An adaptive convolutional neural network based on transmissibility grayscale image for online identification of offshore platform damage pattern

Abstract

In response to the challenges in offshore platform damage pattern recognition, where traditional methods lack adaptability owing to relying on feature extraction and expert knowledge, a novel method for online recognition of offshore platform damage patterns using an improved convolutional neural network (CNN) based on transmissibility grayscale image is proposed. The original transmissibility function is converted into a two-dimensional grayscale image to capture the essential features in the original signal, which is then used as the input for the deep learning model. Subsequently, an attention mechanism is introduced to enable the CNN to focus on key areas of the image, emphasize important feature channels, dynamically adjust feature representations, and enhance its overall performance. Meanwhile, the model is optimized by virtue of variational Bayesian algorithm and early stopping criteria, and finally, the trained offline model is applied to online damage pattern recognition of offshore platforms. Through on indoor vibration monitoring experiments, seven different damage patterns of the platform model are recognized accurately, which show that the accuracy of the improved CNN model is increased by an average of 8.5 % as compared to traditional intelligent methods. Furthermore, the proposed method is applied to the online damage identification of offshore platforms in the field, and the corresponding prediction accuracy reaches 98.9 %, demonstrating the feasibility and generalizability of the proposed method in adaptive feature extraction and precise identification of damage patterns.