Mechanical fault cross-domain diagnosis method based on multi-scale feature adaptive network

Abstract

The application discloses a kind of mechanical fault cross-domain diagnosis methods based on multi-scale feature adaptive network, specifically according to the following steps implementation: step one, obtain source domain data and target domain data;Step two, construct neural network model based on multi-scale feature adaptive block;Step three, the source domain data and target domain data of step one are loaded to neural network model based on multi-scale feature adaptive block and are trained, and the trained neural network model is obtained;Step four, using the trained neural network model to diagnose mechanical fault.The application solves the problem that the feature extracted by neural network in the fault diagnosis method in the prior art is simple, and the algorithm does not have the ability to land due to the overfitting phenomenon.

Description

Technical Field

[0001] This invention belongs to the technical field of mechanical fault diagnosis methods, and relates to a cross-domain diagnosis method for mechanical faults based on multi-scale feature adaptive networks. Background Technology

[0002] With the development of information technology, mechanical fault diagnosis research has mainly been divided into three directions: dynamic mechanism analysis based on mechanical equipment, signal processing-based analysis, and data-driven model analysis.

[0003] With the development of machine learning and deep learning, data-driven model analysis is increasingly becoming a key research focus. This involves using massive amounts of data from mechanical equipment to model and analyze systems. The paper "Deep residual networks for fault diagnosis. IEEE Transactions on Industrial Informatics" (Zhao, M., Zhong, S., Fu, X., Tang, B., & Pecht, M. (2019)) introduces soft thresholding functions and channel attention mechanisms into deep residual networks, significantly improving the model's accuracy in identifying fault signals in noisy environments.

[0004] Network-based methods focus on deep learning neural networks, training them with source domain data, transferring some parameters to the target network by freezing them, and fine-tuning the network parameters using limited data from the engineering domain. Instance-based methods weight the source domain data during training, enabling the classifier to more accurately predict labels aligned with the target domain and reducing the probability of negative transfer. Mapping-based methods use feature extractors to map features extracted from the source and target domains to similar feature spaces, aligning the source and target domain data using different distance functions to reduce data distribution differences. Adversarial methods borrow ideas from GAN networks, using domain discriminators to reduce feature distribution differences between the source and target domains.

[0005] Current fault diagnosis methods based on neural networks mostly focus on optimizing the fault diagnosis results from an algorithmic perspective, but neglect the performance improvements brought about by changes in network structure. Existing methods often employ simple convolutional neural network structures for feature extractors, resulting in shallow networks with no targeted improvements to their internal structure. This leads to simple extracted features, poor generalization ability, and a probability of overfitting, rendering the algorithms unsuitable for practical application. Summary of the Invention

[0006] The purpose of this invention is to provide a cross-domain diagnosis method for mechanical faults based on multi-scale feature adaptive networks, which solves the problems in existing fault diagnosis methods where the features extracted by neural networks are simple and overfitting occurs, resulting in the algorithm's lack of practical application.

[0007] The technical solution adopted in this invention is a cross-domain diagnosis method for mechanical faults based on multi-scale feature adaptive networks, which is implemented according to the following steps:

[0008] Step 1: Obtain source domain data D s ={x i ,y i} and target domain data D t ={x j ,y j}, where x i For source domain data D s The data signal in y i For x i Corresponding data label; x j For target domain data D t The data signal in y j For x j The corresponding data labels, where i,j∈{1,2,3,...,n};

[0009] Step 2: Construct a neural network model based on multi-scale feature adaptive blocks;

[0010] Step 3: Load the source domain data and target domain data from Step 1 into the neural network model based on multi-scale feature adaptive blocks for training, and obtain the trained neural network model.

[0011] Step four: Use the trained neural network model to diagnose mechanical faults.

[0012] The invention is further characterized in that,

[0013] The neural network model based on multi-scale feature adaptive blocks in step two includes a wide convolutional layer, four feature adaptive blocks, a multi-scale feature concatenation layer, and a classifier connected in sequence. The output of each feature adaptive block is also connected to the multi-scale feature concatenation layer.

[0014] The kernel size of a wide convolutional layer is 31;

[0015] The classifier consists of two fully connected layers of classifiers connected in series.

[0016] The feature adaptive block consists of a depthwise separable convolutional layer, a pointwise convolutional layer, a pointwise convolutional layer, and a global flat pooling layer connected in sequence. The global flat pooling layer is connected in parallel to a source domain fully connected layer and a target domain fully connected layer. The source domain fully connected layer and the target domain fully connected layer are jointly connected to an adaptive feature selection layer. The second pointwise convolutional layer is also connected to the adaptive feature selection layer.

