A method and system for training an assembly monitoring model based on federated transfer learning

Abstract

The application discloses a kind of based on federal migration learning's assembly monitoring model training method and system.The method includes: creating multi-task federal learning model library;According to customer task characteristic data, the similarity of current task scene and existing model scene is analyzed, whether the customer task relative to warehouse subnode task is new task is judged;Design fusion weight, selectively weighted fusion model in multi-task federal learning model library, generate the initialization model with prior knowledge for customer node task, downlink model to edge customer node;Customer node trains model using local data, and model is shared to cloud server;Update multi-task federal learning model library.The application adopts the model training mode of edge cloud cooperation, reduces the requirement of edge end to calculation and storage resource;Fully consider the similarity and difference of assembly task, according to customer node task characteristic, generate customized initialization model, reduce model training cost.

Description

Technical Field

[0001] This invention belongs to the field of industrial robot monitoring, specifically relating to a robot assembly monitoring method and system based on machine learning, which is applicable to industrial production scenarios under the background of intelligent manufacturing. Background Technology

[0002] Robot assembly is widely used, especially in flexible manufacturing scenarios, characterized by a variety of tasks and frequent reconfiguration; monitoring methods based on deep learning have the advantages of strong learning ability and low labor costs, but the amount of computation and sample data required to train the model increases the difficulty of application.

[0003] To address the problems in training industrial equipment monitoring models, the industry has proposed an edge-cloud collaborative training framework and a federated learning method. Chinese Patent 202011405412.5 provides a method and system for industrial equipment monitoring based on cloud-edge collaboration. The edge platform is responsible for actual equipment detection, while the cloud platform manages and controls multiple edge platforms, avoiding the aggregation of monitoring data to the cloud platform, improving data transmission efficiency and industrial data security, and reducing network bandwidth pressure. The cloud platform integrates fault detection models from different edge platforms using federated learning technology, obtains a shared model, and distributes it to the edge platforms for use, improving the generalization ability of the fault detection model. The edge platforms detect equipment status, improving response speed and ensuring the effectiveness of detection. This cloud-edge collaborative federated learning method can effectively improve the generalization ability and response speed of the fault detection model, and the federated transfer learning method can protect privacy while considering domain alignment during model transfer. However, robot assembly tasks are diverse, such as shaft and hole assembly and gear assembly, and the tasks have certain similarities, such as assembling the same type of equipment with different materials. Existing federated learning methods do not consider the differences between tasks. Summary of the Invention

[0004] To address the problems of existing technologies, this invention provides an edge-cloud collaborative training method for multi-robot assembly task monitoring models based on federated transfer learning. By fully utilizing the diversity and similarity of actual assembly tasks, a model library is constructed. Using federated transfer learning, existing models are selectively fused to provide an initial model for the current task. The model library is then updated using the model trained for the current task. This improves the utilization and transferability of models across various task scenarios, reduces dependence on sample data, and decreases the computational load during training.

[0005] To address the aforementioned technical problems, this invention provides a method for customizing industrial multi-task monitoring models based on federated transfer learning, comprising the following steps:

[0006] 1. Define a unified neural network model and build a model library. Based on the common types of faults in various typical assembly tasks, a unified neural network structure model is adopted as the robot assembly monitoring model. Considering the differences in assembly tasks, multiple monitoring models with the same structure are used as the basic models for sub-task scenario nodes, forming a multi-task federated learning model library.

[0007] 2. Task Similarity Calculation. Based on the task feature data of the client node, a similarity calculation method is used to obtain the similarity between the current task and the existing basic models in the model library.

[0008] 3. Design a model fusion mechanism. Based on the similarity between the current client node task and the existing base model, compare with a similarity threshold, select child node models with similarity values ​​higher than the threshold as fusion objects, convert their similarity into fusion weights with the client node task, and use federated learning to fuse the models as the initial model for the client node task.

[0009] 4. Model Transfer Training and Sharing. At the client node, a unified data annotation method is used to annotate local data. The initial model obtained from the service node is then fine-tuned using this local data to obtain a model suitable for the current task monitoring, and this model is then shared.

