Development of a coal quality analyzer for application to power plants bad on lar-induced breakdown spectroscopy ☆
Lei Zhang a ,b ,⁎,Yao Gong a ,Yufang Li a ,Xin Wang a ,Juanjuan Fan a ,Lei Dong a ,Weiguang Ma a ,Wangbao Yin a ,b ,⁎,Suotang Jia a ,b
a State Key Laboratory of Quantum Optics and Quantum Optics Devices,Institute of Lar Spectroscopy,Shanxi University,Taiyuan 030006,China b
Collaborative Innovation Center of Extreme Optics,Shanxi University,Taiyuan 030006,China
a b s t r a c t
a r t i c l e i n f o Article history:
Received 1March 2015
Accepted 26September 2015Available online 1October 2015
快门时间Keywords:
Lar-induced breakdown spectroscopy (LIBS)Coal quality analysis
Support vector regression (SVR)Principal component analysis (PCA)Puld lar energy stabilization
It is vitally important for a power plant to determine the coal property rapidly to optimize the combustion pro-cess.In this work,a fully software-controlled lar-induced breakdown spectroscopy (LIBS)bad coal quality analyzer comprising a LIBS apparatus,a sampling equipment,and a control module,has been designed for pos-sible application to power plants for offering rapid and preci coal quality analysis results.A clod-loop feed-back puld lar energy stabilization technology is propod to stabilize the Nd:YAG lar output energy to a pret interval by using the detected lar energy signal so as to enhance the measurement stability and applied in a month-long monitoring experiment.The results show that the lar energy stability has been greatly reduced from ±5.2%to ±1.3%.In order to indicate the complex relationship between the concentrations of the analyte of interest and the corresponding plasma spectra,the support vector regression (SVR)is employed as a non-linear regression method.It is shown that this SVR method combined with principal component a
nalysis (PCA)enables a signi ficant improvement in cross-validation accuracy by using the calibration t of coal samples.The root mean square error for prediction of ash content,volatile matter content,and calori fic value decreas from 2.74%to 1.82%,1.69%to 1.22%,and 1.23MJ/kg to 0.85MJ/kg,respectively.Meanwhile,the corresponding average relative error of the predicted samples is reduced from 8.3%to 5.48%,5.83%to 4.42%,and 5.4%to 3.68%,respectively.The enhanced levels of accuracy obtained with the SVR combined with PCA bad calibration models open up ave-nues for prospective prediction in coal properties.
©2015Elvier B.V.All rights rerved.
1.Introduction
Coal plays an extremely important role in power generation and this role is t to continue.It currently fuels 40%of the electricity worldwide,and this proportion is expected to remain at similar levels over the next 30years.However,for coal-fired power plants,the coal quality impacts not only the combustion ef ficiency,but also the pollutant emissions.Furthermore,it experiences frequent changes due to inherent varia-tions in coal composition,switching of coal supplier and coal blending.So it is advisable that all coals to be fired in a boiler should be analyzed.Conventional coal quality analysis i
s performed by using equipments such as thermo-gravimetric analyzer (TGA).This method for coal sam-ples usually need sample-dissolving,is time consuming,and cannot guide the production timely.The prompt gamma neutron activation analysis (PGNAA)is proved to be an effective on-line coal quality
analysis technique,but it has limitations for extensive application be-cau of expensive price,large size,radiation hazard,and strict regula-tory requirements.Therefore,new analysis methods are required urgently in power industry to obtain better reprentative coal quality rapidly and safely.
Lar-induced breakdown spectroscopy (LIBS)is an emerging ana-lytical spectroscopy technique which can determine the elemental com-position from the line emission of lar generated plasma on the basis of elemental and molecular emission intensities.This technology demon-strates numerous appealing features that distinguish it from conven-tional analytical spectrochemical techniques,such as rapid analysis,no radiation risk,little or no need for sample preparation,simultaneous multi-element analysis,simple apparatus,little sample consumption,etc.Thus,LIBS is chon as the most promising candidate technique that maybe suitable for coal quality analysis under the harsh industrial environment.Over the past few decades,LIBS has already been employed as an effective tool to determine the elemental concentra-
tions in coal [1].Body et al.developed an advanced LIBS instrument for analysis of presd coal,and the absolute measurement repeatability and accuracy for the key inorganic components (such as Al,Si,Mg,Ca,Fe,Na,and K)were typically within ±10%[2].Gaft et al.developed an
Spectrochimica Acta Part B 113(2015)167–173
雷雨剧本☆Selected papers prented at the 8th International Conference on Lar-Induced Breakdown Spectroscopy (LIBS)in Beijing,China,8th-12th September,2014.⁎Corresponding authors.
