CN115545070A - Intelligent diagnosis method for unbalance-like bearing based on comprehensive balance network - Google Patents
Intelligent diagnosis method for unbalance-like bearing based on comprehensive balance network Download PDFInfo
- Publication number
- CN115545070A CN115545070A CN202211113733.7A CN202211113733A CN115545070A CN 115545070 A CN115545070 A CN 115545070A CN 202211113733 A CN202211113733 A CN 202211113733A CN 115545070 A CN115545070 A CN 115545070A
- Authority
- CN
- China
- Prior art keywords
- bearing
- training
- module
- learning module
- classifier
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
Abstract
The invention discloses an intelligent diagnosis method for a quasi-unbalanced bearing based on an integrated balance network, which comprises the following steps of: step (1), data preprocessing: carrying out short-time Fourier transform on each vibration signal in the bearing unbalance data set to obtain a time-frequency spectrum of the signal; step (2), model building: combining the unbalanced distribution learning module, the balanced distribution learning module, the re-weighting module and the classifier re-balancing module to build a comprehensive balance network model; step (3), model training: training a comprehensive balance network model by using time-frequency spectrum data of the unbalanced data set according to a given training step, a loss function and an optimization algorithm; step (4), fault diagnosis: and inputting the time frequency spectrum of the vibration signal of the bearing to be detected into the trained comprehensive balance network model to obtain a fault diagnosis result. The bearing vibration signal feature extraction method is high in bearing vibration signal feature extraction capability and high in bearing fault diagnosis accuracy.
Description
Technical Field
The invention relates to the technical field of diagnosis of similar unbalanced bearings, in particular to an intelligent diagnosis method of a similar unbalanced bearing based on a comprehensive balance network.
Background
In the fields of modern industrial production, processing and manufacturing, transportation and the like, rotary machines are used as the most basic parts of most mechanical equipment and have key functions which cannot be ignored. Rolling bearings are the most widely used mechanical parts in rotating machinery, and the health status of the rolling bearings directly affects the stability and safety of the machine. As the mechanical structure is more and more precise and complex, the failure of the bearing easily causes the functional failure of a large machine, thereby causing great economic loss and even endangering personal safety. Therefore, the method has important significance for monitoring and diagnosing the bearing fault in time and guaranteeing the stable and safe operation of the machine.
In recent decades, experts and scholars have proposed a number of fault diagnosis methods for rolling bearings, which are mainly classified into two categories: the method comprises a feature analysis method based on signal processing and an intelligent diagnosis method based on machine learning. Deep learning is the mainstream method of the current intelligent diagnosis method, can automatically extract data characteristics, has stronger learning ability, can process data with larger quantity and higher dimensionality, and meets higher fault diagnosis requirements.
Despite the great success of deep learning in the field of fault diagnosis, many problems remain to be solved. Deep learning as a data-driven diagnostic method needs to rely on a large amount of data to train a model with good effect, and a bearing monitoring data set in engineering practice is often like unbalanced. The bearing works in a normal state in most of time, so the data volume in the normal state is large; the machine must be shut down as soon as possible to ensure safety when a fault occurs, so the fault status data volume is small. When a traditional deep learning model is trained by adopting a class unbalance data set, overfitting to a few classes of samples is easily caused, a few classes of fault bearings are identified as normal bearings, and an expected bearing fault diagnosis effect cannot be achieved. Therefore, the intelligent diagnosis problem under the class imbalance data has become a research hotspot in the field of mechanical fault diagnosis.
The main method for dealing with the class imbalance problem at present is a rebalance strategy which can be divided into two types of resampling and reweighing. The resampling strategy contains two types, namely oversampling and undersampling. Oversampling is to rebalance the data set by creating enough samples for the minority class, while undersampling is to discard some redundant samples of the majority class to rebalance the data set. The re-weighting strategy increases its classification accuracy by assigning more weight to a few classes in the loss function.
