Objective Distributed fiber optic acoustic sensing (DAS), as a novel vibration sensing technology, leverages single-mode communication fibers to create large-scale, cost-effective sensing arrays. Although this sensing technology is more prone to noise interference compared to traditional strain sensors, it offers a broader response bandwidth and greater durability for long-term deployments. Consequently, it has found widespread applications in various fields, including seismic wave detection, pipeline condition monitoring, perimeter vibration sensing, and more. The DAS system produces a significant amount of data during operation; However, only a fraction of this data contains relevant information. The prevailing approach involves employing machine learning for classification or pattern recognition tasks, thereby maximizing the utilization of the data's value. Since many fiber vibration event recognition methods rely on extracting features from time-frequency domain graphs to accomplish the classification task, and such methods tend to complicate the algorithm due to the incorporation of convolutional neural networks (CNN), this study aims to simplify the process by combining basic data preprocessing with clustering algorithms to categorize vibration signals.

Methods A new method based on time-domain amplitude feature extraction using clustering algorithm for intrusion event recognition is proposed (Fig.2). This method can be used in phase-sensitive optical time-domain reflectometer (Φ-OTDR) to classify the detected vibration signals. Compared with traditional image machine learning algorithms, the signal recognition method proposed in this study requires fewer samples and does not need tedious manual labeling. In this method, firstly, the difference of neighboring data points is calculated to get the maximum value of the difference sequence, and the maximum and envelope values are used to extract key features (Fig.3). Then, the vibration events are classified using hierarchical clustering algorithms. Finally, the effectiveness of the method is verified by evaluation indexes such as V-measure and silhouette coefficient.

Results and Discussions The vibration events simulated in the experiment include wind noise, manual knocking and digging. The experimental results show that the clustering accuracy of the three events can be up to 88.68% (Fig.13), and the indexes of homogeneity, completeness and V-measure are higher than 0.7, indicating that the clusters are well defined and the data points in each cluster are similar to each other; The silhouette coefficient is 0.778, and the clusters are well separated, and the number of data points in each cluster is 596, 504, 543 (Fig.14), and the size of each cluster in the clusters is reasonable. The clustering accuracy, homogeneity, completeness, V-measure, adjusted Rand index and adjusted mutual information are all relatively high, indicating that the clustering results match well with the real labels, and the clusters of clusters have a high degree of similarity of the data points within the clusters, and each cluster contains most of the real data points belonging to the class. The effectiveness of the unsupervised learning-based fiber vibration time-domain feature signal recognition method is verified by recognizing three types of events.

Conclusions This study proposes a feature extraction method based on time-domain amplitude differences to solve the difficulty of poor clustering resolution caused by strong correlation among signal features after conventional time-domain feature extraction of raw data in the Φ-OTDR system. The recognition method based on the proposed feature extraction approach and unsupervised clustering analysis has no requirements with pre-labeling of events as well as large datasets. It has fast calculation speed and effectively addresses the problems of large scale data and much workload for marking data in image recognition machine learning. According to experimental results, the clustering accuracy of the proposed method reaches 88.68%. So it has the function of identifying vibration events caused by intrusion targets. This study provides a novel solution for recognizing the vibration events of Φ-OTDR by machine learning.