[0010] 5. Model Recycling and Model Library Reconstruction. Based on the similarity between the trained customer node task model and the existing base model, a value threshold is compared to determine whether the customer task is a new task relative to the warehouse sub-node tasks. If so, the model is recycled and directly used as an independent base model, representing a new task scenario; otherwise, after model recycling, the generalization fusion weights of the corresponding sub-node models are calculated, and weighted fusion is performed according to the generalization base model method.

[0011] Specifically, the various assembly tasks and typical faults mentioned in step 1 can be categorized into several types based on the differences in task operations, such as shaft and hole assembly, gear assembly, and L-shaped component assembly. Each type of task has operational similarities, and the learned knowledge is generalizable. Assembly operations consist of a combination of many similar actions, and the main faults that occur are similar in different tasks, such as missing assembly parts and obstruction by obstacles. Therefore, the robot assembly monitoring problem is defined as a classification problem.

[0012] Optionally, the unified neural network mentioned in step 1 adopts a convolutional neural network to divide the multi-dimensional robot running time series data into a matrix, design a unified labeling method for typical fault types, and the neural network learns the mapping relationship between robot state and fault through supervised learning.

[0013] Specifically, the multi-task federated learning model library mentioned in step 1 is used to store monitoring models for each sub-task scenario, supporting model reading, updating, and sub-scenario expansion. Assuming that the cloud has already stored monitoring models for n task scenarios, when a new monitoring model for a new task scenario is received, it is numbered n+1 and a new node is allocated to store the model, while the number of monitoring models in the repository is updated.

[0014] Optionally, the task feature data mentioned in step 2 involves the client node collecting normal working time-series data of the task scenario, performing a series of operations such as slicing, scrambling, and reshaping, and then saving it in tensor form.

[0015] Specifically, the similarity calculation method described in step 2 retrieves the base models of all child nodes (1, 2, ..., n) from the multi-task federated learning model library, calculates the maximum classification probability of each node model using the task feature data of the client node, and uses this probability as the similarity [s1, s2, ..., sn] between the client node and the base models of each node in the federated learning model library. n ].

[0016] Specifically, the similarity threshold s mentioned in step 3 thre , is a threshold value determined by humans that reflects the similarity of the model to each sub-node model in the multi-task federated learning model library. It is used to select similar scenario node models for federated learning, reducing the impact of task differences on federated learning. Its specific value can be determined in practice through experience or mathematical relationships.

[0017] Specifically, the fusion weights mentioned in step 3 This refers to the parameter fusion ratio of the model with m highly similar nodes obtained after filtering in federated learning. The weight calculation method is as follows:

[0018]

[0019] Specifically, the federated learning-based model fusion described in step 3 is based on the fusion weights of the selected child node tasks. i∈[1,2,……m], for the filtered child node model parameters α i Weighted fusion yields customized model parameters α. s :

[0020]

[0021] Specifically, the model fine-tuning described in step 4 does not change the model structure; it only uses local data to correct the parameters of the obtained initial model.

[0022] Specifically, the value threshold s mentioned in step 5 valueThis refers to a critical value, determined artificially, that reflects the value of the model relative to the multi-task federated learning model library. It is used to select the update strategy for the multi-task federated learning model library, and its specific value can be determined in practice through experience or mathematical relationships.

[0023] Specifically, the generalization fusion weight mentioned in step 5 is to calculate the generalization fusion weight between the client node sharing model and the basic model of the existing warehouse sub-nodes based on the similarity between the client node sharing model and the basic model of the existing warehouse sub-nodes. The calculation formula is as follows:

[0024]

[0025] s max =max{s1, s2, ..., s} n}

[0026] Specifically, the generalized weighted fusion method in step 5 uses generalized fusion weights to distribute the model parameters α shared by the client nodes. c Highly similar model α in the model library b Perform weighted fusion to update the existing base model:

[0027]

[0028] The present invention also provides a system for implementing the above method. The system includes: a multi-task federated learning model library, a task feature data acquisition module, a task similarity judgment module, a federated model fusion module, a model deployment module, a model recycling module, and a federated model update module.

[0029] The multi-task federated learning model library stores multi-task scene monitoring models. This library consists of base models with child nodes sharing a unified neural network structure, and is updated using a client edge model via a federated model update module.

