A multi-objective domain equipment fault diagnosis method based on AdaDCLF

Abstract

This paper discloses a multi-target domain device fault diagnosis method based on AdaDCLF. This method combines the MK-MMD distance to measure the distribution deviation between the source and target domains, and then adaptively adjusts the domain confusion loss factor, focusing on the importance of the target domain data corresponding to the maximum domain confusion loss. This method, which adaptively adjusts the domain confusion loss factor, enhances the model's generalization capabilities across multiple target domains, significantly improving the diagnostic accuracy of the device fault diagnosis model in these domains. This method addresses the weak generalization capabilities of single-target domain adaptation algorithms across multiple target domains, thereby enabling intelligent diagnosis of device faults.

Description

Technical Field

[0001] The present invention relates to a device fault diagnosis method, in particular to a multi-target domain device fault diagnosis method based on AdaDCLF. Background Art

[0002] With the rapid development of industrial big data and the Internet of Things (IoT) in the context of Industry 4.0, intelligent health management of key industrial equipment is becoming increasingly important and a pressing issue for smart factories. In practical industrial applications, most equipment fault location and repair rely on prior experience or expert knowledge. This diagnostic approach is labor-intensive and time-consuming, resulting in low efficiency. To address this issue, intelligent equipment fault diagnosis, combining artificial intelligence (AI) and fault diagnosis techniques, has become a key branch of intelligent health management. Traditional machine learning methods, including support vector machines (SVMs), decision trees (Random Forests), and artificial neural networks (ANNs), have been widely used in equipment fault diagnosis. With the advent of the big data era, data-driven deep learning equipment fault diagnosis methods, such as convolutional neural networks (CNNs), deep belief networks (DBNs), and long-short term memory (LSTMs), have become increasingly popular. Under the premise of big data, these deep learning-based equipment fault diagnosis methods have been shown to have stronger learning capabilities than machine learning-based methods, resulting in higher diagnostic accuracy. However, equipment fault diagnosis methods based on machine learning and deep learning are based on the assumption that the training dataset and the test dataset used in actual industrial equipment are generated under the same operating conditions. In reality, the distributions of training and test datasets are often different. If these training datasets are used to train models, overfitting is likely to occur, resulting in a poor generalization of the actual test dataset data distribution and reduced performance of the equipment fault diagnosis model. The goal is to ensure that models trained using training datasets, even in the presence of dataset bias, can be applied to data under different operating conditions. This has significant application value for intelligent equipment fault diagnosis and reducing equipment maintenance costs.

[0003] In transfer learning, this problem is called domain adaptation, which involves taking a neural network trained on a source dataset and ensuring good accuracy on a target dataset that is significantly different from the source dataset. In real-world situations, machines typically operate under a variety of working conditions, which requires a training model that can adapt to changing working conditions. Domain adaptation methods can effectively address the problem of misalignment between the source and target domains. The main equipment fault diagnosis methods based on domain adaptation include feature-based domain adaptation, instance-based domain adaptation, and model-based domain adaptation. These domain adaptation methods, based on the powerful representation capabilities of deep learning, learn more transferable features by adapting to the target domain, thereby enhancing the performance of the trained model in the target domain. However, these domain adaptation methods are currently often based on a single target domain. In practical applications, there may be more than one target domain. If only one target domain is adapted, the performance of the trained model in other target domains may be reduced.

