Multi-source domain approximation fault diagnosis method combining real data and mechanism data for fluctuating rotating speed working condition

Abstract

A multi-source domain approximation fault diagnosis method combining real data and mechanism data for fluctuating rotating speed working conditions is mainly characterized in that a dynamic multi-source domain migration strategy is designed by fusing bearing mechanism model data and real data of large fluctuating rotating speed working conditions, and a maximum difference source domain between domains is iterated and approximated by using distance measurement distance; and label space alignment is realized in combination with the minimum class confusion loss, so that the problems of feature drift and label data scarcity caused by rotating speed fluctuation are solved, the cross-domain fault diagnosis precision is remarkably improved, the recognition rate and robustness of bearing faults are improved, and effective diagnosis of the bearing faults under the rotating speed fluctuation is realized.

Description

Technical Field

[0001] The present invention relates to a multi-source domain approximation fault diagnosis method that combines real data and mechanism data for fluctuating rotational speed conditions, and belongs to the technical field of fault diagnosis of rotating machinery vibration signals. Background Art

[0002] With the rapid development of industrial automation and intelligent manufacturing, bearings, as key load-bearing components in rotating machinery, their health status is crucial for the safe operation of equipment. However, in practical applications, due to the diversity of factors such as working conditions, loads, environments, and sensor installation positions, the operating data of bearings often shows significant distribution differences. This inconsistency in distribution makes traditional fault diagnosis models that rely on training in a single data domain prone to problems such as a decrease in recognition rate and insufficient robustness when migrating to new working environments or equipment, resulting in difficulty in effectively diagnosing bearing faults under fluctuating rotational speeds. Summary of the Invention

[0003] The object of the present invention is to provide a multi-source domain approximation fault diagnosis method that combines real data and mechanism data for fluctuating rotational speed conditions, which can effectively diagnose bearing faults under fluctuating rotational speeds.

[0004] To achieve the above object, the present invention provides a multi-source domain approximation fault diagnosis method that combines real data and mechanism data for fluctuating rotational speed conditions, including the following steps:

[0005] S1. Set different bearing health states, manually adjust the frequency converter to randomly vary the motor speed within the set range to form a rotational speed fluctuation condition, and at the same time collect the vibration signals of the bearing under different working conditions;

[0006] S2. Based on Vision-Transformer, divide the mechanism model data at constant rotational speed and the real data at fluctuating rotational speed into time-domain signals with a length of 2400, and perform a fast Fourier transform; the first half of the frequency-domain signals after the fast Fourier transform is used as the samples input to the network; to obtain the signal sequence embedding of the input network, divide the frequency-domain signal samples with a length of 1200 and perform a linear mapping; add Tokens to the signal sequence embedding and add it to the positional signal sequence embedding;

[0007] S3. Extract features from the mechanism data and the actual data through the Vision Transformer module;

[0008] S4. After the feature extraction is completed, take out the initially added learnable Class Tokens and input them into the classifier for classification.

[0009] Further, the bearing health states set in S1 include four categories: normal, inner ring fault, outer ring fault, and rolling element fault; manually adjust the frequency converter to randomly change the motor speed within 500 - 3000 rpm; collect the vibration signals of the bearing under normal, inner ring fault, outer ring fault, and rolling element fault conditions respectively.

[0010] Further, the formula for the fast Fourier transform in S2 is:

[0011]

[0012] where, x n is the time-domain signal, x k is the frequency-domain signal, N is the length of the signal, and i is the imaginary unit;

[0013] The linear mapping is performed using a convolutional neural network, and the formula is:

[0014]

[0015] where, z0 is the embedded sequence after linear mapping, and x is the segmented signal fragment;

[0016] The formula for adding Tokens to the signal sequence embedding and adding it to the position signal sequence embedding is:

[0017]

[0018] where, f is the input of the feature extractor, x class is the added Class Tokens, is the signal sequence embedding of the Nth segment, and E pos is the position signal sequence embedding.

