趙薇玲 章軍輝 陳明亮 李慶 陳大鵬
摘要: 本文探討了人工智能技術(shù)在紡紗質(zhì)量預(yù)測領(lǐng)域的應(yīng)用、創(chuàng)新與不足,介紹了Hadoop技術(shù)為紡紗質(zhì)量預(yù)測建模提供可靠高效的數(shù)據(jù)處理與運(yùn)算平臺,重點(diǎn)闡述了智能建模方法在紡紗質(zhì)量預(yù)測領(lǐng)域的研究進(jìn)展。通過分析得出基于數(shù)據(jù)與知識融合驅(qū)動的人工智能技術(shù),構(gòu)建出多工序關(guān)聯(lián)的混合智能模型,用以準(zhǔn)確描述紗線質(zhì)量與纖維特性、工藝參數(shù)、環(huán)境參數(shù)等之間的非線性映射關(guān)系,可為試紡、過程參數(shù)設(shè)計(jì)、態(tài)勢預(yù)測等環(huán)節(jié)提供指導(dǎo),具有重要的理論研究意義。
關(guān)鍵詞: 紡紗質(zhì)量預(yù)測;人工智能;Hadoop技術(shù);混合智能模型;目標(biāo)優(yōu)化;發(fā)展趨勢
中圖分類號: TS104.7
文獻(xiàn)標(biāo)志碼: A
文章編號: 1001-7003(2023)04-0061-10
引用頁碼:
041109
DOI: 10.3969/j.issn.1001-7003.2023.04.009(篇序)
紡紗是紡織產(chǎn)業(yè)鏈中的首道工序,根據(jù)紡織技術(shù)的不同,可以分為環(huán)錠紡、轉(zhuǎn)杯紡、噴氣紡及其他新型紡紗,不管什么類型的紗線都會受原料質(zhì)量、工藝參數(shù)等因素的影響,進(jìn)而影響到織造、印染等后道工序的品質(zhì)??茖W(xué)合理的紡紗質(zhì)量預(yù)測不僅能夠減少原料浪費(fèi)、提升紗線質(zhì)量,還可以協(xié)調(diào)成本與質(zhì)量的相互關(guān)系。高效準(zhǔn)確的底層數(shù)據(jù)支撐、合理的特征因素及有效的預(yù)測模型都決定了紡紗質(zhì)量預(yù)測的有效性,甚至關(guān)系到試紡、工藝優(yōu)化等任務(wù)的進(jìn)一步開展[1]。毛羽、強(qiáng)度、條干不勻率、斷裂伸長率等紗線質(zhì)量指標(biāo)是紡紗工序中的重要預(yù)測目標(biāo)。決定成紗質(zhì)量的兩個(gè)關(guān)鍵因素是原料性能和紡紗工藝相關(guān)參數(shù),因此預(yù)測模型的輸入特征通常是纖維長度、纖維強(qiáng)度、馬克隆值、斷裂伸長率、長度不勻率等棉纖維屬性,纖維長度、纖維線密度、短毛率、離散系數(shù)等毛纖維屬性及其他纖維屬性,捻度、混紡比等紡紗參數(shù),錠子、羅拉等環(huán)錠紡設(shè)備參數(shù),轉(zhuǎn)杯類型、轉(zhuǎn)杯直徑等轉(zhuǎn)杯紡設(shè)備參數(shù),以及其他機(jī)器部件的速度、某些設(shè)備之間的距離等[2]。人工智能方法和統(tǒng)計(jì)方法是建立紡紗質(zhì)量預(yù)測模型的兩種主要方法,人工智能方法的預(yù)測比統(tǒng)計(jì)方法更加準(zhǔn)確,因此研究人員在人工智能方法應(yīng)用于紡紗制造過程建模研究中取得了許多研究成果[3]。
人工智能技術(shù)在紡紗行業(yè)有多種應(yīng)用,尤其是檢測和預(yù)測紗線的質(zhì)量參數(shù)方面。通過適當(dāng)模擬不同紗線類型的生產(chǎn)過程,智能建??梢愿纳萍喚€性能評估和整體質(zhì)量控制[4]。探討人工智能技術(shù)在紡紗質(zhì)量預(yù)測領(lǐng)域中的應(yīng)用,在理論上可以深化對紗線工藝制造的認(rèn)識,在實(shí)際應(yīng)用中對于發(fā)展新的預(yù)測模型及優(yōu)化方向也具有重要指導(dǎo)意義[5]。
本文介紹了Hadoop技術(shù)在整個(gè)紡紗質(zhì)量預(yù)測建模流程中,作為數(shù)據(jù)處理與運(yùn)算平臺的優(yōu)勢,重點(diǎn)闡述了人工智能技術(shù)在紡紗質(zhì)量預(yù)測領(lǐng)域的研究現(xiàn)狀,最后對紡紗質(zhì)量預(yù)測研究的共性問題及發(fā)展趨勢進(jìn)行了總結(jié)與展望。
1 紡紗工業(yè)大數(shù)據(jù)處理
1.1 數(shù)據(jù)獲取
針對紡紗車間設(shè)備互聯(lián)困難、不同系統(tǒng)間“信息孤島”等問題,一直以來沒有廣泛認(rèn)可的一體化解決方案,隨著網(wǎng)絡(luò)技術(shù)和新興技術(shù)的發(fā)展,多種網(wǎng)絡(luò)連接技術(shù)和通信技術(shù)為設(shè)備間的互聯(lián)及數(shù)據(jù)交互的低時(shí)延、高可靠性提供了基礎(chǔ)設(shè)施保障和技術(shù)支撐?,F(xiàn)有的典型技術(shù)有現(xiàn)場總線、工業(yè)以太網(wǎng)、工業(yè)無線、5G等;新興技術(shù)有邊緣計(jì)算、軟件定義網(wǎng)絡(luò)(SDN)、時(shí)間敏感網(wǎng)絡(luò)(TSN)等[6]?,F(xiàn)場總線結(jié)合
工業(yè)以太網(wǎng)完成紡紗設(shè)備物理層和鏈路層的數(shù)據(jù)傳輸,利用PLC、RFID、智能化設(shè)備接口和人工輔助等多種數(shù)據(jù)采集方式,通過PLC擴(kuò)展模塊、網(wǎng)關(guān)、總線橋、工業(yè)交換機(jī)和以太網(wǎng)模塊等設(shè)備完成現(xiàn)場總線組網(wǎng),總線網(wǎng)絡(luò)向上集成工業(yè)以太網(wǎng),以TCP/IP作為公共協(xié)議,采用“多對一”方式實(shí)現(xiàn)多協(xié)議的集成,是未來紡紗工業(yè)互聯(lián)互通的主要實(shí)施路徑[7]。
1.2 數(shù)據(jù)預(yù)處理
1.2.1 數(shù)據(jù)集成
由于紡紗生產(chǎn)制造工序較長,生產(chǎn)數(shù)據(jù)種類繁多,分布分散,數(shù)據(jù)量大,超越了傳統(tǒng)的存儲方式和數(shù)據(jù)庫管理工具的功能范圍,從而導(dǎo)致企業(yè)面臨數(shù)據(jù)存儲能力不足、數(shù)據(jù)處理繁瑣復(fù)雜等困境。隨著大數(shù)據(jù)技術(shù)的成熟和發(fā)展,依托大數(shù)據(jù)存儲與處理技術(shù)可以實(shí)現(xiàn)海量紡織數(shù)據(jù)的可靠存儲和高效運(yùn)算。Hadoop已經(jīng)成為大數(shù)據(jù)技術(shù)領(lǐng)域中成熟的代表,它具有分布式存儲和計(jì)算的優(yōu)勢,是一個(gè)彈性可擴(kuò)展式的開源軟件框架,用戶根據(jù)應(yīng)用需求在其基礎(chǔ)上構(gòu)建模塊,共同組成Hadoop生態(tài)系統(tǒng)。Hadoop包含Hadoop分布式文件系統(tǒng)和MapReduce計(jì)算引擎兩個(gè)主要組件,其他常用組件有Hbase分布式數(shù)據(jù)庫、Hive數(shù)據(jù)倉庫工具等[8]。