[0004] In summary, finding an effective multi-objective domain equipment fault diagnosis method to enhance the generalization ability of the training model in multi-objective domains has become an urgent problem to be solved. Summary of the Invention

[0005] To solve the above technical problems, the present invention proposes a multi-objective domain equipment fault diagnosis method based on AdaDCLF, which significantly improves the diagnostic accuracy of the equipment fault diagnosis model in multiple objective domains and realizes intelligent diagnosis of equipment faults. The technical solution of the present invention includes the following steps:

[0006] 1) Divide the data into source domain and target domain data according to different working conditions;

[0007] 2) Construct a neural network model for equipment fault diagnosis and initialize its parameters;

[0008] 3) Input the source domain and target domain data into the neural network and calculate the classification loss and domain confusion loss respectively;

[0009] 4) Calculate the domain confusion loss factor through AdaDCLF, obtain the total domain confusion loss based on the domain confusion loss factor, add the classification loss and the total domain confusion loss to obtain the total loss function value, and then perform iterative training to update the model parameters to obtain the final model;

[0010] The AdaDCLF stands for Adaptive domain confusion loss factor. This method adaptively adjusts the domain confusion loss factor between the source domain and each target domain, thereby improving the performance of the equipment fault diagnosis model in multiple target domains. The total loss function value of the equipment fault diagnosis neural network is calculated as follows:

[0011] L=LC +λL DC (1)

[0012] Among them L C represents the source domain classification loss, L DC represents the total domain confusion loss, λ represents the total domain loss penalty factor, the total domain confusion loss is composed of the source domain and each target domain confusion loss, L DC The calculation method is as follows:

[0013]

[0014] Where m represents the number of target domains, j represents the jth target domain, α is the domain confusion loss factor, and α is updated using the AdaDCLF strategy. The update formula is as follows:

[0015]

[0016] Where n represents the number of iterations, γ represents the update learning rate, and d j n represents the MK-MMD distance between the source domain and the jth target domain at the nth iteration. The AdaDCLF method can focus on improving the learning ability of the target domain corresponding to the maximum domain confusion loss. If the target domain with the largest deviation is learned at each iteration, it can be assumed that the model will not perform poorly on all known target domains of equipment fault data. AdaDCLF may reduce the performance of the model on a specific target domain, but for multiple target domains, the AdaDCLF strategy balances the differences in data distribution across target domains and focuses on learning the target domain with the largest deviation, thereby improving the performance of the overall model on multiple target domains. Providing a smaller γ value can make the training of the equipment fault diagnosis model more stable.

[0017] According to the domain confusion loss factor, the total loss function value is calculated as follows:

[0018]

[0019] The overall calculation process of the AdaDCLF strategy is as follows:

[0020] Step 1: Calculate the source domain data classification loss function value L C ;

[0021] Step 2: Calculate the domain confusion loss function value L for each target domain DC ;

[0022] Step 3: Calculate the domain confusion loss factor α at each iteration. This factor focuses on the target domain with the largest domain confusion loss, thereby improving the feature mapping capability on each target domain.

[0023] Step 4: Add the domain confusion loss function values of each target domain according to the domain confusion loss factor, and finally add the classification loss function value to obtain the total loss function value of the neural network;

[0024] Step 5: Iteratively optimize the equipment fault diagnosis model parameters according to the total loss function value, and finally obtain the final model.

[0025] The adaptive update strategy of the domain confusion loss factor optimizes the migration ability of the diagnostic model on multiple target domains and improves the robustness of the model. In fact, AdaDCLF is applicable to both single-target and multi-target domains. When the number of target domains is 1, the AdaDCLF strategy is consistent with the single-target domain adaptation method. The diagnostic model is iteratively optimized through training using the AdaDCLF method to obtain the final model.

[0026] 5) When diagnosing equipment faults, the target domain data is input into the final model to obtain the equipment fault diagnosis results.

[0027] As a preference, the step 1) is specifically as follows: dividing the sensor collected data into source domain data according to the equipment workload and target domain data Where S represents the source domain, T represents the target domain, i represents the i-th sample of the source domain or target domain, j represents the j-th target domain, and N S represents the number of source domain samples, represents the number of samples in the jth target domain; the data conditions of the source domain and the target domain are different.

