Industrial robot cross-process fault diagnosis method and system based on transfer learning

Abstract

The application provides an industrial robot cross-process fault diagnosis method based on transfer learning, relates to the field of industrial robot fault diagnosis, solves the problems of great limitation of current fault diagnosis technology, few samples of diagnosis and low accuracy, and the application is based on a feature-based transfer learning method, which first uses a transfer learning method to learn model parameters, feature representation, task structure and other knowledge of a source domain process to help a learning task of a target domain process, wherein a field self-adaption method is used to transform and align features extracted under different processes, the source domain process and the target domain process are made to have similar distributions through training, the classification ability of the source domain process is continuously improved and the distance between the two fields is continuously close through model training, so that the model can also diagnose faults of the robot target process.

Description

Technical Field

[0001] This invention belongs to the field of industrial robot technology, and more specifically, relates to a method and system for cross-process fault diagnosis of industrial robots based on transfer learning. Background Technology

[0002] In real-world production and daily life, industrial robots are generally applied to various processes, such as welding, handling, and mixing. As the process environment becomes increasingly complex, robot performance is affected to varying degrees, and the characteristics of resulting faults gradually change. Therefore, cross-domain process fault diagnosis is particularly important. Currently, industrial methods for cross-process robot fault diagnosis include physical model-based methods. This method utilizes the physical characteristics, kinematics, and control theory of industrial robots to establish a physical model of the robot. By measuring the robot's state and calculating the physical model, the working state of the robot system is analyzed, thereby achieving fault diagnosis. However, this method is difficult to establish a physical model for complex robot structures and control systems or with unknown structural parameters, thus affecting the accuracy of fault diagnosis. Another method is deep learning-based cross-process fault diagnosis. This method can transfer or fine-tune existing labeled data or models under different operating conditions to achieve cross-condition fault diagnosis. However, this method usually requires a large amount of labeled information in the target domain. In actual industrial production applications, it is difficult to collect a large amount of labeled data for different processes. With a small number of samples, the model struggles to achieve good fault diagnosis results.

[0003] To address the above issues, existing technologies disclose a fault diagnosis method for industrial robots. This application fully utilizes existing data in the control system, eliminating the need for additional physical equipment and leveraging the design theory of the equipment, making it simple and practical. It can easily and effectively identify specific fault causes. However, for complex robot structures and control systems, or in the case of unknown structural parameters, it is difficult to establish a physical model, which can affect the accuracy of robot fault diagnosis. Therefore, using a physical model-based cross-process fault diagnosis method for robot fault diagnosis has significant limitations. Summary of the Invention

[0004] To overcome the limitations of existing technologies in cross-process fault diagnosis, which suffer from low accuracy, and to address the problem of cross-process fault diagnosis, this invention proposes a transfer learning-based method and system for cross-process fault diagnosis of industrial robots. This method effectively extracts cross-domain fault features and significantly improves the accuracy of cross-domain fault diagnosis.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides a cross-process fault diagnosis method for industrial robots based on transfer learning, comprising the following steps:

[0007] S1. Obtain labeled data from the source domain and unlabeled data from the target domain;

[0008] S2. Construct a cross-process fault diagnosis model, including a feature extraction layer, a feature transfer layer, and a loss function layer connected in sequence;

[0009] S3. Input the source domain labeled data and the target domain unlabeled data into the cross-process fault diagnosis model. The feature extraction layer performs feature extraction on the source domain labeled data and the target domain unlabeled data respectively to obtain the source domain process fault features and the target domain process fault features.

[0010] S4. Input the source domain process fault features and the target domain process fault features into the feature transfer layer for feature transfer, and calculate the distance between the source domain process fault features and the target domain process fault features.

[0011] S5. The loss function layer sets a distance loss function based on the distance between the source domain process fault features and the target domain process fault features, sets a triplet loss function based on the feature extraction of labeled data in the source domain, and constructs the total loss function.

[0012] S6. Optimize the total loss function. When the loss function reaches its minimum value, feed back the corresponding network parameters to the feature extraction layer and save them to obtain the trained cross-process fault diagnosis model.

[0013] S7. Obtain the process fault data of the target domain to be diagnosed, input it into the trained cross-process fault diagnosis model, and obtain the diagnosis results of the process fault data of the target domain to be diagnosed.

