CN114580288A - Intelligent fault diagnosis method based on DSECJAN - Google Patents

Intelligent fault diagnosis method based on DSECJAN Download PDF

Info

Publication number
CN114580288A
CN114580288A CN202210225165.3A CN202210225165A CN114580288A CN 114580288 A CN114580288 A CN 114580288A CN 202210225165 A CN202210225165 A CN 202210225165A CN 114580288 A CN114580288 A CN 114580288A
Authority
CN
China
Prior art keywords
domain
fault diagnosis
dsecjan
convolution
model
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
Application number
CN202210225165.3A
Other languages
Chinese (zh)
Inventor
何志伟
刘才明
郑骁蓉
董哲康
高明煜
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hangzhou Dianzi University
Original Assignee
Hangzhou Dianzi University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Hangzhou Dianzi University filed Critical Hangzhou Dianzi University
Priority to CN202210225165.3A priority Critical patent/CN114580288A/en
Publication of CN114580288A publication Critical patent/CN114580288A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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
    • G06F2111/00—Details relating to CAD techniques
    • G06F2111/10—Numerical modelling
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Abstract

The invention discloses an intelligent fault diagnosis method based on DSECJAN, which consists of a convolutional neural network feature extraction module, an extrusion and excitation module, a fault classification module and a joint domain adaptation module and is used for intelligent cross-domain fault diagnosis of equipment. And a one-dimensional convolutional neural network feature extraction module is constructed, then channel features are subjected to self-adaptive enhancement or inhibition through an extrusion and excitation module, and joint domain adaptation is utilized to align the input features and the output labels of the source domain and the target domain so as to realize intelligent fault diagnosis. The model can adaptively enhance channel characteristics important for cross-domain diagnosis and inhibit information ineffective for learning tasks. The method has the advantages that the diagnosis accuracy of the equipment fault diagnosis model on the target domain is remarkably improved, and the problem that the generalization capability of the deep learning intelligent fault diagnosis algorithm on the target domain is not strong is solved.

