Evolutionary Algorithm for Improving Decision Tree with Global Discretization in Manufacturing

Area of Science:

• Manufacturing
• Industrial Internet of Things (IoT)
• Data Science
• Machine Learning

Background:

• The industrial Internet of Things (IoT) generates vast sensor data, necessitating advanced fault detection methods.
• Interpretable machine learning, particularly tree-based algorithms, is crucial for reliable manufacturing and root cause analysis.
• Standard decision trees (C4.5, CART) face a trade-off between predictive accuracy and interpretability.

Purpose of the Study:

• To propose an evolutionary algorithm, DIMPLED, for discretizing multiple attributes to improve decision tree performance.
• To enhance the accuracy and maintain the interpretability of tree-based models for industrial fault detection.
• To address the limitations of single decision trees in handling complex manufacturing data.

Main Methods:

• Developed Decision tree Improved by Multiple sPLits with Evolutionary algorithm for Discretization (DIMPLED).
• Applied DIMPLED to discretize multiple attributes for decision tree construction.
• Evaluated DIMPLED using two real-world sensor datasets from manufacturing environments.

Main Results:

• DIMPLED-enhanced decision trees significantly outperformed traditional C4.5 and CART models.
• The proposed method demonstrated competitive performance against ensemble methods (multiple decision trees).
• DIMPLED achieved a favorable balance between model interpretability and predictive accuracy.

Conclusions:

• DIMPLED offers a superior approach to decision tree-based fault detection in industrial IoT settings.
• The method provides a more interpretable alternative to complex ensemble models while maintaining strong performance.
• DIMPLED contributes to building more reliable and understandable manufacturing systems through advanced data analysis.