[0014] Preferably, the feature extraction layer includes a triplet multi-scale attention convolutional network unit and a multi-scale attention convolutional network unit arranged in parallel; the triplet multi-scale attention convolutional network unit is used to extract features from labeled data in the source domain, and the multi-scale attention convolutional network unit is used to extract features from unlabeled data in the target domain.

[0015] Preferably, in step S3, a triplet MS-SECNN is used to extract features from the labeled data in the source domain, and a multi-scale attention convolutional network (MS-SECNN) is used to extract features from the unlabeled data in the target domain to obtain process fault features in the target domain. The triplet MS-SECNN is pre-trained using the obtained labeled data in the source domain, and the pre-trained labeled data in the source domain is used to extract features from the triplet MS-SECNN. Then, the triplet MS-SECNN is adjusted under the unlabeled process fault data in the target domain and weight sharing is implemented in the feature extraction layer.

[0016] Preferably, the triplet multi-scale attention convolutional network unit includes three multi-scale attention convolutional network sub-units arranged in parallel; the multi-scale attention convolutional network includes a first convolutional layer, a second convolutional layer, a splicing layer and an SENet attention layer connected in sequence, and the output of the first convolutional layer is also connected to the input of the splicing layer.

[0017] Preferably, the feature transfer in step S4 uses domain adaptation to adapt the triplet multi-scale attention convolutional network model from the source domain process to the target domain process. Specifically, the deep features output by the feature extraction layer are mapped to a common high-dimensional feature space in the domain adaptation layer, and the distance between the source domain process fault features and the target domain process fault features is calculated using the MK-MMD method. The MK-MMD Loss is used to judge the effect of domain adaptation.

[0018] Preferably, the domain adaptation refers to the process of adapting the model from the source domain process domain to the target domain process domain, wherein the source domain process and the target domain process may have different distributions or characteristics. The source domain process is the working state of a robot under one process, and the target domain process is the working state of the same robot under different processes or the working state of different robots under the same process.

[0019] Preferably, in step S4, MK-MMD is used to calculate the MMD distance by selecting multiple kernel functions. Different kernel functions will produce different distributions, and linearly combining these distributions can better capture the differences between the source domain process and the target domain process, effectively reducing the distribution difference between the two processes, thereby improving the performance of domain adaptation and the accuracy of fault diagnosis. The specific calculation formula is:

[0020]

[0021] Xs represents the source domain process fault characteristics, Xt represents the target domain process fault characteristics, G represents the number of kernels, and Kθi represents a Gaussian kernel with bandwidth θi.

[0022] Preferably, step S5 uses both the feature extraction triplet loss function and the distance loss function as the model loss. This can improve the model's fault diagnosis performance and reduce the difference between the source and target domains, thereby optimizing the model's parameters and making the model more accurate in fault diagnosis.

[0023] This invention also provides a cross-process fault diagnosis system for industrial robots based on transfer learning, used to implement the above-mentioned diagnosis method, including:

[0024] The data acquisition module is used to acquire labeled data from the source domain and unlabeled data from the target domain.

[0025] The model building module is used to build a cross-process fault diagnosis model, which includes a feature extraction layer, a feature transfer layer and a loss function layer connected in sequence.

[0026] The feature extraction module is used to input labeled data from the source domain and unlabeled data from the target domain into the cross-process fault diagnosis model. The feature extraction layer performs feature extraction on the labeled data from the source domain and the unlabeled data from the target domain to obtain process fault features from the source domain and process fault features from the target domain.

[0027] The feature transfer module is used to input the source domain process fault features and the target domain process fault features into the feature transfer layer for feature transfer and to calculate the distance between the source domain process fault features and the target domain process fault features.

[0028] The loss function setting module is used by the loss function layer to set the distance loss function based on the distance between the source domain process fault features and the target domain process fault features, set the triplet loss function based on the feature extraction of labeled data in the source domain, and construct the total loss function.

[0029] The loss function optimization module is used to optimize the total loss function. When the loss function reaches its minimum value, the corresponding network parameters are fed back to the feature extraction layer and saved to obtain the trained cross-process fault diagnosis model.

[0030] The cross-process fault diagnosis module is used to obtain the target domain process fault data to be diagnosed, input the trained cross-process fault diagnosis model, and obtain the diagnosis results of the target domain process fault data.

[0031] Preferably, in the feature extraction module, the feature extraction layer includes a triplet multi-scale attention convolutional network unit and a multi-scale attention convolutional network unit arranged in parallel; the triplet multi-scale attention convolutional network unit is used to extract features from labeled data in the source domain, and the multi-scale attention convolutional network unit is used to extract features from unlabeled data in the target domain.

