CN114295377B - CNN-LSTM bearing fault diagnosis method based on genetic algorithm - Google Patents
CNN-LSTM bearing fault diagnosis method based on genetic algorithm Download PDFInfo
- Publication number
- CN114295377B CN114295377B CN202111518201.7A CN202111518201A CN114295377B CN 114295377 B CN114295377 B CN 114295377B CN 202111518201 A CN202111518201 A CN 202111518201A CN 114295377 B CN114295377 B CN 114295377B
- Authority
- CN
- China
- Prior art keywords
- fault
- layer
- data
- lstm
- training
- 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
- 238000003745 diagnosis Methods 0.000 title claims abstract description 73
- 238000000034 method Methods 0.000 title claims abstract description 67
- 238000004422 calculation algorithm Methods 0.000 title claims abstract description 28
- 230000002068 genetic effect Effects 0.000 title claims abstract description 27
- 238000012549 training Methods 0.000 claims abstract description 85
- 238000012360 testing method Methods 0.000 claims abstract description 36
- 238000013508 migration Methods 0.000 claims abstract description 8
- 230000005012 migration Effects 0.000 claims abstract description 8
- 238000002372 labelling Methods 0.000 claims abstract description 6
- 238000005070 sampling Methods 0.000 claims abstract description 6
- 238000004088 simulation Methods 0.000 claims abstract description 5
- 230000006870 function Effects 0.000 claims description 31
- 230000008569 process Effects 0.000 claims description 28
- 210000002569 neuron Anatomy 0.000 claims description 24
- 238000005096 rolling process Methods 0.000 claims description 19
- 238000011176 pooling Methods 0.000 claims description 14
- 238000000605 extraction Methods 0.000 claims description 12
- 230000004913 activation Effects 0.000 claims description 10
- 238000012545 processing Methods 0.000 claims description 8
- 238000004458 analytical method Methods 0.000 claims description 6
- 238000011156 evaluation Methods 0.000 claims description 4
- 230000001965 increasing effect Effects 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- 230000000007 visual effect Effects 0.000 claims description 4
- 230000008014 freezing Effects 0.000 claims description 3
- 238000007710 freezing Methods 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 230000009467 reduction Effects 0.000 claims description 3
- 238000013480 data collection Methods 0.000 claims 1
- 238000013527 convolutional neural network Methods 0.000 abstract description 7
- 238000012800 visualization Methods 0.000 description 9
- 238000010586 diagram Methods 0.000 description 7
- 239000011159 matrix material Substances 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 210000000349 chromosome Anatomy 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000008034 disappearance Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000002708 enhancing effect Effects 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 108090000623 proteins and genes Proteins 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000005284 excitation Effects 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 210000004205 output neuron Anatomy 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
Images
Landscapes
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention discloses a CNN-LSTM bearing fault diagnosis method based on a genetic algorithm, which comprises the following steps: firstly, collecting fault data by using a bearing fault simulation experiment platform, and expanding a sample by adopting a data enhancement method based on overlapping sampling; labeling the sample, dividing the expanded sample into a training set and a testing set, and finally standardizing. Selecting structural parameters of the CNN-LSTM fault diagnosis model by using a Genetic Algorithm (GA); and training a fault diagnosis model by using a training set, and performing online diagnosis on the bearing fault by using a test set. And finally, dynamically fine-tuning the fault diagnosis model structure by using a parameter migration method to realize the fault diagnosis of the bearing across working conditions. The method solves the problems that the identification precision of a fault bearing is low by manually extracting features, the accuracy is not high due to the fact that a fault diagnosis model structure is manually selected by experience, and time sequence data are difficult to capture by a single convolutional neural network.
Description
Technical Field
The invention relates to a fault diagnosis method for a rolling bearing, which solves the problems of low accuracy and difficulty in capturing time sequence data by a one-dimensional convolution neural network due to the fact that a fault diagnosis model structure is selected by means of manual experience.
Background
The rolling bearing is one of the most common core basic parts in mechanical equipment, has extremely wide application in various fields, and the health condition of the rolling bearing has great influence on the precision, stability, reliability, service life and the like of the mechanical equipment. When equipment fails, great economic loss can be caused, and even great accidents can be caused. It is difficult to identify faults as a difficulty in the field of fault diagnosis.
In recent years, under the rapid development of signal processing, data mining and artificial intelligence technologies, a fault diagnosis method based on data driving has been applied to the field of fault diagnosis of bearings, which has the advantage that fault diagnosis can be realized only by analyzing monitoring data during the operation of a system without knowing an accurate mathematical model of the system. In recent years, deep learning has been developed, which has a strong processing capability for a large amount of complex data and can sufficiently learn intrinsic characteristics of a failure, and thus has been applied to the field of failure diagnosis. In a plurality of deep learning methods, the convolutional neural network has the advantages of weight sharing, local sensing and the like, so that the number of parameters to be optimized can be reduced, the training speed is greatly improved, and the convolutional neural network has strong anti-noise capability and is widely applied to the field of fault diagnosis of bearings in recent years. But the single convolutional neural network model has the problems of gradient disappearance, difficulty in capturing time series data, too complex calculation and uncertainty in selecting the convolutional neural network structure by means of manual experience, which results in low accuracy.
