Research On Acoustic Emission Signal Recognition Algorithm Based On Improved Neural Networks

Acoustic Emission (AE) is a Non-Destructive Testing (NDT) techniques which records the changes in the internal vibrations of material through the transient elastic waves. AE is suitable for a wide range of researches like material deformation, pipeline leak detection and pressure vessel inspection with real-time, applicability, extensive features. AE signal recognition is the impotant part of AE detection, and only the class of malfunction is precisely dectected can the malfunction be investigated effectively, and the materal loss and safety hazards caused by monitoring error be successfully avoided.At present, AE signal recognition based on Artificial Neural Networks (ANN) has the disadvantages of low information content, easy-to-fall into local optimum, etc. Based on the prior techniques, the dissertation carries out the following aspects of research work:Firstly, AE detection technology has been studied in detail, and AE signal processing methods has been discussed. The dissertation also compares the distinctions between different degrees of collision-friction AE signals through experiments.Secondly, multi-feature extraction of AE signal has been studied. The Hurst index and the Approximate Entropy have been proposed to be added into AE features, in order to analyze and prove its effectiveness from the view of statistical correlation and uncertainty.Thirdly, the algorithm structure of ANN has been studied. The GMM/ANN network has been proposed to be applied to AE recognition, which trains Gaussian Mixture Model and ANN alternatively with their advantages to optimize the network performance.Fourthly, the Chaotic Neutual Networks (CNN) applied in AE recognition has been studied. For the uncertainty and the non-linear property of the AE system, it improves the recognition performance by chaotic Logistic mapping unit.Fifthly, the Deep Belief Networks (DBN) applied in AE recognition has been studied. The dissertation constructs DBN based on Restricted Boltzmann Machines (RBM) model, designing and correcting model parameters to ease the easy-to-fall into local optimum effect. The superiority of DBN by comparison with BP in AE recognition has been proved.