Evaluation of a Multiaxial Optical-Based Shear Sensor Using a Multilayer Perceptron Artificial Neural Network Model

Abstract

Use of tactile shear sensors is increasing, particularly in assistive devices. For example, shear force sensors can monitor forces between a residual limb and prosthetic socket that can result in discomfort, pain, or tissue breakdown. Previous work described a multiaxial shear sensor based on optoelectronic coupling between a broad-spectrum light-emitting diode and a photodiode with bandpass filters corresponding to red, green, and blue (RGB), and broad visible spectrum wavelengths. Shearing is detected based on changes in intensity at specific wavelengths when broad-spectrum light is reflected off a specified color pattern. The goal of this study was to develop a two-output multilayer perceptron (MLP) artificial neural network (ANN) approach for modeling the relationship between the four sensor outputs (RGB and broad-spectrum light) and shear displacement. Shear data from the sensor were collected by displacing in 1-mm increments on a modified computerized numerical control positioning stage for a total range of ±10 mm in the X (medial-lateral) and Y (anterior-posterior) directions. This process was repeated 10 times for a total (n) of 1100 datapoints. A custom hyperparameter tuning algorithm was used to find optimal hyperparameters for the MLP-ANN model. The MLP-ANN algorithm outputs resulted in an R² of X = 0.99 and Y = 0.99, and RMSE of X = 0.072 mm and Y = 0.11 mm. The final averaged 10-fold cross-validation score of both coordinates was 99.16% using randomized 80:20 (training:test) data partitions. The MLP algorithm demonstrated higher average accuracy than comparable single output algorithms reported previously.