Disclosure of Invention
The purpose of the invention is as follows: in order to solve the problems that a bearing fault diagnosis model is selected by means of manual experience, so that the accuracy is low and a traditional One-dimensional convolutional neural network is difficult to capture time sequence data, a CNN-LSTM bearing fault diagnosis method based on a genetic algorithm is provided. Adding an LSTM layer after the convolution layer and the pooling layer to extract the time characteristics of the bearing vibration signal, and selecting the structural parameters of the CNN-LSTM fault diagnosis model by using a Genetic Algorithm (GA); and training the fault diagnosis model by using a training set, carrying out online diagnosis on the bearing fault by using a test set, and observing the diagnosis process by using a visualization technology. And finally, dynamically fine-tuning the fault diagnosis model structure by using a parameter migration method, realizing the fault diagnosis of the bearing across working conditions, and solving the problem of insufficient data of the target domain working conditions with labels.
In order to solve the technical problems, the technical scheme adopted by the invention is as follows:
the invention relates to a CNN-LSTM bearing fault diagnosis method (GA-CNN-LSTM) based on a genetic algorithm, which comprises the following steps:
(1) Simulating six bearing operating states of a bearing normal state, an inner ring fault, an outer ring fault, a rolling body fault, a comprehensive fault and a retainer fault by using a PT500mini mechanical bearing gear fault simulation test bed and acquiring data of corresponding faults;
(2) Preprocessing the collected fault data of the bearing, firstly enhancing, realizing training data expansion by adopting a method of overlapping training sample segmentation, labeling the expanded data, and then standardizing to obtain a data set ofDividing 50% -80% of samples in the front of the data set X into a training set X af The remaining partitions into test sets x ae ;
Wherein X represents the bearing data set, n represents the total number of seismic data in the data set X, and X i Representing the ith vibration data in the data set X, and y representing a label and a fault type;
(3) The method comprises the following steps of performing self-adaptive selection on a CNN-LSTM structure by using a genetic algorithm, taking the parameter of each individual in a population as a CNN-LSTM structure, taking the accuracy of a model as a fitness function of the genetic algorithm, performing fitness evaluation and genetic operation, selecting the individual with the largest output function value as an optimal individual, and selecting the parameter in the optimal individual to construct a CNN-LSTM fault diagnosis model;
(4) Training the training set and the test set in the processed data set to use the CNN-LSTM fault diagnosis model selected by genetic algorithm, and utilizing the training set x af Performing off-line training of the model, and processing the test set x ae Inputting the information into a CNN-LSTM model stored in an offline training stage, carrying out online diagnosis on the bearing fault, and carrying out visual analysis on the diagnosis process;
in some specific embodiments: dividing the first 70% of the samples in the data set X into a training set X af The last 30% is divided into test set x ae ;
In some specific embodiments: the parameters of the data acquisition process in the step (1) are set as follows:
the failure of the inner ring is 0.2-0.6mm of inner ring cracks, the failure of the outer ring is 0.2-0.6mm of outer ring cracks, the failure of the rolling element is 2-5mm of peeling pits, the comprehensive failure is 0.2-0.6mm of cracks on the inner ring and the outer ring, and the failure of the retainer is the fracture of the retainer.
In some more preferred embodiments: the inner ring fault is 0.3mm of inner ring cracks, the outer ring fault is 0.3mm of outer ring cracks, the rolling element fault is 3mm of peeling pits, the comprehensive fault is that 0.3mm of cracks exist on the inner ring and the outer ring, and the retainer fault is that the retainer breaks. And collecting vibration signals of the same type of fault bearing at different rotating speeds and sampling at a sampling frequency of 4.8 KHZ.
In the technical scheme of the invention, the pretreatment process of the step (2) comprises the following steps: and performing data enhancement processing on sample data in the training process, expanding the training data by adopting a method of splitting overlapped training samples, and setting an overlap amount for the selected adjacent training samples. Therefore, the number of samples participating in model training can be increased, and the correlation between adjacent elements can be kept as much as possible, so that the model can learn useful features for classification as much as possible. And labeling the enhanced data with One-hot coding labels, then standardizing, dividing a test set and a training set, and finally standardizing.
The technical scheme of the invention is as follows: the CNN-LSTM structure selection process using the genetic algorithm in the step (3) comprises the following steps:
extracting characteristic information with difference in different types of training samples by increasing the number of neurons in the convolutional layer; the optional range of the number of the convolution layers is 1-5, and the size, the number and the step length of the convolution kernel and the number of LSTM neurons are optimized by using a genetic algorithm; real number coding is used for generating an initial population, one individual in the population represents a CNN-LSTM structure, then operations such as selection, intersection, variation and the like are carried out on the individuals in the population until the maximum iteration number is reached, and the optimal individual is output.
