Dual convolution dynamic domain adaptive equipment fault diagnosis method based on attention mechanism
By employing a dual-convolutional dynamic domain adaptation method based on an attention mechanism and utilizing an adaptive factor ω to adjust distribution differences, the problem of insufficient generalization ability of equipment fault diagnosis models in complex environments is solved, achieving high-accuracy fault diagnosis, which is suitable for the diagnosis of variable operating conditions of industrial production equipment.
Patent Information
- Application Number
- CN202310744126.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-21
- Publication Date
- 2025-10-24
- Estimated Expiration
- 2043-06-21
AI Technical Summary
Existing equipment fault diagnosis methods are prone to overfitting when faced with complex and ever-changing industrial production environments, leading to reduced diagnostic model performance. Furthermore, existing domain adaptation methods cannot effectively align individual similar features, resulting in insufficient generalization ability in target domains with large distributional differences.
We employ a dual-convolutional dynamic domain adaptation method based on an attention mechanism. By dynamically adjusting the edge distribution and conditional distribution through an adaptive factor ω, we can better align the data distributions of the source and target domains. We use a dual-channel convolutional neural network to extract key features and finely align distribution differences using the LMMD metric. We also optimize the loss function by combining the adaptive factor ω to enhance the model's generalization ability.
It significantly improves the diagnostic accuracy of equipment fault diagnosis models in the target domain, enables intelligent diagnosis under varying operating conditions, and solves the problem of insufficient generalization ability of existing methods in target domains with large distribution differences. It is suitable for fault diagnosis of industrial equipment under various operating conditions.
Smart Images
Figure CN116805051B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of industry manufacturing, and relates to a device fault diagnosis method, in particular to a double-convolution dynamic domain adaptive device fault diagnosis method based on an attention mechanism. BACKGROUND
[0002] With the rapid development of industrial Internet of Things, the integration and precision of industrial production equipment are continuously improved, but the complexity and uncertainty are also significantly enhanced, so the fault diagnosis of equipment has always been the most challenging problem in the field. Traditional fault diagnosis methods based on signal analysis include spectrum analysis, auto-regression and moving average model (ARMA), wavelet transform, etc. However, the feature extraction of these methods is manually performed, and the judgment result is too dependent on the prior knowledge of experts, which brings difficulties to the processing of big data and the learning of high-level features. In order to overcome this limitation, many researchers have turned their attention to artificial intelligence-based fault diagnosis methods.
[0003] Machine learning-based algorithms, such as artificial neural network (ANN), support vector machines (SVM), decision tree (Random Forest), etc., have the advantages of high diagnostic accuracy, high intelligence, and low demand for prior knowledge, and are widely used in fault diagnosis. In the face of complex and variable industrial data, deep learning-based algorithms are becoming increasingly popular, such as deep belief network (DBN), particle swarm optimization (PSO), convolutional neural network (CNN), etc. Deep learning can automatically learn fault features from collected data, and a large number of experiments have proved that deep learning has better generalization ability compared to machine learning-based methods. However, the deep learning method needs to use a large amount of labeled and same-distribution training data in the training process, but in the actual industrial production environment, due to the complex and variable operation conditions of industrial production equipment, the distribution of test data and training data is different, and overfitting phenomenon easily occurs in the training process, resulting in reduced performance of the fault diagnosis model.
