Machine Learning-Assisted Data Interpolating Flexible Temperature Sensor Array

Flexible sensors, which can be installed even on nonplanar objects, are expected to collect many datasets for data analyses. Among them, flexible temperature sensor is promising for application of environmental monitoring, healthcare, and robotics. To realize temperature sensor platforms for such applications, sensor stability, reliability, and high sensitivity are necessary to analyze datasets precisely. Furthermore, low power consumption of sensor system is another challenge, which in general each sensor consumes power to operate it, resulting in huge power consumption for Internet of Things. Although many temperature sensor concepts have been reported previously, the sensor that combine such characteristics has yet to be fully developed. To address these challenges, we aim to realize low power, high stability, and high spatial-resolution mapping of temperature distribution using solution-based V₂O₅ nanowire network-based flexible temperature sensor combining with machine learning system. Solution-based process allows to economically fabricate large-scale sensor on a flexible substrate. To make highly stable and relatively high temperature sensor sensitivity, inorganic V₂O₅ nanowire network was selected to form thin film on a flexible substrate. Furthermore, this sensor has almost no hysteresis behavior for temperature sensing. For low power consumption concept, data interpolation is applied to predict the data between sensors with reservoir computing (RC), realizing that low power consumption can be achieved due to less number of sensors required. This enables real-time highly spatial resolution detection of temperature distribution and contact position.<br/>First, fundamental characteristics of the temperature sensor were conducted. Resistance change was measured against temperature changes from 25 to 60 °C by using an environmentally controlled oven. During the measurement, resistance of temperature sensor decreases with increasing temperature due to V₂O₅ nanowire properties. Temperature sweeps (increasing and decreasing) were also conducted to discuss hysteresis of the sensor. To find optimal conditions of the sensor, all measurements were tested by changing annealing temperature conditions from room temperature to 250 °C after V₂O₅ nanowire network formation. The results show that smaller hysteresis is realized when annealing temperature increases. In particular, the sensor annealed at 250 °C shows good properties of sensitivity of ~ -1.5 %/°C and sensitivity difference (i.e. hysteresis) of ~ 0.024 %/°C between temperature sweeps. The long-term reliability of the sensor was evaluated by subjecting it to thermal cycles between 25 °C and 60 °C for over 110 hours in the environmentally controlled oven. During the first 20 hours, the resistance values increased slightly by 0.5 %, corresponding to 0.33 °C fluctuation. However, after 20 hours, the sensor stabilized with no observable resistance shifts, indicating high stability of temperature sensor.<br/>For the data interpolation using RC, the sensors were placed on each side of the cube. A total of six temperature sensors were used to create a device that can simultaneously measure temperature changes on a cube shape. By optimizing the algorithm, detections of temperature and tactile position were successfully detected with recall value of ~90% for temperature (60~100°C) and ~93% for tactile position. Interpolation technique enables temperature and tactile detection not only on the surface where the sensor is placed, but also on the edges of the cube between the sensors.<br/>In summary, this study demonstrated temperature sensing system that combines solution-based process of V₂O₅nanowire and RC analysis. By optimizing fabrication process and data processing algorithm, temperature and tactile sensing were achieved with high stability, high sensitivity, low hysteresis, and high spatial resolution, which is potentially used for next class of low-cost and low power sensing system.