Background
At present: the rolling bearing is one of basic components of mechanical equipment, is an important component of a mechanical transmission system, is an indispensable ring in modern industrial equipment, and the running state and the health degree of the rolling bearing directly influence the normal running of the mechanical equipment and the safety production of enterprises. Failure of the rolling bearing can lead to breakdown of the entire transmission system, causing unpredictable economic losses and even personal injury. Therefore, the effective fault diagnosis method for the rolling bearing is researched, the maintenance plan is arranged and made, the sudden shutdown event is reduced, and the method has great significance for the safe production and cost saving of enterprises.
In the industrial big data era, the intelligent fault diagnosis method gradually reveals advantages in practical engineering application. Meanwhile, the vibration signal is the most common and effective monitoring signal in the rolling bearing monitoring, and the intelligent diagnosis method based on the vibration signal has the capability of identifying and judging the running state of the monitoring equipment and is gradually applied. Although the intelligent diagnosis method based on the vibration signal achieves some practical application results, the industrial application capability is limited due to the fact that the intelligent diagnosis method is highly dependent on a high-quality data set. The intelligent diagnosis model needs a high-quality training data set with balanced categories, however, in actual industrial production, the monitored normal data is far more than fault data, the category imbalance phenomenon of the data set is caused, the identification effect of the intelligent diagnosis model on small categories (such as fault categories) is poor, and the generalization capability of the intelligent diagnosis model is limited. Therefore, under the category unbalanced data set, how to construct an effective rolling bearing intelligent diagnosis model is a problem to be solved.
Through the above analysis, the problems and defects of the prior art are as follows:
(1) the existing intelligent diagnosis method based on vibration signals has larger dependence on high-quality data sets, and the industrial application capability of the intelligent diagnosis method is limited.
(2) The normal data monitored by the existing intelligent diagnosis model is far more than fault data, so that the category imbalance phenomenon of a data set is caused, the identification effect of the intelligent diagnosis model on small categories (such as fault categories) is poor, and the generalization capability of the intelligent diagnosis model is limited.
The difficulty in solving the above problems and defects is: the sample expansion and optimization of the unbalanced-class data set is an idea for solving the problems, however, the effectiveness of the method for expanding and optimizing the unbalanced-class data set is a difficulty, and the expanded and optimized data should have similar statistical distribution characteristics with the original data. Although the traditional up-sampling and down-sampling modes can achieve the effect of class balance on the surface, the training effect of the intelligent diagnosis model is limited due to redundancy or information attenuation caused by the traditional up-sampling and down-sampling modes; the data enhancement method based on the added noise cannot ensure the consistency of the extended data, and the effect of the model is limited.
The significance of solving the problems and the defects is as follows: the quality of the data set has important significance on the effect of the intelligent diagnosis model, the effective optimization and expansion of the class imbalance data set are realized, the class imbalance phenomenon of the original data set is improved, the dependency of the intelligent diagnosis model on the quality of the data set is weakened, the application scene of the intelligent diagnosis model can be expanded, and the industrial energization of the artificial intelligence method is effectively promoted.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides an intelligent fault diagnosis method, system, equipment and application of an unbalance-like rolling bearing.
The invention is realized in such a way, and the intelligent fault diagnosis method for the similar unbalanced rolling bearing comprises the following steps:
acquiring a vibration signal of a rolling bearing, carrying out Zscore standardization on an original signal, and carrying out signal segmentation by using a mobile time window; the Zscore standardization can limit the amplitude range of the data, and is beneficial to the training of a conditional generation countermeasure network.
Constructing a generator and a discriminator combined condition to generate a confrontation network CGAN;
generating an antagonistic network for the established conditions, optimizing network parameters in a cyclic antagonistic training mode until the training is finished, and directionally generating a sample of a category with less data volume by a generator to enhance the data of the training set so as to relieve the category imbalance phenomenon of the original data set; the cyclic confrontation training is beneficial to enabling the generator and the discriminator in the CGAN to form a game relation and facilitating the alternate training of the generator and the discriminator.
Constructing a fault diagnosis model based on a deep convolutional neural network; a fault diagnosis model based on a deep convolutional neural network is constructed, and the deep convolutional neural network has excellent characteristic extraction capability and can effectively extract deep characteristics of signals.
And training the established fault diagnosis model by using the training data set optimized by the CGAN, and after the training is finished, realizing intelligent diagnosis on the unknown label sample by using the fault diagnosis model. The established fault diagnosis model is trained on a data set subjected to sample expansion by the CGAN, the unbalance-like phenomenon of the original data set is relieved, and the fault diagnosis model can be effectively trained.
Further, the raw signal is subjected to Zscore normalization, and the calculation formula of Zscore is as follows:
wherein the content of the first and second substances,
is the mean of the raw data, σ is the standard deviation of the raw data, x
*Data normalized for Zscore.