[0017] The kernel sizes of the depth-separable convolutional layers for the first to fourth feature adaptive blocks are 25, 13, 7, and 3, respectively.

[0018] Step three specifically involves:

[0019] Step 3.1: Input the source domain data and target domain data into the neural network model based on multi-scale feature adaptive blocks and pass them through a wide convolutional layer with a kernel size of 31 to obtain the low-dimensional features corresponding to the source domain data and target domain data, respectively. and

[0020] Step 3.2, convert the low-dimensional features obtained in Step 3.1 into... and By sequentially passing the feature adaptive blocks with four different convolutional kernel sizes, the following results are obtained: Features extracted for different sizes and Features extracted for different sizes

[0021] Step 3.3: The features extracted from different sizes obtained in Step 3.2 are concatenated in a multi-scale feature concatenation layer to obtain multi-scale features of different dimensions.

[0022]

[0023]

[0024] Step 3.4: The multi-scale features of different dimensions obtained in Step 3.3 are used in the classifier for classification training, domain transfer training, class transfer training, and feature orthogonalization regularization training, respectively.

[0025] Step 3.5: Iteratively train the neural network model using source domain data and target domain data until convergence, and obtain the trained neural network model.

[0026] The loss function Loss used in step 3.4 is as follows:

[0027] Loss = Loss class +λLoss ad +γLoss ac +ρLossd

[0028] Among them, Loss class For classification-based cross-entropy loss, Loss ad For consistency loss based on domain alignment, Loss ac For class alignment-based consistency loss, Loss d For regularization terms based on feature orthogonalization loss, λ, γ, and ρ are the weight coefficients of the consistency loss of domain alignment, consistency loss of class alignment, and regularization term of feature orthogonalization loss in the overall loss of network training, respectively, with a magnitude between 0 and 1.

[0029] Classification-based cross-entropy loss class as follows:

[0030]

[0031] Among them, f MFACN For the proposed MFACN network, This represents features based on different dimensions across multiple scales;

[0032] Consistency loss based on domain alignment ad as follows:

[0033]

[0034] Where, k 1 k 2 Representing data fields in different source domains, Indicates k 1 The i-th sample in the domain, Let ||||2|| denote the j-th sample in the target domain, and let Euclidean distance be used as the distance metric function to measure samples from different domains.

[0035] Consistency loss based on class alignment ac as follows:

[0036]

[0037] Among them, because the target domain data is not used in training, k 1 k 2 Representing data fields in different source domains, This indicates that the field k is in category c. 1 The i-th sample; This indicates that the field k is in category c. 2 The j-th sample;

[0038] Regularization term Loss based on feature orthogonalization loss d as follows:

[0039]

[0040] in, Represents multi-scale features of the source domain. This represents the transpose of the multi-scale features of the target domain.

[0041] The beneficial effects of this invention are:

[0042] This invention constructs features through convolutional layers of different sizes, simultaneously establishing low-dimensional and high-dimensional features. The constructed features possess a global perspective, reducing the probability of overfitting caused by local task training. The features extracted from each layer are concatenated, and feature adaptive selection is achieved based on self-learning of weight coefficients. For domain adaptation scenarios, two links are provided for source domain data and target domain data, so that the network can autonomously select effective features that are more compatible with the target domain data. Loss functions are established for data distribution-based transfer and category-based transfer respectively, and cosine similarity is used to make the two as orthogonal as possible, thereby utilizing more features to achieve transfer diagnosis. Attached Figure Description

[0043] Figure 1 This is a flowchart of the cross-domain diagnosis method for mechanical faults based on multi-scale feature adaptive networks according to the present invention;

[0044] Figure 2 This is a structural diagram of the neural network model based on multi-scale feature adaptive blocks in the cross-domain diagnosis method for mechanical faults based on multi-scale feature adaptive networks of this invention;

[0045] Figure 3 This is a structural diagram of the basic feature adaptive block in the cross-domain diagnosis method for mechanical faults based on multi-scale feature adaptive networks in this invention;

[0046] Figure 4 This is a bar chart showing the diagnostic accuracy in Embodiment 3 of the present invention. Detailed Implementation

[0047] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0048] Example 1

[0049] A cross-domain diagnosis method for mechanical faults based on multi-scale feature adaptive networks, the process of which is as follows: Figure 1 As shown, the specific steps are as follows:

[0050] Step 1: Obtain source domain data D s ={x i ,y i} and target domain data D t ={x j ,yj}, where x i For source domain data D s The data signal in y i For x i Corresponding data label; x j For target domain data D t The data signal in y j For x j The corresponding data labels, where i,j∈{1,2,3,...,n};

[0051] Step 2: Construct a neural network model based on multi-scale feature adaptive blocks;

[0052] Step 3: Load the source domain data and target domain data from Step 1 into the neural network model based on multi-scale feature adaptive blocks for training, and obtain the trained neural network model.

