Mike Sulatisky, Sheldon Hill, Bryan Lung Alternative Energy Products, Saskatchewan Research Council ABSTRACT CONTROLLER FORM

Transcription

1 Prelmnary Epermental Verfcaton of an Intellgent Fuel Ar Rato Controller Travs Wens, Rch Burton, Greg Schoenau Dept. of Mechancal Engneerng, Unversty of Saskatchewan Copyrght 2007 SAE Internatonal Mke Sulatsky, Sheldon Hll, Bryan Lung Alternatve Energy Products, Saskatchewan Research Councl ABSTRACT Ths paper presents the epermental verfcaton of a new type of fuel-ar rato controller for spark gnton (SI) engnes. The controller does not requre any ntal calbraton before ts frst use on a new engne beyond some very general nformaton whch s used to generate an ntal fuel map. The controller then contnually updates ths nonlnear fuel map n response to changes n the engne or fuel whle drvng. The controller was mplemented on a 200 Vortec 6L V8 engne n a General Motors 200HD truck wth an OEM (orgnal equpment manufacturer) natural gas converson. Prelmnary results ndcate the controller behaves n a manner comparable to OEM controllers n terms of drvablty and ehaust emssons, at potentally a much lower development cost. INTRODUCTION In modern fuel-njected spark gnton (SI) engnes, the vehcle's ECU (engne control unt) s responsble for determnng the proper amount of fuel to nject nto the ntake []. Devaton from the optmum fuel-ar rato (also known as the FAR or FA rato) can lead to reduced power and fuel effcency and ncreased emssons [2,]. Determnng the proper fuellng s typcally acheved by measurng a number of operatng condtons and referrng to a look-up table, known as a fuel map, to determne the optmum fuellng. Generatng the values n ths fuel map s a tme-consumng and epensve operaton, requrng hours of tme on epensve equpment. Ths process must be repeated for each engne model. In some cases, the fuel map has a basc form of adaptaton to changes n vehcle propertes from vehcle to vehcle and over tme; often, though, t s statc n nature. The objectve of ths paper s to present the results of an epermental test of a new method of adaptve fuel control. Ths approach s capable of controllng the fuelar rato to a hgh precson, whle elmnatng the tmeconsumng and epensve ntal calbraton perod. As the overall goal of ths research program s to reduce the cost of convertng vehcles to gaseous fuels (such as natural gas or hydrogen), the epermental results presented are for a vehcle fuelled by natural gas. A smulaton study was prevously performed n order to show the feasblty of the algorthm and s avalable n reference [4]. As ths prevous paper presents the form of the controller n detal, the current paper wll only brefly descrbe ts operaton. CONTROLLER FORM The goal of the fuel-ar controller s to determne the optmum njecton tme for whch the fuel njectors should reman open. It s typcally assumed that ths can be acheved by mantanng a fuel-ar rato at the ntake that oscllates around the stochometrc pont []. Ths would be a smple matter f one could accurately measure the mass arflow or drectly measure the fuelar rato at the ntake. Unfortunately, t s dffcult to measure the nstantaneous mass arflow of the rapdly fluctuatng flow, and the only practcal fuel-ar rato sensors avalable must be nstalled n the ehaust, as they ndrectly measure the fuel-ar rato by measurng the ecess oygen n the ehaust gases. Locatng the oygen sensor (also known as an ehaust gas oygen (EGO) sensor or heated oygen sensor (HO2S)) n the ehaust ntroduces a delay to the system, whch greatly complcates feedback control. Ths delay s due to the tme taken for the charge mture to travel from the njecton pont to the cylnder, the tme that the mture resdes n the cylnder, and the travel tme between the cylnder and sensor. A typcal fuel-ar controller assumes that the nstantaneous mass arflow can be nferred from such avalable measurements as ntake manfold pressure, engne speed, ntake temperature, engne coolant temperature, etc, wth correctons made va feedback from the oygen sensor. The controller presented n ths paper makes smlar assumptons, although t requres The stochometrc rato s the rato of fuel to ar, such that there s just enough ar to completely combust the fuel. Ths s defned as a relatve ar-fuel rato, λ, of.

