Due to the characteristics of large detection range, high sensitivity and good environmental adaptability, optical fiber sensing networks have been widely applied in large-scale security and regional target positioning. Traditional functions such as identifying the presence of a target and roughly judging its position are no longer sufficient to meet the increasingly demanding detection requirements. Accurately classifying multiple targets and multiple state parameters in a large area has become a research hotspot. Among them, identifying the type of target, calculating the target position, and inferring the target's motion state are the main detection tasks of the sensing network.
SIDELNIKOV O et al. conducted tests on multiple abnormal interference signals within a region, achieving a detection rate of 86.3% by using different frequencies for target classification. However, this method can only perform qualitative classification and cannot provide information on the state parameters of the targets. TEJEDOR J et al. laid out an optical fiber sensing network on pipelines and identified potential engineering operations that could endanger the pipelines by analyzing the differences in vibration signals. They also classified different interferences using intensity thresholds. Tian Miao combined neural networks with the function mode decomposition method to analyze four types of intrusion events, achieving an average recognition rate of 85.2%. Zou Boxian et al. used three-dimensional visualization technology of vibration signals to classify different vibration sources such as white noise, pedestrians, vehicles, and excavators. The simulation analysis showed a correct rate of over 90%. However, the large amount of three-dimensional point cloud data significantly reduced the processing speed. Peng Kuan et al. tested regional intrusion sources based on time/frequency domain differences, achieving a classification accuracy of over 98% for four types of periodic interference sources. Jiang Hong et al. tested five common intrusion interferences using ultra-weak fiber Bragg gratings and classified them based on normalized signal features. In 500 test samples, the recognition rate was over 98%. Pan Ruizhi et al. used fiber Bragg grating tactile sensing technology to achieve target classification, with an algorithm accuracy of 96.6%. However, this method is mainly used for direct contact measurement between the target and the FBG. Although it has high accuracy, its response performance decreases significantly with increasing distance. Wei-hao C et al. used Φ-OTDR technology to obtain target signals, which has the characteristics of high precision and good stability. SUZHEN L et al. used artificial neural networks to measure construction vibrations in optical fiber sensing data, which has the characteristics of high precision and wide coverage. However, this method is mainly used for the recognition of single vibration signals and cannot achieve multi-target classification. Shang Qiufeng et al. combined variational mode decomposition with support vector machine algorithms to identify four types of abnormal signals, achieving an identification accuracy of over 98%. However, due to the use of two algorithms, the processing time for a single set of data was 169 seconds, which was relatively slow.
An identification algorithm based on the characteristics of multi-objective signal parameters was designed. This algorithm marks the features of different targets in terms of amplitude, duration and frequency, achieving signal decoupling in the case of multi-objective signal aliasing. The fiber sensing signal characteristics of four common targets were tested, and quantitative analysis of multi-objective signals was completed. The experimental results show that the mean wavelength amplitude of target 1 is 1.25nm, with a period characteristic of approximately 120ms; the mean wavelength amplitudes of targets 2 and 3 are between 150-350pm, with durations ranging from 1 to 3s; the mean wavelength amplitude of target 4 is over 3.2nm, with a duration of approximately 15s. These features have high recognition accuracy in this algorithm. In the target aliasing test, the average target recognition rate and the mean of the recognition accuracy are both above 80.0%, verifying the feasibility of the proposed algorithm.