TECHNICAL ARTICLE A Virtual Instrumentation Approach to Neural Network-Based Thermistor Linearization on Field Programmable Gate Array K.P.S. Rana, N. Mittra, N. Pramanik, P. Dwivedi, and P. Mahajan Instrumentation and Control Engineering Department, Netaji Subhas Institute of Technology, New Delhi, India Keywords Sensor Linearization, Neural Network, FPGA, Thermistor, Thermal Sensor, Nonlinearity Correspondence K.P.S. Rana, Instrumentation and Control Engineering Department, Netaji Subhas Institute of Technology, New Delhi, India Email: kpsrana1@gmail.com Received: August 22, 2012; accepted: November 19, 2012 doi:10.1111/ext.12011 Abstract Temperature measurement is an important industrial requirement in several applications. Thermistor, in particular, is used to a great extent for this purpose in many industrial applications as it is cost effective, relatively small in size, and has better sensitivity as compared to its counterparts. It offers a moderate range of temperature sensing typically from −55 ◦ C to 125 ◦ C. On the other hand, thermistor is a highly nonlinear sensor as it is characterized by the exponential dependency of resistance on temperature. Effective usage of thermistor thus requires some mechanism for linearization. This paper presents a simple step-by-step, practically implementable artificial neural network (ANN)-based linearization method for thermistor characteristic using a two-layer neural network having two neurons in each layer. The trained feed-forward neural network is implemented on a field programmable gate array (FPGA) on the NI-PXI platform for real-time measurement. Validation of the proposed technique was carried out experimentally using a comparative study. A precise thermocouple-based temperature measurement system was utilized for this purpose. The temperature readings were recorded after allowing both the sensors to settle, and a maximum error of ±0.9 ◦ C was obtained in the experimental measurement range of 5 ◦ C–65 ◦ C. Introduction Industrial instrumentation basically involves mea- surement and control of various physical variables. Measurement of these variables is performed with the help of sensors. One of the most desired properties of the sensing devices is linearity. However, the sensors involved in the measurement of these variables are often nonlinear, and thus they require some mechanism for linearization. Thermistor is one such widely used temperature sensor with nonlinear characteristics. 1 A negative temperature coefficient (NTC) thermistor can be characterized by the fol- lowing resistance temperature relationship given by Eq. 1. Figure 1 shows the variation of the resistance of an industrial grade NTC thermistor (Model No. B57164K0101J000, M/S EPCOS, Munich, Bavaria) with the input temperature having the parameters as β = 3200 K, R 0 = 100 , T 0 = 298 K. As seen clearly, the resistance value varies from 15 to 995  for the input temperature range of 240–361 K. R = R 0 e β  1 T − 1 T 0  (1) where T 0 is the reference temperature, R 0 the resistance at temperature T 0 , β the thermistor constant, T the temperature at a particular instant, and R the resistance at temperature T . As seen from Eq. 1 and Fig. 1, the sensor has an exponential variation of resistance with temperature leading to good amount of nonlinearity. So, it requires some mechanism for linearization. Thermistor lin- earization has been a topic of research for scientists and engineers for a long time. Several techniques have been proposed as solutions for its nonlinearity. An early stage technique made use of simple passive ele- ments like resistances in series or in parallel with the thermistor to linearize its output. 2 As reported, this Experimental Techniques 39 (2015) 23 – 30  2013, Society for Experimental Mechanics 23