Cross-element fault diagnosis method based on lightweight multi-task network

Abstract

The invention relates to a cross-component fault diagnosis method based on a lightweight multi-task network, and aims to solve the technical problem that fault diagnosis of a single component cannot comprehensively reflect the overall state of equipment, and the technical scheme is as follows: collecting operation data under each channel collected by sensors on different components, performing normalization processing, eliminating dimensional difference, and determining the fault diagnosis result; smoothing processing, multi-sensor signal fusion, construction of a cross-element feature sharing network, inputting a training set into the proposed model for training, setting the number of iterations to be 500, selecting a batch size to be 64, and selecting an adam function by an optimizer; according to the method, the lightweight convolutional neural network is adopted as a backbone network, calculation parameters and resource consumption are greatly reduced, a novel parameter-free task attention mechanism is introduced, and on the premise that the model parameter burden is not increased, the key feature weight of the vibration signal is effectively enhanced, and the information extraction efficiency is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment fault diagnosis, and in particular relates to a cross-component fault diagnosis method based on a lightweight multi-task network. Background Art

[0002] With the acceleration of industrialization, the reliability and stability of equipment have become the key to ensuring production safety and efficiency. Fault diagnosis of core components such as plunger pumps and gearboxes is particularly important, and their health status directly affects the operation of the overall system. However, traditional fault diagnosis methods are limited by manual experience and insufficient data, and it is difficult to meet the needs of modern industry. Deep learning technology, with its powerful feature extraction and automatic learning capabilities, provides a new solution for intelligent fault diagnosis. However, deep learning models are highly dependent on data, and large-scale data acquisition is difficult, especially under complex and changeable working conditions. In addition, the fault diagnosis of a single component cannot fully reflect the overall status of the equipment, and cross-device diagnosis has become an inevitable trend, but this also puts higher requirements on model complexity and computational efficiency. Summary of the invention

[0003] The purpose of the present invention is to solve the above problems and provide a cross-component fault diagnosis method based on a lightweight multi-task network.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A cross-component fault diagnosis method based on a lightweight multi-task network comprises the following steps:

[0006] Step 1) Multi-sensor signal acquisition and data preprocessing, the specific steps are as follows:

[0007] Step 1.1) Data collection: collect the operating data of each channel collected by sensors on different components;

[0008] Step 1.2) Normalize the feature data of each channel to eliminate dimensional differences and ensure that the data are compared and analyzed at the same scale;

[0009] Step 1.3) Use the sliding window moving average method to smooth the device status data, reduce noise interference, and improve signal quality;

[0010] Step 2) Multi-sensor signal fusion, the specific steps are as follows:

[0011] Step 2.1) Combine multiple sensor signals in series to construct the original sample space X = {X1, X2};

[0012] Step 2.2) Apply the following mapping formula to map the sample space into the feature space for further analysis:

[0013] Step 2.3) Solve the eigenvalues ​​and eigenvectors of the mapped eigenmatrix, select the eigenvector corresponding to the maximum eigenvalue as the first principal component, and its corresponding eigenvector is the weight of x1 or the transformed basis vector;

[0014] Step 2.4) According to the following fusion formula, the information of x1 and x2 is fused to obtain the fused feature vector: F1=ω 11 zx1+ω 21 zx2 (2);

[0015] Where: 11 represents the eigenvector corresponding to the first principal component, ω 21 represents the eigenvector corresponding to the second principal component. zx1 and zx2 represent the normalized raw data respectively. The value of σ is the value corresponding to when the contribution rate of the first principal component tends to be stable. After many attempts, the value of σ is 3;

[0016] Step 3) Build a cross-component feature sharing network:

[0017] Step 3.1) Lightweight backbone network design:

[0018] The backbone network consists of three convolutional blocks, each of which contains a series of convolutional layers, batch normalization, and ReLU activation functions. Convolutional block 1 extracts preliminary features, convolutional block 2 extracts higher-level features, and convolutional block 3 further extracts detailed high-order features. Each convolutional block gradually extracts multi-level features through appropriate convolution kernel size and channel number configuration;

[0019] Step 3.2) Task Attention Module

[0020] This module is a simple parameter-free attention mechanism for cross-model diagnosis tasks;

[0021] Step 3.3) Dynamically adjusted metric learning

[0022] Introducing angle interval and scale factor into softmax loss to optimize multi-classification performance;

[0023] Step 4) Input the training set into the proposed model for training, set the number of iterations to 500, the batch size to 64, and the optimizer to the adam function.

