Output Prediction of Helical Microfiber Temperature Sensors in Cycling Measurement by Deep Learning

Abstract The inconsistent response curve of delicate micro/nanofiber (MNF) sensors during cycling measurement is one of the main factors which greatly limit their practical application. In this paper, we proposed a temperature sensor based on the copper rod-supported helical microfiber (HMF). The HMF sensors exhibited different light intensity-temperature response relationships in single-cycle measurements. Two neural networks, the deep belief network (DBN) and the backpropagation neural network (BPNN), were employed respectively to predict the temperature of the HMF sensor in different sensing processes. The input variables of the network were the sensor geometric parameters (the microfiber diameter, wrapped length, coiled turns, and helical angle) and the output optical intensity under different working processes. The root mean square error (RMSE) and Pearson correlation coefficient (R) were used to evaluate the predictive ability of the networks. The DBN with two restricted Boltzmann machines (RBMs) provided the best temperature prediction results (RMSE and R of the heating process are 0.9705 °C and 0.9969, while the values of RMSE and R of the cooling process are 0.786 6 °C and 0.997 7, respectively). The prediction results obtained by the optimal BPNN (five hidden layers, 10 neurons in each layer, RMSE=1.126 6 °C, R=0.995 7) were slightly inferior to those obtained by the DBN. The neural network could accurately and reliably predict the response of the HMF sensor in cycling operation, which provided the possibility for the flexible application of the complex MNF sensor in a wide sensing range.