Fault diagnosis method for harmonic reducer based on DANN with variable working conditions and few labels

Abstract

The present invention discloses a method for fault diagnosis of harmonic reducers with variable working conditions and few labels based on DANN. The method comprises: collecting vibration signals of the harmonic reducer; performing aggregation and scaling using a segmented aggregation approximation method; converting using a coordinate transformation method and performing windowed interception sampling; obtaining a GASF image after processing using a two-angle difference cosine function; inputting a training set into a label generation network for training; inputting an unlabeled data image into a processed and generated labeled data image; inputting a source domain and a target domain into a domain adversarial transfer network for training; collecting real-time vibration signals, obtaining real-time GASF images, and sequentially inputting the trained label generation network and the domain adversarial transfer network into the processed network to determine the load working condition and realize fault diagnosis of the harmonic reducer. The method of the present invention can provide a more comprehensive representation for feature extraction, solve the problem of insufficient source domain labels, and is conducive to effective and accurate fault diagnosis results when the working conditions undergo major changes.

Description

Technical Field

[0001] The present invention relates to a speed reducer fault diagnosis method, and in particular to a speed reducer fault diagnosis method based on DANN variable working condition and few-label harmonic wave. Background Art

[0002] Harmonic reducers are widely used in aerospace, engineering machinery, military equipment, wind power generation, communication equipment and many other fields. They are mainly composed of four basic components: a wave generator, a rigid gear (flexible gear) with an internal gear ring, a flexible gear (rigid gear) with an external gear ring and a flexible bearing. Due to its precise structural composition, it has the advantages of large transmission ratio, high load-bearing capacity, high transmission accuracy, high transmission efficiency, smooth movement, simple structure and the ability to transmit motion to confined spaces. Especially in high dynamic performance and high-precision transmission systems, the use of harmonic gear transmission shows its superiority. Due to the importance and irreplaceable nature of harmonic reducers in actual industrial applications, fault diagnosis and condition monitoring of harmonic reducers are carried out to detect, diagnose and predict possible faults in order to "prevent problems before they occur". This is of great practical need and practical significance for ensuring the reliable, continuous and stable operation of equipment and avoiding major accidents.

[0003] A common approach to fault diagnosis under varying operating conditions is to use vibration signals with labeled health status. This approach extracts intrinsic features from vibration signals through deep neural networks, and uses them to narrow the distribution gap between the source and target domains to learn domain-invariant features of signal samples, enabling fault diagnosis under varying operating conditions. However, these approaches assume that vibration signals must have corresponding health status labels, which is inconsistent with actual conditions. In practice, a large number of vibration signals are often collected, but their health status cannot be labeled. This makes conventional fault diagnosis methods unsuitable for real-world situations. Summary of the Invention

[0004] In order to solve the problems existing in the background technology, the present invention provides a harmonic reducer fault diagnosis method based on DANN variable working condition and few labels.

[0005] The technical solution adopted in the present invention is:

[0006] The harmonic reducer fault diagnosis method of the present invention comprises the following steps:

[0007] S1. Vibration signals of the harmonic reducer in a healthy state under four load conditions, including normal, rigid wheel fault, flexwheel fault, and wave generator fault, are collected through sensors installed at the drive end of the test bench. Each vibration signal includes M samples, and each sample has m signal sampling points. The sensor is specifically a vibration sensor.

[0008] S2. For each vibration signal in step S1, use the piecewise aggregation approximation method PAA to aggregate it to reduce the size and then scale it to obtain a preprocessed vibration signal.

[0009] S3. Use the coordinate transformation method to convert the preprocessed vibration signal into a polar coordinate preprocessed vibration signal, and perform window sampling on the polar coordinate preprocessed vibration signal with a step size of 16a to obtain a number of sampling samples, each sample includes 128c signal sampling points, 128c<m; mark a number of sampling samples in each sampling sample as faulty samples, and do not mark the remaining sampling samples as unmarked samples.

[0010] S4. For each sampling sample, use the two-angle difference cosine function to process the one-dimensional sampling sample to obtain a two-dimensional GASF image of the vibration signal of the harmonic reducer; the GASF image obtained from the sampling sample with fault annotation is used as the labeled data image, and the GASF image obtained from the sampling sample without fault annotation is used as the unlabeled data image; the batch of the labeled data image is batch = B, and the batch of the unlabeled data image is batch = μB, where μ is a preset ratio.