Description

Intelligent fault diagnosis method based on DSECJAN
Technical Field
The invention relates to an intelligent fault diagnosis method for equipment, in particular to an intelligent fault diagnosis method based on DSECJAN.
Background
Mechanical equipment is one of key components which can be normally operated and produced in an intelligent factory, and is widely applied to the fields of battery manufacturing, automobile manufacturing, aerospace and the like. Due to the long-term use of the equipment and some human factors, the equipment is likely to malfunction, which causes great economic loss and personnel injury, and therefore, the fault location and maintenance of the mechanical equipment are particularly important. In practical industrial applications, equipment fault location requires sophisticated expertise and significant time cost. Because the intelligent fault diagnosis is combined with the artificial intelligence technology, the time cost can be effectively reduced. The traditional machine learning method comprises a Support Vector Machine (SVM), (support Vector machine), a Random Forest (RF), (random forest), a K-Neighbor (KNN), an Artificial Neural Network (ANN), and the like, and is widely applied to the field of equipment fault diagnosis. These machine learning fault diagnosis methods rely heavily on industry experts with professional knowledge to extract artificial features from original signals, and the methods for extracting features from signals are very different, and the effect of extracting features is also influenced by subjective factors. In addition, as the available manufacturing data has seen explosive growth, the devices have become more complex and the traditional machine learning methods have not been able to meet the current device intelligent troubleshooting needs.
In order to solve the problem of data volume increase, the method based on deep learning is widely applied to the field of intelligent fault diagnosis of equipment and achieves great results. Such as convolutional Neural network cnn (convolutional Neural network), Long-Short Term Memory network LSTM (Long-Short Term Memory), and deep Residual Shrinkage network drsn (deep Residual Shrinkage network). The deep learning intelligent fault diagnosis method can automatically extract fault characteristics and effectively locate faults. However, the deep learning fault diagnosis method needs to satisfy an assumption that the training data set and the test data set need to be obtained under the same working condition. When the deep learning fault diagnosis method is used for real-time diagnosis under different working conditions, the diagnosis performance is greatly reduced.
In conclusion, finding an effective cross-domain device fault diagnosis method to enhance the generalization capability of the training model on the target domain becomes a problem to be solved at present.
Disclosure of Invention
In order to solve the technical problems, the invention provides an intelligent fault diagnosis method based on DSECJAN, which obviously improves the diagnosis accuracy of an equipment fault diagnosis model on a target domain and realizes cross-domain intelligent fault diagnosis of equipment faults.
The technical scheme of the invention comprises the following steps:
the method comprises the following steps: dividing data into source domain data according to different working condition conditions
Figure BDA0003538934090000021
And target domain data
Figure BDA0003538934090000022
Where S denotes a source domain, T denotes a target domain, i denotes the ith sample of the source domain or the target domain, NSRepresenting the number of source domain samples, NTRepresenting a number of target domain samples;
step two: constructing a DSECJAN intelligent fault diagnosis model, and initializing parameters of the DSECJAN intelligent fault diagnosis model;
the DSECJAN model is fully called a deep compression and excitation convolution combined domain adaptive neural network, and the network comprises a convolution feature extractor, a compression and excitation module, a fault classifier and a combined domain adaptive alignment module;
nonlinear modification unit functions RELU are adopted in nonlinear activation layers of the DSECJAN model; a one-dimensional convolution neural network is adopted as a preceding stage feature extraction module of the whole framework; the first layer of convolution adopts convolution kernels larger than 16 x 16 to increase the receptive field, the other convolution layers all use convolution kernels of 3 x 3, and the maximum pooling function is adopted in the convolution feature extractor. The extrusion and excitation module comprises a global average pooling layer, two linear layers and a sigmoid layer, wherein nonlinear correction units are arranged between the linear layers, and the tail part of the extrusion and excitation module is connected with the maximum pooling layer and one linear layer. The joint domain adaptive alignment module optimizes the target using a joint maximum mean distribution metric. The method is characterized in that a normal distribution random initialization method is adopted for initializing parameters of the equipment intelligent fault diagnosis neural network, and the parameters are updated through an Adam algorithm.
Step three: training a DSECJAN model by using source domain and target domain data;
step four: and (4) iteratively optimizing the DSECJAN model according to the target function, and ending the training when the total loss function value is reduced to a certain value or the training times reach a set value to obtain the final intelligent fault diagnosis model.
Step five: and inputting the target domain data into the final model during equipment fault diagnosis to obtain an equipment fault diagnosis result.
Preferably, the DSECJAN model is trained by using source domain and target domain data, and the method specifically comprises the following steps:
preprocessing source domain and target domain data, and calculating a feature map after convolution operation and pooling operation
