On Bayesian Optimization-Based CNN-BiLSTM Network for Multiclass Classification in Distributed Optical Fiber Vibration Sensing Systems

Due to the influence of the false alarm rate, it remains a challenging task to accurately detect potential intrusion behaviors in distributed optical fiber vibration sensing systems. Therefore, accurate multiclass pattern classification is significantly crucial for enhancing the stability and reliability of the system. To this end, we propose a Bayesian optimization (BO)-based hybrid deep learning network in this article that combines a convolutional neural network (CNN) with a bidirectional long short-term memory (BiLSTM) for multiclass pattern recognition and classification. There are mainly two contributions in this work. First, conventional CNN models often fail to capture the temporal dependencies between the acquired signals, which may lead to a decrease in classification accuracy. Moreover, it is still a challenge to directly apply CNN models on 1-D time series as they can only look back on history with a linear size, which may limit the depth of the network. To address these problems, two BiLSTM networks are embedded into the CNN model. Second, as the complexity of the designed network increases, the number of possible combinations of hyperparameters will grow rapidly, thus requiring considerable time for the selection of optimal hyperparameters. To reduce the time cost, we adopt the BO algorithm to automatically calculate the optimal hyperparameters of the designed hybrid model. To evaluate the effectiveness of the proposed scheme, nine different categories of sensing data were selected as data samples and collected from a dual Mach-Zehnder interferometer (DMZI)-based optical fiber perimeter security system. The results show that the mean recognition accuracy of the proposed scheme is 98.76%. Apparently, this research can further guide the development of optical fiber vibration sensing systems and their algorithms in practical applications.