A method for obtaining tactile texture features of surface materials based on federated learning

Abstract

This invention relates to the field of tactile internet technology and provides a method for acquiring tactile texture features of surface materials based on federated learning. The method includes the following steps: performing federated learning; using a deep learning-based image feature extraction method, employing a ResNet residual network model to extract features from surface tactile images; combining federated learning with stacked autoencoder (SAE) to extract tactile texture features, and using an average federated learning algorithm for parameter updates; combining federated learning with gated recurrent network (GRU) to acquire tactile texture features, and using an average federated learning algorithm for parameter updates; saving all model parameters after training, and finally acquiring tactile texture features through a mapping model and reconstructing the tactile texture signal through a decoder. This invention solves the problem of model performance degradation caused by insufficient user data in the tactile internet, while eliminating the need for users to transmit raw data, thus protecting user privacy and data security.

Description

Technical Field

[0001] This invention belongs to the field of tactile internet technology, and in particular relates to a method for obtaining tactile texture features of surface materials based on federated learning. Background Technology

[0002] Federated learning (FL) is a machine learning theory that has emerged in recent years. The concept was first proposed by Google in 2016, initially developed to address the problem of how Android clients can update models using local data. The goal of federated learning is to achieve efficient machine learning across multiple users or servers based on the following three points: (1) protecting data security when users and clients exchange large amounts of data; (2) ensuring terminal data security and user privacy; and (3) ensuring that the algorithm complies with laws and regulations. In practical applications, the machine learning algorithms that FL can use are not limited to traditional convolutional neural networks, multilayer perceptrons, and recurrent neural networks, but also include reinforcement learning algorithms such as random forests. Based on these advantages, federated learning can alleviate the problem of insufficient data that may be encountered in the process of acquiring tactile texture features of surface materials to a certain extent.

[0003] With economic and social development, the number of wearable and portable devices with certain data processing capabilities has increased significantly. These devices typically contain a large amount of data suitable for updating machine learning models. If this data can be used to update models, the user experience of smart devices will be greatly improved. For example, keyboard prediction modules and voice recognition modules can be more accurate and faster, photo albums can accurately select duplicate photos to discard, and navigation software can more accurately determine traffic congestion. However, this highly valuable data often involves user privacy and centralized collection requires a large amount of storage resources. This makes traditional centralized data collection and storage methods unsuitable. To address this issue, federated learning has been proposed and validated both theoretically and experimentally.

[0004] Training a surface tactile texture feature generation model requires a large amount of data to ensure the accuracy of the model. However, in the tactile internet, data transmission between users is often open-loop communication, and each user's data is limited. Therefore, tactile texture feature generation models trained by a single user in the tactile internet are prone to underfitting. To address the problem of low accuracy of cross-modal surface tactile texture feature acquisition models trained by a single user due to limited data, one possible solution is to centralize all users' data on a single server, then train the model and distribute the generated model to all users. However, from a theoretical analysis perspective, this method has the following drawbacks: (1) centralized storage of visual and tactile data requires a large amount of memory; (2) users may be unwilling to share raw data for privacy reasons; (3) the server needs powerful computing capabilities to process large amounts of data; and (4) whenever new tactile data is available for training, the model needs to be retrained with all the data, which consumes a lot of time. In summary, this type of method cannot meet the requirements of high-precision, low-latency tactile data acquisition in the tactile internet. Summary of the Invention

[0005] The purpose of this invention is to provide a method for obtaining tactile texture features of surface materials based on federated learning, which aims to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for obtaining tactile texture features of surface materials based on federated learning includes the following steps:

[0008] Step S1: Perform federated learning;

[0009] Step S2: Deep learning-based image feature extraction method, using the ResNet residual network model to extract features of surface tactile images;

[0010] Step S3: Combine federated learning with stacked autoencoder (SAE) to extract tactile texture features, and use the average federated learning algorithm to update the parameters of the extracted model.

[0011] Step S4: Combine federated learning with gated recurrent network (GRU) to obtain tactile texture features, and use average federated learning algorithm for parameter update;

[0012] Step S5: After training, save all model parameters, and finally obtain tactile texture features through the mapping model and reconstruct the tactile texture signal through the decoder.