[0030] The task feature data acquisition module is used to collect feature data of customer edge tasks. When executing step 2, the raw data of the task scenario to be monitored is taken as input. By running this module, the data is sliced, disordered, and reshaped, and the task feature data is output in tensor form.

[0031] The task similarity assessment module is used to determine the similarity between the customer node task scenario and the existing sub-node task scenario models in the model library. During step 2, the feature data of the customer edge output by the task feature data acquisition module is used as input. By running this module, the similarity between the customer node task and each sub-node task in the multi-task federated learning model library is calculated and output.

[0032] The federated model fusion module is used to fuse models from the multi-task federated learning model library and generate customized models for client node tasks. In step 3, the similarity score output by the task similarity judgment module is used as input. By running this module, fusion objects are filtered based on a similarity threshold, and fusion weights are calculated. Based on these weights, the basic models of the filtered child node tasks are federated and fused, and the resulting customized model is output.

[0033] The model deployment module is used to distribute customized models and corresponding fault labels generated in the federated model fusion module to edge customer nodes.

[0034] The model recycling module is used to recycle customer-localized models. During step 4, running this module recycles the customer's locally customized model from the edge customer node.

[0035] The federated model update module is used to reconstruct the multi-task federated learning model library. During step 4, the fusion weights output by the task similarity judgment module are used as input. By running this module, it is determined whether the client task is a new task, and different strategies are applied to update the multi-task federated learning model library accordingly.

[0036] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0037] This invention constructs an assembly task monitoring model at low cost through knowledge sharing and fusion among multiple assembly robots, ensuring the safety of the assembly process and providing strong support for flexible assembly. It employs an edge-cloud collaborative model training approach, reducing the computational and storage resource requirements of the edge; it fully leverages the similarity of actual assembly tasks, using a federated transfer learning method to achieve knowledge sharing between different individuals through model sharing, avoiding large-scale data exchange and transmission; it fully considers the similarity and differences of assembly tasks, generating customized initialization models based on the task characteristics of customer nodes, and optionally updating the system according to different customer tasks, thereby improving the generalization and accuracy of child node models, reducing model training costs, and improving the utilization of existing model knowledge. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0039] Figure 1 A schematic diagram of the system structure provided in an embodiment of the present invention;

[0040] Figure 2 This is a flowchart of the edge-cloud collaborative training method provided in an embodiment of the present invention;

[0041] Figure 3 This is a system operation flowchart of an embodiment of the present invention. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] This invention provides a method and system for training an assembly monitoring model based on federated transfer learning. The system adopts an edge-cloud collaborative architecture, such as... Figure 1 As shown, it includes one cloud server node and multiple edge client nodes. The cloud server node is used to define a unified global model and uses a multi-task federated transfer learning strategy to provide the initial global model to the edge client nodes; the edge client nodes use local training data and the global model to train client models; the edge-cloud collaboration strategy is used for the interaction of task feature data and model parameters between the cloud server node and the edge client nodes.

[0044] like Figure 2 As shown, the present invention provides a method for training an assembly monitoring model based on federated transfer learning, comprising the following steps:

[0045] S1. A unified neural network model is used to construct assembly monitoring models for each task. Each assembly monitoring model is uploaded to the model library sub-node in the cloud server node to form an initialized multi-task federated learning model library.

[0046] Based on various typical fault scenarios under different industrial robot tasks, sample data was collected from real-time data generated by robot sensors on a Raspberry Pi. A unified neural network model was used to construct assembly monitoring models for each task. These models were then uploaded to a model library sub-node in a cloud server node, forming an initialized multi-task federated learning model library. While the neural network structures in each assembly monitoring model are identical, they differ in parameters (the neural network model parameters mainly include convolutional kernel parameters and the weights and biases of fully connected layers). The fundamental reason for the different model parameters across multiple model library sub-nodes is the significant differences in assembly tasks. The model library is updated based on similarities and differences, resulting in different model families.

[0047] In some embodiments of the present invention, the various assembly tasks and their typical faults can be classified into several categories based on the differences in task operations, such as shaft and hole assembly, gear assembly, L-shaped component assembly, etc. Each category of tasks has operational similarities, and the learned knowledge is generalizable. Assembly operations consist of a combination of many similar actions, and the main faults that occur are similar in different tasks, such as missing assembly parts and obstruction by obstacles. Therefore, the robot assembly monitoring problem is defined as a classification problem.