Figure BDA0003538934090000023
Expressed as:
Figure BDA0003538934090000024
in the formula, KlConvolution kernels of the first convolution layer, blDenotes the offset, x, of the first convolutional layerlRepresenting the input of the first convolutional layer. The feature extractor has four convolution layers in total, and the convolution calculation methods are all consistent with the formula (1). The squeeze and excitation module output is then calculated, and is expressed as:
Figure BDA0003538934090000031
in the formula, W1Represents the weight of the full connection layer FC1,
Figure BDA0003538934090000032
W2represents the weight of the full connection layer FC2,
Figure BDA0003538934090000033
tau is a hyper-parameter controlling the computational cost and capacity, C is the number of channels, N is the length of the input profile, xconvRepresenting the input to the squeeze and fire module,
Figure BDA0003538934090000034
xSEwhich represents the output of the same,
Figure BDA0003538934090000035
the fault classifier adopts a cross entropy loss function and sets a source domain sample label as yiE {1,2, 3.,. m }, m represents the total class number of the samples, and then the fault classification loss is obtained
Figure BDA0003538934090000036
Expressed as:
Figure BDA0003538934090000037
where n represents the batch size of the source domain samples during training, pjRepresenting the probability of the sample being predicted as the jth class, wherein I is a judgment function, if the input is true, the output is 1, otherwise, the output is 0;
the domain alignment penalty of the joint domain adaptation alignment module is expressed as:
Figure BDA0003538934090000038
wherein, P (x) represents the characteristic space distribution of the source domain sample, Q (x) represents the characteristic space distribution of the target domain, l represents the last l-layer network,
Figure BDA0003538934090000039
is a hilbert space feature mapping of the tensor product of the source domain features,
Figure BDA00035389340900000310
source domain data characteristics of the last but one layer network are represented;
Figure BDA00035389340900000311
except that the source domain is replaced with the target domain, meaning similar to that described above,
Figure BDA00035389340900000312
and representing target domain data characteristics of the last but one network.
Figure BDA00035389340900000313
Representing a space of r-order feature product vectors,
Figure BDA00035389340900000314
indicating a desire.
The overall loss function is weighted by the fault classification loss and the domain alignment loss and is expressed as follows:
Figure BDA00035389340900000315
wherein, thetaF，θSE，θC，θDRespectively representing the network parameters of the feature extractor, the extrusion and excitation module, the fault classifier and the feature alignment block. λ represents a weighting factor. And optimizing parameters of the DSECJAN model according to the total loss function value to finish one training. And updating the target by adopting an Adam learning algorithm, wherein a parameter updating formula is as follows:
Figure BDA00035389340900000316
Figure BDA0003538934090000041
Figure BDA0003538934090000042
Figure BDA0003538934090000043
in the formula, μ represents a learning rate.
Preferably, τ is 16.
The invention has the beneficial effects that:
the invention provides an intelligent fault diagnosis model based on a deep extrusion and excitation convolution combined domain adaptive neural network, which consists of a convolution neural network feature extraction module, an extrusion and excitation module, a fault classification module and a combined domain adaptive module and is used for intelligent cross-domain fault diagnosis of equipment. And a one-dimensional convolutional neural network feature extraction module is constructed, then channel features are subjected to self-adaptive enhancement or inhibition through an extrusion and excitation module, and joint domain adaptation is utilized to align the input features and the output labels of the source domain and the target domain so as to realize intelligent fault diagnosis. The model can adaptively enhance channel characteristics important for cross-domain diagnosis and inhibit information ineffective for learning tasks. The method has the advantages that the diagnosis accuracy of the equipment fault diagnosis model on the target domain is remarkably improved, and the problem that the generalization capability of the deep learning intelligent fault diagnosis algorithm on the target domain is not strong is solved.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is the overall structure diagram of the DSECJAN neural network of the present invention;
FIG. 3 is a graph showing the results of diagnosis.
Detailed Description
The invention is described in further detail below with reference to the figures and the embodiments.
As shown in fig. 1, the present invention comprises the following steps.
1) Dividing the data into source domain data and target domain data according to different working conditions;
2) constructing a DSECJAN intelligent fault diagnosis model, and initializing parameters of the DSECJAN intelligent fault diagnosis model;
3) training a DSECJAN model by using source domain and target domain data;
4) iteratively optimizing the DSECJAN model to obtain a final intelligent fault diagnosis model;
5) and inputting the target domain data into the final model during equipment fault diagnosis to obtain an equipment fault diagnosis result.
The step 1) divides the data collected by the sensor into source domain data according to the working condition of the equipment
Figure BDA0003538934090000051
And target domain data
Figure BDA0003538934090000052
Where S denotes a source domain, T denotes a target domain, i denotes the ith sample of the source domain or the target domain, NSRepresenting the number of source domain samples, NTRepresenting the number of samples in the target domain.
The overall structural schematic diagram of the DSECJAN intelligent fault diagnosis model in the step 2) is shown in FIG. 2, and the specific construction steps are as follows:
