Research On Key Issues Of Intelligent Remote Monitoring System For YSU-Ⅱ Rehabilitation Robot

Rehabilitation robots have great advantages in precisely repeating scheduled training movements and improving patient training compliance,which can assist home-bound patients with rehabilitation training in the role of a rehabilitation doctor.However,the current intelligent level of rehabilitation robots is insufficient and cannot completely replace rehabilitation doctors in providing comprehensive rehabilitation guidance to patients.Therefore,constructing an intelligent remote monitoring system based on rehabilitation robots and endowing rehabilitation robots with some essential functions of doctors by utilizing artificial intelligence technology can promote the process of community and home-based rehabilitation.That is of great significance to improve the level of motor function of home-based rehabilitation patients,improve the quality of life of disabled and elderly people,and alleviate the contradiction between supply and demand of medical resources.In this thesis,the YSU-I rehabilitation robot is utilized as an experimental platform to establish the YSU-Ⅱ rehabilitation robot intelligent remote monitoring system.Several artificial intelligence technologies,including speech interaction technology,intelligent Question-Answering(QA)technology,and emotion recognition technology,are employed to enhance the intelligent level of the YSU-I rehabilitation robot.The focus of this work is to conduct algorithmic research on the key issues that exist in the system.Firstly,an instruction text proofreading method for rehabilitation robot training tasks is proposed to solve the problem that speech recognition errors seriously affect the efficiency of the voice-controlled subsystem of YSU-Ⅱ rehabilitation robot.This method establishes an instruction text proofreading corpus for YSU-Ⅱ rehabilitation robot training tasks and provides an error correction data augmentation method to expand the corpus.A Seq2 Seq model based on bidirectional gated recurrent units is constructed to detect and correct erroneous characters in the instruction text.Additionally,a Contextual and Keywords-based Attention(CK Attention)mechanism is designed to enhance the error proofreading performance of the model.The effectiveness of the proposed method is verified on a self-built instruction text corpus.The experimental results show that the proposed data augmentation method and CK Attention mechanism can effectively improve the performance of the text proofreading model,and the model has good practicality in the voice-controlled system.Secondly,a speech recognition text proofreading method for the limb rehabilitation knowledge QA subsystem is proposed to address the issue that speech recognition errors also seriously affect the efficiency of the QA system,and these errors present more complex than those encountered in the voice-controlled system.This method establishes a QA knowledge base by using data crawling technology to collect content related to limb rehabilitation from online medical consultation platforms.Five neural network-based Pinyin coding methods are designed to obtain the phonological information of Chinese characters.Moreover,a speech recognition text proofreading model based on Chinese Semantic and Phonological Information(CSPI)is constructed by using the proposed Pinyin coding methods,which can simultaneously adopt the contextual semantics and phonological information of sentences to correct the recognition errors.The effectiveness of the proposed method is verified on a self-built rehabilitation text proofreading dataset and the public AISHELL-3 dataset.The experimental results show that the phonological information carried by Pinyin contributes to detecting and correcting speech recognition errors,the CSPI-based speech recognition text proofreading model has good practicality in the QA system.Then,an Electroencephalogram(EEG)-based emotion recognition method for remote rehabilitation training is proposed to solve the problem that head shaking of patients causes missing facial expression frames in the remote rehabilitation training videos,which further leads to the inability of facial emotion recognition methods to continuously monitor changes in the emotional state of patients.Specially,a Tunable Q-factor Wavelet Transform(TQWT)-feature extraction method is established to reduce the impact of the non-linearity and non-stationarity of EEG signals on emotion recognition.As a new EEG representation,the TQWT Feature Block Sequence(TFBS)are transformed from TQWTfeatures by integrating the correlation,sub-band complementarity,and temporal information of multi-channel EEG signals.Furthermore,the HCRNN model is constructed to perceive the spatiotemporal information maintained in TFBS,and further predicts the emotion category.The effectiveness of the proposed method is verified on the public SEED dataset.The experimental results show that the developed TQWT features in highfrequency sub-bands are more discriminative than those in low-frequency sub-bands,and the recognition accuracy of HCRNN model with TFBS outperforms existing methods.Finally,most of the proposed methods are combined with existing artificial intelligence technologies to develop an intelligent remote monitoring system for YSU-II rehabilitation robot.This system consists of a voice-controlled subsystem for the rehabilitation robot,a QA subsystem for limb rehabilitation knowledge,and an emotion monitoring system for remote rehabilitation training.The implementation details for each subsystem are described in detail,along with the test results and their effectiveness.