Existing rebalancing strategies may have some negative impact on data classification. In the over-sampling strategy, if the quality of the created new sample is too poor, accurate classification of a few class samples may be affected. In the undersampling strategy, discarding high-quality majority class samples can lose important information and affect the generalization capability of the classification model. In the re-weighting strategy, the difficulty of model optimization is increased when the training data is large in size and extremely unbalanced. Although the rebalancing strategy can improve the overall classification accuracy of the imbalance-like data set, it also weakens the representation capability of the model feature extractor to some extent, since the original distribution of the data is distorted in the rebalancing.
Disclosure of Invention
The invention aims to provide an intelligent diagnosis method for an unbalance-like bearing based on an integrated balance network, which aims to solve the problems in the prior art.
In order to achieve the purpose, the invention provides the following technical scheme: an intelligent diagnosis method for a quasi-unbalanced bearing based on an integrated balance network comprises the following steps:
step (1), data preprocessing: carrying out short-time Fourier transform on each vibration signal in the bearing unbalance data set to obtain a time-frequency spectrum of the signal;
step (2), model building: combining the unbalanced distribution learning module, the balanced distribution learning module, the re-weighting module and the classifier re-balancing module to build a comprehensive balance network model;
step (3), model training: training a comprehensive balance network model by using time-frequency spectrum data of the unbalanced data set according to a given training step, a loss function and an optimization algorithm;
step (4), fault diagnosis: and inputting the time frequency spectrum of the vibration signal of the bearing to be detected into the trained comprehensive balance network model to obtain a fault diagnosis result.
Preferably, in the step (1), the vibration data of various types of states of the bearing form a data set according to a certain unbalance proportion, and the proportion of the normal sample to each type of fault sample is R:1, R >.
Preferably, in the step (2), the unbalanced distribution learning module takes a time-frequency spectrum of the unbalanced data set as an input, the features are extracted by the feature extractor, and the features f are output through global average pooling c ；
In the step (2), the balance distribution learning module adopts an example-class sampler, uses an unbalanced data set in the early stage of training, and adopts a class balance sampler to sample time-frequency spectrum data to continue training after reaching the preset training iteration times; in the class balance sampler, let C denote the total number of classes in the training set, and the sampling probability of the sample in class j can be expressed as
Wherein j belongs to {1,2, …, C }; the collected sample is extracted by a feature extractor, and features f are output through global average pooling r (ii) a The feature extractors in the balanced distribution learning module and the unbalanced distribution learning module have the same structure and share network parameters.
Preferably, in the step (2), the output f of the learning module is distributed by the re-weighting module in an unbalanced manner c And the output f of the balanced distribution learning module r As input, a parameter generator is provided, and as the training process proceeds, a parameter α can be generated according to a certain curve, and the characteristic f is subjected to c And f r Weighting to obtain a weighted feature z:
wherein, W c And W r Respectively representing the sum features f in the re-weighting module c And f r A weight parameter of the corresponding classifier; inputting the weighted feature z into a softmax function to obtain a predicted label, and constructing a weighted cross entropy loss function of the predicted label and a real label:
L＝αE c +(1-α)E r
wherein E is c And E r Respectively represent and feature f c And f r Cross entropy loss of the corresponding classifier.
Preferably, in the step (2), the classifier rebalancing module uses the classifier weight parameter W c And W r As an input, a parameter tau is set, the decision boundary of the classifier is scaled, the classification precision is further improved, and the scaled decision boundary is
Can be expressed as:
wherein W is a classifier weight parameter W by two modules c And W r And longitudinally overlapping.
Preferably, in the step (2), the feature extractor includes, but is not limited to, being constructed by one of a full-connection network, a deep convolution network, a deep confidence network, and a deep residual error network;
in the step (2), the classifier is constructed by one of, but not limited to, a support vector machine, a k-nearest neighbor algorithm, a random forest, a fuzzy system and a deep neural network;
in the step (2), the variation curve of the parameter generator includes, but is not limited to, one of a parabolic increasing strategy, a parabolic attenuation strategy, a cosine increasing strategy, a cosine attenuation strategy, and the like.
Preferably, in the step (3), the given training step is: firstly, training an unbalanced distribution learning module, a balanced distribution learning module and a classifier rebalancing module to obtain an optimal model; after training is finished, scaling the weight parameters of the classifier by adopting tau-normalization, traversing the parameter tau in the range of 0,1, finding out the value of tau with the highest classification precision, and storing the scaled model parameters.