the DSECJAN intelligent fault diagnosis model is called depth extrusion and Excitation Convolution Joint domain adaptive neural network (Deep Squeeze-and-Excitation Convolution node Adaptation Networks). The network specifically comprises a convolution feature extractor, a squeezing and exciting module for enhancing beneficial channel features, a fault classifier for prediction, and a joint domain adaptive alignment module for reducing the joint distribution difference of input features and output labels of a source domain and a target domain. The nonlinear activation layers of the model all adopt nonlinear modification unit functions RELU (rectified Linear Unit). The ultimate goal of the model is to achieve better cross-domain device intelligent fault diagnosis by enhancing channel features that are more beneficial to the diagnostic task. Because the input is a one-dimensional time sequence signal, in order to effectively extract the fault characteristics of the original signal, a one-dimensional convolution neural network is designed as a preceding stage characteristic extraction module of the whole framework. The first layer of convolution adopts a large convolution kernel to increase the receptive field, the other convolution layers all use small convolution kernels, and a maximum pooling function is adopted in a convolution feature extractor to obtain important information irrelevant to the position. The extrusion and excitation module main body comprises a global average pooling layer, two linear layers and a sigmoid layer, wherein nonlinear correction units are arranged between the linear layers, and the tail part comprises a maximum pooling layer and one linear layer. The main body part of the extrusion and excitation module respectively compresses and releases the channel characteristic information globally to acquire the dependency relationship among the channel characteristics, so that the characteristic information more beneficial to the cross-domain diagnosis task is adaptively enhanced. The output shape of the fault classifier is the same as the class number of the samples, and the joint domain adaptation alignment module mainly functions to reduce the joint distribution difference of the input features and the output labels. The method is characterized in that a normal distribution random initialization method is adopted for initializing parameters of the equipment intelligent fault diagnosis neural network, and the parameters are updated through an Adam algorithm.
The DSECJAN model in the step 3) is specifically trained as follows:
1) preprocessing source domain data and target domain data, wherein the source domain data have labels, and the target domain data have no labels;
2) calculating fault classification loss according to the source domain data and the data labels;
let xlFor the input of the ith convolutional layer, the feature map after the convolution and pooling operations can be represented as:
Figure BDA0003538934090000053
in the formula, KlConvolution kernels of the first convolution layer, blIndicating the offset of the first convolutional layer. The feature extractor has four convolution layers in total, and the convolution calculation methods are all consistent with the formula (1). The squeeze and fire module output can be expressed as:
Figure BDA0003538934090000061
in the formula, W1Represents the weight of the full connection layer FC1,
Figure BDA0003538934090000062
W2represents the weight of the full connection layer FC2,
Figure BDA0003538934090000063
τ is a hyper-parameter that controls computation cost and capacity (setting τ to 16 may provide a good balance between accuracy and complexity). N is the length of the input feature map, xconvRepresenting the input to the squeeze and fire module,
Figure BDA0003538934090000064
xSEwhich represents the output of the same,
Figure BDA0003538934090000065
the feature map does not change in size after passing through the extrusion and excitation modules, but the importance of individual channel features has become different.
The fault classifier adopts a cross entropy loss function and sets a source domain sample label as yiE {1,2,3,. eta., m }, m represents the total number of classes of the sample (including normal class and fault class), and then the classification loss is calculated
Figure BDA0003538934090000066
Can be expressed as:
Figure BDA0003538934090000067
wherein n represents the batch size of the source domain sample during training, pjRepresenting the probability of the sample being predicted as class j, I is a decision function, and if the input is true, the output is 1, otherwise 0 is output.
3) Calculating domain alignment loss according to input features and output labels of the source domain and the target domain;
the domain alignment loss is mainly calculated in a joint domain adaptive alignment module, the joint domain adaptive alignment module mainly aims to align the joint distribution of the input features and the output labels of a source domain and a target domain, the feature space distribution metric adopts joint Maximum Mean distribution metric jmmd (joint Maximum Mean distribution), and the domain alignment loss can be expressed as:
Figure BDA0003538934090000068
wherein, P (x) represents the characteristic space distribution of the source domain sample, Q (x) represents the characteristic space distribution of the target domain, l represents the last l-layer network,
Figure BDA0003538934090000069
is a hilbert space feature mapping of the tensor product of the features of the source domain,
Figure BDA00035389340900000610
meaning is similar to that described above, except that the source domain is replaced with the target domain.
Figure BDA00035389340900000611
Representing an order r eigenproduct vector space.
4) Weighting the fault classification loss and the domain alignment loss to obtain a total loss function value;
