Multi-source adversarial domain adaptive fault identification method based on multi-classifier collaborative alignment

Abstract

The invention belongs to the technical field of fault diagnosis, and particularly relates to a multi-source adversarial domain adaptive fault recognition method based on multi-classifier collaborative alignment. According to the method, the domain invariant feature of the sample can be extracted after the data is processed, so that cross-domain fault diagnosis is realized. The method comprises the following steps: firstly, constructing a sub-network for each source domain, learning domain invariant features by adopting domain adversarial training, and forming a fault type feature set by the domain invariant features of different fault types of each source domain; aiming at the target domain features and various fault feature sets in the source domain, measuring and calculating Warisstein distribution differences one by one; then, entropy minimization constraint is carried out on the generated distribution distance vector, and the sub-network is driven to directionally mine fault characterization with a higher identification degree; alignment of multiple classifiers is realized by using an incidence matrix, and a more stable decision boundary is trained. According to the method, the domain invariant features are extracted through the domain adversarial training, the representative features of the fault are learned, and finally the output of the classifier is aligned by using the incidence matrix, so that the success rate of fault diagnosis is improved.

Description

Technical Field

[0001] This invention belongs to the field of axial flow fan fault diagnosis, specifically involving a multi-source adversarial domain adaptive fault identification method based on multi-classifier collaborative alignment. Background Technology

[0002] With the continuous development of modern industry, axial flow fans have become an indispensable piece of machinery in many factories. Axial flow fan equipment is also becoming increasingly complex and systematized, and the property losses caused by fan malfunctions should not be underestimated. Therefore, timely detection and maintenance of abnormal fan conditions are of great significance for ensuring stable equipment operation, reducing maintenance costs, and extending the service life of the fans.

[0003] Currently, fault identification for axial-flow rotating machinery mainly relies on qualitative analysis based on the prior knowledge of maintenance personnel. However, this traditional method is limited by individual differences in subjective experience, which can easily lead to misjudgments or omissions of core faults. Current research on neural network-based fault diagnosis methods focuses on single-source-domain adaptation. However, in reality, machines typically operate under various conditions, such as different speeds or loads. The collected source-labeled data represents the machine's operation under multiple different conditions, and models trained from a single source-domain data cannot adapt to the complex and ever-changing actual working environment. Summary of the Invention

[0004] To address the aforementioned technical challenges, this application constructs a multi-source adversarial domain adaptive fault identification method based on multi-classifier collaborative alignment. It leverages a domain adversarial training mechanism to capture common features (domain-invariant features) between domains and applies entropy minimization constraints to the distribution distance vector between the target features and the feature sets of each fault mode in the source domain, thereby driving the network to converge to a feature space with higher fault identification accuracy. To achieve multi-source domain adaptation when the operating conditions are large, the proposed method uses an association matrix to align the outputs of the target samples in specific domain classifiers of multiple sub-networks, learning a more stable decision boundary.

[0005] To address the aforementioned technical problems, this invention proposes a multi-source adversarial domain adaptive fault identification method based on multi-classifier collaborative alignment, comprising: Sensor devices are installed on the axial flow fan to obtain the high-frequency raw data values ​​of the vibration data; The obtained industrial data is cleaned using the interpolation regression data cleaning method disclosed in this invention, based on the actual situation. After mapping the data to the same feature space, the adversarial fault feature set generation method disclosed in this invention is used to extract key fault features and generate feature sets for various faults. The Wasserstein distance method is used to calculate the feature distribution distance between each fault type feature set and a single sample in the target domain. Minimizing the entropy of the distribution distance vector enables the subnetwork to learn more fault-representative features. The correlation matrix metric method disclosed in this invention is used to measure the differences in prediction results of multiple specific fault classifiers for the same target sample, and the output of the target sample in each specific fault classifier is aligned according to the loss function to learn a more stable decision boundary.

[0006] Optionally, in step 1, a high-frequency sensor of model 6021 from the brand ARIZON is used to collect raw data in the vertical direction at the bearing on the motor side of each axial flow fan. The machine speed is 960 r / min and the sampling frequency is 25.6KHz to obtain the high-frequency raw data value of the vibration data.