[0053] Step four: Use the trained neural network model to diagnose mechanical faults.

[0054] like Figure 2 As shown, the neural network model based on multi-scale feature adaptive blocks in step two includes a wide convolutional layer, four feature adaptive blocks, a multi-scale feature concatenation layer, and a classifier connected in sequence. The output of each feature adaptive block is also connected to the multi-scale feature concatenation layer.

[0055] The kernel size of a wide convolutional layer is 31;

[0056] The classifier consists of two fully connected layers of classifiers connected in series.

[0057] like Figure 3 As shown, the feature adaptive block includes a depthwise separable convolutional layer, a pointwise convolutional layer, a pointwise convolutional layer, and a global flat pooling layer connected in sequence. The global flat pooling layer is connected in parallel to a source domain fully connected layer and a target domain fully connected layer. The source domain fully connected layer and the target domain fully connected layer are jointly connected to an adaptive feature selection layer. The second pointwise convolutional layer is also connected to the adaptive feature selection layer.

[0058] Step three specifically involves:

[0059] Step 3.1: Input the source domain data and target domain data into the neural network model based on multi-scale feature adaptive blocks and pass them through a wide convolutional layer with a kernel size of 31 to obtain the low-dimensional features corresponding to the source domain data and target domain data, respectively. and

[0060] Step 3.2, convert the low-dimensional features obtained in Step 3.1 into... and By sequentially passing the feature adaptive blocks with four different convolutional kernel sizes, the following results are obtained: Features extracted for different sizes and Features extracted for different sizes

[0061] Step 3.3: The features extracted from different sizes obtained in Step 3.2 are concatenated in a multi-scale feature concatenation layer to obtain multi-scale features of different dimensions.

[0062]

[0063]

[0064] Step 3.4: The multi-scale features of different dimensions obtained in Step 3.3 are used in the classifier for classification training, domain transfer training, class transfer training, and feature orthogonalization regularization training, respectively.

[0065] Step 3.5: Iteratively train the neural network model using source domain data and target domain data until convergence, and obtain the trained neural network model.

[0066] The loss function Loss used in step 3.4 is as follows:

[0067] Loss = Loss class +λLoss ad +γLoss ac +ρLoss d

[0068] Among them, Loss class For classification-based cross-entropy loss, Loss ad For consistency loss based on domain alignment, Loss ac For class alignment-based consistency loss, Loss d For regularization terms based on feature orthogonalization loss, λ, γ, and ρ are the weight coefficients of the consistency loss of domain alignment, consistency loss of class alignment, and regularization term of feature orthogonalization loss in the overall loss of network training, respectively, with a magnitude between 0 and 1.

[0069] Classification-based cross-entropy loss class as follows:

[0070]

[0071] Among them, f MFACN For the proposed MFACN network, This represents features based on different dimensions across multiple scales;

[0072] Consistency loss based on domain alignment ad as follows:

[0073]

[0074] Where, k 1 k 2 Representing data fields in different source domains, Indicates k 1 The i-th sample in the domain, Let ||j|| represent the j-th sample in the target domain, and ||j||2 represent the 2-norm. Euclidean distance is used as a distance metric function to measure samples from different domains.

[0075] Consistency loss based on class alignment ac as follows:

[0076]

[0077] Among them, because the target domain data is not used in training, k 1 k 2 Representing data fields in different source domains, This indicates that the field k is in category c. 1 The i-th sample; This indicates that the field k is in category c. 2 The j-th sample;

[0078] Regularization term Loss based on feature orthogonalization loss d as follows:

[0079]

[0080] in, Represents multi-scale features of the source domain. Tr represents the transpose of the multi-scale features of the target domain, and Tr is used to calculate the rank of the matrix.

[0081] The consistency loss functions for domain alignment and class alignment utilize the domain labels and class labels of the data, respectively, aiming to align the data into the regenerated Hilbert space and strive to more accurately narrow the distance between data distributions. The cross-entropy loss function for classification trains a more accurate classifier using the extracted multi-scale features and the labels of the source domain data. The feature orthogonalization loss regularization term, as a supplement to domain alignment and class alignment, avoids overfitting during training, preventing the network from extracting overly specific features.

