Predictive Maintenance: An Autoencoder Anomaly-Based Approach for a 3 DoF Delta Robot

Area of Science:

• Robotics and Automation
• Machine Learning for Predictive Maintenance

Background:

• Predictive maintenance (PdM) is hindered by large datasets lacking run-to-failure (R2F) data, making supervised learning approaches for health index (HI) and remaining useful lifetime (RUL) estimation infeasible.
• Traditional PdM methods struggle when R2F data is unavailable, limiting the ability to identify condition indicators (CIs) and build degradation models.

Purpose of the Study:

• To develop a novel PdM method for a 3 Degrees of Freedom (DoF) delta robot operating in a pick-and-place task.
• To address the challenge of PdM without R2F data by employing autoencoders (AEs) for anomaly detection and RUL prediction.

Main Methods:

• Utilized autoencoders (AEs) for anomaly detection based on signal sequence distribution, crucial for scenarios with no R2F data.
• Employed convolutional layers for effective feature extraction from sequential, nonlinear, and correlated time-series data.
• Integrated a sigmoid function for anomaly probability prediction using CIs from AEs, optimizable via a minimax problem.
• Applied Gaussian process (GP) as a degradation model to calculate RUL using estimated health index (HI) values.

Main Results:

• Successfully demonstrated the capability of AEs to detect anomalies and predict maintenance needs in the absence of R2F data.
• The proposed architecture enabled fault localization within the delta robot system.
• Achieved accurate RUL prediction using Gaussian process regression on health index values derived from AEs.

Conclusions:

• The developed autoencoder-based approach provides a viable solution for predictive maintenance in systems lacking run-to-failure data.
• The method effectively handles complex data characteristics, enabling anomaly detection, fault localization, and RUL estimation for robotic systems.
• This research offers a significant advancement in applying machine learning for robust PdM strategies in industrial applications.