Forearm Controller Based on BP Network and Sub-band Waves

In order to make the artificial forearm controller easier be trained and have higher robust, an adaptive controller for myoelectric signal (MES) is proposed. The control signal, MES, is derived from natural contraction patterns, which can be produced reliably with no subject training. To find features of MES, twenty-five filters with different center frequency and same bandwidth are designed and the feature curves of MES are shown in this paper. Based on the wavelet transform (WT) and the proposed feature curves, an ensemble of gate crossings based representations of MES is proposed. The gate for counting the gate crossings can change along with contractions, and this makes feature extraction adaptive to dissimilar contraction levels. Two-layer perceptron neural networks are used to classify a single site MES based on two features, specifically the gate crossings between 0 and 31.25Hz and the gate crossings between 31.25 and 62.5Hz. Based on the proposed controller scheme, the simulation result displays high accuracy rate.