周紅旭 孫海軍 張雷 王華英
摘要:為了解決懸臂式掘進(jìn)機(jī)當(dāng)截割部被機(jī)身遮擋或粉塵比較嚴(yán)重時引發(fā)的視覺定位失效問題,以磁場強(qiáng)度分量和雙目立體視覺技術(shù)獲取的位姿數(shù)據(jù)作為訓(xùn)練數(shù)據(jù),獲得網(wǎng)絡(luò)參數(shù),提出一種基于一維卷積神經(jīng)網(wǎng)絡(luò)(1D-CNN)的輔助定位方法。結(jié)果表明,1D-CNN對截割部軌跡預(yù)測較好,空間角度俯仰角、偏航角的預(yù)測精度達(dá)到99%以上,總體精度滿足懸臂式掘進(jìn)機(jī)位姿的測量要求。所提方法可以有效預(yù)測掘進(jìn)機(jī)截割部的空間位姿信息,與BP全連接神經(jīng)網(wǎng)絡(luò)相比,具有能自動提取特征、避免過擬合的優(yōu)點(diǎn),為掘進(jìn)機(jī)截割部定位提出了新思路。
關(guān)鍵詞:采礦工程其他學(xué)科;懸臂式掘進(jìn)機(jī);卷積神經(jīng)網(wǎng)絡(luò);磁場定位;位姿測量
中圖分類號:TD421文獻(xiàn)標(biāo)識碼:A
DOI:10.7535/hbkd.2022yx03002
Magnetic field aided positioning technology of roadheader cutting part based on one-dimensional convolution neural network
ZHOU Hongxu SUN Haijun ZHANG Lei WANG Huaying
(1.School of Mathematics and Physics Science and Engineering,Hebei University of Engineering,Handan,Hebei 056038,China;2.Hebei Computational Optical Imaging and Photoelectric Detection Technology Innovation Center,Handan,Hebei 056038,China)
Abstract:In order to solve the problem of visual positioning failure caused by the contilever roadheader when the cutting part is blocked by the fuselage or the dust is serious,this paper proposed an auxiliary positioning method based on one-dimensional convolutional neural network (1D-CNN).The network parameters were obtained by taking the intensity component of the magnetic field and the pose data obtained by binocular stereo vision technology as training data.The experimental results show that the 1D-CNN can predict the trajectory of the cutting part better,and the prediction accuracy of the pitch angle and yaw angle of the space angle is more than 99%.This method can effectively predict the spatial pose information of the cutting part of the roadheader.Compared with the BP fully connected neural network,it has the advantages of automatic feature extraction and avoiding overfitting,and puts forward a new idea for the positioning of the cutting part of the roadheader.