Preferably, in the step (3), the given loss function is a weighted cross entropy loss function of the real label and the predicted label in the re-weighting module.
Preferably, in the step (3), the given optimization algorithm includes, but is not limited to, one of an adaptive moment estimation algorithm, a random gradient descent method, and a root mean square transfer algorithm.
Preferably, in the step (4), the time frequency spectrum of the vibration signal of the bearing to be measured does not need to adopt a class balance sampler when being input to the balanced distribution learning module, and the outputs of the unbalanced distribution learning module and the balanced distribution learning module do not need to be weighted in the re-weighting module, so that the health state of the bearing can be judged according to the output of the re-weighting module.
A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the steps of any of the methods when executing the program.
A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of any of the methods.
A processor for running a program, wherein the program when running performs any of the methods.
Compared with the prior art, the invention has the beneficial effects that: not only data distribution characteristic learning of an original unbalanced data set is considered, but also data distribution characteristic learning of different classes is balanced through a comprehensive resampling and reweighing strategy, and decision boundaries of the classifier are rebalanced by utilizing tau-normalization; the method has the advantages that the data characteristic representation capability and the data classification capability are both considered, and the problem of low fault diagnosis accuracy caused by excessive attention to a few types of samples in the traditional rebalance method is solved; therefore, the bearing vibration signal has strong feature extraction capability and high bearing fault diagnosis accuracy, and can adapt to the condition of extreme unbalanced data.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:
fig. 1 is a flowchart of an implementation of an intelligent diagnosis method for an unbalance-like bearing based on an integrated balance network according to an embodiment of the present invention;
FIG. 2 is a diagram illustrating an arrangement of unbalanced data sets in an embodiment of the present invention;
FIG. 3 is a graph illustrating the classification accuracy of data sets at different imbalance ratios according to an embodiment of the present invention.
FIG. 4 is a distribution diagram of test data after the model is trained using a data set with an imbalance ratio of 200, after T-SNE dimensionality reduction of features of a layer before model output;
FIG. 5 is a confusion matrix for testing the data classification results after training the model using the data set with an imbalance ratio of 200.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings of the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, are within the scope of the present invention. Thus, the following detailed description of the embodiments of the present invention, presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, are within the scope of the present invention.
Referring to fig. 1-3, in an embodiment of the present invention, an intelligent diagnosis method for a quasi-unbalanced bearing based on an integrated balance network includes the following steps:
step (1), data preprocessing: carrying out short-time Fourier transform on each vibration signal in the bearing unbalance data set to obtain a time-frequency spectrum of the signal;
step (2), model building: combining the unbalanced distribution learning module, the balanced distribution learning module, the re-weighting module and the classifier re-balancing module to build a comprehensive balance network model;
step (3), model training: training a comprehensive balance network model by using the time-frequency spectrum data of the unbalanced data set according to a given training step, a loss function and an optimization algorithm;
step (4), fault diagnosis: and inputting the time frequency spectrum of the vibration signal of the bearing to be detected into the trained comprehensive balance network model to obtain a fault diagnosis result.
Preferably, in the step (1), the vibration data of various types of states of the bearing form a data set according to a certain imbalance ratio, and the ratio of the normal sample to each type of fault sample is R:1, R >.
Preferably, in the step (2), the unbalanced distribution learning module takes a time-frequency spectrum of the unbalanced data set as an input, the feature extractor extracts features, and the features f are output through global average pooling c ；
In the step (2), the balance distribution learning module adopts an example-class sampler, uses an unbalanced data set in the early stage of training, and adopts a class balance sampler to sample time-frequency spectrum data to continue training after reaching the preset training iteration times; in the class balance sampler, let C denote the total number of classes in the training set, and the sampling probability of the sample in class j can be expressed as
Wherein j belongs to {1,2, …, C }; the collected sample is extracted by a feature extractor, and features f are output through global average pooling r (ii) a The feature extractors in the learning module with balanced distribution and the learning module with unbalanced distribution have the same structure, andsharing the network parameters.
