Signal Processing and Fault Diagnosis on Structure Vibration Measurement Using a New Composed Deep Learning Model

Abstract

This paper proposes a new approach for identifying the structure faults in offshore jacket platforms. The proposed method is based on the integration of the structural vibration data, a sophisticated data fusion process, and an intelligent diagnosis algorithm. In this new approach, firstly, the temporal convolutional network (TCN) overcomes the efficiency bottleneck of traditional recurrent neural networks in long time series signals, and significantly enhances the ability to capture long-distance dependent features in the structural response signals through the introduction of a dilated convolutional structure. Secondly, the bidirectional gated recurrent unit (BiGRU) network fuses forward and reverse gated recurrent units, which can effectively capture the forward and backward correlated timing features in the structure vibration data. The Attention mechanism then further weights and optimises the BiGRU timing outputs so that the model can automatically focus on the signal pattern that is most discriminative for the fault identification. Furthermore, the global optimisation of the hyperparameters of the TCN and BiGRU with the Artificial Lemming Algorithm (ALA) significantly improves the convergence speed of the entire model and avoid falling into the local optimum, thereby enhancing the prediction accuracy and generalisation performance of the combined model. The validity of the proposed ALA-TCN- BiGRU model is substantiated through a combination of the simulation and experimental validation, thereby substantiating the feasibility of the proposed methodology. The proposed model can achieve an overall detection accuracy of over 98% for the jacket structure, which is superior to several popular diagnosis methods.