Distributed optic-fiber vibration sensing (DVS) system enables distributed measurement of vibration signals. Despite the prevalent utilization of pattern recognition algorithms in practical applications to discern various vibration events, extant pattern recognition features often exhibit fixity and redundancy, thereby inadequately capturing the nuanced characteristics of vibration signals, resulting in a pronounced false alarm rate challenge. This study presents a prototype of a direct detection structure DVS system devised to confront the issues above. Also, a hybrid feature selection methodology is proposed here, integrating the simulated annealing algorithm with the Fisher Score algorithm. Initially, the Fisher Score algorithm is employed to identify an apt initial feature set, subsequently integrating the Fisher Score into the new solution generation phase of the simulated annealing algorithm to optimize the overall efficacy of the feature amalgamation for intrusion vibration signals. Experimental validation of the algorithm underscores its capability to obviate redundant intrusion signal features, evince rapid convergence, and elevate the system's recognition rate from 80.23% to 94.46%.

Objective The distributed fiber-optic vibration event recognition system comprises two key components of the distributed fiber-optic vibration sensing (DVS) system and the vibration event recognition module. This system is highly effective for locating vibration events and has garnered significant attention in applications such as perimeter security, pipeline leakage detection, and earthquake monitoring. In the studies on pattern recognition of intrusion vibration signals collected by DVS, it is essential to handle large volumes of high-dimensional feature vector data, where each component represents a specific characteristic of the data. However, existing pattern recognition features are often fixed and redundant, with some algorithms failing to address the core problem effectively, leading to a high false alarm rate. To enhance pattern recognition accuracy, it is crucial to eliminate irrelevant and redundant features from these high-dimensional vectors and identify the features critical to solving the problem. This article proposes a feature selection algorithm designed to improve the recognition rate of DVS vibration signals.

Methods This study constructs a prototype of a distributed fiber-optic vibration sensing system (Fig.1) and introduces a hybrid feature selection method combining the simulated annealing algorithm with the Fisher Score algorithm (Fig.2). Initially, the Fisher Score algorithm is used to select an appropriate initial feature set. Subsequently, the Fisher Score is embedded into the new solution generation stage of the simulated annealing algorithm to optimize the feature combination for intrusion vibration signals.

Results and Discussions The time-domain characteristics of four types of vibration signals—shear, tapping, shaking, and climbing—were analyzed (Fig.4-5, Tab.1). The algorithm optimization results for two sets of data each for cutting and tapping, as well as shaking and climbing, are presented (Tab.2). The relationship between the fitness values and iteration times of feature subsets selected by different methods on two datasets (Fig.6-7). A comparative analysis of the fitness values for different numbers of features revealed that the highest fitness value,0.9466, was achieved when selecting 10 features (Fig.8). The proposed method demonstrates significant advantages in feature selection, efficiently removing redundant features from the dataset.

Conclusions To address the high false alarm rate in DVS systems used for perimeter security, pipeline leakage detection, and long-distance monitoring, this paper introduces a hybrid feature selection algorithm combining the simulated annealing and Fisher Score algorithms. Experiments on various vibration signals collected by DVS prototypes validate the algorithm. The maximum fitness values of different feature numbers were compared with traditional methods, demonstrating the proposed algorithm's superior performance and faster convergence in removing redundant features. This algorithm significantly improves the recognition rate of DVS systems, offering a valuable solution for reducing the false alarm rate of DVS vibration signals.