920 (8×2400×10%) data are obtained as verification samples, and the rest 1920 (8×2400×10%) data are obtained as test samples.
Step 3, establishing a deep learning network:
Training a convolutional neural network CNN model, wherein the CNN mainly comprises 29 layers, and the structure is shown in figure 5; the training sample two-dimensional sequence data in the step 2 is up-scaled to three-dimensional image data, the three-dimensional image data is input into an image input layer, and training parameters of a specified network are as follows: the image input layer (imageInputLayer) has a feature parameter of [2×2560×1]; the convolution network layers use 6 convolution layers (convolutionLayer), wherein the convolution kernel sizes are [1×8], 6 ReLU activation layers (reluLayer) and 6 batch normalization layers (batchNormalizationLayer), 5 maximum pooling layers (maxPoolingLayer), the pooling kernel sizes are [1×2],1 average pooling layer (maxPoolingLayer), and the pooling kernel sizes are [1×32]; the discard layer (dropoutLayer) discards the probability 0.5; the full connection layer (fullyConnectedLayer), classification function (softmaxLayer) layer and classification (classificationLayer) layer define the output of the network as 8 fault types; the solver uses a "random gradient descent" (SGD). And (3) in the training process, the verification sample in the step (2) is used for verifying the recognition accuracy and generalization capability of the network, the accuracy reaches more than 99%, and the convolutional neural network CNN model at the moment is stored. As shown in FIG. 6, the accuracy of the verification sample is set as a network convergence condition, the loss functions of the training sample and the verification sample are always reduced during network training, the accuracy is always increased, and the training efficiency is higher.
Step 4, verifying a model classification result:
And (3) introducing the verification sample in the step (2) into the convolutional neural network in the step (3) for classification, and finally verifying 8 classification results, wherein the total accuracy of verification is more than 99.3%, as shown in fig. 7.
Example two
A method for constructing a rail transit motor bearing fault degree classification model based on deep learning comprises the following steps:
step1, presetting bearing fault types of a certain model and 8 types of bearing faults:
fault type one: the inner ring is 1.1 mm deep (the inner ring failure degree is 1 level),
Fault type two: the inner ring is 1.5 mm deep (the degree of the inner ring failure is 2 levels),
Fault type three: the inner ring is 1.7 mm deep (the degree of failure of the inner ring is 3 levels),
Fault type four: the inner ring is 2.0 mm deep (the degree of failure of the inner ring is 4 levels),
Fault type five: the rollers are 0.6 mm deep (roller failure level 1),
Fault type six: the roller 1.2 mm is deep (roller failure level 2),
Fault type seven: the roller is 1.5 mm deep (roller failure level 3),
Fault type eight: roller 1.7 mm deep (roller failure degree level 4);
step 2, original signal acquisition and framing processing:
a. The motor of a locomotive of a certain model is arranged on the joint debugging test bed and runs in a no-load mode; collecting continuous Z-axis direction acceleration signals of each fault type bearing at the motor driving end within 240 seconds at the motor rotating speed of 1000-4500 rpm to obtain 8 original signals, dividing each original signal into 2400 parts, adding fault type labels to each part, and enabling each original signal to form an original signal data set composed of 2400 time series data with fault labels; the signal sampling frequency of the single-axis acceleration sensor is 25.6 kHz, so that each time series data with fault labels comprises 2560 sampling points;
b. 2560 sampling points in each time series data with fault labels of each original signal data set are subjected to fast algorithm FFT of discrete Fourier transform to obtain FFT results of 2560 sampling points, so that a time domain value of each sampling point of each time series data with fault labels corresponds to a frequency domain value, each time domain value and the corresponding frequency domain value are combined, and each time series data with fault labels is changed from one dimension (1X 2560) of the time domain to two dimensions (2X 2560) of the time domain frequency domain; in order to realize fast algorithm FFT of discrete Fourier transform, the signal sampling frequency Fs of the single-axis acceleration sensor is more than 2 times of the maximum frequency of the signal, namely the maximum rotating speed of the motor;
c. Data are randomly distributed according to the training set, the verification set and the test set proportion of 8:1:1, and 15360 (8×2400×80%) data are obtained as training samples, 1920 (8×2400×10%) data are obtained as verification samples, and the rest 1920 (8×2400×10%) data are obtained as test samples.