[0048] In some embodiments of the present invention, based on the similarity of assembly robot state data and fault types, a unified network model is adopted for multiple scenarios. The neural network model structure includes "convolutional neural network + fully connected layer" (in this embodiment, a standard convolutional layer is used, but dilated convolutional layers, transposed convolutional layers, etc. can be used as alternatives). Specifically, a network model structure of "3 convolutional layers + 3 fully connected layers" can be adopted. Based on the differences in assembly tasks, the models are trained separately for multiple scenarios to obtain multiple assembly monitoring models with the same structure. Multiple assembly monitoring models with the same structure but different parameters serve as the basic models for sub-task scenario nodes, forming a multi-task federated learning model library.

[0049] In some embodiments of the present invention, a multi-task federated learning model library is used to store assembly monitoring models for each sub-task scenario, supporting model reading, updating, and sub-scenario expansion. Assuming that n assembly monitoring models for each task scenario are already stored in the cloud, when a new monitoring model for a new task scenario is received, it is numbered n+1 and a new node is allocated to store the model, while simultaneously updating the number of assembly monitoring models in the model library.

[0050] S2 defines the task characteristic data of the client node.

[0051] In some embodiments of the present invention, the task feature data is data that is preprocessed and stored in tensor form after being collected by the client node as normal working time sequence data of the task scenario.

[0052] The raw time-series data generated by the client nodes is converted into a unified data format, which can be directly used as input to the task similarity judgment module without additional processing. In some embodiments of this invention, the raw time-series signals stored in array form are sliced, scrambled, and reshaped into a 120*18 tensor format to ensure data format uniformity across various assembly tasks, facilitating model knowledge sharing. Multi-dimensional robot runtime time-series data is segmented into matrices, and a unified annotation method is designed for fault types. The neural network learns the mapping relationship between robot state and faults through supervised learning.

[0053] S3 calculates the similarity between the current task of the client node and the existing models in the model library based on the task feature data and the existing models in the multi-task federated learning model library using the similarity calculation method, and performs model fusion through preset similarity threshold and value threshold.

[0054] The similarity calculation method receives data in a specific format from client nodes, which in some embodiments of this invention is a 120*18 tensor. Then, it retrieves assembly monitoring models from all existing model library child nodes (1,2,...,i,j≤n) from the multi-task federated learning model library, inputs the formatted data from the client nodes, and for each existing assembly monitoring model, extracts the output value of the last layer of the convolutional neural network to calculate the maximum classification probability, which is used as the similarity [s1, s2,...,s] between the client node and the corresponding child node in the multi-task federated learning model library. n ], where n is the number of child nodes in the multi-task federated learning model library, s n It is the similarity between the client node and the nth child node in the multi-task federated learning model library.

[0055] S4, a multi-task federated learning method. Based on the similarity between the current client node task and the existing base model, it compares the results with a preset similarity threshold, selecting those with similarity values ​​higher than the threshold s. thre The model library sub-nodes with similar thresholds are used as the base models of the model library sub-nodes as fusion objects. Their similarity is converted into fusion weights with the customer node task. The federated fusion algorithm is used to fuse the models, generate the initial model customized for the customer node task, and distribute the model to the customer for local training.

[0056] The similarity threshold s thre , is a manually determined threshold value reflecting the similarity between the current task of a client node and the assembly monitoring models of each sub-node in the multi-task federated learning model library. It is used to select similar scenario node models among client nodes for federated learning, reducing the impact of task differences on federated learning. When the similarity exceeds the similarity threshold s... thre When the similarity threshold s is reached, the corresponding model library sub-node has a certain relevance to the current task of the client node and can be used for initial model fusion. Its specific value can be determined through experience or mathematical relationships in practice. In some embodiments of the present invention, the similarity threshold s... thre These are specific values ​​determined by experts in industrial settings.

[0057] In some embodiments of the present invention, samples with similarity higher than a similarity threshold s are selected. thre The model library child nodes are renumbered from 1 to m, and weights are assigned to the nodes based on their similarity to the client nodes; these weights are called fusion weights. m represents the number of child nodes in the selected model library. This fusion weight value is then used for federated model fusion.

