Fused feature vector construction method for loose particle localization on sealed electronic equipment

Sealed electronic equipment (SEE) is the core component of aerospace equipment, and loose particles inside it pose a major threat to the reliable operation of aerospace equipment. The particle impact noise detection method captures loose particle signals through the principle of vibration sound generation, and is currently the mainstream method for detecting loose particles. However, SEE has a large geometric volume and complex internal structure, and multiple sound sensors are required to cover the entire detection range, resulting in multiple related and complementary signals in the same detection. Existing research typically quantifies multiple signals independently, then mechanically concatenates them into fused feature vectors based on channel and feature numbers. However, this construction method, which is determined by manually set numbering rules, lacks systematic justification, and its impact on data set quality, classifier performance, and model stability is not yet clear. In response to the above problems, for the first time, this paper regards the fused feature vector construction method itself as an independent research object for loose particle localization on SEE. It systematically compares various fused feature vector construction methods using channels or features as the basic construction unit and satisfying ordered or disordered rules, focusing on two aspects, basic construction unit selection, and arrangement rule design. The experiment took typical SEE in the aerospace field as the object, created a large number of representative data sets, and compared and evaluated them with classifiers such as kNN, RBF-SVM, random forest, XGBoost, and 1D-CNN. The results indicate that, the fused feature vector construction method using features as the basic construction unit and satisfying ordered rules is overall optimal. Among them, the feature-change-ordered fused feature vector construction method combined with random forest achieved an average classification accuracy of 88.89% and showed good stability. In contrast, the fused feature vector construction methods satisfying disordered rules lead to a significant decrease in classification accuracy and a significant increase in classification accuracy variance for multiple classifiers. Among them, the average classification accuracy achieved by random forest on feature-change-disordered data sets decreases to 85.32%, and the classification accuracy variance increases to 1.0692. In practical applications, the random forest takes each fused feature vector as a single detection sample input, integrate multiple decision trees to jointly discriminate multi-channel and multi-feature information, and output the corresponding location category of loose particles. The research results indicate that, the construction rule of fused feature vectors is not a simple data organization step, but a key factor that directly affects the expression quality of multi-sensor signals, classifier robustness, and reliability of loose particle localization. The conclusion of this paper can provide a methodological basis for multi-sensor information fusion for loose particle detection on SEE, and provide reference for the design of measurement and intelligent diagnostic systems for large and complex equipment.