[0007] Optionally, the interpolation regression data cleaning method involves cleaning the obtained industrial data, specifically as follows: Step 1: For the obtained raw industrial data of the axial flow fan, first calculate the estimated value of the missing data using spline interpolation based on the actual industrial conditions and the data near the missing data, and then fill in the missing data value with the obtained estimated value. Step 2: Based on the actual industrial conditions, fit all the data into a function curve using multiple linear regression. Data points whose error value reaches the error threshold of this curve are considered outliers, and the fitted values ​​are used to replace these outliers, ultimately obtaining a lossless dataset. After completing the above processing, organize the data into training, validation, and test sets to facilitate subsequent model training, calibration, and evaluation steps.

[0008] Optionally, the step involves extracting key fault features and generating feature sets for various fault types using an adversarial fault feature set generation method, specifically: Step 1: Map all cleaned source and target domain data to the same feature space using a shared feature extractor; Step 2: Train a private feature extractor and a specific fault classifier using source domain samples, so that they have the ability to classify faults and extract features, respectively; Step 3: Use domain adversarial learning to train the feature extractor so that it can extract domain-invariant features from the source and target domains; Step 4: Utilize the source domain features learned by each sub-network to form a feature set for each fault type. During the collaborative iteration of the network model and the domain adversarial strategy, the information within the fault type feature set will also be adaptively updated and improved.

[0009] Optionally, the step involves using the minimization of the Wasserstein distribution distance based on features from the source and target domains to learn more fault-representative features, specifically: Step 1: Use Wasserstein distance to measure the degree of difference between individual target domain samples and the characteristics of different fault types in the source domain; Step 2: For a given target sample, calculate the entropy of the distance vector between its features and the feature sets of each fault type in the source domain. Minimizing this entropy allows the network to learn features that are more representative of the faults. Step 3: To prevent a situation where the entropy value is small, but the maximum value in the distribution distance vector cannot match the maximum value of the classifier's prediction vector, this invention assigns weights to the target samples.

[0010] Optionally, the method of minimizing the distributional differences of multiple classifier outputs using the association matrix metric is specifically as follows: Step 1: Use the correlation matrix metric to measure the difference between the output vectors of a target sample and those of multiple specific fault classifiers; Step 2: Minimize the differences between the output vectors of multiple specific fault classifiers to align the outputs of each sub-network's specific fault classifiers and learn a more stable decision boundary. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart illustrating the wind turbine fault analysis method provided in this application embodiment; Figure 2 This is a schematic diagram of the network framework of the wind turbine fault analysis method provided in the embodiments of the present invention; Figure 3 This is a schematic diagram illustrating domain adversarial learning based on the common features of the source and target domains, as provided in an embodiment of the present invention.

[0013] Figure 4 This is a structural block diagram of the wind turbine fault analysis device provided in an embodiment of the present invention. Detailed Implementation

[0014] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] like Figure 1 As shown, Figure 1 This is a flowchart illustrating the wind turbine fault analysis method provided in an embodiment of this application. The wind turbine fault analysis method may include:

[0016] S11: Collect the high-frequency raw data of axial flow fan vibration required by this invention.

[0017] It should be noted that the experimental data of this invention were all collected using a high-frequency sensor (model 6021) from the ARIZON brand on the axial flow fan drive motor. The collected data were placed in different folders according to experimental requirements. The specific steps are as follows:

[0018] Step 1: Install a high-frequency sensor at the bearing on the motor side of the axial flow fan;

[0019] Step 2: Set all axial flow fans to operate at a certain speed n, and set different loads on the axial flow fans to collect different source domain data;

[0020] Step 3: Set the sampling frequency m, and in each source domain, collect raw vibration data for the five states of the axial flow fan, including normal, rotor imbalance, rotor misalignment, reducer failure and sensor failure, according to different working states and save them separately.

[0021] Step 4: Discretize the original data into several short sequences based on a sliding window mechanism with a width of a, and resample the data according to a step size of b points.

[0022] Using the above method, sensors are installed on the bearing of the axial flow fan and different loads are set to simulate data from different source domains. The required data is collected according to different working states and saved separately as required. Finally, the raw data is divided into smaller sequences.

