Abstract—This paper presents a myoelectric knee joint angle estimation algorithm for control of active transfemoral prostheses, based on feature extraction and pattern classification. The feature extraction stage uses a combination of time domain and frequency domain methods (entropy of myoelectric signals and cepstral coefficients, respectively). Additionally, the methods are fused with data from proprioceptive sensors (gyroscopes), from which angular rate information is extracted using a Kalman filter. The algorithm uses a Levenberg-Marquardt neural network for estimating the intended knee joint angle. The proposed method is demonstrated in a normal volunteer, and the results are compared with pattern classification methods based solely on electromyographic data. The use of surface electromyographic signals and additional information related to proprioception improves the knee joint angle estimation precision and reduces estimation artifacts. Keywords - Electromyographic signals, proprioceptive sensors, entropy, cepstral coefficients, Kalman filter, transfemoral prostheses. I. INTRODUCTION LECTRONIC knees can be designed for providing different levels of damping during swing, and for adjusting to different walking speeds, assuming they have the appropriate sensors and control algorithms for estimating the knee joint angle and the walking speed. With the appropriate control algorithm, it is possible to program the prosthesis to allow the knee to flex and extend while bearing a subject’s weight (stance flexion). This feature of normal walking is not possible with conventional prostheses. Electronic knees use some form of computational intelligence to control the resistive torque about the knee. Several research groups have been involved in designing prototype knee controllers. Grimes et al. [1] developed an echo control scheme for gait control, in which a modified Alberto L. Delis is with the Dept. of Electrical Engineering, University of Brasília, Brasília-DF, Brazil, and the Medical Biophysics Center, University of Oriente, Santiago de Cuba, Cuba (lopez_delis@yahoo.com). João L. A. Carvalho, Geovany A. Borges, Icaro dos Santos, and Adson Ferreira da Rocha are with the Department of Electrical Engineering, University of Brasília, Brasília-DF, Brazil (joaoluiz@gmail.com, gaborges@ene.unb.br, adson@unb.br, icaro@ieee.org). Suélia S. Rodrigues is with the UnB-Gama Faculty, University of Brasília, Gama-DF, Brazil (rodrigues.suelia@gmail.com). knee trajectory from the sound leg is played back on the contralateral side. Popovic et al. [2] presented a battery- powered active knee joint actuated by DC motors, together with a finite state knee controller that utilizes robust position tracking control algorithm for gait control. A small number of companies have also developed electronic knee for clinical uses. For example, the Otto Bock C-leg [3] provides adjustable resistance for flexion and extension in swing through onboard intelligence and a special software package. Processing of surface electromyographic (SEMG) signals may be used in actively powered myoelectric prostheses for extracting command signals from muscle in the residual limb [4]. We recently proposed two different algorithms for estimating the intended knee joint angle from SEMG signals measured on opposing muscles of the upper-leg [5],[6]. The first method uses the auto-regressive model for feature extraction and a Levenberg-Marquardt (LM) multi-layer perceptron neural network for pattern classification [5]. The second method uses time-domain and frequency-domain SEMG feature extraction (amplitude histogram and AR model, respectively), self-organizing maps for feature projection, and a LM neural classifier [6]. For the development of an active leg prosthesis that also possesses ankle and foot axes, it is necessary to use other sources of information besides the SEMG signal (e.g. proprioceptive data). This could improve the precision of the prosthesis during movements of knee flexion and extension. Data fusion applied to myoelectric signals and proprioceptive sensors is capable of providing reliable myoelectric control [7]. The level of activity of muscles, either in isometric or isotonic contraction in dynamic limb motion, is the most important process to be recognized in myoelectric control. The combination of time domain features that represent the term of energy in the SEMG signal, with frequency domain features that show the muscle’s level of activation, provides good classification precision, is computationally efficient, and is more robust to electrode displacement [8]. This paper proposes an algorithm for estimation of intended knee joint angle from SEMG signals and proprioceptive sensor data, for the control of active transfemoral leg prostheses (Fig. 1). Two channels of SEMG Fusion of Electromyographic Signals with Proprioceptive Sensor Data in Myoelectric Pattern Recognition for Control of Active Transfemoral Leg Prostheses Alberto López Delis, João Luiz Azevedo de Carvalho, Geovany Araújo Borges, Suélia de Siqueira Rodrigues, Icaro dos Santos, and Adson Ferreira da Rocha E 4755 31st Annual International Conference of the IEEE EMBS Minneapolis, Minnesota, USA, September 2-6, 2009 978-1-4244-3296-7/09/$25.00 ©2009 IEEE