[0024] Furthermore, the task attention mechanism in step 3.2) includes the following steps:

[0025] (1) Calculate the local mean and variance:

[0026]

[0027] In formula (3), N(i,j) represents the 3×3 neighborhood centered at position (i,j), |N(i,j)| is the number of pixels in the neighborhood. For a 3×3 neighborhood, |N(i,j)|=9, X m,n is the pixel value at the neighborhood (m, n) position.

[0028] In formula (4), Xm,n is the pixel value at the neighborhood (m,n). i,j is the local mean of the neighborhood.

[0029] (2) Calculate the new energy factor based on local variance:

[0030]

[0031] (3) Calculate the new attention weight. The Sigmoid function compresses the energy factor to the interval [0,1], making the new energy factor suitable as the attention weight:

[0032] A N(i,j) =σ(E N(i,j) ) (6)

[0033] Where: σ(·) is the Sigmoid function, defined as:

[0034]

[0035] (4) The calculated attention weights are applied to the original feature map, and the values ​​of the original feature map are adjusted in a weighted manner so that important features are enhanced and unimportant features are suppressed:

[0036] X″ i,j =X i,j A N(i,j) (8)

[0037] Where: X″ i,j is the adjusted eigenvalue of position (i, j); X i,j is the eigenvalue of the original feature map at position (i, j), A′ i,j is the attention weight calculated at position (i, j).

[0038] Further, the specific steps of step 3.3) are as follows:

[0039] (1) Calculate the angle: Let the input feature vector be x i , the category label is y i , the center vector of each category is w k , introduce a scale factor α to adjust the size of the feature vector;

[0040] Eigenvector x i and the category center vector w k The angle θ between i,k It can be calculated by the inner product formula:

[0041]

[0042] Since the scale factor α will cancel each other in the numerator and denominator, there is no need to explicitly consider α in actual calculations;

[0043] (2) Angular Loss: The goal of Angular Loss is to minimize the angle between the feature vectors of the same type and the center vector of the category, while maximizing the angle between the feature vectors of different types and the center vector of the category. This ability to distinguish is enhanced by introducing a margin:

[0044]

[0045] Where N is the number of samples in the batch, is the eigenvector x i Its true category center vector y i The cosine of the angle between i,k ) is the eigenvector x i With other categories of center vector w k The cosine value of the angle between them, and m is the margin used to increase the distinction between different categories.

[0046] (3) Introducing a scale factor: By scaling the feature vector when calculating the loss:

[0047] x′ i =axi (11)

[0048] Then calculate the angle loss:

[0049]

[0050] Here, the numerator is the eigenvector x i and the category center vector w k The denominator is the product of their norms.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1) The present invention adopts a lightweight convolutional neural network as the backbone network, which greatly reduces the calculation parameters and resource consumption;

[0053] 2) The present invention introduces a novel parameter-free task attention mechanism (ITA), which effectively enhances the key feature weights of vibration signals and improves the efficiency of information extraction without increasing the burden of model parameters;

[0054] 3) The present invention uses a metric learning strategy to optimize the feature space, so that the learned features present a clearer geometric distribution on the hypersphere, thereby enhancing the discriminability and interpretability of the features. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic diagram of the cross-device signal fusion process of the present invention;

[0056] Figure 2 It is a schematic diagram of the structure framework of the shared knowledge network of the present invention;

[0057] Figure 3 is a curve showing the change of the training loss function value of the network of the present invention;

[0058] Figure 4 is a curve showing the change of the verification loss function value of the network of the present invention; DETAILED DESCRIPTION

[0059] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0060] A cross-component fault diagnosis method based on a lightweight multi-task network comprises the following steps:

[0061] Step 1) Multi-sensor signal acquisition and data preprocessing, the specific steps are as follows:

[0062] Step 1.1) Data acquisition: The vibration signal and acceleration signal of the gearbox are collected across the components, as well as the data of nine channels such as vibration, sound pressure, oil pressure, speed, and current of different fault types of the plunger pump;

[0063] Step 1.2) Normalize the feature data of each channel to eliminate dimensional differences and ensure that the data are compared and analyzed at the same scale;

[0064] Step 1.3) Use the sliding window moving average method to smooth the device status data, reduce noise interference, and improve signal quality;

[0065] Step 2) Multi-sensor signal fusion, such as Figure 1 As shown, the specific steps are as follows:

[0066] Step 2.1) Combine multiple sensor signals in series to construct the original sample space X = {X1, X2};

[0067] Step 2.2) Apply the following mapping formula to map the sample space into the feature space for further analysis:

[0068] Step 2.3) Solve the eigenvalues ​​and eigenvectors of the mapped eigenmatrix, select the eigenvector corresponding to the maximum eigenvalue as the first principal component, and its corresponding eigenvector is the weight of x1 or the transformed basis vector;

[0069] Step 2.4) According to the following fusion formula, the information of x1 and x2 is fused to obtain the fused feature vector: F1=ω 11 zx1+ω 21 zx2 (2);

[0070] Where: 11 represents the eigenvector corresponding to the first principal component, ω 21 represents the eigenvector corresponding to the second principal component. zx1 and zx2 represent the normalized raw data respectively. The value of σ is the value corresponding to when the contribution rate of the first principal component tends to be stable. After many attempts, the value of σ is 3;

[0071] Step 3) Build a cross-component feature sharing network:

[0072] The cross-component feature sharing network first uses a lightweight CNN network as the backbone network, which is composed of three convolutional blocks. Then, a set of attention modules and specific layers are added to each convolutional block network. The feature attention module acts as a feature selector for the task sharing network, and uses a new energy factor to calculate the local energy of each feature map position to determine the attention weight. Finally, the fused shared features are fed back to the branch structure, and a dynamically adjusted metric learning module is added to the multi-branch structure. The metric learning module introduces scale factors and angle intervals. The scale factor α is applied to the normalization process of the feature vector, increasing the angle interval between the feature vectors, so that the learned features can achieve the goal of increasing the compactness within the class and the distinction between classes.

[0073] Step 3.1) Lightweight backbone network design:

[0074] The backbone network is the core part of the lightweight convolutional interpretable shared knowledge framework, responsible for extracting common features from the input data. It consists of three convolution blocks, each of which contains a series of convolutional layers, batch normalization, and ReLU activation functions. Convolution block 1 extracts preliminary features, convolution block 2 extracts higher-level features, and convolution block 3 further extracts detailed high-order features. Each convolution block gradually extracts multi-level features through appropriate convolution kernel size and number of channels. The backbone network adopts a lightweight design to reduce the computational complexity of the model, making it suitable for resource-constrained devices and real-time applications. As a shared feature extractor, it provides a consistent feature basis for different task branches, which helps to improve the efficiency and accuracy of multi-task learning. The details of the convolution blocks can also be found in Figure 2 Seen in.

[0075] Step 3.2) Task Attention Module

[0076] This module is a simple parameter-free attention mechanism for cross-model diagnosis tasks; the specific steps are as follows:

[0077] (1) Calculate the local mean and variance:

[0078]

[0079] In formula (3), N(i,j) represents the 3×3 neighborhood centered at position (i,j), |N(i,j)| is the number of pixels in the neighborhood. For a 3×3 neighborhood, |N(i,j)|=9, X m,n is the pixel value at the neighborhood (m, n) position.

[0080] In formula (4), Xm,n is the pixel value at the neighborhood (m,n). i,j is the local mean of the neighborhood.

[0081] (2) Calculate the new energy factor based on local variance:

[0082]

[0083] (3) Calculate the new attention weight. The Sigmoid function compresses the energy factor to the interval [0,1], making the new energy factor suitable as the attention weight:

[0084] A N(i,j) =σ(E N(i,j) ) (6)

[0085] Where: σ(·) is the Sigmoid function, defined as:

[0086]

[0087] (4) The calculated attention weights are applied to the original feature map, and the values ​​of the original feature map are adjusted in a weighted manner so that important features are enhanced and unimportant features are suppressed:

[0088] X″ i,j =X i,j A N(i,j) (8)

[0089] Where: X″ i,j is the adjusted eigenvalue of position (i, j); X i,j is the eigenvalue of the original feature map at position (i, j), A′ i,j is the attention weight calculated at position (i, j).