Preferably, in the step (2), the output f of the learning module is distributed by the re-weighting module in an unbalanced manner c And the output f of the balanced distribution learning module r As input, a parameter generator is provided, and as the training process proceeds, a parameter α can be generated according to a certain curve, and the characteristic f is subjected to c And f r Weighting to obtain a weighted feature z:
wherein, W c And W r Respectively representing the sum features f in the re-weighting module c And f r A weight parameter of the corresponding classifier; inputting the weighted feature z into a softmax function to obtain a predicted label, and constructing a weighted cross entropy loss function of the predicted label and a real label:
L＝αE c +(1-α)E r
wherein E is c And E r Respectively represent and feature f c And f r Cross entropy loss of the corresponding classifier.
Preferably, in the step (2), the classifier rebalancing module uses the classifier weight parameter W c And W r As an input, a parameter tau is set, the decision boundary of the classifier is scaled, the classification precision is further improved, and the scaled decision boundary is
Can be expressed as:
wherein W is a classifier weight parameter W by two modules c And W r And longitudinally overlapping.
Preferably, in the step (2), the feature extractor includes, but is not limited to, being constructed by one of a full-connection network, a deep convolution network, a deep confidence network, and a deep residual error network;
in the step (2), the classifier is constructed by one of, but not limited to, a support vector machine, a k-nearest neighbor algorithm, a random forest, a fuzzy system and a deep neural network;
in the step (2), the variation curve of the parameter generator includes, but is not limited to, one of a parabolic increasing strategy, a parabolic attenuation strategy, a cosine increasing strategy, a cosine attenuation strategy, and the like.
Preferably, in the step (3), the given training step is: firstly, training an unbalanced distribution learning module, a balanced distribution learning module and a classifier rebalancing module to obtain an optimal model; after training is finished, scaling the weight parameters of the classifier by adopting tau-normalization, traversing the parameter tau in the range of 0,1, finding out the value of tau with the highest classification precision, and storing the scaled model parameters.
Preferably, in the step (3), the given loss function is a weighted cross entropy loss function of the real label and the predicted label in the re-weighting module.
Preferably, in the step (3), the given optimization algorithm includes, but is not limited to, one of an adaptive moment estimation algorithm, a random gradient descent method, and a root mean square transfer algorithm.
Preferably, in the step (4), the time frequency spectrum of the vibration signal of the bearing to be measured does not need to adopt a class balance sampler when being input to the balanced distribution learning module, and the outputs of the unbalanced distribution learning module and the balanced distribution learning module do not need to be weighted in the re-weighting module, so that the health state of the bearing can be judged according to the output of the re-weighting module.
According to the invention content and the attached figure 1, the unbalance-like bearing fault diagnosis method based on the comprehensive balance network specifically comprises the following steps:
step (1): and (4) preprocessing data. And carrying out short-time Fourier transform on each vibration signal in the bearing unbalance data set to obtain a time-frequency spectrum of the signal.
And (3) forming a data set by the vibration data of various types of states of the bearing according to a certain unbalance proportion, wherein the proportion of the normal sample to each type of fault sample is R:1, R >.
Step (2): combining the unbalanced distribution learning module, the balanced distribution learning module, the re-weighting module and the classifier re-balancing module to build a comprehensive balance network model;
the unbalanced distribution learning module takes the time frequency spectrum of the unbalanced data set as input, the features are extracted by the feature extractor, and the features f are output through global average pooling c 。
The balance distribution learning module adopts an example-class sampler, uses an unbalanced data set in the early stage of training, and adopts the class balance sampler to sample the time-frequency spectrum data to continue training after reaching the preset training iteration times. In the class balance sampler, let C denote the total number of classes in the training set, and the sampling probability of the sample in class j can be expressed as
Where j is ∈ {1,2, …, C }. The collected sample is extracted by a feature extractor, and features f are output through global average pooling r . The feature extractors in the balanced distribution learning module and the unbalanced distribution learning module have the same structure and share network parameters.
