Anti-noise transfer learning method based on MMSD measurement and time-frequency enhanced CNN model

Abstract

The invention relates to the technical field of mechanical vibration signal detection, in particular to an anti-noise transfer learning method based on MMSD measurement and a time-frequency enhanced CNN model, and the method comprises the steps: carrying out data collection and sample segmentation, dividing an original signal of collected data into three branches, carrying out the feature extraction of the data of the three branches through a GTFE-Net model, and carrying out the feature extraction of the data of the three branches; aggregating the characteristics of the special region; a CNN model of an end-to-end GNR is adopted as an MMSD measurement diagnosis model, the model is trained, learning parameters of the model are updated through an AdamW optimizer, and the model with the highest accuracy on the verification set is selected as a trained model; and inputting the test set into the well-trained fault diagnosis model, and carrying out multi-angle visual analysis on a diagnosis result. According to the invention, an end-to-end GNR-based GTFE-Net network structure is introduced to improve the performance and accuracy of one-dimensional signal processing in fault diagnosis and monitoring of rotating mechanical equipment, and a Gramian noise reduction strategy is embedded, so that noise and interference components in vibration signals are effectively weakened.

Description

Technical Field

[0001] The present application relates to the technical field of mechanical vibration signal detection, and in particular, to an anti-noise transfer learning method based on MMSD metric and time-frequency enhanced CNN model. Background Art

[0002] With the rapid development of the modern industrial field, the reliability and safety of rotating mechanical equipment have become the focus of attention. These devices play an important role in various application fields, such as wind turbines, aero-engines, and mining machinery. However, due to the complex working environment and high-speed operation, these rotating mechanical equipment often face various potential fault risks. Therefore, timely fault diagnosis and monitoring are crucial for ensuring the reliable operation of the equipment.

[0003] Through training a neural network model, deep learning can automatically learn and extract complex features in one-dimensional signals, and build an accurate model through learning large-scale data. However, due to the complexity of the actual environment and the diversity of equipment working conditions, the collected mechanical vibration signals are often mixed with irrelevant noise, posing challenges to signal processing. Summary of the Invention

[0004] The technical problem to be solved by the present invention is that the collected mechanical vibration signals are often mixed with irrelevant noise, affecting the accuracy of the detection results of the detection model.

[0005] To this end, the present invention provides an anti-noise transfer learning method based on MMSD metric and time-frequency enhanced CNN model.

[0006] The technical solution adopted by the present invention to solve its technical problems is as follows:

[0007] An anti-noise transfer learning method based on MMSD metric and time-frequency enhanced CNN model, comprising the following steps:

[0008] S1 Data acquisition and sample segmentation, dividing the original signal of the collected data into three branches, using the GTFE-Net model to extract features from the data of the three branches respectively, and aggregating the characteristics of the special zones;

[0009] S2 Using an end-to-end GNR CNN model as the MMSD metric diagnosis model, training the model, and updating the learning parameters of the model through the AdamW optimizer, and selecting the model with the highest accuracy on the validation set as the trained model;

[0010] S3 Inputting the test set into the well-trained fault diagnosis model, and performing multi-angle visualization analysis on the diagnosis results.

[0011] Furthermore, the GTFE-Net model combines the Gramian matrix in the spatial domain and the Fourier transform in the frequency domain to enhance feature extraction.

[0012] Furthermore, in step S1, the first branch of the original signal is directly input into the network model for feature extraction and representation; the second branch of the original signal undergoes FFT transformation and then passes through the GNR module for feature map conversion and enhancement; the third branch of the original signal also undergoes FFT transformation, aggregates with the signal obtained from processing the second branch, and then is input into the network model.

[0013] Furthermore, after the GTFE-Net model extracts different features from the original signals of the three branches through multiple convolutional blocks respectively, these features are subjected to adaptive average pooling and concatenation, and then classified through a fully connected layer.

[0014] Furthermore, in step S1, the GNR module converts the one-dimensional signal map into a two-dimensional feature map, then calculates the Gram enhanced feature map based on the two-dimensional feature map, and finally converts the Gram enhanced feature map of the two-dimensional feature map into a one-dimensional enhanced signal.

[0015] Furthermore, in step S2, an end-to-end GNR-based CNN model is used as the backbone of the MMSD metric diagnosis model, and the non-linear ReLU is used as the activation function.