先進(jìn)的紡紗企業(yè)或相關(guān)研究機(jī)構(gòu)已經(jīng)開始應(yīng)用Hadoop技術(shù)構(gòu)建自己的紡紗大數(shù)據(jù)平臺,依托大數(shù)據(jù)架構(gòu)實(shí)現(xiàn)紡織制造執(zhí)行系統(tǒng)的改造升級[9]。邵景峰等[10]基于Hadoop技術(shù)構(gòu)建集成管理平臺,融合集成全流程紡紗生產(chǎn)數(shù)據(jù),利用智能建模技術(shù)挖掘影響紡紗質(zhì)量的關(guān)鍵因素,分析車間運(yùn)行的潛在規(guī)律,為在線質(zhì)量檢測提供技術(shù)支撐。
傳統(tǒng)數(shù)據(jù)庫對于大規(guī)模數(shù)據(jù)存儲有限、處理緩慢,在一些開展數(shù)字化轉(zhuǎn)型的紡紗企業(yè)所構(gòu)建的制造執(zhí)行架構(gòu)模型中,以分布式數(shù)據(jù)庫作為底層服務(wù),集中管理、傳遞和存儲數(shù)據(jù)[11]。馮立增等[12]提出HBase與MySQL雙數(shù)據(jù)庫存儲方式來改進(jìn)現(xiàn)有紡織信息系統(tǒng)的傳統(tǒng)數(shù)據(jù)存儲方式,其數(shù)據(jù)集成與處理平臺設(shè)計(jì)如圖1所示。先將采集到的各工序生產(chǎn)數(shù)據(jù)存入HBase中,便于系統(tǒng)快速更新,基于并行計(jì)算模式經(jīng)過質(zhì)量智能分析后,再將數(shù)據(jù)存入MySQL數(shù)據(jù)庫中進(jìn)行基礎(chǔ)操作。HBase能夠?qū)εc紗線質(zhì)量相關(guān)聯(lián)的各工序參數(shù)快速掃描并獲取相關(guān)數(shù)據(jù),為質(zhì)量預(yù)測等數(shù)據(jù)分析操作提供高效、準(zhǔn)確的數(shù)據(jù)支持,支撐頂層的多維分析與生產(chǎn)應(yīng)用。
圖2是基于Hadoop大數(shù)據(jù)平臺的紡紗質(zhì)量預(yù)測系統(tǒng)總體框架,整體框架一般為三層:數(shù)據(jù)存儲層、數(shù)據(jù)處理層和數(shù)據(jù)應(yīng)用層。數(shù)據(jù)存儲層主要是存儲紡紗生產(chǎn)環(huán)境中產(chǎn)生的數(shù)據(jù),將全流程數(shù)據(jù)進(jìn)行融合,過濾處理后進(jìn)行分布式存儲;數(shù)據(jù)處理層經(jīng)過數(shù)據(jù)預(yù)處理后利用先進(jìn)的智能方法建立紡紗質(zhì)量預(yù)測模型,進(jìn)行快速分析;數(shù)據(jù)應(yīng)用層是根據(jù)預(yù)測結(jié)果為企業(yè)生產(chǎn)提供決策支持。
根據(jù)前述有關(guān)紡織行業(yè)的解決方案可以看出,基于Hadoop技術(shù)的大數(shù)據(jù)驅(qū)動框架,建立紡紗質(zhì)量預(yù)測模型的數(shù)據(jù)處理與分析平臺具有重要意義。以大數(shù)據(jù)平臺為載體,依賴于并行計(jì)算方式,同時(shí)構(gòu)建適應(yīng)于海量數(shù)據(jù)的紡紗質(zhì)量預(yù)測算法與模型,挖掘出紡紗數(shù)據(jù)之間潛在的關(guān)系和價(jià)值,從而根據(jù)結(jié)果進(jìn)行紡織行業(yè)的預(yù)測與研判[13-16]。
1.2.2 數(shù)據(jù)清洗
根據(jù)業(yè)務(wù)和場景應(yīng)用的需求,通過Hive數(shù)據(jù)倉庫工具將SQL語句轉(zhuǎn)換為MapReduce任務(wù),對數(shù)據(jù)集中存在的重復(fù)和極值進(jìn)行剔除,對缺失值采取刪除或填充操作,填充方法包括全局變量等通用方式和專家推理預(yù)測方法[17]。最后,根據(jù)清洗需求進(jìn)行合理的數(shù)據(jù)轉(zhuǎn)換,如圖3所示。
1.2.3 特征優(yōu)化
紡紗生產(chǎn)過程中與紗線質(zhì)量相關(guān)聯(lián)的因素眾多,數(shù)據(jù)量大,如何在海量紡紗數(shù)據(jù)中獲取高質(zhì)量數(shù)據(jù)以減少計(jì)算資源是研究者們關(guān)注的問題之一。在預(yù)測建模前,多數(shù)研究會對相關(guān)特征變量的重要性進(jìn)行評估,評估結(jié)果影響建模分析效率和模型預(yù)測精度。主成分分析(Principal component analysis,PCA)與方差分析(Analysis of variance,ANOVA)是紡紗智能建模前進(jìn)行特征變量選擇的重要方法[18-21]。從處理方式來看,PCA側(cè)重于數(shù)據(jù)量壓縮和降低計(jì)算成本,ANOVA更注重特征與預(yù)測目標(biāo)之間的關(guān)聯(lián)性[22]。其他特征優(yōu)化方法還包括:灰色關(guān)聯(lián)分析[23-25]、敏感性分析[26]、逐步回歸分析[27]及專家經(jīng)驗(yàn)結(jié)合皮爾森系數(shù)[17]等方法。利用這些方式進(jìn)行特征優(yōu)化后,不僅加快了預(yù)測模型的運(yùn)行速度,還提高了模型預(yù)測精確度。
針對維度高、數(shù)據(jù)類型復(fù)雜的情形,聚類分析也有利于快速提取有效數(shù)據(jù)[28-29]。為了提升聚類算法面向大規(guī)模紡紗數(shù)據(jù)的全局尋優(yōu)能力,邵景峰等[30]提出了分布式聚類算法,在面向分布式環(huán)境下該算法與傳統(tǒng)K-means聚類算法相比,體現(xiàn)出全局尋優(yōu)能力更強(qiáng)、收斂平穩(wěn)和速度快的優(yōu)勢。
1.2.4 數(shù)據(jù)變換
max-min標(biāo)準(zhǔn)化和z-score變換是消除不同量綱對模型影響的常用方法。紡紗理論研究通常采用max-min方法進(jìn)行數(shù)據(jù)變換,如下式所示:
x*=x-xminxmax-xmin(1)
式中:x*為紡紗生產(chǎn)數(shù)據(jù)規(guī)范化后的值;x為紡紗生產(chǎn)數(shù)據(jù)原始值;xmin為紡紗生產(chǎn)數(shù)據(jù)最小值;xmax為紡紗生產(chǎn)數(shù)據(jù)最大值。
2 人工智能預(yù)測方法
2.1 基于數(shù)據(jù)驅(qū)動的方法