The reweighting module distributes the output f of the learning module with imbalance c And the output f of the balanced distribution learning module r As input, a parameter generator is set, and as the training process proceeds, the parameter α can be generated according to a certain curve, and the characteristic f is subjected to c And f r Weighting to obtain a weighted feature z:
wherein, W c And W r Respectively representing the sum features f in the re-weighting module c And f r The weight parameter of the corresponding classifier. Inputting the weighted feature z into a softmax function to obtain a predicted label, and constructing a weighted cross entropy loss function of the predicted label and a real label:
L＝αE c +(1-α)E r
wherein, E c And E r Respectively represent and feature f c And f r Cross entropy loss of the corresponding classifier.
Classifier rebalancing module with classifier weight parameter W c And W r Setting a parameter tau as input, scaling the decision boundary of the classifier to further improve the classification precision, and scaling
Can be expressed as:
wherein W is a classifier weight parameter W by two modules c And W r And longitudinally overlapping.
The feature extractor includes, but is not limited to, being constructed by one of a fully connected network, a deep convolutional network, a deep belief network, a deep residual network.
The classifier includes, but is not limited to, one of support vector machine, k-nearest neighbor algorithm, random forest, fuzzy system, and deep neural network.
The variation curve of the parameter generator includes, but is not limited to, one of a parabolic increasing strategy, a parabolic attenuating strategy, a cosine increasing strategy, a cosine attenuating strategy, and the like.
And (3): and (5) training a model. And training the comprehensive balance network model by utilizing the time-frequency spectrum data of the unbalanced data set according to the given training step, the loss function and the optimization algorithm.
The given training steps are: firstly, an unbalanced distribution learning module, a balanced distribution learning module and a classifier rebalancing module are trained to obtain an optimal model. After training is finished, scaling the weight parameters of the classifier by adopting tau-normalization, traversing the parameter tau in the range of 0,1, finding out the value of tau with the highest classification precision, and storing the scaled model parameters.
The given loss function is a weighted cross entropy loss function of the real label and the predicted label in the re-weighting module.
The given optimization algorithm includes, but is not limited to, one of an adaptive moment estimation algorithm (Adam), a stochastic gradient descent method (SGD), a root mean square transfer algorithm (RmsPorp).
And (4): and (5) fault diagnosis. And inputting the time frequency spectrum of the vibration signal of the bearing to be detected into the trained comprehensive balance network model to obtain a fault diagnosis result.
The time frequency spectrum of the vibration signal of the bearing to be detected does not need to adopt a class balance sampler when being input into the balanced distribution learning module, and the outputs of the unbalanced distribution learning module and the balanced distribution learning module do not need to be weighted in the heavy weighting module, so that the health state of the bearing can be judged according to the output of the heavy weighting module.
In order to more clearly understand the technical solution and the effects of the present invention, a detailed description is given below with reference to a specific embodiment.
Taking bearing fault diagnosis as an example, the bearing model is SKF 6205-2RS, a motor is adopted to drive the inner ring of the bearing to rotate, the rotating speed is set to 896rpm, an acceleration sensor is installed on a bearing seat to collect vibration signals of the bearing, the sampling frequency is 10kHz, eight health states are simulated in an experiment, and the method comprises the following steps: health (N), bearing inner ring fault (IR), bearing outer ring fault (OR), bearing rolling element fault (B), bearing inner ring rolling element composite fault (IB), bearing outer ring rolling element composite fault (OB), bearing inner ring outer ring composite fault (IO) and bearing inner ring outer ring rolling element composite fault (IOB). Each type of failure was a crack failure with a width of 0.2 mm. The data set was divided into a training set and a test set in the experiment, and the data set setup is shown in fig. 2. The health samples in the training set are in a plurality of types, including 1000 groups of data, and the rest fault data are sampled according to different unbalance proportions. The test set contains 150 signals per state, each signal having a length of 1024 data points. And training the comprehensive balance network by using a training set, and carrying out model effect test by using a test set.
The data set is then processed using the techniques disclosed herein, with the following specific details.
And (1) preprocessing data.
And carrying out short-time Fourier transform on each vibration signal in the unbalanced data set of the bearing to obtain a time-frequency spectrum of the signal. The number of signal data points is 1024, a Hamming window is selected, the window length is 128, the Hamming window moves 3 points to the right each time, and the time-frequency spectrum size is uniformly adjusted to 32 x 32 to be used as input data of the model. Taking the unbalanced ratio of 200.
