Use of Artificial Neural Networks in the Design of Adaptive Fuzzy Logic Controllers in the manufacturing of Prosthetic Knees

Abstract

This paper presents real-time parameter updates that help to make adaptive machining process robust to changes in exogenous operating conditions. This work supports the commercial imperative for a customizable ‘batch size of one’ implant that does not unreasonably affect operating costs like tool life. A Neural Network (NN) is presented for use within a Computer Numerical Control (CNC) manufacturing cell to determine real-time tool offset adjustments in knee prostheses. This study integrates additional sensor inputs from the CNC cell, including improved force monitoring data for critical tools used in the machining of a Tibial component. The resulting time series force model is used to compare classification performance on Random Forest (RF) and Bi-directional Long Short Term Memory (BiLSTM) Neural Networks. In relation to network training performance, pre-processing improvements are reported for the RF algorithm using a time series data conversion process step. For the case of the BiLSTM model, 2D time series data is converted to a 3D array using a novel projection technique. The accuracy of both of methods is assessed in real-time using tool offset control applied using an adaptive fuzzy logic law. In this use case it has been observed that a RF approach exhibits less overfitting than its BiLSTM counterpart and therefore is a better choice for the dynamic computation of the tool offset height, the actuator output.