支持向量機(jī)(Support vector machines,SVM)實(shí)施結(jié)構(gòu)風(fēng)險(xiǎn)最小化原則,具有良好的模型泛化能力,適用于小樣本建模。在基于SVM的紡紗質(zhì)量預(yù)測研究中,Ghosh等[31]以纖維參數(shù)預(yù)測棉紗的強(qiáng)度、斷裂伸長率、不均勻度和毛羽,SVM模型訓(xùn)練和測試預(yù)測精度都高于ANN模型,并且在噪聲數(shù)據(jù)下SVM比ANN更能保持預(yù)測穩(wěn)定性。項(xiàng)前等[32]和谷有眾等[33]同樣基于小樣本驗(yàn)證了SVM的泛化能力。此外,針對SVM模型核函數(shù)和參數(shù)確定的問題,對于多數(shù)紗線特性,徑向基核函數(shù)略占優(yōu)勢,更為常用[34]。傳統(tǒng)的參數(shù)優(yōu)化方法如經(jīng)驗(yàn)法、試錯(cuò)法和網(wǎng)格搜索法等費(fèi)時(shí)且容易陷入局部最優(yōu),呂志軍等[35]給出了遺傳算法(Genetic algorithm,GA)優(yōu)化參數(shù)的策略,優(yōu)化后的紗線強(qiáng)度預(yù)測模型的預(yù)測精度有所提高,支持向量機(jī)個(gè)數(shù)減少,增強(qiáng)了模型泛化性能,但是模型僅以纖維性能指標(biāo)預(yù)測紗線強(qiáng)度,模型輸入和輸出類型單一。宋楚平等[36]指出這一問題,添加了設(shè)計(jì)參數(shù)如捻度,設(shè)備上機(jī)參數(shù)如牽伸倍數(shù)等作為模型輸入,利用案例推理與GA優(yōu)化的SVM預(yù)測模型建立生產(chǎn)工藝優(yōu)化方案。在后續(xù)研究中,模型輸入輸出多樣性逐漸增強(qiáng),相關(guān)研究考慮的影響因素和紗線質(zhì)量指標(biāo)越來越多,預(yù)測模型也更加復(fù)雜。
大量研究開發(fā)了ANN模型與統(tǒng)計(jì)模型同時(shí)預(yù)測紗線質(zhì)量特性,普遍得出了ANN模型結(jié)果更加可靠精確的結(jié)論,主要?dú)w因于ANN模型具有很強(qiáng)的非線性擬合能力[37-43]。BP網(wǎng)絡(luò)(Back propagation networks,BP)及其變化形式是主流的ANN模型,相關(guān)研究驗(yàn)證了其應(yīng)用于紗線質(zhì)量預(yù)測的可行性和有效性[4,44]。為了克服BP網(wǎng)絡(luò)訓(xùn)練效率低、易陷入局部最優(yōu)等不足,紡織學(xué)者們還將徑向基神經(jīng)網(wǎng)絡(luò)(Radial basis function neural network,RBF)應(yīng)用于紗線質(zhì)量預(yù)測。RBF網(wǎng)絡(luò)有極佳的逼近特性,收斂性好、訓(xùn)練速度快且不存在局部最小問題,模型結(jié)構(gòu)具有適應(yīng)性,相比BP網(wǎng)絡(luò)呈
現(xiàn)出較好的預(yù)測精確度和收斂能力[45-46]。然而李翔等[47]在精紡毛紗的條干不勻率和斷裂強(qiáng)度預(yù)測研究中曾指出,在精度要求相同前提下,BP網(wǎng)絡(luò)對于異常樣本的容錯(cuò)能力更強(qiáng)。
ANN模型結(jié)構(gòu)或重要參數(shù)的優(yōu)化也在相關(guān)文獻(xiàn)中被廣泛討論。例如,比較各類訓(xùn)練算法如LM算法[22,48-50]、貝葉斯正則化算法[51]等對于模型預(yù)測結(jié)果的影響,通過試錯(cuò)法或經(jīng)驗(yàn)法判斷最佳隱層數(shù)量和隱層神經(jīng)元個(gè)數(shù)[52-53]。Ghorbani等[54]研究了不同隱層數(shù)量、隱層神經(jīng)元個(gè)數(shù)、訓(xùn)練算法和激活函數(shù)下的ANN模型,隱層數(shù)量和神經(jīng)元個(gè)數(shù)以預(yù)測精度為評判標(biāo)準(zhǔn),后幾項(xiàng)主要看訓(xùn)練速度和消耗內(nèi)存。結(jié)果表明,具有兩個(gè)隱層,每個(gè)隱層有8個(gè)神經(jīng)元,使用LM算法的網(wǎng)絡(luò)預(yù)測紗線毛羽最準(zhǔn)確。
與SVM模型類似,ANN模型結(jié)合群智算法能夠?qū)崿F(xiàn)模型參數(shù)調(diào)優(yōu),使得模型更加快速準(zhǔn)確地收斂,如基于遺傳算法優(yōu)化的BP網(wǎng)絡(luò)可以提高紗線質(zhì)量預(yù)測模型的預(yù)測精度和穩(wěn)定性,其性能優(yōu)于單一BP網(wǎng)絡(luò)[55]。在GA算法優(yōu)化ANN模型的研究中,研究人員還對GA算法的編碼方式、適應(yīng)度函數(shù)的設(shè)計(jì)、遺傳算子機(jī)理等方面進(jìn)行改進(jìn),如思維進(jìn)化算法(Mind evolutionary algorithm,MEA)[56]、免疫遺傳算法(Immune genetic algorithm,IGA)[57-58]與遺傳模擬退火算法(Genetic simulation annealing algorithm,GSAA)[59]等。這些研究主要解決GA迭代冗余、后期無法成熟收斂的問題,也驗(yàn)證了基于混合優(yōu)化算法的紗線質(zhì)量預(yù)測模型比基于GA的模型預(yù)測精度和泛化性能更好。
除了模型參數(shù)的搜索尋優(yōu),群體智能算法還被用于反演紗線原料性能等參數(shù),即通過重要質(zhì)量特性等目標(biāo)建立反演模型,推導(dǎo)出最佳纖維、工藝等參數(shù)組合[60]。由于紗線質(zhì)量指標(biāo)多且可能相互沖突,需要轉(zhuǎn)化為多目標(biāo)優(yōu)化問題獲取最優(yōu)解集。Barzoki等[61-62]建立了兩個(gè)有關(guān)紗線強(qiáng)度和纖維質(zhì)量的優(yōu)化目標(biāo),旨在利用非支配排序遺傳算法(Non-dominated sorting genetic algorithm Ⅱ,NSGA-Ⅱ)制定紗線質(zhì)量較好且成本較低的配棉方案。在此工作基礎(chǔ)上,Chakraborty等[63]采用包含NSGA-Ⅱ在內(nèi)的四種群智算法對紗線質(zhì)量特性進(jìn)行多目標(biāo)優(yōu)化,粒子群算法的優(yōu)化性能優(yōu)于其他算法。這些研究注重優(yōu)化目標(biāo)的建立和優(yōu)化算法的選取,獲得最優(yōu)解集之后,涉及為企業(yè)提供指導(dǎo)的決策問題較少被提到。
相關(guān)研究嘗試了一些新型優(yōu)化算法,如灰狼算法和帝王蝶算法等。Hadavandi等[64-65]將灰狼算法作為賽洛紡紗線強(qiáng)度預(yù)測模型的權(quán)值優(yōu)化器,與其他三種基于傳統(tǒng)群智算法的模型相比具有更高的預(yù)測精度,并且根據(jù)類似研究得出帝王蝶算法同樣有效。目前新型優(yōu)化算法在紡紗理論研究中應(yīng)用相對較少,其適用性和通用性尚有待深入。
目前深度學(xué)習(xí)模型在紡紗質(zhì)量相關(guān)研究中應(yīng)用相對較少,其主要原因在于樣本數(shù)據(jù)規(guī)模、類型有限。胡臻龍等[66-67]基于卷積神經(jīng)網(wǎng)絡(luò)(Convolutional neural networks,CNN)建立深度神經(jīng)網(wǎng)絡(luò)預(yù)測模型,將纖維、設(shè)備和工藝參數(shù)等多個(gè)特征參數(shù)作為模型輸入,其預(yù)測誤差相比淺層ANN和MLR均控制在1%以內(nèi)。在后續(xù)研究中,考慮到紡紗生產(chǎn)前紡工序?qū)τ诩喚€最終質(zhì)量的影響,又基于長短期記憶網(wǎng)絡(luò)(Long short-term memory,LSTM)建立了考慮紡紗時(shí)序性的深度預(yù)測模型,如圖4所示,所考慮的時(shí)序性體現(xiàn)每個(gè)LSTM單元對應(yīng)紡紗生產(chǎn)不同工序的設(shè)備參數(shù),前紡工序與后紡之間存在的關(guān)聯(lián)及對最終紗線質(zhì)量的影響。其結(jié)果表明,LSTM模型在動態(tài)工序數(shù)據(jù)集上的預(yù)測精度比不考慮時(shí)序性的人工智能模型高。
