CN117516939A - Bearing cross-working condition fault detection method and system based on improved EfficientNetV2 - Google Patents
Bearing cross-working condition fault detection method and system based on improved EfficientNetV2 Download PDFInfo
- Publication number
- CN117516939A CN117516939A CN202311479935.8A CN202311479935A CN117516939A CN 117516939 A CN117516939 A CN 117516939A CN 202311479935 A CN202311479935 A CN 202311479935A CN 117516939 A CN117516939 A CN 117516939A
- Authority
- CN
- China
- Prior art keywords
- fault detection
- improved
- convolution
- dimensional
- module
- 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
- 238000001514 detection method Methods 0.000 title claims abstract description 102
- 238000000034 method Methods 0.000 claims abstract description 36
- 238000005065 mining Methods 0.000 claims abstract description 8
- 230000006870 function Effects 0.000 claims description 49
- 239000013598 vector Substances 0.000 claims description 27
- 238000012549 training Methods 0.000 claims description 23
- 230000008569 process Effects 0.000 claims description 15
- 230000004913 activation Effects 0.000 claims description 10
- 238000011176 pooling Methods 0.000 claims description 9
- 230000006835 compression Effects 0.000 claims description 8
- 238000007906 compression Methods 0.000 claims description 8
- 238000004364 calculation method Methods 0.000 claims description 7
- 230000005284 excitation Effects 0.000 claims description 5
- 239000011800 void material Substances 0.000 claims description 5
- 238000010606 normalization Methods 0.000 claims description 4
- 238000010257 thawing Methods 0.000 claims description 4
- 230000008014 freezing Effects 0.000 claims description 3
- 238000007710 freezing Methods 0.000 claims description 3
- 230000036541 health Effects 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 230000003044 adaptive effect Effects 0.000 claims description 2
- 238000013526 transfer learning Methods 0.000 claims description 2
- 238000000605 extraction Methods 0.000 abstract description 8
- 238000012544 monitoring process Methods 0.000 abstract description 6
- 238000012360 testing method Methods 0.000 description 7
- 230000007246 mechanism Effects 0.000 description 6
- 230000008901 benefit Effects 0.000 description 5
- 238000005070 sampling Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 230000004927 fusion Effects 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000009826 distribution Methods 0.000 description 3
- 230000002708 enhancing effect Effects 0.000 description 3
- 238000013178 mathematical model Methods 0.000 description 3
- 230000005012 migration Effects 0.000 description 3
- 238000013508 migration Methods 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 230000000750 progressive effect Effects 0.000 description 2
- 230000003068 static effect Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 238000009412 basement excavation Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008033 biological extinction Effects 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 230000010339 dilation Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 239000006185 dispersion Substances 0.000 description 1
- 238000010892 electric spark Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000003754 machining Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000009828 non-uniform distribution Methods 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 238000013138 pruning Methods 0.000 description 1
- 238000011897 real-time detection Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Classifications
-
- 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
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
- G06N3/047—Probabilistic or stochastic networks
-
- 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
- G06N3/048—Activation functions
-
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- 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)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Probability & Statistics with Applications (AREA)
- Acoustics & Sound (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention discloses a bearing cross-working condition fault detection method and system based on improved EfficientNetV2, and belongs to the field of bearing fault detection. The method comprises the steps of constructing an improved EfficientNetV2 fault detection model, extracting characteristics of vibration signals of a bearing to obtain signal characteristics, and detecting faults of the bearing according to the signal characteristics; the model comprises a one-dimensional convolution for capturing local features of the vibration signal; the Fused-MBConv1D module is used for carrying out rapid feature extraction according to the local features to obtain first local features; and the MBConv1D module is used for further mining the information in the first local feature to obtain a signal feature. The method can be better adapted to the one-dimensional form of the vibration monitoring signal, and has the capability of mining the local features and the global features of the signal, so that the real-time performance and the accuracy of the model feature extraction are improved.
Description
Technical Field
The invention relates to the technical field of bearing fault detection, in particular to a bearing cross-working condition fault detection method and system based on improved EfficientNetV 2.
Background
The bearing plays a role in reducing friction between the rotating element and the static element, once the bearing fails, the safety and stability operation of the whole mechanical equipment can be affected, and even huge economic loss and catastrophic casualties can be caused under serious conditions, so that the research of the bearing state monitoring and fault detection technology is very important.
The earliest depending experience of bearing fault detection is realized by observing, knocking, hearing, touching and other methods by experts; with the maturity of various technologies, the method based on the mathematical model deeply understands the physical characteristics and behaviors of the bearing, provides more reliable detection results, but is limited by the accuracy requirement of establishing the mathematical model;
the fault detection is widely applied because a complex mathematical model is not required to be established because of the analysis of the sensor sampling state monitoring signal based on a signal processing technology or a machine learning technology, but the existing detection model, such as EfficientNetV2, can be fit and generalization reduced when the problems of insufficient sample to be detected, non-uniform distribution with a model training sample and the like generated by dynamic change of the equipment operation working condition along with the field production demand are faced, and cannot be well adapted to one-dimensional vibration signals.