In some specific embodiments: the first layer of convolution layer adopts 96 convolution kernels with the size of 9 multiplied by 1 and the step length of 4 multiplied by 1; the second convolution layer adopts 25 convolution kernels with the size of 64 multiplied by 1 and the step length of 2 multiplied by 1; the pooling cores of the first layer and the second layer of the maximized pooling layer are both 2 multiplied by 1, and the step length is both 2; the LSTM layer neuron number is 24, the fully-connected layer neuron number is 56, adding a Dropout layer after the fully-connected layer prevents overfitting and adding a batch normalization layer (BN layer) after each convolutional layer allows the training data to remain normalized with the variance and mean changing iteratively.
The technical scheme of the invention is as follows: the CNN-LSTM fault diagnosis model training process in the step (4) is as follows:
the first step is the extraction of fault characteristics, the first part is the input of an original vibration signal of the bearing into a convolution layer, the activation function of the convolution layer adopts a ReLU function, and the dimensionality reduction is carried out through a pooling layer. Then, a second layer of convolution and a pooling layer are performed to generate a plurality of feature maps. And the second part is to segment the obtained feature map along the time axis and input the feature map to the LSTM layer. The activation function of the layer is a Tanh function, and the function used for the cycle time step is a sigmoid function.
And secondly, fault classification is carried out, after fault feature extraction is completed, a data flattening layer is used for flattening and inputting data into a full connection layer to extract comprehensive features, and Softmax is selected as an activation function of the full connection layer, so that classification of various rolling bearing faults can be realized. Using training set x af Training the model selected in the step (3), and processing the test set x ae Inputting the bearing fault into the model for online diagnosisPerforming T-SNE visualization analysis on the diagnosis process, and analyzing the performance of the constructed model;
the method for diagnosing the bearing cross-working-condition faults by using the bearing fault diagnosis method is characterized in that the structure of the obtained CNN-LSTM fault diagnosis model is dynamically fine-tuned by using a parameter migration method, so that the problem that the target domain working condition has label data shortage is solved.
The specific method for diagnosing the fault of the bearing across the working conditions comprises the following steps: the method comprises the steps of dynamically fine-tuning a fault diagnosis model structure by using a characteristic migration method to realize cross-working-condition fault diagnosis of a bearing, pre-training by using 100% of data on a source domain working condition, fine-tuning by using 10% of data on a target domain working condition, freezing a convolution layer and an LSTM layer when fine-tuning is performed, fixing parameters of the convolution layer and the LSTM layer, and replacing the last layer of full connection layer with the number of fault categories, so that the first layers are equivalent to a characteristic extractor, and only the last layer of full connection layer is trained.
Drawings
FIG. 1 is an overall flow diagram of the present invention;
FIG. 2 is a flow chart of the GA-CNN-LSTM fault diagnosis model;
FIG. 3 is a structural diagram of a GA-CNN-LSTM fault diagnosis model;
FIG. 4 is a graph of loss rate and accuracy rate change for the offline training process for data set A;
FIG. 5 is a graph of the change in accuracy for 20 training sessions for data set A;
FIG. 6 is a graph of a confusion matrix of online diagnostic results for data sets A, B, C;
FIG. 7 is a diagram of the T-SNE visualization process during data set A training;
FIG. 8 is a diagram of the T-SNE visualization process during training of data sets B, C;
FIG. 9 is a diagram of a fine-tuning based cross-condition diagnostic confusion matrix.
Detailed Description
The invention is further illustrated by the following examples, without limiting the scope of the invention:
in the embodiment, a PT500mini mechanical bearing gear fault simulation test bed is adopted to simulate bearing fault collection data, the sampling frequency of the data is selected to be 4.8KHz, and the used bearings are divided into a normal state, an inner ring fault, an outer ring fault, a rolling body fault, a comprehensive fault and a retainer fault. The failure of the inner ring is 0.3mm of inner ring cracks, the failure of the outer ring is 0.3mm of outer ring cracks, the failure of the rolling body is 3mm of peeling pits, the comprehensive failure is that 0.3mm of cracks exist on the inner ring and the outer ring, and the retainer fails and is broken.
The sample data has three different rotation speeds: 500r/min,1000r/min and 1500r/min, and is divided into three data sets A, B and C. 2048 data points of the vibration data are selected as one sample, namely a sliding window is 2048, the sliding step length is 28, and 1000 samples are acquired in each state. Randomness the sample set is divided into 7: and 3, dividing the training set and the test set. TABLE 1 bearing test sample compositions
TABLE 1 bearing test sample composition
The overall flow of the method is shown in fig. 1, and the specific implementation steps are as follows:
(1) The data enhancement process is as follows:
wherein,for the rounding-down operator, N is the length of the vibration signal, N is the sample length, η is the overlap ratio for sample expansion, and m is the maximum number of partitionable samples.
