發(fā)布時(shí)間：2018-01-27 00:46

  本文關(guān)鍵詞： 軟測(cè)量建模 球磨機(jī)料位 極限學(xué)習(xí)機(jī) 深度極限學(xué)習(xí)機(jī) 最優(yōu)定界橢球 動(dòng)態(tài)軟測(cè)量建模　出處：《太原理工大學(xué)》2017年碩士論文　論文類型：學(xué)位論文

【摘要】：球磨機(jī)是工業(yè)生產(chǎn)過程中對(duì)物料進(jìn)行研磨破碎的關(guān)鍵設(shè)備,被普遍地使用于冶金、電力、選礦及化工等行業(yè)。其經(jīng)濟(jì)性與內(nèi)部料位相關(guān),料位過低導(dǎo)致當(dāng)前工作效率低,能源利用率不高,料位過高容易造成球磨機(jī)堵磨,存在安全隱患。因此,準(zhǔn)確地測(cè)量筒內(nèi)料位對(duì)實(shí)現(xiàn)球磨機(jī)的優(yōu)化控制至關(guān)重要。但是由于球磨機(jī)的密閉旋轉(zhuǎn)特性,在實(shí)際運(yùn)行過程中,料位很難通過相關(guān)傳感器直接測(cè)量,所以采用數(shù)據(jù)驅(qū)動(dòng)的建模方法,建立軟測(cè)量模型,通過輸入與球磨機(jī)料位相關(guān)的輔助變量來預(yù)測(cè)其料位。傳統(tǒng)的軟測(cè)量建模方法有很多種,包括支持向量機(jī)、偏最小二乘法、神經(jīng)網(wǎng)絡(luò)以及主元回歸分析法等,皆被廣泛地應(yīng)用在球磨機(jī)料位的建模過程中。作為神經(jīng)網(wǎng)絡(luò)的一種,極限學(xué)習(xí)機(jī)(Extreme Learning Machine,ELM)以其簡(jiǎn)潔高效的訓(xùn)練機(jī)制,避免了傳統(tǒng)神經(jīng)網(wǎng)絡(luò)的反向微調(diào)過程,從而提高模型的學(xué)習(xí)速率和泛化性,因此得到廣泛地應(yīng)用。然而,其前饋神經(jīng)網(wǎng)絡(luò)的隱含層輸出是通過某種概率進(jìn)行隨機(jī)選取,造成訓(xùn)練好的模型隨機(jī)性很大,預(yù)測(cè)結(jié)果不穩(wěn)定。此外ELM單隱含層的神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)也限制了其特征提取的能力。為了可以更好地通過無監(jiān)督訓(xùn)練過程學(xué)習(xí)高維數(shù)據(jù)中隱藏的抽象特征表示,研究學(xué)者又提出一種新的深度極限學(xué)習(xí)機(jī)算法(Deep Extreme learningmachine,delm),其采用多個(gè)自編碼器(autoencoder,ae)堆疊而成,逐層利用elm算法進(jìn)行誤差重構(gòu),并將前一層ae的輸出作為后一層ae的輸入,最終通過多層的ae獲得數(shù)據(jù)中更加抽象的特征表示。但是由于delm的每層網(wǎng)絡(luò)權(quán)值都是通過elm算法進(jìn)行計(jì)算得到的,其elm算法的隨機(jī)性,導(dǎo)致網(wǎng)絡(luò)權(quán)值也存在隨機(jī)性而并非最優(yōu),最終造成delm模型訓(xùn)練的不穩(wěn)定。球磨機(jī)在實(shí)際地運(yùn)行過程中,其存在著時(shí)變和工況遷移等復(fù)雜因素的影響。而傳統(tǒng)的軟測(cè)量建模方法,利用已有的離線數(shù)據(jù)進(jìn)行建模,模型一旦建立不再改變,對(duì)于新進(jìn)來的測(cè)試集不能很好地跟蹤當(dāng)前對(duì)象,從而造成模型預(yù)測(cè)性能的下降。因此,必須不斷地對(duì)軟測(cè)量模型進(jìn)行更新和校正。集員估計(jì)是一種在給定數(shù)據(jù)集,模型結(jié)構(gòu)以及噪聲邊界的條件下,描述可行參數(shù)集合的方法,該集合內(nèi)的參數(shù)可以看作是模型參數(shù)辨識(shí)的有效參數(shù)。而最優(yōu)定界橢球算法(optimalboundingellipsoid,obe)是集員估計(jì)理論中經(jīng)典算法之一,將其應(yīng)用在軟測(cè)量模型的參數(shù)優(yōu)化中,可以在給定誤差邊界的條件下,對(duì)模型參數(shù)進(jìn)行約束優(yōu)化,不僅改善模型的魯棒性,還能提高其預(yù)測(cè)精度�；诖�,本文主要做了以下研究:(1)在球磨機(jī)實(shí)驗(yàn)過程中,針對(duì)elm預(yù)測(cè)球磨機(jī)料位結(jié)果不穩(wěn)定的缺點(diǎn),本文采用obe,在誤差未知但有界的條件下,對(duì)訓(xùn)練好的elm網(wǎng)絡(luò)模型進(jìn)行優(yōu)化,提高模型的預(yù)測(cè)準(zhǔn)確度和穩(wěn)定性,并通過實(shí)驗(yàn)證明,該方法的有效性。(2)利用深度網(wǎng)絡(luò)對(duì)球磨機(jī)數(shù)據(jù)進(jìn)行軟測(cè)量建模時(shí),為了更好抽取樣本中最高層次的抽象表達(dá),本文提出一種多層OBE-ELM算法(Multi-Layer OBE-ELM,ML-OBE-ELM),基于自編碼器重構(gòu)思想,采用OBE迭代算法學(xué)習(xí)輸入數(shù)據(jù)的高層特征表示,最后利用ELM算法得到高層特征與樣本標(biāo)簽的關(guān)系式。為了驗(yàn)證該算法的有效性,選用傳統(tǒng)的UCI數(shù)據(jù)集和實(shí)際球磨機(jī)數(shù)據(jù)集作為實(shí)驗(yàn)數(shù)據(jù),分別驗(yàn)證了該算法在回歸和分類中都有較好的預(yù)測(cè)性能。(3)為了解決球磨機(jī),料位中時(shí)變和工況遷移的問題,本文提出基于OBE-PLS的動(dòng)態(tài)軟測(cè)量模型,首先利用離線數(shù)據(jù)訓(xùn)練軟測(cè)量模型,當(dāng)新的查詢樣本到達(dá)時(shí),利用OBE在原有模型的基礎(chǔ)上動(dòng)態(tài)地調(diào)整參數(shù),從而實(shí)現(xiàn)該模型對(duì)查詢樣本的實(shí)時(shí)跟蹤,并通過數(shù)值例子和小型球磨機(jī)實(shí)驗(yàn)對(duì)該方法的有效性進(jìn)行驗(yàn)證。
[Abstract]:The ball mill is a key equipment for grinding of materials in the process of industrial production, is widely used in metallurgy, electricity, mineral and chemical industry. Its economy and the internal material of related material, resulting in low current and low working efficiency, the energy utilization rate is not high, the material level is too high can easily cause plugging ball mill. There are security risks. Therefore, accurate measurement of cylinder material to realize the optimization control