Improved SVDD-WMV Method for Fluidized Bed Multi-Sensor Agglomeration Detection

Multi-sensor based fluidized bed reactor (FBR) agglomeration monitoring system faces the problem with mismatching from different sensors. Moreover, acoustic signals are sensitive to agglomeration as well as the environment interference, so information fusion method is required to improve the stability of fault monitoring systems based on acoustic sensors. In this paper, a support vector data description (SVDD) combined with improved weighted majority voting (WMV) method is proposed for FBR. Firstly, sigmoid function is added to each SVDD model, so the Boolean outputs of SVDD are converted to probability estimations to meet the need of information fusion and improve the detection accuracy. Moreover, a multi-penalty parameter is designed to evaluate classifier in different situations, replacing the single overall penalty parameter in general WMV method. Through the penalty vector, performance of each classifier is added to the prior condition of voting. The proposed method is tested in a pilot device. From the test results, it can be concluded that the conflict handling performance of proposed method is enhanced greatly, and the decision risk is reduced. Compared with that of general method, the detection accuracy of proposed method is improved.