[0016] Furthermore, in step S2, the training process uses the classification cross-entropy loss, the difference metric loss, and the adversarial loss to calculate the total loss of the training set, and the total loss is represented by the objective function based on the domain adaptation method:

[0017] Furthermore, the classification cross-entropy loss term is where G η is the feature extractor, x s represents the data in the source domain dataset D S , and n s is the number of samples in the source domain.

[0018] Furthermore, the difference metric loss term where x T represents the data in the target domain dataset D t , and n s and n t are the number of samples in the source domain and the target domain respectively, is the reproducing kernel Hilbert space, and the tensor product of two Hilbert spaces (H 1 , H 2 ) is still another Hilbert space φ(·) represents the mapping from the original samples to the kernel Hilbert space; k(·) is the kernel function, and generally the Gaussian kernel function is selected.

[0019] Furthermore, the adversarial loss term is where G μ represents the classifier; y i represents the true label of the i-th sample; p i is the probability of correctly predicting the label; L ad_loss is the adversarial loss.

[0020] The beneficial effects of the present invention are as follows. This application introduces an efficient end-to-end CNN model (GTFE-Net) network structure based on GNR to improve the performance and accuracy of one-dimensional signal processing in the fault diagnosis and monitoring of rotating machinery. Through deep learning methods, this application will train the GTFE-Net model using a large amount of experimental data, and through the analysis of one-dimensional signals, achieve accurate detection and prediction of mechanical equipment faults. This will provide important technical support for the maintenance and repair of rotating machinery, improve the reliability, safety, and efficiency of the equipment, and further promote the development of the industrial field.

[0021] This application introduces a new Gramian-based noise reduction strategy, aiming to effectively weaken the noise and interference components in the vibration signal. This strategy can not only significantly reduce the impact of useless noise, but also powerfully enhance the valuable feature information in the vibration signal through low-rank enhancement means.

[0022] GTFE-Net is different from traditional CNNs in structure. It uses three parallel branches to process the input signal and fuses the processed features into the network model. The design of this multi-branch structure enables the network to more fully capture the information interaction and dependency relationships between different branches, thereby improving the model's representation ability and performance. By processing the signal differently, each branch can focus on capturing different features or information of the signal, thereby enhancing the model's processing ability for complex signals.

[0023] The Maximum Mean Square Difference (MMSD) has achieved significant improvement in domain confusion by comprehensively reflecting the mean and variance information of data samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The present invention will be further described below with reference to the drawings and embodiments.

[0025] Figure 1 is a schematic flowchart of the anti-noise transfer learning method based on the MMSD metric and the time-frequency enhanced CNN model in the present invention.

[0026] Figure 2It is a schematic diagram of the framework of the GTFE-Net model in the present invention.

[0027] Figure 3 It is a schematic diagram of the structure of the GNR module in the present invention.

[0028] Figure 4 It is a schematic diagram of the structure of the MMSD metric diagnosis model in the present invention.

[0029] Figure 5 It is a schematic diagram of the experimental bench for collecting the CWRU dataset in Embodiment 1 of the present invention.

[0030] Figure 6 It is a schematic diagram of the confusion matrix of the CWRU dataset in a noisy environment in Embodiment 1 of the present invention.

[0031] Figure 7 It is a schematic diagram of the t-SNE graph of the CWRU dataset in a noise-free environment in Embodiment 1 of the present invention.

[0032] Figure 8 It is a schematic diagram of the confusion matrix of the CWRU dataset in a noise-free environment in Embodiment 1 of the present invention.

[0033] Figure 9 It is a schematic diagram of the experimental platform for collecting laboratory bearing data in Embodiment 2 of the present invention.

[0034] Figure 10 It is a schematic diagram of the gear fault types in Embodiment 2 of the present invention.

[0035] Figure 11 It is a schematic diagram of the confusion matrix of the laboratory dataset in a noisy environment in Embodiment 2 of the present invention.

[0036] Figure 12 It is a schematic diagram of the t-SNE of the laboratory dataset in a noise-free environment in Embodiment 2 of the present invention.

[0037] Figure 13 It is a schematic diagram of the confusion matrix of the laboratory dataset in a noise-free environment in Embodiment 2 of the present invention. Detailed implementation manners

[0038] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0039] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. In addition, the features defined as "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, unless otherwise specified, the meaning of "a plurality" is two or more.

