Advanced Fault Detection in Medical Devices

Comparative Study of Generative Models for Early Detection of Failures in Medical Devices

Fault diagnosis is a seasoned field of research, and many critical medical devices maintain an embedded expert system to self diagnose Increasing complexity of the devices and the variability in the operating environment results in failures which are often hard to predict and prevent by expert systems Majority of medical devices embed intricate electro-mechanical components with varying properties where failure is not linear The increasing complexity of medical devices, particularly in their electronics and software components, presents significant challenges in terms of safety, reliability, and efficacy.Given the increasing complexity of medical devices and the rising incidence of device-related failures and recalls, there is a pressing need to develop and implement advanced generative methods for more effective and reliable fault detection in these devices

Source: FDA MAUDE Database (Retrieved 9th January 2024)

The study investigated predictive and preventive maintenance strategies in medical devices, emphasizing fault detection techniques that utilize data, signal, process, or knowledge-based methods. These innovative techniques are critical for preempting failures that could lead to safety concerns or compromised device performance.

Research Scope

The core of our research involves a comprehensive survey of state-of-the-art Fault Detection and Prediction algorithms. We aim to assess their feasibility and potential application in the medical device sector, thereby contributing to the development of safer and more reliable medical equipment.

As a part of this study we analyzed two generative algorithms (GAN, VAE) and one classic (HMM) for their effectiveness in fault detection. We also evaluated the applicability against 1. Real world Surgical device sensor data 2.Airbus anomaly detection as benchmark dataset

Research Hypothesis

The potential of Generative Adversarial Networks (GANs) is being harnessed to create Data-Driven Digital Twins (DDT), where deep learning models build digital versions of physical assets from sensor data. These DDTs, when incorporated into Prognostic and Health Monitoring frameworks, learn from normal operation data, bypassing the need for historical failure records.

They excel in detecting early-stage faults, discerning between different types of failure modes, and monitoring the deterioration of assets across various operating conditions. Moreover, they can automatically produce health indicators, significantly enhancing the efficiency of predictive maintenance scheduling.

Team Members

  • Binesh Kumar
  • Bahareh Arghavani Nobar
  • Advisor: Dr. Vahid Behzadan

Publications

Submitted: Sadanandan, B., Arghavani Nobar, B., Behzadan, V. (2023). “Analysis of Fault Detection in Medical Devices Leveraging Generative Machine Learning Methods.”