CN114492533A - Construction method and application of variable working condition bearing fault diagnosis model - Google Patents
Construction method and application of variable working condition bearing fault diagnosis model Download PDFInfo
- Publication number
- CN114492533A CN114492533A CN202210113860.0A CN202210113860A CN114492533A CN 114492533 A CN114492533 A CN 114492533A CN 202210113860 A CN202210113860 A CN 202210113860A CN 114492533 A CN114492533 A CN 114492533A
- Authority
- CN
- China
- Prior art keywords
- working condition
- domain
- fault diagnosis
- label
- time sequence
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M13/00—Testing of machine parts
- G01M13/04—Bearings
- G01M13/045—Acoustic or vibration analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Signal Processing (AREA)
- Acoustics & Sound (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention discloses a construction method and application of a variable working condition bearing fault diagnosis model, wherein in the constructed model, a feature extraction module, a generation module, a domain classifier and a label classifier form a whole, additional training data can be effectively generated through the combination of the three components of the label classifier, the domain classifier and the generation module, under the condition of ensuring that the domain classifier can obtain the features with the distinguishing degree in the domain, the domain sensitivity of the domain classifier to different data is reduced through maximizing the domain distinguishing loss, so that the domain classifier is difficult to process and distinguish the data, and in turn, the feature extraction module extracts more domain-independent time sequence features, thereby realizing end-to-end combined training; after the model is continuously trained, the label classifier is refined as much as possible, and the domain classifier is generalized as much as possible, so that the model generalization for the working condition without the bearing is improved, and the fault diagnosis model can accurately judge the fault type under different working conditions.
Description
Technical Field
The invention belongs to the technical field of fault diagnosis, and particularly relates to a construction method and application of a variable working condition bearing fault diagnosis model.
Background
The bearing is an extremely important part in modern mechanical equipment and is widely applied to the fields of transportation, production, manufacturing and the like. When modern industrial technology is rapidly developed, various machine tools are developed towards precision, automation and large-scale, once various bearing parts playing an important role in the machine tools break down, the operation of the machine tools is affected, the defective rate of products is increased, the precision of the products is reduced, and the machine tools are damaged, shut down, troubleshooting, maintenance and maintenance delay time limit to cause huge economic loss and even can endanger the life safety of workers. Therefore, it is of great significance to explore and develop a method capable of quickly and accurately diagnosing bearing faults in various production scenes.
Bearing fault diagnosis technologies are mainly classified into two categories: traditional machine learning Methods (ML) and Deep Learning (DL) based methods. The former needs professionals in multiple fields to perform manual feature extraction, so that the cost is high and the efficiency is low; the latter can realize automatic feature extraction, directly takes the acquired signal as input, realizes end-to-end fault diagnosis, and has certain mobility while having higher accuracy. The current common fault diagnosis network model based on deep learning mainly comprises: convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs); although the above method provides some improvement in efficiency and accuracy compared to ML, some disadvantages still exist, such as that CNN requires more convolutional layers and requires a large amount of marked data; the RNN may have problems of gradient explosion and gradient disappearance; in addition, in the face of the complex situation of the actual production environment and the difference of the production environment, under the condition of variable working conditions of actual industrial application, the data distribution under different working conditions is inconsistent, so that the classification performance of the model can be greatly influenced. The existing method has low generalization on models with different data distributions, and is difficult to have higher accuracy under the condition of variable working conditions.
Disclosure of Invention
Aiming at the defects or improvement requirements of the prior art, the invention provides a construction method and application of a variable working condition bearing fault diagnosis model, which are used for solving the technical problem of low bearing fault diagnosis accuracy rate in a variable working condition scene in the prior art.
In order to achieve the above object, in a first aspect, the present invention provides a method for constructing a variable condition bearing fault diagnosis model, including the following steps:
s1, building a variable working condition bearing fault diagnosis model; the variable working condition bearing fault diagnosis model comprises the following steps: the system comprises a feature extraction module, a generation module, a domain classifier and a label classifier;
the characteristic extraction module is used for extracting time sequence characteristics of input data; the generation module is used for generating generation characteristics similar to the time sequence characteristics; the domain classifier is used for mapping the input time sequence characteristics or the generated characteristics into corresponding prediction working condition categories; the label classifier is used for mapping the input time sequence characteristics or the generated characteristics into corresponding predicted fault category labels;
s2, taking bearing vibration signal data with a real fault category label under the existing working condition as source domain data, and taking bearing vibration signal data without the real fault category label under the new working condition as target domain data to form a training sample set;
s3, inputting the training sample set into a variable working condition bearing fault diagnosis model, and performing first-stage training on the model by simultaneously minimizing characteristic difference loss and first label prediction loss; the characteristic difference loss is used for measuring the difference between the time sequence characteristic of each training sample in the training sample set and the corresponding generated characteristic; the first label prediction loss is used for measuring the difference between a prediction fault class label corresponding to the time sequence characteristic of each training sample in the training sample set and a corresponding real fault class label;
s4, repeating the step S3 to iterate until the iteration number reaches a preset iteration number m; m is more than or equal to 1;
s5, inputting the training sample set into a variable working condition bearing fault diagnosis model, and performing second-stage training on the model by simultaneously minimizing the characteristic difference loss, maximizing the domain discrimination loss and minimizing the second label prediction loss; the domain discrimination loss is used for measuring the sum of discrimination losses of a source domain working condition type, a source domain working condition type and a target domain working condition type, wherein the prediction working condition type corresponding to the time sequence characteristic and the generation characteristic of the source domain data is discriminated as the sum of the discrimination losses of the source domain working condition type and the target domain working condition type; and the second label prediction loss is used for measuring the time sequence characteristics of each training sample in the training sample set and generating the difference between the predicted fault class label corresponding to the characteristics and the corresponding real fault class label.
