Surface EMG Signal Pattern Classification Based On Information Fusion

The surface electromyography (SEMG) signal is a kind of electrophysiological signal which is generated by the neuromuscular activity during voluntary movement and directly reflects the muscle activity. Extracting effective features from SEMG for correct classification is the key problem in the EMG-based prosthesis control system. This paper proposes an approach using information fusion technique to improve the performance of the classification system and investigates the probable solutions for the problems during information fusion.In the first section, a brief overview of current research for the feature extraction techniques and classification methods used in SEMG signals is given and the application of information fusion in pattern recognition is introduced. In the following section, feature extraction for real-time EMG-based prosthesis control is discussed. To achieve the goal of the control system, it is required that the features have good separability and the computational complexity is low. The statistical characteristics of the signals in the time domain, the coefficients of AR model and the singular values of the coefficient matrices of wavelet transform are extracted from the original signals as inputs for each classifier. In the third section, fusions of decision-making level are discussed using Dempster-Shafer evidence theory and fuzzy integral. In the fusion system based on D-S evidence theory, for the conflict evidences, the model of evidence source is modified using averaged evidence and combined using the D-S combination rules. The experimental results show that the problem of confict evidences could be solved effectively when the most evidences are correct. In the fuzzy integral systems, an efficient method for the computation of fuzzy density is given after comparing two approaches for fuzzy density computation, which solves the problem of determining the parameters for the fusion of the decision-making level. Finally, the effects of the two systems, fusion systems using D-S evidence theory and fuzzy integral, are compared. The results reveal that the average accuracies of the two systems are superior to those of using each classifier only. The effect of D-S evidence theory is "the minority subordinates to the majority", where the creditability is not sufficinetly taken into consideration. On the other hand, the fuzzy integral theory simultaneously considers the objective estimation of each evidence and the importance of each evidence into consideration, which illustrates the applicable potential for EMG-based control system.