| With the increasing popularity of intelligent robots in production and life,the society has put forward higher requirements for the ability of robots to interact with the environment.The stable grasping ability of the robot in various environments is the basis for realizing other fine operations.It is of great significance to identify the stability of the robot multi-fingered dexterous hands grasping objects.However,for the tactile time series classification task,the traditional classification algorithms do not consider the potential relationship between the data at different times,resulting in the unsatisfactory classification effect.How to fully extract tactile time series information to improve the effect of grasping stability classification is very challenging.In this paper,based on the preprocessing of tactile data,an improved algorithm for grasping stability classification of multi-fingered dexterous hands based on Long Short-Term Memory(LSTM)is designed and implemented,and the grasping stability classification algorithm based on Graph Neural Network is researched and implemented.The paper focuses on the design and implementation of the multimodal feature fusion grasping stability classification algorithm based on deep learning.Finally,the design and implementation of grasping stability identification system for multi-fingered dexterous hands are completed.The main work of the paper is as follows:(1)An improved multi-fingered dexterous hands grasping stability classification algorithm based on LSTM is proposed.The tactile data is framed by sliding window,the LSTM network is used to extract the features of the data,and the attention weight is added to the output by the residual self-attention mechanism.An improved LSTM classification model is obtained.The experimental results under different network structures and hyperparameter conditions are compared and analyzed,and the classification of the grasping stability of the multi-fingered dexterous hands is realized.(2)The classification algorithm of grasping stability of multi-fingered dexterous hands based on Graph Neural Network is researched and implemented.According to the actual positions of the electrodes on the sensors,the electrode values are converted into graph data.The classification effects of grasping stability of Graph Convolutional Network and Graph Attention Network under different graph structures are studied,and the comparative analysis of experimental results under different activation functions is completed.(3)A multimodal feature fusion grasping stability classification algorithm based on deep learning is proposed.Aiming at the problem of insufficient information provided by single-modal tactile data,a classification model that can fuse multimodal data is designed.The pressure,temperature and electrode value vectors are learned and compressed by Auto-Encoder.The feature level fusion of multimodal data is realized through hidden layer vector splicing,and the classification of grasping stability of multi-fingered dexterous hands is completed,which effectively improves the classification accuracy.(4)The grasping stability identification system of multi-fingered dexterous hands is designed and implemented.The functions of the system are designed in general.On the basis of multimodal feature fusion classification algorithm,the stability identification module and classification model training module are designed,and the interactive user interface is built,and the function test of the system is completed. |