Further preferably, the training method of the first stage comprises the following steps:
s31, respectively extracting the time sequence characteristics of each training sample in the training sample set by adopting a characteristic extraction module, and outputting the time sequence characteristics to a generation module;
s32, generating generation characteristics corresponding to the time sequence characteristics of the training samples by adopting a generation module respectively;
s33, measuring the difference between the time sequence characteristics of each training sample and the corresponding generated characteristics by adopting a domain classifier to obtain characteristic difference loss;
s34, calculating the difference between the predicted fault class label corresponding to the time sequence feature of each training sample in the training sample set and the corresponding real fault class label by adopting a label classifier to obtain the predicted loss of the first label;
and S35, updating the parameters in the variable-operating-condition bearing fault diagnosis model by minimizing the sum of the characteristic difference loss and the first label prediction loss.
Further preferably, the second stage training method comprises the following steps:
s51, respectively extracting the time sequence characteristics of the source domain data and the target domain data by adopting a characteristic extraction module, and outputting the time sequence characteristics to a generation module;
s52, generating generation characteristics corresponding to the time sequence characteristics of the source domain data and the target domain data respectively by adopting a generation module;
s53, measuring the difference between the time sequence characteristics of the source domain data and the target domain data and the corresponding generated characteristics by adopting a domain classifier to obtain characteristic difference loss;
s54, measuring the time sequence characteristics of the source domain data and the prediction working condition categories corresponding to the generation characteristics by adopting a domain classifier, judging the time sequence characteristics of the source domain data and the prediction working condition categories corresponding to the generation characteristics as the sum of judgment losses of the source domain working condition categories and the target domain working condition categories corresponding to the time sequence characteristics and the generation characteristics of the target domain data to obtain domain judgment losses;
s55, measuring time sequence characteristics of the source domain data and the target domain data by adopting a label classifier, and generating a difference between a predicted fault class label corresponding to the characteristics and a corresponding real fault class label to obtain a second label prediction loss;
and S56, updating the parameters in the variable condition bearing fault diagnosis model by simultaneously minimizing the characteristic difference loss, maximizing the domain discrimination loss and minimizing the second label prediction loss.
Further preferably, the feature extraction module is a time convolution network.
Further preferably, the generating module is one generator or a plurality of cascaded generators.
Further preferably, the generator comprises: a plurality of cascaded generation units; the generation unit includes: a fully-connected layer, a batch normalization layer, and a tanhExp activation function layer of the cascade.
Further preferably, the preset number of iterations m is:
m＝(N/C)×(cmax/cmin)
wherein N is the total number of source domain data in the training sample set; c is the fault category number of the source domain data in the training data set; c. CmaxThe maximum value of the sample number of each fault category of the source domain data in the training data set is obtained; c. CminAs training dataAnd collecting the minimum value of the sample numbers of each fault category of the source domain data.
In a second aspect, the invention provides a method for diagnosing a fault of a variable working condition bearing, which comprises the following steps:
the bearing vibration signal data to be diagnosed is input into the variable working condition bearing fault diagnosis model constructed by the construction method of the variable working condition bearing fault diagnosis model provided by the first aspect of the invention, and the time sequence feature is extracted by a feature extraction module based on the variable working condition bearing fault diagnosis model and then input into a label classifier in the variable working condition bearing fault diagnosis model, so that the fault category label of the bearing vibration signal data to be diagnosed is obtained.
In a third aspect, the present invention provides a fault diagnosis system for a variable condition bearing, including: the storage stores a computer program, and the processor executes the computer program to execute the variable-condition bearing fault diagnosis method provided by the second aspect of the invention.
In a fourth aspect, the present invention provides a machine-readable storage medium storing machine-executable instructions which, when invoked and executed by a processor, cause the processor to carry out the method of constructing a variable regime bearing fault diagnosis model provided in the first aspect of the present invention and/or the method of diagnosing a variable regime bearing fault provided in the second aspect of the present invention.
Generally, by the above technical solution conceived by the present invention, the following beneficial effects can be obtained:
1. the invention provides a construction method of a variable working condition bearing fault diagnosis model, wherein in the constructed model, a feature extraction module, a generation module, a domain classifier and a label classifier form a whole, additional training data can be effectively generated through the combination of the three components of the label classifier, the domain classifier and the generation module, and under the condition of ensuring that the domain classifier can obtain the features with the distinguishing degree in the domain, the domain sensitivity of the domain classifier to different data is reduced through maximizing the domain distinguishing loss, so that the domain classifier is difficult to process and distinguish the data, and in turn, the feature extraction module extracts more domain-independent time sequence features, thereby realizing end-to-end combined training; after the model is continuously trained, the label classifier is refined as much as possible, and the domain classifier is generalized as much as possible, so that the model generalization of the working condition (namely, unknown domain) without the bearing is improved, and the fault type can be accurately judged by the fault diagnosis model under different working conditions.
