Partial set cross-domain fault diagnosis method based on multi-level weighted dynamic domain adversarial network
Through a multi-level weighted dynamic domain adversarial network, the problems of noise weight and unrelated source samples in cross-domain fault diagnosis are solved, and higher diagnostic accuracy and stability are achieved, which are suitable for fault diagnosis of mechanical equipment.
Patent Information
- Application Number
- CN202311339033.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-16
- Publication Date
- 2025-08-19
- Estimated Expiration
- 2043-10-16
AI Technical Summary
The existing partial set adaptive diagnostic methods are difficult to avoid noise weights in cross-domain fault diagnosis, resulting in unstable model performance and difficult to effectively filter unrelated source samples.
A multi-level weighted dynamic domain adversarial network is adopted to extract data features through feature extractors, combine multi-level weight computing strategies and global and local domain discriminators to implement source sample filtering in parallel, dynamically adjust the relative importance of global and local adversarial adaptation processes, and build a dynamic domain adversarial framework to maximize positive migration.
Effectively filtering of unrelated source samples enhances the model's correction performance on noise weights, and improves the accuracy and stability of cross-domain diagnostics, especially when the target domain data type is a subset of the source tag space.
Smart Images
Figure CN117312922B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of cross-domain fault diagnosis, and in particular to a partial set cross-domain fault diagnosis method based on a multi-level weighted dynamic domain adversarial network. Background Art
[0002] The convergence of digitalization, intelligence, industrialization, and informatization has become a key focus of manufacturing industry upgrades, transformations, and structural optimization. This is driving the continuous advancement of mechanical equipment toward intelligent, scalable, and multifunctional capabilities. To ensure the smooth operation of mechanical equipment, health monitoring and fault diagnosis of key components are particularly important. In the field of mechanical fault diagnosis, traditional diagnostic methods based on manual feature extraction are labor-intensive and costly, making them inadequate for the increasingly complex data volumes. Consequently, intelligent fault diagnosis methods capable of processing large amounts of monitoring data have been developed in recent years. Among various intelligent fault diagnosis methods, deep learning-based methods have attracted considerable attention due to their ability to automatically extract features from large amounts of data. However, most reliable deep learning models assume that training and test data follow the same distribution, which is not always true in industrial systems. Domain-adaptive transfer learning techniques offer a new tool to address this issue. Transfer learning techniques learn transferable diagnostic knowledge from a source domain and reuse it in another related target task, thereby enabling cross-domain fault diagnosis. However, in practice, it has been found that while existing domain-adaptive-based intelligent fault diagnosis methods hold promise for cross-domain diagnosis scenarios, they often assume symmetric source and target label spaces. Unfortunately, in industrial scenarios, it is often difficult to collect a target dataset that is completely identical to the source data in terms of all health status types. Given that fault samples accumulate in the source domain over the long-term monitoring process of equipment, the target label space is more likely to be a subset of the source label space.
[0003] In recent years, domain adaptation methods have been developed in partial set scenarios. A common approach is to estimate the label distribution of target samples through a classifier or predictor to obtain class-level or instance-level weights to evaluate the importance of source data. However, a single-level source data weighting strategy relies too much on the predictor's accurate estimation of the target pseudo-label distribution. Incorrect distribution predictions cause some shared categories to be assigned noise weights, making it difficult to maintain stable performance during the application of the above method. By integrating different weighting strategies, the diagnostic model's over-reliance on the pseudo-label distribution given by a single predictor can be reduced. However, the fusion of multi-level weights may weaken the correction effect of different weighting strategies on the noise weights of source data during training, causing the model to over-focus on irrelevant source classes. Therefore, in order to solve the above problems, it is necessary to explore a new partial set cross-domain fault diagnosis method. Summary of the Invention
[0004] 1. Technical problem to be solved by the invention
[0005] The present invention aims to provide a partial set cross-domain fault diagnosis method based on a multi-level weighted dynamic domain adversarial network, addressing the difficulty of existing partial set adaptive diagnosis methods in avoiding noise weights. This method integrates a multi-level weighting strategy to mitigate the impact of noise weights during the training process of existing methods, effectively filtering out irrelevant source samples, and achieving reliable partial set adaptive results.
