<p>This paper proposes a novel method for time-series window prediction in non-stationary signals by integrating an improved non-negative matrix factorization (NMF) framework with frequency enhancement. First, the Short-Time Fourier Transform (STFT) is applied to decompose the non-stationary signal in the time-frequency domain, where the optimal factorization order is determined using a magnitude correlation algorithm. Next, an enhanced NMF algorithm is employed to extract the optimal feature matrix from the STFT magnitude spectrogram. Subsequently, a frequency-averaging technique is introduced to further refine the feature matrix, which is then iteratively updated subject to a self-contained smoothing constraint. Based on the feature-regularized NMF, a secondary weight matrix is derived to predict the optimal signal-to-noise ratio (SNR) time-series windows. Extensive simulations and experiments on multiple non-stationary signal datasets demonstrate that the proposed method not only achieves optimal SNR but also accurately identifies the best-performing time windows under diverse operating conditions.</p>