Step 3, establishing a deep learning network:
Training a long-short-time memory network LSTM model, wherein the LSTM network mainly comprises 8 layers, and the structure is shown in figure 8; inputting the training sample two-dimensional sequence data in the step 2 into a sequence input layer, and designating training parameters of a network as follows: the characteristic parameter of the sequence input layer (sequenceInputLayer) is 2; the first long-short-term memory layer (LSTMLAYER) includes 10 hidden layer units; the second long-short-term memory layer (LSTMLAYER) includes 100 hidden layer units; the two discard layers (dropoutLayer) discard a probability of 0.2; the full connection layer (fullyConnectedLayer), classification function (softmaxLayer) layer and classification (classificationLayer) layer define the output of the network as 8 fault types; the solver uses "moment estimation solution" (adam). And (3) in the training process, the verification sample in the step (2) is used for verifying the recognition accuracy and generalization capability of the network, the accuracy reaches more than 99%, and the LSTM model of the network with the time length and short time memory is stored. The accuracy of the verification sample is set as the network convergence condition, as shown in fig. 9, the loss functions of the training sample and the verification sample are always reduced during the network training, the accuracy is always increased, and the training efficiency is higher.
Step 4, verifying a model classification result:
And (3) introducing the verification sample in the step (2) into the long-short-time memory network in the step (3) for classification, and finally verifying 8 classification results, wherein the total accuracy of verification is more than 98.3%, as shown in fig. 10.
The scope of the present invention is not limited to the above two embodiments, and various modifications and alterations of the present invention will become apparent to those skilled in the art, and any modifications, improvements and equivalents within the spirit and principle of the present invention are intended to be included in the scope of the present invention.
Claims (6)
1. A method for constructing a rail transit motor bearing fault degree classification model based on deep learning is characterized by comprising the following steps: the method comprises the following steps:
Step 1, presetting bearing fault types and quantity: the number of the types of the preset bearing faults and the severity is A;
step 2, original signal acquisition and framing processing:
a. The track traffic motor is arranged on the joint debugging test bed and runs in a no-load mode; collecting Z-axis direction acceleration signals of a bearing at a transmission end by using a single-axis acceleration sensor, respectively collecting continuous Z-axis direction acceleration signals within a T time range of 1000-4500 rpm under each bearing fault type to obtain A original signals, dividing each original signal into B parts, adding fault type labels to each part, and enabling each original signal to form an original signal data set composed of B time sequence data with fault labels; the signal sampling frequency of the single-axis acceleration sensor is Fs, so that each time series data with the fault tag comprises N sampling points, and N=T/B xFs;
b. N sampling points in each time series data with fault labels of each original signal data set are subjected to fast algorithm FFT of discrete Fourier transform to obtain FFT results of the N sampling points, so that a time domain value of each sampling point of each time series data with fault labels corresponds to a frequency domain value, each time domain value and the corresponding frequency domain value are combined, and each time series data with fault labels is changed from time domain one-dimensional 1 XN to time domain frequency domain two-dimensional 2 XN, and the original N-point sequences are sequentially decomposed into a series of short sequences; the symmetrical property and the periodic property of the exponential factor in the discrete Fourier transform calculation are fully utilized, and further, the discrete Fourier transform values corresponding to the short sequences are obtained and combined, so that the purposes of deleting repeated calculation, reducing multiplication operation and simplifying the structure are achieved; in order to realize fast algorithm FFT of discrete Fourier transform, the signal sampling frequency Fs of the single-axis acceleration sensor needs to be more than 2 times of the maximum frequency of the signal, namely the maximum rotating speed (rps) of the motor, so as to meet the shannon sampling theorem;
