Session: 06-17-01 AI Technology for Ocean Engineering - I

Submission Number: 180868

Cnn-Based Fault Detection Using Vibration Image Transformations With Applications to Offshore Machinery 

An effective fault detection system in offshore machinery is crucial for improving safety, minimising downtime, and reducing costs. A reliable classifier should be able to identify faults and provide timely alerts to machine operators. Recent studies show that deep learning methods, particularly neural networks, reveal a high potential for condition monitoring in offshore machinery. However, the use of vibration signal image transformations as a convolutional neural network (CNN) input for fault detection has not been widely investigated. This paper presents the performance benchmark of CNN models applied to image representations of vibration data. Data used in this paper comes from the open-source NASA repository. This study involves preprocessing vibration signals and converting them into four different image types: grayscale, RGB, continuous wavelet transform scalograms (CWT), and short-time Fourier transform spectrograms (STFT). It also includes training and evaluating the performance of CNN-based classifiers to detect bearing faults in offshore machinery. The results show that the best performing classifiers detect faults with an accuracy of more than 98%. CNNs using CWT scalograms and STFT spectrograms as input achieved the highest accuracy, whereas models trained on grayscale and RGB images showed lower performance. This work provides a performance comparison of CNN models based on different inputs for offshore machinery fault detection. In addition, the study demonstrates data preprocessing methods, a convolutional neural network architecture, and evaluation approaches.

Presenting Author: Pawel Klis University of Stavanger

Presenting Author Biography: Industrial PhD candidate at the University of Stavanger.

Authors:

Pawel Klis University of Stavanger
Dan Sui University of Stavanger
Yihan Xing University of Stavanger