Keywords: other disciplines of mining engineering;boom-type roadheader;convolutional neural network;magnetic field positioning;pose measurement
掘進(jìn)機(jī)是井下煤礦掘進(jìn)施工的核心設(shè)備。隨著煤炭開采設(shè)備向智能化、無人化方向的發(fā)展,對掘進(jìn)機(jī)自動化水平提出了更高要求 [1-3]。掘進(jìn)機(jī)實(shí)時位姿測量技術(shù)是未來煤礦智能化的關(guān)鍵技術(shù)之一[4]。國內(nèi)外眾多研究人員對掘進(jìn)機(jī)位姿檢測進(jìn)行了大量研究,其中具有代表性的工作包括基于雙目立體視覺的掘進(jìn)機(jī)位姿測量技術(shù)[5]、基于全站儀的掘進(jìn)機(jī)機(jī)身位姿參數(shù)測量技術(shù)[6]、基于超寬帶技術(shù)的掘進(jìn)機(jī)自主定位定向技術(shù)[7-8]、慣導(dǎo)傳感器掘進(jìn)機(jī)位姿測量技術(shù)[9-10]以及基于iGPS掘進(jìn)機(jī)位姿測量技術(shù)[11-12]。這些技術(shù)雖然采用了現(xiàn)代先進(jìn)的物理方法和信息技術(shù),但由于煤礦井下掘進(jìn)機(jī)工作環(huán)境惡劣,導(dǎo)致掘進(jìn)機(jī)工作面存在高濕、高粉塵、低照度、振動等問題?;谝曈X定位方法采用圖像處理技術(shù)對掘進(jìn)機(jī)的位姿信息進(jìn)行解算,能夠非接觸、實(shí)時地得到掘進(jìn)機(jī)的位姿信息[13-17]。然而,當(dāng)截割部被機(jī)身遮擋或粉塵嚴(yán)重時,這種定位方法會失效。磁場定位技術(shù)正好可以彌補(bǔ)這種不足,利用傳感器的靈敏度檢測磁源和傳感器之間相對位置的變化,具有緊湊、低成本、節(jié)能和安全等優(yōu)點(diǎn)。
傳統(tǒng)磁場定位技術(shù)采用的是磁偶極子(MD)模型[18-22]進(jìn)行定位,但由于存在鐵磁物質(zhì),在井下使用時常會受到復(fù)雜磁場的嚴(yán)重干擾,使傳統(tǒng)定位模型失效。為此。本文采用一維卷積神經(jīng)網(wǎng)絡(luò)(1D-CNN)算法解決視覺測量定位失效下掘進(jìn)機(jī)截割部的位姿測量問題,通過在掘進(jìn)機(jī)懸臂上固定強(qiáng)磁鐵,利用高精度磁場傳感器,獲得空間磁場強(qiáng)度信息,同時利用深度學(xué)習(xí)算法,計(jì)算磁鐵位置坐標(biāo),得出掘進(jìn)機(jī)截割部的位姿信息。
11D-CNN磁場輔助定位技術(shù)
1.1一維卷積神經(jīng)網(wǎng)絡(luò)簡介
卷積神經(jīng)網(wǎng)絡(luò)(CNN)是多層感知器(MLP)網(wǎng)絡(luò)的一種變體,于1980年首次被使用[23],自推出以來其在檢測和分類等問題上有著良好表現(xiàn)且得到了廣泛應(yīng)用。如圖1所示,一個基本的卷積神經(jīng)網(wǎng)絡(luò)由輸入層、多個隱藏層(卷積、歸一化、池化的重復(fù))以及全連接和輸出層組成。卷積層主要提取相關(guān)特征,減少參數(shù)數(shù)量;池化層進(jìn)一步減少通過二次采樣的特征參數(shù),簡化網(wǎng)絡(luò),并保留盡可能多的有效信息;全連接層主要連接所有特征值,輸出最終值[24-25]。
1D-CNN是一種前饋神經(jīng)網(wǎng)絡(luò),屬于經(jīng)典的深度神經(jīng)網(wǎng)絡(luò)[26]。其輸入為一維數(shù)據(jù),卷積核也相應(yīng)采用一維結(jié)構(gòu),每個卷積層和池化層的輸出也為一維特征矢量。與其他機(jī)器學(xué)習(xí)模型相比,基于一維卷積神經(jīng)網(wǎng)絡(luò)的滾動軸承自適應(yīng)故障診斷算法具有強(qiáng)大的抗噪能力,被廣泛用于各種場景下的時間序列建模。卷積神經(jīng)網(wǎng)絡(luò)通常用于圖像識別和目標(biāo)檢測及分類等[27]。CNN結(jié)構(gòu)獨(dú)特,具有良好的處理網(wǎng)絡(luò)結(jié)構(gòu)特征數(shù)據(jù)的能力,可以有效解決煤礦井下磁場復(fù)雜、非線性等因素造成的數(shù)據(jù)處理困難等問題[28]。
