CADM+: Confusion-Based Learning Framework With Drift Detection and Adaptation for Real-Time Safety Assessment

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Real-time safety assessment (RTSA) is crucial for dynamic systems in industrial and electronic applications.
• Challenges include complex data streams, high labeling costs, and the need for continuous model adaptation.
• Existing methods struggle with the rapid, evolving nature of data in dynamic environments.

Purpose of the Study:

• To propose a novel confusion-based learning framework, CADM+, for effective RTSA.
• To address the challenges of data drift and concept evolution in dynamic systems.
• To enhance the accuracy and efficiency of safety assessments in real-time applications.

Main Methods:

• Introduced the confusion-and-detection method plus (CADM+) framework.
• Utilized cosine similarity to quantify conceptual confusion between existing and new data concepts.
• Employed the change in standard deviation within a cosine similarity window for drift detection.
• Applied extreme value theory (EVT) to establish drift detection thresholds.

Main Results:

• Theoretical analysis demonstrated the asymptotic increase of cosine similarity during drift.
• Empirical results showed the approximate independence of standard deviation changes from the number of trained samples.
• CADM+ achieved superior performance in RTSA tasks compared to state-of-the-art algorithms.
• The framework effectively handles data drift by updating models with uncertain samples.

Conclusions:

• CADM+ provides a robust and effective solution for real-time safety assessment in dynamic systems.
• The proposed method offers improved drift detection capabilities by leveraging conceptual confusion.
• The framework is well-suited for applications requiring continuous monitoring and adaptation of dynamic systems.