Agglomeration Detection Of Polythene-Making In A Fluid Bed With SVDD Based Multi-Sensor Information Fusion Method

Gas-phase polyethylene production process has been widely used because of the stable,simple,economical and safe technical advantages.Fluidized bed reactors are often used as its core reactor.However,in the process of large-scale continuous production with gas-solid fluidized bed,the hydrodynamic properties of the reactor will change due to the problems of dead corner,local static electricity or uneven distribution of reaction heat,resulting in agglomeration of particles,melting and flaking,and eventually forming caking problems,which will cause the shutdown of the reactor,interrupt the operation,inflict economic damage and even threaten the security of the equipment.In previous work,aiming at the incomplete collection problem of single sensor to fluidized beds,a multi-channel acoustic sensor monitoring system was designed.Multiple acoustic sensors of the same type affixed to different locations of the wall monitor the bed and collect signals.On this foundation,the early warning method of agglomeration fault of polymerized reaction of FBR based on multi-channel acoustic wave is studied in this paper.Firstly,the acoustic signal is decomposed by the singular value decomposition(SVD)under the Hankel matrix to extract features and compare the magnitude of the singular value between the normal and the agglomeration signals and to confirm that the energy distribution of the acoustic signal will be vastly different between diverse conditions,i.e.,singular values of acoustic signal under abnormal conditions are much higher than that under normal condition.The method can effectively extract the characteristics of agglomeration signal,has less undetermined parameters and the merits of simplicity,which is suitable for feature extraction of industrial signals.Secondly,according to the lack of abnormal labeled samples in the practical application,an agglomeration warning model is established by support vector data description(SVDD).In order to transform the Boolean output from SVDD model into probability value to adapt to information fusion,a probability conversion function is designed and the parameters of the conversion function are optimized by using the parameter maximum likelihood method based on the expectation maximization algorithm,which can be realized without abnormal fault samples.The process greatly enhances the descriptiveness of model output and reduces the influence of subjectivity of labels.Finally,in order to solve the problem of multi-sensor association,two information fusion methods,DS evidence theory and majority weight voting(WMV),are used to fuse the decision output of each sensor’s classifier.DS theory can effectively reduce the uncertainty of fault diagnosis and WMV can alleviate the impact of sensor performance differences on the diagnosis results,which improve the reliability of the final detection results.Furthermore,an improved method of WMV is proposed to correct its sensitivity to small agglomeration,which decrease the overall false alarm rate.Subsequently,the validity of the method proposed was tested on a pilot plant.The experimental results show that this method can detect agglomeration 40-60 minutes earlier than the traditional temperature-pressure method.Compared with the single sensor method,it has better reliability.By the contrast of the output results of each signal,the location of agglomeration can be estimated,which provides the guidance for the analysis of agglomeration and the application of control.