Labeling the expanded data to obtain a data set ofAnd dividing the data set X into training set and test set, selecting the first 70% of samples to be divided into training set X af The last 30% is divided into test set x ae Then, one-hot encoding tag, [ 1000 00 ] is performed]Denotes a normal bearing, [01 000]Indicating a rolling element failure bearing, [ 001 00]Indicating a failed bearing, [ 000 10 ]]Indicating faulty bearing of inner ring, [ 000 01 0]Indicating a faulty outer ring bearing, [ 000 001]Indicating a cage failed bearing. And finally, carrying out standardization.
(2) The normalization process is as follows:
wherein,for a standardized set of training data>For the standardized test data set, x af Represents a training data set, x ae Representing a test data set, σ f Is the standard deviation of the training data set, A is the number of samples in the training set, and->Is the mean of the training data set.
And 3, selecting parameters of the CNN-LSTM fault diagnosis model by using a genetic algorithm:
the invention adopts a mode of increasing the number of neurons in the convolutional layer to extract characteristic information with difference in training samples of different types, fixes the number of layers of the convolutional layer to be 2 layers, and uses a GA algorithm to perform optimization on the size, the number and the step length of the convolutional layer and the number of the neurons of the LSTM. The encoding method selects real number encoding, selects a roulette with a high fitness and a high probability of being selected as the selection operation of the GA, selects the most basic single point intersection in the intersection, and selects basic bit variation for mutating a certain gene randomly specified by the variation probability of the individual encoding as the variation operation of the GA. The flow chart is shown in FIG. 2
(1) Real number encoding was used to generate initial populations with population parameters as shown in table 2 below. One individual in the population represents one CNN-LSTM structure;
(2) The selectable range of the chromosome number is 10-30, the selectable range of the evolution generation number is 10-30, the selectable range of the cross rate is 0.6-0.95, and the selectable range of the variation rate is 0.001-0.1. The invention preferably sets the chromosome number as 20, the base factor as 8, the evolution generation number as 20, the cross rate as 0.9 and the variation rate as 0.01;
(3) Constructing a CNN-LSTM model, and taking genes on a chromosome as the CNN-LSTM model;
(4) Taking the accuracy of the model as a fitness function of the GA;
(5) Carrying out fitness evaluation and carrying out genetic operation;
(6) And judging whether the algorithm meets the maximum evolution algebra, and if so, ending the process. And if not, repeating the step 5 until the maximum evolution algebra is met.
(7) And (4) outputting the individual fitness of each generation by the program, and selecting the individual outputting the maximum function value as the optimal individual.
TABLE 2 population parameters
The optimal individuals selected were: the first layer of convolution layer adopts 96 convolution kernels with the size of 9 multiplied by 1 and the step length of 4 multiplied by 1; the second convolution layer adopts 25 convolution kernels with the size of 64 multiplied by 1 and the step length of 2 multiplied by 1; the pooling cores of the first layer and the second layer of the maximized pooling layer are both 2 multiplied by 1, and the step length is both 2; the number of neurons in the LSTM layer is 24, the number of neurons in the full connection layer is 56, a Dropout layer is added behind the full connection layer, a Dropout regularization technology is used for preventing overfitting, some neurons are randomly subtracted according to a certain proportion, and parameters of the neurons which are not taken away and the weight are updated. And a batch of normalization layers (BN layers) are added after each convolution layer, so that the training data can be kept in standardization under the condition that the variance and the mean value are continuously changed in an iterative manner, the training and convergence speed of the network is accelerated, the gradient explosion is controlled, and the gradient disappearance is prevented.
And 4, building a CNN-LSTM fault diagnosis model by using the optimal individuals selected in the step 3 and training the model, firstly extracting fault characteristics, inputting the original vibration signals of the bearing into a convolutional layer as the first part, adopting a ReLU function as the activation function of the convolutional layer, and reducing the dimension through a pooling layer. Then, a second layer of convolution and a pooling layer are carried out to generate a plurality of feature maps, and the process of convolution pooling feature extraction is as follows:
Y i out =f max (Y j in ,Y j+1 in )
wherein, X i out Is the output of the ith neuron, X i in Is the input of the ith neuron, f (-) is a nonlinear excitation function, ω ij For an input signal X i in Connection weight to jth neuron, b j Is the output bias. Y is j in Is the output value of the jth neuron of the input feature plane, f max (. Is the maximum value of a function, f ave (. As the mean value of a function, Y i out Is the output value of the ith neuron of the output feature plane.