[0028] Preferably, in step 2), the device fault diagnosis neural network specifically includes four convolutional layers, a pooling layer, an adaptive pooling layer, a fully connected hidden layer, and a Softmax layer. The convolutional layer is used to extract device data signal features, and each convolutional layer is followed by a batch normalization and LeakyRelu layer. Both the pooling layer and the adaptive pooling layer use a maximum pooling strategy for feature downsampling. The adaptive pooling layer is placed after the fourth convolutional layer and connected to the fully connected hidden layer, and finally output through the Softmax layer.

[0029] Preferably, in step 3), the classification loss is calculated as follows:

[0030]

[0031] Where S represents the source domain, i represents the i-th sample in the source domain, and N S represents the number of source domain samples, J uses the cross entropy loss function, represents the data of the i-th sample in the source domain, Represents the data label of the i-th sample in the source domain. The domain confusion loss between the source domain and the j-th target domain is calculated as follows:

[0032]

[0033] The beneficial effects of the present invention are as follows: the present invention proposes an AdaDCLF method for adaptively adjusting domain confusion loss factors for multiple target domains, measures the distribution deviation between the source domain and the target domain through MK-MMD, and then adaptively adjusts the domain confusion loss factor, focusing on the importance of the target domain data corresponding to the largest domain confusion loss. The total domain confusion loss is calculated according to the domain confusion loss factor, and then added to the source domain classification loss to obtain the total loss function value. The model parameters are iteratively updated and optimized through back propagation. Based on the powerful learning ability of the deep learning model, the domain invariance representation of the source domain and the target domain is learned. The AdaDCLF method for adaptively adjusting the domain confusion loss factor enhances the generalization ability of the model in multiple target domains, significantly improves the diagnostic accuracy of the equipment fault diagnosis model in multiple target domains, and solves the problem that the single-target domain adaptation algorithm has weak generalization ability in multiple target domains, thereby realizing intelligent diagnosis of equipment faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of the method of the present invention;

[0035] Figure 2 This is the overall structure diagram of the neural network of the present invention;

[0036] Figure 3 This is the overall flow chart of the AdaDCLF algorithm of the present invention;

[0037] Figure 4 Diagnosis result diagram. DETAILED DESCRIPTION

[0038] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] like Figure 1 As shown, the present invention includes the following steps.

[0040] 1) Divide the data into source domain and target domain data according to different working conditions;

[0041] 2) Construct a neural network model for equipment fault diagnosis and initialize its parameters;

[0042] 3) Input the source domain and target domain data into the neural network and calculate the classification loss and domain confusion loss respectively;

[0043] 4) Calculate the domain confusion loss factor through AdaDCLF, add the classification loss and domain confusion loss according to the domain confusion loss factor to obtain the total loss function value, and then perform iterative training to update the model parameters to obtain the final model;

[0044] 5) When diagnosing equipment faults, the target domain data is input into the final model to obtain the equipment fault diagnosis results.

[0045] The step 1) divides the sensor collected data into source domain data according to the device workload and target domain data Where S represents the source domain, T represents the target domain, i represents the i-th sample of the source domain or target domain, j represents the j-th target domain, and N S represents the number of source domain samples, represents the number of samples in the jth target domain. The number of target domains is set to 3, and the data conditions of the source domain and target domain are different.

[0046] The overall structure diagram of the equipment fault diagnosis model in step 2) is as follows Figure 2 As shown, the specific construction steps are as follows:

[0047] The device fault diagnosis neural network specifically consists of four convolutional layers, one pooling layer, one adaptive pooling layer, one fully connected hidden layer, and one softmax layer. The convolutional layers are used to extract device data signal features. Each convolutional layer is followed by batch normalization and LeakyReLU layers. Both the pooling and adaptive pooling layers use a max pooling strategy for feature downsampling. The adaptive pooling layer is placed after the fourth convolutional layer and connected to the fully connected hidden layer. The output is then passed through the softmax layer.

[0048] The parameters of the equipment fault diagnosis neural network are initialized using the normal distribution random initialization method, and the parameters are updated using the Adam algorithm.