And (2) building a model.
(1) The feature extractor adopts a lenet-5 convolutional neural network to design two convolutional layers and two pooling layers, the convolutional kernel of the convolutional layers is 5*5, and the convolutional kernel of the pooling layers is 2*2. And a batch of standardized layers are arranged behind each convolution layer, the characteristic value distribution is pulled back to the standard normal distribution, and ReLU activation function layers are connected among convolution layers.
(2) The classifier adopts a fully-connected network, a layer is designed, then a softmax activation function is connected, and finally an 8-dimensional vector is output by the model to represent the category of input data.
(3) Selecting a parabolic increasing strategy for the change curve of the parameter generator, wherein T represents the training iteration number, and T max Representing the maximum number of iterations, α can be expressed as:
and (3) training a model.
In the training process, the first 8 iterations of the balanced distribution learning module adopt the time-frequency spectrum data of the unbalanced data set for training, and then adopt the class balance sampler to sample the time-frequency spectrum data for continuous training. In the process of back propagation, a random gradient descent algorithm (SGD) is adopted to optimize the model, the learning rate is 0.001, the iteration times are 200, the loss tends to be balanced, and the model training is finished. After training is finished, scaling weight parameters of the classifier by adopting tau-normalization, uniformly taking 20 values in the range of [0,1] as tau, finding out the tau with the highest classification precision, and storing scaled model parameters.
And (4) diagnosing faults.
Fig. 3 shows that the classification accuracy of the data set is given under different imbalance proportions, and the accuracy of the method provided by the invention can be more than 98% under the imbalance proportion of 200. In the case of extreme imbalance of 500.
Taking the unbalance ratio of 200. It can be seen that the 8 types of health status data of the bearings are respectively aggregated into one cluster, and there are obvious gaps between the clusters, and only a few feature points are misclassified. And then inputting the mapping characteristics into a classifier to obtain a classification label, comparing the classification label with the real state label to obtain a fault diagnosis result of each data of the test set, and representing the fault diagnosis result by using a confusion matrix, as shown in fig. 5. It can be seen that the fault diagnosis accuracy of the method of the present invention is very high, and even under the unbalanced ratio of 200.
In conclusion, the comprehensive balance network provided by the invention not only considers the data distribution characteristic learning of the unbalanced data set, but also integrates various rebalance strategies, improves the feature extraction capability of the unbalanced-like bearing data set, relieves the problems caused by class imbalance, and realizes accurate fault diagnosis of the bearing under the class imbalance.
A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the steps of any of the methods when the program is executed.
A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of any of the methods.
A processor for running a program, wherein the program when running performs any of the methods.
Finally, it should be noted that: although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that changes may be made in the embodiments and/or equivalents thereof without departing from the spirit and scope of the invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (10)
1. An intelligent diagnosis method for a similar unbalance bearing based on a comprehensive balance network is characterized by comprising the following steps: the method comprises the following steps:
step (1), data preprocessing: carrying out short-time Fourier transform on each vibration signal in the bearing unbalance data set to obtain a time-frequency spectrum of the signal;
step (2), model building: combining the unbalanced distribution learning module, the balanced distribution learning module, the re-weighting module and the classifier re-balancing module to build a comprehensive balance network model;
step (3), model training: training a comprehensive balance network model by using time-frequency spectrum data of the unbalanced data set according to a given training step, a loss function and an optimization algorithm;
step (4), fault diagnosis: and inputting the time frequency spectrum of the vibration signal of the bearing to be detected into the trained comprehensive balance network model to obtain a fault diagnosis result.
2. The intelligent diagnosis method for the unbalance-like bearing based on the comprehensive balance network as claimed in claim 1, wherein: in the step (1), vibration data of various types of states of the bearing form a data set according to a certain unbalance proportion, and the proportion of the normal sample to each type of fault sample is R:1, R >.