And the second part is to segment the obtained feature map along the time axis and input the feature map to the LSTM layer. The activation function of the layer selects a Tanh function, and the function used for the cycle time step selects a sigmoid function. The LSTM layer is used for capturing long-term dependency relationship among time sequence data, firstly, the previous redundant sequence information is selectively abandoned through a forgetting gate, then new sequence information is selectively recorded by utilizing an input gate and an input node, and finally predicted sequence information is output by utilizing an output gate. The process of extracting features of the LSTM layer is as follows:
f t =σ(W f ·[h t-1 ,x t ]+b f )
i t =σ(W i ·[h t-1 ,x t ]+b i )
o t =σ(W o ·[h t-1 ,x t ]+b o )
h t =o t *tanh(C t )
wherein x is t For input at time t, f t To forget the output of the gate, i t Andas an output of the input gate, C t For output of status update, o t And h t To output the output of the gate, σ is a sign function, W f ,W i ,W C And W o Weight matrix, b f ,b i ,b C And b o Is a bias vector.
And finally, fault classification is carried out, after fault feature extraction is finished, a data flattening layer is used for flattening and inputting data to a full connection layer to extract comprehensive features, and Softmax is selected as an activation function of the full connection layer, so that the classification of various rolling bearing faults can be realized, and the process comprises the following steps:
wherein z is l+1 j Is the value of the jth output neuron at level l + 1, w ij l Is the weight between the ith neuron of the l layer and the jth neuron of the l +1 layer, b l j Bias of all neurons in layer l to the jth neuron in layer l +1, a l i For activation of the ith neuron at layer l, k represents the total number of classification tasks and q represents the probability distribution of the final output.
Using training set x af Training the model selected in the step (3), then storing model parameters with high structure accuracy so as to be used for an online diagnosis part, and processing the test set x ae Inputting the data into a model stored in an off-line training stage, carrying out on-line bearing fault diagnosis, and carrying out T-SNE visual analysis and analysis on the diagnosis processThe performance of the constructed model was analyzed.
The optimizer chooses Adam because its learning rate can be adaptively optimized. The method can continuously and iteratively update the weight of the neural network according to the training data, and can avoid the learning rate of each parameter from being dynamically adjusted by local optimization. The learning rate of the Adam algorithm is set to 0.01 and the dropout rate is set to 0.2. The iteration times are selected to be 15 times, a block diagram of a fault diagnosis model provided by the invention is shown in fig. 3, diagnosis is carried out on a data set A, fig. 4 is an offline training result of a training set in the data set A, the horizontal axis represents the iteration times, the vertical axis represents the accuracy and the loss rate respectively, the solid line represents the accuracy of the training set, the dotted line represents the loss rate of the training set, the accuracy gradually rises with the increase of the training iteration times, the accuracy finally reaches 99%, and the loss rate continuously falls and finally approaches to 0. In order to eliminate the influence of randomness, 20 times of training are performed by using the above model, and fig. 5 shows the results of 20 times of training. The accuracy of the obtained model test is stable between 98% and 100%, the highest accuracy is 100%, the average accuracy is 99.1%, and the result proves that the model structure fault diagnosis accuracy obtained by the invention is higher. The trained model is validated using a test set. Firstly, online diagnosis is performed by using the test set in the data set A, and fig. 6 (a) is a confusion matrix of online evaluation results, so that the accuracy is 99.5%, most samples are correctly classified, only 3 samples are wrongly classified, and the diagnosis precision is high. In order to verify the generalization of the test set, the accuracy of the test set on the remaining data sets B and C is shown in fig. 6 (B) and 6 (C), the diagnostic effect on the data set B reaches 99.83%, and the diagnostic effect on the data set C reaches 100%, which proves that the model of the invention has good generalization.
And the T-SNE is used for visualizing the characteristic extraction process of the online fault diagnosis, so that the relation between the neural network layers can be clearly represented, and the performance of the constructed model is analyzed. And selecting the diagnosis process of the data set A for visualization, outputting the characteristics of each layer by using a T-SNE algorithm, and performing dimension reduction visualization, wherein FIG. 7 is a T-SNE visualization process diagram of the data set A. It can be seen from fig. 7 (a) that the data just input into the model is difficult to distinguish between the categories due to the redundancy of the vibration signal itself. Clearly distributed clutter. After the first convolutional layer feature extraction, as shown in fig. 7 (b), a part of features are separated, the data clutter is obviously improved, but a lot of samples are not gathered and scattered between adjacent categories. Next, as shown in fig. 7 (c), the samples of the respective categories are more gathered, and a part of the features of the input data are effectively separated and gathered, but a part of the features are still not separated in the CNN output layer, and a small number of samples are wrongly separated, but the degree of clutter is greatly improved. Through the last LSTM layer, the time sequence characteristics of sample data are extracted, as can be seen from fig. 7 (d), the characteristics have obvious 10-cluster distribution, and the extraction and classification of ten fault characteristics of the bearing are very obvious, so that the model provided by the invention can effectively diagnose and classify the fault of the bearing.
