This study presents a method for supervised classification of multi-channel surface electromyography (SEMG) signals with the aim of recognizing drivers' lumbar muscle fatigue during prolonged driving. An experiment was carried out to investigate the SEMG manifestations of 8 drivers' lumbar muscle fatigue with recording of SEMG from 4 locations over lumbar erector spinae. Based on the wavelet packet transform (WPT) and continuous wavelet transform (CWT) of each SEMG segment, a representation space composed of 176-dimension features was extracted to classify three muscle fatigue statuses. The 176D features were calculated from Shannon entropy and relative energy of wavelet packets, along with instantaneous median frequency (IMDF), mean frequency (IMNF), and energy (IE) from CWT of wavelet packet (4, 15). The classification was performed by a C typed support vector machine (SVM) with a radial basis function (RBF) kernel, which was compared with a linear kernel. Parameters of SVM were optimized with the grid search method. Results: Correct classification rate (CCR) of the testing set was around 82.69 % (1.46 %)-an average (STD) value from 10 successive tests using a RBF-SVC, while the accuracy dropped to 78.94 % (1.63 %) with a linear kernel. Optimum parameters (c, gamma) for the RBF kernel were identified to be (110, 0.082), which affected the classification capacity in a serious way. The AUC (normalized area under the curve) values (0-1) of receiver operating characteristic (ROC) curves for the 10 successive tests were all above 0.9, which proved our method to be reliable and promising in a detection system of drivers' lumbar muscle fatigue.
展开▼