[0090] Step 3.3) Dynamically adjusted metric learning

[0091] Introduce angle interval and scale factor into softmax loss to optimize multi-classification performance. The specific steps are as follows:

[0092] (1) Calculate the angle: Let the input feature vector be x i , the category label is y i , the center vector of each category is w k , introduce a scale factor α to adjust the size of the feature vector;

[0093] Eigenvector x i and the category center vector w k The angle θ between i,k It can be calculated by the inner product formula:

[0094]

[0095] Since the scale factor α will cancel each other in the numerator and denominator, there is no need to explicitly consider α in actual calculations;

[0096] (2) Angular Loss: The goal of Angular Loss is to minimize the angle between the feature vectors of the same type and the center vector of the category, while maximizing the angle between the feature vectors of different types and the center vector of the category. This ability to distinguish is enhanced by introducing a margin:

[0097]

[0098] Where N is the number of samples in the batch, is the eigenvector x i Its true category center vector y i The cosine of the angle between i,k ) is the eigenvector x i With other categories of center vector wk The cosine value of the angle between them, and m is the margin used to increase the distinction between different categories.

[0099] (3) Introducing a scale factor: By scaling the feature vector when calculating the loss:

[0100] x′ i =axi (11)

[0101] Then calculate the angle loss:

[0102]

[0103] Where the numerator is the eigenvector x i and the category center vector w k The denominator is the product of their norms.

[0104] In the calculation of the loss function, a scale factor is introduced to affect the size of the feature vector, thereby adjusting the learning process. On this basis, by increasing the angle interval between feature vectors and adding an angle penalty term to the loss function, feature vectors of different categories are kept as far apart as possible during the training process, thereby significantly enhancing the performance and feature representation capabilities of the model.

[0105] Step 4) Input the training set into the proposed model for training, set the number of iterations to 500, select the batch size of 64, and select the adam function as the optimizer. The initial learning rate is 0.001. Selecting an appropriate learning rate can optimize the model training process and enable the network to converge quickly in the right direction. When the loss value fluctuates within a certain range, stop training. Figure 3-4 The changing curves of the diagnostic accuracy and loss function value of the multi-task learning network are shown.

[0106] Lightweight model evaluation and verification

[0107] We verified the number of parameters, floating point numbers and test time of each model under different convolution kernels. The number of parameters represents the total amount of learnable parameters contained in the model during the training phase, which reflects the degree of model storage requirements; the floating point calculation amount is used as a measure of the total number of floating point operations required during the execution of the model; the test time can directly reflect the efficiency of the model in processing data. These three indicators together constitute the evaluation criteria for evaluating the complexity of the model. The higher the number of parameters and FLOPs, the larger the model. The optimal parameter combination was obtained through a large number of experimental verifications. Compared with other methods, the three indicators of this method are the smallest, which shows that the model is more stable than other models on cross-component data sets. Table 1 shows the average recognition performance of each model in Task 1 and Task 2, among which the FLOPs with a convolution kernel size of 1*1*1 is the lowest, which also proves the lightweight construction of the model. At the same time, the test time of the present invention is shorter than that of other methods, about 0.1573 seconds.

[0108] Table 1 Lightweight comparison of models with different convolution sizes

[0109]