[0006] 2. Technical solution
[0007] The technical solution of the present invention is:
[0008] The partial set cross-domain fault diagnosis method based on a multi-level weighted dynamic domain adversarial network of the present invention divides source domain data and target domain data, extracts feature expressions of the data through a feature extractor, and calculates multi-level weights of source samples using feature expressions and a multi-level weight calculation strategy; then, source sample filtering is implemented in parallel through global and local domain discriminators; finally, the relative importance of the global and local adversarial adaptive processes is dynamically adjusted through a dynamic domain adversarial factor to maximize the promotion of positive migration.
[0009] The present invention comprises the following steps:
[0010] Step 1: Obtain historical monitoring data to construct a source domain dataset, and collect samples of different states in the target scenario to construct a target domain dataset;
[0011] Step 2: Establish a multi-level weighted dynamic domain adversarial model framework, which includes a feature extractor, a classifier, an auxiliary predictor, a global domain discriminator, and a local domain discriminator;
[0012] Step 3: Input the divided source domain training data and target domain training data into the feature extractor to obtain the feature expression of the source domain and target domain samples;
[0013] Step 4: Input the feature expression of the target domain sample into the classifier, and input the feature expression of the source domain and target domain samples into the auxiliary predictor to obtain the class-level and instance-level weights of the source domain data respectively;
[0014] Step 5: Input the feature representations of the source and target domain samples into the domain adversarial framework, and implement the class-level weighted local adversarial adaptation and instance-level weighted adversarial adaptation processes in parallel to obtain the global domain discrimination loss value and the local domain discrimination loss value;
[0015] Step 6: Calculate the dynamic domain adversarial factor based on the loss values of global domain adversarial adaptation and local domain adversarial adaptation, and dynamically adjust the relative importance of the global domain discriminator and the local domain discriminator during training;
[0016] Step 7: Perform model training based on the above optimization objectives. After training, save the final feature extractor and classifier parameters for diagnosis of target domain test samples.
[0017] Step 8: Load the diagnostic model parameters, input the target domain test sample data into the feature extractor to obtain feature expression, and then input it into the classifier to achieve fault prediction.
[0018] Furthermore, the feature extractor includes four layers of convolution and average pooling layers, the classifier includes three layers of fully connected layers, the auxiliary predictor includes two layers of fully connected layers and one layer of LeakySoftmax activation layer, and the global domain discriminator and the local domain discriminator include three layers of fully connected layers.
[0019] Furthermore, step 4 specifically includes the following steps:
[0020] Step 4.1: Input the feature expression of the target domain data into the classifier G y Get a |C s |-dimensional predicted output value Based on the predicted output value Obtain the class-level weight w(x s,k );
[0021] Step 4.2: Input the feature expressions of source domain data and target domain data into the auxiliary predictor Further quantify the instance-level weight of source domain data
[0022] Furthermore, step 5 specifically includes the following steps:
[0023] Step 5.1: Input the feature expressions of source and target domain samples into the global domain discriminator to implement global domain adaptation, and embed instance-level weights based on the global domain discriminator to implement source domain sample filtering;
[0024] In step 5.2, the feature expressions of the source domain and target domain samples are input into the local domain discriminator to implement local domain adaptation, and the class-level weights are embedded in the local domain discriminator to achieve class-level filtering of source domain samples.
[0025] Furthermore, the dynamic resistance factor is calculated as follows:
[0026]
[0027] Where d A,g (D s ,D t )and Represents the A-distance of global and local distributions, corresponding to the loss of the weighted global domain discriminator and the weighted local domain discriminator loss
[0028] Furthermore, in step 6, the dynamic adversarial factor is recalculated after each iteration and the dynamic domain adversarial loss is embedded to adaptively adjust the relative importance of global and local adaptation. The calculation formula of the dynamic domain adversarial loss is as follows:
[0029]
[0030] 3. Beneficial effects
[0031] (1) Aiming at the partial set fault diagnosis scenario that is more in line with the actual industrial system, the present invention proposes a partial set cross-domain fault diagnosis method based on a multi-level weighted dynamic domain adversarial network, which effectively solves the partial set cross-domain diagnosis problem where the target domain data type is only a part of the source label space.