[0058] Fusion weights i∈[1,2,……m] represents the parameter fusion ratio of each highly similar node obtained after filtering in the federated model fusion module. This weight is determined based on the proportion of the corresponding similarity value in the sum of similarities.

[0059]

[0060] In the formula, s i s is the similarity score of the i-th model among the m models selected from the model library's child nodes. m It is the similarity of the m-th model among the m models selected from the model library's sub-nodes;

[0061] The federated fusion algorithm, that is, using fusion weights i∈[1, 2, ..., m], for the parameters α of the existing basic models of the selected model library sub-nodes. i We perform weighted fusion to obtain the customized model parameters α for the client nodes. s The specific formula is as follows:

[0062]

[0063] Where, α i The parameters of the existing assembly monitoring model in the model library sub-nodes can be the convolution kernel parameters of a convolutional neural network in some embodiments of the present invention, but the weights and biases of the fully connected layers of the model can also be used as the objects of the federated fusion algorithm of the present invention. Generally, the convolution kernel parameters are selected, and it can be extended to include convolution kernel parameters, weights, and biases simultaneously.

[0064] S5, Transfer Learning. The client node receives the client-customized model parameters α. s Subsequently, the client node uses local data to fine-tune the model, generating a locally personalized model after a short period of local training, thus obtaining a model suitable for the client node's current task monitoring. This model can both transfer knowledge from the cloud to the client and achieve model localization on the client side.

[0065] In some embodiments of the present invention, the model is fine-tuned without changing the model structure. A unified data annotation method is used to annotate the local data at the client node, and the parameters of the obtained initial model are corrected using only the local data.

[0066] S6, Design a model recycling and update mechanism. Determine whether a client node's task type is a new task by using similarity and a value threshold. If the client node's task has a maximum similarity s with the models in the model library's child nodes... max =max{s1, s2, ..., s} n} less than the preset value threshold s valueIf the client node task is a new task, then the locally trained model of the client node is recycled and added to the multi-task federated learning model library as a new node task. It is numbered n+1 and allocated to a new node for storage. After recycling, the model directly serves as an independent base model, representing a new task scenario. If the maximum similarity s... max Greater than the value threshold s value If the model is considered to be highly similar to the model of the existing model library child node, then the model trained locally by the client node is recycled and weighted and fused with the model of the existing model library child node to generalize the existing child node model.

[0067] The value threshold s value This is a manually determined threshold value reflecting the value of a client node's current task relative to the multi-task federated learning model library. It is defined when there exists a model with a similarity to an existing model library sub-node that exceeds this value threshold. value If the similarity between the client node task and existing sub-node tasks in the model library is high, it is considered an existing task; conversely, if the similarity between all corresponding existing sub-node tasks in the model library is lower than the value threshold s, it is considered an existing task. value If this occurs, the client node's task is considered a new task. Value threshold s value The specific value can be determined through experience or mathematical relationships in practice.

[0068] The generalized fusion weight and generalized base model method, after retrieving the customer model, when it is determined to be an old task scenario, uses the maximum similarity to construct a generalized fusion weight between the customer model and the corresponding model repository sub-node model. Using the generalized fusion weight and the customer model, the corresponding model repository sub-node model is then updated using generalization.

[0069] In some embodiments of the present invention, the weighted fusion algorithm for generalizing existing task models is as follows: based on the maximum similarity s between the shared model of the client node and the model of the existing model library sub-nodes. max To determine the generalization fusion weights between the shared model trained locally at the client node and the existing child node model.

[0070]

[0071] Use generalized fusion weights The model parameters α shared by the client nodes c High similarity model parameter α in the multi-task federated learning model library b Perform weighted fusion to update the models in the multi-task federated learning model library:

[0072]

[0073] This invention also provides a customized system for industrial multi-task monitoring models based on federated transfer learning, comprising:

[0074] The Multi-Task Federated Model Learning Library stores multi-task scene monitoring models. This library consists of base models with child nodes sharing a unified neural network structure, and is updated using client edge models via the Federated Model Update module.

[0075] The task feature data acquisition module is used to collect feature data of customer edge tasks. When executing step 2, the raw data of the task scenario to be monitored is taken as input. By running this module, the data is sliced, disordered, and reshaped, and the task feature data is output in tensor form.

