A weight recognition method for movable objects in sealed cavity based on supervised learning. (December 2022)

• Record Type: Journal Article
• Authors: Lv, Bingze; Wang, Guotao; Li, Shuo; Wu, Yaoyang; Wang, Guoqiang
• Is Part Of: Measurement. Volume 205(2023)
• Publication Date: 2022-12
• Subjects: Weight recognition -- Machine Learning -- Qualitative analyses -- Quantitative analyses -- SVM -- MLP; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2022.112149
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: As for the qualitative measurement of the movable object in the unobservable sealed cavity, the qualitative analysis is introduced to narrow the weight range of the object and significantly improves the weight recognition effect for specific weight level. The weight recognition problem is innovatively transformed into classification and regression problems by using supervised learning methods. From the perspectives of time domain, frequency domain and time–frequency domain, the feature combination that can effectively characterize the weight information of movable object is calculated, and the feature selection is carried out by using the feature analysis method based on Random Forest importance ranking and Pearson correlation. A specific weight recognition method is designed to increase the weight prediction effect of the certain weight level, which can be applied to the priority recognition of objects with relatively serious hazard levels. We also design a kind of discriminant rule based on majority voting, which greatly enhance the practicability of this method to verify the model. Abstract: Weight information has important reference value for the analysis of the fault source of measurement and industry. When the object in the sealing cavity collides with the inner wall, it will produce undetectable weak sound signal containing weight information. Thus this paper presents a novel method to recognize the weight information of movable object in a sealed cavity. In this study, we compared and analyzed a variety of sound features, and a method combining qualitative and quantitative analysis is proposed to recognize the feature datasets that retained by feature selection. The support vector machine (SVM) algorithm was first adopted to determine the approximate range of object weight. Thereafter, the Multilayer Perceptron (MLP) based regression model was constructed to recognize the exact weight. Finally, weight detection experiment of movable particles from the sealed electronic components is utilized to verify the effectiveness of the proposed method.