Research On Motion Intention Based On Wearable Data Human Lower Limb Pattern And Phase Recognition

With the continuous development and popularization of wearable technology,the research on motion intention based on wearable data has become a research hotspot in the field of medical and industrial integration.It has important application value in the fields of human-computer interaction,rehabilitation engineering,exoskeleton robot and intelligent auxiliary equipment.Based on the support of the national key research and development plan ’stroke upper and lower limb rehabilitation robot design’ and other projects,this paper studies the basic theory and method of human lower limb motion intention recognition technology.The purpose of this study is to explore the application of human lower limb motion pattern and phase recognition based on wearable data in motion intention recognition.To this end,this paper first studies the acquisition and signal processing methods of lower limb motion information based on wearable sensors(inertial sensors and pressure sensors).Then,based on the collected lower limb motion data,how to establish an effective human lower limb motion pattern recognition model to distinguish different lower limb motion patterns is studied.Finally,a lower limb motion intention recognition framework is constructed by using these models,which can better identify and predict the motion pattern and motion phase of human lower limbs.The framework can be used to optimize the motion control of rehabilitation institutions such as exoskeleton robots,and provide more accurate and personalized rehabilitation programs for rehabilitation institutions.The main research contents of this paper are as follows:(1)In the process of designing human lower limb motion information acquisition system and signal and processing,aiming at the problem of redundancy between multi-point pressure sensor acquisition signals,a pressure sensor deployment scheme based on mutual information method is proposed to ensure information integrity while reducing the number of sensors and minimizing data redundancy.At the same time,aiming at the problem of time error in the original data moving average filtering,an improved moving average filtering algorithm is proposed,which can eliminate the time error before and after filtering as much as possible under the premise of obtaining the same quality smoothing effect.(2)In the process of constructing a lower limb motion pattern recognition framework based on the combination of feature engineering and machine learning,aiming at the shortcomings of the traditional recursive feature elimination method,which does not consider the correlation within the features,an improved recursive feature selection method based on cross-validation is proposed to search for the best feature subset.Additionally,parameter optimization was performed on various traditional machine learning algorithms such as K-Nearest Neighbors,Support Vector Machines,and Decision Trees.Finally,the effectiveness of the feature selection and model optimization algorithms was confirmed through experiments,and the most suitable machine learning algorithm for the dataset in this study was determined to be the Weighted K-Nearest Neighbors(GKNN)model.(3)In response to the limitations of traditional feature extraction methods and the tendency for overfitting in Convolutional Neural Network(CNN)models when dealing with small sample problems,a human lower limb motion pattern recognition model based on the combination of CNN and Weighted K-Nearest Neighbors(GKNN)is proposed.This model extracts features from raw motion data using CNN and performs classification recognition using GKNN.Experiments have shown that this model effectively addresses the small sample problem and achieves a significant performance improvement in lower limb motion pattern recognition tasks,with an overall recognition rate of 96.6%.(4)In the process of building the lower limb motion phase recognition model,improvements were made to the traditional Relief-F feature selection algorithm.An enhanced Relief-F feature selection algorithm based on temporal relationships was introduced to select more discriminative feature subsets,thereby enhancing the model’s construction effectiveness.At the same time,an improved HMM phase recognition method based on dwell time is adopted to improve the accuracy of phase recognition.Experimental results demonstrate that this approach achieved excellent recognition performance in phase recognition tasks for six different motion patterns.(5)Aiming at the lag and delay problems of the lower limb motion phase recognition model,a motion phase prediction model(BMAD)based on Bi-LSTM is proposed.The model improves the performance of the model through the multi-head attention mechanism and the DNN layer to predict the phase of the next possible movement of the human body.Finally,a motion intention recognition framework is constructed by combining the lower limb motion recognition model and the lower limb motion phase prediction model.Experiments show that the BMAD model has good prediction accuracy,which can reach 92.3%.The proposed motion intention recognition framework can accurately identify and predict the motion pattern and motion phase of the lower limbs.