[0076] The task similarity assessment module is used to determine the similarity between the customer node task scenario and the existing sub-node task scenario models in the model library. During step 2, the feature data of the customer edge output by the task feature data acquisition module is used as input. By running this module, the similarity between the customer node task and each sub-node task in the multi-task federated learning model library is calculated and output.

[0077] The federated model fusion module is used to fuse models from the multi-task federated learning model library and generate customized models for client node tasks. In step 3, the similarity score output by the task similarity judgment module is used as input. By running this module, fusion objects are filtered based on a similarity threshold, and fusion weights are calculated. Based on these weights, the basic models of the filtered child node tasks are federated and fused, and the resulting customized model is output.

[0078] The model deployment module is used to distribute customized models and corresponding fault labels generated in the federated model fusion module to edge customer nodes.

[0079] The model recycling module is used to recycle customer-localized models. During step 4, this module is run to recycle the customer's locally customized model from the edge customer node.

[0080] The federated model update module is used to reconstruct the multi-task federated learning model library. During step 4, the fusion weights output by the task similarity judgment module are used as input. By running this module, it is determined whether the client task is a new task, thereby updating the multi-task federated learning model library.

[0081] The specific system operation flow of this invention is as follows: Figure 3 As shown:

[0082] S1, Send Model Request. The edge client node collects normal operation data for this task scenario and converts the data into a specific format. When the edge client node sends a request for a customized monitoring model to the system, it also sends data in a specific format to the system's client node task identification module.

[0083] S2, Multi-task Federated Learning. Based on customer task feature data, existing models are selectively fused to provide a good initial model for customer nodes. This process includes the following steps: 1) Based on the customer's request for a customized monitoring model, the task similarity judgment module is run. This module runs multiple existing base models using customer data as input parameters and outputs the similarity between the customer node and each sub-node task in the multi-task federated model learning library. 2) Based on the customer node data, the task similarity module is run to obtain the similarity between the customer node and each sub-node task in the multi-task federated learning model library. This similarity is compared with a similarity threshold, and only the corresponding sub-node tasks with similarities above the threshold are retained as fusion objects. Their similarity is then converted into fusion weights with the customer node task. 3) Based on the fusion weights of the customer node task, the corresponding sub-node models in the multi-task federated model learning library are weighted and fused to generate a customized model for the customer node task.

[0084] S3, Model Deployment. The generated customized model and corresponding fault labels are deployed to edge client nodes via the TCP / IP protocol.

[0085] S4, Transfer Learning. After receiving the customized model, the edge client node uses local data to fine-tune the model and generate a locally personalized model.

[0086] S5, Model Recycling. The system sends a request to the cloud server node via TCP / IP protocol and uploads the localized model to complete the model recycling process.

[0087] S6, Model Library Update. Based on the fusion weights of the client node task with existing tasks, determine whether the node task belongs to a new task. If it belongs to a new task, the client-localized trained model is retrieved and added to the multi-task federated learning model library as the new node task. If it belongs to an existing task, the client-localized trained model is retrieved and weightedly fused with the corresponding base model to generalize the corresponding base model.

[0088] The method and system provided in this invention adopt an edge-cloud collaborative model training approach, reducing the requirements of edge computing and storage resources; it achieves knowledge sharing between different individuals through model sharing, avoiding large-scale data exchange and transmission; it fully considers the similarity and differences of assembly tasks, and generates customized initialization models based on the task characteristics of customer nodes, reducing model training costs and improving the utilization level of existing model knowledge.

