本文關鍵詞： 統(tǒng)計建模 小波 核方法 支持向量機 谷氨酸 發(fā)酵 廣義可加模型　出處：《江南大學》2011年博士論文　論文類型：學位論文

【摘要】：在當今信息時代,各種統(tǒng)計方法層出不窮,統(tǒng)計知識得到越來越多的應用。例如,統(tǒng)計的多尺度建模無論是在理論統(tǒng)計學還是在應用統(tǒng)計學中現(xiàn)都已成為熱門課題,這無論對統(tǒng)計方法還是其在各個應用科學領域的發(fā)展都起著沖擊作用;基于核的學習方法引起了數(shù)據(jù)分析領域的一場革命;廣義可加模型高度的靈活性,為有效揭示數(shù)據(jù)間所隱含的各種關系提供了一種有效的方法。在化工領域,一個有效的過程模型的建立,對研究如何科學規(guī)劃生產(chǎn)工藝,進而實現(xiàn)生產(chǎn)過程的優(yōu)化意義重大。 針對常規(guī)預測函數(shù)模型存在未將預測時域的優(yōu)化從總體上考慮的不足,在統(tǒng)計的多尺度建模方面研究后,基于小波多尺度的特性而提出了基于小波基函數(shù)和Hammerstein模型的預測函數(shù)模型,其內(nèi)部模型參數(shù)可以通過不斷辨識,自適應的進行校正。利用小波的緊支局部性和多尺度分析特性,既保證了整體誤差性能的優(yōu)化,又突出了重要擬合點的逼近要求,并實現(xiàn)了優(yōu)化變量的集結。理論分析和仿真應用表明,該方法有更好的跟蹤性和抗模型失配性能。 (1)針對如何提高核方法的建模精度的同時還要兼顧建模速度的問題,通過核方法研究,結合小波分析的理論,提出了小波融合核的建模方法。該方法具有小波多分辨率分析和核方法對輸入維數(shù)不敏感的特點,理論上在保證建模精度的前提下,有更快的建模速度。在此基礎上,分別通過一維函數(shù)和化工生產(chǎn)數(shù)據(jù)進行了仿真研究,仿真結果也驗證了算法的有效性。(2)由可分Hilbert空間與L~2 ( R )的等價性,利用內(nèi)積同構的線性算子,可以把L~2 ( R )中子空間的小波尺度函數(shù)折算為Hilbert空間中子空間的小波尺度函數(shù)�；谥С窒蛄繖C核函數(shù)的條件和小波多分辨率理論,在Hilbert空間構造出Morlet小波核函數(shù)。通過仿真實驗,與傳統(tǒng)的RBF核函數(shù)相比較,該尺度再生核函數(shù)具有更高的精度和更好的泛化能力。(3)在應用融合核支持向量機建模以提高模型的泛化能力和精度時,為避免在進行核融合時,支持向量機稀疏性的缺失,提出了將數(shù)據(jù)映射到稀疏特征空間進行研究。通過仿真研究表明,所建模型在保證稀疏性的前提下,能提高建模精度,從而驗證了算法的有效性,有良好的應用意義。 針對谷氨酸發(fā)酵過程復雜,如何解決難以建立有效的模型來指導生產(chǎn)過程優(yōu)化的現(xiàn)狀的研究中,發(fā)現(xiàn)廣義可加模型(GAM)能為谷氨酸的發(fā)酵過程提供行之有效的建模方法。利用該方法可以方便的分析不同的建模變量對谷氨酸產(chǎn)量的影響并從中得出與谷氨酸產(chǎn)量間的關系。研究中,基于15批次發(fā)酵實驗數(shù)據(jù),通過對不同影響因素的分析,最終選擇三個顯著影響因素(時間T、溶氧DO和氧攝取率OUR)來構建GAM模型,這一模型可以對谷氨酸的發(fā)酵過程解釋97%。該模型的構建成功,為研究發(fā)酵過程中不同因素對谷氨酸產(chǎn)量的影響提供了基礎。該模型不僅為根據(jù)在線數(shù)據(jù)預測谷氨酸產(chǎn)量提供了可行有效的方法,而且為發(fā)酵過程中在線故障診斷提供了新思路。在谷氨酸發(fā)酵過程故障診斷的方法研究中,提出了基于GAMs和Bootstrap方法的故障診斷方法。該方法能只依靠顯著觀測變量就可對發(fā)酵過程的狀態(tài)是否正常做出判斷,并能初步給出故障源相關的觀測變量。該方法只有很少的參數(shù)需要確定和調(diào)整,在發(fā)酵過程中,一方面能及時的對故障狀態(tài)進行報告,另一方面為排除故障源提供必要的參考信息,從而為發(fā)酵過程的正常運行提供了可靠的保障。 總之,隨著計算機技術的快速普及和廣泛發(fā)展,面對著數(shù)據(jù)和信息爆炸的挑戰(zhàn),為迅速有效地將數(shù)據(jù)提升為信息、知識和智能,統(tǒng)計建模方法在工業(yè)領域的研究意義重大。
[Abstract]:In today's information age, various statistical methods emerge in an endless stream of statistical knowledge, get more and more applications. For example, multiscale modeling statistics both in theoretical statistics or applied statistics have now become a hot topic, both the statistical methods and the application in various fields of science development plays a role in learning impact; method based on the kernel caused a revolution in the field of data analysis; generalized additive models can be highly flexibility, provides an effective method for revealing the implicit various relationships among data. In the chemical field, establish an effective process model, to study how to scientifically plan the production process, so as to realize the the optimization of production process.
Aiming at the shortage will not consider optimizing the overall prediction horizon from the existence of the conventional predictive function model, research in statistical aspects of multiscale modeling, feature based on wavelet multi-scale and proposes the prediction function model of wavelet function and Hammerstein model based on the internal model parameters can be through continuous identification, adaptive correction. Analysis of the characteristics of using compactly supported wavelets and multiscale, both to ensure the optimization of overall error performance, and some important points fitting, and the optimal parameters. The theoretical analysis and simulation show a better tracking performance and anti model mismatch performance of this method.