In this regard, the present invention is directed to improving EfficientNetV2 and introducing a migration learning strategy to overcome the above-described problems.
Disclosure of Invention
In view of the above, the invention provides a bearing cross-working condition fault detection method and system based on improved EfficientNetV 2. The method is suitable for one-dimensional signals, enhances the characteristic excavation capability and the cross-working condition generalization performance of the model, and improves the instantaneity and the accuracy of bearing fault detection.
In order to achieve the above purpose, the present invention adopts the following technical scheme:
on one hand, the invention discloses a bearing cross-working condition fault detection method based on improved EfficientNetV2, which comprises the following steps:
an improved EfficientNetV2 fault detection model is constructed and used for extracting characteristics of vibration signals of the bearing to obtain signal characteristics, and fault detection is carried out on the bearing according to the signal characteristics; wherein,
the improved EfficientNet V2 fault detection model sequentially comprises a one-dimensional convolution module, a Fused-MBConv1D module, an MBConv1D module and an output module;
the one-dimensional convolution is used for capturing local characteristics of the vibration signal;
the Fused-MBConv1D module sequentially comprises a first one-dimensional convolution, a first one-dimensional expansion convolution, a first one-dimensional SE attention module, a second one-dimensional convolution and a first Shortcut connection layer, and is used for rapidly extracting features according to the local features to obtain first local features;
the MBConv1D module sequentially comprises a one-dimensional depth convolution, a second one-dimensional expansion convolution, a second one-dimensional SE attention module, a one-dimensional point-by-point convolution and a second Shortnut connection layer, and is used for further mining information in the first local feature to obtain a signal feature.
Preferably, the one-dimensional convolution slidingly performs convolution operation on the vibration signal by defining the size and step size of a convolution kernel to capture local characteristics of the signal, and the calculation formula is as follows:
in the method, in the process of the invention,is the output feature of the j-th region of layer 1>Is the weight of the jth region of the ith convolution kernel of the ith layer, +.>Is the offset of the j-th area of the convolution kernel of the first layer, K represents the size of the convolution kernel, x l(i,j) Is the output characteristic of the ith convolution kernel of the first layer in the jth region.
Preferably, a batch normalization layer and a Swish activation function layer are added after convolution operation of the model, and the calculation formula of the Swish activation function is as follows:
Swish(X)＝X·Sigmoid(βX)
where X is a feature vector and β is an adjustable parameter used to control the shape of the function, sigmoid (·) represents the Sigmoid activation function.
Preferably, the one-dimensional expansion convolution layer is used for injecting holes into a standard one-dimensional convolution kernel to increase the receptive field of the model without reducing the resolution, and the expression is as follows:
in the method, in the process of the invention,is a feature of the jth region of the first layer, K is denoted as the convolution kernel original size, and r is denoted as the void fraction.
Preferably, the one-dimensional SE attention module is configured to perform adaptive weighted calibration on feature vectors of each channel, and the process includes:
compressing the feature vector through global average pooling to obtain a compressed feature vector:
wherein X= [ X ] 1 ,x 2 ,…,x c ]For one of them is led toThe eigenvector of the track, H is the length of eigenvector X, F sq (X) represents compressing X, wherein Z is a characteristic value after compression, a plurality of characteristic values Z after compression are spliced to obtain a compressed characteristic vector Z of the channel,
the relation among the characteristics of each channel in the reconstruction process of the compressed characteristic vector Z is learned through two full-connection layers to obtain the weight of each channel,
S＝F ex (Z,ω)＝σ(ω 2 δ(ω 1 Z))
omega in 1 、ω 2 Respectively representing the weights of two full-connection layers, delta is a ReLU function, sigma is a Sigmoid function, F ex (Z, ω) represents the excitation phase, S is the channel weight vector and is finally normalized to [0,1 ]]A section;
weighting the corresponding channel feature vector X by using the weight S in the normalized weight matrix S scale (X, s) to obtain a calibration vectorThe following formula is shown:
preferably, the output module of the improved EfficientNetV2 fault detection model comprises a one-dimensional convolution layer, an average pooling layer and a full connection layer, and is used for classifying health states according to the signal characteristics to obtain a bearing fault detection result.
