Abstract
Villa elevator noise is predominantly concentrated in the 20–500 Hz frequency range, with a reverberation time of 0.2 s, which is far shorter than the 0.4 s of conventional elevators. This makes it challenging for adaptive filters to guarantee effective noise reduction performance. A novel deep learning-based active noise cancellation method is proposed based on Dominant Transmission Path and Complex-spectrum Network (DTPC-Net). First, a transmission path model was established, and the primary noise source and its dominant propagation direction were identified based on coherence analysis and energy distribution. Subsequently, a one-dimensional modeling strategy oriented toward the dominant propagation direction is proposed to address the high computational complexity of three-dimensional modeling. Furthermore, a convolutional recurrent network was designed to estimate the complex spectrum of noise signal. A hierarchical long short-term memory module and a Squeeze-and-Excitation attention module were incorporated to capture dynamically varying noise features. In addition, a predictive compensation mechanism based on blank frame padding was proposed to train the neural network and mitigate system delay. Finally, noise was reduced by 10.29 dB and 8.76 dB when the villa elevator operates at 0.4 m/s and 1 m/s, respectively. And the corresponding A-weighted reductions were 7.62 dBA and 7.54 dBA.