[0004] The method of transfer learning can learn knowledge from other distribution different related data sets, build a robust model for the current fault classification task, and effectively solve the problems of insufficient labeled data and weak generalization ability. As an important branch of transfer learning, domain adaptation is aimed at the problem of the same task scene and different data distribution. Unsupervised domain adaptation can align the data distribution of the source domain and the target domain by learning the domain invariant features of the two domains, so as to train the model using the labeled data of the source domain to realize the fault diagnosis of the unlabeled data of the target domain. However, most of the existing fault diagnosis methods only align the edge distribution of the features, that is, only the overall distribution can be aligned, and the single similar feature cannot be aligned in a more fine-grained manner. A few methods that consider both edge distribution and conditional distribution, such as joint distribution adaptation, assign the same weight to edge distribution and conditional distribution, which obviously cannot achieve the best effect. SUMMARY
[0005] In view of the shortcomings of the prior art, the application provides a double convolution dynamic domain adaptation equipment fault diagnosis method based on an attention mechanism. The edge distribution and the conditional distribution are dynamically adjusted by an adaptive factor ω, so that the data distribution of the source domain and the target domain is better aligned, the generalization ability of the model is enhanced, the diagnostic accuracy of the equipment fault diagnosis model on the target domain is significantly improved, and intelligent diagnosis of equipment faults can be realized.
[0006] Step one, divide the collected data into labeled source domain data and unlabeled target domain data according to the working conditions and unlabeled target domain data wherein, represents the i-th sample of the source domain, represents the corresponding label, represents the i-th sample of the target domain, n s represents the number of source domain samples, n t represents the number of target domain samples. Since the working conditions of the source domain data and the target domain data are different, the distributions of the two are also different.
[0007] Step two, construct an equipment fault diagnosis neural network model, including a feature extractor, a domain discriminator and a classifier.
[0008] The feature extractor is a dual-channel convolutional neural network based on an attention mechanism, which is used to extract transferable features of input data, and includes a dual-channel convolutional layer, a feature fusion layer, an attention mechanism module, and a pooling layer. The dual-channel convolutional layer includes two convolutional channels with different depths, which respectively extract low-dimensional and high-dimensional features of the input data. The obtained features are normalized in batches and activated by a MetaAconC function, then fused in the feature fusion layer, and input into the attention mechanism module. The attention is focused on key feature information through global maximum pooling. Finally, the features are input into the pooling layer to complete down-sampling through global average pooling strategy, and the transferable features are obtained.
[0009] The domain discriminator includes two groups of cascaded fully connected hidden layers, a Relu activation function, and a dropout layer, which is used to determine whether the output features of the feature extractor belong to the source domain or the target domain.
[0010] The classifier includes a fully connected hidden layer and a Softmax activation function, which is used to determine the fault type of the input data of the feature extractor.
[0011] Step three, initialize the device fault diagnosis neural network model parameters, input the data collected in step one, extract the transferable features through the feature extractor, and then input them into the domain discriminator and the classifier respectively, calculate the classifier loss, the domain discriminator loss, and the difference measurement loss.
[0012] Since the labels of the target domain data are often not available in actual applications, the classifier loss is only calculated for the source domain data:
[0013]
[0014] wherein C represents the number of fault categories of sample labels in the source domain data, represents the probability that sample x i belongs to category c, G f represents the feature extractor, G y represents the classifier; G y (G f (x i )) represents the predicted label of sample x i .
[0015] The transferable features of the source domain and the target domain data are input into the domain discriminator to obtain domain discrimination prediction labels, and the cross-entropy loss is calculated with the real domain labels as the domain discriminator loss:
[0016]
[0017] wherein G d represents the domain discriminator, L d represents the cross-entropy loss, G d (Gf (x i ))、d i respectively represent the domain discriminative prediction label and the real domain label of the i-th sample.
[0018] The LMMD metric loss is calculated by using the source domain real label and the target domain prediction label:
[0019]
[0020] wherein and respectively represent the transferable features of the source domain sample and the target domain sample , and d represents the LMMD distance between the two.
[0021] The global difference metric d between the distribution distance of the source domain sample and the target domain sample is calculated by the similarity metric A-g (D s ∪D t ):
[0022] d A-g (D s ∪D t )=2(1-2(L g )) (4)
[0023] The local difference metric d l (D s ∪D t ) is calculated according to the LMMD metric criterion:
[0024]
[0025] The adaptive factor ω is calculated according to the global difference metric and the local difference metric:
[0026]
[0027] The importance of the two domain edge distribution and conditional distribution is dynamically adjusted by using the adaptive factor ω, and the following domain adaptation loss is designed:
[0028] L D =ωL g +(1-ω)L lmmd (7)
[0029] As a preferred, the sampling normal distribution random initialization method is used to initialize the equipment fault diagnosis neural network model parameters.