[0013] Furthermore, the specific approach of step S1 is as follows:

[0014] The users form a fixed set of K clients, each with a fixed dataset of tactile and visual textures of surface materials. At the beginning of each round, the server distributes the model to all K users. The m selected users perform local model gradient calculations based on the model they received in this round and their local tactile texture datasets, and send the updates to the server. The server applies the received updates to the global model and sends the new model to all clients, repeating this process.

[0015] Furthermore, the specific approach of step S2 is as follows:

[0016] The ResNet50 residual network model consists of four large modules, each containing 3, 4, 6, and 3 smaller modules respectively. Each smaller module consists of 3 convolutional layers, which, together with the initial convolutional layer and the final fully connected layer, form a 50-layer convolutional neural network with residual connections. After several convolutions and pooling, the image dimension is reduced. The data from the last layer of the convolutional network is unfolded to obtain the image's feature vector, which is then connected to a fully connected layer or a softmax layer to complete the classification task. The effectiveness of the feature extraction method is verified by the classification accuracy.

[0017] Furthermore, the specific approach of step S3 is as follows:

[0018] For tactile texture feature extraction of surface materials based on stacked autoencoders, when combined with federated learning, the tactile texture signal compression and reconstruction model based on stacked autoencoders is updated through continuous model parameter transfer between the client and server, with the goal of minimizing the reconstruction error. Taking the outermost autoencoder as an example, the optimization objective is expressed as:

[0019]

[0020] The objective function is specifically expressed as:

[0021]

[0022]

[0023] The loss is calculated using model parameters Θ to obtain the force-tactile signal of the original surface material and the reconstructed force-tactile signal. The error; assuming there are K clients partitioning the data in a federated learning system. For the set of indexes of data points on client k, Therefore, the objective function is rewritten as:

[0024]

[0025]

[0026] The algorithm for updating the texture feature extraction model of surface material based on stacked autoencoder using average federated learning is as follows: The objective is to minimize F k (Θ), in each round, m users are randomly selected to form a client set S. t Users participate in this round of model updates. Users use local data to update and calculate the model gradient, upload the calculated model gradient to the client to calculate the mean, and then the client sends the new model parameters to the server. The server aggregates the received parameters and then distributes them to all users. This process is repeated until the model converges.

[0027] Furthermore, the specific approach of step S4 is as follows:

[0028] For the acquisition of tactile texture features of surface materials based on gated recurrent networks, the optimization objective when combined with federated learning is expressed as:

[0029]

[0030] The error function is specifically expressed as follows:

[0031]

[0032]

[0033] The error between the tactile texture features obtained from the visual texture image features of the surface material and the actual tactile texture features is calculated using model parameters Φ; assuming there are K clients partitioning the data. For the set of indexes of data points on client k, Therefore, the objective function is rewritten as:

[0034]

[0035]

[0036] The average federated learning algorithm is selected for parameter updates.

[0037] Furthermore, the specific update process for parameter updates in the average federated learning algorithm is as follows:

[0038] Each client in the K client set is labeled by k. B is the subgroup size of the local group, E is the number of iterations performed locally, and η is the learning rate.

[0039] The server executes: Initializes Θ0 and sends it to the client. For each round t=1,2,..., it executes m←max(C·K,1)S. t←A random set of m clients for each in k∈S t Clients in the process execute in parallel;

[0040]

[0041]

[0042] Client update (k,Θ), executed on the client. Will Divided into several small groups Execute for each local iteration from the 1st to the Eth, for each implement Then the parameter Θ is sent to the server.

[0043] Furthermore, the specific approach of step S5 is as follows:

[0044] After the training process is completed, the final updated model parameters are retained. The extracted visual image features are used to obtain tactile texture features through a mapping model. The obtained tactile texture features are then used to reconstruct the tactile texture signal through a decoder. The original force tactile signal is compared with the tactile signal obtained by the federated learning-based distributed surface tactile texture feature acquisition method to verify the feasibility of the method.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] This federated learning-based method for acquiring tactile texture features of surface materials combines federated learning with the acquisition of tactile texture features of surface materials to propose a distributed method. This method can solve the underfitting problem of the tactile texture data acquisition model of surface materials caused by insufficient data from a single user, while also protecting user privacy and data security by eliminating the need for users to transmit their original data. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the federated learning framework.

[0048] Figure 2 Reconstruct error maps for surface material tactile texture signals based on federated learning (5 users are selected in each round).