3. The intelligent diagnosis method for the unbalance-like bearing based on the comprehensive balance network as claimed in claim 1, wherein: in the step (2), unbalance is madeThe distribution learning module takes the time-frequency spectrum of the unbalanced data set as input, the features are extracted by the feature extractor, and the features f are output through global average pooling c ；
In the step (2), the balance distribution learning module adopts an example-class sampler, uses an unbalanced data set in the early stage of training, and adopts a class balance sampler to sample time-frequency spectrum data to continue training after reaching a preset training iteration number; in the class balance sampler, let C denote the total number of classes in the training set, and the sampling probability of the sample in class j can be expressed as
Wherein j belongs to {1,2, …, C }; the collected sample is extracted by a feature extractor, and features f are output through global average pooling r (ii) a The feature extractors in the balanced distribution learning module and the unbalanced distribution learning module have the same structure and share network parameters.
4. The intelligent diagnosis method for the unbalance-like bearing based on the comprehensive balance network as claimed in claim 3, wherein: in the step (2), the reweighting module distributes the output f of the learning module in an unbalanced manner c And the output f of the balanced distribution learning module r As input, a parameter generator is set, and as the training process proceeds, the parameter α can be generated according to a certain curve, and the characteristic f is subjected to c And f r Weighting to obtain a weighted feature z:
wherein, W c And W r Respectively representing the sum features f in the re-weighting module c And f r A weight parameter of the corresponding classifier; inputting the weighted feature z into a softmax function to obtain a predicted label, and constructing a weighted cross entropy loss function of the predicted label and a real label:
L＝αE c +(1-α)E r
wherein E is c And E r Respectively represent and feature f c And f r Cross entropy loss of the corresponding classifier.
5. The intelligent diagnosis method for the unbalance-like bearing based on the comprehensive balance network as claimed in claim 1, wherein: in the step (2), the classifier rebalancing module uses the classifier weight parameter W c And W r As an input, a parameter tau is set, the decision boundary of the classifier is scaled, the classification precision is further improved, and the scaled decision boundary is
Can be expressed as:
wherein W is a classifier weight parameter W by two modules c And W r And longitudinally overlapping.
6. The intelligent diagnosis method for the unbalance-like bearing based on the comprehensive balance network as claimed in claim 1, wherein: in the step (2), the feature extractor includes, but is not limited to, being constructed by one of a full-connection network, a deep convolutional network, a deep confidence network, and a deep residual error network;
in the step (2), the classifier is constructed by one of, but not limited to, a support vector machine, a k-nearest neighbor algorithm, a random forest, a fuzzy system and a deep neural network;
in the step (2), the variation curve of the parameter generator includes, but is not limited to, one of a parabolic increasing strategy, a parabolic attenuation strategy, a cosine increasing strategy, a cosine attenuation strategy, and the like.
7. The intelligent diagnosis method for the unbalance-like bearing based on the comprehensive balance network as claimed in claim 1, wherein: in the step (3), the given training step is as follows: firstly, training an unbalanced distribution learning module, a balanced distribution learning module and a classifier rebalancing module to obtain an optimal model; after training is finished, scaling the weight parameters of the classifier by adopting tau-normalization, traversing the parameter tau in the range of [0,1], finding out the tau value with the highest classification precision, and storing the scaled model parameters;
in the step (3), the given loss function is a weighted cross entropy loss function of the real label and the predicted label in the re-weighting module;
in the step (3), the given optimization algorithm includes, but is not limited to, one of an adaptive moment estimation algorithm, a stochastic gradient descent method, and a root mean square transfer algorithm;
in the step (4), the time frequency spectrum of the vibration signal of the bearing to be detected does not need to adopt a class balance sampler when being input into the balanced distribution learning module, and the outputs of the unbalanced distribution learning module and the balanced distribution learning module do not need to be weighted in the re-weighting module, so that the health state of the bearing can be judged according to the output of the re-weighting module.
8. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the steps of the method of any of claims 1 to 7 are implemented when the program is executed by the processor.
9. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of the method according to any one of claims 1 to 7.
