Switching Stabilization and Transient Artifact Suppression in a Hybrid CNN-SFANC-FxNLMS Controller

Frame-level classification fluctuations in convolutional neural network (CNN)-based selective fixed-filter active noise control (SFANC) can trigger abrupt controller-filter switching and click/pop transients. This study proposes a switching-stabilization strategy for a hybrid CNN–SFANC–FxNLMS framework that combines entropy-adaptive Kalman smoothing of class decisions with coefficient-difference-dependent linear crossfading (FADE50). Switching artifacts are assessed using robust POP-event detection, switch-aligned temporal analysis, and relative short-time Fourier transform (STFT) measures. Numerical tests using recorded aircraft and traffic noise cover abrupt, gradual, and multistage cyclic transitions, with FxNLMS adaptation disabled to isolate switching effects. Relative to RAW_HARD, KALMAN_FADE50 reduces representative-event POP intensity by 92.74–97.84% and representative stripe score by 94.46–95.09%. In the cyclic scenario, detected events decrease from 892 to 823 (7.74%).