[0049] The classification loss in step 3) is calculated after the Softmax layer, while the domain confusion loss is calculated before the Softmax layer. The classification loss only includes the classification loss of the source domain data, because the labels of the target domain data are often not available in practical applications. Specifically, the classification loss target is optimized according to the following formula:

[0050]

[0051] Where S represents the source domain, i represents the i-th sample in the source domain, and N S represents the number of source domain samples, J uses the cross-entropy loss function, x represents the source domain data, and y represents the source domain data label. By optimizing this classification loss target, the equipment fault diagnosis model can better fit the complex nonlinear mapping from source domain diagnostic signals to fault types.

[0052] The domain confusion loss function value in step 3) represents the degree of confusion between the source and target domains after extracting common features. The greater the degree of confusion, the more aligned the source and target domains are after extracting common features. By optimizing the domain confusion loss, the neural network can learn domain-invariant representations of the source and target domains, thereby improving the diagnostic accuracy of the equipment fault diagnosis model in the target domain. The domain confusion loss is measured using the MMD (Maximum Mean Discrepancy) distance. MMD can measure the distance between two distributions in the reproducing kernel Hilbert space (RKHS). The MMD distance calculation formula is as follows:

[0053]

[0054] Among them, H k Use kernel k to represent RKHS. Usually, Gaussian kernel is used as kernel. φ(*) represents the mapping to RKHS. tr represents the trace of the matrix. M is the coefficient matrix. M is calculated as follows:

[0055]

[0056] The idea of MMD is based on samples from the source and target domains. By finding the mapping function φ in the sample space, and then finding the mean of the function values of samples with different distributions on φ, the difference between the two means can be obtained to obtain the mean distance between the source and target domain data distributions corresponding to φ. MMD is used as a test statistic to determine whether the two distributions are the same. If the MMD distance is small enough, the two distributions can be considered to be the same, otherwise they are considered to have a large deviation. However, for practical applications, the parameter selection of each kernel is crucial to the final performance of the feature mapping. In order to better select the Gaussian kernel parameters, MK-MMD (Multi-Kernel Maximum Mean Discrepancy) is used instead of MMD. MK-MMD uses a convex combination of m kernels to provide an effective estimate of the mapping. The kernel selection calculation method is as follows:

[0057]

[0058] Where u represents the u-th kernel, β u Represents the u-th kernel weight parameter, and d represents the number of kernels. MK-MMD can be used to calculate the mean distance between the feature maps of the source domain and the single target domain, and this distance is used as the loss function value of the single target domain. The calculation method is as follows:

[0059]

[0060] The model is iteratively optimized through MK-MMD, so that the model has domain-invariant representation capabilities for both the source and target domains.

[0061] The AdaDCLF in step 4) stands for Adaptive Domain Confusion Loss Factor. This method can adaptively adjust the domain confusion loss factor between the source domain and each target domain, thereby improving the performance of the equipment fault diagnosis model in multiple target domains. The total loss function value of the equipment fault diagnosis neural network is calculated as follows:

[0062] L=L C +λL DC (6)

[0063] Among them L C represents the source domain classification loss, L DC represents the total domain confusion loss, λ represents the total domain loss penalty factor, the total domain confusion loss is composed of the source domain and each target domain confusion loss, L DC The calculation method is as follows:

[0064]

[0065] Where m represents the number of target domains, j represents the jth target domain, α is the domain confusion loss factor, and α is updated using the AdaDCLF strategy. The update formula is as follows:

[0066]

[0067] Where n represents the number of iterations, and γ represents the updated learning rate. The AdaDCLF method can focus on improving the learning ability of the target domain corresponding to the maximum domain confusion loss. If the target domain with the largest deviation is focused on learning at each iteration, it can be assumed that the model will not perform poorly on all known target domains of equipment fault data. AdaDCLF may reduce the performance of the model in a specific target domain, but for multiple target domains, the AdaDCLF strategy balances the differences in data distribution across target domains and focuses on learning the target domain with the largest deviation, thereby improving the performance of the overall model in multiple target domains. In addition, providing a smaller γ value can make the training of the equipment fault diagnosis model more stable.

