For time-series models in safety-critical domains, you can now extract interpretable concepts that appear in both time and frequency domains automatically, with better alignment to what the model actually uses for decisions.
CENDRe is a method for understanding what patterns CNNs learn from time-series data by extracting interpretable concepts in both time and frequency domains. Unlike existing approaches, it automatically determines how many concepts to extract and produces localizations that align with regions the model actually uses for predictions, making it useful for critical applications like fault diagnosis.