[0011] S5. Take all labeled data images and all unlabeled data images as a training set, and divide them into a training set and a validation set according to a preset ratio for training and validation; input the training set into a label generation network based on semi-supervised learning SSL for training to obtain a trained label generation network.

[0012] S6. Input each unlabeled data image into the trained label generation network, and then output the generated label data image of each unlabeled data image after processing, that is, generate artificial labels.

[0013] S7. Take all labeled data images and generated labeled data images as source domains, specifically divide the preset proportion into training set and validation set for training and validation; take all unlabeled data images of vibration signals under another load condition other than the load condition of the source domain obtained through steps S2-S4 as target domains, input the source domain and target domain into the domain adversarial transfer network DANN for training, and obtain the trained domain adversarial transfer network DANN.

[0014] S8. Collect the real-time vibration signal of the harmonic reducer during operation through the sensor, perform the same operation on the real-time vibration signal as the vibration signal in steps S2-S3 to obtain several real-time sampling samples, perform the same operation on each real-time sampling sample as the sampling sample in step S4 to obtain several real-time GASF images, input each real-time GASF image into the trained label generation network and the trained domain adversarial transfer network DANN in sequence for processing to determine the load condition of the real-time vibration signal, and realize the fault diagnosis of the harmonic reducer under variable working conditions and few labels.

[0015] In step S2, each vibration signal is a time series signal in a Cartesian coordinate system. The time series signal is aggregated using the segmented aggregation approximation method PAA and then scaled to the [-1, 1] interval to complete the preprocessing of the vibration signal.

[0016] In step S3, the preprocessed vibration signal is converted into a polar coordinate preprocessed vibration signal using a coordinate transformation method. Specifically, the preprocessed vibration signal in the Cartesian coordinate system is converted into a polar coordinate system using an inverse cosine method. That is, for each timestamp (x, y) of the preprocessed vibration signal in the Cartesian coordinate system, the timestamp (x, y) is converted into a radius and angle in the polar coordinate system, as follows:

[0017]

[0018]

[0019] Where r is the radius of the polar coordinate system after the timestamp (x, y) is converted, θ is the angle of the polar coordinate system after the timestamp (x, y) is converted; x and y are the horizontal and vertical coordinates of the timestamp in the Cartesian coordinate system, respectively.

[0020] In step S5, the training set is input into the label generation network based on semi-supervised learning SSL for training to obtain a trained label generation network, as follows:

[0021] S5.1. Input each labeled data image in the training set into the label generation network based on semi-supervised learning SSL for conventional supervised training. The label generation network based on semi-supervised learning SSL outputs the CrossEntropyLoss cross entropy loss l for each labeled data image. s .

[0022] S5.2. Augment image enhancement processing is performed on each unlabeled data image in the training set based on the label generation network of semi-supervised learning SSL. Augment image enhancement processing includes strong enhancement processing and weak enhancement processing.

[0023] S5.3. According to each unlabeled data image that has been weakly enhanced, the argmax prediction probability of the Softmax prediction result of the unlabeled data image that has been weakly enhanced is obtained after processing by the label generation network based on semi-supervised learning SSL, and the Softmax prediction result of the unlabeled data image that has been weakly enhanced is output until the argmax prediction probability is greater than a preset benchmark.

[0024] S5.4. According to each unlabeled data image that has been strongly enhanced, the softmax prediction result of the unlabeled data image that has been strongly enhanced is obtained after processing based on the label generation network of semi-supervised learning SSL. The softmax prediction results obtained after each unlabeled data image has been weakly enhanced and strongly enhanced are subjected to cross entropy loss CrossEntropyLoss to obtain the cross entropy loss l of each unlabeled data image. u ; Extracting generalized features through consistency regularization and pseudo-labeling.

[0025] S5.5, according to the CrossEntropyLoss cross entropy loss l of each labeled data image s and the CrossEntropyLoss cross entropy loss l for each unlabeled data image u The final loss of the label generation network is obtained, and the reverse gradient of the final loss is propagated to the label generation network to optimize and update the parameters of the label generation network, and finally a trained label generation network based on semi-supervised learning SSL is obtained.

[0026] The strong enhancement processing is specifically performing RandAugment processing and Control Theory Data Enhancement CTAugment processing; the weak enhancement processing is specifically performing standard rotation and / or shift processing.

