A deep neural network for electrical resistance calibration of self-sensing carbon fiber polymer composites compatible with edge computing structural monitoring hardware electronics

Abstract

The self-sensing ability of materials, in particular carbon fiber polymer composites (SSCFPC), is a must-have requirement when designing a structural monitoring network for remote assessment of structural serviceability. This work presents a study using an Artificial Deep Neural Network (ADNN) wherein is evaluated the electrical resistance ( R) output of specimens subjected to an unchanged deformation state of 2.86% strain for prolonged periods of time. Six ADNN architectures are evaluated with varying numbers of neurons on pre-defined hidden layers, sharing the same four data inputs and one output. The dataset is based on 3276 data points collected during the experimental campaign of an innovative electrode design embedded in SSCFPC specimens. The effect of the number of iterations and the architecture of the neural network is investigated in proposed ADNN models. Simple moving average, and moving Standard Deviation, [Formula: see text], are determined and plotted in terms of z-score to assist in performance evaluation of proposed ADNN models. The optimal ADNN architecture is found among six proposed architectures and for each of the four SSCFPC mixtures. Results show the proposed model architectures are able to predict values of R with greater accuracy than traditional regression mathematical methods when traditional statistical coefficients are used. However, when analyzing data in a time-series manner, results show further research is needed to achieve optimal accuracy results. The analysis presented focused on the structural monitoring network infrastructure and hardware electronics compatibility for further development of this type of SSCFPC as a self-sensing composite material with ability of automatic calibration and suitable for real-time data acquisition and artificial intelligence modeling.