[0040] In the description of the present invention, it should be noted that unless otherwise clearly defined and limited, the terms "mounted", "connected", and "connected" should be understood in a broad sense. For example, it may be a fixed connection, a detachable connection, or an integral connection; it may be a mechanical connection or an electrical connection; it may be directly connected or indirectly connected through an intermediate medium, and it may be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0041] Refer to Figures 1-13 , an anti-noise transfer learning method based on the MMSD metric and the time-frequency enhanced CNN model, comprising the following steps,

[0042] S1 Data acquisition and sample segmentation

[0043] The system vibration data is collected and transmitted to a digital device. Subsequently, the input raw signal is processed. The GNR and FFT modules are introduced into the GTFE-Net model to operate on the raw signal to capture more abundant feature information, and the raw signal is divided into three branches for processing.

[0044] GTFE-Net is a neural network model for signal processing, especially suitable for time series analysis. This network combines the Gramian matrix in the spatial domain and the Fourier transform in the frequency domain to enhance feature extraction. The model first normalizes the input signal, then converts the one-dimensional signal into a two-dimensional signal through the Gramian block, calculates the Gramian matrix, and then converts it back to a one-dimensional signal. At the same time, the Fourier block transforms the signal to the frequency domain through the fast Fourier transform and normalizes it. Next, the model processes the original signal, the Gramian-enhanced signal, and the Fourier-enhanced signal through multiple convolutional blocks respectively to extract different features. These features are pooled and concatenated adaptively, and then classified through a fully connected layer. GTFE-Net utilizes the information in the spatial and frequency domains, enhances the effect of feature extraction, and is suitable for complex signal processing tasks.

[0045] Although the design of the loss function plays an important role in enabling the feature extractor to capture domain-invariant and discriminative features in the source and target domains, the performance of the feature extractor itself also plays a crucial role. To mine more representative information in complex vibration signals, directly adopting a deeper network is an intuitive approach, but this may also bring some negative impacts, such as overfitting and slow convergence speed. To overcome the problem of difficult feature extraction in vibration, a GTFE-Net model with time-frequency domain enhancement is proposed. The model architecture of the GTFE-Net model is as Figure 1 shown.

[0046] Refer to Figure 2 , first, the first branch of the original signal is directly input into the network model for feature extraction and representation. Secondly, the second branch of the original signal undergoes FFT transformation to enhance the valuable feature information in the signal. The third branch of the original signal also undergoes FFT transformation and is aggregated with the signal obtained after the second branch is processed by FFT transformation and GNR processing, and then input into the network model. Finally, the feature maps extracted from the three channels are aggregated together.

[0047] By combining these features on different channels, this paper can obtain a more comprehensive and rich feature representation, improving the performance of the feature extractor. Specifically, this application introduces the GNR and FFT modules to operate on the original signal to capture more abundant feature information. By combining these features on different channels, this paper can obtain a more comprehensive and rich feature representation, improving the performance of the feature extractor. By introducing a multi-channel feature extraction structure similar to this, this paper aims to overcome the limitations in vibration signal feature extraction and can better capture domain-invariant features in the source and target domains. This method will provide a more powerful and robust feature representation for subsequent domain adaptation tasks, thereby enhancing the overall performance and generalization ability.

[0048] Among them, the working principle of GNR is as Figure 3 shown. First, the signal after FFT transformation is input into GNR. The operation of GNR is divided into three steps. The first step is to convert the one-dimensional signal map into a two-dimensional feature map. Then, the second step is to calculate the Gram enhanced feature map based on the two-dimensional feature map. Finally, the third step is to convert the two-dimensional enhanced feature map obtained in the second step into a one-dimensional enhanced signal. The design of this series of steps aims to effectively extract the key feature information of the signal and effectively remove noise and interference by enhancing the feature map, thereby improving the effect of signal processing.

[0049] The GNR is proposed to solve the deficiency of feature expression in convolutional neural networks, enabling the network to handle different types of features more flexibly. It turns the one-dimensional signal into a two-dimensional signal matrix, and through Gram enhancement processing, its time-frequency domain expression is strengthened, making its features more obvious. Finally, the two-dimensional signal matrix enhanced in the time-frequency domain is converted back into a one-dimensional signal, which can effectively improve the anti-noise performance of the model.

[0050] S2 Model Design and Model Training

[0051] This application introduces the concept of maximum mean square deviation difference, integrating the mean and variance information of data samples, successfully dealing with the confusion problem between different domains, and achieving remarkable progress. This comprehensive method provides a more thorough understanding of the distribution characteristics of data, accurately captures the differences between different domains, provides an effective way to solve the domain confusion problem, and is expected to further improve the model performance.