[0027] The final loss Loss is as follows:

[0028] Loss = l s +λl u

[0029] Among them, λ is a preset hyperparameter.

[0030] In step S7, the source domain and the target domain are input into the domain adversarial transfer network DANN for training to obtain the trained domain adversarial transfer network DANN, as follows:

[0031] S7.1. The domain adversarial transfer network DANN includes a feature extractor Gf, a category classifier Gy, and a domain discriminator Gd. First, the source domain and the target domain are input into the feature extractor Gf for feature extraction to obtain several domain-invariant features with strong generalization ability; the feature extractor Gf outputs the feature extraction loss.

[0032] S7.2. Input the invariant features of each domain into the category classifier Gy for fault classification, and the category classifier Gy outputs the category classification loss.

[0033] S7.3. Input each domain-invariant feature into the domain discriminator Gd to determine whether the domain-invariant feature comes from the source domain or the target domain. The domain discriminator Gd outputs the domain discrimination loss.

[0034] S7.4. The domain adversarial transfer network DANN optimizes and updates the parameters of the feature extractor Gf, category classifier Gy, and domain discriminator Gd by minimizing the feature extraction loss and category classification loss while maximizing the domain discrimination loss until the domain adversarial transfer network DANN converges, thereby obtaining the trained domain adversarial transfer network DANN.

[0035] The domain adversarial transfer network DANN also adds a gradient reversal layer GRL between the feature extractor Gf and the domain discriminator Gd, so that the domain classification loss is maximized while minimizing the fault label classification loss, turning the stochastic gradient descent into gradient ascent. The function of the gradient reversal layer GRL is achieved by inverting the gradient of this layer.

[0036] The beneficial effects of the present invention are:

[0037] 1. GASF and windowed image processing methods are used to convert time domain signals into images, thereby providing a more comprehensive representation for feature extraction.

[0038] 2. Using an SSL-based label generation module, we solve the previously unconsidered problem of insufficient source domain labels by generating artificial labels; and extract generalized features through consistency regularization and pseudo-labeling, both of which are applied to the generation of source domain labels to realize intelligent fault detection in actual industry.

[0039] 3. The adversarial domain adaptation method DANN is introduced to measure and optimize the data distribution across domains, which is conducive to effective and accurate fault diagnosis results when major changes occur in working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] like Figure 1 As shown, the harmonic reducer fault diagnosis method of the present invention includes the following steps:

[0043] S1. Vibration signals of the harmonic reducer in a healthy state under four load conditions, including normal, rigid wheel fault, flexwheel fault, and wave generator fault, are collected through sensors installed at the drive end of the test bench. Each vibration signal includes M samples, and each sample has m signal sampling points. The sensor is specifically a vibration sensor.

[0044] S2. For each vibration signal in step S1, use the piecewise aggregation approximation method PAA to aggregate it to reduce the size and then scale it to obtain a preprocessed vibration signal.

[0045] In step S2, each vibration signal is a time series signal in a Cartesian coordinate system. The time series signal is aggregated using the segmented aggregation approximation method PAA and then scaled to the [-1, 1] interval to complete the preprocessing of the vibration signal.

[0046] S3. Use the coordinate transformation method to convert the preprocessed vibration signal into a polar coordinate preprocessed vibration signal, and perform window sampling on the polar coordinate preprocessed vibration signal with a step size of 16a to obtain a number of sampling samples, each sample includes 128c signal sampling points, 128c<m; mark a number of sampling samples in each sampling sample as faulty samples, and do not mark the remaining sampling samples as unmarked samples.

[0047] In step S3, the preprocessed vibration signal is converted into a polar coordinate preprocessed vibration signal using a coordinate transformation method. Specifically, the preprocessed vibration signal in the Cartesian coordinate system is converted into a polar coordinate system using an inverse cosine method. That is, for each timestamp (x, y) of the preprocessed vibration signal in the Cartesian coordinate system, the timestamp (x, y) is converted into a radius and angle in the polar coordinate system, as follows:

[0048]

[0049]

[0050] Where r is the radius of the polar coordinate system after the timestamp (x, y) is converted, θ is the angle of the polar coordinate system after the timestamp (x, y) is converted; x and y are the horizontal and vertical coordinates of the timestamp in the Cartesian coordinate system, respectively.

