Multi-level and multi-scale cross attention network of wavelet packet transform for supersonic inlet unstart prediction

Abstract

Wavelet packet transform has demonstrated excellent performance as a time–frequency analysis method in machine fault diagnosis. However, the inherent non-linearity and multi-scale characteristics of these signals pose significant challenges to the parameter selection of wavelet packet transform. This paper introduces a novel approach by integrating attention blocks with multi-level wavelet packet transform. This integration allows for the extraction of multi-level and multi-scale features from the time–frequency domain of non-linear signals. The proposed method provides an efficient solution to enhance the feature representation of nonlinear signals through time–frequency analysis with different decomposition levels and scales. The effectiveness of this innovative approach is validated through experiments conducted on a real inlet model dataset. The results demonstrate that the proposed model achieved comprehensive testing accuracy of 99.48% and 99.36% in two experimental schemes, respectively, indicating that its recognition performance is superior to other existing supersonic inlet unstart prediction methods.