Mechanical equipment related measuring point transfer learning method based on multi-scale convolutional network

Abstract

The application discloses a kind of mechanical equipment associated measuring point transfer learning method based on multi-scale convolution network, first, the labeled source domain data is used as the input of the source domain of domain learning network to extract associated measuring point features, and network structure and weight are shared to the domain learning network of target domain, then the unlabeled target domain data is used as the input of the target domain of domain learning network to extract target measuring point features;The difference between the fields is measured using the maximum mean square difference method for associated measuring point features, target measuring point features, and combined with the classification loss of the domain learning network of source domain, the difference loss between fields to construct a joint loss function as optimization goal, iterative training using back propagation algorithm, update target domain learning model parameters, predicted sample class, finally obtain the target domain learning model with excellent diagnostic effect, realize the knowledge transfer between mechanical equipment associated measuring point, target measuring point.The present application effectively extracts deep features in data through multi-scale convolutional neural network, and realizes the knowledge transfer between different measuring points using the maximum mean square difference, so as to use associated measuring point for fault diagnosis in the case of inconvenient sensor installation, large data distribution difference and lack of label information, effectively improve the model generalization.

Description

Technical Field

[0001] This invention relates to a transfer learning method for associated measurement points of mechanical equipment based on multi-scale convolutional networks, belonging to the field of fault diagnosis of mechanical equipment. Background Technology

[0002] The continued growth of the manufacturing industry relies on the efficient operation of various mechanical equipment. However, in industrial settings, mechanical equipment operating under prolonged and high-load conditions is prone to deterioration and even malfunction if it lacks professional maintenance. This not only affects the lifespan of the equipment and production efficiency but can also lead to personal injury or death in severe cases.

[0003] In recent years, with the rapid development of artificial intelligence technology, especially the rise of deep learning, it has shown unprecedented potential in the field of mechanical equipment fault diagnosis. Deep learning, by constructing complex neural network models, can automatically extract hidden features and patterns from massive amounts of data, thereby providing accurate predictions and diagnoses for mechanical equipment faults. However, the complex layout and operating conditions of mechanical equipment in industrial settings often result in missing data for the required locations, ultimately leading to a lack of key data for deep learning models and hindering their application in fault diagnosis.

[0004] Therefore, how to effectively utilize deep transfer learning to extract diagnostic knowledge from associated measurement points of mechanical equipment and effectively transfer it to target measurement points has become a current research hotspot and challenge. Summary of the Invention

[0005] This invention provides a transfer learning method for associated measurement points of mechanical equipment based on multi-scale convolutional networks. By using an improved convolutional neural network for transfer learning, it successfully extracts equipment fault information and adapts the fault feature domain to different measurement points of the equipment. This realizes the transfer of diagnostic knowledge between associated measurement points and target measurement points, and solves the problems of inconvenient sensor installation and poor knowledge transfer diagnostic capabilities in mechanical equipment diagnosis.

[0006] The technical solution of this invention is:

[0007] A transfer learning method for mechanical equipment associated measurement points based on multi-scale convolutional networks includes: collecting associated measurement point data of equipment fault points as source domain data and collecting target point data to be diagnosed as target domain data according to the transfer task; normalizing both source and target domain data, then using a sliding window mechanism to segment the normalized data to obtain samples, and determining the corresponding category labels for the samples, ultimately obtaining the source domain dataset and the target domain dataset; wherein, the category labels include normal category labels and different fault category labels; constructing a domain learning network; and jointly training the source and target domains through a shared network structure: first, using the labeled source domain data as the domain learning network of the source domain. The input of the learning network is used to extract the features of the associated measurement points, and the network structure and weights are shared with the domain learning network of the target domain. Then, the unlabeled target domain data is used as the input of the domain learning network of the target domain to extract the features of the target measurement points. The maximum mean square difference method is used to measure the difference between the features of the associated measurement points and the features of the target measurement points. The joint loss function is constructed by combining the classification loss of the domain learning network of the source domain and the difference loss between the domains as the optimization objective. The backpropagation algorithm is used for iterative training to update the parameters of the target domain learning model and predict the sample category. Finally, a target domain learning model with excellent diagnostic performance is obtained, realizing the knowledge transfer between the associated measurement points and the target measurement points of mechanical equipment.

[0008] The sliding window mechanism method is as follows:

[0009]

[0010] In the formula, n represents the number of samples obtained by sliding window, l is the sample length, and s is the window step size.