[0051] S4. For each sampling sample, use the two-angle difference cosine function to process the one-dimensional sampling sample to obtain a two-dimensional GASF image of the vibration signal of the harmonic reducer; the GASF image obtained from the sampling sample with fault annotation is used as the labeled data image, and the GASF image obtained from the sampling sample without fault annotation is used as the unlabeled data image; the batch of the labeled data image is batch = B, and the batch of the unlabeled data image is batch = μB, where μ is a preset ratio.

[0052] S5. Take all labeled data images and all unlabeled data images as a training set, and divide them into a training set and a validation set according to a preset ratio for training and validation; input the training set into a label generation network based on semi-supervised learning SSL for training to obtain a trained label generation network.

[0053] In step S5, the training set is input into the label generation network based on semi-supervised learning SSL for training to obtain a trained label generation network, as follows:

[0054] S5.1. Input each labeled data image in the training set into the label generation network based on semi-supervised learning SSL for conventional supervised training. The label generation network based on semi-supervised learning SSL outputs the CrossEntropyLoss cross entropy loss l for each labeled data image. s .

[0055] S5.2. Augment image enhancement processing is performed on each unlabeled data image in the training set based on the label generation network of semi-supervised learning SSL. Augment image enhancement processing includes strong enhancement processing and weak enhancement processing.

[0056] S5.3. According to each unlabeled data image that has been weakly enhanced, the argmax prediction probability of the Softmax prediction result of the unlabeled data image that has been weakly enhanced is obtained after processing by the label generation network based on semi-supervised learning SSL, and the Softmax prediction result of the unlabeled data image that has been weakly enhanced is output until the argmax prediction probability is greater than a preset benchmark.

[0057] S5.4. According to each unlabeled data image that has been strongly enhanced, the softmax prediction result of the unlabeled data image that has been strongly enhanced is obtained after processing based on the label generation network of semi-supervised learning SSL. The softmax prediction results obtained after each unlabeled data image has been weakly enhanced and strongly enhanced are subjected to cross entropy loss CrossEntropyLoss to obtain the cross entropy loss l of each unlabeled data image. u ; Extracting generalization features through consistency regularization and pseudo-labeling.

[0058] S5.5, according to the CrossEntropyLoss cross entropy loss l of each labeled data image s and the CrossEntropyLoss cross entropy loss l for each unlabeled data image u The final loss of the label generation network is obtained, and the reverse gradient of the final loss is propagated to the label generation network to optimize and update the parameters of the label generation network, and finally a trained label generation network based on semi-supervised learning SSL is obtained.

[0059] The strong enhancement processing is specifically performing RandAugment processing and Control Theory Data Augmentation CTAugment processing; the weak enhancement processing is specifically performing standard rotation and / or shift processing.

[0060] The final loss is as follows:

[0061] Loss = l s +λl u

[0062] Among them, λ is a preset hyperparameter.

[0063] S6. Input each unlabeled data image into the trained label generation network, and then output the generated label data image of each unlabeled data image after processing, that is, generate artificial labels.

[0064] S7. Take all labeled data images and generated labeled data images as source domains, specifically divide the preset proportion into training set and validation set for training and validation; take all unlabeled data images of vibration signals under another load condition other than the load condition of the source domain obtained through steps S2-S4 as target domains, input the source domain and target domain into the domain adversarial transfer network DANN for training, and obtain the trained domain adversarial transfer network DANN.

[0065] In step S7, the source domain and the target domain are input into the domain adversarial transfer network DANN for training to obtain the trained domain adversarial transfer network DANN, as follows:

[0066] S7.1. The domain adversarial transfer network DANN includes a feature extractor Gf, a category classifier Gy, and a domain discriminator Gd. First, the source domain and the target domain are input into the feature extractor Gf for feature extraction to obtain several domain-invariant features with strong generalization ability; the feature extractor Gf outputs the feature extraction loss.

[0067] S7.2. Input the invariant features of each domain into the category classifier Gy for fault classification, and the category classifier Gy outputs the category classification loss.

[0068] S7.3. Input each domain-invariant feature into the domain discriminator Gd to determine whether the domain-invariant feature comes from the source domain or the target domain. The domain discriminator Gd outputs the domain discrimination loss.

[0069] S7.4. The domain adversarial transfer network DANN optimizes and updates the parameters of the feature extractor Gf, category classifier Gy, and domain discriminator Gd by minimizing the feature extraction loss and category classification loss while maximizing the domain discrimination loss until the domain adversarial transfer network DANN converges, thereby obtaining the trained domain adversarial transfer network DANN.