In order to verify the generalization, T-SNE visualization is carried out on the data sets B and C, and FIGS. 8 (a) and 8 (B) show the T-SNE visualization process of the data sets B and C, and it can be seen that 10 clusters of distribution are already obviously formed through feature extraction of the last LSTM layer, and ten kinds of fault feature extraction and classification of the bearing are very obvious.
Selecting 1 normal data and 3 fault data as follows: normal data, 0.3mm inner ring fault data, 3mm rolling element fault data and 0.3mm inner and outer ring comprehensive fault data are used as source domain data sets. The target domain is 1 normal data and 5 fault data under different working conditions from the source domain, and the normal data and the fault data are respectively as follows: 0.3mm inner ring fault data, 3mm rolling element fault data, 0.3mm inner and outer ring comprehensive fault data, 0.3mm outer ring fault data and retainer fault. The number of fault species in the target domain is 2 more than that in the target domain. Freezing the convolution layer and the LSTM layer, fixing the parameters of the convolution layer and the LSTM layer, replacing the last full-link layer with the number of fault categories, and training only the last full-link layer.
(1) The data set A is selected as a source domain data set, 100% of data is pre-trained on a training set, the result is shown in the following figure 9 (a), the accuracy rate is 99.75%, 10% of data is finely adjusted on the training set on the data set B, the test is carried out on a test set, as shown in the following figure 9 (B), the overall accuracy rate is up to 99.83%, and only 1 group of rolling element faults are classified wrongly.
(2) The same fine-tuning was performed on the training set at 10% data on data set C and the testing was performed on the test set with an overall accuracy as high as 100% as shown in fig. 9 (C) below.
Therefore, the CNN-LSTM structure selected by the GA is used for carrying out cross-working condition migration to obtain higher accuracy, if a working condition is changed in actual engineering, the model needs to be trained again, a large amount of time is consumed, and in the actual engineering, the obtained data is small sample data and cannot support the model training to converge to the optimal state, so that a migration learning method is used, a large amount of time can be saved, training can be carried out on the model with trained parameters, and a large amount of data is not needed.
Claims (6)
1. A CNN-LSTM bearing fault diagnosis method based on genetic algorithm is characterized by comprising the following steps:
(1) Collecting six kinds of running state data sets of normal state of a bearing, inner ring fault, outer ring fault, rolling element fault, comprehensive fault and retainer fault by using a PT500mini mechanical bearing gear fault simulation test bed;
(2) Preprocessing collected bearing fault data, namely firstly performing enhancement processing, realizing training data expansion by adopting a method of splitting overlapped training samples, labeling the expanded data to obtain a data set, dividing 50-80% of samples in the front of the data set into a training set, and dividing the rest into a test set; then carrying out standardization;
wherein a bearing dataset is represented, a total number of vibration data in the dataset is represented, a first vibration data in the dataset is represented, a label and a fault category are represented;
(3) The method comprises the following steps of performing self-adaptive selection on a CNN-LSTM structure by using a genetic algorithm, taking the parameter of each individual in a population as a CNN-LSTM structure, taking the accuracy of a model as a fitness function of the genetic algorithm, performing fitness evaluation and genetic operation, selecting the individual with the largest output function value as an optimal individual, and selecting the parameter in the optimal individual to construct a CNN-LSTM fault diagnosis model;
(4) Training a training set and a test set in the processed data set to use a CNN-LSTM fault diagnosis model selected by a genetic algorithm, performing offline training of the model by using the training set, inputting the processed test set into the CNN-LSTM model stored in an offline training stage, performing online diagnosis on the bearing fault, and performing visual analysis on the diagnosis process;
the process of using genetic algorithm to carry out CNN-LSTM fault diagnosis model structure selection in the step (3) comprises the following steps:
extracting characteristic information with difference in training samples of different classes by adopting a mode of increasing the number of neurons in the convolutional layer; the optional range of the number of the convolution layers is 2, and the size, the number and the step length of the convolution kernel and the number of LSTM neurons are optimized by using a genetic algorithm; generating an initial population by using real number coding, wherein one individual in the population represents a CNN-LSTM structure, then performing operations such as selection, intersection, variation and the like on the individuals in the population until the maximum iteration number is reached, and outputting an optimal individual;
the number of the convolution layers is 2, and the first convolution layer adopts 96 convolution kernels with the size of 9 multiplied by 1 and the step length of 4 multiplied by 1; the second convolution layer adopts 25 convolution kernels with the size of 64 multiplied by 1 and the step length of 2 multiplied by 1; the pooling cores of the first layer and the second layer of the maximized pooling layer are both 2 multiplied by 1, and the step length is both 2; the number of neurons in the LSTM layer is 24, the number of neurons in the full connection layer is 56, a Dropout layer is added behind the full connection layer to prevent overfitting, and a batch normalization layer (BN layer) is added behind each convolution layer to keep the training data standardized under the condition that the variance and the mean value are changed iteratively;
the diagnosis is to dynamically fine-tune a fault diagnosis model structure by using a parameter migration method for the obtained CNN-LSTM fault diagnosis model, so that the problem of insufficient data with labels in the target domain working condition is solved;
the method comprises the steps of dynamically fine-tuning a fault diagnosis model structure by using a parameter migration method to realize cross-working-condition fault diagnosis of a bearing, pre-training by using 100% of data on a source domain working condition, fine-tuning by using 10% of data on a target domain working condition, freezing a convolution layer and an LSTM layer when fine-tuning is performed, fixing parameters of the convolution layer and the LSTM layer, replacing the last layer of full connection layer with the number of fault categories, and training only the last layer of full connection layer.