during source domain pre-training, fault classification loss is used to optimize the target, and during domain adaptation training, fault classification loss and domain alignment loss are used to optimize the target. The overall loss function may be weighted from the fault classification loss and the domain alignment loss, and is expressed as follows:
Figure BDA00035389340900000612
wherein, thetaF，θSE，θC，θDRespectively representing the network parameters of the feature extractor, the extrusion and excitation module, the fault classifier and the feature alignment block. λ represents a weighting factor.
5) And optimizing parameters of the DSECJAN model according to the total loss function value to finish one training.
The training process of the DSECAN model comprises source domain sample pre-training in which fault classification loss is used for optimizing the target and domain adaptation training between the source domain and the target domain in which fault classification loss and domain alignment loss are used for optimizing the target. The goal is updated by adopting an Adam learning algorithm (Adaptive motion Estimation), and the parameter updating formula is as follows:
Figure BDA0003538934090000071
Figure BDA0003538934090000072
Figure BDA0003538934090000073
Figure BDA0003538934090000074
in the formula, μ represents a learning rate.
In the step 4), the DSECJAN model needs to be iteratively optimized according to the target function, and the training is finished when the total loss function value is reduced to a certain value or the training times reach a set value, so that the final intelligent fault diagnosis model is obtained.
And 5) inputting the target domain data into the final model during equipment fault diagnosis to obtain an equipment fault diagnosis result.
The invention provides an intelligent fault diagnosis model based on a deep extrusion and excitation convolution combined domain adaptive neural network in steps 2) and 3), which consists of a convolution neural network feature extraction module, an extrusion and excitation module, a fault classification module and a combined domain adaptive module and is used for intelligent cross-domain fault diagnosis of equipment. Specifically, a one-dimensional convolutional neural network feature extraction module is constructed, then channel features are adaptively enhanced or inhibited through an extrusion and excitation module, and joint domain adaptation is utilized to align the source domain and target domain input features and output labels in joint distribution so as to realize intelligent fault diagnosis. The model can adaptively enhance channel characteristics important for cross-domain diagnosis and inhibit information ineffective for learning tasks. The method has the advantages that the diagnosis accuracy of the equipment fault diagnosis model on the target domain is remarkably improved, and the problem that the generalization capability of the deep learning intelligent fault diagnosis algorithm on the target domain is not strong is solved.
The specific embodiment of the invention:
the experimental raw data set had 12 collected data under different working conditions for a total of 12000 samples. The data are divided into source domain data and target domain data according to different working conditions, the source domain data have 1000 samples, and each target domain data have 1000 samples. The source domain data is labeled, and the target domain data is unlabeled. The source domain data is numbered 0, the other 11 target domain data are numbered 1 to 11, and the source domain data is divided into 4: 1 ratio randomly divides the original training set and test set. And training by using source domain data, carrying out migration generalization of the DSECJAN model on the target domain through the source domain data and the target domain data, and finally inputting the target domain data into a final model to obtain a final fault diagnosis result.
The experimental environment of the invention is as follows: CPU is
Figure BDA0003538934090000081
CoreTMThe intelligent equipment fault diagnosis method based on the DSECJAN is characterized in that i7-6700K @4.00GHz, GPU is GTX1080 Ti, video memory is 11GB, the system is a Windows 10 operating system, a deep learning framework is Pythroch 1.8, and the intelligent equipment fault diagnosis method based on the DSECJAN is tested by utilizing Python 3.8.
In order to verify the effectiveness of the intelligent fault diagnosis method for the equipment based on DSECJAN, the method is compared with four typical Domain Adaptation models, namely a convolutional Neural network CNN (convolutional Neural network), a deep Domain Confusion network DDC (deep Domain fusion network), a joint Domain Adaptation network JAN (Joint Domain Adaptation network) and a deep-countermeasure Neural network DANN (deep adaptive Neural network). In particular, the CNN model is trained only by labeled samples of the source domain, and other models are jointly trained by labeled samples of the source domain and unlabeled samples of the target domain. For comparison, the structural design of the models adopts a CNN feature extractor, the number of network layers in the CNN is consistent with DSECJAN, and the models and the super-parameters of the DSECJAN are also set to be consistent. In order to objectively perform comprehensive analysis and comparison on each method, each model is completely trained 20 times, the models are trained for 300 epochs once, then each model is tested to obtain 20 diagnosis results, and the final results are averaged. The diagnostic result is shown in fig. 3, and it can be seen from the graph that DSECJAN achieves the highest diagnostic accuracy on other diagnostic tasks except for diagnostic tasks 0-3 and 3-0, and the reason why DSECJAN performs slightly worse on diagnostic tasks 0-3 and 3-0 is that the two operating condition data distributions have larger difference. The average diagnosis accuracy of the DSECJAN is the highest, so that the intelligent equipment fault diagnosis method based on the DSECJAN is an effective intelligent fault diagnosis method, and the equipment fault diagnosis model can be well and generally applied to target domains with different working conditions.

