基于改进MobileNetV3-BiLSTM的Φ-OTDR管道微泄漏识别研究

Micro-leak identification in gas pipelines using Φ-OTDR and an improved MobileNetV3-BiLSTM network

摘要:

针对埋地燃气管道微泄漏不易被监测的问题，通过分布式光纤振动传感系统采集泄漏振动信号，并采用连续小波变化转换为二维图像。提出了改进型 MobileNetV3-BiLSTM（Improved MobileNetV3-BiLSTM，以下简称 IMV3-BiLSTM）网络，对模拟工况下的实时泄漏信号进行了识别。网络对泄漏时间序列具有可记忆性，同时具备轻量化识别特点。泄漏实验结果表明，该网络可实现四种泄漏工况下95.96 %的总体分类准确率；在0.1 MPa泄漏压力下，实现了1/16英寸孔径下泄漏的可靠识别。相较于传统卷积神经网络识别准确率提升约10 %-12 %，平均训练速度显著提升，有望满足燃气管道低压、小孔径泄漏工况下的监测需求。

Abstract:

To address the difficulty of monitoring small leaks in buried gas pipelines, leakage vibration signals are collected by a distributed fiber optic vibration sensing system. These signals are converted into two-dimensional images using continuous wavelet transform. An improved MobileNetV3-Bidirectional Long Short-Term Memory (BiLSTM) network is proposed to identify real-time leakage signals under simulated working conditions. The network possesses memory capabilities for leakage time series and features lightweight recognition. Experimental results show that the network achieves an overall classification accuracy of 95.96% under four leakage conditions. Under a leakage pressure of 0.1 MPa, it realizes reliable identification of micro-leakage with a 1/16 inch aperture. Compared with traditional Convolutional Neural Networks (CNNs), the recognition accuracy improves by about 10%–12%, and the average training speed is significantly increased. The proposed method is expected to meet the monitoring requirements for gas pipeline leakage under low pressure and small aperture conditions.