1.2位姿估計(jì)模型架構(gòu)
本文構(gòu)建了如圖2所示的1D-CNN位姿估計(jì)模型,input_1和input_2為輸入層,訓(xùn)練數(shù)據(jù)經(jīng)過3層一維卷積神經(jīng)網(wǎng)絡(luò)卷積,每層卷積都經(jīng)過批標(biāo)準(zhǔn)化(batch normalization,BN)使其加快收斂、防止過擬合。其中3層卷積層的一維卷積核大小均為3,卷積核的數(shù)量依次為16,32和64,模型使用Adam方法訓(xùn)練參數(shù)。使用SeLU激活函數(shù)后輸入下一卷積層,不斷改善訓(xùn)練參數(shù),訓(xùn)練時設(shè)置batch_size=64,epoch=1 000。由于特征較少,因而無需采用池化層,添加平展層將張量重新整形為矢量,數(shù)據(jù)經(jīng)卷積后引入激活函數(shù)連接輸出層,得到懸臂式掘進(jìn)機(jī)截割部的位姿估計(jì)。將樣本數(shù)據(jù)輸入1D-CNN模型進(jìn)行訓(xùn)練,并利用Huber Loss損失函數(shù)方法對模型進(jìn)行參數(shù)優(yōu)化,導(dǎo)入測試集檢驗(yàn)?zāi)P偷念A(yù)測效果。經(jīng)過不斷優(yōu)化改進(jìn),得到理想的1D-CNN位姿估計(jì)模型。
為驗(yàn)證構(gòu)建的1D-CNN模型,采用如圖3所示的BP全連接神經(jīng)網(wǎng)絡(luò)位姿估計(jì)模型,對掘進(jìn)機(jī)截割部預(yù)測結(jié)果進(jìn)行對比,input_1和input_2為輸入層,其中2組磁場數(shù)據(jù)在輸入前同樣進(jìn)行標(biāo)準(zhǔn)化,對數(shù)據(jù)進(jìn)行某種統(tǒng)一處理。2組數(shù)據(jù)經(jīng)過2層全連接網(wǎng)絡(luò)進(jìn)行傳遞,神經(jīng)元數(shù)分別為128和256。之后,同樣添加平展層將張量重新整形為矢量,將網(wǎng)絡(luò)層連接進(jìn)行特征聯(lián)合,之后再經(jīng)過2層全連接網(wǎng)絡(luò),其神經(jīng)元數(shù)分別為256和3。為了與1D-CNN測量結(jié)果進(jìn)行對比與分析,將BP全連接神經(jīng)模型同樣采用Adam方法訓(xùn)練參數(shù),使用SeLU激活函數(shù),不斷改進(jìn)訓(xùn)練參數(shù),訓(xùn)練時設(shè)置batch_size=64,epoch=1 000,同樣采用Huber Loss損失函數(shù)方法對模型進(jìn)行參數(shù)優(yōu)化。
1.3性能指標(biāo)評價
2模擬礦井測量實(shí)驗(yàn)
2.1實(shí)驗(yàn)環(huán)境
為了驗(yàn)證本文方法在實(shí)際應(yīng)用中的效果,對掘進(jìn)機(jī)截割部位進(jìn)行了定位。實(shí)驗(yàn)地點(diǎn)選擇在峰峰集團(tuán)有限公司中澳培訓(xùn)基地的模擬礦井,掘進(jìn)機(jī)機(jī)型為EBZ-135型,巷道斷面4.5 m×2.8 m,2組平行掘進(jìn)方向的跟頭支護(hù)梁中心距為1.5 m,將2組傳感器分別固定在支護(hù)梁中心線距前端2 m處。將磁源點(diǎn)固定在截割臂上部,磁源點(diǎn)中心沿截割部中軸線距離截割頭中心10 cm處,采用3D打印技術(shù)制作磁源隔熱和消振保護(hù)裝置。采用實(shí)驗(yàn)掘進(jìn)機(jī)移動和運(yùn)轉(zhuǎn)三維空間的磁源點(diǎn)移動軌跡,模擬礦井生產(chǎn)現(xiàn)場溫度、濕度、振動、淋水及粉塵等環(huán)境。
2.2數(shù)據(jù)采集與處理