[0023] Based on the above discussion, in an optional embodiment of this application, the process for acquiring the required high-frequency raw vibration data may include: In an optional embodiment of S11, the axial fan speed is set to n=960r / min, the sensor sampling frequency is set to m=25.6KHz, a sliding window with a width of a=2048 is used, and the data is resampled every b=200 points.

[0024] S12: Clean the data using the interpolation regression data cleaning method disclosed in this invention, based on the actual situation.

[0025] It should be noted that the industrial vibration data after resampling may contain some omissions or errors. Therefore, this invention employs an interpolation regression data cleaning method for the resampling industrial data. The specific steps include:

[0026] Step 1: For the resampled industrial data, first calculate the estimated value of the missing value using cubic spline interpolation based on the actual industrial conditions and the m data points to the left and right of the missing data. Then, use the calculated estimated value to fill in the missing data value. The basic principle of spline interpolation is as follows: Suppose there are n+1 known data points on the interval [a,b], which divide [a,b] into n subintervals [x0,x1], [x1,x2], ..., [xn−1,xn]. A cubic interpolation polynomial is inserted between every two points. Based on the following conditions: the second derivative is a linear function; the function values ​​at the interpolation points are known; the first and second derivatives are continuous; and the boundary conditions are met, the coefficients of each polynomial can be solved to obtain the final expression.

[0027] Step 2: Based on standard industrial conditions, fit all data into a function curve using multiple linear regression. When the error between the industrial data point and the fitted data point at a certain moment reaches the error threshold n, the industrial data at that point is considered outlier, and the fitted data is used to replace the outlier, thus obtaining a lossless dataset.

[0028] It should be noted that the error threshold is obtained by the golden section method. Set the initial interval [a, b], and calculate the fault diagnosis accuracy corresponding to the endpoint values ​​of different intervals according to the steps of the golden section method. Continuously narrow the interval until the interval size is reduced to z. At this time, the average value of the left and right endpoints of the corresponding interval is the error threshold.

[0029] Step 3: After completing the above processing, organize the data into training set, validation set and test set to facilitate subsequent model training, calibration and evaluation steps.

[0030] Based on the above discussion, in an optional embodiment of this application, the data cleaning process using the interpolation regression data cleaning method may include: In an optional embodiment of S12, missing data is supplemented using cubic spline interpolation, taking m=25 data points to the left and right of the missing data for interpolation to obtain the missing values. When fitting the data using multiple linear regression, the initial interval of the golden section method is set to [0%, 10%], z=0.01%, and an error threshold is obtained. Finally, the fitted data is used to replace outliers whose errors exceed the error threshold. When performing multiple linear regression, the function degree is generally set to 4. The original industrial data processed by the interpolation regression data cleaning method is a lossless dataset.

[0031] S13: Divide the processed fault data into different source domains, extract key fault features using an adversarial fault feature set generation method, and generate feature sets for various types of faults.

[0032] It should be noted that since the source and target domain data we collected were acquired under different loads on the axial flow fan, the method used to extract fault features in the source domain may not be able to accurately extract features in the target domain. Therefore, we utilize an adversarial approach and employ an adversarial fault feature set generation method to extract key faults. The specific steps are as follows:

[0033] Step 1: Divide the processed fault data into different source domains according to the collection conditions;

[0034] Step 2: Map all cleaned source and target domain data to the same feature space using a shared feature extractor;

[0035] Step 3: Train a private feature extractor and a specific fault classifier using source domain samples, enabling each sub-network to classify faults and extract features. The basic idea is to backpropagate the classification loss from the source domain samples to update the parameters of the private feature extractor and the specific fault classifier until they have the ability to classify faults and extract features.

[0036] The source domain sample classification loss for each sub-network is:

[0037] k∈1,2, …,K, K is the number of source domains, E represents the cross-entropy loss, and C represents the cross-entropy loss. k A specific fault classifier for the k-th source domain, and an F-feature extractor.