c. Randomly distributing data according to the proportion of the training set, the verification set and the test set in each original signal data set of 8:1:1, and finally obtaining AXnXB x 80% data as training samples, wherein AXnXB x 10% data are used as verification samples, and the rest AXnXB x 10% data are used as test samples;
Step 3, establishing a deep learning network:
Training a deep learning network model for processing classification tasks, wherein the model is a convolutional neural network CNN or a long and short memory network LSTM;
a. the processing procedure of the convolutional neural network CNN network comprises the following steps: changing the ascending dimension of the two-dimensional sequence data into three-dimensional image data to be input into an image input layer; then, extracting image features by using a convolution layer, wherein the convolution layer, a batch normalization layer, a ReLU activation layer and a pooling layer are used as a complete feature extraction process; when needed, the method can be used as a whole for expansion to form a plurality of convolution characteristic extraction processes; after the feature extraction is completed, a discarding layer is used for preventing the model from being over fitted; then using the fully connected layer for combining the features found before; the classification function layer Softmax performs weighting and classification calculation on the output fully-connected characteristics; the classification layer outputs the result of the network training as the fault type;
b. Long and short term memory network LSTM network processing procedure: inputting two-dimensional sequence data into a sequence input layer; the long-term dependency relationship between time steps of the sequence data is learned by the long-term and short-term memory layer; discarding the layer, and randomly neglecting hidden layer nodes with a certain probability; when needed, the long and short time memory layer and the discarding layer can be used as a whole to be expanded to form a multi-layer memory dependent process; randomly ignoring some hidden nodes when each network iterates to prevent the network from being over-fitted; the full connection layer is used for combining the features found before; the classification function layer Softmax performs weighting and classification calculation on the output fully-connected characteristics; the classification layer outputs the result of the network training as the fault type;
c. Inputting the data in the training sample in the step 2 into a CNN or LSTM network for training, and designating training parameters of the network; verifying the identification accuracy of the current training network by the data in the verification sample in the step 2, achieving the required accuracy, and storing the CNN or LSTM network model at the moment;
Step 4, verifying a model classification result:
The verification sample in the step 2 is led into the CNN model or the LSTM model formed in the step 3 for classification, so that a classification result of fault type and severity degree is obtained, and the overall accuracy of the classification result is obtained; if the total accuracy of the classification does not meet the user index requirement, returning to the step 3 to adjust the depth or the layer number of the CNN or LSTM network, and modifying the training parameters until the total accuracy of the classification meets the requirement.
2. The method for constructing the deep learning-based rail transit motor bearing fault degree classification model is characterized by comprising the following steps of: in the step 1, the type and the number of the preset bearing faults are 8, and the types and the number of the preset bearing faults are respectively as follows:
fault type one: the inner ring is 1.1 mm deep, and the failure degree of the inner ring is 1 level;
fault type two: the inner ring is 1.5 mm deep, and the fault degree of the inner ring is 2 levels;
fault type three: the inner ring is 1.7 mm deep, and the fault degree of the inner ring is 3 levels;
fault type four: the inner ring is 2.0mm degrees deep, and the fault degree of the inner ring is 4 levels;
fault type five: the roller is 0.6 mm degrees deep, and the failure degree of the roller is 1 grade;
fault type six: roller 1.2 mm is deep, the degree of failure of the roller is 2 grade;
Fault type seven: the roller is 1.5 mm deep, and the failure degree of the roller is 3 stages;
fault type eight: the roller is 1.7 mm deep, the degree of failure of the roller is 4 stages.