E-mail address:k1226@ (L.Zhang),ywb65@ (W.
Yin).
/10.1016/j.sab.2015.09.021怎么折战斗机
0584-8547/©2015Elvier B.V.All rights
rerved.
Contents lists available at ScienceDirect
Spectrochimica Acta Part B
j o u r na l h om e p a g e :w w w.e l s e v i e r.c o m /l o c a t e /s a b
industrial LIBS machine for quantitative analysis of the ash content in coal and the standard deviatio
n (SD)for ash content analysis was 0.21%[3].Deguchi et al.developed an automated LIBS unit and applied it in one-month long monitoring experiment,and the result showed that both Na and K reached the ppb level [4].Wallis ployed LIBS to perform chemical analysis of low-ash lignite,of which the five inorganic components were shown to be reproducible between sample pellets at a 95%con fidence level [5].Mateo et al.prented the capabil-ity of LIBS for process control in a thermal power plant through quanti-tative compositional characterization of the coal ud for combustion [6].Yao et al.directly correlated the spectra with ash content by multi-variate analysis method rather than using the LIBS quanti fied metal in-formation then converting to ash content [7].Lu et al.established a principal component analysis (PCA)model to directly determine vol-atile matter content from LIBS spectra with venteen samples for cali-bration and three samples for prediction [8].Yuan et al.propod a partial least squares (PLS)model bad on dominant factor for coal ele-mental concentration measurement using LIBS [9].However,there still leave veral key issues to be resolved,including the long-term measurement stability,the measurement accuracy,etc.Long-term measurement stability is an important criterion for judging the reliabil-ity and repeatability of apparatus,and determines the recalibration and maintenance period.There are many factors that in fluence the long-term measurement stability of LIBS.Factors such as the CCD wavelength and nsitivity shift caud by thermal effect,lens contamination,vibra-tion induced micro-
displacement,and ambient humidity,could be overcome by using temperature control,high pressure gas cleaning,vi-bration isolation,and in flatable als,respectively.However,the output power of the Q-switched puld lar that acts as the excitation source in LIBS,will drift slowly due to the aging of xenon lamp and power sup-ply,deliquescence of Q-switched crystal,and thermal deformation,and result in incread measurement error and shorter maintenance period.Therefore,how to stabilize the output energy of the puld lar so as to suppress the slow drift and ensure a long-term measurement stability,is a key issue to be resolved in LIBS apparatus.Another key issue with regard to the measurement accuracy is to minimize the in fluence that might be caud by factors such as fluctuations in uncontrollable exper-imental parameters and the physical and chemical matrix effect.Hence,how to u a reasonable data analysis method to improve the measure-ment accuracy remains as a major bottleneck for practical application.
In this study,we apply this LIBS technique to the development of a new and versatile analyzer for rapid,accurate,and long-term coal qual-ity (including the volatile matter,the calori fic value and the ash)analy-sis in a coal-fired power plant.A clod-loop feedback puld lar energy stabilization device is propod to stabilize the Nd:YAG lar output energy to a pret interval by using the detected lar energy
signal.Support vector regression (SVR)[10,11]combined with PCA [12]is employed to establish calibration models to realize more accu-rate measurements.2.Experimental 2.1.LIBS apparatus
The LIBS-bad coal quality analyzer is an integrated and fully soft-ware controlled analysis apparatus,which can be divided into three portions:the optical system,the analysis chamber,and the control module.Details of the portions are described as follows.
