Integrated control system of leg prostheses based on deep learning neural network

• Main Authors: Widhiada, I. Wayan, Kusuma, I. Gusti Bagus Wijaya, Diputra, Anak Agung Gede Pradnyana, Winata, I. Made Putra Arya, Budiarsa, I. Nyoman, Indra, Cokorda Gede
• Format: Conference Proceeding
• Language: English
• Subjects: Amputation; Artificial neural networks; Control systems; Deep learning; Gait; Human motion; Joints (anatomy); Knee; Machine learning; Multilayer perceptrons; Multilayers; Neural networks; Prostheses; Recurrent neural networks; Sensors; Thigh
• ISSN: 0094-243X; 1551-7616
• DOI: 10.1063/5.0199791

Physical disabilities cause significant distress in people’s daily routine, especially for leg amputees as their movement is restricted. The civilized society has the responsibility to provide them with better prosthetics so they can live a normal life. Previously, prosthetic legs were developed for lower leg amputees. Today’s electronic prosthesis provides a wide range of mobility compared to a mechanical prosthesis. Powered prosthetics are quite expensive and hence not widely used. We have proposed a low-cost control system above-knee powered prosthetics leg that is reliable, and the complexity of the model is less. The prostheses works by taking input from a sensor placed in the model and six FSR402 sensor and an IMU sensor that record force muscle activity of the thigh during variation of leg movement. The value of motion of the joint knee angle is determined by sensors. In this paper, we are discussing and comparing potential deep learning control system namely artificial neural network (ANN) as baseline and recurrent neural network-long short term memory (RNN-LSTM) to predict human gait cycle. A multi-layer perceptron (MLP) method is implemented. In general, network architecture consists of input layer which is represents by six inputs of the FSR402 sensor and an IMU sensor, fully connected layers, and an output layer as joint knee angle. As the result model of ANN as baseline has metric value MAE and R squared are 6.21 and 68.7% respectively whereas model of RNN-LSTM has metric value MAE and R squared are 4.99 and 96.9% respectively. The result indicates the RNN-LSTM model is most likely to perform fewer errors based on the dataset. RNN-LSTM outperforms the task with a greater value compared to the ANN model as the baseline. This present study suggests that deep learning with the recurrent neural network-long short term memory (RNN-LSTM) is suitable for the prediction of human gait cycle movement from the FSR402 and an IMU signal in the overall walking task.