of ball mill is very important. But the mill closed rotation characteristics, in the actual operating process, the material level is very difficult by the relevant sensor, so the data driven modeling method, a soft measurement model, through the input and load of ball mill the auxiliary variables to predict the material level. There are many kinds of soft measurement of traditional modeling methods, including support vector machines, partial least squares, neural networks and principal component regression analysis, all Is widely used in the modeling process of ball mill bit. As a kind of neural network, the extreme learning machine (Extreme Learning Machine, ELM) with its simple and efficient training mechanism, to avoid the reverse process of fine-tuning the traditional neural network model, so as to improve the learning rate and generalization, therefore widely used however, the feedforward neural network hidden layer output is randomly selected by some probability, caused by the trained model of great randomness, the forecasting result is stable. In addition the ability of ELM neural network with one hidden layer also limit the feature extraction. In order to be better through unsupervised learning process abstract characteristics. In the high dimensional data representation, research scholars and put forward a new algorithm of machine learning depth limit (Deep Extreme learningmachine, delm), which uses multiple self encoder (autoencoder, AE) Stacked layer by layer, using elm algorithm for error reconstruction, and will enter a layer of AE AE as the output of the previous layer, the final features more abstract data obtained by multi AE said. But because each layer of the network weights of delm are calculated by elm algorithm, the random elm algorithm, also cause network weights are random and not optimal, resulting in delm model training is not stable. In the actual operation of the ball mill process, the influences of time-varying and migration condition of complex factors. The traditional methods of soft measurement modeling, off-line modeling by using existing data, model once no longer change, the new incoming test set cannot properly track the object, resulting in a decline in the prediction performance. Therefore, must be updated and correction of the soft measurement model of set membership estimation is constantly. In a given data set, the model structure and the noise boundary conditions, method of describing the feasible parameter set, the parameters in the collection can be regarded as effective parameter identification of model parameters. The optimal bounding ellipsoid algorithm (optimalboundingellipsoid, OBE) is a member of one of the classic estimation algorithm theory, its application in parameter optimization of soft measurement model, can in the given error boundary conditions for constrained optimization of the model parameters, not only improve the robustness of the model, but also improve the prediction accuracy. Based on this, this paper mainly do the following research: (1) in the ball mill in the course of the experiment, the elm prediction of ball mill level unstable result the shortcomings of the OBE, the error is unknown but bounded under the condition of the trained elm network model optimization, improve the prediction accuracy and stability, and proved by experiments, the 娉曠殑鏈夋晥鎬，

本文編號(hào)：1467075