Research On The Integrated Modeling Technique For Acoustic Emission Signals And Its Application In Particle Measurement

Hydrodynamics parameters of fluidization, behaviour of reaction and properties in fluidized bed reactors are strongly affected by the movement and characters of the particles. And the characters of the particles would change with the development of reaction process. Fluctuation of the layer height and the particle size has strongly effect on some buildingup reactions. It is therefore vital important to measure parameters on line and study on the hydromechanical behavior of the fluidized bed. Besides, batch stirrred-tank reactors are also common-used facility for synthesizing chemicals such as granular sodium percarbonate.The marked characteristic of these reaction processes is that the the characteristic parameters can not remain stationary, it would change dynamically according to the reaction conditions. Until now, it is still a challenge task to measure the particle size or concentration of such dynamic system on line by traditional method.To measure the characteristic parameters by passive acoustic emission (AE) signals is a novel measuring technology due to its sensitive, environmental security, non-contact, non-invasive and real-time online features.This dissertation centered on how to measure the average particle size or solid concentration(also named as concentration) by asessing the AE signals that originated from the fluidized bed or stirring vessel. The AE signals were firstly decomposed by wavelet (or wavelet packet). Energies had been extracted from the details and approximations of AE signals. Soft sensing model employecd with modern data processing and modeling technology had finally been established for determining parameters such as the average particle size in a fluidized bed or concentration in a stirring vessel. These algorithms were not only unambiguous but also achieved high accuracy. There are significant applied cost and technical value to measure the particle size or concentration by AE signals by employing these soft sensing models. The dissertation centered on the following points and achieved corresponding achievements:1) Particle size measuring technology had been reviewed. The state of the arts of measuring characteristic parameters by AE signals as well as the processing method for AE signals had been discussed.2) Theory on wavelet or waveletpacket analysis as well as AE signal decomposing, denoising and reconstructing had been introduced. Algorithm on how to denoise AE signals by classic wavelet or wavepacket had also been discussed. State of the ats of modern modeling technique and its application in characteristic parameters determining by AE signals had been reviewed.3) Average particle size measuring technology in a model fluidized bed by passive AE signals had been proposed by principal component analysis, neural networks and wavelet (wavepacket) decomposing. The relationship between acoustic emission signals and average particle size had been established. The AE signals originated by various particle sizes were decomposed by wavelet analysis (WLA) or wavepacket analysis (WLPA). Energies, the summations of absolute wavelet coefficients of detail signals and approximation signals, were used as recognition pattern. Principal components analysis (PCA) had been used to eliminate the complex relativity and the number of variables. A multi-layer feed forward neural network (MLFN) for regression had been established, in which the principal components were used as inputs of the neural networks and the average particle size was used as output. Factors such as the type of the wavelet or the decomposed level that influence on the prediction accuracy had been investigated. The results showed that the Sym8 WLA-PCA-MLFN model achieved high accuracy on average particle size regression by acoustic emission signals.Four-dimensional energy pattern had been finally obtained after the original AE signals had been 2-level decomposed by Haar waveletpacket. Both the radical base function neural network based on Haar waveletpacket analysis and principal components analysis (Haar WLPA-PCA- RBFN) and the Haar WLPA-RBFN model achieved high accuracy when they were used to predict the average particle size in a model fluidized bed. The regularization radical base neural networks can be constructed easily due to only one parameter is needed.4) Models had been propounded for classifying average particle size in a stirring vessel by acoustic emission signals. The AE signals had been 2-level decomposed by Sym2 WLP. Standardized energy pattern had been obtained after the absolute coefficients of the detail signals were summed. The variables had been selected and verified by stepwise discriminate analysis and mahalanobis statistic. The Sym2 WLPA-Bayes model or the Sym2 WLPA-MDis model achieved high accuracy when they were used to classify the average particle size in a stirring vessel according to the AE signals on condition that the concentration and rotating rate remained changeless.It could also be used to predict the concentration of a stirring vessel on condition that the rotating speed and the average particle size remained changless.5) A Sym2 WLPA-PCA-LSSVM (least square support vectors machine) model had been proposed for classifying the average particle sizes or concentrations according to the AE signals originated from multi-concentrations and multi-particle sizes of a stirring vessel. Factors that influence on the precision had been discussed. The accuracy of prediction as well as validation of this Sym2 WLPA-PCA-LSSVM model is superior to discriminant analysis model of the Sym2 WLPA-PCA-MDis.The Sym2 WLPA-PCA-LSSVM model could also achieve high accuracy when it was applied to measure the average particle size of the fluidized bed.The LSSVM model is suitable even if the number of the individual of a training set is not large enough. It had no bearing up on the problem of "overfitting" or "lack of study".6) On condition that the relationship between AE signals and concentration or average particle size in a stirring vessel was unclear, Cluster analysis was often needed to the datasets that lack of prior knowledge. Meaningful revelations had been obtained through cluster analysis. For example, there was clear cluster structure between average particle size and energy pattern of AE signals on condition that the concentration and other parameters remain changeless. Statistical method was therefore suitable for relating AE signals and average particle size in a stirring vessel.AE signals had been firstly 2-level decomposed by waveletpacket of Sym2, standard energies obtained from the detail signals were used for cluster analysis. Results of hierarchical clustering showed there was good relationship between energies and the average particle sizes on condition that the concentration remained changeless. Similarly, Results of hierarchical clustering also indicated that the energies had relationship with the concentrations if the average particle size and other parameters remained changeless.There wasn't distinguishable boundary between two adjacent average particle sizes, and the FCM algorithm often converged to local minima, A Sym2 WLPA-GA-FCM algorithm had been proposed for clustering AE signals. Clustering results based on GA-FCM algorithm were in agreement with practical class if the average particle size or the concentration remained changeless.Neither hierarchical clustering nor GA-FCM clustering analysis showed there was clear relationship between the AE signals and the concentrations if the AE signals had been originated from multi-concentration and multi-particle size. This indicated that the relationship between the AE signals and the concentrations could not be classified by statistical methods.