[0030] Step four, iterative training is carried out by using the source domain samples with labels and the target domain samples without labels, and the distribution difference between the source domain and the target domain samples is reduced by using an adaptive factor during the training process, and the total loss function of the equipment fault diagnosis neural network model is:
[0031] L total = L y + λL D = L y + λ(ωL g +(1-ω)L lmmd ) (8)
[0032]
[0033] Wherein, lambda represents a trade-off parameter, used to control the domain adaptation loss L D The weight in the whole loss function, epoch and totalepochs respectively represent the current iteration training times and the total iteration training times set.
[0034] As preferred, the equipment fault diagnosis neural network model parameters are updated by the SGD random gradient descent method during the optimization process.
[0035] Step five, the equipment running data collected by the sensor is input into the model trained in step four, and the fault diagnosis result is obtained.
[0036] As preferred, the equipment running data collected by the sensor is subjected to fast Fourier transform to obtain a corresponding one-dimensional frequency domain vibration signal, and then input into the trained equipment fault diagnosis neural network model for fault diagnosis.
[0037] The present application has the following beneficial effects:
[0038] 1. For the working condition with large difference, the method can autonomously select more suitable features for migration through the adversarial training between the feature extractor, the classifier and the domain discriminator, and better fault diagnosis results are obtained. The feature extractor adopts double-channel convolution and attention mechanism, so that the extracted features are more critical. The domain discrimination loss is used to measure the global distribution difference between the source domain and the target domain, and the LMMD measurement criterion is used to more finely align the local distribution difference between the source domain and the target domain, compared with the method of only considering the global distribution difference or the local distribution difference, better migration effect can be obtained while only a small amount of parameter quantity is increased.
[0039] 2. The adaptive factor is used to stably and quantitatively dynamically adjust the proportion of marginal distribution and conditional distribution differences in the overall data distribution difference, so that the model can better align the data distribution between the source domain and the target domain, enhance the generalization ability of the model in the target domain with large distribution differences, significantly improve the diagnostic accuracy of the equipment fault diagnosis model, and solve the problem that common domain adaptation algorithms have poor generalization ability for target domains with large differences. The model still has good diagnostic effects when facing variable working conditions, and can be widely used in fault diagnosis tasks of production equipment in various industrial fields under variable working conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 This is a flow chart of the equipment fault diagnosis method;
[0041] Figure 2 Schematic diagram of the neural network model structure constructed in the embodiment;
[0042] Figure 3 This is the flow chart of the dynamic domain adaptation algorithm. DETAILED DESCRIPTION
[0043] The present invention will be further explained below with reference to the accompanying drawings;
[0044] like Figure 1 As shown in FIG, the dual-convolution dynamic domain adaptation device fault diagnosis method based on the attention mechanism includes the following steps:
[0045] Step 1: Divide the collected data into labeled source domain data according to the working conditions and unlabeled target domain data in, represents the i-th sample in the source domain, Indicates the corresponding label, represents the i-th sample in the target domain, n s Indicates the number of source domain samples, n t Indicates the number of target domain samples.
[0046] Step 2: Build Figure 2 The equipment fault diagnosis neural network model shown includes a feature extractor, a domain discriminator, and a classifier.
[0047] Step 3: Initialize the parameters of the equipment fault diagnosis neural network model using the random initialization method of the sampled normal distribution. Input the data collected in step 1, extract the transferable features through the feature extractor, and then input them into the domain discriminator and classifier respectively to calculate the classifier loss, domain discriminator loss and difference metric loss.
[0048] The classifier loss shown is calculated only on the source domain data:
[0049]
[0050] wherein C represents the number of fault categories of sample labels in the source domain data, denotes the probability that sample x i belongs to category c, G f represents a feature extractor, G y represents a classifier; G y (G f (x i )) denotes the predicted label of sample x i .