[0019] Further, the specific process of S3 is:

[0020] To achieve the approximation of the multi-source domain to the target domain, during the feature extraction process, the Wasserstein distance metric is performed on the tokens between multiple source domains and the target domain in each layer. It should be noted that the network used in this paper is different from the traditional VIT that takes pictures as input. In this paper, the one-dimensional frequency-domain signal is used as the network input, which changes the encoding method of the two-dimensional position embedding of the traditional VIT. The processed signal sequence embedding is input into the Vision-Transformer Blocks for feature extraction. The Transformer Block includes layer normalization LN, multi-layer perceptron MLP, and multi-head attention mechanism. The feature extraction process is represented by the following formula:

[0021] f l' = MSA(LN(f l-1 )) + f l-1 ;

[0022] f l = MLP(LN(f l ')) + f l ';

[0023] Among them, f l is the output of the l-th Vision transformer module; f l ' is the output of the multi-head attention mechanism.

[0024] Furthermore, the specific process of S4 is as follows:

[0025] S4.1. After the feature extraction is completed, the initially added learnable Class Tokens are taken out and input into the classifier for classification. This is because, during the feature extraction process, the Class Tokens can interact with the global features, which enables the Class Tokens to better reflect the global information of the signal. Moreover, the Class tokens are further optimized in each layer. When the feature extraction is completed, the Class Tokens become high-dimensional vectors representing the global information of the sample. At the same time, in order to achieve domain alignment between each source domain and the target domain, the distance metric between multiple source domains and the target domain is also performed in each layer. The distance metric formula is as follows:

[0026]

[0027] Among them, (P r , P g ) is the set of all joint distributions γ(x, y), and the marginal probabilities of γ(x, y) are P r and P g ;

[0028] S4.2. While the feature extraction is in progress, the multi-source domains are approximated and aligned layer by layer. In this process, the two source domains with the largest difference from the target domain are aligned. Compared with the method of aligning with all source domains simultaneously, this strategy enables the method to approximate all source domains in the correct direction. The calculation process of the multi-source domain approximation loss is as follows:

[0029]

[0030] Among them, S is the source domain feature, and y is the target domain feature; the multi-source domain approximation loss is L ds , is the feature of the i-th sample of a single source domain, n is the number of samples, S', S'' are the two source domains for domain approximation in each iteration, is the feature of the i-th sample in two source domains;

[0031] S4.3. After prediction by the label classifier, the logits of the target domain are calculated for class confusion loss, and the calculation process of class confusion loss is as follows:

[0032] S4.3-1. To alleviate the negative impact caused by the overconfidence of the network, the logits of the target domain are temperature recalibrated:

[0033]

[0034] where, is the result of the re-calibrated probability that the i-th instance belongs to the j-th class, Z ij is the output of the classifier, |c| is the number of source domain classes, and T is a hyperparameter; Z ij′ is the classifier prediction of the j'-th class of the i-th instance;

[0035] S4.3-2. Evaluate the pairwise correlation between fault classes to reveal the correlation between classes:

[0036]

[0037] where, C jj′ is the correlation between the j-th class and the j'-th class; represents the probability that the examples in each batch come from the j-th class;

[0038] S4.3-3. Evaluate the uncertainty weight, which is used to measure class confusion and perform class correlation annotation:

[0039]

[0040] where, is the uncertainty of the i-th sample; W ii is used to quantify the importance of the i-th instance for class confusion modeling, B is the sample size, and W is the corresponding diagonal matrix; thus, a preliminary characterization of class confusion is obtained:

[0041]

[0042] S4.3-4. Standardize the prediction with class correlation annotation and calculate the class confusion loss, and the formula is as follows:

[0043]

[0044] S4.3-5. Feature extraction is performed layer by layer on the physical model data of different constant rotational speed conditions, and the Class Tokens of the domain-invariant features reflecting bearing faults under various conditions obtained through feature extraction are input into the classifier, thereby obtaining the classification prediction results of the source domain samples; the Softmax cross-entropy loss is used to evaluate the cross-entropy between the predicted probability and the actual sample data, and the obtained label classification loss L C is superimposed with other losses and finally completes the training of the model through optimization calculation. The label classification loss is defined as follows:

[0045]

[0046] where, is the indicator function, is the k-th value of the prediction distribution, and the number of label types is defined as K; x s represents the output of the source domain classifier, represents the source domain sample label; and y s are the source domain features and labels respectively;

[0047] S4.3-6. In order to enable the model to obtain richer multi-source domain features, the Class Tokens of the constant rotational speed condition features from multiple source domains are respectively used to measure the distance from the Class Tokens of the target domain fluctuating rotational speed features. The sum of the distances of the Class Tokens after feature extraction is used as the difference amount between the multiple source domains and the target domain, and the obtained distance loss L wd is superimposed with other losses and finally completes the training of the model through optimization calculation. The distance loss is defined as:

[0048]

[0049] where, is the output of the feature extractor.