[0032] (2) This paper designs a parallel sample filtering scheme based on multi-level weights to filter outlier samples and avoid negative transfer. It reduces the dependence of sample transferability on the output of a single predictor and improves the performance of correcting noise weights during model training.
[0033] (3) Building on the above, we design a dynamic domain adversarial framework with a parallel sample filtering scheme to enhance the forward transfer of shared categories. This framework reduces the contribution of outliers from both global and local perspectives in a parallel manner and eliminates the marginal and conditional distribution differences between shared categories. The relative importance of global and local adaptation is dynamically adjusted to maximize the domain adaptation performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 It is a schematic diagram of a partial set of diagnostic scenarios solved by the present invention;
[0035] Figure 2 This is a schematic diagram of a partial set cross-domain fault diagnosis process based on a multi-level weighted dynamic domain adversarial network of the present invention;
[0036] Figure 3 This is a schematic diagram of the diagnostic framework based on a multi-level weighted dynamic domain adversarial network of the present invention;
[0037] Figure 4 This is a confusion matrix diagram of the fault diagnosis effect of the multi-level weighted dynamic domain adversarial network according to the present invention. DETAILED DESCRIPTION
[0038] In order to further understand the content of the present invention, the present invention is described in detail with reference to the accompanying drawings and embodiments.
[0039] Example 1
[0040] against Figure 1In the partial set fault diagnosis scenario shown, a fault detection model is established using a multi-level weighted dynamic domain adversarial network, under the condition that the target sample contains only a portion of the fault categories of the source samples. First, the source domain data and target domain data are divided, and the feature representation of the data is extracted by a feature extractor. Using the feature representation and a multi-level weight calculation strategy, the multi-level weights of the source samples are calculated to avoid the negative impact of irrelevant source samples on the model. Then, source sample filtering is implemented in parallel using global and local domain discriminators to improve the method's ability to correct noisy labels. Finally, the dynamic domain adversarial factor dynamically adjusts the relative importance of the global and local adversarial adaptive processes to maximize the promotion of positive transfer. The present invention designs a parallel source sample filtering scheme based on multi-level weights to fully filter irrelevant source data and enhance the performance of the model in correcting noise weights during training. Based on this, a dynamic domain adversarial framework with a multi-level source sample filtering scheme is constructed to maximize the positive transfer of shared classes. In the partial set migration fault diagnosis scenario where the target operating condition fault sample type is a subset of the source domain, this method can effectively filter out irrelevant source categories and reduce their disturbance to the diagnosis model. At the same time, it can promote the migration effect between partially shared categories and achieve higher cross-domain diagnosis accuracy.
[0041] This example uses the Case Western Reserve University bearing experimental dataset (CWRU) to illustrate the method provided by the present invention in detail. The detailed description of the dataset is shown in Table 1.
[0042] Table 1 Detailed description of the example data set
[0043]
[0044] Ten status types 0-9 under the same load are selected as source domain data, while the target domain data only contains some fault types from 0-9 under a certain load. The detailed diagnosis task description is shown in Table 2 below.
[0045] Table 2 Detailed description of diagnostic tasks in the embodiment
[0046]
[0047] See Figure 2 , the present invention comprises the following steps:
[0048] Step 1: In the CWRU dataset, data under one working condition is used as the source domain data, and data under another different working condition is used as the target domain data. The source domain data and target domain data are divided into a training dataset and a test dataset respectively.
[0049] Step 2. Establish a dynamic domain adversarial model framework based on multi-level weights, which includes a feature extractor, a classifier, an auxiliary predictor, a global domain discriminator, and a local domain discriminator; the feature extractor includes four layers of convolution and average pooling layers, the classifier includes three layers of fully connected layers, the auxiliary predictor includes two layers of fully connected layers and one layer of LeakySoftmax activation layer, and the global domain discriminator and the local domain discriminator include three layers of fully connected layers.
[0050] Step 3: Input the divided source domain training data and target domain training data into the feature extractor to obtain the feature expression of the source domain and target domain samples;
[0051] Step 4: Input the feature expression of the target domain sample into the classifier G y , the feature expressions of source and target domain samples are input into the auxiliary predictor To obtain the class-level and instance-level weights of the source domain data respectively; specifically, the following steps are included:
[0052] Step 4.1: Express the features from the target domain data Input classifier G y Get a |C s |-dimensional predicted output value Based on this, the class-level weight of the source domain data can be expressed as:
[0053]
[0054] in Represents the output vector The kth element of .
