Browse Teach Curate

Frequency-Domain Filtering Layer for Efficient Time-Series Process Monitoring

How replacing a Transformer's self-attention with a learned frequency-domain filter yields a model that captures long-range temporal structure in multivariate monitoring logs at sub-quadratic cost. The student learns why self-attention is both inaccurate and expensive for semantically-thin step-wise logs, and how a discrete-Fourier-transform filtering layer addresses both the accuracy and efficiency limitations for industrial process monitoring.

K-2SO Text

IEEE TRANSACTIONS 1 DeepFilter: A Transformer-style Framework for Accurate and Efficient Process

DeepFilter, a Transformer-style neural architecture for industrial process monitoring that replaces the self-attention layer with an efficient frequency-domain filtering layer. The layer maps each ch…