Loose Particle Material Identification for Sealed Electronic Devices Using Pulse Endpoint Detection and CEEMDAN Feature Optimization

Loose particles inside aerospace-sealed electronics cause circuit short circuits and contact faults. Particle Impact Noise Detection (PIND) relies on piezoelectric acoustic emission (AE) sensors to capture collision pulses, yet raw sensor signals are heavily contaminated by background noise, leading to severe time–frequency feature aliasing and low particle material recognition accuracy. This work proposes a sensing signal optimization method combining pulse endpoint detection and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) decomposition for PIND acoustic-sensing systems. First, a frequency-domain variance dual-threshold algorithm extracts valid collision pulses from noisy sensor output and eliminates invalid noise segments. Second, CEEMDAN reconstruction with kurtosis-based IMF screening suppresses high-frequency impulsive noise and low-frequency trend components, and seven-dimensional time–frequency-fused features are extracted for classification. A two-hidden-layer back-propagation (BP) neural network identifies four typical contaminants: copper particles, solder particles, rubber particles, and epoxy particles. Comparative tests against EMD, EEMD, and wavelet thresholding show that the proposed CEEMDAN-based method raises overall classification accuracy from 71.8% (no denoising) to 85.1%. This approach improves the discrimination performance of PIND acoustic-sensing platforms and supports aerospace-packaging defect tracing.