Further, the generator inputs Gaussian white noise and a category label, the generator outputs a simulated mechanical vibration signal, the discriminator inputs a real mechanical vibration signal subjected to Zscore standardization and a simulated mechanical vibration signal generated by the generator, and the discriminator outputs a probability value that the current signal is the real mechanical vibration signal; the generator and the discriminator are both formed by 3 layers of convolutional neural networks, both use the tanh activation function, and the expression of the activation function is as follows:
where x is the input to the activation function, e is the natural logarithm base, and y is the output of the activation function.
Further, the cyclic confrontation training means that in a training period, after the parameters of the discriminator are updated for 5 times, the parameters of the generator are updated for 1 time, and the next training period is entered after one complete training is completed.
Further, the constructed fault diagnosis model is composed of three layers of convolutional neural networks, input samples are original vibration signals and small sample class signals generated by generator directional supplement, the output is rolling bearing running state class, a softmax function is used as network output, and a softmax calculation formula is as follows:
wherein, P (c)i) Softmax output normalized probability representing class i sample, e is the base of the natural logarithm, oiK represents the total number of classes of the dataset classification for the corresponding network output before the class isofmax layer.
Further, the generator-oriented compensation in the input generates data which is fault category data with a small number of samples in the original data set.
Further, the loss function of the fault diagnosis model is a cross entropy loss function, and the specific calculation formula is as follows:
where loss is the specific value of the cross entropy loss function, y is the tag information desired to be output,
is the actual output label information.
It is a further object of the invention to provide a computer device comprising a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to perform the steps of:
acquiring a vibration signal of a rolling bearing, carrying out Zscore standardization on an original signal, and carrying out signal segmentation by using a mobile time window;
constructing a generator and a discriminator combined condition to generate a confrontation network CGAN;
generating an antagonistic network for the established conditions, optimizing network parameters in a cyclic antagonistic training mode until the training is finished, and directionally generating a sample of a category with less data volume by a generator to enhance the data of the training set so as to relieve the category imbalance phenomenon of the original data set;
constructing a fault diagnosis model based on a deep convolutional neural network;
and training the established fault diagnosis model by using the training data set optimized by the CGAN, and after the training is finished, realizing intelligent diagnosis on the unknown label sample by using the fault diagnosis model.
Another object of the present invention is to provide an intelligent fault diagnosis system for an unbalanced-like rolling bearing, which implements the intelligent fault diagnosis method for an unbalanced-like rolling bearing, the intelligent fault diagnosis system comprising:
the specific category data generation module is used for generating a confrontation network CGAN by using conditions and generating specific category data in a directional mode so as to reduce the class unbalance phenomenon of the data set;
and the health state acquisition module is used for training based on the optimized data set by utilizing the deep convolutional neural network, acquiring fault characteristic information and identifying the health state of the rolling bearing.
Another object of the present invention is to provide an intelligent fault diagnosis method for mechanical equipment, which uses the intelligent fault diagnosis method for unbalanced-like rolling bearings.
By combining all the technical schemes, the invention has the advantages and positive effects that: the method uses a Conditional generation countermeasure Network (CGAN), directionally generates specific category data to reduce the class imbalance phenomenon of a data set, simultaneously trains based on the optimized data set by using a deep convolutional neural Network, acquires fault characteristic information and identifies the health state of a rolling bearing, solves the problem that an intelligent diagnosis model cannot be effectively trained under the class imbalance condition, improves the generalization capability of the intelligent diagnosis model, and provides a feasible scheme for realizing industrial energization of a deep learning technology in the production practice.
The invention uses the vibration signal of the rolling bearing, and has the characteristics of easy acquisition and simple and easy operation; the fault category data with few samples are directionally generated through the CGAN, an original category unbalance data set is optimized, the problem of category unbalance caused by few fault data and many normal data in an actual industrial scene is solved, the effectiveness of network training is guaranteed, and the robustness of a network is improved. The method can enhance the reliability of the state identification of the rolling bearing, has the characteristics of low cost, simplicity, practicability and the like, provides an effective and convenient solution for using an intelligent diagnosis method in an actual industrial scene, and has strong engineering practical value and application prospect. Fig. 5 is a comparison of the use effect, in which (a) in fig. 5 is the classification result of the convolutional neural network commonly used at present without the optimization processing of the present invention on the data set, and (b) in fig. 5 is the processing result by using the present invention, and it can be seen from the comparison of the confusion matrix that the confusion matrix achieves better results (the classification effect of each class is represented on the diagonal line) after the present invention is used.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Aiming at the problems in the prior art, the invention provides an intelligent fault diagnosis method, system, equipment and application of a similar unbalanced rolling bearing, and the invention is described in detail below with reference to the accompanying drawings.