[0052] S21 Construction of MMSD Metric Diagnostic Model

[0053] Referring to Figure 4 , this application uses the non-linear ReLU as the activation function and focuses on unsupervised transfer learning, where the source domain samples are labeled while the target domain samples are not. By adopting the CNN model with end-to-end GNR as the backbone of the MMSD metric diagnostic model, this application can extract fault features and, through sub-domain adaptation, more accurately align and learn the fine-grained features of the source domain and the target domain while maintaining the overall distribution consistency, so as to improve the effect of the domain adaptation task.

[0054] S22 Model Training

[0055] The training process uses categorical cross-entropy loss, difference metric loss, and adversarial loss to calculate the total loss of the training set.

[0056] and respectively represent the source domain dataset D S and D t the target domain dataset, n s and nt are the number of samples in the source domain and the target domain respectively. Due to changes in working conditions, such as factors like environment, task settings, data collection methods, etc., the data distributions of the source domain D S and the target domain D t will be different. This difference leads to a distribution shift between the source domain and the target domain. Among them, the feature extractor G η extracts domain-invariant features F d =(G η (X S )∪G η (X T )) to alleviate the domain shift. MMD is one of the commonly used methods of statistical distance in domain adaptation. However, when MMD uses a non-feature kernel, it may lead to loss of information about the distribution. To solve this problem, this paper introduces a new distribution metric method - Maximum Mean Square Difference (MMSD), and MMSD has better transfer diagnosability. Therefore, the objective function based on the domain adaptation method usually includes the classification cross-entropy loss term the difference metric loss term λMMSD(G η (X S ),G η (X T )) and the adversarial loss term L ad_loss :

[0057]

[0058] When calculating the classification cross-entropy loss term , first find the values of the variables G f , G μ to minimize this expression, represents the cross-entropy loss between the predicted result of the input after passing through the model and the target label .

[0059]

[0060] Among them, is the reproducing kernel Hilbert space, and the tensor product of two Hilbert spaces (H 1 ,H 2 ) is still another Hilbert space φ(·) represents the mapping of the original sample to the kernel Hilbert space; k(·) is the kernel function, and generally the Gaussian kernel function is selected.

[0061]

[0062] Among them, G μ represents the classifier; y idenotes the true label of the i-th sample; p i is the probability of correctly predicting the label; L ad_loss is the adversarial loss.

[0063] S23 Update the model parameters

[0064] First, calculate the model output through forward propagation, then calculate the value of the loss function,接着进行反向传播计算梯度,最后使用AdamW优化器更新模型的参数,AdamW在更新时不仅考虑梯度信息,还添加了权重衰减项以防止过拟合,从而逐步优化模型性能。(原句中“接着进行反向传播计算梯度”表述有误,推测是想说“then perform backpropagation to calculate the gradients”,以下翻译按此修正)First, calculate the model output through forward propagation, then calculate the value of the loss function, then perform backpropagation to calculate the gradients, and finally use the AdamW optimizer to update the model parameters. AdamW not only considers the gradient information during the update but also adds a weight decay term to prevent overfitting, thereby gradually optimizing the model performance.

[0065] S3 Fault detection, diagnosis, and result analysis. By inputting the test set into a well-trained fault diagnosis model, perform multi-angle visual analysis on the diagnosis results.

[0066] Example 1

[0067] The following uses an example to illustrate the effectiveness and superiority of the anti-noise transfer learning method based on the MMSD metric and the time-frequency enhanced CNN model:

[0068] In this example, a public dataset provided by Case Western Reserve University (CWRU) is used to evaluate the diagnostic effectiveness of the method proposed in this paper. The public dataset is measured by the experimental bench as shown in Figure 5 The experimental bench consists of a 2 hp motor, a torque sensor / encoder, a dynamometer, and control electronics. This bearing dataset contains three fault types: inner race fault (IF), outer race fault (OF), and ball fault (BF). Each fault type contains three fault diameters of 7 mils, 14 mils, and 21 mils under different fault conditions. Therefore, this bearing dataset contains one normal state and eight fault states.

