基于Multi-Head Attention的BiLSTM改进模型光纤多事件识别

Multi-Head Attention-Based BiLSTM Improved Model for Fiber Optic Multi-Event Recognition

摘要: 针对相位敏感光时域反射计(φ-OTDR)分布式光纤传感系统对多事件(如挖掘、敲打、振动等)进行检测和识别的问题，提出一种基于双向长短期记忆网络(BiLSTM)和多头注意力机制(Multi-Head Attention)结合，并用Lookahead优化算法进行改进的深度学习网络模型。将获取的历史记录数据中的训练集进行处理后导入到改进后的预测模型进行训练，利用BiLSTM处理序列信号的能力，再结合多头注意力机制将输入向量分成多个子空间(即多个头)，每个头独立进行计算权重并加权求和，最后将所有头的输出拼接并线性变换得到最终结果。测试集数据进行模拟仿真，将BiLSTM-Multi-Head Attention改进模型与神经网络(CNN)、长短期记忆网络(LSTM)、双向长短期记忆网络及未改进的BiLSTM-Multi-Head Attention等模型进行对比，研究了对光纤多事件的分类识别能力，验证了该模型相对其他基准模型有更好的预测精度。

Abstract: In order to solve the problem of detecting and identifying multiple events (such as mining, tapping, vibration, etc.) in the phase-sensitive optical time domain reflectometer (φ-OTDR) distributed optical fiber sensing system, a deep learning network model based on the combination of Bidirectional Long Short-Term Memory Network (BiLSTM) and Multi-Head Attention mechanism and improved by Lookahead optimization algorithm is proposed. The training set in the obtained historical data is processed and imported into the improved prediction model for training, and the ability of BiLSTM to process sequence signals is used, and then the input vector is divided into multiple subspaces (i.e., multiple heads) by combining with the multi-head attention mechanism, and each head independently calculates the weights and weights the sum, and finally the output of all heads is spliced and linearly transformed to obtain the final result. The improved BiLSTM-Multi-Head Attention model was compared with the Convolutional neural network (CNN), Long Short-Term Memory network (LSTM), bidirectional long short-term memory network and the unimproved BiLSTM-Multi-Head Attention models. The model’s ability to classify and recognize the classification for multiple events in optical fibers is investigated to verify that the model achieves better prediction accuracy.