Claims

1. A cross-component fault diagnosis method based on a lightweight multi-task network, characterized in that: The steps include: Step 1) Multi-sensor signal acquisition and data preprocessing, the specific steps are as follows: Step 1.1) Data collection: collect the operating data of each channel collected by sensors on different components; Step 1.2) Normalize the feature data of each channel to eliminate dimensional differences and ensure that the data are compared and analyzed at the same scale; Step 1.3) Use the sliding window moving average method to smooth the device status data, reduce noise interference, and improve signal quality; Step 2) Multi-sensor signal fusion, the specific steps are as follows: Step 2.1) Combine multiple sensor signals in series to construct the original sample space X = {X1, X2}; Step 2.2) Apply the following mapping formula to map the sample space into the feature space for further analysis: Step 2.3) Solve the eigenvalues ​​and eigenvectors of the mapped eigenmatrix, select the eigenvector corresponding to the maximum eigenvalue as the first principal component, and its corresponding eigenvector is the weight of x1 or the transformed basis vector; Step 2.4) According to the following fusion formula, the information of x1 and x2 is fused to obtain the fused feature vector: F1=ω 11 xz1+ω 21 zx2 (2); Where: 11 represents the eigenvector corresponding to the first principal component, ω 21 represents the eigenvector corresponding to the second principal component. zx1 and zx2 represent the normalized raw data respectively. The value of σ is the value corresponding to when the contribution rate of the first principal component tends to be stable. After many attempts, the value of σ is 3; Step 3) Build a cross-component feature sharing network: Step 3.1) Lightweight backbone network design: The backbone network consists of three convolutional blocks, each of which contains a series of convolutional layers, batch normalization, and ReLU activation functions. Convolutional block 1 extracts preliminary features, convolutional block 2 extracts higher-level features, and convolutional block 3 further extracts detailed high-order features. Each convolutional block gradually extracts multi-level features through appropriate convolution kernel size and channel number configuration; Step 3.2) Task Attention Module This module is a simple parameter-free attention mechanism for cross-model diagnosis tasks; Step 3.3) Dynamically adjusted metric learning Introducing angle interval and scale factor into softmax loss to optimize multi-classification performance; Step 4) Input the training set into the proposed model for training, set the number of iterations to 500, the batch size to 64, and the optimizer to the adam function.

2. A cross-component fault diagnosis method based on a lightweight multi-task network according to claim 1, characterized in that: The task attention mechanism in step 3.2) includes the following steps: (1) Calculate the local mean and variance: In formula (3), N(i,j) represents the 3×3 neighborhood centered at position (i,j), |N(i,j)| is the number of pixels in the neighborhood. For a 3×3 neighborhood, |N(i,j)|=9, X m,n is the pixel value at the neighborhood (m, n) position. In formula (4), Xm,n is the pixel value at the neighborhood (m,n). i,j is the local mean of the neighborhood. (2) Calculate the new energy factor based on local variance: (3) Calculate the new attention weight. The Sigmoid function compresses the energy factor to the interval [0,1], making the new energy factor suitable as the attention weight: A N(i,j) =σ(E N(i,j) ) (6) Where: σ(·) is the Sigmoid function, defined as: (4) The calculated attention weights are applied to the original feature map, and the values ​​of the original feature map are adjusted in a weighted manner so that important features are enhanced and unimportant features are suppressed: X″ i,j =X i,j A N(i,j) (8) Where: X″ i,j is the adjusted eigenvalue of position (i, j); X i,j is the eigenvalue of the original feature map at position (i, j), A′ i,j is the attention weight calculated at position (i, j);.

3. A cross-component fault diagnosis method based on a lightweight multi-task network according to claim 1, characterized in that: The specific steps of step 3.3) are as follows: (1) Calculate the angle: Let the input feature vector be x i , the category label is y i , the center vector of each category is w k , introduce a scale factor α to adjust the size of the feature vector; Eigenvector x i and the category center vector w k The angle θ between i,k It can be calculated by the inner product formula: Since the scale factor α will cancel each other in the numerator and denominator, there is no need to explicitly consider α in actual calculations; (2) Angular Loss: The goal of Angular Loss is to minimize the angle between the feature vectors of the same type and the center vector of the category, while maximizing the angle between the feature vectors of different types and the center vector of the category. This ability to distinguish is enhanced by introducing a margin: Where N is the number of samples in the batch, is the eigenvector x i Its true category center vector y i The cosine of the angle between i,k ) is the eigenvector x i With other categories of center vector w k The cosine value of the angle between them, and m is the margin used to increase the distinction between different categories. (3) Introducing a scale factor: By scaling the feature vector when calculating the loss: x′ i =axi (11) Then calculate the angle loss: Where the numerator is the eigenvector x i and the category center vector w k The denominator is the product of their norms.

Images