[0051] The transferable features of the source domain and the target domain data are input into a domain discriminator to obtain a domain discrimination predicted label, and a cross-entropy loss is calculated with a real domain label as a domain discriminator loss:
[0052]
[0053] wherein G d represents a domain discriminator, L d represents a cross-entropy loss, G d (G f (x i )) and d i respectively represent the domain discrimination predicted label and the real domain label of the i th sample.
[0054] The source domain real label and the target domain predicted label are used to calculate a LMMD metric loss:
[0055]
[0056] wherein and respectively represent the transferable features of the source domain sample and the target domain sample , and d represents the LMMD distance between the two:
[0057]
[0058] wherein represents the weight of the i th source domain sample belonging to category c, represents the weight of the i th target domain sample belonging to category c, and respectively represent the features of the source domain sample and the target domain sample . The LMMD can be used to calculate the feature mapping distance value between the source domain and the target domain, and by reducing the distribution difference of each category activation, the transfer ability is enhanced.
[0059] Compared to the more common MMD (Maximum Mean Discrepancy) distance, LMMD prioritizes the distribution of subdomains within the same category, aligning the source and target domain data distributions at a finer granularity. By iteratively optimizing the model with LMMD, the model achieves domain-invariant representation capabilities across both the source and target domains.
[0060] By similarity measurement Calculate the global difference metric d between the distribution distance between source domain samples and target domain samples A-g (D s ∪D t ):
[0061] d A-g (D s ∪D t )=2(1-2(L g )) (5)
[0062] Calculate the local difference metric d according to the LMMD metric criterion l (D s ∪D t ):
[0063]
[0064] The adaptive factor ω is calculated based on the global difference metric and the local difference metric:
[0065]
[0066] The adaptive factor ω is used to dynamically adjust the importance of the marginal distribution and conditional distribution of the two domains, and the following domain adaptation loss is designed:
[0067] L D =ωL g +(1-ω)L lmmd (8)
[0068] Step 4: Figure 3 As shown in the figure, labeled source domain samples and unlabeled target domain samples are used for iterative training. The parameters of the equipment fault diagnosis neural network model are updated by the SGD stochastic gradient descent method. During the training process, the adaptive factor is used to reduce the distribution difference between the source domain and target domain samples. The total loss function of the equipment fault diagnosis neural network model is:
[0069] L total =L y +λL D =L y +λ(ωL g +(1-ω)L lmmd ) (9)
[0070]
[0071] wherein, λ represents a trade-off parameter, used to control the domain adaptation loss L D The weight in the whole loss function, epoch and totalepochs represent the current iteration training times and the total iteration training times set respectively. The automatic updating strategy of the trade-off parameter optimizes the generalization ability of the diagnosis model, and improves the robustness of the model.
[0072] Step five, the fast Fourier transform is carried out on the equipment operation data collected by the sensor, one-dimensional frequency domain vibration signals corresponding are obtained, and the model trained in step four is input, so that the fault diagnosis result is obtained.
