One-class Classification-based Acoustic Inspection Method for Canned Foods

It is significant to inspect whether the vacuum degree of food container meets the standard. This paper proposes to treat it as a one-class classification problem. And for this, we present a one-class classification algorithm based on semi-non-negative matrix factorization: the classifier only needs to be learned from the dataset of qualified products, and then it can be used to judge whether the vacuum of the detected product is qualified or not. The detection results show that the proposed method can not only acquire the highest detection accuracy, but also correctly distinguishes the unqualified types that are easy to be misjudged by traditional methods. Introduction The three-piece steel cans are widely used as packaging containers to store food. Generally the food containers required to be vacuum for preventing the stored food from premature deterioration. It is significant to inspect whether the vacuum degree of food container meets the standard. The vacuum degree of the container is closely related to its internal pressure. Furthermore, the sound generated from the forced vibration of the container cover reflects the stress of the container cover. This inspired people to analyze the vacuum degree of the container by the sound produced by the vibration of the container cover. Hence the acoustic technology has been widely applied to detect the vacuum degree of canned foods in recent years [1,2,3]. At present, acoustic-based vacuum inspection technology for canned food mainly adopts spectrum peak method [1,2], that is to judge the vacuum of the product is qualified or unqualified according to whether the spectral peak frequency of the sound generated from the vibration of the container cover is within the appropriate range. The traditional method relies on the artificial experience to extract spectral peak frequency as the feature for inspecting the quality of canned foods. However, the sound emitted by the vibration of an object is usually a composite signal composed of many components with multiple frequencies [4], which may result in the fact that the spectral peak frequency does not afford sufficient discrimination between abnormal and normal cans. It was found from the practical application that the spectral peak frequencies of some unqualified canned foods are close to that of the qualified ones. These cause the misjudgment of the spectrum peak method. Actually the purpose of the inspection is only to distinguish whether a product is qualified or unqualified, without having to distinguish the specific types of disqualification. Therefore this paper proposes to treat the inspection as a one-class classification problem [5]. One-class classification is the task of discerning unusual samples in data. Typically, it is treated as an unsupervised learning problem where the anomalous samples are not known a priori and it is assumed that the majority of the training dataset consists of normal data [6]. The purpose then is to learn a model that accurately describes the “normality”. Deviations from this description are then deemed to be anomalies. It is widely used in text classification, spam detection, outlier images, video anomalies, machine fault detection, and so on [9]. The common classical one-class classification methods include the support vector machine-based one-class classifier (OneClass-SVM) [7], the auto-encoder-based one-class classifier (OneClass-AutoEncoder) [8]. International Conference on Modeling, Analysis, Simulation Technologies and Applications (MASTA 2019) Copyright © 2019, the Authors. Published by Atlantis Press. This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/). Advances in Intelligent Systems Research, volume 168