懸臂式掘進(jìn)機(jī)截割部位姿測量實(shí)驗(yàn)?zāi)M圖如圖4所示,掘進(jìn)機(jī)截割部位姿數(shù)據(jù)來源于雙目立體視覺測量方案。根據(jù)掘進(jìn)機(jī)的結(jié)構(gòu)組成與運(yùn)動特點(diǎn)構(gòu)建視覺位姿測量結(jié)構(gòu),在掘進(jìn)機(jī)機(jī)身安裝立體視覺相機(jī)和紅外識別標(biāo)志物,建立掘進(jìn)機(jī)位姿解算數(shù)學(xué)模型,解算出巷道坐標(biāo)系下機(jī)身與截割部位姿信息,采集到的部分雙目圖像如圖5所示。在機(jī)臂安裝無源永磁體,在機(jī)身左右兩側(cè)安裝高精度三軸數(shù)字磁場計(jì)采集磁場信號數(shù)據(jù)。通過模擬截割部切割煤層的運(yùn)動,共采集2 228組磁場強(qiáng)度信號與其對應(yīng)的紅外測量位姿,選取80%的數(shù)據(jù)作為網(wǎng)絡(luò)訓(xùn)練集,剩余的20%數(shù)據(jù)作為模型測試用,其中將掘進(jìn)機(jī)截割部的位置坐標(biāo)、俯仰角、偏航角、翻滾角作為網(wǎng)絡(luò)輸出量,將三軸數(shù)字磁場計(jì)采集的磁場強(qiáng)度和方向作為輸入量。由于采集到的數(shù)據(jù)處于不同的強(qiáng)度范圍,因而需要將磁場強(qiáng)度信號按式(5)進(jìn)行歸一化處理:
3結(jié)果與分析
掘進(jìn)機(jī)截割部位置測試結(jié)果如圖6所示,俯仰角、偏航角、翻滾角測試結(jié)果如圖7—圖9所示。
圖6同時給出了基于雙目立體視覺的紅外定位軌跡與基于1D-CNN的磁場定位軌跡,圖中藍(lán)色實(shí)線為紅外定位結(jié)果,橙色空心圓為磁場定位結(jié)果。由圖6可見,磁場定位與紅外定位結(jié)果一致性非常高,表明對于位置測量來說,本文提出的基于1D-CNN網(wǎng)絡(luò)模型的磁場定位方法是可行的。
圖7 a)、圖8 a)、圖9 a)和圖7 b)、圖8 b)、圖9 b)分別為利用1D-CNN磁場定位方法和利用BP神經(jīng)網(wǎng)絡(luò)磁場定位方法得到的掘進(jìn)機(jī)俯仰角、偏航角、翻滾角的測試結(jié)果。作為比較,圖中同時給出了基于紅外雙目立體視覺的定位結(jié)果,以藍(lán)色圓點(diǎn)表示,并將其視為真實(shí)值,基于神經(jīng)網(wǎng)絡(luò)的磁場定位模型測量結(jié)果用橙色三角表示??梢钥闯觯诰W(wǎng)絡(luò)模型的磁場定位測量值與基于紅外立體視覺的定位結(jié)果吻合較為一致。經(jīng)計(jì)算,基于1D-CNN的磁場定位方法獲得的掘進(jìn)機(jī)截割部俯仰角、偏航角、翻滾角測量精度分別為99.24%,99.59%和91.74%,基于BP神經(jīng)網(wǎng)絡(luò)磁場定位方法得到的掘進(jìn)機(jī)截割部俯仰角、偏航角、翻滾角的測量精度分別為98.90%,99.33%和89.69%??梢姡途蜻M(jìn)機(jī)截割部姿態(tài)角測量而言,基于1D-CNN的磁場定位方法比基于BP神經(jīng)網(wǎng)絡(luò)的磁場定位方法測量精度高。
為了進(jìn)一步評價本文算法的有效性,采用多種評價指標(biāo)進(jìn)行定量分析,比較1D-CNN與BP神經(jīng)網(wǎng)絡(luò)預(yù)測模型對掘進(jìn)機(jī)截割部3個空間角度(俯仰角、偏航角、翻滾角)測量的優(yōu)劣,對比結(jié)果見表1—表3。
從表1—表3可以看出,與BP網(wǎng)絡(luò)模型相比,1D-CNN模型截割部俯仰角評價度量指標(biāo)MAE,MAPE和RMSE分別降低了0.096 1,0.004 5和0.151 9,準(zhǔn)確率R提高了0.003 4;截割部偏航角評價度量指標(biāo)MAE和RMSE分別降低了0.275 1和0.333 4,準(zhǔn)確率R提高了0.002 6,但MAPE提高了0.030 6;截割部翻滾角度量指標(biāo)MAE,MAPE和RMSE分別降低了0.090 0,5.135 0和0.111 5,準(zhǔn)確率R提高了0.020 5。1D-CNN模型算法在一定程度上提高了預(yù)測性能,具有更高的擬合度、更高的精度和估測能力。
