Health monitoring is a process of collecting data in real-time from various assets using sensors, and other IoT devices that provide valuable insights to organizations for making informed decisions. In this process, deep learning plays a vital role by enabling organizations to learn from complex data for making such decisions, especially in the context of fault diagnosis, and predictive maintenance of the assets.
While deep learning methods show tremendous predictive performance, they are heavily dependent on the availability of a diverse class-balanced dataset for their effective training. However, in real-world scenarios, it is difficult to collect such diverse datasets as industrial assets are safety critical. These assets mostly operate in normal or healthy conditions, and fault events are rare, making it costly to obtain fault data. Therefore, there is a significant gap in the number of different category samples which deteriorates the fault diagnosis performance of the deep learning models leading to misclassification of an asset’s health state.
To solve this problem, your task will be to explore data augmentation techniques using deep generative models for generating artificial fault samples to train deep learning architectures for reliable fault diagnosis of industrial assets under missing sample or limited data scenarios. Recently, diffusion-based generative models have emerged as a powerful tool due to their capability to generate high diversity samples. The goal of diffusion models is to learn a diffusion process that generates the probability distribution of a given dataset, which may be utilized for generating artificial asset fault samples. Therefore, in this project you will
Undergraduate
None
Mechatronics Teaching (405.822, Lab)