The "DOLPHINS" Project: A Low-Cost Real-Time Multivariate Process Control From Large Sensor Arrays Providing Sparse...

Area of Science:

• Industrial Engineering
• Data Science
• Machine Learning

Background:

• Predictive maintenance is crucial for minimizing industrial downtime.
• Leveraging existing Programmable Logic Controller (PLC) data offers a cost-effective approach.
• The DOLPHINS project aimed to implement real-time predictive maintenance in an industrial pilot plant.

Purpose of the Study:

• To develop and validate a real-time predictive maintenance system.
• To utilize Machine Learning (ML) techniques on sparse binary data from existing sensors.
• To avoid additional investments in new sensors and infrastructure.

Main Methods:

• The project utilized data from PLCs connected to on/off switches.
• Multivariate Data Analysis (MDA) models, including PCA, PLS-DA, SVM, XGBoost, and SIMCA, were tested.
• A combination of Principal Component Analysis (PCA) and Similarity-based Approach (SIMCA) was selected as the optimal approach.

Main Results:

• The combined PCA and SIMCA models achieved sensitivity and specificity values exceeding 90%.
• The system provides maintenance alerts 30 minutes to 3 hours before potential breakdowns.
• Model performance is continuously improved through weekly updates and daily data collection.

Conclusions:

• The DOLPHINS project successfully established a real-time predictive maintenance method.
• The approach effectively utilizes existing PLC data and ML for preventive maintenance.
• This strategy significantly enhances machinery reliability and production continuity.