[0050] By fusing the bearing mechanism model data and the real data of the large fluctuating rotational speed conditions, designing a dynamic multi-source domain migration strategy, using distance measurement to iteratively approximate the maximum difference source domain between domains, and combining the minimum class confusion loss to achieve label space alignment, the present invention solves the problems of feature drift caused by rotational speed fluctuations and scarce labeled data, significantly improves the cross-domain fault diagnosis accuracy, increases the recognition rate and robustness of bearing faults, and realizes the effective diagnosis of bearing faults under rotational speed fluctuations. Brief Description of the Drawings

[0051] Figure 1 is the working flow chart of the method of the present invention;

[0052] Figure 2It is a bar chart of the diagnostic accuracy of the existing method and the present invention in the transfer task in the embodiments of the present invention;

[0053] Figure 3 It is a box plot and distribution plot of the diagnostic accuracy of multiple times in the task in the embodiments of the present invention;

[0054] Figure 4 It is a characteristic point distribution curve graph of the source domain and target domain of different faults in the embodiments of the present invention:

[0055] Figure 5 It is a confusion matrix graph of the classification results in the transfer task in the embodiments of the present invention;

[0056] Figure 6 It is a diagnostic accuracy result graph of the existing technology and the present invention in the transfer task in the embodiments of the present invention;

[0057] Figure 7 It is a dimensionality reduction graph of the output of the feature extractor in the embodiments of the present invention;

[0058] Figure 8 It is a confusion matrix graph drawn according to the classification results of the embodiments of the present invention. Detailed implementation manners

[0059] The present invention will be further described below with reference to the accompanying drawings.

[0060] As Figure 1 shown, a multi-source domain approximation fault diagnosis method for combined real data and mechanism data under fluctuating rotational speed conditions includes the following steps:

[0061] S1. Set different bearing health states, manually adjust the frequency converter to make the motor speed randomly change within the set range to form a rotational speed fluctuation condition, and at the same time collect the vibration signals of the bearing under different conditions;

[0062] S2. Based on Vision-Transformer, divide the mechanism model data at constant rotational speed and the real data at fluctuating rotational speed into time-domain signals with a length of 2400, and perform fast Fourier transform; the first half of the frequency-domain signals after the fast Fourier transform is used as the samples input to the network; to obtain the signal sequence embedding of the input network, divide the frequency-domain signal samples with a length of 1200 and perform linear mapping; add Tokens to the signal sequence embedding and add it to the position signal sequence embedding;

[0063] S3. Extract features from the mechanism data and actual data through the Vision Transformer module;

[0064] S4. After the feature extraction is completed, take out the initially added learnable Class Tokens and input them into the classifier for classification.

[0065] In this embodiment, a vibration signal is collected and intelligently diagnosed for a specially designed bearing test bench under rotational speed fluctuations, and the test bench data and the deep groove ball bearing dynamics model are used to verify the effectiveness of the present invention.

[0066] During data collection, an acceleration sensor was installed directly above the bearing housing, and an LMS signal tester was used to collect acceleration signals. The sampling frequency of both sets of data was 12.8 kHz. The dynamics model data set included a total of 7 operating conditions: 500 rpm (V1), 1000 rpm (V2), 1500 rpm (V3), 1800 rpm (V4), 2000 rpm (V5), 2500 rpm (V6), 3000 rpm (v7); the test bench data included 8 operating conditions: 500 rpm (R1), 1000 rpm (R2), 1500 rpm (R3), 1800 rpm (R4), 2000 rpm (R5), 2500 rpm (R6), 3000 rpm (R7), and a time-varying rotational speed condition fluctuating within the range of 500 - 3000 rpm (R8); the dynamics model data was used as the source domain, and the test bench data was used as the target domain; the dynamics model included 4 types of fault: normal (NC), inner ring fault 0.2 mm (IF0.2), outer ring fault 0.2 mm (OF0.2), inner ring fault 0.4 mm (IF0.4 mm); the test bench data included six types of faults: NC, IF0.2, IF0.4, OF0.2, rolling element fault 0.2 mm (RF0.2), rolling element fault 0.4 mm (RF0.4); the details of the experimental tasks are shown in Table 1:

[0067] Table 1

[0068]

[0069] Three types of fault diagnosis methods were used for comparison with the present invention, including closed-set domain adaptation methods, namely domain adversarial neural network (DANN) and multi-source adversarial network (MDAN); a partial-set domain adaptation method based on class-weighted alignment transfer network (CWATN); and an unknown fault diagnosis method called universal domain adaptation method (UDAM). The diagnostic accuracies of the above methods and the method described in the invention in the transfer task are as Figure 2 shown. It can be seen that when the transfer task is closed-set domain adaptation, the above methods all have relatively high diagnostic accuracies. When performing unknown fault diagnosis, the diagnostic accuracies of DAAN, MDAN, and CWATN all decreased. In the transfer task, compared with other methods, the method shown in the present invention can maintain more stable and higher diagnostic accuracies.

[0070] As Figure 3The box plots and distribution diagrams of the multiple diagnostic accuracies of the method under five tasks are shown. A total of five migration tasks were set up in the experiment, and multiple repeated experiments were conducted. It can be found that the results of most repeated experiments in the five tasks were around 98%-98.5%, and a few experimental results were lower than 97%, but still higher than 95.5%.

[0071] In order to verify the ability of the present invention to perform cross-domain feature extraction on the target domain through multiple source domains, the outputs of the feature extractors of tasks 1, 3, and 6 are visualized by dimension reduction. Figure 4 As shown in FIG. 1 , the present invention sets the sample points of different subdomains of the same fault to the same shape and the same color, and uses the distribution curve to represent the distribution of the feature points of the source domain and the target domain of different faults. By observing the distribution curve and the scatter plot, it can be found that Figure 4 In the closed set tasks shown in (a) and (d), most of the source and target domains can be well aligned. Although the OF 0.4 feature distribution curve is offset, its sample points can still be well aligned and clustered. The offset of the OF 0.4 feature distribution curve may be due to the confusion of a few sample points between OF 0.4 and IF 0.2. Figure 4 In the partial set experiments shown in (b) and (d), the present invention can well align the corresponding sample clusters in the source domain and the target domain, and the sample clusters of various types of faults in the source domain can be clustered normally. Figure 4 In the open set tasks shown in (c) and (d), unknown faults are clustered into very small sample clusters. This may be because the classifier is added with unknown class output bits when the method is used to identify unknown faults, and the source domain is filled with blank samples. It can be observed that known faults and unknown faults can be clearly separated, and the sample clusters of known fault samples can be correctly clustered.

[0072] like Figure 5 (a) to (e) show the confusion matrices of the classification results of the present invention in five migration tasks. There are five types of faults, namely 1, 2, 3, 4, and 5, namely NC, IF0.2, OF0.2, RF0.2, and IF0.4. It can be seen that in tasks 1 and 4, the present invention can correctly classify most samples, but a small amount of confusion is prone to occur in the classification of OF0.2 and RF0.2. In task 2, a small number of OF0.2 are still misclassified as RF0.2, and one sample is misclassified into the fifth category. In task 3, there are individual samples of the three types of faults in the target domain that are misclassified. In task 5, it can be found that OF0.2 is easily confused with unknown class faults. In summary, it can be seen that OF0.2 and RF0.2 have classification confusion in multiple tasks, which shows that the representative feature extraction ability of the present invention for these two groups of samples needs to be improved. Although the present invention has misclassification of a very small number of samples, it can accurately classify most samples.

[0073] In order to verify the effectiveness of the present invention in cross-domain fault diagnosis of mechanism data and real data, three different virtual data item real data migration tasks are set. The specific settings of the three migration tasks are shown in Table 2:

[0074] Table 2

[0075]

[0076]

