Phase-sensitive optical time domain reflectometer (Φ-OTDR) based distributed optical fiber sensing system has been widely used in many fields such as long range pipeline pre-warning, perimeter security and structure health monitoring. However, the lack of event recognition ability is always being the bottleneck of Φ-OTDR in field application. An event recognition method based on deep learning is proposed in this paper. This method directly uses the temporal-spatial data matrix from Φ-OTDR as the input of a convolutional neural network (CNN). Only a simple bandpass filtering and a gray scale transformation are needed as the pre-processing, which achieves real-time. Besides, an optimized network structure with small size, high training speed and high classification accuracy is built. Experiment results based on 5644 events samples show that this network can achieve 96.67% classification accuracy in recognition of 5 kinds of events and the retraining time is only 7 min for a new sensing setup.

Traditional event recognition methods for signal collected by Φ-OTDR sensing system is difficult to identify the event category accurately in field application. Deep-learning-based event recognition method can achieve high classification accuracy but needs massive scale computation and long-term training. An event recognition method based on transfer training which can build a high-precision event recognition network quickly is proposed in this paper. The raw data collected by Φ-OTDR only needs simple bandpass filtering and scaling according to the size of the input layer of the pre-trained network. The initial network is created by freezing the front structure of the pre-trained network and only the rest layers are trained. The experiment result based on 4254 samples from a 8 kinds event data set showed that through freezing one-fifth of the pre-trained AlexNet, which is trained on the ImageNet data set, and retraining the rest parts by Nvidia GTX1050Ti, which contains only 768 CUDA cores, for less than 5 minutes can achieve the best classification accuracy, which is about 96.16%. When the training data set reduces to only 1146 samples, the method can still achieve 95.56% classification accuracy. It provides a way to quickly build a high-accuracy network for a new filed application.

Thanks to the development of machine learning and deep learning, data-driven pattern recognition based on neural network is a trend for Φ-OTDR system intrusion event recognition. The data-driven pattern recognition needs a large number of samples for training. However, in some scenarios, intrusion signals are difficult to collect, resulting in the lack of training samples. At the same time, labeling a large number of samples is also a very time-consuming work. This paper presents a few-shot learning classification method based on time series transfer and cycle generative adversarial network (CycleGAN) data augmentation for Φ-OTDR system. By expanding the rare samples based on time series transfer and CycleGAN, the number of samples in the dataset can finally meet the requirement of network training. The experimental result shows that even when the training set has two minor classes with only two samples, the average accuracy of the validation set with 5 classification tasks can still reach 90.84%, and the classification accuracy of minor classes can reach 79.28% with the proposed method.

A phase-sensitive optical time domain reflectometer (Φ-OTDR) can be used for pipeline security. However, the sensing distance (less than 20 km) of traditional Φ-OTDR is too short for the needs of typical oil and gas pipeline monitoring applications (30–50 km). A simple structure Φ-OTDR system utilizing long pulse, balanced amplified detector and heterodyne detection is proposed in this paper and the sensing range is thereby increased to 60 km. Through analyzing the sensing principle of Φ-OTDR, a novel locating strategy is proposed to maintain the locating accuracy at a few meters when a long pulse (5 µs) is used. The increased pulse width deteriorates the time series of each sensing point seriously. In order to eliminate the deterioration, a data processing technique combining wavelet and empirical mode decomposition is applied in this system. The experiment results show that the sensing distance can be increased to 60 km and the locating accuracy is maintained at 6.8 m.

In the processing of recognizing Φ-OTDR event signals by deep learning models, a large amount of annotated data is required for model training. However, the data annotation is a major challenge for a lot of unlabeled data and the labeling cost are high. This article proposes a valuable signal data selection method for Φ-OTDR sensing system. The data samples that are beneficial for training from the unlabeled data set is selected based on their uncertainty values. Through only labeling these valuable data samples and training the classification model with them, the labeling cost can be saved a lot and the classification model can keep its classification accuracy. The experiments show that this method can save 83.77% on manual labeling cost and the classification accuracy can still reach 96.37%, which is similar to the classification accuracy of the model trained by labeling all the data.