[0055] Step 4.2: Express the characteristics of source domain samples and the characteristic expression of the target domain samples Input auxiliary predictor To get instance-level weights first, Through the LeakySoftmax activation function Convert the sample feature expressions of the source domain and the target domain into |C s |A vector of dimension:
[0056]
[0057] Among them, f represents the characteristic expression of the input sample, f c is the cth element of f, The sum of the output elements of is less than 1. It can accurately identify source domain samples with high confidence. The sum of the output values is close to 1. For the target sample, The sum of the output values is close to 0. Based on this, we define The output is |C s |Dimensionally The sum of the elements of The smaller the value of , the smaller the sample x i The higher the probability of belonging to the target class, the higher the instance weight should be assigned. Therefore, the instance-level weight is calculated as
[0058]
[0059] Step 5: Input the feature expressions of the source and target domain samples into the domain adversarial framework, and implement the class-level weighted local adversarial adaptation and instance-level weighted adversarial adaptation processes in parallel to obtain the global domain discrimination loss value and the local domain discrimination loss value. This specifically includes the following steps:
[0060] Step 5.1: First, the feature representations of the source and target domain samples are input into the global domain discriminator to implement global domain adaptation. Considering that the source domain data contains irrelevant categories, instance-level weights are embedded on the basis of the global domain discriminator to implement source domain sample filtering. The loss function of the weighted global domain discriminator is as follows:
[0061]
[0062] In step 5.2, the feature representations of the source and target domain samples are input into the local domain discriminator to implement local domain adaptation. Considering the inequality between the source and target domains, class-level weights are embedded in the local domain discriminator to implement class-level source domain sample filtering, effectively reducing the importance of irrelevant source classes and enhancing the model's ability to correct noise weights. The loss of the weighted local domain discriminator is as follows:
[0063]
[0064] Where w(x s,k ) represents the class-level weight of the k-th source class, Represents the probability that the sample belongs to the kth class.
[0065] Step 6: Calculate the dynamic domain adversarial factor based on the loss values of the global domain adversarial adaptation and the local domain adversarial adaptation, and dynamically adjust the relative importance of the global domain discriminator and the local domain discriminator during training. Specifically, the following steps are included:
[0066] A dynamic adversarial factor is calculated to measure the relative importance of global and local distributions. This factor is continuously learned and updated during the training process, enabling the network to adapt to changes in distribution differences. It is calculated as follows:
[0067]
[0068] Where d A,g (D s ,D t )and represent the A-distance of the global and local distributions, respectively. In the described diagnostic method, they are calculated as and The dynamic adversarial factor is embedded in the domain adversarial loss of the global and local domain discriminators to dynamically adjust the relative importance of global and local adaptation. The dynamic domain adversarial loss can be defined as:
[0069]
[0070] The dynamic adversarial factor is recalculated after each iteration and embedded into the dynamic domain adversarial loss to adaptively adjust the relative importance of global and local adaptation.
[0071] Step 7: Perform model training based on the above optimization objectives. After training, save the final feature extractor and classifier parameters for diagnosis of target domain test samples.
[0072] Step 8: Load the diagnostic model parameters, input the data of the target domain test samples into the feature extractor to obtain feature expressions, and then input them into the classifier to achieve fault prediction.
[0073] Table 3 shows the comparative results of the diagnostic accuracy of the method of the present invention and intelligent recognition methods such as deep convolutional neural network, domain adaptive neural network and partial domain adversarial adaptive network after five independent runs of the embodiment of the present invention. Specifically analyzing the diagnostic results of the method of the present invention, the average diagnostic accuracy of the method of the present invention is the highest, and the overall classification accuracy is 99.61%. Specifically, the multi-classification confusion matrix on the C8 task is as follows Figure 4 As shown in Table 3, in the diagnosis task where the source domain and target domain categories are not equal and the target domain categories are only a partial set of the source domain categories, the proposed method can effectively improve the domain adaptation effect and accurately distinguish different fault states of the bearing.
[0074] Table 3 Comparison of diagnostic results of the embodiments (%)
[0075]
[0076] The above is a schematic description of the present invention and its embodiments, which is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs a structure and embodiment similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.
