A Study on Thermal Detection Based on Support Vector Machine Using Dynamic Time Warping and Application to Optical Fiber Sensor

This study investigates a thermal detection method using the classification of a support-vector machine and the application to an optical fiber sensor. Since the conventional simple threshold determination requires to adjust the threshold level according to the environments but does not consider features of the temperature-time relation, it sometimes causes false positives. False positives of an abnormal temperature rise can be avoided when the threshold level is set enough high, however it still remains a problem not to be optimal for early detection. To solve such problems, the proposed method in this work takes into accounts features of the temperature-time relation. Its features exclude unnecessary fluctuations by calculating the differences, emphasize the effect of temperature rise by using a rectified linear unit function, express the similarity by the dynamic time warping distance, and are used for the support-vector machine classification of machine learning. Proof-of-principle simulations are conducted and experiments using an optical fiber sensor are also successfully performed. The abnormal temperature rise is early detected as the abnormal state before reaching to the maximum temperature. Nevertheless, the rise-and-fall temperature profile is effectively detected as the normal state. These results show that the proposed method considering temporal factors improves the classification quality.