2 fewer sensors to determne the operatng pont and requres only mnmal ntal calbraton. The controller, presented n reference [4], s a modfed type of nonlnear model nverse controller. The nternal model of the engne s composed of two parts. The frst s a non-lnear but statc model classfyng whether the ntake ar-fuel rato s rch or lean. Ths model takes nputs of manfold pressure, P m, engne speed, N e, and njector pulse wdth, t, each scaled to the range of 0 to, and outputs a value between 0 and, transtonng at ts estmate of the stochometrc fuel-ar rato. Ths model can be thought of as an estmated output of a bang-bang oygen sensor f t could be placed n the ntake, rather than the ehaust. A generalzed neural network [] s used for ths model, wth some weghts strategcally set to zero. A generalzed neural network (GNN) has an archtecture dfferng from the famlar multlayer perceptron (MLP). Whle an MLP has a number of layers wth each layer s nputs comng from the prevous layer s outputs, a GNN can be more easly thought of as a strng on neurons, as shown n Fgure. Each neuron s nput s the output of every neuron to the left of t. Thus, wth four nputs (ncludng a to enable a bas), the frst non-nput neuron,, has four nputs. The net, 6, has fve nputs and so on. Notce that the output, ŷ, has many nputs but, unlke an MLP, t has drect connectons to the network nputs, P m, N e, and t. sensor (HO2S) mounted n the ehaust, as well as the lag n the sensor tself. As there can be consderably varablty of the number of njecton events n ths delay tme, t s treated as a range of probable values. Upon each njecton event, the controller samples nputs of manfold pressure, engne speed, and the HO2S output. The oygen sensor output s converted to a bnary value ( for rch and 0 for lean) and s placed, together wth P m and N e, n a crcular buffer. The controller then checks f the oygen sensor output has been constant for the range of njecton events n the delay model. If so, the delayed oygen sensor output s matched to the approprate delayed nputs of P m, N e, and t from the buffer and the neural network model s updated usng standard back-propagaton on-lne learnng [6,7]. If the oygen sensor output s not constant for the range of njectons n the delay model, no learnng takes place for that sample. The controller then algebracally nverts the neural network to determne the stochometrc njector pulse wdth, t s, whch t estmates wll result n a stochometrc fuel-ar rato for the currently measured manfold pressure and engne speed. Ths algebrac nverson s typcally mpossble to acheve for most types of neural networks, but the generalzed neural network archtecture may be nverted as long as certan weghts are constraned to be zero. The output of the neural network, ŷ, s gven by yˆ = sg( W N 2 N ) () where W n s the weght appled to the nth nput, n, of a network wth N neurons. The sgmod functon sg()=(- e ) - s used whch ranges from 0 to and wth a transton pont of sg(0)=0., whch corresponds to an estmate of the stochometrc pont. In the specfc case of ths controller, the nputs are =P m, 2 =N e, =t and 4 =. The above equaton then takes the form of yˆ = sg( W P or m +... N 2 N e N t ) 4... (2) Fgure : A generalzed neural network. The nputs to each neuron are the outputs of each of the neurons to the left of t. Ths network has 4 nput neurons on the left, hdden neurons, and output neuron on the rght. The network used for the epermental verfcaton has 4 hdden neurons. The second part of the model s a delay, whch takes nto account the transport delays as the fuel travels from the ntake, through the engne and to the heated oygen y ˆ = sg( WPm 2Ne t 4 + ( Pm,Ne, t )) () f we combne the contrbuton of the hdden neuron outputs nto Φ. The goal of nvertng the network s to solve for a t that results n ŷ =0.=sg(0), gven operatng ponts of P m and N e. After applyng the nverse of sg(.), Equaton can be rearranged algebracally to the form