[0032] Preferably, in the feature extraction module, the triplet multi-scale attention convolutional network unit includes three multi-scale attention convolutional network sub-units arranged in parallel.

[0033] Preferably, in the feature extraction module, the multi-scale attention convolutional network includes a first convolutional layer, a second convolutional layer, a splicing layer, and an SENet attention layer connected in sequence, and the output of the first convolutional layer is also connected to the input of the splicing layer.

[0034] Preferably, in the feature transfer module, the source domain process fault features and the target domain process fault features are input into the feature transfer layer for feature transfer, and the specific formula for calculating the distance between the source domain process fault features and the target domain process fault features is as follows:

[0035]

[0036] Xs represents the source domain process fault characteristics, Xt represents the target domain process fault characteristics, G represents the number of kernels, and Kθi represents a Gaussian kernel with bandwidth θi.

[0037] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0038] Compared with some existing transfer learning-based fault diagnosis methods, such as transfer component analysis and joint distribution adaptation, the industrial robot cross-process fault diagnosis method and system proposed in this invention extracts transfer features more clearly, minimizes the domain offset between the source process and the target process, and has good generalization ability and high-precision cross-process fault diagnosis capability. It is of great significance for the fault diagnosis application of robots in actual industrial production. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the steps of a cross-process fault diagnosis method for industrial robots based on transfer learning.

[0040] Figure 2 The diagram shows the model structure of a method and system for cross-process fault diagnosis of industrial robots based on transfer learning.

[0041] Figure 3 This diagram shows the component modules of a cross-process fault diagnosis system for industrial robots based on transfer learning. Detailed Implementation

[0042] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0043] To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual size of the product. The descriptions of the directions of parts such as "up" and "down" are not intended to limit this patent.

[0044] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0045] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0046] Example 1

[0047] This embodiment provides a cross-process fault diagnosis method for industrial robots based on transfer learning, such as... Figure 1 As shown, it includes the following steps:

[0048] S1. Obtain labeled data from the source domain and unlabeled data from the target domain;

[0049] S2. Construct a cross-process fault diagnosis model, including a feature extraction layer, a feature transfer layer, and a loss function layer connected in sequence;

[0050] S3. Input the source domain labeled data and the target domain unlabeled data into the cross-process fault diagnosis model. The feature extraction layer performs feature extraction on the source domain labeled data and the target domain unlabeled data respectively to obtain the source domain process fault features and the target domain process fault features.

[0051] S4. Input the source domain process fault features and the target domain process fault features into the feature transfer layer for feature transfer, and calculate the distance between the source domain process fault features and the target domain process fault features.

[0052] S5. The loss function layer sets a distance loss function based on the distance between the source domain process fault features and the target domain process fault features, sets a triplet loss function based on the feature extraction of labeled data in the source domain, and constructs the total loss function.

[0053] S6. Optimize the total loss function. When the loss function reaches its minimum value, feed back the corresponding network parameters to the feature extraction layer and save them to obtain the trained cross-process fault diagnosis model.

[0054] S7. Obtain the process fault data of the target domain to be diagnosed, input it into the trained cross-process fault diagnosis model, and obtain the diagnosis results of the process fault data of the target domain to be diagnosed.

[0055] This embodiment first obtains labeled source domain data and unlabeled target domain data through a data acquisition module. The obtained labeled source domain data is used to pre-train the network. The resulting model serves as a feature extractor in the feature extraction module to extract features from the labeled source domain data. During extraction, a triplet loss function is constructed for the source domain process data to obtain source domain process fault features. Feature extraction is then performed on the unlabeled target domain process fault data to obtain target domain process fault features. The extracted source domain process fault features are then transferred to the target domain process using a domain-adaptive method, and the distance between the source and target domain process fault features is calculated. A distance loss function is set based on this distance. Combined with the feature extraction triplet loss function, a total loss function is constructed. The total loss function is optimized, and when it reaches its minimum value, the corresponding network parameters are fed back to the feature extraction layer and saved, resulting in a trained cross-process fault diagnosis model. Finally, the target domain process fault data to be diagnosed is input into the trained cross-process fault diagnosis model to obtain the diagnostic results for the target domain process fault data. This embodiment can effectively extract fault features across domains, significantly improving the accuracy of cross-domain fault diagnosis.