[0049] Figure 3 Error map for reconstructing tactile texture signal of surface material (single user).

[0050] Figure 4 This is a schematic diagram illustrating how the mapping error changes with the number of communication rounds.

[0051] Figure 5 This is a schematic diagram of the original and acquired force tactile signals. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0054] An embodiment of the present invention provides a method for obtaining tactile texture features of surface materials based on federated learning, comprising the following steps:

[0055] Step S1: Perform federated learning;

[0056] Step S2: Deep learning-based image feature extraction method, using the ResNet residual network model to extract features of surface tactile images;

[0057] Step S3: Combine federated learning with stacked autoencoder (SAE) to extract tactile texture features, and use the average federated learning algorithm to update the parameters of the extracted model.

[0058] Step S4: Combine federated learning with gated recurrent network (GRU) to obtain tactile texture features, and use average federated learning algorithm for parameter update;

[0059] Step S5: After training, save all model parameters, and finally obtain tactile texture features through the mapping model and reconstruct the tactile texture signal through the decoder.

[0060] In a preferred embodiment of the present invention, the specific approach of step S1 is as follows:

[0061] The users form a fixed set of K clients, each with a fixed dataset of tactile and visual textures of surface materials. At the beginning of each round, the server distributes the model to all K users. The m selected users perform local model gradient calculations based on the model they received in this round and their local tactile texture datasets, and send the updates to the server. The server applies the received updates to the global model and sends the new model to all clients, repeating this process.

[0062] In a preferred embodiment of the invention, federated learning aims to perform distributed model training among multiple users who possess data suitable for model training. The system has a central node, but the central node does not collect users' local samples or personal privacy data. A basic network diagram is shown below. Figure 1As shown. In each round of communication, the central node transmits model parameters or intermediate results (such as gradients) to the user, constructing a global model, thereby simultaneously achieving data privacy protection and computing power sharing. In federated learning and the acquisition of tactile texture features of surface materials, a synchronous update scheme is assumed, which is carried out cyclically in a certain number of rounds.

[0063] In a preferred embodiment of the present invention, the specific approach of step S2 is as follows:

[0064] The ResNet50 residual network model consists of four large modules, each containing 3, 4, 6, and 3 smaller modules respectively. Each smaller module consists of 3 convolutional layers, which, together with the initial convolutional layer and the final fully connected layer, form a 50-layer convolutional neural network with residual connections. After several convolutions and pooling, the image dimension is reduced. The data from the last layer of the convolutional network is unfolded to obtain the image's feature vector, which is then connected to a fully connected layer or a softmax layer to complete the classification task. The effectiveness of the feature extraction method is verified by the classification accuracy.

[0065] In this embodiment of the invention, preferably, an advanced image processing network structure, Residual Network (ResNet), is used for feature extraction from surface tactile texture images. Specifically, ResNet proposes a solution through data preprocessing and the use of batch regularization layers and residual modules within the network. ResNet can significantly reduce model parameters and computational load while only slightly decreasing accuracy. Compared to traditional convolutional neural networks where data is passed layer by layer through parameters, ResNet allows certain layers to skip the next layer of neurons, directly connecting to adjacent layers without computation. This approach weakens the strong connections between layers and better preserves image information. Furthermore, by introducing batch regularization, ResNet can suppress the amplification of small parameter changes as the network depth increases, making the network more adaptable to parameter sizes. Therefore, ResNet, with these advantages, can mitigate the performance degradation caused by model depth.

[0066] In a preferred embodiment of the present invention, the specific approach of step S3 is as follows:

[0067] For tactile texture feature extraction of surface materials based on stacked autoencoders, when combined with federated learning, the tactile texture signal compression and reconstruction model based on stacked autoencoders is updated through continuous model parameter transfer between the client and server, with the goal of minimizing the reconstruction error. Taking the outermost autoencoder as an example, the optimization objective is expressed as:

[0068]

[0069] The objective function is specifically expressed as:

[0070]

[0071]

[0072] The loss is calculated using model parameters Θ to obtain the force-tactile signal of the original surface material and the reconstructed force-tactile signal. The error is the same for other automatic encoders and decoders that further extract features; assuming there are K clients partitioning the data in a federated learning system, For the set of indexes of data points on client k, Therefore, the objective function is rewritten as:

[0073]

[0074]

[0075] The algorithm for updating the texture feature extraction model of surface material based on stacked autoencoder using average federated learning is as follows: The objective is to minimize F k (Θ), in each round, m users are randomly selected to form a client set S. t Users participate in this round of model updates. Users use local data to update and calculate the model gradient, upload the calculated model gradient to the client to calculate the mean, and then the client sends the new model parameters to the server. The server aggregates the received parameters and then distributes them to all users. This process is repeated until the model converges.