3. The method for constructing the deep learning-based rail transit motor bearing fault degree classification model is characterized by comprising the following steps of: in the step 2:
a. Collecting continuous Z-axis direction acceleration signals of each fault type bearing at the motor driving end within 240 seconds at the motor rotating speed of 1000-4500 rpm to obtain 8 original signals, dividing each original signal into 2400 parts, adding fault type labels to each part, and enabling each original signal to form an original signal data set composed of 2400 time series data with fault labels; the signal sampling frequency of the single-axis acceleration sensor is 25.6 kHz, so that each time series data with fault labels comprises 2560 sampling points;
b. 2560 sampling points in each time series data with fault labels of each original signal data set are subjected to fast algorithm FFT of discrete Fourier transform to obtain FFT results of 2560 sampling points, so that a time domain value of each sampling point of each time series data with fault labels corresponds to a frequency domain value, each time domain value and the corresponding frequency domain value are combined, and each time series data with fault labels is changed into 2X 2560 in two dimensions from one dimension of the time domain to two dimensions of the frequency domain; in order to realize fast algorithm FFT of discrete Fourier transform, the signal sampling frequency Fs of the single-axis acceleration sensor is more than 2 times of the maximum frequency of the signal, namely the maximum rotating speed of the motor;
c. Data are randomly distributed according to the training set, the verification set and the test set proportion of 8:1:1, and 15360=8×2400×80% of data are obtained as training samples, 1920=8×2400×10% of data are taken as verification samples, and the rest 1920=8×2400×10% of data are taken as test samples.
4. The method for constructing the deep learning-based rail transit motor bearing fault degree classification model is characterized by comprising the following steps of: in the step 3:
Training a convolutional neural network CNN model, wherein the CNN model consists of 29 layers, the training sample two-dimensional sequence data in the step 2 is up-scaled into three-dimensional image data, the three-dimensional image data is input into an image input layer, and training parameters of the network are specified as follows: the image input layer imageInputLayer has a feature parameter of [2×2560×1]; the convolutional network layers use 6 convolutional layers convolutionLayer, the convolutional kernel sizes are [1×8], 6 ReLU active layers reluLayer and 6 batch normalization layers batchNormalizationLayer,5 maximum pooling layers maxPoolingLayer, the pooling kernel sizes are [1×2],1 average pooling layer maxPoolingLayer, and the pooling kernel sizes are [1×32]; discard layer dropoutLayer discard probability 0.5; the full connection layer fullyConnectedLayer, classification function softmaxLayer, and classification classificationLayer layers define the output of the network as 8 fault types; the solver uses a "random gradient descent" SGD; in the training process, the verification sample in the step 2 is used for verifying the recognition accuracy and generalization capability of the network, the accuracy reaches more than 99%, and the convolutional neural network CNN model at the moment is stored; and setting the accuracy of the verification sample as a network convergence condition.
5. The method for constructing the deep learning-based rail transit motor bearing fault degree classification model is characterized by comprising the following steps of: in the step 3:
Training a long-short-term memory network LSTM model, wherein the LSTM network mainly comprises 8 layers, inputting the training sample two-dimensional sequence data in the step 2 into a sequence input layer, and designating the training parameters of the network as follows: the characteristic parameter of the sequence input layer sequenceInputLayer is 2; the first long and short term memory layer LSTMLAYER includes 10 hidden layer units; the second long and short term memory layer LSTMLAYER includes 100 hidden layer units; two discard layers dropoutLayer discard a probability of 0.2; the full connection layer fullyConnectedLayer, classification function softmaxLayer, and classification classificationLayer layers define the output of the network as 8 fault types; the solver solves for adam using "moment estimates; in the training process, the verification sample in the step 2 is used for verifying the recognition accuracy and generalization capability of the network, the accuracy reaches more than 99%, and the LSTM model of the network with the time length and short time memory is stored; and setting the accuracy of the verification sample as a network convergence condition.
