A proof-of-concept classifier for acoustic signals from the knee joint on a FPAA

Abstract

A proof-of-concept low-power analog classifier for assessing acoustic signals from the knee joint on a reconfigurable Field Programmable Analog Array (FPAA) is presented in this paper. Knee joint sounds are measured using piezoelectric (contact) microphones and processed using the front end analog filters. A single layer of neural network composed of Vector Matrix Multiplication (VMM) and Winner-Take All (WTA) is used for the classification. A simple classifier detecting an anterior cruciate ligament injury is implemented here. Measurement from a single subject's healthy and injured knees are used here as an input. The FPAA is fabricated in a 350nm CMOS process. A bank of 12 parallel filters is used for feature extraction and a 12×2 VMM-WTA is used as a classifier. The compiled system, front-end and the classifier, consumes a power of 15.29μW with a power supply of 2.5 V.