2. In the construction method of the variable working condition bearing fault diagnosis model, the training of the whole model is divided into two stages, the first stage is mainly based on the data in the original training sample set to pre-train the model, so that a generation module in the model can preliminarily generate characteristic data according to the data distribution of the existing bearing characteristics; and in the second stage, the model is further trained based on the original training sample set and the newly generated training samples, and the domain sensitivity of the model to the data of the source domain and the target domain is reduced under the condition that the domain classifier can obtain the distinguishing features in the domain, so that the model generalization to the working condition (namely an unknown domain) without the bearing is improved. The bearing data in the invention belongs to high-frequency time sequence signals, because of the cyclic stability and the discreteness of the data, the quality of the generated bearing data of the generator is poor in m pre-iterations, if the generated data and the time sequence data are considered to be the same working condition, the domain discriminator is easy to be not converged, and through the training of two stages, the problems of better data generation effect and effective avoidance of the non-convergence of the domain discriminator are realized.
3. According to the construction method of the variable working condition bearing fault diagnosis model, provided by the invention, the periodicity of bearing fault impact components and the change of the fault impact frequency along with the change of the bearing working condition are considered, the TCN model is used as a feature extraction network, the bearing feature representation under the variable scale is extracted through the causal convolution of the cavity, and the time relevance is considered to be favorable for capturing the constant working condition feature, so that the variable working condition bearing fault diagnosis model supports variable length input, the model is favorable for learning the relevance representation among multi-scale signals, and the accuracy of cross-domain fault diagnosis is improved. In addition, in the adopted generation model, the use of the TanhExp activation function enables the model to have a stable gradient, and the cascaded generators also effectively solve the problem of poor generated data quality caused by less samples in the industrial problem.
4. Compared with the traditional bearing fault diagnosis method, the variable working condition bearing fault diagnosis method provided by the invention does not need manual feature extraction, and can automatically carry out bearing fault diagnosis without manual judgment only by acquiring corresponding data and applying the corresponding data to a model after training the variable working condition bearing fault diagnosis model.
5. The model adopted by the variable working condition bearing fault diagnosis method provided by the invention comprises the domain classifier, and the domain classifier can greatly improve the generalization of the model on an invisible domain by using a domain generalization method, so that the fault diagnosis can be accurately carried out under the variable working condition, and the method has stronger universality and practicability.
Drawings
FIG. 1 is a flowchart of a method for constructing a variable condition bearing fault diagnosis model according to embodiment 1 of the present invention;
FIG. 2 is a schematic structural diagram of a variable condition bearing fault diagnosis model provided in embodiment 1 of the present invention;
FIG. 3 is a schematic diagram of a hole causal convolution according to embodiment 1 of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and 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. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
Examples 1,
A method for constructing a variable working condition bearing fault diagnosis model comprises the steps of taking original time sequence data collected by a sensor as input, then using a feature extractor based on a time convolution network to carry out automatic feature extraction, then using a domain generalization method generated by a countermeasure object based on a depth domain to carry out data enhancement, continuously training to obtain an accurate label classifier, and further establishing the variable working condition bearing fault diagnosis model.
Specifically, as shown in fig. 1, the method comprises the following steps:
s1, building a variable working condition bearing fault diagnosis model;
specifically, as shown in fig. 2, the variable condition bearing fault diagnosis model includes: the system comprises a feature extraction module, a generation module, a domain classifier and a label classifier;
the characteristic extraction module is used for extracting time sequence characteristics of input data; the generation module is used for generating generation characteristics similar to the time sequence characteristics; the domain classifier is used for mapping the input time sequence characteristics or the generated characteristics into corresponding prediction working condition categories; the label classifier is used for mapping the input time sequence characteristics or the generated characteristics into corresponding predicted fault category labels;
the feature extraction module may be a Time Convolutional Network (TCN), RNN, bidirectional LSTM, or other model. In this embodiment, it is preferable to perform automatic feature extraction on the timing signals collected by the sensor by using a TCN, which uses a one-dimensional full convolution network and in which the length of each hidden layer is the same as that of the input layer, and uses zero padding to ensure that the subsequent layers have the same length. The TCN simultaneously uses the hole causal convolution to ensure that the output at the time t only comes from the convolution result of the earlier time elements at the time t and the previous layer, and obtains input from each step d away from the time t, so that the network traces back to (k-1) d time steps (wherein k is the size of an inner core and d is an expansion factor), and all the input in effective historical data is used in the output prediction sequence when the receptive field of each layer grows exponentially.
The architecture of the time convolutional network TCN will be described in several respects:
sequence modeling task: given an input sequence, the corresponding output can be predicted, i.e. find a function f (build a convolution network) that will make a vector X of 1 element upT+1Mapping to another vector Y of 1 element litresT+1I.e. by
Wherein, XT+1＝{x0，x1，...，xT-1，xT}，
It includes one constraint (causal constraint): when the predicted time T is less than or equal to T, only inputs from the same time point and earlier time points can be used, and inputs from time points later than T cannot be used. The goal of sequence modeling is to find a function f that minimizes the true tag y0，...，yTAnd predicted results
With the loss in between.
