Paweł Prokop

Condition monitoring of rotating machinery increasingly demands anomaly detection executed on the machine, under bounded computation, memory and delay. Compact one-class detectors built on learnable Sinc filter banks are candidates for such deployment, because their filters remain interpretable. How their design choices trade detection quality against those bounds under changing rotational speed has not been established. Interpretable Sinc filtering, speed dependence and edge deployment have been addressed separately, leaving the exchange between them open. The novelty of this work is a controlled comparative analysis of how the design parameters of such a detector influence detection performance, realized filter selectivity and estimated Edge AI requirements across different rotational speeds. Single-channel vibration acceleration was recorded under a nominal and a controlled eccentric-mass condition. Log-band-energy features were reconstructed by an autoencoder fitted on nominal data and five parameters were varied individually. Three findings were obtained. First, the parameter effects were strongly speed-dependent. Second, the parameters loaded different resources instead of a single complexity axis. The shortest kernel reduced computational cost by approximately 76% while producing a substantially wider realized passband. Third, the learned center frequencies occupied regions dominated by either condition and shifted nonmonotonically. Consequently, the filter bank did not simply converge on the strongest anomaly-related spectral regions or scale rigidly with rotational speed. Detector parameters therefore form distinct trade-offs set by the operating condition rather than a universal optimum, supporting comparison and configuration of detector architectures with respect to resource demands before validation on specific Edge AI hardware.