The con figuration and the schematic of the fully aled optical sys-tem are shown in Fig.1(a)and (b),respectively.The beam of a Nd:YAG puld lar operates at 1064nm and delivers 70mJ in 8ns pul duration at a repetition rate of 10Hz.The lar beam pass through a half-wave plate (HWP)and is divided into two beams by a polarizing beam splitter (PBS).The intensity ratio of the two beams can be adjust-ed by rotating the HWP conveniently.The combination of HWP and PBS not only permits the lar source works at the most stable condition with a single pul energy of 70mJ,but also provides the optimal abla-tion energy of 60mJ/pul for LIBS analysis.The re flected lar (~10mJ/pul)travels to an energy meter for continuous power monitoring.The transmitted light is re flected by a 45°silver-covered mirror and pass through a beam expander (BE)vertically.This beam expander is utilized to reduce both the divergence angle and the diffraction effects,resulting in better focusing effect.Then the light is focud at the sample surface using a 15mm diameter and 20
0mm focal length plano-convex quartz lens.The focal point is located 7.5mm below the sample surface in order to avoid air breakdown above the sample surface.A tiny portion of the sample is ablated to form the plasma,and the plasma spectra are collect-ed by an optical fiber and detected by a dual-channel spectrometer that covers spectral ranges of 190–366nm and 586–1116nm with a typical spectral resolution of 0.7nm.In this investigation,the acquisition time and delay time were of 10ms and 2μs,respectively.Since the temper-ature of the charge coupled device (CCD)within the spectrometer is ris-ing continuously during the scan time and results in increasing wavelength shift (0.15pixel/°C)and intensity change (1%/°C),a simple chiller made by coiling copper tubing around the spectrometer shell and connected with an external cooling water circulation system was devel-oped to stabilized the CCD temperature at 20°C within 0.1°C.In order to prevent the window from contamination,an automatic valve was placed in front of the measurement head by shutting this valve when any abnormality occurs to affect the LIBS.Additionally,shock absorbers and an air conditioner were employed to enable the optical system to work outdoors under extreme
environment.
(a)(b)
PM HWP
Mirror BE Lar
PBS
Lens Valve
Optical Fiber
Lens
Plasma
Fig.1.Optical system of the LIBS Unit:(a)the con figuration,(b)the schematic.
168L.Zhang et al./Spectrochimica Acta Part B 113(2015)167–173
A schematic reprentation of the analysis chamber is shown in Fig.2(a).It mainly includes a sample stage,a turntable and a jet pump.Here,the sample stage has a spring inside which can guarantee the same distance between the focusing lens and the sample surface of different thickness.In order to make each lar pul have a fresh ac-tion point and achieve uniform sample of points,the sample is placed on the stage that is driven by a stepping motor and the rotation speed is t appropriately according to the single measurement time.The jet pump that is driven by the high speed compresd air is utilized to create a continuous negative pressure in the nozzle near the plasma,so as to suck up all the aerosols that generated during the LIBS analysis.It is no-table that the air flow rate should keep constant to avoid any incread signal fluctuation.The discharge outlet path rves as a dust cleaner for removing the aerosols excited by the powerful lar puls from the lo-cation above the lar focus spot quickly so as to prevent the accumula-tion of aerosol deposition on the optical lens.Moreover,during the LIBS measurement,the aerosols can cau a great attenuation of lar en-ergy via the classical collisional absorption mechanism as well as a re-duction in the fiber collection ef ficiency [13].
The control module can control the entire system involving mov-ing the sample,firing the lar,stabili
zing the puld lar energy,storing the spectrum from each spectrometer channel and generat-ing the resulting analysis results.A schematic of the clod-loop feedback puld lar energy stabilization device is shown in Fig.3.The vertical direction of the lar that acts as the negative feedback signal is probed by an energy meter,which then generates a 0–5V DC voltage signal that corresponds to the current lar energy level.The voltage signal is measured by the I/O analog input terminal of the spectrometer and transmitted to the computer,which then converts the voltage signal into the energy value through a linear proportional conversion.Then the program determines whether the value of the energy is within the pre-t range.If it exceeds,the
PID toolkit of LabVIEW will generate a corresponding control vari-able according to the deviation quantity,and the power supply volt-age of the xenon lamp will be changed by adjusting the analog output voltage of the spectrometer.This adjustment will not stop until the energy level is turned back to the middle of the pre-t range.Hereinto,the analog signal generated from the spectrometer instead of the potentiometer is employed to change the xenon lamp voltage.In order to prevent any accidental damage on the xenon lamp or lar power supply,the analog output voltage of the spectrometer is restricted to be less than a maximum value limit.Moreover,all the PID parameters should be optimized according to the output energy stability and the respon time for stabilization.2.2.Coal samples
The 550air dried coal samples ud in this study were provided by China Coal Yangjian Washery Co.Ltd.that gathered from different mines in Shuozhou.Following the standard sampling procedure,all the coal samples were prepared with a grain size of ~100μm.Due to the extremely small ablated mass being vaporized and exited by the lar pul,the LIBS precision and accuracy are heavily dependent on the homogeneity of the sample.Therefore,the powdery samples were pretreated by pressing them into compact and smooth pellets (40mm diameter,7mm thickness)using a tablet machine with a pressure of 25MPa for 1min.Each pellet weights 5g by using an analytical balance,of which the precision is 0.05mg at low load.The typical presd coal sample pellet is shown in Fig.2(b).Additionally,in order to make each lar pul have a fresh action point and achieve uniform sample of points,the sample was fixed on a rotation stage driven by a stepping motor and the rotation speed was t appropriately according to the measurement time.All the coal samples were determined following Chine standards GB/T 212-2008.In this work,a total of 500
samples
(a) (b)
Stepping Motor
Fig.2.(a)A schematic reprentation of the analysis chamber;(b)a typical presd coal sample pellet.
