Adaptive optimal kernel time-frequency representation for nonstationary signals
Abstract
Traditional bilinear time–frequency analysis methods employ fixed kernel functions, which makes it difficult to effectively suppress cross-term interference while maintaining high time–frequency resolution. To solve this problem, this paper proposes an adaptive optimal kernel (AOK) time–frequency analysis method based on hybrid kernel functions. In the ambiguity domain, the complementary characteristics of the Gaussian kernel, Choi-Williams kernel, and Slepian kernel are integrated to construct a novel hybrid kernel function. Meanwhile, the dynamic particle group optimization algorithm is used to adaptively optimize the weight coefficient. The elliptic support region of the Slepian kernel preserves the transient features of weak impact signals; the Choi-Williams kernel suppresses linear cross-terms within the elliptic domain; and the Gaussian kernel improves the global time–frequency concentration capability. By adjusting the weighting coefficients in stages, the hybrid kernel function achieves an optimal balance between weak feature extraction in the early fault stage and global focusing performance in the middle and late fault stages. Based on the CWRU bearing fault dataset, compared with the short-time Fourier transform, the time–frequency map generated by the AOK method can clearly display narrow-band high-brightness strips of fault characteristic frequencies. The background presents a uniform low-amplitude distribution, which effectively eliminates cross-term interference and background noise. The AOK method exhibits remarkable advantages in both time–frequency resolution and interference suppression performance.