[0056] Example 2

[0057] This embodiment provides a cross-process fault diagnosis method for industrial robots based on transfer learning, such as... Figure 2 As shown, it includes the following steps:

[0058] S1. Obtain labeled data from the source domain and unlabeled data from the target domain;

[0059] S2. Construct a cross-process fault diagnosis model, including a feature extraction layer, a feature transfer layer, and a loss function layer connected in sequence;

[0060] S3. Input the source domain labeled data and the target domain unlabeled data into the cross-process fault diagnosis model. The feature extraction layer performs feature extraction on the source domain labeled data and the target domain unlabeled data respectively to obtain the source domain process fault features and the target domain process fault features.

[0061] The feature extraction layer includes a triplet multi-scale attention convolutional network unit and a multi-scale attention convolutional network unit arranged in parallel. The triplet multi-scale attention convolutional network unit is used to extract features from labeled data in the source domain, and the multi-scale attention convolutional network unit is used to extract features from unlabeled data in the target domain.

[0062] The triplet multi-scale attention convolutional network unit consists of three multi-scale attention convolutional network sub-units arranged in parallel.

[0063] The multi-scale attention convolutional network consists of a first convolutional layer, a second convolutional layer, a splicing layer, and a SENet attention layer connected in sequence. The output of the first convolutional layer is also connected to the input of the splicing layer.

[0064] S4. Input the source domain process fault features and the target domain process fault features into the feature transfer layer for feature transfer, and calculate the distance between the source domain process fault features and the target domain process fault features; the specific formula is:

[0065]

[0066] Xs represents the source domain process fault characteristics, Xt represents the target domain process fault characteristics, G represents the number of kernels, and Kθi represents a Gaussian kernel with bandwidth θi.

[0067] S5. The loss function layer sets a distance loss function based on the distance between the source domain process fault features and the target domain process fault features, sets a triplet loss function based on the feature extraction of labeled data in the source domain, and constructs the total loss function.

[0068] S6. Optimize the total loss function. When the loss function reaches its minimum value, feed back the corresponding network parameters to the feature extraction layer and save them to obtain the trained cross-process fault diagnosis model.

[0069] S7. Obtain the process fault data of the target domain to be diagnosed, input it into the trained cross-process fault diagnosis model, and obtain the diagnosis results of the process fault data of the target domain to be diagnosed.

[0070] This embodiment first acquires labeled data in the source domain and unlabeled data in the target domain. The network is pre-trained using the labeled source domain data. Feature extraction for the source domain and target domain uses a triplet multi-scale attention convolutional network (Triplet MS-SECNN) structure and a multi-scale attention convolutional network (MS-SECNN) structure, respectively. The triplet MS-SECNN is used to extract features from the labeled source domain data. After extraction, the distance between anchor sample data and positive sample data becomes closer, making the anchor sample data farther from negative sample data. Feature extraction for the unlabeled target domain data uses a multi-scale attention convolutional network (MS-SECNN). The pre-trained model is used as the feature extractor. Because a triplet network is used for training, a triplet loss function (Triplet...) is applied. Domain Adaptation Loss (MK-MMD) is used as the loss function for the pre-trained model. After pre-training, the distance between source domain fault data and normal data increases, thereby achieving the effect of fault classification for the robot. The pre-trained model is then adjusted under the target domain process, allowing features from the source domain process to be transferred to the target domain process. Weight sharing is implemented in the feature extraction layer. The deep feature vectors output from the feature extraction layer are mapped to a common high-dimensional feature space in the domain adaptation layer. The distance between the source domain and target domain fault features (MK-MMD) is calculated, and the distance loss function (MK-MMD Loss) is used to judge the effect of domain adaptation. The domain adaptation method enables the target domain process to learn the features and parameters of the source domain, allowing for fault diagnosis in the target domain process under the existing model. Multi-kernel maximum mean difference (MK-MMD) is used to measure the difference between the feature distributions of the source and target domain processes. The triplet loss function (Triplet Loss) and the distance loss function (MK-MMD) are combined. The overall loss function of the model, along with the source and target process technologies, improves the model's fault diagnosis performance. It also narrows the gap between the source and target process technologies, thus optimizing model parameters and achieving optimal results. When the overall loss function is minimized, the corresponding network parameters are fed back to the feature extraction layer and saved, resulting in a well-trained cross-process fault diagnosis model. Finally, the target process fault data is input into this model to obtain the diagnostic results. This embodiment effectively extracts cross-domain fault features, significantly improving the accuracy of cross-domain fault diagnosis.