[0011] The domain learning network is specifically designed as follows: First, it uses a multi-scale convolutional neural network to extract data features and introduces a self-attention module to enhance the representation ability of the features; then, it fuses feature information at different scales; finally, it introduces the maximum mean square difference method in the adaptive layer to measure the feature difference between the source domain and the target domain.

[0012] The multi-scale convolutional neural network includes two branches: the first branch uses a five-layer stacked structure of "convolutional layer + batch normalization layer + ReLU activation function + max pooling layer", and the kernel size of the first layer in the first branch is larger than the kernel size of the other layers in the first branch; the second branch uses a four-layer stacked structure of "convolutional layer + batch normalization layer + ReLU activation function + max pooling layer", and the kernel size of the first layer in the second branch is larger than the kernel size of the other layers in the second branch.

[0013] The five convolutional layers in the first branch use kernel sizes of 64, 3, 3, 3, and 3, respectively; the four convolutional layers in the second branch use kernel sizes of 32, 3, 3, and 3, respectively.

[0014] The joint loss function is expressed as follows:

[0015] loss = clc_Loss + mmsd * λ;

[0016] Where clc_Loss represents the cross-entropy loss function for source domain training, λ is the adjustment coefficient, and mmsd represents the inter-domain difference loss obtained based on the maximum mean square difference.

[0017] The beneficial effects of this invention are as follows: The mechanical correlation measurement point transfer learning method proposed in this invention, based on a multi-scale convolutional neural network, has significant innovative value and practicality, providing a new method and approach for mechanical equipment fault diagnosis. This model effectively extracts deep features from the data through a multi-scale convolutional neural network and utilizes maximum mean square difference to achieve knowledge transfer between different measurement points. Therefore, even in situations where sensor installation is inconvenient, data distribution varies significantly, and label information is lacking, fault diagnosis can be performed using correlated measurement points, effectively improving the model's generalization ability. Attached Figure Description

[0018] Figure 1 This is a diagram of the transfer learning framework for related measurement points in this invention;

[0019] Figure 2 The training accuracy for two different transfer tasks;

[0020] Figure 3 A scatter plot for t_SNE visualization of the target domain. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0022] In industrial settings, signal acquisition of mechanical equipment fault locations often faces challenges such as limited sensor installation locations and high acquisition costs. While a large amount of data from associated measurement points related to the fault point can be obtained, data on the actual fault location is often scarce, leading to poor fault diagnosis. Furthermore, in industrial settings, it is impossible to establish diagnostic models for all measurement points of mechanical equipment. Therefore, it is necessary to use a fault diagnosis model for one measurement point to identify faults at another. The inherent signal coupling phenomenon in vibration signals provides a theoretical basis and practical possibility for knowledge transfer between measurement points. Therefore, based on the ease of data acquisition and accessibility to the target measurement point, adjacent measurement points in the mechanical equipment where vibration signals are relatively easy to collect are considered source domain measurement points (i.e., associated measurement points), while the equipment fault location is considered the target domain measurement point. By utilizing the rich information contained in the known associated measurement point data for knowledge transfer, the trained model can be effectively applied to the diagnostic task of the target measurement point, solving the problems of insufficient target measurement point data and lack of label information. Specific optional embodiments are given below:

[0023] Example 1: As Figure 1-3 As shown, a transfer learning method for mechanical equipment associated measurement points based on multi-scale convolutional networks includes: collecting associated measurement point data of equipment fault points as source domain data and collecting target point data to be diagnosed as target domain data according to the transfer task; normalizing both source and target domain data, then using a sliding window mechanism to segment the normalized data to obtain samples, and determining the corresponding category labels of the samples as true labels, finally obtaining the source domain dataset and the target domain dataset; wherein, the category labels include normal category labels and different fault category labels; constructing a domain learning network; and jointly training the source and target domains through a shared network structure: firstly, using the labeled source domain data as input to the domain learning network of the source domain for... Features of associated measurement points are extracted, and the network structure and weights are shared with the domain learning network of the target domain. Then, the unlabeled target domain data is used as the input of the domain learning network of the target domain to extract target measurement point features. The maximum mean square difference method is used to measure the inter-domain difference between the associated measurement point features and the target measurement point features. A joint loss function is constructed by combining the classification loss of the source domain's domain learning network and the inter-domain difference loss as the optimization objective. The backpropagation algorithm is used for iterative training to update the parameters of the target domain learning model and predict the sample category. Finally, a target domain learning model with excellent diagnostic performance is obtained, realizing knowledge transfer between associated measurement points and target measurement points of mechanical equipment. The predicted category is compared with the real label to calculate the accuracy.