通過比較1D-CNN和BP神經(jīng)網(wǎng)絡(luò)算法俯仰角、偏航角、翻滾角的平均絕對誤差、平均絕對百分誤差、均方根誤差及預(yù)測準(zhǔn)確率,可以看出1D-CNN算法在各項(xiàng)性能指標(biāo)方面的表現(xiàn)都很優(yōu)秀,且與基于紅外雙目立體視覺的定位結(jié)果相吻合。因此,本文所提出的基于一維卷積神經(jīng)網(wǎng)絡(luò)的磁場輔助定位方法對懸臂式掘進(jìn)機(jī)截割部定位精確、有效,具有良好的實(shí)用性。
4結(jié)語
1)本文提出了一種基于一維卷積神經(jīng)網(wǎng)絡(luò)的煤礦井下復(fù)雜磁場環(huán)境下的掘進(jìn)機(jī)截割部位姿測量方法,并與紅外雙目立體視覺定位方法和典型BP神經(jīng)網(wǎng)絡(luò)預(yù)測結(jié)果進(jìn)行了對比。由實(shí)驗(yàn)結(jié)果可知:在掘進(jìn)機(jī)安裝無源永磁體及在左右機(jī)身安裝三軸數(shù)字磁場計(jì)測量磁場信號,通過構(gòu)架的神經(jīng)網(wǎng)絡(luò)模型訓(xùn)練磁場信號數(shù)據(jù),可有效預(yù)測掘進(jìn)機(jī)截割部的位姿變化,為視覺定位掘進(jìn)機(jī)位姿預(yù)測提供了一種新的輔助定位方法。
2)采用本文構(gòu)建的一維卷積神經(jīng)網(wǎng)絡(luò)模型對掘進(jìn)機(jī)截割部位姿進(jìn)行預(yù)測,預(yù)測值與實(shí)測值吻合度較高,MAE,MAPE,RMSE和R2這4項(xiàng)預(yù)測模型指標(biāo)的結(jié)果比BP神經(jīng)網(wǎng)絡(luò)要好,預(yù)測性能較常規(guī)方法也有一定的提高,說明本文所提出的輔助定位方法是可行的。
3)為進(jìn)一步提高基于一維卷積神經(jīng)網(wǎng)絡(luò)懸臂式掘進(jìn)機(jī)磁場輔助定位方法的可行性,未來應(yīng)該從機(jī)身模塊化入手,將磁源與機(jī)臂進(jìn)行融合,在實(shí)際生產(chǎn)中實(shí)現(xiàn)穩(wěn)定高效運(yùn)行。
參考文獻(xiàn)/References:
[1]周開平.煤礦井下懸臂式掘進(jìn)機(jī)回轉(zhuǎn)機(jī)構(gòu)優(yōu)化設(shè)計(jì)[J].煤炭技術(shù),2021,40(9):177-180.ZHOU Kaiping.Optimization design of slewing mechanism of cantilever roadheader in coal mine[J].Coal Technology,2021,40(9):177-180.
[2]趙亦輝,趙友軍,周展.綜采工作面采煤機(jī)智能化技術(shù)研究現(xiàn)狀[J].工礦自動化,2022,48(2):11-18.ZHAO Yihui,ZHAO Youjun,ZHOU Zhan.Research status of intelligent technology of shearer in fully mechanized working face[J].Industry and Mine Automation,2022,48(2):11-18.
[3]王國法.煤礦智能化最新技術(shù)進(jìn)展與問題探討[J].煤炭科學(xué)技術(shù),2022,50(1):1-27.WANG Guofa.New technological progress of coal mine intelligence and its problems[J].Coal Science and Technology,2022,50(1):1-27.
[4]焦承堯,王永強(qiáng),張幸福,等.煤礦智能化綜采裝備系統(tǒng)及其可靠性分析[J].煤炭工程,2022,54(2):166-171.JIAO Chengyao,WANG Yongqiang,ZHANG Xingfu,et al.Research on intelligent fully mechanized coal mining equipment system and its reliability[J].Coal Engineering,2022,54(2):166-171.
