Physics-Aware Generative Demasking: Spatially Conditioned Diffusion for Robust Transient Detection in Industrial...

Area of Science:

• Acoustic signal processing
• Machine learning for industrial monitoring

Background:

• Detecting transient sounds like connector insertion clicks is crucial for automotive assembly quality.
• High-intensity, non-stationary industrial noise presents a significant challenge for accurate sound detection.

Purpose of the Study:

• To develop a robust method for detecting transient acoustic events amidst severe industrial noise.
• To enhance the precision and reliability of quality control in automotive assembly lines.

Main Methods:

• A physics-aware generative demasking framework integrating acoustic spatial priors with conditional diffusion modeling.
• Development of a spatially conditioned diffusion probabilistic model (SC-DPM) using ambient reference signals as physical constraints.
• Extraction of discriminative temporal patterns via causal random convolutional kernels and local proportion of positive values (LPPV) pooling.

Main Results:

• The SC-DPM effectively disentangles target transient sounds from background noise, reconstructing high-fidelity spectro-temporal features.
• Experiments on real-world datasets achieved 93.3% accuracy in detecting transient click sounds.
• The proposed 'restore-then-classify' paradigm demonstrated significant robustness against acoustic variability.

Conclusions:

• The developed framework offers a scalable methodology for precise industrial monitoring in extreme noise conditions.
• This approach significantly enhances the ability to detect critical acoustic signals for improved manufacturing quality control.