Detecting and Explaining Performance Differences and Cycle Time Variability in Industrial Robots Using Multivariate Machine Learning Models

Abstract

Modern material handling relies heavily on automation, so it is important to understand why some robots complete tasks faster and more efficiently than others. This thesis looks at performance differences and cycle time variation in customized gantry robots built by Cimcorp Oy. The analysis uses machine learning models and real-world data collected from eight robots over a long period of operation.

The full dataset was divided into eight groups based on different working conditions, and two operating states were chosen for detailed study. Each dataset included numeric signal data from individual robot cycles. Time-based features were removed to focus on internal signals, and all signal names were anonymised to protect the client’s confidentiality.

A supervised XGBoost regression model was used to predict cycle times in both selected conditions. The model gave high predictive accuracy, with R² values of 0.92 and 0.77. To understand which signals had the biggest impact on cycle time, SHAP (SHapley Additive exPlanations) and Sobol sensitivity analysis were used. PCA (Principal Component Analysis) was also used to find patterns in robot behaviour and show how different robots performed under the same conditions.

These results show that machine learning, combined with explainable AI methods, can help predict and understand differences in robot performance. The model's reliability was further confirmed by testing on new unseen data, and an additional efficiency ranking method was developed using predicted versus actual cycle times to highlight robot level performance differences. This approach supports improved monitoring, decision making, and operational efficiency in industrial automation systems.