2. The genetic algorithm-based CNN-LSTM bearing fault diagnosis method of claim 1, wherein: in the step (2), the first 70% of samples of the data set are divided into training sets, and the last 30% of samples are divided into testing sets.
3. The genetic algorithm-based CNN-LSTM bearing failure diagnosis method according to claim 1, wherein the parameters of the data collection process in step (1) are set as:
the failure of the inner ring is 0.2-0.6mm of inner ring cracks, the failure of the outer ring is 0.2-0.6mm of outer ring cracks, the failure of the rolling element is 2-5mm of peeling pits, the comprehensive failure is 0.2-0.6mm of cracks on the inner ring and the outer ring, and the failure of the retainer is the fracture of the retainer.
4. The genetic algorithm-based CNN-LSTM bearing failure diagnosis method according to claim 3, wherein the parameters of the data acquisition process in step (1) are set as:
the failure of the inner ring is 0.3mm of inner ring cracks, the failure of the outer ring is 0.3mm of outer ring cracks, the failure of the rolling body is 3mm of peeling pits, the comprehensive failure is that 0.3mm of cracks exist on the inner ring and the outer ring, and the failure of the retainer is that the retainer breaks; and collecting vibration signals of the same type of fault bearings at different rotating speeds, and sampling at the sampling frequency of 4.8 KHZ.
5. The genetic algorithm-based CNN-LSTM bearing failure diagnosis method according to claim 1, wherein the preprocessing process in the step (2) is:
carrying out data enhancement processing on sample data in a training process, realizing training data capacity expansion by adopting a method of splitting overlapped training samples, and setting an overlap amount for selected adjacent training samples; and labeling the enhanced data with One-hot codes, then dividing a training set and a testing set, and finally carrying out standardization.
6. The genetic algorithm-based CNN-LSTM bearing fault diagnosis method according to claim 1, wherein the process of training the CNN-LSTM fault diagnosis model in step (4) is as follows:
the off-line training and the on-line diagnosis of the CNN-LSTM fault diagnosis model are divided into two parts;
the method comprises the steps that firstly, fault features are extracted, the first part of the extraction is that original vibration signals of a bearing are input into a convolution layer, the convolution layer activation function adopts a ReLU function, and dimensionality reduction is carried out through a pooling layer; then, a second layer of convolution and a pooling layer are carried out to generate a plurality of feature maps, the second part of extraction is to divide the obtained feature maps along a time axis and input the feature maps into an LSTM layer, a Tanh function is selected as an activation function of the layer, and a sigmoid function is selected as a function for a cycle time step;
secondly, fault classification is carried out, after fault feature extraction is finished, a data flattening layer is used for flattening and inputting data into a full connection layer to extract comprehensive features, and Softmax is selected as an activation function of the full connection layer, so that classification of various rolling bearing faults can be realized; and (4) performing off-line training on the model selected in the step (3) by using a training set, inputting the processed test set into the model, performing on-line bearing fault diagnosis, and performing T-SNE visual analysis on the diagnosis process.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111518201.7A CN114295377B (en) | 2021-12-13 | 2021-12-13 | CNN-LSTM bearing fault diagnosis method based on genetic algorithm |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111518201.7A CN114295377B (en) | 2021-12-13 | 2021-12-13 | CNN-LSTM bearing fault diagnosis method based on genetic algorithm |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114295377A CN114295377A (en) | 2022-04-08 |
CN114295377B true CN114295377B (en) | 2023-04-18 |
Family
ID=80967678
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111518201.7A Active CN114295377B (en) | 2021-12-13 | 2021-12-13 | CNN-LSTM bearing fault diagnosis method based on genetic algorithm |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114295377B (en) |
Families Citing this family (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114662680B (en) * | 2022-05-24 | 2022-08-30 | 杭州谨煜科技有限公司 | Bearing fault diagnosis method and system based on Gaussian Noise CNN model under variable working conditions |
CN115062867A (en) * | 2022-07-26 | 2022-09-16 | 中车长春轨道客车股份有限公司 | Bearing fault prediction method based on Teager energy spectrum and improved CNN model |
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 |
