Analysis of CNN Internal Representations for Non-Stationary Processes in Rolling Bearings

This paper looks into interpretability problem and investigates whether intermediate Convolutional Neural Network (CNN) activations contain linearly recoverable information re- lated to analytic signal envelopes extracted from narrow-band vibration components. It is believed to help better understand algorithm’s and features’ design. A Multiscale CNN architecture (MSCNN) is employed with Frequency Attention (FA) after the initial convolution, where convolutional kernel processes features alongside the frequency axis while aggregating temporal informa- tion, in order to predict bearings’ Health Indicators (HI). Slight improvement in accuracy is observed across 6 testing bearings in parallel with attention weights’ extraction, which force the model to focus on certain frequency bands. A linear regression is intentionally overfitted on the intermediate CNN activations of individual samples to test whether the envelope of a narrow-band filtered vibration signal lies within the linear span of the learned features. High in-sample R^2 scores at deeper layers confirm that the network’s representations linearly contain the target envelope, while attention moderately improves representational quality at shallower layers. The results suggest that frequency- aware inductive biases enable earlier emergence of physically meaningful representations