Thus, TCN should be based on two principles: the output produced is the same length as the input and satisfies causal constraints. To implement the former, the TCN uses a one-dimensional full-convolution network (FCN), and where the length of each hidden layer is the same as the length of the input layer, and zero padding is used to ensure that subsequent layers have the same length. To realize the latter, the TCN uses a hole Causal convolution (scaled cause convolution), as shown in fig. 3, firstly it ensures that the output at time t comes only from the convolution result of time t and the earlier time element of the previous layer, and obtains the input from each d step of distance t, and makes the network trace back to (k-1) d time steps, and the prediction sequence output while the receptive field of each layer grows exponentially uses all the inputs in the valid history data, so that the convolution network with less depth can obtain a very large valid history. The hole causal convolution operation f(s) for the finally defined sequence element s is:
in addition, it should be noted that the idea of residual concatenation expresses the output as a linear superposition of the input and a non-linear transformation of the input: that is, a non-linear variation function is used to describe the input and output of a network, the input is X, and the output is Fr(x)，FrOperations such as convolution, activation, etc. are generally included; when an input is forcibly added to the output of the function, G (x) can be used to describe the relation between the input and the output, and at the moment, G (x) can be clearly split into Fr(x) And linear superposition of X. The residual block can allow each layer to learn to modify the identity mapping, avoid gradient disappearance and guarantee long-term effective use history.
Based on this, each layer of the TCN contains multiple filters for feature extraction, and therefore, in the TCN, a generic residual module is used instead of a convolutional layer. The residual block in the TCN is composed of the following components in sequence: the hole causal convolution, weight normalization (modifying data size), ReLU activation function, and dropout operations prevent overfitting and do this twice. Considering that the input and output can have different widths in TCN, using an additional 1 × 1 convolution ensures the same length.
After TCN is used for feature extraction, a new Domain Generalization (DG) method is used so that the model is applied to a new target domain without adjustment, and the original training data of a source domain is increased by using the synthetic data of an unknown domain so as to reduce the difference between different domains, thereby greatly enhancing the universality of the diagnosis model.
The generation module, the domain classifier and the tag classifier in this embodiment are all differentiable deep neural networks. And training a label classifier and a domain classifier to respectively predict class labels (fault class labels) and domain labels (working condition class labels) of the input data. The learning training goal of the generation module is to transform the input data. The conversion generated by the generation module is designed to be the same as the disturbance of the input label, the disturbance is combined with the original input to generate new data, the loss of the label classifier is minimized, and the label classifier is refined and accurate as much as possible, so that the label classifier can identify the input data and classify the input data accurately. Under the condition that the domain classifier can acquire the features with distinguishing degrees in the domain, the domain sensitivity of the domain classifier to different data is reduced by maximizing the domain distinguishing loss (one of the losses of the domain classifier), so that the domain classifier is difficult to process and distinguish the data, after the model is continuously trained, the label classifier is refined as much as possible, and the domain classifier is generalized as much as possible, so that the model generalization to the working condition (namely, unknown domain) where the bearing is not seen is improved, and the fault diagnosis model can accurately judge the fault type under different working conditions. In this embodiment, the training generation module is used to generate feature data according to the data distribution of the existing bearing features, and the generated feature data and the original feature data have the same fault category label.
The domain classifier can perform multi-class classification on the source domain (multiple source domains can be accessed) to ensure that maximization of domain classification loss does not simply force the composite data to fall into another domain distribution. Assuming there are three source domains, maximization given a synthetic instance from the first domain
(zi denotes the operating condition class, i ═ 1,2,3), which is essentially minimizing z1, while giving equal gradients to maximize both z2 and z 3. Thus, neither gradient of z2 nor z3 is dominant.
The combination of the label classifier, the domain classifier and the generation module can effectively generate additional training data, in turn allows more field-independent label classifiers to be learned, and the end-to-end joint training is performed, so that the variable condition bearing fault diagnosis model provided by the invention directly works at a pixel level, and the interpretability of the model is greatly improved.
Specifically, the generating module may be one generator or a plurality of cascaded generators.
The generator comprises: a plurality of cascaded generation units; the generation unit includes: a fully-connected layer, a batch normalization layer, and a tanhExp activation function layer of the cascade. Multiple groups of data in one iteration can be normalized on the level of the batch through a batch normalization layer so as to prevent abnormal values obtained by sampling from influencing the stability of a generator; a smoother boundary can be provided through the TanhExp activation function, the calculation speed of the lightweight neural network can be improved through simplified operation, and the method is robust to noise.
S2, taking bearing vibration signal data with a real fault category label under the existing working condition as source domain data, and taking bearing vibration signal data without the real fault category label under the new working condition as target domain data to form a training sample set;
specifically, bearing vibration signal data under different existing working conditions are obtained through a sensor and serve as a source domain data set. And acquiring bearing vibration signal data under a new working condition through a sensor to serve as target domain data.
S3, inputting the training sample set into a variable working condition bearing fault diagnosis model, and performing first-stage training on the model by simultaneously minimizing characteristic difference loss and first label prediction loss; the characteristic difference loss is used for measuring the difference between the time sequence characteristic of each training sample in the training sample set and the corresponding generated characteristic; the first label prediction loss is used for measuring the difference between a prediction fault class label corresponding to the time sequence characteristic of each training sample in the training sample set and a corresponding real fault class label;
specifically, in the first stage, the tag classifier loss is a first tag prediction loss; the first tag predicted loss expression is as follows:
wherein N is the number of samples in the training sample set; y iso (i)A real fault category label of the ith sample in the training sample set;
the predicted fault class label of the ith sample in the training sample set is the output of the label classifier. D1(. is) yo (i)And with
The difference metric function of (2) may be a distance calculation function (e.g., euclidean distance), a cross entropy loss function, or other loss functions.