XLV:xenon lamp voltage; HVP: high voltage pul; EM: Energy Meter; NEL:negative feedback lar
Fig.3.Experimental tup for lar energy stabilization.XLV:xenon lamp voltage;HVP:high voltage pul;EM:energy meter;NEL:negative feedback larNEL:negative feedback lar.
169
L.Zhang et al./Spectrochimica Acta Part B 113(2015)167–173
were lected as calibration samples and the remaining 50samples were lected for prediction.2.3.Data processing method
Since the basic characteristics of a lar induced plasma are its high temperature and energy propagation in the form of a shock wave,the corresponding signal riously fluctuates with the uncontrollable envi-ronmental changes [14].In order to effectively achieve better noi re-duction and improve the signal to noi ratio (SNR)of LIBS spectra,the spectral peak fitting with a combination of Lorentzian and linear functions was ud in the prent work.It is notable that the back-ground subtraction should be carried out prior to perform the spectral fitting procedure.The original spectrum recorded by the analyzer com-pared with the fitted one is displayed in Fig.4.The original spectrum of coal reveals most of the principal components such as C,H,O,N,Si,Al,Fe,Ca,Mg,Ti,Li,K,Na,etc.However,due to the complexity of peak fitting,only some resp
ective elemental emission lines (listed in Table 1)were identi fied in the spectrum.Obviously,spectral fitting leads to effective enhancement in both the resolution and the SNR.The individual spectral features (atomic and ionic emission lines)are in-tegrated by using MATLAB.To further reduce the signal uncertainties that resulted from fluctuations of lar energy and other factors,the spectral lines of the element under analysis are normalized by rationing
to the total plasma intensity,which could be calculated by integrating the total respon from the spectrometer channels [15].
出人头地的典故
After intensity normalization,discarding the highest and lowest 10%of the data values sorted in ascending order has the further bene fit of preventing erroneous data in fluencing the mean value.Discarding the outlying data reduces the potential changes in size,position,and plasma intensity fluctuations caud by dust particles.Yet it still leaves 80in-tensity readings for each element for the calculation of a reliable mean value [16].Such value is directly related to the elemental information.
For proximate analysis,the coal properties should have a linear rela-tionship with the information of relevant elements in coal.So the line intensities of relevant elements composition were lected to c
onstruct the dominant factors of coal quality.However,the theoretical relation-ship between the intensity and the elemental concentration can be weakened by the uncontrollable experimental parameter fluctuations and the physical/chemical matrix effects.Therefore,to improve the forecasting ability of the model,the curved effects could be modeled by non-linear methods.SVR is an ef ficient non-parametric nonlinear re-gression technique bad on support vector machines,which shows many unique advantages in solving nonlinear and high dimensional patterns [17,18].This method shows excellent forecasting performance due to its particular design of minimizing structural risk.Furthermore,it can ek the best compromi according to the limited sample informa-tion in the complexity of the model and learning ability.For nonlinear problem,SVR us the nonlinear mapping function to map sample to higher dimensional space where linear regression is performed.Thus,linear regression in the output space corresponds to nonlinear regres-sion in the low dimensional input space.Details of our SVR implemen-tation procedure are discusd below.
The calculations pertaining to the multivariate non-linear calibration were performed using a SVR MATLAB toolbox [19].A Radial Bias Func-tion (RBF)kernel was ud for non-linear regression.Prior to SVR leave-one-out cross-validation,the line intensities of relevant element composition of the ash content,volatile matter content and calori fic value were linearly scaled such that the intensity values
were distribut-ed between −1and 1.The main advantage of scaling is to avoid the dominance of speci fic intensity values that reside in greater numeric ranges over tho having smaller numeric values and the numerical dif ficulties during the calculation.Then the optimal model parameters that provide the smallest root-mean-square error cross-validation (RMSECV)were obtained using a particle swarm optimization [20].In this paper,we ud 5cross validation to estimate the accuracy of each parameter combination and to decide the best parameters for the opti-mization problem.Then an effective SVR regression calibration model was established to realize the quantitative analysis of the properties of coal including ash content,volatile matter content,and caloric value.In-cidentally,in order to comprehensively consider the complicated chem-ical composition and structure of coal,more uful information that related to ash content,volatile matter content and calori fic value were lected by PCA.The extracted principal components were ud as the input data for the SVR model.Meanwhile,the methods can
also
Fig.4.Typical averaged original spectra of coal obtained by the two-channel spectrometer compared to the fitted ones.