[0070] The domain adversarial transfer network DANN also adds a gradient reversal layer GRL between the feature extractor Gf and the domain discriminator Gd, so that the domain classification loss is maximized while minimizing the fault label classification loss, turning the stochastic gradient descent into gradient ascent. The function of the gradient reversal layer GRL is achieved by inverting the gradient of this layer.

[0071] S8. Collect the real-time vibration signal of the harmonic reducer during operation through the sensor, perform the same operation on the real-time vibration signal as the vibration signal in steps S2-S3 to obtain several real-time sampling samples, perform the same operation on each real-time sampling sample as the sampling sample in step S4 to obtain several real-time GASF images, input each real-time GASF image into the trained label generation network and the trained domain adversarial transfer network DANN in sequence for processing to determine the load condition of the real-time vibration signal, and realize the fault diagnosis of the harmonic reducer under variable working conditions and few labels.

[0072] In step S3, the polar coordinate preprocessed vibration signal is windowed and sampled with a step size of 64 to obtain several samples, each containing 512 signal sampling points. In steps S5 and S7, the network is trained and validated by dividing the data set into a training set and a data set in a ratio of 8:2. In step S7, the network maximizes the loss term under the influence of the gradient reversal layer (GRL). Comparative testing shows that the model selection domain loss influence factor δ = 0.3.

[0073] The method of the present invention converts the time domain signal of the harmonic reducer into a two-dimensional image by designing a signal windowing and interception method and a GASF signal conversion method, thereby improving the extraction efficiency of signal features in the deep network. A label generation method based on semi-supervised learning is designed to generate corresponding artificial prediction labels for signal samples without health status annotations, making them suitable for subsequent fault diagnosis tasks. A domain adversarial neural network model based on the adversarial generation algorithm is designed, which greatly improves the precision and accuracy of harmonic reducer fault diagnosis under variable operating conditions.

Claims

1. A fault diagnosis method for harmonic reducers based on DANN variable working conditions and few labels, characterized by: The method comprises the following steps: S1. Vibration signals of the harmonic reducer under four load conditions, including normal, rigid wheel fault, flexspline fault, and wave generator fault, are collected through sensors. S2. For each vibration signal in step S1, use the piecewise aggregation approximation method PAA to aggregate and scale it to obtain a preprocessed vibration signal; S3. Use a coordinate transformation method to convert the preprocessed vibration signal into a polar coordinate preprocessed vibration signal, perform windowing and sampling on the polar coordinate preprocessed vibration signal with a step size of 16 Å to obtain a number of sampling samples, each sample including 128 signal sampling points; mark a number of sampling samples as fault-marked samples, and do not mark the remaining sampling samples as unmarked samples; S4. For each sampling sample, use the two-angle difference cosine function to process the sampling sample to obtain a GASF image of the vibration signal of the harmonic reducer; the GASF image obtained from the sampling sample with fault marking is used as a labeled data image, and the GASF image obtained from the sampling sample without fault marking is used as an unlabeled data image; S5. Take all labeled data images and all unlabeled data images as training sets, input the training sets into a label generation network based on semi-supervised learning (SSL) for training, and obtain a trained label generation network. S6. Input each unlabeled data image into the trained label generation network, and then output a generated label data image for each unlabeled data image; S7. Use the labeled data images and the generated labeled data images together as the source domain, use the vibration signal under a load condition other than the source domain obtained through steps S2-S4 as the target domain, input the source domain and the target domain together into a domain adversarial transfer network (DANN) for training, and obtain the trained domain adversarial transfer network (DANN); S8. Collect the real-time vibration signal of the harmonic reducer during operation through the sensor, perform the same operation on the real-time vibration signal as the vibration signal in steps S2-S3 to obtain several real-time sampling samples, perform the same operation on each real-time sampling sample as the sampling sample in step S4 to obtain several real-time GASF images, input each real-time GASF image into the trained label generation network and the trained domain adversarial transfer network DANN in sequence for processing to determine the load condition of the real-time vibration signal, and realize the fault diagnosis of the harmonic reducer under variable working conditions and few labels.

2. The method for fault diagnosis of harmonic reducer based on DANN variable working condition and few labels according to claim 1 is characterized by: In the step S2, each vibration signal is a time series signal in a Cartesian coordinate system. The time series signal is aggregated using the piecewise aggregation approximation method PAA and then scaled to the [-1, 1] interval to complete the preprocessing of the vibration signal.