10. A processor, characterized in that the processor is configured to run a program, wherein the program when running performs the method of any of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211113733.7A CN115545070A (en) | 2022-09-14 | 2022-09-14 | Intelligent diagnosis method for unbalance-like bearing based on comprehensive balance network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211113733.7A CN115545070A (en) | 2022-09-14 | 2022-09-14 | Intelligent diagnosis method for unbalance-like bearing based on comprehensive balance network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115545070A true CN115545070A (en) | 2022-12-30 |
Family
ID=84727339
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211113733.7A Pending CN115545070A (en) | 2022-09-14 | 2022-09-14 | Intelligent diagnosis method for unbalance-like bearing based on comprehensive balance network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115545070A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116451124A (en) * | 2023-04-28 | 2023-07-18 | 哈尔滨工程大学 | Unbalanced radiation source signal identification method based on decoupling characterization learning |
| CN116499748A (en) * | 2023-06-27 | 2023-07-28 | 昆明理工大学 | Bearing fault diagnosis method and system based on improved SMOTE and classifier |
-
2022
- 2022-09-14 CN CN202211113733.7A patent/CN115545070A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116451124A (en) * | 2023-04-28 | 2023-07-18 | 哈尔滨工程大学 | Unbalanced radiation source signal identification method based on decoupling characterization learning |
| CN116451124B (en) * | 2023-04-28 | 2023-09-08 | 哈尔滨工程大学 | Unbalanced radiation source signal identification method based on decoupling characterization learning |
| CN116499748A (en) * | 2023-06-27 | 2023-07-28 | 昆明理工大学 | Bearing fault diagnosis method and system based on improved SMOTE and classifier |
| CN116499748B (en) * | 2023-06-27 | 2023-08-29 | 昆明理工大学 | Bearing fault diagnosis method and system based on improved SMOTE and classifier |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109918752B (en) | Mechanical fault diagnosis method, equipment and medium based on migration convolution neural network | |
| CN110162018B (en) | Incremental equipment fault diagnosis method based on knowledge distillation and hidden layer sharing | |
| CN110361176B (en) | Intelligent fault diagnosis method based on multitask feature sharing neural network | |
| CN110647923B (en) | Variable working condition mechanical fault intelligent diagnosis method based on self-learning under small sample | |
| CN111914883B (en) | Spindle bearing state evaluation method and device based on deep fusion network | |
| CN110502991B (en) | Internal combustion engine health monitoring method and system based on random convolutional neural network structure | |
| CN112577747A (en) | Rolling bearing fault diagnosis method based on space pooling network | |
| CN106980822B (en) | A kind of rotary machinery fault diagnosis method based on selective ensemble study | |
| CN115545070A (en) | Intelligent diagnosis method for unbalance-like bearing based on comprehensive balance network | |
| CN113505655B (en) | Intelligent bearing fault diagnosis method for digital twin system | |
| CN110334580A (en) | The equipment fault classification method of changeable weight combination based on integrated increment | |
| CN110082136B (en) | Rotary machine fault diagnosis method based on cloud genetic algorithm optimization support vector machine | |
| CN111351665B (en) | Rolling bearing fault diagnosis method based on EMD and residual error neural network | |
| CN106374465B (en) | Short-term wind power forecast method based on GSA-LSSVM model | |
| CN110333076B (en) | Bearing fault diagnosis method based on CNN-Stacking | |
| CN112729834B (en) | Bearing fault diagnosis method, device and system | |
| CN113567131A (en) | Bearing fault diagnosis method based on S transformation and miniature convolution neural network model | |
| CN111753891A (en) | Rolling bearing fault diagnosis method based on unsupervised feature learning | |
| CN112651426A (en) | Fault diagnosis method for rolling bearing of wind turbine generator | |
| CN113032917A (en) | Electromechanical bearing fault detection method based on generation countermeasure and convolution cyclic neural network and application system | |
| CN116956215A (en) | Fault diagnosis method and system for transmission system | |
| CN116561517A (en) | Unbalanced intelligent fault quantitative diagnosis method based on reward optimization deep reinforcement learning | |
| CN116010818A (en) | Gear box fault diagnosis method based on boundary enhancement prototype network | |
| CN117475191A (en) | Bearing fault diagnosis method for feature alignment domain antagonistic neural network | |
| CN113869451A (en) | Rolling bearing fault diagnosis method under variable working conditions based on improved JGSA algorithm |
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 |