[0024] Furthermore, the sliding window mechanism method is as follows:

[0025]

[0026] In the formula, n represents the number of samples obtained by sliding window, l is the sample length, and s is the window step size.

[0027] Furthermore, the domain learning network specifically comprises: first, extracting data features using a multi-scale convolutional neural network and introducing a self-attention module to enhance the representational ability of the features; then, fusing feature information at different scales; and finally, introducing the maximum mean square difference method in the adaptive layer to measure the feature difference between the source domain and the target domain, so as to guide domain adaptation in the transfer learning process.

[0028] Furthermore, the multi-scale convolutional neural network includes two branches: the first branch uses a five-layer stacked structure of "convolutional layer + batch normalization layer + ReLU activation function + max pooling layer", and the kernel size of the first layer in the first branch is larger than the kernel size of the other layers in the first branch; the second branch uses a four-layer stacked structure of "convolutional layer + batch normalization layer + ReLU activation function + max pooling layer", and the kernel size of the first layer in the second branch is larger than the kernel size of the other layers in the second branch.

[0029] As can be seen from the above technical solution, the domain learning network constructed in this invention uses CNN networks of different scales, enabling the network to represent both local and large-scale features in the data well, and extracting richer features. On this basis, the introduction of a self-attention mechanism can further enhance the expressive power of features. The self-attention mechanism allows the network to dynamically adjust the attention weights according to the correlation between various elements in the data when processing data, which helps the model focus on the most important features, thereby improving the model's ability to identify key information. Then, the multi-scale features processed by the self-attention mechanism are subjected to point-by-point dot product to achieve feature fusion. Finally, the maximum mean square difference method is used in the adaptive layer to measure the difference between the source domain and the target domain. The maximum mean square difference method can comprehensively reflect the feature information of the data samples in the Hilbert space of the reproduction kernel, thereby enhancing domain adaptation.

[0030] Furthermore, the five convolutional layers in the first branch use kernel sizes of 64, 3, 3, 3, and 3, respectively; the four convolutional layers in the second branch use kernel sizes of 32, 3, 3, and 3, respectively. Because the vibration frequency of mechanical equipment is high, the data sampling rate is set relatively large, and the number of samples is also large, the periodicity of the signal and the impact characteristics of the vibration are considered when performing convolution. Therefore, in the technical solution of this invention, the first layer in both branches uses large convolutional kernels, which can effectively acquire the characteristics of the vibration signal and suppress high-frequency noise while reducing the computational load of the network.

[0031] The steps for calculating the maximum mean square difference are as follows:

[0032] Based on the feature set X of the associated measurement points and the feature set Y of the target measurement points extracted by the network, as well as the parameters σ and weights ωt of the Gaussian kernel function, the mixture RBF kernel matrix is ​​calculated. The formula for calculating the RBF kernel matrix is ​​as follows:

[0033]

[0034] Among them, X·X T , X·Y T Y·Y T It is the dot product matrix between sample features, and diag(X·X) and diag(Y·Y) represent the vectors composed of the diagonal elements of the dot product matrices X·X and Y·Y, respectively.

[0035] Next, based on the three hybrid RBF kernel matrices K of the hybrid RBF kernel function... XX K XY K YY The formula for calculating the maximum mean square difference is as follows:

[0036]

[0037] Where trace_X and trace_Y are the traces of matrices X and Y, respectively; m and n represent the number of samples in X and Y, respectively.

[0038] Furthermore, the joint loss function is expressed as follows:

[0039] loss = clc_Loss + mmsd * λ;

[0040] in, Let y represent the cross-entropy loss function for training in the source domain. i It is the actual category label, p i λ is the probability of belonging to class i predicted by the source domain learning model; λ is the adjustment coefficient; and mmsd represents the inter-domain difference loss obtained based on the maximum mean square difference.

[0041] Example 2: A mechanical correlation measurement point transfer learning method based on a multi-scale convolutional neural network is applied to rotating equipment. The following is a detailed explanation using a bearing as an example:

[0042] (1) The proposed method was tested using bearing vibration signal data publicly available from Case Western Reserve University in the United States. This data included data from the motor drive end and fan end, encompassing four categories: normal (NO), inner race fault (IF), rolling element fault (BF), and outer race fault (OF). Data with a sampling rate of 12000 Hz and a fault size of 0.007 inches was selected for the experiment. The drive end measurement point was used as the source domain measurement point, and its vibration data for the four different categories were used as the training set. The fan end measurement point was used as the target measurement point, and its vibration data for the four different categories were used as the test set. The specific dataset settings are shown in Table 1.