(1) in order to improve the modeling accuracy of kernel methods but also the modeling speed, by the nuclear method, combined with wavelet analysis theory, put forward the modeling method of wavelet fusion kernel. This method has the features of wavelet multiresolution analysis and kernel method is not sensitive to the input dimension theory, under the premise of ensuring modeling the accuracy of modeling, faster. On this basis, we have studied the one-dimensional function and chemical production data, the simulation results verify the validity of the algorithm. (2) when the Hilbert space and L~ 2 (R) of equivalence, using linear operator product isomorphism, can L~2 (R) wavelet scale function and wavelet scale function conversion neutron space as a subspace in Hilbert space. Conditions of the support vector kernel function and wavelet multi-resolution theory based on Hilbert space structure Morlet The nuclear wave function. Through the simulation experiment, compared with the traditional RBF kernel function, the scaling reproducing kernel function has higher accuracy and better generalization ability. (3) in the application of nuclear fusion support vector machine modeling to improve the generalization ability and the precision of the model, in order to avoid nuclear fusion timely, lack of support the sparse vector machine, the research data is mapped to the sparse feature space. The simulation results show that the model under the premise of guaranteeing sparsity, can improve the modeling accuracy, which verifies the validity of the algorithm, has good application significance.
In view of the glutamic acid fermentation process is complex, research how to solve it is difficult to establish an effective model to guide the status of production process optimization, find the generalized additive model (GAM) can provide effective modeling method for the fermentation process of glutamic acid. The method can affect convenient modeling and analysis of the different variables to the yield of glutamic acid and from and that the yield of glutamic acid. Among the studies, 15 batch fermentation based on experimental data, through the analysis of the influence of different factors, the final choice of the three significant factors (T, DO and dissolved oxygen uptake rate OUR) to construct the GAM model, this model can explain the fermentation process of glutamic acid was successfully constructed the 97%. model the influence of different factors to provide a basis for research on the fermentation process of glutamic acid production. The model not only for predicting the yield of glutamic acid according to online data provides a feasible The effective method, and provides a new idea for online fault diagnosis in the process of fermentation. In the research of fault diagnosis method of glutamic acid fermentation process, this paper presents a fault diagnosis method based on GAMs and Bootstrap method. This method can only rely on significant variables can be the state of the fermentation process is normal judgment, and observation variables the preliminary fault source. This method gives only a few parameters need to be determined and adjusted in the fermentation process, on the one hand to the failure state of the report, on the other hand, to provide the necessary information for troubleshooting source, so as to provide a reliable guarantee for the normal running of the fermentation process.
In short, with the rapid popularization and extensive development of computer technology, facing the challenge of data and information explosion, the research of statistical modeling is of great significance in the industrial field for rapidly and effectively upgrading data to information, knowledge and intelligence.

【學位授予單位】：江南大學
【學位級別】：博士
【學位授予年份】：2011
【分類號】：C81

【引證文獻】

相關博士學位論文 前2條

1 孟憲勇;圖模型基礎理論研究[D];東北師范大學;2012年

2 陳進東;基于模糊在線支持向量回歸的建模與預測控制研究[D];江南大學;2013年

相關碩士學位論文 前1條

1 韓順成;在線支持向量機在發(fā)酵過程建模中的應用[D];江南大學;2013年

，

本文編號：1474885