Preferably, the improved EfficientNetV2 fault detection model is trained and detected under a cross-working condition based on transfer learning, and the training process comprises the following steps:
training the improved EfficientNetV2 based on a source domain working condition sample to obtain source domain fault detection model parameters;
and constructing the same improved EfficientNetV2 fault detection model in a target domain according to the model parameters, freezing parameters of a one-dimensional convolution, a Fused-MBConv1D module and an MBConv1D module, thawing parameters of an output module, and performing fine adjustment on the parameters of the output module by utilizing a target domain sample.
Preferably, constructing a joint loss function according to a cross entropy loss function and a center distance loss function, and training the improved EfficientNetV2 fault detection model according to the joint loss;
the joint loss function expression is:
L All ＝ηL+η C L C
wherein eta, eta C Is the weight of the two loss functions,
l is a cross entropy loss function, expressed as:
wherein N is the number of categories, y ic As a sign function, if the true class of sample i is c, y ic =1, otherwise y ic ＝0；p ic The prediction probability of the sample i belonging to the category c;
L C as a center distance loss function, the expression is:
in the formula, h i Representing a sample of the i-th class,representing the center samples of each class, n is the number of classes.
On the other hand, the invention also discloses a bearing cross-working condition fault detection system based on the improved EfficientNet V2, which stores the improved EfficientNet V2 fault detection model in the bearing cross-working condition fault detection method based on the improved EfficientNet V2, extracts the characteristics of the vibration signals of the bearing by using the improved EfficientNet V2 fault detection model to obtain signal characteristics, and performs fault detection on the bearing according to the signal characteristics.
The invention discloses a bearing cross-working condition fault detection method and system based on improved EfficientNetV2, which have the following effects compared with the prior art:
(1) The improved EfficientNetV2 fault detection model can be better adapted to one-dimensional state detection signals, so that calculation cost and characteristic loss during conversion into a two-dimensional image are avoided, and the real-time performance of fault detection is improved;
(2) The invention improves the EfficientNetV2 fault detection model, has a plurality of feature extraction modules, and can excavate the global features and local features of signals so as to improve the feature expression capability of the model;
(3) In addition, the invention effectively solves the problem of poor performance of the fault detection model under the condition of few samples in variable working conditions by utilizing the model migration strategy.
Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and claims thereof as well as the appended drawings.
The technical scheme of the invention is further described in detail through the drawings and the embodiments.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are required to be used in the embodiments or the description of the prior art will be briefly described below, and it is obvious that the drawings in the following description are only embodiments of the present invention, and that other drawings can be obtained according to the provided drawings without inventive effort for a person skilled in the art.
FIG. 1 is a schematic block diagram of the Fused-MBConv1D layer of the present invention;
FIG. 2 is a schematic block diagram of MBConv1D layer in the present invention;
FIG. 3 is a schematic diagram of the overall structure of the improved EfficientNetV2 fault detection model of the present invention;
FIG. 4 is a flow chart of training and detection of the improved EfficientNetV2 fault detection model of the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The embodiment of the invention discloses a bearing cross-working condition fault detection method based on improved EfficientNetV2, which can be used for mining local features and global features of signals, can be better adapted to real-time detection of one-dimensional signals, and can be used for accurately identifying the health state of a bearing under the condition of few samples in the cross-working condition. The following is a description of specific examples.
Example 1
The method specifically comprises the following steps:
an improved EfficientNetV2 fault detection model is constructed and used for extracting characteristics of vibration signals of the bearing and detecting faults of the bearing according to the extracted characteristics;
in general, for different sensors, the collectable bearing state monitoring signals include temperature signals, acoustic emission signals, static signals, oil signals, vibration signals and the like, wherein the fault detection is performed based on the vibration signals in consideration of the fact that the vibration signals are directly related to the working state of the bearing.
Collecting vibration signals of the bearing in a normal state and an abnormal state under the working conditions of a source domain and a target domain, and intercepting a sufficient number of short-time samples from samples collected continuously for a long time to construct a data set and adding corresponding state labels, wherein each sample at least comprises data of one rotation period so as to ensure that the data has enough state information; the target domain samples are randomly divided into training sets and test sets.
Further, the improved EfficientNetV2 fault detection model in the invention is obtained by improving the EfficientNetV2 model. The EfficientNet V2 is an efficient convolutional neural network architecture, is now used for two-dimensional image processing, and is suitable for one-dimensional signal analysis by performing pruning on the basic architecture of the EfficientNet V2 in consideration of unnecessary calculation overhead and feature loss caused by converting one-dimensional signals into two-dimensional images.