[0068] According to the domain confusion loss factor, the total loss function value is calculated as follows:

[0069]

[0070] To illustrate the AdaDCLF strategy more specifically and intuitively, the overall calculation process is as follows: Figure 3 As shown, the calculation steps are as follows:

[0071] Step 1: Calculate the source domain data classification loss function value L C ;

[0072] Step 2: Calculate the domain confusion loss function value L for each target domain according to formulas (2)-(5) DC ;

[0073] Step 3: Calculate the domain confusion loss factor α at each iteration according to formula (8). This factor focuses on the target domain with the largest domain confusion loss, thereby improving the feature mapping ability on each target domain.

[0074] Step 4: Add the domain confusion loss function values of each target domain according to the domain confusion loss factor, and finally add the classification loss function value to obtain the total loss function value of the neural network;

[0075] Step 5: Iteratively optimize the equipment fault diagnosis model parameters according to the total loss function value, and finally obtain the final model.

[0076] The adaptive update strategy for the domain confusion loss factor optimizes the diagnostic model's transferability across multiple target domains, improving its robustness. AdaDCLF is applicable to both single-target and multi-target domains. When the number of target domains is one, the AdaDCLF strategy is consistent with the single-target domain adaptation method. The AdaDCLF method is used to iteratively optimize the diagnostic model to obtain the final model.

[0077] In the step 5), when diagnosing equipment faults, the target domain data is input into the final model to obtain equipment fault diagnosis results.

[0078] Step 4) of the present invention proposes an AdaDCLF method for adaptively adjusting the domain confusion loss factor for multiple target domains. The distribution deviation between the source domain and the target domain is measured by MK-MMD, and then the domain confusion loss factor is adaptively adjusted, focusing on the importance of the target domain data corresponding to the maximum domain confusion loss. The total domain confusion loss is calculated based on the domain confusion loss factor, and then added to the source domain classification loss to obtain the total loss function value. The model parameters are iteratively updated and optimized through backpropagation. Based on the powerful learning ability of the deep learning model, the domain invariance representation of the source domain and the target domain is learned, which enhances the generalization ability of the model in multiple target domains, significantly improves the diagnostic accuracy of the equipment fault diagnosis model in multiple target domains, and solves the problem that the single-target domain adaptation algorithm has weak generalization ability in multiple target domains, thereby realizing intelligent diagnosis of equipment faults.

[0079] Specific embodiments of the present invention:

[0080] The original experimental dataset consists of 6,000 samples collected under four different operating conditions. The data is divided into a source domain and three target domains according to the different operating conditions. The source domain data contains 3,000 samples, and each target domain contains 1,000 samples. The source domain data is labeled, while the target domain data is unlabeled. The source domain data is labeled 0, and the three target domain data are labeled 1 to 3. The original training set and test set are randomly divided into a 4:1 ratio within the source domain data. The model is trained using the source domain data, and the AdaDCLF method is used to transfer and generalize the source and target domain data to multiple target domains. Finally, the target domain data is input into the final model to obtain the final fault diagnosis results.

[0081] The experimental environment of the present invention is: CPU is CoreTMi7-6700K@4.00GHz, GPU is GTX1080Ti, video memory is 11GB, the system is Windows 10 operating system, the deep learning framework is Pytorch1.8, and Python3.8 is used to implement the multi-target domain device fault diagnosis method test based on AdaDCLF.