大量紡紗理論研究集中在多個(gè)特征參數(shù)與紗線質(zhì)量之間的非線性逼近,沒有根據(jù)工業(yè)數(shù)據(jù)的動態(tài)時(shí)序性特點(diǎn)考慮有關(guān)質(zhì)量的態(tài)勢預(yù)測問題。在同一批原料的前提下,紡紗加工過程產(chǎn)生的數(shù)據(jù)隨時(shí)間變化,如機(jī)器部件的速度等特征參數(shù),而這些數(shù)據(jù)的實(shí)時(shí)性可能映射出更加準(zhǔn)確的紗線質(zhì)量,使得生產(chǎn)模擬更加真實(shí)。利用流行的時(shí)序分析方法,如LSTM、門控循環(huán)單元網(wǎng)絡(luò)(Gated recurrent unit,GRU)等建立多工序關(guān)聯(lián)的人工智能模型,學(xué)習(xí)工序之間、時(shí)序之間的相關(guān)性,預(yù)測未來的質(zhì)量態(tài)勢,還可延伸至工藝優(yōu)化、質(zhì)量異常追蹤和在線質(zhì)量檢測等領(lǐng)域。
目前的紡紗質(zhì)量預(yù)測研究主要應(yīng)用于試紡和工藝優(yōu)化兩方面。在基于數(shù)據(jù)驅(qū)動的智能建模技術(shù)體系中,結(jié)合群體智
能算法建立的混合智能模型比單一模型具有更大的靈活性,是目前提高預(yù)測性能的普遍做法[27]。
2.2 基于知識驅(qū)動的方法
模糊方法建立在模糊規(guī)則之上,模糊推理通過制定模糊邏輯將給定輸入映射為輸出,相對于人工神經(jīng)網(wǎng)絡(luò)來說適用于不精確、模糊的或不完備的數(shù)據(jù)建模,不需要大量示例數(shù)據(jù)來訓(xùn)練模型。最常用的兩種模糊推理算法為Mamdani型和Sugeno型[68]。
研究人員將模糊方法應(yīng)用到紡紗質(zhì)量預(yù)測領(lǐng)域,通過模糊規(guī)則提取和模糊系統(tǒng)結(jié)構(gòu)優(yōu)化等方面,主要解決如何提高模型精度、保持算法效率,同時(shí)兼顧完備性和魯棒性及提高可解釋性等問題。
在針對模糊規(guī)則優(yōu)化和模糊系統(tǒng)結(jié)構(gòu)參數(shù)優(yōu)化方面,主要通過群智算法和神經(jīng)網(wǎng)絡(luò)這兩種優(yōu)化技術(shù)以混合方式實(shí)現(xiàn)。如遺傳算法可以從數(shù)值數(shù)據(jù)中自動生成模糊規(guī)則[27],還可以進(jìn)行隸屬函數(shù)參數(shù)調(diào)優(yōu)[69],提高模糊系統(tǒng)預(yù)測精度。神經(jīng)網(wǎng)絡(luò)結(jié)合模糊系統(tǒng)同時(shí)具有易于表達(dá)和自適應(yīng)學(xué)習(xí)的能力,其中自適應(yīng)模糊神經(jīng)推理系統(tǒng)(Adaptive neuro fuzzy inference system,ANFIS)在與SVM、ANN的對比中體現(xiàn)出了學(xué)習(xí)能力強(qiáng)和預(yù)測精度高的優(yōu)點(diǎn)[70-72]。
前述解決問題的總體思想是利用數(shù)據(jù)驅(qū)動方式強(qiáng)大的尋優(yōu)能力實(shí)現(xiàn)知識驅(qū)動方法中結(jié)構(gòu)或參數(shù)的優(yōu)化,即數(shù)據(jù)調(diào)優(yōu)的知識驅(qū)動方法,旨在減少依靠專家制定的知識規(guī)則。知識驅(qū)動方法與數(shù)據(jù)驅(qū)動方法的協(xié)同驅(qū)動可以體現(xiàn)為數(shù)據(jù)調(diào)優(yōu)的知識驅(qū)動,知識增強(qiáng)的數(shù)據(jù)驅(qū)動及并聯(lián)或串聯(lián)的互補(bǔ)結(jié)合[73]。在紡紗質(zhì)量預(yù)測研究中,神經(jīng)模糊系統(tǒng)是知識與數(shù)據(jù)的互補(bǔ)并聯(lián)結(jié)合。除此之外,Ghanmi等[74-75]利用了一種級聯(lián)模式下的神經(jīng)模糊混合模型來預(yù)測全局質(zhì)量指標(biāo),根據(jù)烏斯特統(tǒng)計(jì)數(shù)據(jù)建立幾種紗線特性與整體質(zhì)量指數(shù)的規(guī)則,將ANN預(yù)測的多種紗線質(zhì)量特性作為第二階段模糊專家系統(tǒng)的輸入,預(yù)測輸出綜合質(zhì)量指標(biāo)。這個(gè)兩階段預(yù)測方法也體現(xiàn)了知識與數(shù)據(jù)的串聯(lián)互補(bǔ)結(jié)合思想。
研究人員利用各種數(shù)據(jù)驅(qū)動方式致力于降低專家經(jīng)驗(yàn)的依賴性,但并不能完全由數(shù)據(jù)策略代替現(xiàn)有規(guī)則。知識驅(qū)動方法可解釋性強(qiáng)、執(zhí)行效率高,但存在難以自學(xué)習(xí)、獲取知識困難等缺點(diǎn);數(shù)據(jù)驅(qū)動方法通用性強(qiáng)、可持續(xù)學(xué)習(xí),但需要具備高質(zhì)量數(shù)據(jù)、強(qiáng)大算力等基本條件??梢娭R驅(qū)動方法與數(shù)據(jù)驅(qū)動方法各有優(yōu)勢和不足,將數(shù)據(jù)驅(qū)動方法與知識驅(qū)動方法相結(jié)合,如利用數(shù)據(jù)驅(qū)動方法的預(yù)測分析,結(jié)合知識驅(qū)動方法的智能決策等,有望為紡紗領(lǐng)域研究提供更多思路。綜上所述,近年來相關(guān)研究中典型人工智能預(yù)測方法的對比如表1所示。
2.3 模型評估
預(yù)測模型評估指標(biāo)主要包括:
1) 預(yù)測精度。通過比較實(shí)測值與預(yù)測值的一致性或近似誤差來評估模型的預(yù)測精度[55],常用的統(tǒng)計(jì)學(xué)指標(biāo)有相關(guān)系數(shù)(R)、決定系數(shù)(R2)、平均絕對誤差、均方誤差、均方根誤差等。
2) 泛化能力(穩(wěn)定性)。利用交叉驗(yàn)證或適當(dāng)分配數(shù)據(jù)集方式多次測試,以檢驗(yàn)?zāi)P偷姆夯芰Α?/p>
3) 收斂速度。一般以算法整體響應(yīng)時(shí)間為主。
3 共性問題及發(fā)展趨勢
如何建立一個(gè)合理有效的紡紗質(zhì)量預(yù)測模型一直是紡織領(lǐng)域的一個(gè)熱點(diǎn)、難點(diǎn)問題?;诤A考徏啍?shù)據(jù)集成處理、現(xiàn)有模型適用性較低及各加工工序之間的強(qiáng)相關(guān)性無法被有效表達(dá)等視角,本文對紡紗質(zhì)量預(yù)測的共性問題及發(fā)展趨勢進(jìn)行了總結(jié)與展望。