The improved EfficientNet V2 fault detection model comprises a one-dimensional convolution module, a Fused-MBConv1D module, an MBConv1D module and an output module in sequence;
in one embodiment, the improved EfficientNetV2 fault detection model is formed by stacking 1 one-dimensional convolution Conv1D, 1 one-dimensional fusion mobile rollover bottleneck convolution Fused-MBConv1D module, 1 one-dimensional mobile rollover bottleneck convolution MBConv1D module and 1 output module;
1. one-dimensional convolution Conv1D
The one-dimensional convolution Conv1D is used for capturing local characteristics of the vibration signal; the input signal is slidingly convolved by defining the size and step size of the convolution kernel to capture local features of the signal for subsequent processing and extraction of the signal. And the calculation formula is as follows,
in the method, in the process of the invention,is the output feature of the j-th region of layer 1>Is the weight of the jth region of the ith convolution kernel of the ith layer, +.>Is the offset of the j-th region of the layer i convolution kernel, K represents the width of the convolution kernel,x l(i,j) is the output characteristic of the ith convolution kernel of the first layer in the jth region;
further, to reduce the data distribution differences between different layers, a batch normalization layer (Batch Normalization) is added after all the convolution layers, so that the data obeys the normal distribution with the mean value of 0 and the variance of 1; all convolutions include one-dimensional convolutions, dilated convolutions, depth convolutions, and point-by-point convolutions.
In one embodiment, a Swish activation function is also introduced for non-linear transformation, and in the same case, a Swish activation function with a smoother curve than a ReLU activation function helps to avoid gradient extinction. The formula is shown below, where β is an adjustable parameter used to control the shape of the function;
Swish(X)＝X·sigmoid(βX)
the BN layer and the Swish activation function are accessed after all convolution operations in the model, and are not described in detail below.
2. One-dimensional fusion mobile overturning bottleneck convolution module
The one-dimensional fusion mobile overturning bottleneck convolution module can be expressed as Fused-MBConv1D and is used for primarily and rapidly mining local features and global features of signals;
specifically, the Fused-MBConv1D module consists of a first one-dimensional convolution, a first one-dimensional expansion convolution (DilatedConv 1D), a first one-dimensional SE (Squeeze and Excitation) attention mechanism, a second one-dimensional convolution, a first Shortcut connection, and the like; as shown in fig. 1;
1) First one-dimensional convolution
The EfficientNetV2 authors found that in the shallow network of the model, the deep convolution was instead less efficient than the normal convolution, and therefore proposed that the Fused-MBConv module structurally replaced the one-dimensional deep separable convolution in the MBConv module with the normal convolution. The method also takes this point into consideration, replaces the one-dimensional depth separable convolution in the MBConv1D module with the one-dimensional convolution, and more specifically combines the depth convolution and the point-by-point convolution into one-dimensional convolution operation.
2) First one-dimensional dilation convolution
Injecting holes in standard one-dimensional convolution kernels without degradationThe receptive field of the model is increased with resolution, thereby enhancing the ability of the model to capture a wider range of features. The formula is shown below, whereinIs a feature of the jth region of the first layer, K is denoted as the convolution kernel original size, and r is denoted as the void fraction.
3) First one-dimensional SE (Squeeze-and-specification) attention module
In order to realize the self-adaptive calibration of the channel characteristics, a one-dimensional SE attention mechanism is used for capturing global characteristics of each channel, calculating the channel weight, compressing, exciting and calibrating the global characteristics of each channel, and enhancing the expression capacity of the network on the global characteristics.
The SE mechanism is divided into three processes, namely compression (carried out through global average pooling), excitation (each channel weight is learned through two layers of fully connected layers) and weighting, so that the attention to important characteristic channels is enhanced,
and compressing the feature vector through global average pooling to obtain a compressed feature vector. Let x= [ X ] 1 ,x 2 ,…,x c ]For each channel feature vector, H is the length of feature vector X, F sq (X) represents that X is compressed, z is a characteristic value after compression, and the compressed characteristic values of all channels can be spliced into a one-dimensional compressed characteristic vector;
at F ex The (Z, omega) excitation stage learns the weight of each channel through two fully connected layers, omega 1 、ω 2 Representing the weights of two fully connected layers respectively, delta is a ReLU function, sigma is a Sigmoid function, S is a channel weight vector and is finally normalized to [0,1 ]]A section; the compression characteristic values of all channels can be spliced into one-dimensional compression of the feature vector;
S＝F ex (Z,ω)＝σ(ω 2 δ(ω 1 Z))
weighting the corresponding channel feature vector X by using the weight S in the normalized weight matrix S scale (X, s) to obtain a calibration vectorThe following formula is shown:
4) Second one-dimensional convolution
A second one-dimensional convolution is introduced to further aggregate and extract features weighted by SE attention mechanisms.
5) First Shortcut connection
Fused-MBConv1D has a Shortcut connection structure so that input and output can be directly connected, the characterization capability of the network on global features and local features is improved, and the risk of gradient disappearance or explosion in the model training process is reduced.