[0043] Table 1 Experimental Dataset Design

[0044]

[0045] The designed transfer tasks consist of two sets: Task 1 (transfer from the drive end 0HP to the fan end 0HP) and Task 2 (transfer from the drive end 1HP to the fan end 1HP). The deep transfer learning method enables the diagnostic knowledge of the source domain measurement points (motor drive end) to be applied to the prediction and diagnosis of the target domain measurement points (fan end).

[0046] (2) Read the vibration data of the source domain measuring point and the target measuring point, normalize the data, and then perform data segmentation on the normalized data. The step size of the sliding window is designed to be 1024, the size of each sample is 1024 data points, the data of each state is divided into 100 samples, and 400 samples are obtained in the source domain and the target domain respectively. Each of them contains four types of bearing state data. Generate corresponding labels according to the bearing state, and finally form the source domain dataset and the target domain dataset.

[0047] (3) The constructed domain learning network uses two CNNs with different kernel sizes to extract domain features. Then, the features extracted by the two CNNs at different scales are processed using self-attention layers, and the extracted features are then multiplied point-by-point to obtain the fused features. Average pooling is used to reduce the data volume of the fused domain features. The specific settings of the domain learning network are shown in Table 2. After the above processing, the source and target domains obtain their respective multi-scale fused features. During joint training, maximum mean square difference is added to the adaptive layer to adapt the domains between related measurement points, adjusting the network parameters to adapt to the feature distributions of the two domains. The overall framework for jointly training the source and target domains through a shared network structure is as follows: Figure 1 As shown.

[0048] Table 2 Domain Learning Networks

[0049]

[0050] (4) Network Training and Diagnostic Knowledge Transfer: Some specific network settings are as follows: batch size is set to 32, learning rate is set to 0.01, stochastic gradient descent (SGD) is selected as the optimizer, the total number of training rounds is set to 300, and the adjustment coefficient of the maximum mean square difference is set to 30. First, the labeled driving dataset of the source domain data is used as the input of the domain learning network of the source domain to extract the features of the associated measurement points, and the network structure and weights are shared with the domain learning network of the target domain. Then, the unlabeled target domain dataset is used as the input of the domain learning network of the target domain to extract the features of the target measurement points. In the adaptive layer, the output of two linear layers is used for domain adaptation: the maximum mean square difference (MMSD) between the source domain and the target domain in these two layers is calculated, and it is combined with the cross-entropy loss function of the source domain training as the final optimization objective for backpropagation and update iteration. Finally, a target domain learning model with excellent diagnostic performance is obtained, realizing the knowledge transfer between the associated measurement points of mechanical equipment and the target measurement points; the target domain learning model is used to predict the category of the unlabeled target domain data, and the predicted category is compared with the real label to calculate the accuracy. The training process for the two transfer tasks is described in [link to training process]. Figure 2 As can be seen, the training process of the method proposed in this invention is stable, and the model can transfer source domain diagnostic knowledge to fault diagnosis in the target domain.

[0051] To compare the advantages of the proposed method, the transfer task between the associated measurement points from the motor drive end to the motor fan end was performed using ordinary CNN, DDC (Deep Domain Confusion), and the method proposed in this invention. The results are shown in Table 3:

[0052] Table 3. Diagnostic Results of Migration Tasks

[0053] Migration task CNN DDC The method proposed in this invention 1 (0HP from drive end to fan end) 83.71％ 91.86％ 99.75％ 2 (1HP from drive end to fan end) 78.91％ 84.9％ 99.50％ Average accuracy 81.31％ 88.38％ 99.63％

[0054] As shown in Table 3, both the classic transfer learning method DDC and the method proposed in this invention significantly improve accuracy compared to ordinary CNN models, indicating that transfer learning is effective in diagnosing data from different test points. The method proposed in this invention achieves the highest accuracy, improving upon DDC by 11%, with an average accuracy of 99.63%, fully demonstrating its superiority. Based on the target domain learning model, category prediction is performed on unlabeled target domain data. The t-SNE visualization scatter plot of the predicted test results is shown below. Figure 3 As shown, the data of each category is effectively distinguished, and the distance between each category is large, indicating a significant recognition effect.

[0055] Therefore, based on the above, the present invention has the following advantages:

[0056] 1. Multi-scale feature extraction and fusion: Employing multi-scale convolutional networks can capture deep features across different receptive fields, effectively enhancing the data's expressive power. The feature fusion process comprehensively considers both local and large-scale features, enabling the model to understand the data more comprehensively.