[0069] For the dataset of this experiment, bearing vibration signals with an acceleration sensor sampling frequency of 48 kHz and motor loads of 0 hp, 1 hp, 2 hp, and 3 hp are used. Four signals from different sensors are selected in the experimental samples, and the sample length is selected as 1024 data points. When dividing the data, 80% of the data is selected as the training set, and 20% of the dataset is the test set. In this experiment, gear data with a load of 1.5 A and rotation speeds of 900 rpm, 1200 rpm, and 1500 rpm are selected, corresponding to three transmission states of 0, 1, and 2 respectively.

[0070] Table 1 Diagnostic accuracy of each model on the CWRU dataset

[0071] Tasks Proposed EWSnet WDCNN MobileNet MCSwint 0-1 1.0000 0.8851 0.5402 1.0000 0.9885 0-2 1.0000 0.8966 0.5455 1.0000 0.9770 1-2 1.0000 1.0000 0.8415 0.9885 1.0000 1-0 1.0000 1.0000 0.7224 0.9770 0.9770 2-0 1.0000 0.7701 0.8636 0.9655 0.8736 2-1 1.0000 1.0000 0.5372 1.0000 0.9885

[0072] According to the data in Table 1, it can be clearly observed that the model proposed in this paper shows the highest accuracy in most transfer tasks. This means that the method in this paper has achieved excellent results on various tasks and datasets, and generally presents good and stable performance. This consistent performance shows the robustness and credibility of the method in this paper.

[0073] Comparison results of ablation study in Table 2

[0074]

[0075]

[0076] According to the results in Table 2, we can observe the performance of different models in a noisy environment with SNR = -2. In the experiment, we used four models: Model1 represents the method proposed in this paper, Model2 lacks GTFE-NET, Model3 lacks maximum mean square difference, and Model4 lacks both GTFE-NET and MMSD. It can be seen from the results that whether lacking the GTFE model or MMSD, the accuracy of the model has decreased to varying degrees.

[0077] Comparison results of ablation study in the network model in Table 3

[0078] Tasks 0 -2 -4 -6 Proposed 1.0000 1.0000 0.9885 0.9057 GTE 1.0000 0.9434 0.9245 0.8679 GT 1.0000 0.9434 0.8868 0.7925

[0079] Accuracy of the model in a noisy environment in Table 4

[0080] Tasks SNR = 0 SNR = -1 SNR = -2 SNR = -3 SNR = -4 SNR = -5 SNR = -6 0-1 1.0000 0.9811 0.9885 0.9655 0.9425 0.9057 0.9057 0-2 1.0000 1.0000 0.9885 0.977 0.954 0.8679 0.8491 1-2 1.0000 1.0000 1.0000 0.9811 0.9885 0.9623 0.9057 1-0 1.0000 0.9885 0.9885 0.977 0.9425 0.8868 0.8302 2-0 1.0000 0.9655 0.9623 0.9655 0.9195 0.8868 0.8491 2-1 1.0000 1.0000 1.0000 0.9885 0.9885 0.9434 0.8679

[0081] In further ablation experiments, the FFT module in GTFE was removed to obtain a network that only performs time-domain enhancement based on GNR, which we call GTE. Then, we removed the Gramian bridge from GNR in GTE to obtain a CNN model without time-domain enhancement, which we call GT. According to the results in Table 3, we found that in the noise-free case, the accuracy of both GTE and GT models reached 100%. However, once noise was added, especially in a strong noise environment, their accuracy began to gradually decline, even less than 90%. By observing Table 4 and Figure 6 , it can be concluded that the model proposed in this paper shows good performance in various noise intensity environments, indicating its reliability in transfer tasks. Among them Figure 6 the small sub-figures a-f in are the confusion matrices of the model proposed in this paper in noise environments with signal-to-noise ratios SNR of -1, -2, -3, -4, -5, and -6 respectively in various noise intensity environments.