1) 由于對分散的紡紗生產(chǎn)數(shù)據(jù)缺乏集成管理,紡紗質(zhì)量預(yù)測缺少高效高可用的底層數(shù)據(jù)支撐,同時(shí)預(yù)測算法和模型也不具備面向海量數(shù)據(jù)的分析能力。建立基于Hadoop技術(shù)的紡紗質(zhì)量預(yù)測系統(tǒng),融合集成全流程紡紗生產(chǎn)數(shù)據(jù),利用大數(shù)據(jù)關(guān)聯(lián)分析技術(shù)挖掘影響紡紗質(zhì)量的關(guān)鍵因素,同時(shí)提出適應(yīng)海量數(shù)據(jù)的紡紗質(zhì)量預(yù)測模型,提高模型在海量數(shù)據(jù)環(huán)境下的穩(wěn)定性和精確度。
2) 以往研究方法聚焦于多種特征參數(shù)與多個(gè)紗線質(zhì)量特性的非線性逼近問題,沒有考慮利用具有時(shí)序性的生產(chǎn)過程數(shù)據(jù)進(jìn)行未來的態(tài)勢預(yù)測。基于LSTM、GRU等時(shí)序分析方法建立多工序關(guān)聯(lián)的人工智能模型,提前發(fā)現(xiàn)質(zhì)量異常,及時(shí)調(diào)整方案,維護(hù)生產(chǎn)過程。
3) 從企業(yè)角度出發(fā),人工智能方法應(yīng)用在紡紗質(zhì)量預(yù)測領(lǐng)域不能只關(guān)注制造過程建模,更重要的是智能決策?;诩徏喼圃旃に嚨念I(lǐng)域知識與預(yù)測結(jié)果,綜合成本、質(zhì)量、資源等各種優(yōu)化目標(biāo),利用群智算法、模糊方法及先進(jìn)的深度強(qiáng)化學(xué)習(xí)等決策技術(shù),為企業(yè)提供決策支持,實(shí)現(xiàn)制造過程模擬、優(yōu)化和決策的生產(chǎn)閉環(huán)。
4) 數(shù)據(jù)驅(qū)動方法與知識驅(qū)動方法各有優(yōu)勢和不足,任何一方都不能被完全替代,知識與數(shù)據(jù)的融合驅(qū)動才是未來人工智能方法在紡織領(lǐng)域走向?qū)嶓w化的關(guān)鍵。利用知識與數(shù)據(jù)驅(qū)動方法的互補(bǔ)結(jié)合方式,或是其他先進(jìn)的協(xié)同算法,以知識引導(dǎo)數(shù)據(jù)模型的產(chǎn)生,再基于數(shù)據(jù)模型生成新知識,形成知識與數(shù)據(jù)的交替迭代。
4 結(jié) 語
本文以紡紗質(zhì)量預(yù)測建模流程為導(dǎo)向,對人工智能技術(shù)在紡紗質(zhì)量預(yù)測領(lǐng)域中的發(fā)展現(xiàn)狀進(jìn)行了研究,分析了Hadoop技術(shù)作為紡紗質(zhì)量預(yù)測建模分析平臺的優(yōu)勢,重點(diǎn)分析對比了智能建模方法在紡紗質(zhì)量預(yù)測領(lǐng)域的應(yīng)用,最后基于平臺、算法和模型角度總結(jié)了紡紗質(zhì)量預(yù)測研究的共性問題及發(fā)展趨勢。
參考文獻(xiàn):
[1]董奎勇, 于偉東. 基于BP神經(jīng)網(wǎng)絡(luò)的紡紗質(zhì)量預(yù)報(bào)模型[J]. 東華大學(xué)學(xué)報(bào)(自然科學(xué)版), 2005(2): 88-92.
DONG Kuiyong, YU Weidong. Yarn property prediction models based on BP neural network[J]. Journal of Donghua University (Natural Science), 2005(2): 88-92.
[2]HE Z L, XU J, TRAN K P, et al. Modeling of textile manufacturing processes using intelligent techniques: A review[J]. International Journal of Advanced Manufacturing Technology, 2021, 116(1/2): 39-67.
[3]GURUPRASAD R, BEHERA B K. Soft computing in textiles[J]. Indian Journal of Fibre & Textile Research, 2010, 35(1): 75-84.
[4]EL-GEIHEINI A, ELKATEB S, ABD-ELHAMIED M R. Yarn tensile properties modeling using artificial intelligence[J]. Alexandria Engineering Journal, 2020, 59(6): 4435-4440.
[5]宇宏達(dá), 吳麗莉, 陳廷. 群體智能算法在紡織領(lǐng)域的應(yīng)用[J]. 紡織導(dǎo)報(bào), 2021(1): 93-96.
YU Hongda, WU Lili, CHEN Ting. Applications of swarm intelligence algorithm in textiles[J]. China Textile Leader, 2021(1): 93-96.
[6]《工業(yè)物聯(lián)網(wǎng)互聯(lián)互通白皮書》在無錫發(fā)布[J]. 模具工業(yè), 2018, 44(10): 10-11.
Industrial Internet of Things Connectivity White Paper was released in Wuxi[J]. Die & Mould Industry, 2018, 44(10): 10-11.
[7]汪松松, 彭來湖, 戴寧, 等. 基于工業(yè)互聯(lián)網(wǎng)的針織機(jī)械互聯(lián)互通結(jié)構(gòu)研究[J]. 紡織學(xué)報(bào), 2020, 41(1): 165-173.
WANG Songsong, PENG Laihu, DAI Ning, et al. Research on knitting machinery interconnection and interoperability structure based on industrial internet[J]. Journal of Textile Research, 2020, 41(1): 165-173.
[8]NAIDU D, THAKUR A. A systematic literature review of big data and the hadoop frameworks[J]. International Journal of Early Childhood Special Education, 2022, 14(2): 2969-2973.
[9]程隆棣, 張潔, 張紅霞, 等. 棉紡智能化紡紗關(guān)鍵技術(shù)芻議[J]. 紡織導(dǎo)報(bào), 2021(6): 50-53.
CHENG Longdi, ZHANG Jie, ZHANG Hongxia, et al. Discussion on the key technology of intelligent cotton spinning[J]. China Textile Leader, 2021(6): 50-53.
[10]邵景峰, 賀興時(shí), 王進(jìn)富, 等. 大數(shù)據(jù)環(huán)境下的紡織制造執(zhí)行系統(tǒng)設(shè)計(jì)[J]. 機(jī)械工程學(xué)報(bào), 2015, 51(5): 160-170.