3. The method for fault diagnosis of harmonic reducer based on DANN variable working condition and few labels according to claim 2 is characterized by: In step S3, the preprocessed vibration signal is converted into a polar coordinate preprocessed vibration signal using a coordinate transformation method. Specifically, the preprocessed vibration signal in the Cartesian coordinate system is converted into a polar coordinate system using an inverse cosine method. That is, for each timestamp (x, y) of the preprocessed vibration signal in the Cartesian coordinate system, the timestamp (x, y) is converted into a radius and angle in the polar coordinate system, as follows: Where r is the radius of the polar coordinate system after the timestamp (x, y) is converted, θ is the angle of the polar coordinate system after the timestamp (x, y) is converted; x and y are the horizontal and vertical coordinates of the timestamp in the Cartesian coordinate system, respectively.

4. The method for fault diagnosis of harmonic reducer based on DANN variable working condition and few labels according to claim 1 is characterized by: In step S5, the training set is input into the label generation network based on semi-supervised learning SSL for training to obtain a trained label generation network, as follows: S5.

1. Input each labeled data image in the training set into the label generation network based on semi-supervised learning SSL for supervised training. The label generation network based on semi-supervised learning SSL outputs the CrossEntropyLoss cross entropy loss l for each labeled data image. s ; S5.

2. Augment image enhancement is performed on each unlabeled data image in the training set based on the label generation network of semi-supervised learning (SSL). Augment image enhancement includes strong enhancement and weak enhancement. S5.

3. Based on each weakly enhanced unlabeled data image, obtain the argmax prediction probability of the softmax prediction result of the weakly enhanced unlabeled data image after processing based on the label generation network of semi-supervised learning SSL, and output the softmax prediction result of the weakly enhanced unlabeled data image when the argmax prediction probability is greater than a preset benchmark; S5.

4. According to each unlabeled data image that has been strongly enhanced, the softmax prediction result of the unlabeled data image that has been strongly enhanced is obtained after processing based on the label generation network of semi-supervised learning SSL. The softmax prediction results obtained after each unlabeled data image has been weakly enhanced and strongly enhanced are subjected to cross entropy loss CrossEntropyLoss to obtain the cross entropy loss l of each unlabeled data image. u ; S5.5, according to the CrossEntropyLoss cross entropy loss l of each labeled data image s and the CrossEntropyLoss cross entropy loss l for each unlabeled data image u The final loss of the label generation network is obtained, and the reverse gradient of the final loss is propagated to the label generation network to optimize and update the parameters of the label generation network, and finally a trained label generation network based on semi-supervised learning SSL is obtained.

5. The method for fault diagnosis of harmonic reducer based on DANN variable working condition and few labels according to claim 4 is characterized by: The strong enhancement processing is specifically performing RandAugment processing and Control Theory Data Enhancement CTAugment processing; the weak enhancement processing is specifically performing rotation and / or shift processing.

6. The method for fault diagnosis of harmonic reducer based on DANN variable working condition and few labels according to claim 4 is characterized by: The final losses are as follows: Loss=l s +λl u Among them, Loss is the final loss; λ is the preset hyperparameter.

7. The method for fault diagnosis of harmonic reducer based on DANN variable working condition and few labels according to claim 1 is characterized by: In step S7, the source domain and the target domain are input into the domain adversarial transfer network DANN for training to obtain the trained domain adversarial transfer network DANN, as follows: S7.

1. The domain adversarial transfer network (DANN) consists of a feature extractor Gf, a category classifier Gy, and a domain discriminator Gd. First, the source domain and the target domain are input into the feature extractor Gf for feature extraction to obtain several domain-invariant features. The feature extractor Gf outputs the feature extraction loss; S7.

2. Input the domain-invariant features into the category classifier Gy, which outputs the category classification loss. S7.

3. Input the domain-invariant features of each domain into the domain discriminator Gd, and the domain discriminator Gd outputs the domain discrimination loss; S7.

4. The domain adversarial transfer network DANN optimizes and updates the parameters of the feature extractor Gf, category classifier Gy, and domain discriminator Gd by minimizing the feature extraction loss and category classification loss while maximizing the domain discrimination loss until the domain adversarial transfer network DANN converges, thereby obtaining the trained domain adversarial transfer network DANN.

Images