Overview

In our comprehensive research project, we aimed to address the challenges in fault detection for critical devices, specifically focusing on surgical staplers’ dataset and utilizing the Airbus dataset as a benchmark. The medical device industry has witnessed significant advancements by incorporating electronics to enhance the safety and performance of life-saving medical devices. However, this integration of complex electro-mechanical systems introduced new failure modes that traditional testing protocols struggle to identify and mitigate.

To tackle this issue, we leveraged generative machine learning models, namely the Hidden Markov Model (HMM), Variation Autoencoder (VAE), and Generative Adversarial Network (GAN), as innovative fault detection methods. Our study applied rigorous design principles and experimental methodologies to evaluate the effectiveness of these models.

One key aspect of our analysis involved conducting a Window Size Analysis, employing downsampling techniques to systematically explore the influence of different window sizes on model performance. This emphasis on window size aimed to understand the trade-offs associated with varying temporal contexts in anomaly detection.

Furthermore, our research delved into model evaluation, highlighting the advantages of utilizing larger datasets in terms of accuracy metrics. The evaluation process involved comparing the performance of the generative models on both the Airbus Helicopter Accelerometer dataset and the Surgical Stapler dataset. Our focus on accuracy metrics allowed us to identify the model with the highest accuracy rate in both datasets.

The context of our research lies in the evolving landscape of technology, where connectivity and data collection, coupled with advanced sensors, facilitate preventive and predictive maintenance strategies. These strategies rely on fault detection techniques, including data, signal, process, or knowledge-based methods, to prevent potential failures that could compromise device safety or performance.

Through our independent study, we aimed to contribute to the state of the art in Fault Detection and Prediction algorithms. Our feasibility study sought to understand the applicability of these advanced techniques to the unique challenges presented by medical devices, particularly focusing on surgical staplers. This research not only provides valuable insights into fault detection methods but also lays the groundwork for the integration of predictive and preventive maintenance strategies in the medical device industry.

Team

Binesh Kumar

Bahareh Arghavani Nobar

Advisor: Dr. Vahid Behzadan

GitHub: N/A

Publications: Sadanandan, B., Arghavani Nobar, B., Behzadan, V. (2023). “Comparative Study of Generative models for early detection of
failures in medical devices.” Accepted at the ICMHI 2024

Sponsor: Medtronic Plc