Loose Particle Localization Method for Sealed Electronic Equipment Based on Improved Random Forest

Existing localization model based on normal random forest achieved limited classification accuracy, the author proposes an improved random forest, and applies it in the loose particle localization method for sealed electronic equipment. Specifically, multiple classification regression trees (CARTs) that more than predetermined are trained, and three reserved sample subsets are used to calculate the average classification accuracy achieved by each CART. Based on this, CARTs are arranged in descending order. Meanwhile, the vectorial inner product method is used to calculate the inner product values among the CARTs. The overall classification accuracy achieved by the random forest and the grid search method are used to determine the optimal inner product threshold. The inner product values between each pair of CARTs are compared with the inner product threshold, and the one in pair of CARTs with an inner product value smaller than the inner product threshold that achieves a lower average classification accuracy is marked as erasable. On this basis, the average classification accuracies achieved by CARTs and the correlations between them are comprehensive considered, and those CARTs with higher correlation and lower classification accuracy are deleted until the number of remaining CARTs reaches the predetermined. They are used to construct a high-performance improved random forest, from which a better localization model is obtained. Experimental results show that, the precision, recall and F1-value obtained by the localization model based on improved random forest are substantially improved compared with the previous one. Meanwhile, it achieves a classification accuracy of 90.72%, which is 6.4 higher than the previous one of 84.25%. This fully demonstrates the superiority and stability of the proposed improved random forest, and also shows the feasibility and practicality of applying it to solve the loose particle localization problem. Theoretically, it can be applied to the acoustic emission source localization or fault source localization research in similar fields, and provide valuable reference for algorithm optimization in machine learning.