[0071] Example 3

[0072] This embodiment provides an industrial robot cross-process fault diagnosis system based on transfer learning, used to implement the diagnosis method described in Embodiment 1 or 2, such as... Figure 3 As shown, it includes:

[0073] The data acquisition module is used to acquire labeled data from the source domain and unlabeled data from the target domain.

[0074] The model building module is used to build a cross-process fault diagnosis model, which includes a feature extraction layer, a feature transfer layer and a loss function layer connected in sequence.

[0075] The feature extraction module is used to input labeled data from the source domain and unlabeled data from the target domain into the cross-process fault diagnosis model. The feature extraction layer performs feature extraction on the labeled data from the source domain and the unlabeled data from the target domain to obtain process fault features from the source domain and process fault features from the target domain.

[0076] The feature transfer module is used to input the source domain process fault features and the target domain process fault features into the feature transfer layer for feature transfer and to calculate the distance between the source domain process fault features and the target domain process fault features.

[0077] The loss function setting module is used by the loss function layer to set the distance loss function based on the distance between the source domain process fault features and the target domain process fault features, set the triplet loss function based on the feature extraction of labeled data in the source domain, and construct the total loss function.

[0078] The loss function optimization module is used to optimize the total loss function. When the loss function reaches its minimum value, the corresponding network parameters are fed back to the feature extraction layer and saved to obtain the trained cross-process fault diagnosis model. The cross-process fault diagnosis module is used to obtain the target domain process fault data to be diagnosed, input it into the trained cross-process fault diagnosis model, and obtain the diagnosis results of the target domain process fault data to be diagnosed.

[0079] In the feature extraction module, the feature extraction layer includes a triplet multi-scale attention convolutional network unit and a multi-scale attention convolutional network unit arranged in parallel; the triplet multi-scale attention convolutional network unit is used to extract features from labeled data in the source domain, and the multi-scale attention convolutional network unit is used to extract features from unlabeled data in the target domain.

[0080] The triplet multi-scale attention convolutional network unit consists of three multi-scale attention convolutional network sub-units arranged in parallel.

[0081] The multi-scale attention convolutional network consists of a first convolutional layer, a second convolutional layer, a splicing layer, and a SENet attention layer connected in sequence. The output of the first convolutional layer is also connected to the input of the splicing layer.

[0082] In the feature transfer module, the source domain process fault features and the target domain process fault features are input into the feature transfer layer for feature transfer. The specific formula for calculating the distance between the source domain process fault features and the target domain process fault features is as follows:

[0083]

[0084] Xs represents the source domain process fault characteristics, Xt represents the target domain process fault characteristics, G represents the number of kernels, and Kθi represents a Gaussian kernel with bandwidth θi.

[0085] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0086] The above embodiments of the present invention are merely illustrative examples and are not intended to limit the implementation of the invention. Those skilled in the art can make other variations and modifications based on the above description. It is impossible to exhaustively list all possible implementations here. All obvious variations or modifications derived from the technical solutions of the present invention are still within the protection scope of the present invention.

Claims

1. A method for cross-process fault diagnosis of industrial robots based on transfer learning, characterized in that, Includes the following steps: S1. Obtain labeled data from the source domain and unlabeled data from the target domain; S2. Construct a cross-process fault diagnosis model, including a feature extraction layer, a feature transfer layer, and a loss function layer connected in sequence; S3. Input the source domain labeled data and the target domain unlabeled data into the cross-process fault diagnosis model. The feature extraction layer performs feature extraction on the source domain labeled data and the target domain unlabeled data respectively to obtain the source domain process fault features and the target domain process fault features. S4. Input the source domain process fault features and the target domain process fault features into the feature transfer layer for feature transfer, and calculate the distance between the source domain process fault features and the target domain process fault features. S5. The loss function layer sets a distance loss function based on the distance between the source domain process fault features and the target domain process fault features, sets a triplet loss function based on the feature extraction of labeled data in the source domain, and constructs the total loss function. S6. Optimize the total loss function. When the loss function reaches its minimum value, feed back the corresponding network parameters to the feature extraction layer and save them to obtain the trained cross-process fault diagnosis model. S7. Obtain the process fault data of the target domain to be diagnosed, input it into the trained cross-process fault diagnosis model, and obtain the diagnosis results of the process fault data of the target domain to be diagnosed.