As shown in fig. 1, the intelligent fault diagnosis method for the unbalance-like rolling bearing provided by the invention comprises the following steps:
s101: acquiring a vibration signal of a rolling bearing, carrying out Zscore standardization on an original signal, and carrying out signal segmentation by using a mobile time window;
s102: constructing a condition generation countermeasure network (CGAN) combining a generator and a discriminator;
s103: generating a confrontation network for the established conditions, optimizing network parameters in a cyclic confrontation training mode until the training is finished, and directionally generating a sample of a category with less data quantity by a generator to enhance the data of the training set so as to relieve the category imbalance phenomenon of the original data set;
s104: constructing a fault diagnosis model based on a deep convolutional neural network;
s105: and training the established fault diagnosis model by using the training data set optimized by the CGAN, and after the training is finished, realizing intelligent diagnosis on the unknown label sample by using the fault diagnosis model.
Persons skilled in the art of the intelligent fault diagnosis method for the unbalanced-like rolling bearing provided by the invention can also use other steps to implement, and the intelligent fault diagnosis method for the unbalanced-like rolling bearing provided by the invention in fig. 1 is only a specific embodiment.
As shown in fig. 2, the intelligent fault diagnosis system for the unbalance-like rolling bearing provided by the invention comprises:
a specific category data generating module 1, configured to generate a countermeasure Network (CGAN) using a condition, and directionally generate data of a specific category to reduce a class imbalance phenomenon of a data set;
and the health state acquisition module 2 is used for training based on the optimized data set by utilizing the deep convolutional neural network, acquiring fault characteristic information and identifying the health state of the rolling bearing.
The technical solution of the present invention is further described below with reference to the accompanying drawings.
As shown in fig. 3, the intelligent fault diagnosis method for the unbalance-like rolling bearing provided by the invention specifically comprises the following steps:
the method comprises the following steps: acquiring a vibration signal of a rolling bearing, carrying out Zscore standardization on an original signal, and carrying out signal segmentation by using a mobile time window;
the calculation formula of the Zscore labeling of the original signal is as follows:
wherein the content of the first and second substances,
is the mean of the raw data, σ is the standard deviation of the raw data, x
*Normalized output for Zscore.
Step two: constructing a condition generation countermeasure network (CGAN) combining a generator and a discriminator;
the generator inputs Gaussian white noise and a category label, the output is a simulated mechanical vibration signal, the input of the discriminator is the real mechanical vibration signal after the Zscore standardization processing and the simulated mechanical vibration signal generated by the generator, and the output of the discriminator is the probability value that the current signal is the real mechanical vibration signal. The generator and the discriminator are both formed by 3 layers of convolutional neural networks, both use the tanh activation function, and the expression of the activation function is as follows:
where x is the input to the activation function, e is the natural logarithm base, and y is the output of the activation function.
Step three: generating a countermeasure network for the conditions established in the step two, optimizing network parameters in a cyclic countermeasure training mode until the training is finished, and directionally generating a sample of a category with less data quantity by a generator to enhance the data of the training set so as to relieve the category imbalance phenomenon of the original data set;
the cyclic confrontation training in the third step means that in a training period, after the parameters of the discriminator are updated 5 times, the parameters of the generator are updated 1 time, the generator tries to generate a simulation signal similar to the real signal through continuous parameter updating, and the discriminator tries to distinguish the simulation signal generated by the generator from the original real signal through continuous parameter updating. The two form a game relation with each other, and the generation capability of the generator and the discrimination capability of the discriminator are improved through continuous and cyclic training.
Step four: constructing fault diagnosis model based on deep convolution neural network
The fault diagnosis model constructed in the step four is composed of three layers of convolutional neural networks, input samples of the fault diagnosis model are original vibration signals and small sample class signals generated by directional supplement of a generator, output is a rolling bearing running state class, a softmax function is used as network output, and a softmax calculation formula is as follows:
wherein, P (c)i) Softmax output normalized probability representing class i sample, e is the base of the natural logarithm, oiK represents the total number of classes of the dataset classification for the corresponding network output before the class isofmax layer. Meanwhile, cross entropy is used as a loss function, and a cross entropy loss function calculation formula is as follows:
where loss is the specific value of the cross entropy loss function, y is the tag information desired to be output,
is the actual output label information.
Step five: and training the fault diagnosis model established in the fourth step by using the training data set optimized by the CGAN, and after the training is finished, realizing intelligent diagnosis on the unknown label sample by using the fault diagnosis model.
