时间窗口 Noise2Noise：一种用于机械系统振动与冲击信号盲去噪的自监督方法
Time-Window Noise2Noise: A Self-Supervised Method for Blind Denoising of Vibration and Impact Signals in Mechanical Systems
浏览论文内容
中文总结 AI 辅助
本研究提出自监督盲去噪方法TWN2N，无需先验噪声信息，在两类基准上实现显著SNR增益，提升了模态参数识别精度。
中文摘要 AI 辅助
机械系统中的振动和冲击测量值总会被噪声污染，这会降低从这些测量值推导得到的所有物理量的精度，比如模态参数和接触力。经典滤波器只有在预先知晓噪声统计特性时才能衰减噪声，且在极低信噪比（SNR）或有色噪声场景下常常失效。本研究提出了时间窗口 Noise2Noise（TWN2N），这是一种用于确定性机械系统产生的信号盲去噪的自监督方法。该方法既不需要干净的参考信号，也不需要成对的含噪观测值，更不需要对噪声进行任何先验表征：通过从每个时间时刻的时间邻域重建该时刻，同时对该时刻本身进行掩码处理，迫使紧凑自编码器学习底层动力学而非噪声。该方法对任何确定性动力学系统均适用，计算量轻量（推理时每个样本仅需一次前向传播），并在两个具有精确已知真值的基准上进行了评估：线性3自由度系统的脉冲响应和非线性Hunt-Crossley冲击。在白噪声、粉噪声、布朗噪声和量化噪声场景下，经多次重复实验，TWN2N在无任何先验噪声信息的情况下，取得了具有统计显著性的SNR增益（p < 0.001），且与最优调优的Savitzky-Golay滤波器和小波（VisuShrink）滤波器进行了基准对比。该方法的优势可传递到识别任务：在输入SNR为15 dB时，原始信号中无法识别的模态参数，经去噪后恢复至其参考值的4%以内。
英文摘要
Vibration and impact measurements in mechanical systems are invariably corrupted by noise, which degrades every quantity derived from them, such as modal parameters and contact forces. Classical filters attenuate noise only when its statistics are known a priori and often fail at very low signal-to-noise ratio (SNR) or under spectrally coloured noise. This work presents Time-Window Noise2Noise (TWN2N), a self-supervised method for blind denoising of signals produced by deterministic mechanical systems. The method requires neither a clean reference signal nor paired noisy observations, nor any prior characterization of the noise: by reconstructing each time instant from its temporal neighbourhood while masking the instant itself, a compact autoencoder is forced to learn the underlying dynamics rather than the noise. The formulation is general for any deterministic dynamical system, is computationally lightweight (a single feed-forward pass per sample at inference), and is assessed on two benchmarks with exactly known ground truth: the impulse response of a linear 3-degree-of-freedom system and a nonlinear Hunt-Crossley impact. Across white, pink, brown and quantization noise, and over repeated realizations, TWN2N yields statistically significant SNR gains (p < 0.001) without any prior noise information, and it is benchmarked against optimally-tuned Savitzky-Golay and wavelet (VisuShrink) filters. The benefit propagates to identification: at 15 dB input SNR, modal parameters that are unidentifiable from the raw signal are recovered within 4% of their reference values after denoising.