SHAO Jingfeng, HE Xingshi, WANG Jinfu, et al. Design of textile manufacturing execution system based on big data[J]. Journal of Mechanical Engineering, 2015, 51(5): 160-170.
[11]葛陳鵬, 季承, 糜娜, 等. 基于制造執(zhí)行系統(tǒng)的紡紗企業(yè)智能化轉(zhuǎn)型探討[J]. 棉紡織技術(shù), 2021, 49(10): 42-46.
GE Chenpeng, JI Cheng, MI Na, et al. Intelligent transformation discussion of spinning enterprise based on manufacturing execution system[J]. Cotton Textile Technology, 2021, 49(10): 42-46.
[12]馮立增, 宋久祥, 魏敬典, 等. 基于HBase的雙數(shù)據(jù)庫在紡織信息化平臺的改造與實(shí)現(xiàn)[J]. 工業(yè)儀表與自動化裝置, 2020(2): 123-127.
FENG Lizeng, SONG Jiuxiang, WEI Jingdian, et al. Transformation and implementation of double database based on HBase in textile information platform[J]. Industrial Instrumentation & Automation, 2020(2): 123-127.
[13]劉佩全. 紡織行業(yè)數(shù)據(jù)處理過程中大數(shù)據(jù)技術(shù)的選擇與應(yīng)用[J]. 紡織導(dǎo)報(bào), 2017(3): 90-91.
LIU Peiquan. Application of big data technology in textile industry for data processing[J]. China Textile Leader, 2017(3): 90-91.
[14]鄭健. 基于大數(shù)據(jù)的棉紡織設(shè)備數(shù)據(jù)管理系統(tǒng)開發(fā)[J]. 上海紡織科技, 2015, 43(6): 77-80.
ZHENG Jian. Study on cotton textile equipment data management system development based on big data[J]. Shanghai Textile Science & Technology, 2015, 43(6): 77-80.
[15]張潔, 呂佑龍, 汪俊亮, 等. 大數(shù)據(jù)驅(qū)動的紡織智能制造平臺架構(gòu)[J]. 紡織學(xué)報(bào), 2017, 38(10): 159-165.
ZHANG Jie, L Youlong, WANG Junliang, et al. Big-data-driven framework for intelligent textile manufacturing[J]. Journal of Textile Research, 2017, 38(10): 159-165.
[16]SASAKI Y. A survey on IoT big data analytic systems: Current and future[J]. IEEE Internet of Things Journal, 2022, 9(2): 1024-1036.
[17]JIANG H, SONG J X, ZHANG B W, et al. Prediction of yarn unevenness based on BMNN[J]. Journal of Engineered Fibers and Fabrics, 2021, 16: 15589250211037978.
[18]DORAN E C, SAHIN C. The prediction of quality characteristics of cotton/elastane core yarn using artificial neural networks and support vector machines[J]. Textile Research Journal, 2020, 90(13/14): 1558-1580.
[19]FAULKNER W B, HEQUET E F, WANJURA J, et al. Relationships of cotton fiber properties to ring-spun yarn quality on selected High Plains cottons[J]. Textile Research Journal, 2012, 82(4): 400-414.
[20]MOGHASSEM A R, FALLAHPOUR A R. Processing parameters optimization of draw frame for Rotor spun yarn strength using gene expression programming (GEP)[J]. Fibers and Polymers, 2011, 12(7): 970-975.
[21]DEMIRYUREK O, KOC E. The mechanism and/or prediction of the breaking elongation of polyester/viscose blended open-end rotor spun yarns[J]. Fibers and Polymers, 2009, 10(5): 694-702.
[22]SU T L, KUO C F J, WEI H J. Prediction system of open-end rotor spinning process based on LM-neural network for bamboo charcoal fibers[J]. Fibers and Polymers, 2011, 12(5): 657-663.
[23]曹繼鵬, 張志丹, 張明光, 等. 生條中短纖維含量與成紗指標(biāo)的相關(guān)性分析[J]. 紡織學(xué)報(bào), 2018, 39(12): 30-36.
CAO Jipeng, ZHANG Zhidan, ZHANG Mingguang, et al. Correlation analysis on relationship between short fiber content of card sliver and yarn parameters[J]. Journal of Textile Research, 2018, 39(12): 30-36.
[24]YIN X G, YU W D. The virtual manufacturing model of the worsted yarn based on artificial neural networks and grey theory[J]. Applied Mathematics and Computation, 2007, 185(1): 322-332.
[25]張羽彤, 沈卓爾, 代利花, 等. 小規(guī)模樣本精梳毛紡紗線質(zhì)量預(yù)測[J]. 毛紡科技, 2020, 48(12): 1-5.
ZHANG Yutong, SHEN Zuoer, DAI Lihua, et al. Prediction of worsted yarn quality with small-scale samples[J]. Wool Textile Journal, 2020, 48(12): 1-5.
[26]KHAN Z, LIM A E K, WANG L J, et al. An artificial neural network-based hairiness prediction model for worsted wool yarns[J]. Textile Research Journal, 2009, 79(8): 714-720.
[27]SHAHRABI J, HADAVANDI E, SOLTANI P. Developing an intelligent fiber migration simulator in spun yarns using a genetic fuzzy system[J]. Fibers and Polymers, 2013, 14(5): 844-853.
[28]YILDIRIM P, BIRANT D, ALPYILDIZ T. Data mining and machine learning in textile industry[J]. WIREs Data Mining and Knowledge Discovery, 2018, 8(1): e1228.
[29]MOZAFARY V, PAYVANDY P. Application of data mining technique in predicting worsted spun yarn quality[J]. Journal of the Textile Institute, 2014, 105(1): 100-108.
[30]邵景峰, 賀興時(shí), 王進(jìn)富, 等. 基于海量數(shù)據(jù)的紡紗質(zhì)量異常因素識別方法[J]. 計(jì)算機(jī)集成制造系統(tǒng), 2015, 21(10): 2644-2652.
SHAO Jingfeng, HE Xingshi, WANG Jinfu, et al. Identification method for abnormal factors of spinning quality based on massive data[J]. Computer Integrated Manufacturing Systems, 2015, 21(10): 2644-2652.
[31]GHOSH A, CHATTERJEE P. Prediction of cotton yarn properties using support vector machine[J]. Fibers and Polymers, 2010, 11(1): 84-88.
[32]項(xiàng)前, 楊建國, 程隆棣. 基于支持向量機(jī)的紗線質(zhì)量預(yù)測[J]. 紡織學(xué)報(bào), 2008(4): 43-46.
XIANG Qian, YANG Jianguo, CHENG Longdi. Yarn quality prediction by support vector machine method[J]. Journal of Textile Research, 2008(4): 43-46.
[33]谷有眾, 高衛(wèi)東, 盧雨正, 等. 應(yīng)用遺傳算法優(yōu)化支持向量回歸機(jī)的噴氣渦流紡紗線質(zhì)量預(yù)測[J]. 紡織學(xué)報(bào), 2016, 37(7): 142-148.
GU Youzhong, GAO Weidong, LU Yuzheng, et al. Prediction of vortex yarn properties based on hybrid genetic algorithm and support vector regression[J]. Journal of Textile Research, 2016, 37(7): 142-148.
[34]ABAKAR K A A, YU C W. Performance of SVM based on PUK kernel in comparison to SVM based on RBF kernel in prediction of yarn tenacity[J]. Indian Journal of Fibre and Textile Research, 2014, 39(1): 55-59.
[35]呂志軍, 楊建國, 項(xiàng)前. 基于遺傳算法參數(shù)優(yōu)化的紗線質(zhì)量預(yù)測技術(shù)[J]. 東華大學(xué)學(xué)報(bào)(自然科學(xué)版), 2012, 38(5): 519-523.