6. The method for constructing the deep learning-based rail transit motor bearing fault degree classification model is characterized by comprising the following steps of: in the step 4: and (3) importing the verification sample in the step (2) into a convolutional neural network or a long-short-time memory network for classification, and finally verifying 8 classification results and verifying the overall accuracy.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111409465.9A CN114330096B (en) | 2021-11-25 | 2021-11-25 | Rail transit motor bearing fault degree classification model construction method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111409465.9A CN114330096B (en) | 2021-11-25 | 2021-11-25 | Rail transit motor bearing fault degree classification model construction method based on deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114330096A CN114330096A (en) | 2022-04-12 |
CN114330096B true CN114330096B (en) | 2024-04-19 |
Family
ID=81047517
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111409465.9A Active CN114330096B (en) | 2021-11-25 | 2021-11-25 | Rail transit motor bearing fault degree classification model construction method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114330096B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115115037B (en) * | 2022-08-30 | 2022-11-11 | 无锡沃尔德轴承有限公司 | Bearing fault probability analysis device |
CN115758258B (en) * | 2022-11-22 | 2023-12-12 | 西南交通大学 | Traction transformer fault diagnosis method based on vibration signals and domain knowledge |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111538947A (en) * | 2020-05-18 | 2020-08-14 | 中车永济电机有限公司 | Method for constructing wind power generator bearing fault classification model |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108388860B (en) * | 2018-02-12 | 2020-04-28 | 大连理工大学 | Aero-engine rolling bearing fault diagnosis method based on power entropy spectrum-random forest |
-
2021
- 2021-11-25 CN CN202111409465.9A patent/CN114330096B/en active Active
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111538947A (en) * | 2020-05-18 | 2020-08-14 | 中车永济电机有限公司 | Method for constructing wind power generator bearing fault classification model |
Non-Patent Citations (2)
Title |
---|
基于注意力机制BiLSTM的设备智能故障诊断方法;王太勇;王廷虎;王鹏;乔卉卉;徐明达;;天津大学学报(自然科学与工程技术版);20200427(06);全文 * |
基于深度学习的电机轴承故障诊断研究;许爱华;杜洋;袁涛;;组合机床与自动化加工技术;20200320(03);全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN114330096A (en) | 2022-04-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111709448B (en) | Mechanical fault diagnosis method based on migration relation network | |
Zou et al. | Bearing fault diagnosis method based on EEMD and LSTM | |
CN114330096B (en) | Rail transit motor bearing fault degree classification model construction method based on deep learning | |
CN107144430B (en) | A kind of Method for Bearing Fault Diagnosis based on incremental learning | |
CN113469060A (en) | Multi-sensor fusion convolution neural network aeroengine bearing fault diagnosis method | |
CN107817106A (en) | Fault Diagnosis of Roller Bearings based on Bayes's residual transform singular value decomposition Gaussian Mixture HMM framework | |
CN102252843B (en) | Assessment method for rolling bearing performance variation | |
Li et al. | Joint attention feature transfer network for gearbox fault diagnosis with imbalanced data | |
CN108709744B (en) | Motor bearing fault diagnosis method under variable load working condition | |
CN112541233A (en) | Rotary machine fault diagnosis method based on improved convolutional neural network | |
CN113339204B (en) | Wind driven generator fault identification method based on hybrid neural network | |
CN113139278A (en) | Deep learning-based bearing residual life prediction method and system | |
Cao et al. | Deterioration state diagnosis and wear evolution evaluation of planetary gearbox using vibration and wear debris analysis | |
CN107776606B (en) | Fault detection method for shaft temperature monitoring system | |
CN114757365A (en) | High-speed railway roadbed settlement prediction and early warning method based on deep learning | |
CN116358871A (en) | Rolling bearing weak signal composite fault diagnosis method based on graph rolling network | |
CN116956215A (en) | Fault diagnosis method and system for transmission system | |
CN114861349A (en) | Rolling bearing RUL prediction method based on model migration and wiener process | |
CN116625683A (en) | Wind turbine generator system bearing fault identification method, system and device and electronic equipment | |
Zhang et al. | Fault diagnosis with bidirectional guided convolutional neural networks under noisy labels | |
CN112729825A (en) | Method for constructing bearing fault diagnosis model based on convolution cyclic neural network | |
Li et al. | Transformer-based meta learning method for bearing fault identification under multiple small sample conditions | |
CN110160781B (en) | Test set reconstruction and prediction method for rotary machine fault classification | |
CN117093938A (en) | Fan bearing fault detection method and system based on deep learning | |
Niu et al. | A hybrid bearing prognostic method with fault diagnosis and model 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 |