[0082] Figure 7 and Figure 8respectively show the t-SNE results of the CWRU dataset in a noise-free environment and the confusion matrix of the CWRU dataset in a noise-free environment, where Figure 7 the small subgraphs a-f in Figure 8 are the confusion matrices of the model proposed in this paper in a noise-free environment with signal-to-noise ratios SNR of -1, -2, -3, -4, -5, and -6 respectively under various noise intensity environments; Figure 7 and Figure 8 the small subgraphs a-f in

[0083] Example 2

[0084] This example uses the laboratory bearing dataset to illustrate the effectiveness and superiority of the anti-noise transfer learning method based on the MMSD metric and the time-frequency enhanced CNN model:

[0085] The laboratory gear dataset contains 5 types of faults, including normal, tooth missing, spalling, pitting, and wear. The dataset is sampled at a frequency of 10 kHz under four different loads of 0.5A, 1.5A, 2.5A, and 3.5A. In this experiment, the gear data at a load of 1.5A and rotational speeds of 900 rpm, 1200 rpm, and 1500 rpm are selected, corresponding to three transfer states of 0, 1, and 2 respectively. 6 groups are randomly selected for each working condition in this paper for experimental verification. 6 groups of transfer experiments are designed and carried out on the NEFU dataset, and 6 transfer processes between 0, 1, and 2 are experimentally verified. Figure 9 is the data acquisition experimental platform, Figure 10 are pictures of various gear fault types.

[0086] Table 5 Diagnostic accuracies of each model of the laboratory dataset

[0087] Tasks Proposed EWSnet WDCNN MobileNet MCSwint 0-1 1.0000 1.0000 0.8864 0.8864 0.9773 0-2 1.0000 1.0000 0.5455 0.9318 1.0000 1-2 1.0000 1.0000 0.8636 0.9773 1.0000 1-0 1.0000 1.0000 0.8409 0.9545 1.0000 2-0 1.0000 0.9318 0.7273 0.9318 1.0000 2-1 1.0000 0.9773 0.75 0.9773 1.0000

[0088] Table 6 Comparison results of ablation studies

[0089] Tasks Model1 Model2 Model3 Model4 0-1 0.8864 0.7500 0.7955 0.5455 0-2 0.8409 0.6136 0.6818 0.5682 1-2 0.9773 0.6591 0.8182 0.6136 1-0 0.9318 0.7045 0.8409 0.6591 2-0 0.7955 0.5455 0.6591 0.5909 2-1 0.8864 0.5455 0.7045 0.5682

[0090] According to the results in Table 6, the performance of different models can be observed in a noise environment with SNR = -2. Among them, Model1 represents the method proposed in this paper, Model2 is the model lacking GTFE-NET, Model3 is the model lacking the maximum mean square difference, and Model4 is the model lacking both GTFE-NET and the maximum mean square difference. By observing the results, it is found that in this dataset, whether lacking GTFE or MMSD will lead to varying degrees of decline in the model accuracy.

[0091] Table 7 Comparison results of ablation studies in network models

[0092]

[0093]

[0094] According to the results in Table 7, the accuracies of the GTE and GT models are both 100% in the noise-free case. However, once noise is added to the original signal, their accuracies start to decline gradually. Especially in strong noise, the accuracy is less than 75%. Particularly in the GT model, its accuracy in a strong noise environment has dropped to less than 65%. By observing Table 8 and Figure 11 , where Figure 11 the small sub-figures a - f respectively show the confusion matrices of the laboratory dataset in noise environments with signal-to-noise ratios SNR of -1, -2, -3, -4, -5, -6 at each noise intensity level. We can draw the conclusion that the model proposed in this paper demonstrates good performance in noise environments at each noise intensity level, indicating its reliability in transfer tasks.

[0095] Figure 12 and Figure 13 respectively show the t-SNE results of the laboratory dataset in a noise-free environment and the confusion matrix of the CWRU dataset in a noise-free environment. Among them, Figure 12 the small sub-figures a - f respectively show the confusion matrices of the model proposed in this paper in noise environments with signal-to-noise ratios SNR of -1, -2, -3, -4, -5, -6 at each noise intensity level in a noise-free environment; Figure 13 the small sub-figures a - f respectively show the confusion matrices of 6 transfer processes of the laboratory dataset in a noise-free environment.

[0096] Table 8 Accuracy of the model in a noise environment

[0097] Tasks 0 -1 -2 -3 -4 -5 -6 0-1 1.0000 0.9091 0.8864 0.8409 0.8409 0.8182 0.7955 0-2 1.0000 0.8864 0.8409 0.8864 0.8409 0.7273 0.7243 1-2 1.0000 0.9861 0.9773 0.8409 0.875 0.8333 0.7955 1-0 1.0000 0.9318 0.9318 0.9318 0.8636 0.8182 0.7273 2-0 1.0000 0.9773 0.7955 0.7955 0.7727 0.7045 0.7045 2-1 1.0000 0.9091 0.8864 0.7955 0.9091 0.7955 0.7955

[0098] In summary, this paper proposes an anti-noise transfer learning method based on the MMSD metric and the time-frequency enhanced CNN model. First, a difference representation metric that can comprehensively reflect the mean and variance information of data samples in the reproducing kernel Hilbert space is constructed to improve the distribution consistency between the target domain and the source domain. Second, in order to further improve the model performance, the GTFE-Net model is introduced, which can extract rich and transferable features in the signal. Finally, through the architecture fusion and verification of two datasets, the effectiveness of the proposed method is verified. These improvements enable the method in this paper to achieve remarkable results in unsupervised cross-domain fault diagnosis.