[0073] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A double convolution dynamic domain adaptive equipment fault diagnosis method based on an attention mechanism, characterized in that: Specifically comprising the following steps: Step one, according to the working condition, the collected data is divided into labeled source domain data and unlabeled target domain data Wherein, represents the i-th sample of the source domain, represents the corresponding label, represents the i-th sample of the target domain, n s represents the number of source domain samples, n t represents the number of target domain samples; Step two, constructing a device fault diagnosis neural network model, including a feature extractor, a domain discriminator and a classifier; The feature extractor uses two different depth convolution channels to extract low-dimensional and high-dimensional features of the input data, performs attention calculation after fusion, and outputs transferable features; the domain discriminator is used to judge whether the output features of the feature extractor belong to the source domain or the target domain; and the classifier is used to judge the fault type of the input data of the feature extractor; Step three, initializing the device fault diagnosis neural network model parameters, inputting the data collected in step one, extracting transferable features through the feature extractor, and then inputting the domain discriminator and the classifier respectively, calculating the classifier loss, the domain discriminator loss and the difference measurement loss; The classifier loss is only calculated for the source domain data: wherein C represents a number of failure classes of sample labels in the source domain data, denotes a sample x i belongs to a class c, G f represents a feature extractor, G y represents a classifier; G y (G f (x i )) denotes a predicted label of a sample x i ; The domain discriminator loss is obtained by calculating the cross-entropy loss of the domain discrimination prediction label and the real domain label: wherein G d represents a domain discriminator, L d represents a cross-entropy loss, G d (G f (x i )), d i respectively represent the domain discrimination prediction label and the real domain label of the i-th sample. The LMMD measurement loss is calculated using the source domain real label and the target domain prediction label: wherein and respectively represent transferable features of source domain samples target domain samples , d represents LMMD distance between them. By similarity measure Computing a global difference measure d of distribution distance between source domain samples and target domain samples A-g (D s ∪D t ): d A-g (D s ∪D t )=2(1-2(L g )) (4) According to the LMMD metric criterion, a local difference measure d is calculated l (D s ∪D t ): The adaptive factor ω is calculated according to the global difference measurement and the local difference measurement: The importance of the edge distribution and the conditional distribution of the two domains is dynamically adjusted using the adaptive factor ω, and the following domain adaptation loss is designed: L D = ωL g + (1 - ω)L lmmd (7) Step four, using the source domain samples with labels and the target domain samples without labels for iterative training, and using the adaptive factor to reduce the distribution difference between the source domain and the target domain samples during the training process, and the total loss function of the device fault diagnosis neural network model is: L total = L y + λL D = L y + λ(ωL g + (1 - ω)L lmmd ) (8) where λ denotes a trade-off parameter for controlling the domain adaptation loss L D The weight in the whole loss function, epoch and totalepochs represent the current iteration training times and the total iteration training times set respectively. Step five, inputting the device running data collected by the sensor into the model trained in step four to obtain the fault diagnosis result.
2. The method of claim 1, wherein the method is based on an attention mechanism and double convolution dynamic domain adaptation device fault diagnosis. The feature extractor is a dual-channel convolutional neural network based on an attention mechanism, which is used to extract transferable features of input data, including a dual-channel convolutional layer, a feature fusion layer, an attention mechanism module and a pooling layer; wherein the dual-channel convolutional layer includes two convolution channels with different depths, which extract low-dimensional and high-dimensional features of the input data respectively, and the obtained features are fused in the feature fusion layer after batch normalization and MetaAconC function activation, and then enter the attention mechanism module, concentrate the attention on the key feature information through global maximum pooling, and finally enter the pooling layer to complete the down-sampling through the global average value pooling strategy to obtain the transferable features.
3. The method of claim 1, wherein the method comprises: The domain discriminator includes two groups of cascaded fully connected hidden layers, a Relu activation function and a dropout layer.
4. The device fault diagnosis method based on the attention mechanism and double convolution dynamic domain adaptation according to claim 1, wherein: The classifier includes a fully connected hidden layer and a Softmax activation function.
5. The device fault diagnosis method based on the attention mechanism and double convolution dynamic domain adaptation according to claim 1, wherein: In step three, the device fault diagnosis neural network model parameters are initialized by the sampling normal distribution random initialization method.
6. The device fault diagnosis method based on the attention mechanism and double convolution dynamic domain adaptation according to claim 1, wherein: In the optimization process, the device fault diagnosis neural network model parameters are updated by the SGD stochastic gradient descent method.
7. The device fault diagnosis method based on the attention mechanism and double convolution dynamic domain adaptation according to claim 1, wherein: The device running data collected by the sensor is subjected to fast Fourier transform to obtain a corresponding one-dimensional frequency domain vibration signal, which is then input into the trained device fault diagnosis neural network model for fault diagnosis.