Quickly build a high-precision classifier for Φ-OTDR sensing system based on transfer learning and support vector machine

Abstract

The event recognition method based on deep learning for Φ-OTDR sensing system can guarantee robustness by automatically mining features, but relies on massive scale computation and a long time for data acquisition and model training. A method based on transfer learning and support vector machine is proposed in this paper, which can quickly build a high-precision classifier with a few training samples and a common device without GPU. In the pre-processing, the raw data just needs simple bandpass filtering and scaling operation. Based on transfer learning, the features of class identification are extracted by pre-trained AlexNet and then be directly used to construct SVM classifier without any feature selection operation. Besides, the rigorous explanation through visualization approaches validates the ability of automatic feature extraction to ensure the reliability of this simple method. The experiment based on 4254 samples with 8 event types shows that based on the extracted features and 13 s training by a portable computer with Inter i5-7300HQ without GPU, the support vector machine classifier can achieve 94.67% validation accuracy. Even if the scale of training data declines to 1146, 90.82% classification accuracy can still be kept. It is an easily generalized method to quickly build a high-precision classifier in a new field application.