[0082] Example 2

[0083] Based on Example 1, the kernel sizes of the depth-separable convolutional layers of the first to fourth feature adaptive blocks are 25, 13, 7, and 3, respectively.

[0084] Example 3

[0085] Building upon Examples 1 and 2, this example utilizes the bearing dataset from the University of Paderborn, Germany. This dataset contains real damage caused by human error and accelerated life. By controlling three parameters—rotation speed, load torque, and bearing radial force—bearing data is recorded under four operating conditions, including vibration signals and motor current signals, with a sampling frequency of 64kHz. The operating parameters are shown in Table 1. Human error is categorized into three types: electrical discharge machining (EDM) damage, electrical engraving damage, and drilling damage. Combining bearing type, fault location, fault severity, and damage type, twelve fault types are identified, as shown in Table 2. Serial number 1 in Table 1 is used as the target domain, and the other data are used as the source domain for experimental analysis. With the addition of a healthy state, a total of thirteen fault types are achieved, as shown in Table 2.

[0086] Table 1. Bearing Data Set Operating Conditions from the University of Paderborn, Germany

[0087] 0 1500 0.7 1000 1 900 0.7 1000 2 1500 0.1 1000 3 1500 0.7 400

[0088] Table 2. Fault Types in Bearing Dataset from the University of Paderborn, Germany

[0089] KA01 Outer ring 1 Electrical Discharge Machining KA03 Outer ring 2 EDM KA05 Outer ring 1 EDM KA06 Outer ring 2 EDM KA07 Outer ring 1 drilling KA08 Outer ring 2 drilling KA09 Outer ring 2 drilling KI01 Inner circle 1 Electrical Discharge Machining KI03 Inner circle 1 EDM KI05 Inner circle 1 EDM KI07 Inner circle 2 EDM KI08 Inner circle 2 EDM

[0090] In this embodiment, the dataset provides a signal with a relatively high sampling frequency of 64kHz, meaning that the signal length acquired per second is 64,000. To facilitate the deep learning network in better learning the information contained in the signal and extracting fault features suitable for the task requirements, the original data signal is segmented into segments of 1024 data lengths using a sliding window.

[0091] The signal, with a length of 1024, is fed into the neural network model based on multi-scale feature adaptive blocks trained in Example 2 to extract the corresponding signal features. First, low-dimensional features are extracted through a wide convolutional layer. This low-dimensional feature establishes signal features in different dimensions through feature adaptive blocks with different convolutional kernel sizes. After all features have been extracted, they are merged to obtain multi-scale dimensional features. On the one hand, this ensures that the network can extract features from different dimensions, guaranteeing the globality of the constructed features; on the other hand, it fully utilizes the computational resources of each feature adaptation block, avoiding redundant computation for constructing multi-scale features and saving network training resources. The core feature adaptation block of this network implements an adaptive selection process for features from different domains. The feature adaptation block uses two different fully connected layers to separate source and target domain data, and combines a channel attention mechanism to adaptively learn feature map parameters that fit the source and target domains. These parameters are then applied to multi-scale dimensional features, enabling feature selection. This allows features beneficial to task analysis to receive more weight, improving the accuracy of the network's final diagnosis.

[0092] The loss function of this invention consists of four parts: classification-based cross-entropy loss, domain-alignment-based consistency loss, class-alignment-based consistency loss, and feature orthogonalization-based regularization loss. The classification-based cross-entropy loss calculates the network's classification loss using the diagnostic results of the source domain features and the actual source domain data labels. The domain-alignment-based consistency loss maps features to the same space by calculating distances within the Hilbert space of the extracted features. Similarly, the class-alignment-based consistency loss maps features to the same regenerated Hilbert space by calculating distances within the Hilbert space of the extracted features; compared to the domain-alignment-based consistency loss, this class-alignment-based loss achieves more accurate feature mapping. The feature orthogonalization-based regularization loss maximizes the network's ability to extract features by orthogonalizing domain and class features, applying them to different loss calculations to reduce their correlation and improve the network's diagnostic capabilities.

[0093] After the network training converges, its performance is evaluated based on the classification accuracy of the target domain signals. To verify the superiority of the proposed algorithm, it is compared with state-of-the-art network models in deep learning and baseline methods in transfer learning. The diagnostic accuracy is shown in Table 3, and the data visualization is as follows: Figure 4 As shown in Table 3, eight deep learning network models and five transfer learning baseline methods are compared. The data demonstrates that the proposed Multi-Scale Feature Adaptive Network (MFACN) outperforms the other six deep learning networks in all transfer learning methods, achieving top-ranking fault diagnosis accuracy across all approaches.