[0076] In a preferred embodiment of the present invention, the specific approach of step S4 is as follows:

[0077] For the acquisition of tactile texture features of surface materials based on gated recurrent networks, the optimization objective when combined with federated learning is expressed as:

[0078]

[0079] The error function is specifically expressed as follows:

[0080]

[0081]

[0082] The error between the tactile texture features obtained from the visual texture image features of the surface material and the actual tactile texture features is calculated using model parameters Φ; assuming there are K clients partitioning the data. For the set of indexes of data points on client k, Therefore, the objective function is rewritten as:

[0083]

[0084]

[0085] The average federated learning algorithm is selected for parameter updates.

[0086] In this embodiment of the invention, preferably, since the visual texture image feature extraction of the surface material used in this invention is performed through a neural network pre-trained on ImageNet, the image feature extraction network does not need to participate in federated learning training. After the parameters of the stacked autoencoder are stored in each client, the parameters of the gated recurrent network are updated. Similar to the stacked autoencoder, this invention uses an average federated learning algorithm for parameter updates.

[0087] In a preferred embodiment of the present invention, the specific update process for parameter updating by the average federated learning algorithm is as follows:

[0088] Each client in the K client set is labeled by k. B is the subgroup size of the local group, E is the number of iterations performed locally, and η is the learning rate.

[0089] The server executes: Initializes Θ0 and sends it to the client. For each round t=1,2,..., it executes m←max(C·K,1)S. t ←A random set of m clients for each in k∈S t Clients in the process execute in parallel;

[0090]

[0091]

[0092] Client update (k,Θ), executed on the client. Will Divided into several small groups Execute for each local iteration from the 1st to the Eth, for each implement Then the parameter Θ is sent to the server.

[0093] In a preferred embodiment of the present invention, the specific approach of step S5 is as follows:

[0094] After the training process is completed, the final updated model parameters are retained. The extracted visual image features are used to obtain tactile texture features through a mapping model. The obtained tactile texture features are then used to reconstruct the tactile texture signal through a decoder. The original force tactile signal is compared with the tactile signal obtained by the federated learning-based distributed surface tactile texture feature acquisition method to verify the feasibility of the method.

[0095] Example 1

[0096] Simulation Experiment: The following analysis uses simulation experiment data to examine the applicability of the proposed method for obtaining surface material tactile texture features based on federated learning to force tactile signals. The software used in the simulation experiment is PyCharm software with the Keras framework.

[0097] Numerical validation is performed on a method combining federated learning with tactile texture feature extraction from surface materials. In this simulation, the invention considers five users (K=5) and a central server. The HapTex database is divided into six groups, each containing 80 data samples collected from 20 materials. Five data samples are distributed to the five clients for local model updates, and the remaining one group serves as a test sample. The simulation experiments use TensorFlow Federated, an open-source framework for machine learning and other decentralized data computations, on a desktop computer with an Intel(R) Core(TM) i7-7700 CPU 3.6GHz and 16GB RAM. The optimizer is the gradient descent solver Adam. In this invention, the tactile texture feature extraction model for surface materials is first trained, i.e., the users collaboratively train a stacked autoencoder / decoder with a compression ratio of N / M=100, meaning the compressed tactile texture signal has a dimension of 10. The activation function is ReLU, and the model structure is {1000, 512, 10}.

[0098] During the communication process, this invention attempted two update methods:

[0099] One approach involves communicating with all five users (m=5) at a time. After the server distributes the initialized model to the users in Zhejiang West, each user calculates the model update using their local data and then passes the gradient to the server. The server aggregates the data and distributes the new model to all users. This update cycle repeats, with an average communication time of 0.89 seconds per round. After training, all model parameters are saved. At this point, the stacked autoencoder's relative percentage errors for reconstructing the normal and tangential force signals in the test set are 8.66% and 9.45%, respectively.

