MEDEP: Maintenance Event Detection for Multivariate Time Series Based on the PELT Approach

Area of Science:

• Data Science
• Industrial Engineering
• Machine Learning

Background:

• Predictive Maintenance (PdM) is crucial for Industry 4.0 manufacturing.
• Collecting labeled maintenance data is challenging due to human involvement and time constraints.
• Existing condition monitoring datasets often lack sufficient labeled maintenance event data.

Purpose of the Study:

• To introduce MEDEP, a novel framework for automatic maintenance event detection.
• To address the scarcity of labeled maintenance data for PdM models.
• To offer new research possibilities for existing condition monitoring datasets.

Main Methods:

• Developed MEDEP, a maintenance event detection framework using the Pruned Exact Linear Time (PELT) approach.
• Extended PELT with a heuristic method including mean thresholding for multivariate time series and distribution threshold analysis.
• Validated MEDEP on the Microsoft Azure Predictive Maintenance dataset and welding industry data.

Main Results:

• MEDEP demonstrated superior performance in detecting maintenance events.
• Achieved an average false-positive (FP) rate of approximately 10%.
• Exhibited high sensitivity and accuracy in experimental outcomes.

Conclusions:

• MEDEP effectively detects maintenance events from condition monitoring data deviations.
• The framework offers a promising solution for data scarcity in PdM.
• MEDEP enhances research possibilities by providing automatic labeling for datasets.