2. The method for cross-process fault diagnosis of industrial robots based on transfer learning according to claim 1, characterized in that, The feature extraction layer includes a triplet multi-scale attention convolutional network unit and a multi-scale attention convolutional network unit arranged in parallel. The triplet multi-scale attention convolutional network unit is used to extract features from labeled data in the source domain, and the multi-scale attention convolutional network unit is used to extract features from unlabeled data in the target domain.

3. The method for cross-process fault diagnosis of industrial robots based on transfer learning according to claim 2, characterized in that, The triplet multi-scale attention convolutional network unit consists of three multi-scale attention convolutional network sub-units arranged in parallel.

4. The method for cross-process fault diagnosis of industrial robots based on transfer learning according to claim 3, characterized in that, The multi-scale attention convolutional network consists of a first convolutional layer, a second convolutional layer, a splicing layer, and a SENet attention layer connected in sequence. The output of the first convolutional layer is also connected to the input of the splicing layer.

5. A method for cross-process fault diagnosis of industrial robots based on transfer learning according to claim 1 or 4, characterized in that, The source domain process fault features and the target domain process fault features are input into the feature transfer layer for feature transfer. The specific formula for calculating the distance between the source domain process fault features and the target domain process fault features is as follows: Xs represents the source domain process fault characteristics, Xt represents the target domain process fault characteristics, G represents the number of kernels, and Kθi represents a Gaussian kernel with bandwidth θi.

6. A cross-process fault diagnosis system for industrial robots based on transfer learning, used to implement the diagnosis method according to any one of claims 1-5, characterized in that, include: The data acquisition module is used to acquire labeled data from the source domain and unlabeled data from the target domain. The model building module is used to build a cross-process fault diagnosis model, which includes a feature extraction layer, a feature transfer layer and a loss function layer connected in sequence. The feature extraction module is used to input labeled data from the source domain and unlabeled data from the target domain into the cross-process fault diagnosis model. The feature extraction layer performs feature extraction on the labeled data from the source domain and the unlabeled data from the target domain respectively to obtain the process fault features of the source domain and the process fault features of the target domain. The feature transfer module is used to input the source domain process fault features and the target domain process fault features into the feature transfer layer for feature transfer and to calculate the distance between the source domain process fault features and the target domain process fault features. The loss function setting module is used by the loss function layer to set the distance loss function based on the distance between the source domain process fault features and the target domain process fault features, set the triplet loss function based on the feature extraction of labeled data in the source domain, and construct the total loss function. The loss function optimization module is used to optimize the total loss function. When the loss function reaches its minimum value, the corresponding network parameters are fed back to the feature extraction layer and saved to obtain the trained cross-process fault diagnosis model. The cross-process fault diagnosis module is used to obtain the target domain process fault data to be diagnosed, input the trained cross-process fault diagnosis model, and obtain the diagnosis results of the target domain process fault data.

7. The industrial robot cross-process fault diagnosis system based on transfer learning according to claim 6, characterized in that, In the feature extraction module, the feature extraction layer includes a triplet multi-scale attention convolutional network unit and a multi-scale attention convolutional network unit arranged in parallel; the triplet multi-scale attention convolutional network unit is used to extract features from labeled data in the source domain, and the multi-scale attention convolutional network unit is used to extract features from unlabeled data in the target domain.

8. The industrial robot cross-process fault diagnosis system based on transfer learning according to claim 7, characterized in that, The triplet multi-scale attention convolutional network unit consists of three multi-scale attention convolutional network sub-units arranged in parallel.

9. The industrial robot cross-process fault diagnosis system based on transfer learning according to claim 8, characterized in that, In the feature extraction module, the multi-scale attention convolutional network includes a first convolutional layer, a second convolutional layer, a splicing layer, and an SENet attention layer connected in sequence. The output of the first convolutional layer is also connected to the input of the splicing layer.

10. A cross-process fault diagnosis system for industrial robots based on transfer learning according to claim 6 or 9, characterized in that, In the feature transfer module, the source domain process fault features and the target domain process fault features are input into the feature transfer layer for feature transfer. The specific formula for calculating the distance between the source domain process fault features and the target domain process fault features is as follows: Xs represents the source domain process fault characteristics, Xt represents the target domain process fault characteristics, G represents the number of kernels, and Kθi represents a Gaussian kernel with bandwidth θi.

Images