The following sets of specific embodiments to further describe the technical solution of the present invention:
the data set used in the embodiment of the invention comprises four different bearing running states, wherein the four different bearing running states are totally five states of inner ring fault (slight), outer ring fault (slight), inner ring fault (severe), outer ring fault (severe) and normal, which are respectively represented by categories 1,2,3,4 and 5, and the data quantity proportion of each category of the training set is as follows: 1: 1: 1: 1: the test set data size was 400 (80 test specimens per class).
The method comprises the following steps: collecting vibration signals of a rolling bearing, carrying out Zscore standardization on the original signals, carrying out signal segmentation by utilizing a moving time window, selecting the length of the moving time window as 2000 data points, overlapping the moving window, and setting the window moving step length as 500 data points;
the calculation formula of the Zscore labeling of the original signal is as follows:
wherein the content of the first and second substances,
is the mean of the raw data, σ is the standard deviation of the raw data, x
*Normalized output for Zscore.
Step two: constructing a condition generation countermeasure network (CGAN) combining a generator and a discriminator;
the generator inputs Gaussian white noise and a category label, the output is a simulated mechanical vibration signal, the input of the discriminator is the real mechanical vibration signal after the Zscore standardization processing and the simulated mechanical vibration signal generated by the generator, and the output of the discriminator is the probability value that the current signal is the real mechanical vibration signal. The generator and the discriminator are both formed by 3 layers of convolutional neural networks, the structural diagrams of the constructed CGAN networks are shown in FIG. 4, both the structures use tanh activation functions, and the expression of the activation functions is as follows:
where x is the input to the activation function, e is the natural logarithm base, and y is the output of the activation function.
Step three: generating a countermeasure network for the conditions established in the step two, optimizing network parameters in a cyclic countermeasure training mode until the training is finished, and directionally generating a sample of a category with less data quantity by a generator to enhance the data of the training set so as to relieve the category imbalance phenomenon of the original data set;
the cyclic confrontation training in the third step means that in a training period, after the parameters of the discriminator are updated for 5 times, the parameters of the generator are updated for 1 time, the next round of training period is started after the parameters are updated, and the CGAN training period is set to 3000 times. The generator attempts to generate a simulated signal similar to the real signal by constantly updating the parameters, and the discriminator attempts to distinguish the simulated signal generated by the generator from the original real signal by constantly updating the parameters. The two form a game relation with each other, and the generation capability of the generator and the discrimination capability of the discriminator are improved through continuous and cyclic training.
Step four: constructing fault diagnosis model based on deep convolution neural network
The fault diagnosis model constructed in the step four is composed of a three-layer convolutional neural network, input samples of the fault diagnosis model are original vibration signals and small sample class signals (fault class signals of the small samples) generated by the generator in a directional supplement mode, output of the fault diagnosis model is a rolling bearing running state class, a softmax function is used as network output, the softmax output is set to be a 5-dimensional vector, and a softmax calculation formula is as follows:
wherein, P (c)i) Softmax output normalized probability representing class i sample, e is the base of the natural logarithm, oiFor the corresponding network output before the class isofmax layer, k represents the total number of classes for the dataset class, and in this embodiment k has a value of 5. Meanwhile, cross entropy is used as a loss function, and a cross entropy loss function calculation formula is as follows:
where loss is the specific value of the cross entropy loss function, y is the tag information desired to be output,
is the actual output label information.
Step five: and training the fault diagnosis model established in the fourth step by using the training data set optimized by the CGAN, and testing on the test set after the training is finished. The results of the model test are shown in fig. 5, in which (a) of fig. 5 shows the effect of direct classification without using the method of the present invention, and (b) of fig. 5 shows the effect of classification using the method of the present invention. As can be seen from fig. 5 (a), the imbalance of the data set causes that only data with a large amount of data is emphasized in the network training, so that the recognition capability of normal data (category 5) is strong, the recognition capability of fault data with less data is weak, and the recognition accuracy is low; as can be seen from (b) of fig. 5, after the method is used, the original class imbalance data set is optimally supplemented, and the recognition capability of the model for the fault classes with fewer samples is improved.
It should be noted that the embodiments of the present invention can be realized by hardware, software, or a combination of software and hardware. The hardware portion may be implemented using dedicated logic; the software portions may be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or specially designed hardware. Those skilled in the art will appreciate that the apparatus and methods described above may be implemented using computer executable instructions and/or embodied in processor control code, such code being provided on a carrier medium such as a disk, CD-or DVD-ROM, programmable memory such as read only memory (firmware), or a data carrier such as an optical or electronic signal carrier, for example. The apparatus and its modules of the present invention may be implemented by hardware circuits such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of hardware circuits and software, e.g., firmware.
The above description is only for the purpose of illustrating the present invention and the appended claims are not to be construed as limiting the scope of the invention, which is intended to cover all modifications, equivalents and improvements that are within the spirit and scope of the invention as defined by the appended claims.