The loss of the domain classifier is a feature difference loss, and the expression of the feature difference loss is as follows:
wherein N is the number of samples in the training sample set; f. ofg (i)Generating characteristics for the ith sample in the training sample set, namely the output of the characteristic extraction module; f. ofo (i)The time sequence characteristic of the ith sample in the training sample set, namely the output of the generating module; df(. is) fg (i)And fo (i)The difference metric function of (2) may be a distance calculation function (e.g., euclidean distance), a cross entropy loss function, or other loss functions.
In summary, the loss function of the variable working condition bearing fault diagnosis model in the first stage is as follows: l is1＝α1loss1+(1-α1)lossfBy minimizing L1Namely, the training of the first stage is completed; wherein alpha is1The value in this example is 0.5 for the balance weight of the first stage.
S4, repeating the step S3 to iterate until the iteration number reaches a preset iteration number m; m is more than or equal to 1;
the domain classifier is trained in advance in step S3 so that feature data can be generated from the data distribution of the existing bearing features, and the generated feature data has the same class label as the original feature data.
Preferably, considering the difference in the number of samples between different fault classes, the number of iteration rounds is considered by averaging the number of samples in each class, and considering the maximum difference in the number of samples between different classes, so as to ensure that the model can sufficiently learn the features of the minimum class of samples. Specifically, the preset iteration number m is:
m＝(N/C)×(cmax/cmin)
wherein N is the total number of source domain data in the training sample set; c is the fault category number of the source domain data in the training data set; c. CmaxThe maximum value of the sample number of each fault category of the source domain data in the training data set is obtained; c. CminIs the minimum of the sample numbers of each fault category of the source domain data in the training dataset.
S5, inputting the training sample set into a variable working condition bearing fault diagnosis model, and performing second-stage training on the model by simultaneously minimizing the characteristic difference loss, maximizing the domain discrimination loss and minimizing the second label prediction loss; the domain discrimination loss is used for measuring the sum of discrimination losses of a source domain working condition type, a source domain working condition type and a target domain working condition type, wherein the prediction working condition type corresponding to the time sequence characteristic and the generation characteristic of the source domain data is discriminated as the sum of the discrimination losses of the source domain working condition type and the target domain working condition type; and the second label prediction loss is used for measuring the time sequence characteristics of each training sample in the training sample set and generating the difference between the predicted fault class label corresponding to the characteristics and the corresponding real fault class label.
Specifically, in the second stage, the generated feature data of the same label input by the label classifier and the domain classifier and the original bearing feature data are directly subjected to mixed training; and alternately inputting the source domain data and the target domain data in the training sample set into a working condition bearing fault diagnosis model for training.
The label classifier loss is a second label prediction loss; the expression for the second tag predicted loss is as follows:
wherein M is the sum of the number of samples in the training sample set and the number of newly generated samples; y is(i)Is composed of the feature samples corresponding to the original samples in the training sample set and the newly generated feature samplesTrue fault category labels of the ith feature sample in the sample set;
and (3) a predicted fault class label of the ith feature sample in the sample set formed by the feature sample corresponding to the original sample in the training sample set and the newly generated feature sample, namely the output of the label classifier. D2(. is) y(i)And
the difference metric function of (2) may be a distance calculation function (e.g., euclidean distance), a cross entropy loss function, or other loss functions.
The loss of the domain classifier is the characteristic difference loss and the domain discrimination loss; the expression for the domain discrimination loss is as follows:
wherein M is the sum of the number of samples in the training sample set and the number of newly generated samples; d(i)And
respectively training the domain type of the prediction working condition type label of the ith characteristic sample in the sample set formed by the characteristic sample corresponding to the original sample in the sample set and the newly generated characteristic sample, and the real domain type corresponding to the domain type label; dd(. is) d(i)And
the difference metric function of (2) may be a distance calculation function (e.g., euclidean distance), a cross entropy loss function, or other loss functions.
In summary, the loss function of the variable working condition bearing fault diagnosis model in the second stage is as follows: l is2＝α2loss2+(1-α2)(βlossf-(1-β)lossd) By minimizing L2I.e. completing the second stageTraining; wherein alpha is2Beta is the balance weight of the second stage; in this example, α2The value is 0.4 and the value of beta is 0.6.
The invention provides a variable working condition bearing fault diagnosis model based on a time convolution network. Firstly, a time sequence signal acquired by a sensor is used as input, after time sequence feature extraction is carried out through feature extraction module extraction, a antagonism thought is introduced, data enhancement is carried out based on a generation module, a domain classifier and a label classifier to reduce the difference between domains, and a model is decomposed into a part which is specific to the domain and general to the domain, so that the model for bearing fault diagnosis under variable working conditions is established, and accurate fault diagnosis under the variable working conditions is further realized.
Examples 2,
A fault diagnosis method for a variable-condition bearing comprises the following steps:
the bearing vibration signal data to be diagnosed is input into the variable working condition bearing fault diagnosis model constructed by the construction method of the variable working condition bearing fault diagnosis model provided by the embodiment 1 of the invention, and is input into a label classifier in the variable working condition bearing fault diagnosis model after time sequence feature extraction is carried out on the basis of a feature extraction module in the variable working condition bearing fault diagnosis model, so that a fault category label of the bearing vibration signal data to be diagnosed is obtained.