[0089] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A training method for an assembly monitoring model based on federated transfer learning, characterized in that, The method includes: A unified neural network model is used to construct assembly monitoring models for each task. Each assembly monitoring model is then uploaded to the model library sub-node in the cloud server node to form an initialized multi-task federated learning model library. Define the task characteristic data of the client node; Based on task feature data and existing base models in the multi-task federated learning model library, a similarity calculation method is used to calculate the similarity between the current task of the client node and the existing base models of the model library sub-nodes. The similarity calculation method is as follows: retrieve the base models of all model library sub-nodes from the multi-task federated learning model library, use the task feature data of the client node to calculate the maximum classification probability of the model of each model library sub-node, and use it as the similarity between the client node and the base models of each sub-node in the multi-task federated learning model library. Based on the similarity between the current client node task and the existing base model, a preset similarity threshold is compared, and child node models with similarity values ​​above the threshold are selected as fusion objects. The similarity of child node models with similarity values ​​above the threshold is converted into fusion weights with the client node task. The federated fusion algorithm is used to fuse the models, generate a customized initial model for the client node task, and distribute the model to the client for local training. The client node uses local data to fine-tune the initial model obtained from the cloud server node to obtain a model suitable for the client node's current task monitoring. The client node then uploads this model suitable for the client node's current task monitoring to the cloud server node to update the multi-task federated learning model library and share model knowledge with other client nodes. Based on the similarity between the trained client node task model and the existing base model, a preset value threshold is compared to determine whether the client task is a new task relative to the model library sub-node task. If so, the client node's locally trained model is recycled as a new node task and added to the multi-task federated learning model library. Otherwise, the client node's locally trained model is recycled and weighted and fused with the existing model library sub-node model to generalize and update the existing model library sub-node model.

2. The assembly monitoring model training method based on federated transfer learning according to claim 1, characterized in that, The neural network model includes a convolutional neural network and fully connected layers.

3. The assembly monitoring model training method based on federated transfer learning according to claim 1, characterized in that, The task characteristic data of the client node is data that is stored in tensor form after preprocessing the original time-series data generated by the client node under normal operating conditions.

4. The assembly monitoring model training method based on federated transfer learning according to claim 1, characterized in that, The fusion weight is obtained by calculating the proportion of the corresponding similarity value in the sum of similarities, including: Filter out those with similarity higher than the similarity threshold The model library child nodes were renumbered from 1 to... The nodes are then assigned weights based on the similarity between the selected model library child nodes and the customer nodes; this is called the fusion weight. , The number of child nodes in the selected model library: In the formula, Selected from child nodes of the model library In the model, the first The similarity between the models Selected from child nodes of the model library In the model, the first The similarity between the models.

5. The assembly monitoring model training method based on federated transfer learning according to claim 4, characterized in that, The use of federated fusion algorithms for model fusion refers to the use of fusion weights. The parameters of the existing basic models in the selected model library sub-nodes. Weighted fusion is performed to obtain customized model parameters for customer nodes. : The parameters are the parameters of the existing basic models in the model library child nodes.

6. The assembly monitoring model training method based on federated transfer learning according to claim 1, characterized in that, Before performing local training on client nodes, the local data needs to be labeled using a unified data labeling method on the client nodes.

7. The assembly monitoring model training method based on federated transfer learning according to claim 1, characterized in that, The fine-tuning does not change the model structure; it only uses local data to correct the parameters of the obtained initial model.

8. The assembly monitoring model training method based on federated transfer learning according to claim 5, characterized in that, Value threshold This refers to a critical value, determined manually, that reflects the value of the model relative to the multi-task federated learning model library. It is used to select the update strategy for the multi-task federated learning model library, and its specific value can be determined through experience or mathematical relationships in practice. The generalization fusion weight and generalization update base model method are as follows: after retrieving the customer model, when it is determined to be an old task scenario, the maximum similarity is used to construct the generalization fusion weight between the customer model and the corresponding model library sub-node model; the generalization fusion weight and the customer model are used to perform generalization update on the corresponding model library sub-node model. The generalization fusion weight is calculated based on the similarity between the client node's shared model and the basic model of the existing model library sub-nodes, thus determining the generalization fusion weight between the client node's shared model and the basic model of the existing model library sub-nodes. The calculation formula is as follows: 。 9. A training system for an assembly monitoring model based on federated transfer learning, characterized in that, The system for implementing the method of any one of claims 1-8 comprises: A multi-task federated model learning library is used to store multi-task scenario monitoring models; The task feature data acquisition module is used to collect task feature data from edge client nodes; The task similarity judgment module is used to judge the similarity between the client node task scenario and the existing sub-node task scenario model in the multi-task federated model learning library. The federated model fusion module is used to fuse models from the multi-task federated model learning library and generate customized models for client node tasks. The model deployment module is used to distribute customized models and corresponding fault labels generated in the federated model fusion module to edge client nodes; The model recycling module is used to recycle customer-localized models; The federated model update module is used to update the multi-task federated model learning library.

Images