【作者】 叶廷东；

【导师】 刘桂雄；

【作者基本信息】 华南理工大学 ， 机械电子工程， 2010， 博士

【摘要】 实时、高准确度建模和动态校正理论方法是实现制造工业过程的在线、高准确度测量的关键,这对先进制造、仪器科学技术发展具有重要学术价值和实际意义。论文以“多传感信息建模与动态校正方法研究”为题,系统地研究多传感信息预处理方法、多传感信息建模解耦和预测补偿方法,并进一步进行网络化检测系统研制和相关实验与初步应用试验。论文得到教育部产学研结合项目(2007A090302039)和广州市科技计划项目(2005Z3-D0341)资助。论文对多传感信息预处理与建模校正方法的国内外研究进展和研究热点进行分析,确定论文将利用PLS和NPLS的相关分析、数据简化和多元回归能力,实现传感信息预处理与非线性建模,并结合时间序列分析方法、曲线拟合建模方法和小波多尺度分析方法,开展可在线快速计算的传感信息解耦与动态预测补偿方法研究。主要工作包括:①研究一种基于多项式外模型-内模型NPLS的多传感信息预处理与建模方法。在多项式外模型-内模型NPLS建模前端,引入基于PLS的预处理实现变量筛选,可使多项式外模型-内模型NPLS建模方法更有实用意义;②研究基于变量投影重要性-PLS回归系数的多传感信息变量筛选方法。该方法综合考虑变量投影重要性VIP指标、PLS回归系数对自变量的解释作用,有别于以往单一VIP指标作为变量筛选条件易出现误筛的现象。并提出以PLS回归模型拟合误差增量ΔE l作为变量筛选指标,无需逐个地考察每个自变量的重要性,具有计算量少的特点;③研究一种基于多项式外模型-内模型NPLS的双层非线性回归建模方法。该模型很好地表达了反应变量与解释变量之间、解释潜变量和反应潜变量之间以及反应变量相互之间的非线性关系,模型显式稳健,较好地解决了单独内外模型NPLS方法在应用中难于确定非线性项的问题;④提出一种简便的多传感信息尺度特征估计方法。该方法对所有传感信息仅进行一次N ( N≥6)尺度分解,求得分辨误差εij和分辨误差阈值ξ,进而完成多传感信息量的尺度特征估计,过程相对比较简单;⑤提出一种基于尺度逼近的多传感信息自适应插值解耦方法。根据各传感信息分辨级和在预估准确度目标下确定的分辩阈值δ,确定不同插值方法,完成多传感信息解耦计算;⑥提出一种提高传感动态性能的基于小波计算的传感信息动态预测模型。模型由多分辨近似树原理,利用àtrous算法进行在线小波分解计算,借助小波分析的低通滤波效应,有效抑制噪声干扰,应用基于滑动窗口的多项式预测算法SWPM和基于AR预测模型的并行Kalman递推估计算法REPK算法,分别对平滑层、分辩层信息进行动态预测,有效地利用各分解层信息特点,提高传感系统的动态性能;⑦系统研究REPK的实现算法和滚动混合式预测算法。REPK算法使用两个Kalman滤波器,交替进行AR模型参数的递推辨识与时变数据中真实信号的最优估计,能根据测量数据的最新分辨信息d j ,t实时修正AR模型参数进行预测,具有良好的计算一致性和收敛性,可推广应用到其它平稳时间序列信号的预测估计中;所提出的滚动混合式预测算法,能够克服长延迟传感信息预测中直接多步预测间隔时间过长问题,将一次长时间预测,分解为若干次直接多步预测,由实测数据开始,用前一次预测得到的数据实现后一次预测模型参数的滚动修正,使得最终预测信息,是由实测数据滚动修正预测获得的,降低预测误差。⑧结合检测通用化、智能化和网络化要求,设计一种基于嵌入式智能检测节点的网络化检测系统结构模型。研制用于多传感网络化检测的嵌入式智能检测节点,节点采用DSP和ARM微处理器为核心芯片,将所有传感量转换为频率信号,提高信号的抗干扰能力;用ARM的嵌入式微型因特网互联技术进行通信接口设计,在uClinux操作系统中引入IPv6通信模式,提高通信的安全性、可靠性和可扩展性。并讨论网络检测平台的软件结构与运行机制、基于XML的跨平台数据交换技术、基于XML数据的检测平台实时数据库技术等几个关键技术的解决方法。论文还开展相关仿真实验及应用试验,仿真结果表明,基于多项式外模型-内模型NPLS的多传感信息预处理与建模方法,可在少用拟合自变量的情况下,提高预测准确度(分别提高56.2%和24.7%);基于尺度逼近的多传感信息自适应插值解耦方法在预估准确度目标θ为0.1%下,通过分辩阈值δ计算,取δ=2-4,解耦时间50.4 ms,该方法与神经网络解耦方法相比,具有通用性好、收敛性好、运算速度较快等特点;基于小波计算的传感信息动态预测补偿方法,利用小波快速计算算法进行一次分解的时间为54.3ms,进行一次预测补偿的时间为127.0ms,具有良好的计算实时性;对低延迟传感信息进行直接三步预测时,准确度为0.538%。在发酵液及乙醇精馏中的检测试验初步应用结果表明,应用传感信息解耦与动态预测补偿技术后,使基于嵌入式智能检测节点的网络化检测系统具有较高检测准确度和较好实时性能,液态乙醇浓度的最大检测误差为-1.9%,传感检测响应时间从20s提高到1.3s。这些都表明本论文所研究的理论方法正确性、有效性,成果可推广到其它先进制造过程等应用领域。更多还原