CN115964661B (en) * | 2023-01-04 | 2023-09-08 | 东北林业大学 | Rotary machine fault diagnosis method and system based on domain-by-domain network |
CN115876475B (en) * | 2023-02-06 | 2023-06-09 | 山东大学 | Fault diagnosis method, system, equipment and storage medium |
CN116933114B (en) * | 2023-06-12 | 2024-06-21 | 浙江大学 | CNN-LSTM-based direct-current micro-grid detection method and device |
CN117236381B (en) * | 2023-11-08 | 2024-02-02 | 智能制造龙城实验室 | Cutter wear monitoring method based on triplet length short-time memory neural network |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112067294A (en) * | 2019-09-20 | 2020-12-11 | 宫文峰 | Rolling bearing intelligent fault diagnosis method based on deep learning |
CN112577747A (en) * | 2020-12-07 | 2021-03-30 | 东南大学 | Rolling bearing fault diagnosis method based on space pooling network |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109447092B (en) * | 2018-10-25 | 2021-08-20 | 哈尔滨工程大学 | Method for extracting ice-to-ice passage based on sea ice scene classification |
CN109598341A (en) * | 2018-11-21 | 2019-04-09 | 济南浪潮高新科技投资发展有限公司 | A kind of detection of convolutional neural networks training result and method for visualizing based on genetic algorithm |
CN110398369A (en) * | 2019-08-15 | 2019-11-01 | 贵州大学 | A kind of Fault Diagnosis of Roller Bearings merged based on 1-DCNN and LSTM |
CN112746934B (en) * | 2020-12-31 | 2022-03-15 | 江苏国科智能电气有限公司 | Method for diagnosing fan fault through self-association neural network |
CN113177629A (en) * | 2021-03-25 | 2021-07-27 | 江苏大学 | Method and system for diagnosing blockage fault in operation of combined harvester |
-
2021
- 2021-12-13 CN CN202111518201.7A patent/CN114295377B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112067294A (en) * | 2019-09-20 | 2020-12-11 | 宫文峰 | Rolling bearing intelligent fault diagnosis method based on deep learning |
CN112577747A (en) * | 2020-12-07 | 2021-03-30 | 东南大学 | Rolling bearing fault diagnosis method based on space pooling network |
Also Published As
Publication number | Publication date |
---|---|
CN114295377A (en) | 2022-04-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN114295377B (en) | CNN-LSTM bearing fault diagnosis method based on genetic algorithm | |
CN110674752B (en) | Tool wear state identification and prediction method based on hidden Markov model | |
CN110609524B (en) | Industrial equipment residual life prediction model and construction method and application thereof | |
CN111145042A (en) | Power distribution network voltage abnormity diagnosis method adopting full-connection neural network | |
Sani et al. | Drop-out prediction in higher education among B40 students | |
CN112756759B (en) | Spot welding robot workstation fault judgment method | |
CN112765890B (en) | Dynamic domain adaptive network-based multi-working-condition rotating machine residual life prediction method | |
CN112308147A (en) | Rotating machinery fault diagnosis method based on integrated migration of multi-source domain anchor adapter | |
CN110455512B (en) | Rotary mechanical multi-integration fault diagnosis method based on depth self-encoder DAE | |
CN112763215B (en) | Multi-working-condition online fault diagnosis method based on modular federal deep learning | |
CN115859077A (en) | Multi-feature fusion motor small sample fault diagnosis method under variable working conditions | |
CN113887342A (en) | Equipment fault diagnosis method based on multi-source signals and deep learning | |
CN112541510A (en) | Intelligent fault diagnosis method based on multi-channel time series data | |
CN113989550A (en) | Electric vehicle charging pile operation state prediction method based on CNN and LSTM hybrid network | |
CN116384224A (en) | Aero-engine life prediction method based on conditional parameter dynamic convolutional neural network | |
CN115290326A (en) | Rolling bearing fault intelligent diagnosis method | |
CN115165366A (en) | Variable working condition fault diagnosis method and system for rotary machine | |
CN116108346A (en) | Bearing increment fault diagnosis life learning method based on generated feature replay | |
CN117034143A (en) | Distributed system fault diagnosis method and device based on machine learning | |
CN113505639B (en) | Rotary machine multi-parameter health state assessment method based on TPE-XGBoost | |
CN114997211A (en) | Cross-working-condition fault diagnosis method based on improved countermeasure network and attention mechanism | |
CN115876467A (en) | Pseudo label transfer type two-stage field self-adaptive rolling bearing fault diagnosis method | |
CN113551904A (en) | Gear box multi-type concurrent fault diagnosis method based on hierarchical machine learning | |
Singh et al. | Predicting the remaining useful life of ball bearing under dynamic loading using supervised learning | |
Li et al. | CapsNet-Enabled Multimodal Data Fusion: A Case Study in RUL Estimation of Turbofan Engine |
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 |