[0099] Inspired by the above ideal embodiments of the present invention, through the above description, relevant staff can completely make various changes and modifications without departing from the technical idea of the present invention. The technical scope of the present invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A noise-resistant transfer learning method based on MMSD metric and time-frequency enhanced CNN model, characterized in that: The following steps are included: S1 data collection and sample segmentation: the original signal of the collected data is divided into three branches, and the GTFE-Net model is used to extract features from the data of the three branches respectively, and then the characteristics of the special zone are aggregated; S2 uses the end-to-end GNR CNN model as the MMSD metric diagnosis model, trains the model, and updates the model's learning parameters through the AdamW optimizer, and selects the model with the highest accuracy on the validation set as the trained model; S3 inputs the test set into a well-trained fault diagnosis model and performs multi-angle visual analysis of the diagnosis results.

2. The noise-resistant transfer learning method based on MMSD metric and time-frequency enhanced CNN model according to claim 1, characterized in that: The GTFE-Net model combines the Gramian matrix in the spatial domain and the Fourier transform in the frequency domain to enhance feature extraction.

3. The noise-resistant transfer learning method based on MMSD metric and time-frequency enhanced CNN model according to claim 2, characterized in that: In step S1, the first original signal is directly input into the network model for feature extraction and representation; the second original signal is transformed by FFT, and then the feature map is converted and enhanced through the GNR module; the third original signal is also transformed by FFT, aggregated with the second processed signal, and then input into the network model.

4. The noise-resistant transfer learning method based on MMSD metric and time-frequency enhanced CNN model according to claim 3, characterized in that: The GTFE-Net model extracts different features from the original signals of the three branches through multiple convolution blocks, and then classifies these features through a fully connected layer after adaptive average pooling and splicing.

5. The noise-resistant transfer learning method based on MMSD metric and time-frequency enhanced CNN model according to claim 3, characterized in that: In step S1, the GNR module converts the one-dimensional signal map into a two-dimensional feature map, then calculates the Gram enhanced feature map based on the two-dimensional feature map, and finally converts the Gram enhanced feature map of the two-dimensional feature map into a one-dimensional enhanced signal.

6. The noise-resistant transfer learning method based on MMSD metric and time-frequency enhanced CNN model according to claim 1, characterized in that: In step S2, the end-to-end GNR CNN model is used as the backbone of the MMSD metric diagnosis model, and the nonlinear ReLU is used as the activation function.

7. The noise-resistant transfer learning method based on MMSD metric and time-frequency enhanced CNN model according to claim 1, characterized in that: In step S2, the training process uses classification cross entropy loss, difference metric loss and adversarial loss to calculate the total loss of the training set, and the total loss is expressed by the objective function based on the domain adaptation method:

8. The noise-resistant transfer learning method based on MMSD metric and time-frequency enhanced CNN model according to claim 7, characterized in that: The categorical cross entropy loss term is Among them, G η is the feature extractor, x s Represents the source domain dataset D S The data in n s The number of samples in the source domain.

9. The noise-resistant transfer learning method based on MMSD metric and time-frequency enhanced CNN model according to claim 7, characterized in that: The difference metric loss term where x T Represents the target domain dataset D t The data in n s and n t are the number of samples in the source domain and the target domain respectively, To reproduce the kernel Hilbert space, the tensor product of two Hilbert spaces (H1, H2) is still another Hilbert space φ(·) represents the mapping from the original sample to the kernel Hilbert space; k(·) is the kernel function, and the Gaussian kernel function is generally used.

10. The noise-resistant transfer learning method based on MMSD metric and time-frequency enhanced CNN model according to claim 7, characterized in that: The adversarial loss term is Among them G μ represents the classifier; y i represents the true label of the i-th sample; p i is the probability of predicting the correct label; L ad_loss To combat losses.

Images