L Zhijun, YANG Jianguo, XIANG Qian. GA based parameters optimization on prediction method of yarn quality[J]. Journal of Donghua University (Natural Science), 2012, 38(5): 519-523.
[36]宋楚平, 蔡彬彬, 李少芹. 基于SVM的紗線生產(chǎn)工藝智能優(yōu)化研究[J]. 毛紡科技, 2015, 43(7): 13-17.
SONG Chuping, CAI Binbin, LI Shaoqin. Intelligent optimization study of yarn production processes based on SVM[J]. Wool Textile Journal, 2015, 43(7): 13-17.
[37]AZIMI B, TEHRAN M A, MOJTAHEDI M R M. Prediction of false twist textured yarn properties by artificial neural network methodology[J]. Journal of Engineered Fibers and Fabrics, 2013, 8(3): 97-101.
[38]GHAREHAGHAJI A A, SHANBEH M, PALHANG M. Analysis of two modeling methodologies for predicting the tensile properties of cotton-covered nylon core yarns[J]. Textile Research Journal, 2007, 77(8): 565-571.
[39]HAGHIGHAT E, JOHARI M S, ETRATI S M, et al. Study of the hairiness of polyester-viscose blended yarns. Part III: Predicting yarn hairiness using an artificial neural network[J]. Fibres and Textiles in Eastern Europe, 2012, 20(1): 33-38.
[40]ALMETWALLY A A, IDREES H M F, HEBEISH A A. Predicting the tensile properties of cotton/spandex core-spun yarns using artificial neural network and linear regression models[J]. Journal of the Textile Institute, 2014, 105(11): 1221-1229.
[41]OZKANZKAN I, KUVVETLI Y, BAYKAL P D, et al. Comparison of the neural network model and linear regression model for predicting the intermingled yarn breaking strength and elongation[J]. Journal of the Textile Institute, 2014, 105(11): 1203-1211.
[42]DEMIRYUREK O, KOC E. Predicting the unevenness of polyester/viscose blended open-end rotor spun yarns using artificial neural network and statistical models[J]. Fibers and Polymers, 2009, 10(2): 237-245.
[43]ABD-ELLATIF S A M. Optimizing sliver quality using Artificial Neural Networks in ring spinning[J]. Alexandria Engineering Journal, 2013, 52(4): 637-642.
[44]王雯, 費(fèi)樹岷, 鄭波, 等. 基于BP網(wǎng)絡(luò)的成紗質(zhì)量預(yù)測的數(shù)據(jù)挖掘[J]. 上海紡織科技, 2006(12): 8-10.
WANG Wen, FEI Shumin, ZHENG Bo, et al. Data mining of yarn quality based on BP neural network[J]. Shanghai Textile Science & Technology, 2006(12): 8-10.
[45]袁利華. 基于HVI指標(biāo)與RBF神經(jīng)網(wǎng)絡(luò)的紗線質(zhì)量預(yù)測[J]. 絲綢, 2014, 51(11): 17-21.
YUAN Lihua. Yarn quality prediction based on HVI indicator and RBF neural network[J]. Journal of Silk, 2014, 51(11): 17-21.
[46]TURHAN Y, TOPRAKCI O. Comparison of high-volume instrument and advanced fiber information systems based on prediction performance of yarn properties using a radial basis function neural network[J]. Textile Research Journal, 2013, 83(2): 130-147.
[47]李翔, 彭志勤, 金鳳英, 等. 基于神經(jīng)網(wǎng)絡(luò)的精紡毛紗性能預(yù)測模型比較[J]. 紡織學(xué)報(bào), 2011, 32(3): 51-56.
LI Xiang, PENG Zhiqin, JIN Fengying, et al. Comparing prediction models for worsted yarn performances based on neutral networks[J]. Journal of Textile Research, 2011, 32(3): 51-56.
[48]ERBIL Y, BABAARSLAN O, ILHAN I. A comparative prediction for tensile properties of ternary blended open-end rotor yarns using regression and neural network models[J]. Journal of the Textile Institute, 2018, 109(4): 560-568.
[49]AHMAD G G. Using artificial neural networks with graphical user interface to predict the strength of carded cotton yarns[J]. Journal of the Textile Institute, 2016, 107(3): 386-394.
[50]FAROOQ A, CHERIF C. Development of prediction system using artificial neural networks for the optimization of spinning process[J]. Fibers and Polymers, 2012, 13(2): 253-257.
[51]MALIK S A, FAROOQ A, GEREKE T, et al. Prediction of blended yarn evenness and tensile properties by using Artificial Neural Network and multiple linear regression[J]. Autex Research Journal, 2016, 16(2): 43-50.
[52]查劉根, 謝春萍. 應(yīng)用四層BP神經(jīng)網(wǎng)絡(luò)的棉紗成紗質(zhì)量預(yù)測[J]. 紡織學(xué)報(bào), 2019, 40(1): 52-56.
ZHA Liugen, XIE Chunping. Prediction of cotton yarn quality based on four-layer BP neural network[J]. Journal of Textile Research, 2019, 40(1): 52-56.
[53]李惠軍, 朱磊. 基于BP人工神經(jīng)網(wǎng)絡(luò)的紗線毛羽預(yù)測研究[J]. 棉紡織技術(shù), 2011, 39(1): 32-34.
LI Huijun, ZHU Lei. Research of yarn hairiness forecast based on BP Artificial Neural Network[J]. Cotton Textile Technology, 2011, 39(1): 32-34.
[54]GHORBANI V, VADOOD M, JOHARI M S. Prediction of polyester/cotton blended rotor-spun yams hairiness based on the machine parameters[J]. Indian Journal of Fibre & Textile Research, 2016, 41(1): 19-25.
[55]劉彬, 項(xiàng)前, 楊建國, 等. 基于遺傳神經(jīng)網(wǎng)絡(luò)的紗線質(zhì)量預(yù)測[J]. 東華大學(xué)學(xué)報(bào)(自然科學(xué)版), 2013, 39(4): 504-508.
LIU Bin, XIANG Qian, YANG Jianguo, et al. Combining the genetic algorithm with Artificial Neural Networks for yarn quality forecasting[J]. Journal of Donghua University (Natural Science), 2013, 39(4): 504-508.
[56]吳志剛, 方璐, 郁崇文. 基于思維進(jìn)化神經(jīng)網(wǎng)絡(luò)的細(xì)紗斷裂強(qiáng)度預(yù)測[J]. 棉紡織技術(shù), 2020, 48(2): 6-9.
WU Zhigang, FANG Lu, YU Chongwen. Prediction of spun yarn breaking tenacity based on mind evolutionary neural network[J]. Cotton Textile Technology, 2020, 48(2): 6-9.
[57]查劉根, 謝春萍. 免疫遺傳神經(jīng)網(wǎng)絡(luò)算法的紗線原料性能反演[J]. 棉紡織技術(shù), 2019, 47(2): 18-23.
ZHA Liugen, XIE Chunping. Property inversion of yarn raw material for immune genetic neural network algorithm[J]. Cotton Textile Technology, 2019, 47(2): 18-23.
[58]查劉根, 謝春萍. 基于免疫遺傳算法的BP神經(jīng)網(wǎng)絡(luò)在紗線條干預(yù)測上的應(yīng)用[J]. 絲綢, 2019, 56(2): 19-26.
ZHA Liugen, XIE Chunping. Application of BP neural network in yarn evenness prediction based on immune genetic algorithm[J]. Journal of Silk, 2019, 56(2): 19-26.
[59]谷有眾, 劉建立, 高衛(wèi)東. 基于遺傳模擬退火神經(jīng)網(wǎng)絡(luò)的紗線強(qiáng)力預(yù)測[J]. 上海紡織科技, 2016, 44(1): 45-49.