[0082] In order to verify the effectiveness of the multi-target domain equipment fault diagnosis method based on AdaDCLF, this method is compared with the network-based domain adaptation and MK-MMD-based domain adaptation methods (without AdaDCLF strategy). The models are trained according to different methods, and the target domain data is input into the final model to obtain the final target domain diagnosis results. Figure 4 The horizontal axis represents different operating conditions, and the last column represents the mean accuracy. Although the multi-target domain equipment fault diagnosis results based on AdaDCLF may not be as accurate as the single-target domain adaptation method in individual target domains, the overall mean accuracy is the highest. This is because AdaDCLF comprehensively considers the MK-MMD distance of each target domain and focuses on target domains with large deviations, thereby improving network performance. This shows that the multi-target domain equipment fault diagnosis method based on AdaDCLF is an effective diagnostic method that can well generalize the equipment fault diagnosis model to multiple target domains.

Claims

1. A multi-objective domain device fault diagnosis method based on AdaDCLF, characterized by: The method specifically comprises the following steps: 1) Divide the data into source domain and target domain data according to different working conditions; 2) Construct a neural network model for equipment fault diagnosis and initialize its parameters; 3) Input the source domain and target domain data into the neural network and calculate the classification loss and domain confusion loss respectively; 4) Calculate the domain confusion loss factor through AdaDCLF, obtain the total domain confusion loss based on the domain confusion loss factor, add the classification loss and the total domain confusion loss to obtain the total loss function value, and then perform iterative training to update the model parameters to obtain the final model; The AdaDCLF stands for Adaptive domain confusion loss factor. This method adaptively adjusts the domain confusion loss factor between the source domain and each target domain, thereby improving the performance of the equipment fault diagnosis model in multiple target domains. The total loss function value of the equipment fault diagnosis neural network is calculated as follows: L=L C +λL DC (1) Among them L C represents the source domain classification loss, L DC represents the total domain confusion loss, λ represents the total domain loss penalty factor, the total domain confusion loss is composed of the source domain and each target domain confusion loss, L DC The calculation method is as follows: Where m represents the number of target domains, j represents the jth target domain, α is the domain confusion loss factor, and α is updated using the AdaDCLF strategy. The update formula is as follows: Where n represents the number of iterations, γ represents the update learning rate, represents the MK-MMD multi-core maximum mean difference distance between the source domain and the j-th target domain at the n-th iteration; According to the domain confusion loss factor, the total loss function value is calculated as follows: Among them, S represents the source domain, T represents the target domain, i represents the i-th sample of the source domain or target domain, j represents the j-th target domain, and N S represents the number of source domain samples, Represents the data of the i-th sample in the source domain, represents the data label of the i-th sample in the source domain; L C represents the classification loss, L DC represents the domain confusion loss; 5) When diagnosing equipment faults, the target domain data is input into the final model to obtain the equipment fault diagnosis results.

2. The multi-objective domain device fault diagnosis method based on AdaDCLF according to claim 1 is characterized in that: The step 1) is specifically as follows: the sensor collected data is divided into source domain data according to the equipment workload and target domain data represents the number of samples in the jth target domain; the data conditions of the source domain and the target domain are different.

3. The multi-objective domain device fault diagnosis method based on AdaDCLF according to claim 1 is characterized in that: In step 2), the device fault diagnosis neural network specifically includes four convolutional layers, a pooling layer, an adaptive pooling layer, a fully connected hidden layer, and a Softmax layer; the convolutional layer is used to extract device data signal features, and the convolutional layer is followed by a batch normalization and a LeakyRelu layer; the pooling layer and the adaptive pooling layer both use the maximum pooling strategy for feature downsampling; The adaptive pooling layer is placed after the fourth convolutional layer and connected to the fully connected hidden layer, and finally output through the Softmax layer.

4. The multi-objective domain device fault diagnosis method based on AdaDCLF according to claim 1 is characterized in that: In step 3), the classification loss is calculated as follows: Where S represents the source domain, i represents the i-th sample in the source domain, and N S represents the number of source domain samples, J uses the cross entropy loss function, Represents the data of the i-th sample in the source domain, represents the data label of the i-th sample in the source domain; the domain confusion loss between the source domain and the j-th target domain is calculated as follows: Among them, MKMMD() represents the multi-core maximum mean difference distance calculation.

Images