[0094] Table 3 Comparison Experiments of Deep Learning Networks and Transfer Learning Methods

[0095]

Claims

1. A cross-domain diagnosis method for mechanical faults based on multi-scale feature adaptive networks, characterized in that, The specific steps are as follows: Step 1: Obtain source domain data D s ={ x i , y i } and target domain data D t ={ x j , y j },in, x i For source domain data D s Data signals in y i for x i Corresponding data labels; x j For target domain data D t Data signals in y j for x j The corresponding data labels, among which, The source domain data and the target domain data are both bearing signal data under different operating conditions; Step 2: Construct a neural network model based on multi-scale feature adaptive blocks, including a wide convolutional layer, four feature adaptive blocks, a multi-scale feature concatenation layer, and a classifier connected in sequence. The output of each feature adaptive block is also connected to the multi-scale feature concatenation layer. The feature adaptive block includes a depthwise separable convolutional layer, a pointwise convolutional layer, a pointwise convolutional layer, and a global flat pooling layer connected in sequence. The global flat pooling layer is connected in parallel to a source domain fully connected layer and a target domain fully connected layer. The source domain fully connected layer and the target domain fully connected layer are jointly connected to an adaptive feature selection layer. The second pointwise convolutional layer is also connected to the adaptive feature selection layer. Step 3: Load the source domain data and target domain data from Step 1 into the neural network model based on multi-scale feature adaptive blocks for training, and obtain the trained neural network model. Step four: Use the trained neural network model to diagnose mechanical faults.

2. The cross-domain diagnosis method for mechanical faults based on multi-scale feature adaptive networks according to claim 1, characterized in that, The kernel size of the wide convolutional layer in step two is 31. The classifier comprises two fully connected layers of classifiers connected in series.

3. The cross-domain diagnosis method for mechanical faults based on multi-scale feature adaptive networks according to claim 2, characterized in that, The kernel sizes of the depth-separable convolutional layers in the first to fourth feature adaptive blocks are 25, 13, 7, and 3, respectively.

4. The cross-domain diagnosis method for mechanical faults based on multi-scale feature adaptive networks according to claim 1, characterized in that, Step three specifically involves: Step 3.1: Input the source domain data and target domain data into the neural network model based on multi-scale feature adaptive blocks and pass them through a wide convolutional layer with a kernel size of 31 to obtain the low-dimensional features corresponding to the source domain data and target domain data, respectively. and ; Step 3.2, convert the low-dimensional features obtained in Step 3.1 into... and By sequentially passing the feature adaptive blocks with four different convolutional kernel sizes, the following results are obtained: Features extracted for different sizes , , and Features extracted for different sizes , , ; Step 3.3: The features extracted from different sizes obtained in Step 3.2 are concatenated in a multi-scale feature concatenation layer to obtain multi-scale features of different dimensions. ； Step 3.4: The multi-scale features of different dimensions obtained in Step 3.3 are used in the classifier for classification training, domain transfer training, class transfer training, and feature orthogonalization regularization training, respectively. Step 3.5: Iteratively train the neural network model using source domain data and target domain data until convergence, and obtain the trained neural network model.

5. The cross-domain diagnosis method for mechanical faults based on multi-scale feature adaptive networks according to claim 4, characterized in that, The loss function trained in step 3.4 Loss as follows: in, Loss class For classification-based cross-entropy loss, Loss ad For consistency loss based on domain alignment, Loss ac For consistency loss based on category alignment, Loss d For regularization terms based on feature orthogonalization loss, λ, γ, and ρ are the weight coefficients of the consistency loss of domain alignment, consistency loss of class alignment, and regularization term of feature orthogonalization loss in the overall loss of network training, respectively, and their magnitudes range from 0 to 1.

6. The cross-domain diagnosis method for mechanical faults based on multi-scale feature adaptive networks according to claim 5, characterized in that, The classification-based cross-entropy loss Loss class as follows: in, For the proposed MFACN network, This represents features based on different dimensions across multiple scales; Consistency loss based on domain alignment Loss ad as follows: in, Indicates the first in the source domain i One sample, Indicates the first in the target domain j One sample, The 2-norm is used to represent the distance between samples from different domains, which is measured by Euclidean distance. Consistency loss based on class alignment Loss ac as follows: In this case, the target domain data is not used in the training. , Representing data fields in different source domains, Indicates category c Central Region The i One sample; Indicates category c Central Region The j One sample; Regularization term based on feature orthogonalization loss Loss d as follows: in, Represents the multi-scale features of the source domain. This represents the transpose of the multi-scale features of the target domain.

Images