The method for diagnosing the fault of the variable working condition bearing in the embodiment comprises the following steps:
1) preparing a data set:
the method comprises the steps of obtaining original data under different working conditions through a vibration acceleration sensor at a bearing seat to be detected, and preparing a plurality of source domain data sets with different fault type labels. Combining the actual operating condition of a certain bearing, to different rotational speeds, radial load combination operating modes, obtaining multiplex condition data such as operating mode A, operating mode B, operating mode C, every operating mode has the trouble of multiclass, includes: normal, outer ring failure, inner ring failure, and rolling element failure. And constructing data labels corresponding to the category labels for the obtained data. Carrying out signal interception through a sliding time window, and segmenting the long-time sequence into equal-length time sequence segments as network input; in this example, the time window is 2048 samples long, and the truncated data is stored as ndarray.
Loading data of multiple working conditions and dividing a training set and a test set; and carrying out standardization processing on the divided data in the training set, and selecting partial working condition combinations as source domains, for example, data of the working condition A and the working condition B as the source domains, and partial working conditions as target domains, for example, data of the working condition C.
2) Training a variable working condition bearing fault diagnosis model: the method of the embodiment 1 of the invention is adopted to train the fault diagnosis model of the bearing under variable working conditions. Wherein the process of feature extraction is as follows:
and inputting the time domain vibration acceleration data in the training set into a 2-layer residual error module which has the same expansion factor and is connected with the residual error. It is easy to understand that, the expansion coefficient d of the residual block is 2, the kernel size k is 3, the input original signal passes through the output of two convolutional layers, and is added to the input of the residual block, the channel widths of the input and output of the internal blocks of the residual block except the beginning and the end are the same, and are num _ filters, and the input and output widths of the first convolutional layer of the first residual block and the second convolutional layer of the last residual block are adjusted by 1 × 1 convolution.
The whole TCN feature extractor comprises a cavity cause and effect convolution layer, a weight normalization layer, a ReLU activation function layer, a Dropout layer, a cavity cause and effect convolution layer, a weight normalization layer, a ReLU activation function layer and a Dropout layer which are sequentially connected, and original signals are sequentially subjected to feature extraction through the layers; wherein, the purpose of the weight normalization layer is to normalize the input of the hidden layer to prevent gradient explosion; the purpose of the Dropout layer is to introduce regularization to prevent overfitting. Feature extraction is achieved through the TCN feature extractor and serves as input of training of the cross-working-condition-domain generalization network.
The specific process is the same as embodiment 1, and is not described herein.
3) Fault diagnosis of variable working condition bearing: inputting bearing data of an unknown domain (working condition C) which needs fault diagnosis in a test set into a trained variable working condition bearing fault diagnosis model, extracting characteristics through a time convolution network TCN, and directly inputting the extracted characteristics into a label classifier to perform automatic bearing fault diagnosis, so that cross-working condition bearing fault diagnosis under variable working conditions and unknown working conditions is realized.
In summary, in the embodiment, the original time series data acquired by the sensor is used as input, then the feature extractor based on the time convolution network is used for automatic feature extraction, then the domain generalization method generated by the countermeasure object based on the depth domain is used for improving the model generalization of the cross-working-condition fault diagnosis, the original data and the generated data are trained to obtain the accurate label classifier, then the classification network for bearing fault diagnosis under the variable working condition is established, and finally the fault diagnosis under the variable working condition is realized. Compared with the traditional bearing fault diagnosis method, the method has the advantages that manual feature extraction is not needed, and automatic analysis and diagnosis can be realized only by collecting relevant data and inputting the data into a training network; compared with other deep learning methods, the method supports variable-length input and has more stable gradient; the used domain generalization method also enables the model to adapt to the variable working condition, and enhances the generalization of the model and the application value of the actual scene.
The related technical scheme is the same as embodiment 1, and is not described herein.
Examples 3,
A variable condition bearing fault diagnostic system comprising: the fault diagnosis method for the variable-condition bearing comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the fault diagnosis method for the variable-condition bearing provided by the embodiment 2 of the invention when executing the computer program.
The related technical scheme is the same as embodiment 2, and is not described herein.
Examples 4,
A machine-readable storage medium storing machine-executable instructions which, when invoked and executed by a processor, cause the processor to implement a method of constructing a variable regime bearing fault diagnosis model as provided in embodiment 1 of the invention and/or a method of diagnosing a variable regime bearing fault as provided in embodiment 2 of the invention.
The related technical scheme is the same as that of embodiment 1 and embodiment 2, and is not described herein.
It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.