3. One-dimensional mobile overturning bottleneck convolution module
In this embodiment, the one-dimensional mobile rollover bottleneck convolution module may be denoted as MBConv1D, which is a main feature extraction layer of the network, for further mining of richer local and global feature information.
The first one-dimensional convolution operation in the Fused-MBConv1D is replaced by a one-dimensional depth separable convolution operation, namely MBConv1D, and compared with the Fused-MBConv1D, the method has the advantages that the parameter is increased, but more efficient feature extraction and information fusion are realized;
namely MBConv1D, comprising a one-dimensional depth convolution, a second one-dimensional expansion convolution, a second one-dimensional SE attention module, a one-dimensional point-by-point convolution and a second Shortnut connection, wherein the one-dimensional depth convolution and the one-dimensional point-by-point convolution form a one-dimensional depth separable convolution, as shown in figure 2;
the one-dimensional depth separable convolution decomposes a complete convolution operation into two steps of one-dimensional depth convolution and one-dimensional point-by-point convolution, and specifically, when the MBConv1D is utilized to further mine the features, the data sequentially passes through the one-dimensional depth convolution, the one-dimensional expansion convolution, the SE attention mechanism and the one-dimensional point-by-point convolution; the method comprises the steps of firstly, independently carrying out convolution operation on each channel by depth convolution, carrying out one-dimensional expansion convolution and SE attention mechanism, then, merging multi-channel depth convolution output by using convolution check with a size of 1 by point-to-point convolution, and enabling a model to be more sensitive to characteristic change in signals by independently modeling channel characteristics and space characteristics, thereby improving the characteristic expression capability of the model.
Example two
On the basis of the scheme disclosed by the embodiment I, the improved EfficientNetV2 fault detection model further comprises an output module, a detection module and a detection module, wherein the output module is used for directly identifying and outputting a fault diagnosis result;
the output module comprises: a one-dimensional convolution layer, an average pooling layer and a full connection layer.
The one-dimensional convolution layer is used for further extracting local features of the signals and integrating the features;
the important features are reserved and the unimportant features are removed through the average pooling operation, so that the functions of downsampling and overfitting prevention are achieved, the formula is shown as follows, wherein W is the width of a pooling layer;
adding a full connection layer at the end of the module, and remapping the separable features learned in the high-dimensional space to a mark space of the sample; the feature classification was performed by following the log (softmax) function at the fully connected layer, the formula is shown below.
The bearing fault detection model constructed by the invention can be better adapted to one-dimensional state monitoring signals, has the capability of mining global and local characteristics of the signals, and improves the real-time performance and accuracy of fault detection.
However, as the actual running environment tends to be complex, the working condition and condition of the bearing often dynamically change according to the on-site production requirement, the bearing fault detection often faces the problems of fewer test sets and non-identical distribution of data of the training set, so that the existing majority of detection methods have an unbreakable force when the running state of the bearing is identified.
The method is based on the migration learning of the model, and the source domain fault detection model which completes training under the working condition of sufficient data and labels is applied to the detection task of the target domain, so that the problem that the sampling state monitoring signal of the target working condition is insufficient to train a new detection model is avoided, and the generalization performance of the fault detection method under the cross-working condition scene is improved.
In addition, because the model input is a one-dimensional signal and the model parameter scale is smaller, a progressive strategy is not needed in the model optimization process, so that the application selects a Ranger optimization algorithm to train an improved EfficientNetV2 fault detection model. The Ranger algorithm combines the advantages of RAdam and LookAhead optimization algorithms, can utilize a dynamic rectifier to adjust Adam self-adaptive momentum according to changes, can also maintain a training process of updating a stable model of a slow parameter, and has the advantages of high convergence rate, stable parameter, good generalization performance and the like.
The specific training procedure comprises:
improving the EfficientNetV2 fault detection model based on the known working condition data and the label training until the iteration times or the detection accuracy reaches a preset threshold value, stopping training, and obtaining a source domain detection model;
as shown in fig. 3, the same improved afflicientnetv 2 fault detection model framework is further built under the target domain working condition, and the source domain detection model parameters are used for initializing the target domain detection model; the freezing improvement improves model parameters in the operation of the one-dimensional convolution Conv1D, fused-MBConv1D module and the MBConv1D module in the EfficientNet V2, and only the parameters in the output module are unfrozen for fine adjustment; retraining the thawing parameters by using the target domain training set sample so as to quickly adapt to a target domain detection task;
and finally, as shown in fig. 4, inputting the test set sample of the target domain into the improved EfficientNetV2 fault detection model of the target domain, and verifying the actual detection effect of the model under the condition of less samples under the cross-working condition.
The improved EfficientNet V2 obtained based on the EfficientNet V2 is used as a feature extractor of the fault detection model, so that the fault detection accuracy of the model under the working conditions of a source domain and a target domain is improved.
