Machine learning methods for damage detection of thermoplastic composite pipes under noise conditions

Machine learning methods for damage detection of thermoplastic composite pipes (TCPs) under noise conditions are presented, which combine the random decrement technique (RDT) with random forest (RF) and long and short-term memory (LSTM) networks. RDT is applied first to process the measured noisy strain response data of the TCP under random excitation. Then the RF or LSTM method is used to conduct the damage localization and severity estimation of the pipe. The applicability of the proposed methods is verified by means of numerical and experimental studies. The numerical example consists of a TCP subjected to internal pressure and random wave excitation considering several noise levels. The damages are simulated as circular holes on different layers of the pipe and varying severity, characterized by their radii and depths. The damage detection is carried out using RDT-RF and RDT-LSTM methods. The experimental studies consist of laboratory tests of a TCP model using Fiber Bragg Grating sensors. The damage cases, simulated as cracks with different lengths and depths on the reinforcement layer, are discussed. Both the numerical simulation and experimental tests show that the proposed RDT-RF and RDT-LSTM methods have an excellent performance in damage detection of TCPs.