[0100] Another approach involves randomly selecting three users (m=3) in each round of communication. These selected users calculate the model gradient using their local surface material tactile texture data and upload it to the server. The server then updates the model and distributes the new model to all five users. This update cycle continues until the error stabilizes and the model update stops. At this point, each user possesses a final, fitted tactile texture feature extraction model for the surface material. This update mode considers that some users may not be able to participate in each round of updates during actual communication. Furthermore, this user selection method effectively reduces the amount of data transmitted in each round of updates, saving communication resources and further alleviating the computational pressure on the server nodes. Results show that under these conditions, the surface material tactile feature extraction model can still converge. Once a stable state is reached, the model is saved. At this point, the relative percentage error of the test set is: normal force 8.73, tangential force 9.52.

[0101] For comparison, this invention conducted a simulation experiment with only one user participating. The model structure was the same as described above, meaning the tactile texture feature extraction model for the surface material was updated and calculated only using the user's local data. The error curve is shown below. Figure 3 As shown, the final training error differs significantly from the test error. At this point, the relative percentage errors for the test set are: normal force 12.19% and tangential force 13.23%. Compared to training with multiple users, the reconstruction error of the surface material force tactile signal feature extraction model trained by a single user is higher. This demonstrates that federated learning can be used to address the problem of low accuracy in surface material feature acquisition models in the tactile internet due to insufficient tactile data from a single user.

[0102] The ResNet network extracts features from visual texture images into a low-dimensional space. These extracted features are then used in a gated recurrent network (GRU) to perform visual texture modality transformation from surface materials. The activation function in the tactile texture feature acquisition model is ReLU. The visual texture features are fed into the GRU as input to the transformation model to acquire the visual texture features of the surface materials. To prevent overfitting, L2 regularization is introduced in each layer of the transformation model with a regularization coefficient of 0.001. Dropout layers are also introduced in the GRU and fully connected layers, with a dropout ratio of 0.5. This means that in each gradient update, 50% of the parameters are randomly selected and do not participate in the model update, thus mitigating overfitting to some extent.

[0103] This invention employs two user selection modes: the central server communicates with all 5 users in each round, or with 3 randomly selected users in each round. The invention tests a model for acquiring visual texture features of surface materials and tactile texture features using features extracted by the ResNet model. The error variation curve of the visual texture feature acquisition model extracted by the ResNet model under the condition of communicating with 5 users in each round is shown below. Figure 4 As shown. From Figure 4 As can be seen, after 100 rounds of communication, the model gradually stabilizes, and the error is close to that during intensive training.

[0104] In each round of communication, updates from all users are collected. After 200 rounds of iterative updates, the model is saved. In the test set, the relative percentage errors between the tangential force and normal force obtained from the visual texture image features of surface materials and the original signals are 15.44 for the tangential force and 17.99 for the normal force. These errors are similar to those in the centralized learning approach, demonstrating the feasibility of the distributed model training scheme based on federated learning.

[0105] Updates from 3 users were collected each time, and the model was saved after 200 rounds of updates. The relative percentage errors between the tangential and normal forces obtained from the visual texture image features of the surface material in the test set and the original tactile texture signals were 16.67 for the tangential force and 18.27 for the normal force. Compared with selecting five users per round, the model error increased when selecting 3 users per round because the average number of training sessions per user was reduced by 40%.

[0106] The above experimental data demonstrates that the federated update mechanism enables federated learning to achieve results similar to centralized model training without transmitting the original surface material visual-tactile data. This update method fully utilizes user computing resources while effectively protecting user data security and privacy. Furthermore, the error when only three users participate in training per round is higher than the error when all users participate in training per round. This indicates that federated learning can solve the underfitting problem of surface tactile texture feature acquisition models caused by insufficient data from individual users.

[0107] The tactile texture feature acquisition model for surface materials constructed in the above manner acquires a set of tactile texture signals of the surface material, such as... Figure 5 As shown, from Figure 5 It can be seen that the tactile texture signal of the surface material obtained from the visual image can fit the changing trend of the tactile texture signal of the surface material directly collected.