Claims (10)
1. A construction method of a variable working condition bearing fault diagnosis model is characterized by comprising the following steps:
s1, building a variable working condition bearing fault diagnosis model; the variable working condition bearing fault diagnosis model comprises the following steps: the system comprises a feature extraction module, a generation module, a domain classifier and a label classifier;
the characteristic extraction module is used for extracting time sequence characteristics of input data; the generation module is used for generating generation characteristics similar to the time sequence characteristics; the domain classifier is used for mapping the input time sequence characteristics or the generated characteristics into corresponding prediction working condition categories; the label classifier is used for mapping the input time sequence characteristics or the generated characteristics into corresponding predicted fault category labels;
s2, taking bearing vibration signal data with a real fault category label under the existing working condition as source domain data, and taking bearing vibration signal data without the real fault category label under the new working condition as target domain data to form a training sample set;
s3, inputting the training sample set into the variable working condition bearing fault diagnosis model, and performing first-stage training on the model by simultaneously minimizing characteristic difference loss and first label prediction loss; the feature difference loss is used for measuring the difference between the time sequence features and the corresponding generated features of each training sample in the training sample set; the first label prediction loss is used for measuring the difference between a prediction fault class label corresponding to the time sequence feature of each training sample in the training sample set and a corresponding real fault class label;
s4, repeating the step S3 to iterate until the iteration number reaches a preset iteration number m; m is more than or equal to 1;
s5, inputting the training sample set into the variable working condition bearing fault diagnosis model, and performing second-stage training on the model by simultaneously minimizing the characteristic difference loss, maximizing the domain discrimination loss and minimizing the second label prediction loss; the domain discrimination loss is used for measuring the sum of discrimination losses of a source domain working condition type, a source domain working condition type and a target domain working condition type, wherein the source domain working condition type is a prediction working condition type corresponding to a time sequence characteristic and a generation characteristic of source domain data; and the second label prediction loss is used for measuring the time sequence characteristics of each training sample in the training sample set and the difference between the predicted fault class label corresponding to the generated characteristic and the corresponding real fault class label.
2. The method for constructing the variable condition bearing fault diagnosis model according to claim 1, wherein the training method in the first stage comprises the following steps:
s31, respectively extracting the time sequence characteristics of each training sample in the training sample set by adopting the characteristic extraction module, and outputting the time sequence characteristics to the generation module;
s32, generating generation characteristics corresponding to the time sequence characteristics of the training samples by adopting the generation module respectively;
s33, measuring the difference between the time sequence characteristics of each training sample and the corresponding generated characteristics by adopting the domain classifier to obtain characteristic difference loss;
s34, measuring the difference between a predicted fault category label corresponding to the time sequence characteristics of each training sample in the training sample set and a corresponding real fault category label by using the label classifier to obtain a first label prediction loss;
and S35, updating parameters in the variable-condition bearing fault diagnosis model by minimizing the sum of the characteristic difference loss and the first label prediction loss.
3. The method for constructing the variable condition bearing fault diagnosis model according to claim 1, wherein the training method in the second stage comprises the following steps:
s51, respectively extracting the time sequence characteristics of the source domain data and the target domain data by using the characteristic extraction module, and outputting the time sequence characteristics to the generation module;
s52, generating generation characteristics corresponding to the time sequence characteristics of the source domain data and the target domain data by adopting the generation module respectively;
s53, measuring the difference between the time sequence characteristics of the source domain data and the target domain data and the corresponding generated characteristics by adopting the domain classifier to obtain the characteristic difference loss;
s54, measuring the sum of discrimination losses of the time sequence characteristics of the source domain data and the predicted working condition types corresponding to the generated characteristics by adopting the domain classifier, wherein the discrimination losses are judged as the sum of the discrimination losses of the source domain working condition types and the discrimination losses of the time sequence characteristics of the target domain data and the predicted working condition types corresponding to the generated characteristics, and the domain discrimination losses are obtained;
s55, measuring time sequence characteristics of source domain data and target domain data by using the label classifier, generating the difference between a predicted fault class label corresponding to the characteristics and a corresponding real fault class label, and obtaining the predicted loss of the second label;
s56, updating parameters in the variable condition bearing fault diagnosis model by simultaneously minimizing the characteristic difference loss, maximizing the domain discrimination loss and minimizing the second label prediction loss.
4. The method for constructing the variable condition bearing fault diagnosis model according to any one of claims 1 to 3, wherein the feature extraction module is a time convolution network.
5. The method for constructing the variable condition bearing fault diagnosis model according to any one of claims 1 to 3, characterized in that the generating module is one generator or a plurality of cascaded generators.
6. The method for constructing the variable condition bearing fault diagnosis model according to claim 5, wherein the generator comprises: a plurality of cascaded generation units; the generation unit includes: a fully-connected layer, a batch normalization layer, and a tanhExp activation function layer of the cascade.
7. The method for constructing the variable condition bearing fault diagnosis model according to any one of claims 1 to 3, wherein the preset iteration number m is as follows:
m＝(N/C)×(cmax/cmin)
wherein N is the total number of the source domain data in the training sample set; c is the fault category number of the source domain data in the training data set; c. CmaxThe maximum value of the sample number of each fault category of the source domain data in the training data set is obtained; c. CminIs the minimum value of the sample number of each fault category of the source domain data in the training data set.
8. A fault diagnosis method for a variable working condition bearing is characterized by comprising the following steps: inputting bearing vibration signal data to be diagnosed into a variable working condition bearing fault diagnosis model established by the variable working condition bearing fault diagnosis model establishment method according to any one of claims 1 to 7, and inputting the extracted time sequence characteristics of a characteristic extraction module in the variable working condition bearing fault diagnosis model into a label classifier in the variable working condition bearing fault diagnosis model so as to obtain a fault category label of the bearing vibration signal data to be diagnosed.
9. A variable condition bearing fault diagnostic system, comprising: a memory storing a computer program and a processor executing the computer program to perform the variable condition bearing fault diagnosis method of claim 8.