In this embodiment, a joint loss function is constructed when training the improved EfficientNetV2 fault detection model.
The joint loss function consists of a cross entropy loss function and a center distance loss function, wherein the cross entropy loss reflects the deviation between a predicted label and a real label and represents training accuracy; the center distance loss reflects the intra-class aggregation and inter-class dispersion of the sample features after improved EfficientNetV2 extraction, representing feature extraction quality. Therefore, the combined loss function reflects the training effect of the model more comprehensively, and is beneficial to enhancing the discrimination accuracy of the model and improving the generalization of the model.
Wherein the cross entropy loss function is as follows,
where N is the number of categories, y ic As a sign function, if the true class of sample i is c, y ic =1, otherwise y ic ＝0；p ic The prediction probability of the sample i belonging to the category c;
the center distance loss function formula is shown below,
wherein h is i Representing a sample of the i-th class,representing various typesCenter sample, n is the number of categories;
the joint loss function is represented as follows, where η, η C Weights for two loss functions;
L All ＝ηL+η C L C
to verify the effectiveness of the detection method of the present application, the present example uses the published dataset of kesixi Chu Da (Case Western Reserve University, CWRU) as an example.
First, an experimental dataset is constructed. The CWRU data set is used for carrying out data sampling on a normal bearing and a manual damage fault bearing adopting electric spark machining, the driving end vibration acceleration data in the CWRU data set is used for carrying out experimental verification of the embodiment, and the cross-working condition tasks of fault detection under four radial loads of 0HP, 1HP, 2HP, 3HP and the like are set, and the number of each working condition data set and the number of the cross-working condition fault detection task are shown in the following table. Under each operating condition, a dataset was constructed as follows: according to the fault position and damage level of the bearing, 10 bearing state types can be divided, wherein the bearing position comprises three types of rolling body faults, inner ring faults and outer ring faults, the damage level is divided into three types of 7 mil, 14 mil and 21 mil, and the normal state of the bearing is also provided; under the working condition of a source domain, randomly intercepting 100 samples from continuous long-time sampling samples of each state type, wherein each sample comprises 1024 non-overlapping data points, and the 1024 non-overlapping data points are all used as a training set; under the target domain working condition, according to 10:1, randomly intercepting 10 samples from continuous long-time sampling samples of each state type to serve as a training set, simulating a few sample scene, intercepting 30 non-repeated samples to serve as a test set, wherein each sample also comprises 1024 non-overlapping data points;
secondly, an improved EfficientNetV2 fault detection model is constructed, wherein parameters involved comprise the number of channels, the convolution kernel size, the convolution step length, the void ratio and the like of each module, and in the embodiment, parameter setting examples are given as shown in the following table;
| module | Convolution kernel size | Step size | Void fraction | Number of channels |
| One-dimensional convolution module | 3 | 1 | - | 16 |
| Fused-MBConv1D module | 3 | 1 | 8 | 32 |
| MBConv1D module | 3 | 1 | 32 | 64 |
| Output module | 1 | 2 | - | 128 |
Thirdly, calculating a source domain data joint loss function, and optimizing parameters of an improved EfficientNetV2 fault detection model by using a Ranger algorithm to realize joint loss minimization;
fourthly, initializing model parameters of a target domain by utilizing a source domain improved EfficientNetV2 fault detection model, and fine-tuning thawing layer parameters of the model by using training set data of the target domain;
fifthly, detecting target domain test set data by using an improved EfficientNetV2 fault detection model after training; and (3) finishing all the cross-working condition detection tasks, and verifying the actual effect of the method in the cross-working condition fault detection, as shown in the following table.
| Cross-domain detection task numbering | Target domain test set joint error | Target domain test set detection accuracy |
| A→B | 3.3085e-04 | 1.0000 |
| A→C | 6.1091e-07 | 1.0000 |
| A→D | 7.4303e-04 | 1.0000 |
| B→A | 0.0010 | 1.0000 |
| B→C | 0.0000e+00 | 1.0000 |
| B→D | 3.7550e-07 | 1.0000 |
| C→A | 0.0012 | 1.0000 |
| C→B | 1.9789e-07 | 1.0000 |
| C→D | 1.1921e-09 | 1.0000 |
| D→A | 0.0653 | 0.9950 |
| D→B | 0.0058 | 1.0000 |
| D→C | 4.2341e-06 | 1.0000 |
In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant points refer to the description of the method section.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (9)
1. The bearing cross-working condition fault detection method based on the improved EfficientNetV2 is characterized by comprising the following steps of:
an improved EfficientNetV2 fault detection model is constructed and used for extracting characteristics of vibration signals of the bearing to obtain signal characteristics, and fault detection is carried out on the bearing according to the signal characteristics; wherein,
the improved EfficientNet V2 fault detection model sequentially comprises a one-dimensional convolution module, a Fused-MBConv1D module, an MBConv1D module and an output module;
the one-dimensional convolution is used for capturing local characteristics of the vibration signal;
the Fused-MBConv1D module sequentially comprises a first one-dimensional convolution, a first one-dimensional expansion convolution, a first one-dimensional SE attention module, a second one-dimensional convolution and a first Shortcut connection layer, and is used for rapidly extracting features according to the local features to obtain first local features;
the MBConv1D module sequentially comprises a one-dimensional depth convolution, a second one-dimensional expansion convolution, a second one-dimensional SE attention module, a one-dimensional point-by-point convolution and a second Shortnut connection layer, and is used for further mining information in the first local feature to obtain a signal feature.
