The existing research on loose particle detection only focuses on the identification of loose particle signals and component signals in laboratory scenarios, ignoring mixed signals and oversized component signals in real application scenarios, which reduces the reliability of detection results. And the existing research regards the pulse perspective as the main perspective, carries out feature calculation on pulses and constructs data to train classifiers, ignoring the signal perspective that provide feedback on the randomness motion of loose particles and periodic motion of components. Based on this, the authors took the above four types of signals as the research object, in addition to the loose particle signals and component signals, the mixed signals and oversized component signals were supplied, and the loose particle detection research was carried out from both signal and pulse perspectives. From the signal perspective, the authors directly carried out feature calculation on four types of signals, created the signal data set and trained multiple representative classifiers. From the pulse perspective, the authors first used the three-threshold pulse extraction algorithm to extract pulses from four types of signals, then carried out feature calculation on pulses, created the pulse data set and trained the same multiple classifiers. The experimental results show that, multiple representative classifiers all achieved higher accuracy on the signal data set, all achieve higher and more stable F1-Score on the four categories in the signal data set, and the random forest performs well in multiple classifiers. Therefore, for the accurate detection of loose particles inside sealed electronic components, the signal perspective is the optimal perspective, and the random forest is the optimal classifier worth considering.

Fault Diagnosis,week signal,supervised machine learning,measurement