Improved Generalization in Signal Identification With Unsupervised Spiking Neuron Networks for Fiber-Optic Distributed Acoustic Sensor

Abstract

More and more machine learning techniques are involved in fiber-optic distributed acoustic sensor (DAS) to improve its intelligent recognition ability. However, the supervised methods need sufficient and comprehensive labeled data, or they may present poor generalization when the test and the training data are inconsistent. Thus, in this paper a novel unsupervised Spiking Neural Network (SNN) is introduced into DAS for the first time to improve its generalization in different scenes and time-varying backgrounds. The theoretical mechanism, construction and optimization of the new network are first discussed in details; then its effectiveness is validated through three groups of field tests, using small dataset A with balanced, typical and consistent data, bad datasets B and C full of inconstant and atypical data, and full imbalanced dataset D respectively. It shows that the proposed SNN using the unsupervised learning mechanism, spiking-timing dependent plasiticity (STDP), is more stable than the mainstream supervised CNN when trained with small number of samples or imbalanced dataset, what often happens in practice. Especially in the most challenging case full of atypical or inconsistent samples, its generalization and anti-overfitting abilities are proved to be much better, where the average accuracy in the random test using bad datasets B and C has been increased by 18.0% and 14.4%. In the field test with full dataset, its recognition perform is also stably better than the compared CNN but with a longer time.