There is an issue with the timely application of artificial intelligence to humankind, particularly to the disabled. In this paper, we propose a surface electromyography signal recognition system for prosthetic application to persons with disabled hands, which achieves simplicity, reliability, accuracy and a short response time by increasing the performance of each part of the system and the coordination between the interconnected components. First, a surface electromyography signal acquisition system was designed on the basis of the cost and processing speed. Second, a method called the ‘extreme value’ was carried out on the original signal containing five continuous movements, by separating the signal into isolated segments representing different postures, which made application of the system for daily use possible. Third, on the basis of time-domain, chaos-theory and time–frequency-domain analysis methods, four features, namely the average amplitude, fractal dimension, maximum Lyapunov exponent and wavelet coefficient, were extracted from four possible arm locations to be determined. Furthermore, the average amplitude from the extensor digitorum and the wavelet coefficient from the flexor pollicis longus were determined as the final features after comparing the clustering effects of the extracted features. Finally, a new strategy for classifying the different postures based on a back-propagation neural network was introduced to obtain an average system accuracy of 82.77% for five continuous movements.