[5]陳慎金,成龍,王鵬江,等.基于掘進(jìn)機(jī)位姿測量系統(tǒng)的自主標(biāo)定方法誤差分析[J].煤炭學(xué)報(bào),2018,43(9):2647-2652.CHEN Shenjin,CHENG Long,WANG Pengjiang,etal.Error analysis of autonomous calibration strategy based on roadheader attitude measurement system[J].Journal of China Coal Society,2018,43(9):2647-2652.
[6]朱信平,李睿,高娟,等.基于全站儀的掘進(jìn)機(jī)機(jī)身位姿參數(shù)測量方法[J].煤炭工程,2011(6):113-115.
[7]符世琛,李一鳴,楊健健,等.基于超寬帶技術(shù)的掘進(jìn)機(jī)自主定位定向方法研究[J].煤炭學(xué)報(bào),2015,40(11):2603-2610.FU Shichen,LI Yiming,YANG Jianjian,et al.Research on autonomous positioning and orientation method of roadheader based on ultra wide-band technology[J].Journal of China Coal Society,2015,40(11):2603-2610.
[8]FU Shichen,LI Yiming,ZONG Kai,etal.Ultra-wideband pose detection method based on TDOA positioning model for boom-type road- header[J].AEU-International Journal of Electronics and Communications,2019,99:70-80.
[9]張旭輝,劉博興,張超,等.掘進(jìn)機(jī)全站儀與捷聯(lián)慣導(dǎo)組合定位方法[J].工礦自動化,2020,46(9):1-7.ZHANG Xuhui,LIU Boxing,ZHANG Chao,et al.Roadheader positioning method combining total station and strapdown inertial navigation system[J].Industry and Mine Automation,2020,46(9):1-7.
[10]陶云飛,楊健健,李嘉賡,等.基于慣性導(dǎo)航技術(shù)的掘進(jìn)機(jī)位姿測量系統(tǒng)研究[J].煤炭技術(shù),2017,36(1):235-237.TAO Yunfei,YANG Jianjian,LI Jiageng,et al.Research on position and orientation measurement system of heading machine based on inertial navigation technology[J].Coal Technology,2017,36(1):235-237.
[11]陶云飛,宗凱,張敏駿,等.基于iGPS的掘進(jìn)機(jī)單站多點(diǎn)分時機(jī)身位姿測量方法[J].煤炭學(xué)報(bào),2015,40(11):2611-2616.TAO Yunfei,ZONG Kai,ZHANG Minjun,et al.Aposition and orientationmeasurement method of single-station,multipoint and time-sharing forroadheaderbody based on iGPS[J].Journal of China Coal Society,2015,40(11):2611-2616.
[12]賈文浩,陶云飛,張敏駿,等.基于iGPS的煤巷狹長空間中掘進(jìn)機(jī)絕對定位精度研究[J].儀器儀表學(xué)報(bào),2016,37(8):1920-1926.JIA Wenhao,TAO Yunfei,ZHANG Minjun,et al.Research on absolute positioning accuracy of roadheader based on indoor global posi-tioning system in narrow and long coal tunnel[J].Chinese Journal of Scientific Instrument,2016,37(8):1920-1926.
[13]DU Yuxin,Tong Minming,LIU Ting,et al.Visual measurement system for roadheaders pose detection in mines[J].Optical Engineering,2016,55(10):104107.
[14]楊文娟,張旭輝,馬宏偉,等.懸臂式掘進(jìn)機(jī)機(jī)身及截割頭位姿視覺測量系統(tǒng)研究[J].煤炭科學(xué)技術(shù),2019,47(6):50-57.YANG Wenjuan,ZHANG Xuhui,MA Hongwei,et al.Research on position and posture measurement system of body and cutting head for boom-type roadheader based on machine vision[J].Coal Science and Technology,2019,47(6):50-57.
