Research On Multi-source Information Fusion Perception Technology Of Lower Limb Assist Suit

Thesis focuses on the problem of human motion intention perception based on multisource information,carries out the research on the motion perception of lower limb assist suit,and establishes the mapping relationship between multi-source signals and lower limb motion state by fusing and processing surface electromyography signals,joint angle signals and plantar pressure signals.So as to provide an effective information source for the lower limb assist suit.Firstly,the lower limb sensor layout scheme is designed and experiments are carried out to collect multi-source data.Through research and analysis,suitable lower limb muscles were screened for surface EMG signal acquisition.Two angle sensors were used to provide knee joint angle signals,and film pressure sensors were used to provide foot pressure signals to determine the overall scheme of lower limb sensing layout.In thesis,a variety of lower limb movement modes are selected and data acquisition is carried out to collect the surface EMG signals of human lower limbs such as daily walking on the ground,going up / down stairs,going up / down slopes,squatting and so on.Secondly,due to the high signal-to-noise ratio of the angle and pressure signals when the lower limbs are moving,the surface EMG signals are specially filtered and de-noised,and the signals are de-noised and reconstructed based on discrete wavelet transform.Faced with the inherent inconsistency of dimensions and sampling rates among signals of multiple modes,a signal normalization model is designed to achieve the normalization of multi-source signals.Faced with the problem of EMG signal feature extraction,an extraction method of directional metric features and a feature optimization method based on correlation information fusion are designed,which effectively improve the recognition efficiency and accuracy of the perception system.Based on multi-source fusion signals,the classification performance of various classifiers is compared,and it is verified that the modified k-nearest neighbor classifier(M-k NN)and empirical wavelet neural network(EW-NN)designed in Thesis have better recognition effect on each action pattern than other classifiers,reaching the recognition rates of 95.7% and 97.07% respectively.In addition,in order to give full play to the advantages of multi-source information,Thesis constructs a multi-classifier fusion system(MCFS)based on Bayesian theory.The model can realize distributed decision-making on multi-source input signals,and then realize motion perception through a multidimensional decision fusion layer.The recognition rate of the system reaches 98.33% for fixed motion patterns,94.3% for different motion environments,and 94.7% for different motion gaits.