Optical fiber sensors play an increasingly important role in safety monitoring areas in the smart Internet of Things (IoT). Particularly, a fiber-optic distributed acoustic sensor (fiber-optic DAS) based on the phase-sensitive optical time-domain reflectometry (Φ?OTDR) technology provides a highly dense, cost-effective, and continuous environment measurement way over a wide range. All kinds of vibration sources can be sensed and located with high sensitivity and precision utilizing the widely laid ordinary telecommunication cables, and thus fiber-optic DAS has been applied in various ground listening applications, such as natural disaster prediction of ocean-floor seismic activity, volcanic events, and earthquake, energy exploration in oil and gas industry, and civil infrastructure monitoring in the pipelines, railways, and perimeters. It leads to a new generation of large-scale fiber-optic IoT for ground and underwater listening technology. From the current research status in China and abroad, DAS is becoming mature in its hardware performance, such as the demodulation fidelity, sensing distance, detection bandwidth, and sensitivity, which are all approaching their perfection. However, with the rapid advance of DAS applications, the complicated and ever-changing environments for large-scale monitoring have brought about challenges of high false alarm rates due to its advantages of high sensitivity. It is difficult to achieve high-precision detection, recognition, and positioning of perceived vibration and acoustic targets, which has become the biggest technical bottleneck restricting the large-scale application of DAS technology. In recent years, driven by the development of advanced signal processing and artificial intelligence (AI) technology, the signal processing methods of fully intelligent DAS with high accuracy and real-time performance in practical complex environments have become a research hotspot and focus in the field of fiber-optic sensing. The signal processing method in DAS plays a crucial and decisive role in improving the intelligent perception ability of the entire system.

We review the current research status of signal processing methods in smart fiber-optic DAS entering the deep learning stage, from mainstream supervised learning to unsupervised, semi-supervised, and transfer learning, from single-source detection to multi-source aliasing detection, and from single-task recognition or localization to simultaneous implementation of recognition and localization tasks, and we predict possible research directions for further improving the intelligent processing performance and perception ability of DAS in the future. Firstly, the typical fiber-optic DAS system structure and its vibration/sound sensing mechanism (Fig. 2), and the smart DAS and its signal processing architecture in smart city monitoring applications (Fig. 3) are introduced. Then, the signal processing methods based on deep learning are explained in detail, which includes the main stream of supervised learning methods based on multi-dimensional information extraction, and semi-supervised, unsupervised learning, and cross-scene transfer learning methods in DAS. For the supervised learning method, it includes DAS signal recognition models based on temporal information extraction, such as one-dimensional convolutional neural networks (1D-CNNs) (Fig. 4), multi-scale convolutional neural networks (MS-CNNs) (Fig. 5), multi-scale and contextual temporal relationship mining methods (Figs. 6-7), and the two-dimensional recognition models based on time-frequency (Figs. 8-11), time-space (Figs. 12-14), and space-frequency (Fig. 15) information extraction technologies. Besides, some other supervised methods are also included, for example, recognition models based on attention-based long short-term memory (Fig. 16) and the fusion of manual features and deep features. It proves that the combination of traditional empirical rules and deep learning networks can further reduce the false alarm rate of the system. In response to the problem of insufficient labeled samples in new scenarios in practical applications, several semi-supervised recognition methods based on the 1D-SSGAN (one-dimensional semi-supervised generative adversarial network), SSAE (sparse stacked autoencoder), and FixMatch models have been involved to achieve accurate recognition of DAS signals with a small amount of labeled data and a large amount of unlabeled data. Furthermore, the SNN-based DAS unsupervised learning network (Fig. 17) and the cross-scene transfer learning network based on AlexNet+SVM (Fig. 18) also appear to improve the generalization ability of DAS signal recognition methods. In order to evaluate the performance of these recognition models, we introduce seven indicators for evaluating the recognition accuracy and four indicators for the processing time of the algorithms. The above key DAS recognition methods and their performance are statistically compared in Table 2. At last, the new challenges of smart DAS sensing, from single-source detection to multi-source aliasing detection, from target recognition to localization, and from a single task to multi-task processing, as well as other methods to enhance its intelligent perception capabilities, have also been introduced.

Further improvement of signal processing and its sensing capabilities still faces new challenges and opportunities and will open a new chapter in fully intelligent DAS. Stable, accurate, real-time, and efficient signal recognition in DAS in new complicated application scenarios remains a research hotspot in the field of distributed fiber-optic sensing in the future, including: 1) improving the generalization ability of DAS recognition models in cross scenarios; 2) significant improvement in real-time processing capabilities in DAS; 3) improvement of multi-task processing ability in DAS; 4) implementation of high-performance on-chip DAS.