[0038] Step 4: The proposed method uses domain adversarial learning to train the feature extractor, and utilizes a specific optimization algorithm to minimize the optimization objective until the maximum number of iterations M, enabling it to extract domain-invariant features from both the source and target domains. The domain adversarial process is as follows: Figure 3 As shown. The basic principle of domain adversarial learning is as follows: Feature extractor F needs to extract high-level features from source and target domain samples that can confuse the domain discriminator, while discriminator D needs to distinguish whether features are extracted from the source or target domain. To ensure the network extracts domain-invariant features, the loss of the domain discriminator must be minimized. Parameters of the training domain discriminator maximizing the loss of the domain discriminator To update the parameters of the feature extractor When the domain discriminator cannot effectively identify the domain label (source domain / target domain) of the input features, the extracted features are determined to be domain-invariant features. Defining the source domain label as 0 and the target domain label as 1, the domain classification loss of each specific domain discriminator is:

[0039] Where k∈1,2,…,K, K is the number of domain-specific discriminators, D represents the number of samples in the k-th source domain and the number of samples in the target domain, and D k k-th specific domain discriminator.

[0040] Step 5: Utilize the source domain features learned by each sub-network to form feature sets for each fault type. Based on domain adversarial mechanisms and iterative network training, dynamically adjust the feature distribution information within the fault type feature sets.

[0041] Using the above method, a private feature extractor and a specific fault classifier are trained using source domain data, enabling them to perform feature extraction and fault classification in the source domain, respectively. Then, using the idea of ​​domain adversarial analysis, domain-invariant features are extracted, and the source domain features learned by each sub-network are used to form feature sets for each fault type.

[0042] Based on the above discussion, in an optional embodiment of this application, the process of extracting key fault features and generating feature sets for various faults using the adversarial fault feature set generation method described above can be... include: In an optional embodiment of S13, for the domain adversarial network, the network architecture of its feature extractor is: 64Conv3(4*1), BN3(64), MP3(2*1), 64Conv4(4*1), BN4(64), MP4(2*1), 64Conv5(4*1), BN5(64), MP5(2*1); the network architecture of the fault classifier is: FC1(832), FC2(256), FC3(c); the network architecture of the specific domain discriminator is: FC4(832), FC5(128), FC6(64), FC7(32), FC8(16), FC9(32), FC10(64), FC11(128), FC12(832).

[0043] S14: Calculate the Wasserstein distance between the target feature set and the fault feature set, and use the entropy minimization strategy to process the distribution distance vector to determine the weight of the target sample.

[0044] It should be noted that after extracting key faults through the above steps, the feature extractor can extract domain-invariant features, but the differences between features of different fault types are still not significant enough. Therefore, this invention uses the Wasserstein distance method to calculate the feature distribution distance between each fault type feature set and a single sample in the target domain. Then, it minimizes the entropy of the distribution distance vector to enable the sub-network to learn more fault-representative features. Simultaneously, it calculates the matching value between the distribution distance vector and the classifier prediction result vector to assign weights to the target samples. Specifically, this can be divided into the following steps:

[0045] Step 1: Use the Wasserstein distribution distance to measure the degree of difference between a single target domain sample and the characteristics of different fault types in the source domain. The calculation method is as follows: For a given target sample, the distribution distance between its features and the feature sets of each fault type in the source domain is:

[0046] Where j (∈ 1,2, …, c), the number of samples of each fault type in the source domain in each batch is m. p C represents the number of fault categories, W(∙) represents the Wasserstein distance metric, and M j,i This represents the feature of the i-th sample in the feature set of the j-th fault type in the source domain of a batch. Let t be the t-th sample in the target domain.

[0047] Then, the distribution distance obtained for each fault type in the source domain is normalized to obtain the normalized distribution distance:

[0048] Step 2: Based on the normalized distance described above, obtain the vector of distance between the target feature and the feature sets of each fault type in the source domain:

[0049] Step 3: Obtain the vector The entropy is minimized, reducing the distribution distance between the target sample and the source domain feature set of the same fault type features, and increasing the distribution distance of different fault type features, so that the network can learn more representative fault features.

[0050] For a given target sample, the entropy of the distance vector between its features and the feature sets of each fault type in the source domain is:

[0051] Step 4: Assign weights to the target samples to prevent the maximum value in the distribution distance vector from not matching the maximum value in the classifier's prediction vector.

