Towards Safety in Machine Learning Using Validity Frames

As Machine Learning (ML) models are integrated into complex systems, their role in safety-critical scenarios and human-in-the-loop environments becomes increasingly significant. These scenarios require that the ML model be trustworthy and predictable. However, ML models are often considered black boxes, making them difficult to validate. As a result, ensuring system safety becomes considerably more difficult. In this paper, we propose the validity frame concept to manage the ML lifecycle and improve trust in the model. Within the frame, we capture key factors that influence a ML model’s performance throughout its lifecycle, from data collection and training to model deployment. These factors are explicitly linked back to the model requirements to ensure traceability. We present this approach using a practical case study involving anomaly detection in a Brushless Direct Current (BLDC) motor. This demonstrates how the validity frame enables lifecycle monitoring and informed design decisions. We discuss the viability of this approach, its limitations, and provide ideas for future work.