【Abstract】 Real-time and high precision modeling &dynamic compensation method is a key to realize online and high precision measurement of manufacturing process, it has important academic value and practical significance in promoting the development of advanced manufacture and instrument technology. With the title“Study on Multiple Sensor Information Modeling &Dynamic Compensation”, the thesis systematically studies sensing information preprocessing, decoupling, prediction compensation method, and farther develops a networking measurement system, carries through correlative experiments and primary application. The thesis is supported by Guangzhou Science and Technology Planning Project (No.2005Z3-D0341) and Industry-Academy-Research Project of Education Ministry (2007A090302039).The thesis first analyzes the domestic and international researches on sensing information preprocessing &modeling method. It confirms the thesis will use correlation analysis, data reduction and multiple regressions ability of PLS and NPLS method to realize sensing information preprocessing and nolinar-modeling. And then the thesis combines time series analysis method ,curve fitting and wavelet-multiscale method together to develop online-rapid decoupling and prediction compensation method of sensing information The main work includes the following parts:I. It studies a NPLS preprocessing and modeling method based on outer-inner polynomial model. Before outer-inner polynomial NPLS modeling, importing variable selection based PLS can make the outer-inner polynomial NPLS modeling method have more practical significance.II. It studies a multiple sensor information variable selection method based on VIP-PLS regression coefficient. The method considers synthetically VIP index and PLS regression coefficient interpretative action on independent variables, and differing from the former method with single VIP filtration index, it doesn’t take place the phenomena of wrong filtration easily. And the method brings forward using error incrementΔEl of PLS model as variable filtering index, it needn’t review each independent variable’s importance, and it has virtue of small calculation work.III. It studies a double non-linearization PLS regression modeling method based on outer-inner polynomial model. The built model is explicit, steady and can express non-linear relation between explanatory variables and responsive variables, explanatorily latent variables and responsively latent variables, and among responsive variables commendably, it solves problem about the nonlinear terms is hard to confirmed in modeling process of outer polynomial NPLS model.IV. It brings forward a handy scale estimation method of multi-sensing information. The method just processes an N ( N≥6) scale decomposing for all sensing information, and works out resolution errorεij and resolution error thresholdξ, then it can fulfill scale estimation of multi-sensing information variables, its process is relatively simple.V. It brings forward a adaptive interpolation decoupling method of multi-sensing information based on scale approximation. The method select different interpolation method to fulfill decoupling calculation of multi-sensing information by resolution of each sensing information and resolution thresholdδcalculated under preestimating precision target.VI. It brings forward a dynamic prediction model of sensing information based on wavelet calculation to improve dynamic sensing characteristic. Based on multi-resolution approximation tree principle, the model usesàtrous arithmetic to process online wavelet-decomposing calculation, it can restrain noise disturbance effectively in virtue of low-pass filtering effect of wavelet analysis. The model uses Sliding Window Polynomial Model (SWPM) arithmetic, and Recursive Estimator based on Parallel Kalman (REPK) arithmetic of AR prediction model to dynamically predict scale information and detail information respectively, it can make use of each decomposed information’s characteristic effectively and improve dynamic performance of sensing system.VII. It systemically studies REPK arithmetic and composite-scroll prediction arithmetic. REPK arithmetic uses two Kalman filter to recursively identify parameters of AR model and optimally estimate true signal in time-varying data, it uses fresh resolution information d j ,tof measurement data to real-time amend parameters of AR model and predict, the arithmetic has good calculation consistency and convergence, and can be extended to prediction of other stationary time series signal. The proposed composite-scroll prediction arithmetic can overcome long interval problem in direct multi-step prediction method about prediction of long delay sensing information, it divides a long-time prediction to several direct multi-step prediction, and it starts from measured data, uses forward prediction data to amend parameters of afterward prediction model, and the final prediction data is scroll-amendatorily calculated from measured data, it decreases prediction error.VIII. Combined with request of measurement generalization, intelligentization and networking, it designs a networking structure model of measurement system based on embedded intelligent agent. The intelligent agent is used for networking measurement of multi-sensing information, it uses DSP and ARM as kernel chip, transforms all sensing information to frequency signal, and it increases antijamming ability of signal. The agent realizes the design of network communication by ARM embedded mini-internet technology, imports IPv6 communication mode in uClinux operation system, and it increases security, reliability and expansibility of communication. It also discusses resolvents of several pivotal technology about soft structure and operational mechanism, cross platform exchange technology based on XML, and real-time database technology of measurement platform based on XML data.The thesis also carries through correlative emulational experiments and applicational trial. The emulational result shows the NPLS preprocessing and modeling method based on outer-inner polynomial model can improve predictional precision(improves 56.2% and 24.7% respectively) with less fitting independent variables. After resolution thresholdδis calculatedδ=2-4 under preestimating precision targetθ=0.1%, the decoupling time of proposed adaptive interpolation decoupling method is 50.4 ms, the decoupling method has good generalization, convergence and rapid calculation speed compared with NN decoupling method. Useing wavelet rapid calculation arithmetic, the dynamic prediction model of sensing information based on wavelet calculation processes one time decomposition need 54.3ms, one time prediction compensation need 127.0ms, it has good real-time characteristic, and its precision is 0.538% when it uses direct three-step prediction method for low delay sensing information. After the decoupling and dynamic prediction compensation technology is used, the primary application result in measurement of ferment liquid and ethanol rectification shows the networking measurement system based on embedded intelligent agent has high measurement precision and good real-time characteristic, the maximal measurement error of liquid ethanol concentration is±1.9%, and improves responsing time of ethanol sensors from 20s to 1.3s. these show the researched theory and methods are correctness and validity, and can be extended to other advanced manufacturing processes.更多还原