[0052] Let the predicted vector of a target sample in the k-th fault classifier be denoted as 𝑃 t,k,i Then vector 𝑉 𝐷𝑖𝑠 The positional matching value between the minimum value and the maximum value of the vector t,k,i:

[0053] In this context, 𝑎𝑟𝑔𝑚𝑖𝑛𝑛𝑉𝐷𝑖𝑠, 𝑛 represents 𝑉 𝐷𝑖𝑠 The position of the minimum value of the element is 𝑎𝑟𝑔𝑚𝑎𝑥𝑛𝑃 𝑡,𝑘,𝑖,𝑛 Indicates 𝑃 𝑡,𝑘,𝑖,𝑛 The position of the maximum value of the elements in the middle.

[0054] The weight of the target sample is then calculated as follows:

[0055] Step 5: Calculate the entropy loss of the feature set distribution distance vector for all target samples in different sub-networks:

[0056] Where k∈1,2,…,K, K is the number of source domains, and n is the number of source domains. t This represents the number of samples in the target domain.

[0057] Using the above method, the feature distribution distance between each fault type feature set and a single sample in the target domain is calculated. Then, the entropy of the distribution distance vector is minimized so that the sub-network learns features that are more representative of the fault. At the same time, the matching value between the distribution distance vector and the classifier prediction result vector is calculated, and weights are assigned to the target samples.

[0058] Based on the above discussion, in an optional embodiment of this application, the process of calculating the feature distribution distance between the fault type feature set and the target domain samples using the Wasserstein distance method, then minimizing the entropy of the distribution distance vector, and finally assigning weights to the target samples may include: In an optional embodiment of S14, the batch size of the experiment is set to 75, the optimizer is Adam, the network learning rate is set to 0.001, and the training is performed for 200 epochs.

[0059] S15: Use the correlation matrix alignment method to align the outputs of the multi-classifier.

[0060] It should be noted that this application uses the correlation matrix alignment method to align the outputs of specific fault classifiers in each sub-network because the differences in operating conditions across source domains lead to variations in the output results of the same target sample across different source domain fault classifiers. Utilizing the correlation matrix to align the multi-classifier outputs mainly includes the following steps:

[0061] Step 1: Based on the output vectors P1, P2, ... P of the target sample in k specific fault classifiers k The correlation matrix A is obtained as follows:

[0062] Step 2: Calculate the target sample n t The differences between the output vectors of K specific fault classifiers:

[0063] Where A is a square matrix of size C×C, and C is the number of fault types. Let A a，b Let be a value in the a-th row and b-th column of matrix A.

[0064] Step 3: Calculate the loss for aligning the multi-classifiers:

[0065] Where m < n, a ≠ b, n t Let P be the number of target samples, K be the number of specific fault classifiers, and P be the number of samples. i,m Let represent the output vector of the i-th target sample in the m-th fault classifier, (·). T The transpose of a vector is [·]. a,b This represents the value in the a-th row and b-th column of the matrix.

[0066] Step 4: In the training process, apply the optimization algorithm to minimize the overall optimization index, and stop updating when the cumulative number of iterations reaches M.

[0067] Based on the above discussion, in an optional embodiment of this application, the output of the multi-classifier based on the above-mentioned correlation matrix alignment method may include: In an optional embodiment of S15, the Adam optimizer is used when optimizing the objective function, and the number of iterations M is set to 20.

[0068] refer to Figure 2 , Figure 2 This is a schematic diagram of the network framework for the wind turbine fault analysis method provided in an embodiment of the present invention.

[0069] refer to Figure 4 , Figure 4 A schematic diagram of the network framework for the wind turbine fault analysis method provided in the embodiments of this application.

[0070] In summary, this invention discloses a multi-source adversarial domain adaptive fault identification method based on multi-classifier collaborative alignment. It utilizes a domain adversarial training mechanism to capture common features across domains (domain-invariant features) and applies entropy minimization constraints to the distribution distance vectors between the target features and the feature sets of each fault mode in the source domain, thereby driving the network to converge to a feature space with higher fault identification accuracy. To achieve multi-source domain adaptation when the operating conditions are wide-ranging, the proposed method uses an association matrix to align the outputs of target samples in specific domain classifiers across multiple sub-networks, learning a more stable decision boundary.