GU Youzhong, LIU Jianli, GAO Weidong. Prediction of yarn strength based on genetic simulated annealing BP neural network[J]. Shanghai Textile Science & Technology, 2016, 44(1): 45-49.
[60]DAS S, GHOSH A, MAJUMDAR A, et al. Yarn engineering using hybrid artificial neural network-genetic algorithm model[J]. Fibers and Polymers, 2013, 14(7): 1220-1226.
[61]BARZOKI P K, VADOOD M, JOHARI M S. Multi-objective optimization of rotorcraft compact spinning core-spun yarn properties[J]. Journal of Textiles and Polymers, 2018, 6(1): 47-53.
[62]GHOSH A, DAS S, BANERJEE D. Multi objective optimization of yarn quality and fibre quality using evolutionary algorithm[J]. Journal of The Institution of Engineers (India): Series E, 2013, 94(1): 15-21.
[63]CHAKRABORTY S, DIYALEY S. Multi-objective optimization of yarn characteristics using evolutionary algorithms: A comparative study[J]. Journal of The Institution of Engineers (India): Series E, 2018, 99(2): 129-140.
[64]HADAVANDI E, MOSTAFAYI S, SOLTANI P. A Grey Wolf Optimizer-based neural network coupled with response surface method for modeling the strength of siro-spun yarn in spinning mills[J]. Applied Soft Computing, 2018, 72: 1-13.
[65]SOLTANI P, HADAVANDI E. A monarch butterfly optimization-based neural network simulator for prediction of siro-spun yarn tenacity[J]. Soft Computing, 2019, 23(20): 10521-10535.
[66]HU Z L, ZHAO Q, WANG J. The prediction model of cotton yarn intensity based on the CNN-BP neural network[J]. Wireless Personal Communications, 2018, 102(2): 1905-1916.
[67]HU Z L, ZHAO Q, WANG J. The prediction model of worsted yarn quality based on CNN-GRNN neural network[J]. Neural Computing and Applications, 2019, 31(9): 4551-4562.
[68]HAGHIGHAT E, JOHARI M S, ETRATI S M, et al. Study of the hairiness of polyester-viscose blended yarns. Part IV: Predicting yarn hairiness using fuzzy logic[J]. Fibres & Textiles in Eastern Europe, 2012, 20(3): 39-42.
[69]VADOOD M, BARZOKI P K, JOHARI M S. Multi objective optimization of rotorcraft compact spinning system using fuzzy-genetic model[J]. Journal of the Textile Institute, 2017, 108(12): 2166-2172.
[70]NURWAHA D, WANG X H. Prediction of rotor spun yarn strength using support vector machines method[J]. Fibers and Polymers, 2011, 12(4): 546-549.
[71]ADMUTHE L S, APTE S. Adaptive neuro-fuzzy inference system with subtractive clustering: A model to predict fiber and yarn relationship[J]. Textile Research Journal, 2010, 80(9): 841-846.
[72]MAJUMDAR A, CIOCOIU M, BLAGA M. Modelling of ring yarn unevenness by soft computing approach[J]. Fibers and Polymers, 2008, 9(2): 210-216.
[73]蒲志強(qiáng), 易建強(qiáng), 劉振, 等. 知識和數(shù)據(jù)協(xié)同驅(qū)動的群體智能決策方法研究綜述[J]. 自動化學(xué)報(bào), 2022, 48(3): 627-643.
PU Zhiqiang, YI Jianqiang, LIU Zhen, et al. Knowledge-based and data-driven integrating methodologies for collective intelligence decision making: A survey[J]. Acta Automatica Sinica, 2022, 48(3): 627-643.
[74]GHANMI H, GHITH A, BENAMEUR T. Prediction of rotor-spun yarn quality using hybrid artificial neural network-fuzzy expert system model[J]. Indian Journal of Fibre & Textile Research, 2019, 44(1): 31-38.
[75]GHANMI H, GHITH A, BENAMEUR T. Ring spun yarn quality prediction using hybrid neural networks[J]. Journal of the Textile Institute, 2023, 114(1): 66-74.
Research progress on yarn quality prediction based on artificial intelligence technology
ZHAO Weiling1,3,4, ZHANG Junhui2,3,4, CHEN Mingliang1,3,4, LI Qing1,3, CHEN Dapeng1,3
(1.School of Integrated Circuits, University of Chinese Academy of Sciences, Beijing 100049, China; 2.School of Electrical and AutomaticEngineering, Changshu Institute of Technology, Suzhou 215500, China; 3.Wuxi IoT Innovation Center Co., Ltd.,Wuxi 214135, China; 4.Institute of Microelectronic Technology of Kunshan, Suzhou 215347, China)
Abstract:
The yarn production process is a complex multi-step process, and the yarn quality is affected by raw material properties, process parameters and equipment parameters. The spinning mills attempt to predict and control the yarn quality in advance through the spinning data, so they can adjust the production process parameters in time according to the individual needs of customers, and achieve the goal of reducing raw material waste, improving yarn quality and even reducing costs. With the development of artificial intelligence technologies such as big data and intelligent modeling, artificial intelligence technology has been gradually applied in the spinning industry. Aiming at the problem of yarn quality prediction, researchers have conducted a lot of research on platform framework, algorithms and models, accelerating the application of artificial intelligence technology in the spinning industry.
Most of the statistical correlation methods such as simple mathematical models and multiple linear regression have certain idealized assumptions, which are strongly dependent on production experience and involve obvious subjective factors. At present, researchers are committed to the application of artificial intelligence methods in the field of spinning, and propose a variety of spinning quality prediction models based on neural networks. Compared with the traditional method, the self-learning and adaptive ability of neural networks can quickly learn the nonlinear relationship between fiber parameters, process and equipment parameters and yarn quality indicators. The model has high prediction accuracy and generalization ability. Aiming at the parameter optimization problem of neural networks, the related research combines the improved swarm intelligence algorithm to realize model parameter tuning, accelerate model convergence and improve prediction accuracy. Aiming at the problem of spinning small sample modeling, related research uses support vector machine combined with swarm intelligence algorithm to propose a small sample modeling method with strong adaptability. Related research also applies the swarm intelligence algorithm to the inversion of yarn raw material parameters, and obtains the optimal solution set by converting it into a multi-objective optimization problem. There are also some researches devoted to the design of computing platform framework for yarn quality prediction modeling, which mainly uses Hadoop technology to provide reliable and efficient underlying technical support for yarn quality prediction process. In addition, the fuzzy method is another important method in the research of yarn quality prediction. The related research has gone through simple fuzzy logic to the combination of swarm intelligence algorithm, neural network and fuzzy system, which preliminarily reflects the fusion idea of the knowledge-driven method and the data-driven method, and provides more ideas for the research of the spinning field.
At present, although artificial intelligence technology has accumulated a lot of achievements in the field of yarn quality prediction, there are still some common problems in the existing research, which are mainly reflected in the fact that the model does not have the ability to adapt to massive data, and that the research lacks sequential situation prediction problems and intelligent decision-making. Therefore, researchers still need to carry out in-depth research on these problems and continue to tap the application potential of artificial intelligence technology in the field of yarn quality prediction.
Key words:
yarn quality prediction; artificial intelligence; Hadoop technology; hybrid intelligent model; objective optimization; development trend
收稿日期:
2022-07-15;
修回日期:
2023-03-03
基金項(xiàng)目:
江蘇省博士后科研資助計(jì)劃項(xiàng)目(2020Z411)
作者簡介:
趙薇玲(1998),女,碩士研究生,研究方向?yàn)槿斯ぶ悄?、工業(yè)大數(shù)據(jù)建模。通信作者:章軍輝,博士,zhangjunhui@ime.ac.cn。