2. The bearing cross-working condition fault detection method based on improved EfficientNet V2 of claim 1, wherein the one-dimensional convolution slidingly convolves the vibration signal by defining the size and step size of a convolution kernel to capture the local characteristics of the signal, and the calculation formula is:
in the method, in the process of the invention,is the output characteristic, omega, of the j-th region of the layer 1 i l(j) Is the weight of the jth region of the ith convolution kernel of the ith layer, +.>Is the offset of the j-th area of the convolution kernel of the first layer, K represents the size of the convolution kernel, x l(i,j) Is the output characteristic of the ith convolution kernel of the first layer in the jth region.
3. The bearing cross-working condition fault detection method based on improved EfficientNet V2 of claim 1, wherein a batch normalization layer and a Swish activation function layer are added after convolution operation of the improved EfficientNet V2 fault detection model, and the Swish activation function calculation formula is as follows:
Swish(X)＝X·Sigmoid(βX)
where X is a feature vector and β is an adjustable parameter used to control the shape of the function, sigmoid (·) represents the Sigmoid activation function.
4. The improved Efficient NetV2 based bearing cross-condition fault detection method of claim 1, wherein the first/second one-dimensional expansion convolution layer is used for injecting holes in a standard one-dimensional convolution kernel to increase the receptive field of a model without reducing resolution, and the expression is:
in the method, in the process of the invention,is a feature of the jth region of the first layer, K is denoted as the convolution kernel original size, and r is denoted as the void fraction.
5. The method for detecting bearing cross-working condition fault based on improved EfficientNetV2 according to claim 1, wherein the one-dimensional SE attention module is used for performing adaptive weighted calibration on feature vectors of each channel, and the process comprises the following steps:
compressing the channel feature vector through global average pooling to obtain a compressed feature vector:
wherein X= [ X ] 1 ,x 2 ,…,x c ]For the eigenvector of one of the channels, H is the length of the eigenvector X, F sq (X) represents compressing X, wherein Z is a characteristic value after compression, a plurality of characteristic values Z after compression are spliced to obtain a compressed characteristic vector Z of the channel,
the relation among the characteristics of each channel in the reconstruction process of the compressed characteristic vector Z is learned through two full-connection layers to obtain the weight of each channel,
S＝F ex (Z,ω)＝σ(ω 2 δ(ω 1 Z))
omega in 1 、ω 2 Respectively representing the weights of two full-connection layers, delta is a ReLU function, sigma is a Sigmoid function, F ex (Z, omega) tableShowing the excitation stage, S is the channel weight vector and is finally normalized to [0,1 ]]A section;
weighting the corresponding channel feature vector X by using the weight S in the normalized weight matrix S scale (X, s) to obtain a calibration vectorThe following formula is shown:
6. the bearing cross-working condition fault detection method based on the improved Efficient NetV2 of claim 1, wherein an output module of the improved Efficient NetV2 fault detection model comprises a one-dimensional convolution layer, an average pooling layer and a full connection layer, and is used for classifying health states according to the signal characteristics to obtain a bearing fault detection result.
7. The method for detecting bearing cross-working condition fault based on improved EfficientNet V2 according to claim 1, wherein the improved EfficientNet V2 fault detection model is trained and detected under the cross-working condition based on transfer learning, and the training process comprises the following steps:
training the improved EfficientNetV2 fault detection model based on a source domain working condition sample to obtain source domain fault detection model parameters;
and constructing the same improved EfficientNetV2 fault detection model in the target domain according to the source domain fault detection model parameters, freezing the parameters of the one-dimensional convolution, the Fused-MBConv1D module and the MBConv1D module, thawing the parameters of the output module, and performing fine adjustment on the parameters of the output module by utilizing a target domain sample.