Claims
1. A partial set cross-domain fault diagnosis method based on a multi-level weighted dynamic domain adversarial network, characterized in that: The following steps are involved: Step 1: Obtain historical monitoring data to construct a source domain dataset, and collect samples of different states in the target scenario to construct a target domain dataset; Step 2: Establish a multi-level weighted dynamic domain adversarial model framework, which includes a feature extractor, a classifier, an auxiliary predictor, a global domain discriminator, and a local domain discriminator; Step 3: Input the divided source domain training data and target domain training data into the feature extractor to obtain the feature expression of the source domain and target domain samples; Step 4: Input the feature expression of the target domain sample into the classifier G y , the feature expressions of source and target domain samples are input into the auxiliary predictor To obtain the class-level and instance-level weights of the source domain data respectively; It includes the following sub-steps: Step 4.1: Express the characteristics of the target domain data Input classifier G y Get a |C s |-dimensional predicted output value Based on this, the class-level weight of the source domain data can be expressed as: in Represents the output vector The kth element of ; Step 4.2: Express the characteristics of source domain samples and the characteristic expression of the target domain samples Input auxiliary predictor To get instance-level weights first, Through the LeakySoftmax activation function Convert the sample feature expressions of the source domain and the target domain into |C s |A vector of dimension: Among them, f represents the characteristic expression of the input sample, f c is the cth element of f, The sum of the output elements of is less than 1; the trained It can accurately identify source domain samples with high confidence. The sum of the output values is close to 1; For the target sample, The sum of the output values is close to 0; based on this, we define The output is |C s |Dimensionally The sum of the elements of The smaller the value of , the smaller the sample x i The higher the probability of belonging to the target class, the higher the instance weight should be assigned; therefore, the instance-level weight is calculated as: Step 5: Input the feature representations of the source and target domain samples into the domain adversarial framework, and implement the class-level weighted local adversarial adaptation and instance-level weighted adversarial adaptation processes in parallel to obtain the global domain discrimination loss value and the local domain discrimination loss value; Step 6: Calculate the dynamic domain adversarial factor based on the loss values of global domain adversarial adaptation and local domain adversarial adaptation, and dynamically adjust the relative importance of the global domain discriminator and the local domain discriminator during training; Step 7: Perform model training based on the above optimization objectives. After training, save the final feature extractor and classifier parameters for diagnosis of target domain test samples. Step 8: Load the diagnostic model parameters, input the target domain test sample data into the feature extractor to obtain feature expression, and then input it into the classifier to achieve fault prediction.
2. The partial set cross-domain fault diagnosis method based on a multi-level weighted dynamic domain adversarial network according to claim 1 is characterized by: The feature extractor includes four layers of convolution and average pooling layers, the classifier includes three layers of fully connected layers, the auxiliary predictor includes two layers of fully connected layers and one layer of LeakySoftmax activation layer, and the global domain discriminator and the local domain discriminator include three layers of fully connected layers.
3. The partial set cross-domain fault diagnosis method based on a multi-level weighted dynamic domain adversarial network according to claim 2 is characterized by: Step 5 specifically includes the following steps: Step 5.1: Input the feature expressions of source and target domain samples into the global domain discriminator to implement global domain adaptation, and embed instance-level weights based on the global domain discriminator to implement source domain sample filtering; In step 5.2, the feature expressions of the source domain and target domain samples are input into the local domain discriminator to implement local domain adaptation, and the class-level weights are embedded in the local domain discriminator to achieve class-level filtering of source domain samples.
4. The method for partial set cross-domain fault diagnosis based on a multi-level weighted dynamic domain adversarial network according to claim 3 is characterized by: The calculation formula of dynamic resistance factor is as follows: Where d A,g (D s ,D t )and Represents the A-distance of global and local distributions, corresponding to the loss of the weighted global domain discriminator and the weighted local domain discriminator loss 5. The method for partial set cross-domain fault diagnosis based on a multi-level weighted dynamic domain adversarial network according to claim 4 is characterized in that: In step 6, the dynamic adversarial factor is recalculated after each iteration and the dynamic domain adversarial loss is embedded to adaptively adjust the relative importance of global and local adaptation. The calculation formula of the dynamic domain adversarial loss is as follows: