Signal Pattern Recognition And Confidence Evaluation For Loose Particle Detection Of Sealed Electronic Devices

One critical factor that may reduce the reliability of sealed electronic devices is loose particles, which need to be solved urgently in aerospace domain. Given the particle impact noise detection(PIND) for sealed electronic components, the loose particle detection technology gets rapid development for sealed electronic devices. The reliability of aerospace system has been guaranteed to some extent, but there are some shortcomings. Considering that the loose particle detection signal is complex as well as various, the component signal is easy to be misjudged into the particle signal, the performance of material recognition model is poor, and there is no effective method to evaluate confidence for loose particle detection, this dissertation has made an intense study of signals pattern recognition and confidence evaluation approaches for loose particle detection of sealed electronic devices.Considering that the sealed electronic device is heavy in weight, large in volume and complicate in inner structure, the loose particle detection signal is complex as well as various. Analyzing statistically the time and frequency characteristics of detection signals including particle signal, component signal, background noise and interference pulses, a coupled detection model was established. Based on particle-plate impact model, the particle material, particle size, mechanical test conditions, propagation distance and sensor positioning are considered as factors affecting particle signals, and the influence rules of them were investigated with single factor analysis and significance test methods. The results are of great value for pattern recognition and confidence measure evaluation.Considering that the types of component signals were classified with incomplete knowledge, which led to be misjudged into particle signal, a recognition method based clustering analysis was proposed for component signals. According the periodicity of component signals and randomness of particle signals, the types of different component signals were classified systematically based on time sequences for the first time. Aiming to integrate the different types of signals, after pulses extraction using energy threshold method and pulses superposition within the reference cycle, the original signal was transformed into pulse-phase distribution. On this basis, the clustering structure was determined, and a modified k-means algorithm was proposed for component signal recognition. The time sequence( pulse-phase distribution) for single channel recording and the similarity of the trend distribution of synchronized pulses for multichannel recordings are selected as features, the recognition rate achieves about 92% for the particle and component signals.Considering the poorly repeatable and overlapping features of different material particle signals, a recognition method based including hidden Markov model(HMM) and Mel frequency cepstral coefficient(MFCC) was proposed for particle materials. The HMM was introduced to particle material recognition from speech recognition technology. Based on probability statistical models, this method obtained the probability values that belong to different material models and then made a decision. Combined with the feature extraction method of MFCC, a filter bank based sensitive frequency was designed and an algorithm of modified MFCC was proposed. Finally, the material recognition model was generated based on HMM and modified MFCC, which can recognize the four types of materials including wire, chip, aluminum and tin. The experimental results show that classification performance of HMM and modified MFCC outperforms than the other conventional methods, the recognition rate achieves about 91%.Considering that the number of samples was limited, the parameters of model were uncertain and the confidence level was hard to control, a method of confidence measure evaluation based on conformal predictor(CP) was proposed. The time sequence for single channel recording and the similarity of the trend distribution for multichannel recordings were selected as features, which mapped into the K-nearest neighbors and conformal predictor CP-KNN algorithm, and then give the confidence measure for component signal recognition. Selected ―re-recognition‖ rather than ―class division‖as a guiding principle and modified MFCCs as the features, after dimension reduction analysis, concex hull covering two-dimensional model for material recognition was constructed. Combining the inner points anslysis and the CP-KNN algorithm, the confidence measure model was realized, which has such advantages as rejection feature, nearest neighbor recognition and credibility analysis for material recognition.