Table 1
The list of characteristic emission lines for spectral fitting and model establishment.Characteristic line (nm)
Organic elements Characteristic line (nm)
Mineral elements Characteristic line (nm)
Mineral elements Characteristic line (nm)
Mineral elements C(I)193.09Si(I)288.16Fe(I)229.4Ti(I)334.9C(I)247.86Si(I)198.9Fe(I)358.12Ti(I)334.94H(I)656.28Si(I)212.41Fe(I)259.94Ti(I)336.12O(I)776.7Si(I)220.8Fe(I)262.35Li(I)670.3O(I)844.1Si(I)221.67Fe(I)278.81K(I)765.9N(I)742.4Si(I)243.51Al(I)309.27K(I)769.2N(I)743.9Mg(I)280.27Al(I)237.31Na(I)818.33N(I)746.4
Mg(I)285.21Ca(II)315.89Na(I)819.48
Mg(II)279.55
蹦蹦猴
Ca(II)317.93
170L.Zhang et al./Spectrochimica Acta Part B 113(2015)167–173
reduce the redundancy in spectral data and interpret the data relative to
a subt of spectral variations[21].
3.Results and discussion
3.1.Long-term running performance of lar energy
As mentioned above,a clod-loop feedback puld lar energy locking device is integrated into the analyzer by using the lar energy feedback signal to stabilize the Nd:YAG lar output energy so as to en-sure the measurement stability.A one-month long-term test was car-ried out to verify the feasibility of the propod clod-loop feedback puld lar energy stabilization technology.A comparison between the stabilized lar energy and the free running one is shown in Fig.5. It can be en that after stabilization,the relative standard deviation (RSD)of the lar energy stability has been greatly reduced from ±5.2%to±1.3%.
3.2.Ash content
The ash content of coal is an important index for asssing the coal quality.As we know the main ash forming elements of coal include Si,
Al,Fe,Ca,Mg,Ti,K,and Na.Therefore,12characteristic emission lines of the ash forming elements were lected and the78corresponding quadratic and cross variables were directly input for development of the SVR calibration model.However,deviations originated from sources including the lf-absorption effect and the inter-element interference would deteriorate the theoretical relationship.Thus the line intensity of a specific element may not only result from itlf,but also from other elements prent in the plasma.On this basis,considering the complicated structure of coal,more correlations between ash content and other spectral information may not be possibly included in the dominant factor model[22,23].Therefore,in addition to the ash forming elements,more uful and less interference information related to other elements that prent in coal(such as C,H,O,etc.)could be lected from the LIBS spectra by PCA.It is particularly important tofind the rel-evant spectral features of the ash content,which are then extracted to establish the calibration model of ash content by SVR.For example, the correlation coefficients between the136quadratic and cross vari-ables of16characteristic emission lines of the main elements and the ash content were calculate
d using PCA and shown in Fig.6.As we can e,all the variables have a correlation coefficient larger than0.02.For comparison,a minimum correlation coefficient of0.85was lected to reduce the size of input variables to78,which is the same as the simple SVR described above.In Fig.6,the data points that satisfy the‘discard’criterion,and therefore are rejected,are shown as open circles.50un-known samples were ud as validation samples to evaluate the performance of LIBS coupled with the above two SVR models.A com-parison between the results of tho of the simple SVR model and the SVR combined with PCA model for ash content is prented
in格罗安达犬
Fig.5.Comparison of the long-term running performance of lar output energy at60mJ/
pul before and after
沙棘的副作用
stabilization.
Fig.6.Comparison of the correlation coefficients between the input variables that related to
the quadratic and cross terms of the characteristic elemental emission lines and the ash con-
tent.Shown in open circles are tho with less correlation coefficient,as defined in the
动车可以带酒吗text.
Fig.7.Ash content analysis results using(a)the SVR model and(b)the SVR combined
with PCA model.
171
L.Zhang et al./Spectrochimica Acta Part B113(2015)167–173
