Naval warships and marine vessels rely on integrated equipment and systems that operate under extreme critical conditions, which make them highly susceptible to failures that lead to major accidents resulting in loss of life and assets. To prevent major accidents during military or maritime exercises, it is imperative to have an efficient fault prognosis that can anticipate potential failures. This study proposes a systematic approach to develop a diagnostic model via thermal imaging for real-time machinery health monitoring. Considering the challenges associated with thermal image classification in state-of-the-art models due to the unique nature of thermal images, this study proposes ThermoNet, an AI (Artificial Intelligence) diagnostic model specifically tailored for thermal images to enhance feature extraction and classification accuracy. ThermoNet comprises CNN (Convolutional Neural Networks) integrated with dilated convolutions, CBAM (Convolutional Block Attention Module), and global average pooling. The non-availability of a thermal image dataset required for developing an effective diagnostic model for real-time machinery health assessment has further led to the development of a physics-informed computational thermal image generation model that represents various machinery health conditions (normal, defective, and degraded). These thermal images were validated against real thermal images. Experimental results indicate that ThermoNet outperforms other state-of-the-art models, with an accuracy of 0.987, a precision of 0.986, and a recall of 0.987. Furthermore, our model operates with an extremely low inference latency of 0.0017 s, significantly faster than other state-of-the-art models. The unique contribution of this study entails the development of a physics based model that can generate a condition series thermal image dataset that represents the different machinery health conditions, development of the ThermoNet model, and a comparative analysis for real-time machinery health prediction for prognosis and diagnosis.
Khan et al. (Tue,) studied this question.