[15]張超,張旭輝,杜昱陽,等.基于雙目視覺的懸臂式掘進(jìn)機(jī)位姿測量技術(shù)研究[J].煤炭科學(xué)技術(shù),2021,49(11):225-235.ZHANG Chao,ZHANG Xuhui,DU Yuyang,et al.Measuring technique of cantilever roadheader position based on binocular stereo vision[J].Coal Science and Technology,2021,49(11):225-235.
[16]YANG Wenjuan,ZHANG Xuhui,MA Hongwei,et al.Laserbeams-based localization methods for boom-type roadheader using underground camera non-uniform blur model[J].IEEE Access,2020,8:190327-190341.
[17]劉勇.基于機(jī)器視覺的掘進(jìn)機(jī)截割頭姿態(tài)檢測系統(tǒng)[J].煤礦機(jī)械,2020,41(3):187-189.LIU Yong.Attitude detection system of cutting head of roadheader based on machine vision[J].Coal Mine Machinery,2020,41(3):187-189.
[18]單珊,周穗華,戴忠華,等.強(qiáng)跟蹤漸進(jìn)更新擴(kuò)展卡爾曼濾波器及其在磁偶極子跟蹤中的應(yīng)用[J].海軍工程大學(xué)學(xué)報(bào),2022,34(1):105-112.SHAN Shan,ZHOU Suihua,DAI Zhonghua,et al.Strong tracking progressive extended Kalman filter and its application in magnetic dipole tracking[J].Journal of Naval University of Engineering,2022,34(1):105-112.
[19]楊勇,崔麗珍,郭倩倩,等.煤礦井下基于Attention機(jī)制的LSTM地磁定位算法[J].測繪通報(bào),2022(1):72-78.YANG Yong,CUI Lizhen,GUO Qianqian,et al.LSTM geomagnetic positioning algorithm based on attention mechanism in underground coal mine[J].Bulletin of Surveying and Mapping,2022(1):72-78.
[20]WU Faye,ROBERT N M,F(xiàn)REY D D,et al.Enhanced magnetic localization with artificial neural network field models[C]//2013 IEEE International Conference on Robotics and Automation.Karlsruhe,Germany:IEEE,2013:1560-1565.
[21]CHEN Shengzhi,ZHU Minghua,ZHANG Qing,et al.Accurate magnetic object localization using artificial neural network[C]//2019 15th International Conference on Mobile Adhoc and Sensor Networks (MSN).Shenzhen:IEEE,2019:25-30.
[22]HU Guanghui,WAN Hong,LI Xinxin.Ahigh-precision magnetic-assisted heading angle calculation method based on a 1D convolutional neural network (CNN) in a complicated magnetic environment[J].Micromachines,2020,11(7):642-648.
[23]彭成,李鳳娟,蔣金元.基于深度學(xué)習(xí)的故障診斷與預(yù)測方法綜述[J].現(xiàn)代電子技術(shù),2022,45(3):111-120.PENG Cheng,LI Fengjuan,JIANG Jinyuan.Summary of fault diagnosis and prediction methods based on deep learning[J].Modern Electronics Technique,2022,45(3):111-120.
[24]LI Wenjun,CUI Zhiwen,JIN Minjun,et al.Dynamic temperature measurement with a dual-thermocouple sensor based on a dual-head one-dimensional convolutional neural network[J].Measurement,2021,182:109679.
[25]CUI Hua,BAI Jie.A new hyperparameters optimization method for convolutional neural networks[J].Pattern Recognition Letters,2019,125:828-834.
[26][ZK(]周飛燕,金林鵬,董軍.卷積神經(jīng)網(wǎng)絡(luò)研究綜述[J].計(jì)算機(jī)學(xué)報(bào),2017,40(6):1229-1251.ZHOU Feiyan,JIN Linpeng,DONG Jun.Review of convolutional neural network[J].Chinese Journal of Computers,2017,40(6):1229-1251.
[27]FUKUSHIMA K.Neocognitron:A selforganizing neural network model for a mechanism of pattern recognition unaffected by shift in position[J].Biological Cybernetics,1980,36(4):193-202.
[28]LIANG Jiefeng,JING Tianjun,NIU Huanna,et al.Two-terminal fault location method of distribution network based on adaptive convolution neural network[J].IEEE Access,2020,8:54035-54043.