Claims (3)

1. An intelligent fault diagnosis method based on DSECJAN is characterized by comprising the following steps:
the method comprises the following steps: dividing data into source domain data according to different working condition conditions
Figure FDA0003538934080000011
And target domain data
Figure FDA0003538934080000012
Where S denotes a source domain, T denotes a target domain, i denotes the ith sample of the source domain or the target domain, NSRepresenting the number of source domain samples, NTRepresenting a number of target domain samples;
step two: constructing a DSECJAN intelligent fault diagnosis model, and initializing parameters of the DSECJAN intelligent fault diagnosis model;
the DSECJAN model is fully called a deep compression and excitation convolution combined domain adaptive neural network, and the network comprises a convolution feature extractor, a compression and excitation module, a fault classifier and a combined domain adaptive alignment module;
nonlinear modification unit functions RELU are adopted in nonlinear activation layers of the DSECJAN model; a one-dimensional convolution neural network is adopted as a preceding stage feature extraction module of the whole framework; the first layer of convolution adopts convolution kernels larger than 16 multiplied by 16 to increase the receptive field, the other convolution layers all use convolution kernels of 3 multiplied by 3, and the maximum pooling function is adopted in the convolution feature extractor; the extrusion and excitation module comprises a global average pooling layer, two linear layers and a sigmoid layer, wherein nonlinear correction units are arranged between the linear layers, and the tail part of the extrusion and excitation module is connected with the maximum pooling layer and one linear layer; the joint domain adaptive alignment module optimizes the target by adopting joint maximum mean distribution measurement; the method comprises the following steps that a normal distribution random initialization method is adopted for initializing parameters of the equipment intelligent fault diagnosis neural network, and the parameters are updated through an Adam algorithm;
step three: training a DSECJAN model by using source domain and target domain data;
step four: iteratively optimizing the DSECJAN model according to a target function, and ending training when the total loss function value is reduced to a certain value or the training times reach a set value to obtain a final intelligent fault diagnosis model;
step five: and inputting the target domain data into the final model during equipment fault diagnosis to obtain an equipment fault diagnosis result.
2. The DSECJAN-based intelligent fault diagnosis method according to claim 1, wherein: training a DSECJAN model by using source domain and target domain data, specifically:
preprocessing source domain and target domain data, and calculating a feature map after convolution operation and pooling operation
Figure FDA0003538934080000013
Expressed as:
Figure FDA0003538934080000014
in the formula, KlConvolution kernels of the first convolution layer, blDenotes the offset of the first convolution layer, xlAn input representing the first convolutional layer; the feature extractor has four convolution layers in total, and the convolution calculation methods are all consistent with the formula (1)(ii) a The squeeze and excitation module output is then calculated, and is expressed as:
Figure FDA0003538934080000015
in the formula, W1Represents the weight of the full connection layer FC1,
Figure FDA0003538934080000016
W2represents the weight of the full connection layer FC2,
Figure FDA0003538934080000017
tau is a hyper-parameter controlling the computational cost and capacity, C is the number of channels, N is the length of the input profile, xconvRepresenting the input to the squeeze and fire module,
Figure FDA0003538934080000018
xSEto represent the output of the same,
Figure FDA0003538934080000019
the fault classifier adopts a cross entropy loss function and sets a source domain sample label as yiE {1,2, 3.,. m }, m represents the total class number of the samples, and then the fault classification loss is obtained
Figure FDA0003538934080000021
Expressed as:
Figure FDA0003538934080000022
where n represents the batch size of the source domain samples during training, pjRepresenting the probability of the sample being predicted as the jth class, wherein I is a judgment function, if the input is true, the output is 1, otherwise, the output is 0;
the domain alignment penalty of the joint domain adaptation alignment module is expressed as:
Figure FDA0003538934080000023
wherein, P (x) represents the characteristic space distribution of the source domain sample, Q (x) represents the characteristic space distribution of the target domain, l represents the last l-layer network,
Figure FDA0003538934080000024
is a hilbert space feature mapping of the tensor product of the features of the source domain,
Figure FDA0003538934080000025
source domain data characteristics of the last but one layer network are represented;
Figure FDA0003538934080000026
except that the source domain is replaced with the target domain, meaning similar to that described above,
Figure FDA0003538934080000027
representing target domain data characteristics of the last but one layer network;
Figure FDA0003538934080000028
representing a space of r-order feature product vectors,
Figure FDA0003538934080000029
indicating a desire;
the overall loss function is weighted by the fault classification loss and the domain alignment loss and is expressed as follows:
Figure FDA00035389340800000210
wherein, thetaF，θSE，θC，θDRespectively representing feature extractor, squeeze and excitation module, fault classifier, feature alignment blockThe network parameters of (a); λ represents a weighting factor; optimizing DSECJAN model parameters according to the total loss function value to complete one-time training; and updating the target by adopting an Adam learning algorithm, wherein a parameter updating formula is as follows:
Figure FDA00035389340800000211
Figure FDA00035389340800000212
Figure FDA00035389340800000213
Figure FDA00035389340800000214
in the formula, μ represents a learning rate.
3. The DSECJAN-based intelligent fault diagnosis method according to claim 1, wherein: and tau is 16.
CN202210225165.3A 2022-03-09 2022-03-09 Intelligent fault diagnosis method based on DSECJAN Pending CN114580288A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210225165.3A CN114580288A (en) 2022-03-09 2022-03-09 Intelligent fault diagnosis method based on DSECJAN