[0057] 2. Enhanced attention mechanism: An attention module is introduced to adjust the weights of features, enabling the network to adaptively focus on important feature parts, further enhancing the feature representation ability and robustness.

[0058] 3. Calculation of maximum mean difference: The maximum mean square difference (MMSD) is used as a metric for domain adaptation. Compared with some common metrics such as MMD, MMSD can more comprehensively reflect the feature differences between source domain data and target domain data, thereby effectively reducing the distribution differences between domains.

[0059] 4. Transfer of associated measurement points: The proposed method does not rely on the label information of the target domain. Through domain adaptation and feature learning, it has strong adaptability and can be effectively applied to the knowledge transfer of associated measurement points of mechanical equipment.

[0060] These advantages enable the transfer learning method based on multi-scale convolutional networks to excel in addressing the problems of inconvenient sensor installation and poor model transfer diagnostic performance in mechanical equipment diagnosis, successfully realizing knowledge transfer between associated measurement points and target measurement points of mechanical equipment.

[0061] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for transfer learning of associated measurement points in mechanical equipment based on multi-scale convolutional networks, characterized in that, include: Based on the migration task, collect associated measurement point data of equipment fault points as source domain data, and collect target point data to be diagnosed as target domain data. Both the source domain data and the target domain data are normalized. Then, a sliding window mechanism is used to segment the normalized data to obtain samples and determine the corresponding category labels for the samples, thus obtaining the source domain dataset and the target domain dataset. The category labels include normal category labels and different fault category labels. A domain learning network is constructed. The source and target domains are jointly trained by sharing a network structure: First, labeled source domain data is used as input to the source domain learning network to extract associated measurement point features. The network structure and weights are then shared with the target domain learning network. Next, unlabeled target domain data is used as input to the target domain learning network to extract target measurement point features. The maximum mean square difference method is used to measure the inter-domain differences between associated and target measurement point features. A joint loss function is constructed using the classification loss of the source domain learning network and the inter-domain difference loss as the optimization objective. Iterative training is performed using the backpropagation algorithm to update the target domain learning model parameters and predict sample categories. Finally, a target domain learning model with excellent diagnostic performance is obtained, realizing knowledge transfer between associated and target measurement points of mechanical equipment. The domain learning network is specifically designed as follows: First, it uses a multi-scale convolutional neural network to extract data features and introduces a self-attention module to enhance the representation ability of the features; then, it fuses feature information at different scales; finally, it introduces the maximum mean square difference method in the adaptive layer to measure the feature difference between the source domain and the target domain. The multi-scale convolutional neural network includes two branches: the first branch uses a five-layer stacked "convolutional layer + batch normalization layer + ReLU activation function + max pooling layer", and the kernel size of the first layer in the first branch is larger than the kernel size of the other layers in the first branch; the second branch uses a four-layer stacked "convolutional layer + batch normalization layer + ReLU activation function + max pooling layer", and the kernel size of the first layer in the second branch is larger than the kernel size of the other layers in the second branch; The steps for calculating the maximum mean square difference are as follows: Based on the feature set of associated measurement points extracted from the network and target measurement point feature set and the parameters of the Gaussian kernel function. and weight Calculate the hybrid RBF kernel matrix; the formula for calculating the RBF kernel matrix is ​​as follows: ； ； ； in, , , It is the dot product matrix between sample features. , Represent the dot product matrix respectively A vector consisting of the diagonal elements; Next, based on the three hybrid RBF kernel matrices of the hybrid RBF kernel function... , , The formula for calculating the maximum mean square difference is as follows: ； in, It is a matrix traces; , They represent , The number of samples.

2. The mechanical equipment associated measurement point transfer learning method based on multi-scale convolutional networks according to claim 1, characterized in that, The sliding window mechanism method is as follows: ； In the formula, This indicates the number of samples obtained by the sliding window. s is the sample length and s is the window step size.

3. The mechanical equipment associated measurement point transfer learning method based on multi-scale convolutional networks according to claim 1, characterized in that, The five convolutional layers in the first branch use kernel sizes of 64, 3, 3, 3, and 3, respectively; the four convolutional layers in the second branch use kernel sizes of 32, 3, 3, and 3, respectively.

4. The mechanical equipment associated measurement point transfer learning method based on multi-scale convolutional networks according to claim 1, characterized in that, The joint loss function is expressed as follows: ； in, This represents the cross-entropy loss function used for training in the source domain. It is an adjustment factor. This represents the difference loss between domains obtained based on the maximum mean square difference.

Images