8. The bearing cross-working condition fault detection method based on improved EfficientNet V2 according to claim 1, wherein a joint loss function is constructed according to a cross entropy loss function and a center distance loss function, and the improved EfficientNet V2 fault detection model is trained according to the joint loss function;
the joint loss function expression is:
L All ＝ηL+η C L C
wherein eta, eta C Is the weight of the two loss functions,
l is a cross entropy loss function, expressed as:
wherein N is the number of categories, y ic As a sign function, if the true class of sample i is c, y ic =1, otherwise y ic ＝0；p ic The prediction probability of the sample i belonging to the category c;
L C as a center distance loss function, the expression is:
in the formula, h i Represents a class i sample, c hi Representing the center samples of each class, n is the number of classes.
9. An improved EfficientNet V2-based bearing cross-working condition fault detection system, which is characterized in that an improved EfficientNet V2 fault detection model in the improved EfficientNet V2-based bearing cross-working condition fault detection method according to any one of claims 1-8 is stored, the improved EfficientNet V2 fault detection model is utilized to extract characteristics of vibration signals of a bearing, signal characteristics are obtained, and fault detection is carried out on the bearing according to the signal characteristics.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311479935.8A CN117516939A (en) | 2023-11-08 | 2023-11-08 | Bearing cross-working condition fault detection method and system based on improved EfficientNetV2 |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311479935.8A CN117516939A (en) | 2023-11-08 | 2023-11-08 | Bearing cross-working condition fault detection method and system based on improved EfficientNetV2 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117516939A true CN117516939A (en) | 2024-02-06 |
Family
ID=89765663
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311479935.8A Pending CN117516939A (en) | 2023-11-08 | 2023-11-08 | Bearing cross-working condition fault detection method and system based on improved EfficientNetV2 |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117516939A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117892258A (en) * | 2024-03-12 | 2024-04-16 | 沃德传动（天津）股份有限公司 | Bearing migration diagnosis method based on data fusion, electronic equipment and storage medium |
| CN117892258B (en) * | 2024-03-12 | 2024-06-07 | 沃德传动（天津）股份有限公司 | Bearing migration diagnosis method based on data fusion, electronic equipment and storage medium |
-
2023
- 2023-11-08 CN CN202311479935.8A patent/CN117516939A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117892258A (en) * | 2024-03-12 | 2024-04-16 | 沃德传动（天津）股份有限公司 | Bearing migration diagnosis method based on data fusion, electronic equipment and storage medium |
| CN117892258B (en) * | 2024-03-12 | 2024-06-07 | 沃德传动（天津）股份有限公司 | Bearing migration diagnosis method based on data fusion, electronic equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110059601B (en) | Intelligent fault diagnosis method for multi-feature extraction and fusion | |
| CN109765053B (en) | Rolling bearing fault diagnosis method using convolutional neural network and kurtosis index | |
| CN108319962B (en) | Tool wear monitoring method based on convolutional neural network | |
| CN108805083B (en) | Single-stage video behavior detection method | |
| CN108510153B (en) | Multi-working-condition rotary machine fault diagnosis method | |
| Che et al. | Hybrid multimodal fusion with deep learning for rolling bearing fault diagnosis | |
| CN107341452A (en) | Human bodys' response method based on quaternary number space-time convolutional neural networks | |
| CN107451565B (en) | Semi-supervised small sample deep learning image mode classification and identification method | |
| CN111539132B (en) | Dynamic load time domain identification method based on convolutional neural network | |
| CN110197205A (en) | A kind of image-recognizing method of multiple features source residual error network | |
| CN113392931B (en) | Hyperspectral open set classification method based on self-supervision learning and multitask learning | |
| CN113203566B (en) | Motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN | |
| CN113034483B (en) | Cigarette defect detection method based on deep migration learning | |
| CN112766283B (en) | Two-phase flow pattern identification method based on multi-scale convolution network | |
| CN113780242A (en) | Cross-scene underwater sound target classification method based on model transfer learning | |
| CN113392881A (en) | Rotary machine fault diagnosis method based on transfer learning | |
| CN116342894B (en) | GIS infrared feature recognition system and method based on improved YOLOv5 | |
| CN112799128A (en) | Method for seismic signal detection and seismic phase extraction | |
| CN103646256A (en) | Image characteristic sparse reconstruction based image classification method | |
| CN110930378A (en) | Emphysema image processing method and system based on low data demand | |
| CN108596044B (en) | Pedestrian detection method based on deep convolutional neural network | |
| CN113887342A (en) | Equipment fault diagnosis method based on multi-source signals and deep learning | |
| CN115358259A (en) | Self-learning-based unsupervised cross-working-condition bearing fault diagnosis method | |
| CN114118162A (en) | Bearing fault detection method based on improved deep forest algorithm | |
| CN117516939A (en) | Bearing cross-working condition fault detection method and system based on improved EfficientNetV2 |
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 |