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210225165.3A CN114580288A (en) 2022-03-09 2022-03-09 Intelligent fault diagnosis method based on DSECJAN

Publications (1)

Publication Number Publication Date
CN114580288A true CN114580288A (en) 2022-06-03

Family

ID=81778195

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210225165.3A Pending CN114580288A (en) 2022-03-09 2022-03-09 Intelligent fault diagnosis method based on DSECJAN

Country Status (1)

Country Link
CN (1) CN114580288A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115510926A (en) * 2022-11-23 2022-12-23 武汉理工大学 Cross-machine type diesel engine combustion chamber fault diagnosis method and system
CN115965057A (en) * 2022-11-28 2023-04-14 北京交通大学 Similar brain continuous learning fault diagnosis method for train transmission system
CN116401596A (en) * 2023-06-08 2023-07-07 哈尔滨工业大学（威海） Early fault diagnosis method based on depth index excitation network

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115510926A (en) * 2022-11-23 2022-12-23 武汉理工大学 Cross-machine type diesel engine combustion chamber fault diagnosis method and system
CN115965057A (en) * 2022-11-28 2023-04-14 北京交通大学 Similar brain continuous learning fault diagnosis method for train transmission system
CN115965057B (en) * 2022-11-28 2023-09-29 北京交通大学 Brain-like continuous learning fault diagnosis method for train transmission system
CN116401596A (en) * 2023-06-08 2023-07-07 哈尔滨工业大学（威海） Early fault diagnosis method based on depth index excitation network
CN116401596B (en) * 2023-06-08 2023-08-22 哈尔滨工业大学（威海） Early fault diagnosis method based on depth index excitation network

Similar Documents

Publication Publication Date Title
CN114580288A (en) Intelligent fault diagnosis method based on DSECJAN
WO2020244134A1 (en) Multi-task feature sharing neural network-based intelligent fault diagnosis method
CN110378435B (en) Apple leaf disease identification method based on convolutional neural network
CN108876044B (en) Online content popularity prediction method based on knowledge-enhanced neural network
CN114358124B (en) New fault diagnosis method for rotary machinery based on deep countermeasure convolutional neural network
CN111582396A (en) Fault diagnosis method based on improved convolutional neural network
CN115100709B (en) Feature separation image face recognition and age estimation method
CN113239897B (en) Human body action evaluation method based on space-time characteristic combination regression
CN114266301A (en) Intelligent power equipment fault prediction method based on graph convolution neural network
CN112784920A (en) Cloud-side-end-coordinated dual-anti-domain self-adaptive fault diagnosis method for rotating part
CN115659174A (en) Multi-sensor fault diagnosis method, medium and equipment based on graph regularization CNN-BilSTM
CN114740361A (en) Fuel cell voltage prediction method based on long-short term memory neural network model
CN116362325A (en) Electric power image recognition model lightweight application method based on model compression
CN114707698A (en) Long-term power load prediction method based on comprehensive shape error and time error
CN111695611B (en) Bee colony optimization kernel extreme learning and sparse representation mechanical fault identification method
CN112504682A (en) Chassis engine fault diagnosis method and system based on particle swarm optimization algorithm
CN111078895A (en) Remote supervision entity relation extraction method based on denoising convolutional neural network
CN117034003A (en) Full life cycle self-adaptive fault diagnosis method, system, equipment and medium for aerospace major product manufacturing equipment
CN114841076A (en) Power battery production process fluctuation abnormity detection method based on space-time diagram model
CN117407772B (en) Method and system for classifying training multi-element time sequence data by supervising and comparing learning network model
CN115290326A (en) Rolling bearing fault intelligent diagnosis method
CN111612733A (en) Convolutional neural network optimization method for medical image data analysis
CN115409217A (en) Multitask predictive maintenance method based on multi-expert hybrid network
CN115640531A (en) Fault diagnosis method based on residual error learning and attention mechanism fusion
CN115578325A (en) Image anomaly detection method based on channel attention registration network

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