A Lightweight Spatial–Temporal Deep Learning Network for Multiple Events Classification in Optical Fiber Perimeter Security System

Leveraging deep learning (DL) techniques, Artificial Intelligence (AI) can revolutionize optical fiber perimeter security systems, one of the cornerstones of next-generation Internet of Things (IoT) infrastructure, through automated classification of numerous events. However, enhancing recognition and detection performance with existing DL methodologies often necessitates substantial computational resources, posing challenges for hardware-constrained platforms. Furthermore, as DL-based neural network complexity increases, the number of potential hyperparameter combinations grows exponentially, rendering manual tuning arduous and time-consuming. To address these challenges, this study introduces a novel lightweight DL classification model optimized using a Crested Porcupine Optimizer (CPO) algorithm. Specifically, built upon an improved ShuffleNet architecture and a Gated Recurrent Unit module, this model effectively extracts spatial and temporal features while minimizing computational complexity. The CPO-based algorithm automatically determines the optimal hyperparameter combination for the neural network, significantly enhancing tuning efficiency. To validate the model’s efficacy, nine unique sensing patterns were collected from an optical fiber perimeter security system as original data samples. Experimental results demonstrate that these patterns can be accurately classified with a mean accuracy of 98.10% and a mean processing time of only 0.192 s. Moreover, Compared to other intelligent detection methods in optical fiber perimeter security field, the proposed approach exhibits superior performance.