[0077] The diagnostic accuracy results of the five methods in three migration tasks are shown in Figure 2. Figure 6 As shown. It can be seen that the diagnostic accuracy of MDAN is lower than that of the other three methods, which shows that in the adversarial network with multi-source domain adaptation, the domain differences between multiple sets of mechanism model data and real data bring great difficulties to the feature extraction of the network. In Task 2, the diagnostic accuracy of MDAN is still lower than that of the other four methods. At the same time, it can be found that the diagnostic accuracy of DANN has decreased. In Task 3, the diagnostic accuracy of DANN continues to decrease and the diagnostic accuracy of MDAN is higher than that of DANN. This is because the difference in label space makes it more difficult for the network to achieve domain alignment. In addition, by observing the diagnostic accuracy bar chart, it can be found that the diagnostic accuracy of CWATN and UDAN is higher than that of the first two methods. Because the diagnostic labels CWATN and UDAN optimize the network for the problem of differences in the spatial distribution of source and target domain labels. Finally, the comparative experimental results of the three migration tasks show that the present invention has higher diagnostic accuracy than the other four methods, which also verifies the effectiveness of the present invention.

[0078] The feature extractor output dimensionality reduction diagrams of the three tasks are as follows Figure 7 As shown. By observing Figure 7 The dimensionality reduction diagram of the closed set task shown in (a) shows that various samples can be well clustered. However, some target domain sample points of NC are outside the sample cluster. This may be because the characteristics of normal bearings are not obvious and under the influence of other vibration sources and speed fluctuations of the equipment, the feature distribution of some samples is easily different from that of other samples. Figure 7 In the partial cluster shown in (b), the source domain and the target domain can achieve good domain alignment, and a small number of NC samples in the target domain are still outside the sample cluster. Figure 7In the open set task shown in (c), it can be found that unknown class faults can be well separated, and most samples can be well clustered. It is worth noting that in the three groups of tasks, there are mostly multiple sample clusters in the source domain samples, and the target domain sample points will align with some of the sample clusters. This problem is caused by the existence of multiple constant rotational speed working conditions in the source domain. Generally speaking, the feature clustering and alignment effects in this part are not as good as those in Experiment 1 because there is a greater difference between the mechanism model data and the real data. However, although there are a small number of sample points outside the sample clusters in all three tasks, most of the sample clusters can still be well clustered and aligned. This is sufficient to prove the effectiveness of the present invention.

[0079] To further show the classification situation of the method described in the present invention in different tasks, the classification results of the present invention are plotted into a confusion matrix. Among them, 1, 2, 3, and 4 represent the four types of faults NC, IF0.2, OF0.2, and IF0.4 respectively. It can be found that in Figure 8 the closed set task shown in (a), the classification accuracy of the present invention for the first three types of samples can reach more than 95%, and the classification accuracy for IF0.4 needs to be improved. In Figure 8 the partial set domain adaptation shown in (b), NC and IF0.2 are easily misclassified as OF 0.2. In the open set domain adaptation, the classification accuracy of IF0.2 is higher than that of the other three types of samples, and the samples of NC and OF0.2 are easily confused with the samples of IF0.4. Compared with the first two tasks, the diagnostic accuracy of the present invention decreases in the Figure 8 open set domain adaptation task shown in (c), but the overall accuracy can still reach more than 90%.

Claims

1. A multi-source domain approximation fault diagnosis method combining real data and mechanism data for fluctuating rotational speed conditions, characterized in that It includes the following steps: S1. Set different bearing health states, manually adjust the frequency converter to make the motor speed randomly change within the set range to form a speed fluctuation working condition, and at the same time collect the vibration signals of the bearing under different working conditions; S2. Based on Vision-Transformer, divide the mechanism model data with constant speed and the real data with fluctuating speed into time-domain signals with a length of 2400, and perform fast Fourier transform; the first half of the frequency-domain signals after the fast Fourier transform is used as the sample input to the network; to obtain the signal sequence embedding of the input network, divide the frequency-domain signal samples with a length of 1200 and perform linear mapping; add Tokens to the signal sequence embedding and add it to the position signal sequence embedding; S3. Extract features from the mechanism data and the actual data through the Vision Transformer module; S4. After the feature extraction is completed, take out the initially added learnable Class Tokens and input them into the classifier for classification.

2. The multi-source domain approximation fault diagnosis method for combined real data and mechanism data under fluctuating rotational speed conditions according to claim 1, characterized in that The bearing health states set in S1 include four categories: normal, inner ring fault, outer ring fault, and rolling element fault; manually adjust the frequency converter to make the motor speed randomly change within 500 - 3000 rpm; collect the vibration signals of the bearing under normal, inner ring fault, outer ring fault, and rolling element fault working conditions respectively.