[0108] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A method for acquiring tactile texture features of surface materials based on federated learning, characterized in that, Includes the following steps: Step S1: Perform federated learning; Step S2: Deep learning-based image feature extraction method, using the ResNet residual network model to extract features of surface tactile images; Step S3: Combine federated learning with stacked autoencoder (SAE) to extract tactile texture features, and use the average federated learning algorithm to update the parameters of the extracted model. Step S4: Combine federated learning with gated recurrent network (GRU) to obtain tactile texture features, and use average federated learning algorithm for parameter update; Step S5: After training, save all model parameters, and finally obtain tactile texture features through the mapping model and reconstruct the tactile texture signal through the decoder; The specific approach for step S3 is as follows: For tactile texture feature extraction of surface materials based on stacked autoencoders, when combined with federated learning, the tactile texture signal compression and reconstruction model based on stacked autoencoders is updated through continuous model parameter transfer between the client and server, with the goal of minimizing the reconstruction error. Taking the outermost autoencoder as an example, the optimization objective is expressed as: ； The objective function is specifically expressed as: ； ； Loss is achieved using model parameters Calculate the force-tactile signal of the original surface material and the reconstructed force-tactile signal. Error; assuming there are errors in the federated learning system. Each client partitions the data. For the client The index set of the data points Therefore, the objective function is rewritten as: ； ； The algorithm for updating the texture feature extraction model of surface material based on stacked autoencoder using average federated learning is as follows: The objective is to minimize... Randomly selected in each round A client set consists of 10 users. Users participate in this round of model updates. Users use local data to update and calculate the model gradient, upload the calculated model gradient to the client to calculate the mean, and then the client sends the new model parameters to the server. The server aggregates the received parameters and then distributes them to all users. This process is repeated until the model converges.

2. The method for obtaining surface material tactile texture features based on federated learning according to claim 1, characterized in that, The specific approach for step S1 is as follows: Users form a fixed group containing A collection of clients, each with a fixed dataset of tactile and visual information about surface materials; at the beginning of each round, the server distributes the model to all clients. One user, of which Each selected user performs local model gradient calculations based on the model and its local haptic texture dataset received in this round of updates, and sends the update to the server; the server applies the received update to the global model update, sends the new model to all clients, and repeats this process.

3. The method for obtaining surface material tactile texture features based on federated learning according to claim 1, characterized in that, The specific approach for step S2 is as follows: The ResNet50 residual network model consists of four large modules, each containing 3, 4, 6, and 3 smaller modules respectively. Each smaller module consists of 3 convolutional layers, which, together with the initial convolutional layer and the final fully connected layer, form a 50-layer convolutional neural network with residual connections. After several convolutions and pooling operations, the image dimension is reduced. The data from the last layer of the convolutional network is unfolded to obtain the image's feature vector, which is then connected to a fully connected layer or a softmax layer to complete the classification task. The effectiveness of the feature extraction method is verified by the classification accuracy.

4. The method for obtaining surface material tactile texture features based on federated learning according to claim 1, characterized in that, The specific approach for step S4 is as follows: For tactile texture feature acquisition of surface materials based on gated recurrent networks, the optimization objective when combined with federated learning is expressed as: ； The error function is specifically expressed as follows: ； ； Using model parameters Calculate the error between the tactile texture features obtained from the visual texture image features of the surface material and the actual tactile texture features; assume that... Each client partitions the data. For the client The index set of the data points Therefore, the objective function is rewritten as: ； ； The average federated learning algorithm is selected for parameter updates.

5. The method for obtaining surface material tactile texture features based on federated learning according to claim 1, characterized in that, The specific update process for parameter updates in the average federated learning algorithm is as follows: Each client in the client set passes through mark This refers to the size of a local group. Number of iterations performed locally The learning rate; Server execution: Initialization And send to the client every round implement A random set of clients for each Clients in the process execute in parallel; Client update ; ； Client update Execute on the client Will Divided into several small groups For each local number from the 1st to the 2nd Each iteration is executed, for each implement Then the parameters Send to the server.

6. The method for obtaining surface material tactile texture features based on federated learning according to claim 1, characterized in that, The specific approach for step S5 is as follows: After the training process is completed, the final updated model parameters are retained. The extracted visual image features are used to obtain tactile texture features through a mapping model. The obtained tactile texture features are then used to reconstruct the tactile texture signal through a decoder. The original force tactile signal is compared with the tactile signal obtained by the federated learning-based distributed surface tactile texture feature acquisition method to verify the feasibility of the method.

Images