10. A machine-readable storage medium storing machine-executable instructions which, when invoked and executed by a processor, cause the processor to carry out the method of constructing a variable regime bearing fault diagnosis model according to any one of claims 1 to 7 and/or the variable regime bearing fault diagnosis method of the invention according to claim 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210113860.0A CN114492533A (en) | 2022-01-30 | 2022-01-30 | Construction method and application of variable working condition bearing fault diagnosis model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210113860.0A CN114492533A (en) | 2022-01-30 | 2022-01-30 | Construction method and application of variable working condition bearing fault diagnosis model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114492533A true CN114492533A (en) | 2022-05-13 |
Family
ID=81478490
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210113860.0A Pending CN114492533A (en) | 2022-01-30 | 2022-01-30 | Construction method and application of variable working condition bearing fault diagnosis model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114492533A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116226676A (en) * | 2023-05-08 | 2023-06-06 | 中科航迈数控软件(深圳)有限公司 | Machine tool fault prediction model generation method suitable for extreme environment and related equipment |
| CN116910570A (en) * | 2023-09-13 | 2023-10-20 | 华能新能源股份有限公司山西分公司 | Wind turbine generator system fault monitoring and early warning method and system based on big data |
| CN116956048A (en) * | 2023-09-19 | 2023-10-27 | 北京航空航天大学 | Industrial equipment fault diagnosis method and device based on cross-domain generalized label |
| CN117197546A (en) * | 2023-08-24 | 2023-12-08 | 佛山科学技术学院 | Fault diagnosis method and computer equipment based on causal mechanism |
-
2022
- 2022-01-30 CN CN202210113860.0A patent/CN114492533A/en active Pending
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116226676A (en) * | 2023-05-08 | 2023-06-06 | 中科航迈数控软件(深圳)有限公司 | Machine tool fault prediction model generation method suitable for extreme environment and related equipment |
| CN116226676B (en) * | 2023-05-08 | 2023-07-21 | 中科航迈数控软件(深圳)有限公司 | Machine tool fault prediction model generation method suitable for extreme environment and related equipment |
| CN117197546A (en) * | 2023-08-24 | 2023-12-08 | 佛山科学技术学院 | Fault diagnosis method and computer equipment based on causal mechanism |
| CN116910570A (en) * | 2023-09-13 | 2023-10-20 | 华能新能源股份有限公司山西分公司 | Wind turbine generator system fault monitoring and early warning method and system based on big data |
| CN116910570B (en) * | 2023-09-13 | 2023-12-15 | 华能新能源股份有限公司山西分公司 | Wind turbine generator system fault monitoring and early warning method and system based on big data |
| CN116956048A (en) * | 2023-09-19 | 2023-10-27 | 北京航空航天大学 | Industrial equipment fault diagnosis method and device based on cross-domain generalized label |
| CN116956048B (en) * | 2023-09-19 | 2023-12-15 | 北京航空航天大学 | Industrial equipment fault diagnosis method and device based on cross-domain generalized label |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114492533A (en) | Construction method and application of variable working condition bearing fault diagnosis model | |
| Li et al. | Self-attention ConvLSTM and its application in RUL prediction of rolling bearings | |
| CN108875771B (en) | Fault classification model and method based on sparse Gaussian Bernoulli limited Boltzmann machine and recurrent neural network | |
| CN111709448A (en) | Mechanical fault diagnosis method based on migration relation network | |
| CN111914883B (en) | Spindle bearing state evaluation method and device based on deep fusion network | |
| CN111241744B (en) | Low-pressure casting machine time sequence data abnormity detection method based on bidirectional LSTM | |
| CN111680788A (en) | Equipment fault diagnosis method based on deep learning | |
| CN115563536A (en) | Rolling bearing fault diagnosis method based on subdomain self-adaptation | |
| CN116448419A (en) | Zero sample bearing fault diagnosis method based on depth model high-dimensional parameter multi-target efficient optimization | |
| CN114034486B (en) | Pump mechanical equipment bearing fault diagnosis method based on unsupervised transfer learning | |
| Zang et al. | Markov chain-based feature extraction for anomaly detection in time series and its industrial application | |
| CN111637045B (en) | Fault diagnosis method for air compressor of ocean platform | |
| CN115600150A (en) | Multi-mode gearbox fault diagnosis method based on deep migration learning | |
| CN112731890A (en) | Power plant equipment fault detection method and device | |
| CN113110398A (en) | Industrial process fault diagnosis method based on dynamic time consolidation and graph convolution network | |
| CN115510950A (en) | Aircraft telemetry data anomaly detection method and system based on time convolution network | |
| Li et al. | Rolling bearing fault diagnosis under variable working conditions based on joint distribution adaptation and SVM | |
| CN116467653A (en) | Loom abnormal data processing method based on probability distribution and XGBoost decision algorithm | |
| CN115438897A (en) | Industrial process product quality prediction method based on BLSTM neural network | |
| CN114169091A (en) | Method for establishing prediction model of residual life of engineering mechanical part and prediction method | |
| CN117592543A (en) | Aeroengine rolling bearing fault diagnosis method based on self-supervision learning | |
| CN117290771A (en) | Rotary machine fault diagnosis method for generating countermeasure network based on improved auxiliary classification | |
| Wei et al. | A novel data augmentation and composite multi-scale network for mechanical fault diagnosis | |
| CN117056226A (en) | Cross-project software defect number prediction method based on transfer learning | |
| CN114143210B (en) | Command control network key node identification method based on deep learning |
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 |