3. The multi-source domain approximation fault diagnosis method for combined real data and mechanism data under fluctuating rotational speed conditions according to claim 1, characterized in that, The formula for the fast Fourier transform in S2 is: where x n is the time-domain signal, x k is the frequency-domain signal, N is the length of the signal, and i is the imaginary unit; The linear mapping is to perform linear mapping using a convolutional neural network, and the formula is: where z0 is the embedded sequence after linear mapping, and x is the segmented signal segment; The formula for adding Tokens to the signal sequence embedding and adding it to the position signal sequence embedding is: Among them, f is the input of the feature extractor, and x class is the added class Tokens, is the signal sequence embedding of the Nth segment, E pos is the position signal sequence embedding.

4. The multi-source domain approximation fault diagnosis method for combined real data and mechanism data under fluctuating rotational speed conditions according to claim 1, characterized in that The specific process of S3 is: Input the processed signal sequence embedding into Vision-Transformer Blocks for feature extraction. The Transformer Block includes layer normalization LN, multi-layer perceptron MLP, and multi-head attention mechanism. The feature extraction process is represented by the following formula: f l ′ = MSA(LN(f l-1 )) + f l-1 ; f l = MLP(LN(f l ′)) + f l ′; Among them, f l is the output of the l-th Vision transformer module; f l ' is the output of the multi-head attention mechanism.

5. The multi-source domain approximation fault diagnosis method for combined real data and mechanism data under fluctuating rotational speed conditions according to claim 1, characterized in that The specific process of S4 is: S4.

1. When the feature extraction is completed, the Class Tokens become high-dimensional vectors representing the global information of the samples. At the same time, in order to achieve domain alignment between each source domain and the target domain, distance metrics are also performed between multiple source domains and the target domain at each layer. The distance metric formula is as follows: where (P r , P g ) is the set of all joint distributions γ(x, y) whose marginal probabilities are P r and P g ; S4.

2. During the feature extraction, gradually approximate and align multiple source domains. The calculation process of the multi-source domain approximation loss is as follows: where S is the source domain feature and y is the target domain feature; the multi-source domain approximation loss is L ds , is the feature of the i-th sample of a single source domain, n is the number of samples, S′ and S″ are the two source domains for domain approximation in each iteration, are the features of the i-th sample of the two source domains; S4.

3. After prediction by the label classifier, the logits of the target domain are calculated for class confusion loss. The calculation process of the class confusion loss is as follows: S4.3-1. To alleviate the negative impact caused by the network being overconfident, the logits of the target domain are subjected to temperature recalibration: wherein, is the result after re - calibrating the probability that the i - th instance belongs to the j - th class, Z ij is the classifier output, |c| is the number of source - domain classes, and T is a hyperparameter; Z ij′ is the classifier prediction for the j'-th class of the i - th instance; S4.3-2. Evaluate the pairwise correlation between fault categories: where C jj′ is the correlation between class j and class j'; represents the probability that an example in each batch comes from the j-th class; S4.3-3. Evaluate the uncertainty weight: Among them, is the uncertainty of the i-th sample; W ii is used to quantify the importance of the i-th instance for class confusion modeling, B is the sample size, and W is the corresponding diagonal matrix; thus, a preliminary characterization of out-of-class confusion is obtained: S4.3-4. Perform normalization processing and class confusion loss calculation on the prediction marked with class correlation. The formula is as follows: S4.3-5. Feature extraction is performed layer by layer on the physical model data of different constant rotational speed conditions, and the Class Tokens of domain-invariant features reflecting bearing faults under various conditions obtained through feature extraction are input into the classifier to obtain the classification prediction results of source domain samples; the Softmax cross-entropy loss is used to evaluate the cross-entropy between the predicted probability and the actual sample data, and the obtained label classification loss L C is superimposed with other losses and finally completes the training of the model through optimization calculation. The label classification loss is defined as follows: Among them, is an indicator function, is the k-th value of the predicted distribution, and the number of label types is defined as K; x s represents the output of the source domain classifier, represents the source domain sample label; and y s are the source domain feature and label respectively; S4.3-6. The sum of the distances of the Class Tokens after feature extraction is used as the difference amount between the multi-source domain and the target domain, and the obtained distance loss L wd is superimposed with other losses and finally completes the training of the model through optimization calculation. The distance loss is defined as: Among them, is the output of the feature extractor.

Images