基于LSTM神经网络的管道缺陷模式识别方法研究

针对复杂环境下,管道振动信号特征微弱难以提取的问题,提出一种基于长短时记忆网络(LSTM)深度学习神经网络的管道缺陷模式识别方法;首先利用改进型自适应噪声的完全集合经验模态分解(ICEEMDAN)对采集的原始信号进行分解得到若干个固有模态函数(IMF)分量,随后根据信息熵理论计算IMF分量的近似熵作为管道典型状态的特征值构造特征向量集合,然后构造LSTM深度学习神经网络训练模型并调节深度神经网络在训练过程中的相关参数进行网络的结构优化,最后将特征向量输入到LSTM神经网络模型进行训练和识别;结果表明:针对管道振动信号特征微弱难以提取的问题,该方法对管道缺陷模式识别的准确率达到了95％,在消除管道振动信号的背景噪声、挖掘特征信息和保证识别准确性方面优势明显.

Research on the Recognition Method for Pipeline Defect Pattern Based on LSTM Neural Network

Aiming at the difficulty in feature extraction from pipeline vibration signal in complex environments, a pipeline defect pattern recognition method based on Long Short-Term Memory network (LSTM) deep learning neural network is proposed here. Firstly, the collected original signal is decomposed for several intrinsic modal function (IMF) components with the Improved Complete Ensemble Empirical Mode Decomposition with adaptive noise (ICEEMDAN). Then the approximate entropy of the IMF component is calculated according to the information entropy theory as the eigenvalues of the pipeline running state to construct the feature vector set. And then the typical LSTM deep learning neural network training model is constructed and the relevant parameters of the deep neural network amid the training process are adjusted to optimize the network structure. Finally, the feature vector is input to the LSTM neural network model for training and recognition. The research results show that: for the problem that the pipeline vibration signal features are weak and difficult to extract, the accuracy of the method for pipeline defect pattern recognition has reached 95%, and it has obvious advantages in eliminating the background noise of the pipeline vibration signal, mining feature information, and ensuring recognition accuracy.