Claims

1. A multi-source adversarial domain adaptive fault identification method based on multi-classifier collaborative alignment, characterized in that, Includes the following steps: Sensor devices are installed on the axial flow fan to obtain the high-frequency raw data values ​​of the vibration data; The obtained industrial data is cleaned using the interpolation regression data cleaning method disclosed in this invention, based on the actual situation. The cleaned data is divided into different source domains and target domains according to different working conditions. The shared feature extractor maps the data of all source domains and target domains to the same feature space. The adversarial fault feature set generation method disclosed in this invention is used to extract key fault features and generate feature sets of various faults. The Wasserstein distance metric is used to calculate the feature distribution distance between each fault type feature set and a single sample in the target domain. Minimizing the entropy of the distribution distance vector enables the subnetwork to learn more fault-representative features. The correlation matrix metric method disclosed in this invention is used to measure the differences in prediction results of multiple specific fault classifiers for the same target sample, and the output of the target sample in each specific fault classifier is aligned according to the loss function to learn a more stable decision boundary.

2. The multi-source adversarial domain adaptive fault identification method based on multi-classifier collaborative alignment according to claim 1, characterized in that... High-frequency raw data values ​​of vibration data are acquired through sensing devices installed on the axial flow fan, including: Under different operating conditions, high-frequency sensors were used to collect raw data in the vertical direction at a sampling frequency of 2.56KHz at the bearing side of the axial flow fan motor with a machine speed of 960 r / min. The collected data included five fault types: normal, rotor imbalance, rotor misalignment, reducer fault, and sensor fault.

3. The multi-source adversarial domain adaptive fault identification method based on multi-classifier collaborative alignment according to claim 1, characterized in that... The acquired data was processed using interpolation and regression data cleaning methods, including: This method addresses the characteristics of the raw industrial data obtained from axial flow fans. First, it uses cubic spline interpolation to calculate estimated values ​​for missing data based on actual industrial conditions and then fills in the missing data. Next, it removes outlier and invalid data values ​​and uses polynomial interpolation to calculate estimated values, replacing the outliers with these estimated values ​​to obtain a lossless dataset. This processing significantly reduces potential disturbances to subsequent analysis stages, thus comprehensively ensuring the accuracy and robustness of the data.

4. The multi-source adversarial domain adaptive fault identification method based on multi-classifier collaborative alignment according to claim 1, characterized in that... The adversarial-based fault feature set generation method disclosed in this invention is used to extract key fault features and generate feature sets for various faults, including: First, the preprocessed lossless dataset is mapped to the same feature space using a shared feature extractor. Then, the data in each domain is used to learn their common features through domain adversarial learning using a specific domain discriminator and a specific fault classifier. The features learned in each domain constitute feature clusters for various types of faults. As the domain adversarial learning process progresses and the network iterates repeatedly, the amount of information carried by the above feature clusters will be continuously updated and enhanced.

5. The multi-source adversarial domain adaptive fault identification method based on multi-classifier collaborative alignment according to claim 1, characterized in that... Based on the minimization of Wasserstein distribution distance using features from the source and target domains, fault features with higher discriminative power are extracted. Specifically, this involves: The Wasserstein distance metric is used to calculate the feature distribution distance between the feature set of each fault type and the single sample feature of the target domain. While deeply exploring the heterogeneity among the features of each fault mode, the weight coefficient of the target domain sample is determined by quantifying the consistency index between the distribution distance vector and the classifier prediction result vector.

6. The multi-source adversarial domain adaptive fault identification method based on multi-classifier collaborative alignment according to claim 1, characterized in that... Minimizing the distributional differences of multiple classifier outputs using the association matrix metric includes: Due to differences in the operating conditions of various source domains, the output results of the same target sample on different specific fault classifiers may vary. The correlation matrix metric is used to measure the differences between the output vectors of a target sample across multiple specific fault classifiers. Minimizing the differences between the output vectors of multiple specific fault classifiers aligns the outputs of each sub-network's specific fault classifier, leading to the learning of a more stable decision boundary.

Images