3 = )). (4) Ths cannot be solved analytcally because Φ s a nonlnear functon of t. However, f we constran the weghts assocated wth t to zero, as shown n Fgure 2, Φ s now only a functon of P m and N e. Ths means that, whle the nputs to the last neuron are lnear wth t, the transton boundary s a nonlnear functon of P m and N e. The closed form soluton for the model s estmate of the stochometrc njecton tme s t = ( WPm 2Ne 4 + ( Pm,Ne. () W s )) Ths fuel map can be thought of as a lnear functon, augmented by the nonlnear term Φ. One addtonal advantage to the form of neural network used for ths controller s that the lnear terms n Equaton may be used to set up an ntal fuel map. Ths ntal fuel map s a lnear one, set such that, at the mamum engne speed and ntake manfold pressure, the njectors are open nearly contnually (for ths eperment, they were set to be open 90% of the mamum), and at zero manfold pressure, the njector pulse wdth s zero. The hdden neurons are then ntalzed followng Nguyen and Wdrow s method [8]. Thus, unlke many engne control unts whch requre a larger number of sensors, such as ntake and coolant temperatures, a complete control system for natural gas can be composed of manfold pressure sensor, crank poston sensor (for engne speed measurement and tmng of njecton pulses), ehaust gas oygen sensor (a wde range sensor s not requred), processor, and assocated drvers and wrng. Furthermore, the only engne-specfc data requred to develop and ntalze a controller for a new engne are number of cylnders, appromate mamum engne speed, appromate mamum ntake manfold sensor readng, appromate oygen sensor transton readng, and crank poston sensor encodng scheme. EXPERIMENTAL IMPLEMENTATION Fgure 2: The weghts connectng t to the hdden neurons are set to zero (shown as dashed lnes). Ths allows the network to be algebracally nverted. Fnally, a proportonal-ntegral (PI) controller value s added to ths stochometrc value to correct for shortterm fuellng errors as well as generate the oscllatons requred for proper ehaust catalyst operaton: t ( k) = t ( k) + K ( HO2S( k)) + K ( HO2S( )) (6) s p k = 0 where K p and K are the proportonal and ntegral gans, whch are both functons of HO2S(k) (postve for lean HO2S readngs, negatve for rch), and k s the nde of the tme step. t s s then sent to the njector drver and stored n a crcular buffer. The process s repeated for each njecton. The test vehcle selected for ths demonstraton was a 200 General Motors 200HD pckup truck wth a 200 Vortec V8 6L engne and automatc transmsson, converted to run on natural gas (see Fgure ). A controller, developed by the Saskatchewan Research Councl for a prevous project, was programmed wth the new algorthm. The controller (shown n Fgure 4) ncludes sensor and njector drvers, three mcrocontrollers to generate the pulse sgnals, and a Motorola MPC CPU to handle pulse wdth computaton. The truck's orgnal manfold pressure, crank poston, and ehaust gas oygen sensors were used, as well as the standard fuel njectors. The OEM fuel controller was not permanently deactvated, so that drect comparsons could be made between the OEM system and the proposed controller by flppng a swtch. At all tmes the OEM ECU mantaned control of all functons other than fuel control, such as spark tmng, dle ar control, transmsson control, etc.

4 EXPERIMENTAL RESULTS Ths secton presents epermental data recorded whle drvng the "Hot 0" secton of the Federal Test Procedure (FTP) drvng cycle, whch s shown n Fgure. The neural network was rentalzed mmedately before startng the test, whch was performed wth the engne, oygen sensors, and catalyst warm. Fgure : Vehcle used for epermental tests, a 200 GM 200HD truck wth a Vortec V8 6L natural gas/gasolne b-fuel engne. Fgure : The drvng cycle used n ths test was the "Hot 0" secton of the FTP (Federal Test Procedure) profle. Ths drvng cycle s 0 s long and ncorporates moderate acceleraton and brakng, as well as freeway-style drvng. Fgure 4: The controller used to mplement the algorthm s based on a MPC CPU and ncludes nterfacng crcutry for drvng the njectors as well as readng varous types of sensors. The neural network control algorthm was mplemented wth neurons and a tranng rate of The delay was estmated to be between 8 and 8 njectons, measured by cuttng fuel flow and measurng the tme taken for the oygen sensor to transton from a rch readng to a lean one [9]. The PI controller was set wth K p =0 (ntegral control only) and K =-0.00 ms/sample for rch HO2S readngs and 0.0 ms/sample for lean. These unsymmetrcal ntegral gans were used to bas the fuel-ar rato slghtly rch for the reducton of ntrogen odes (NO). The trace of the relatve fuel-ar rato, λ, s shown n Fgure 6, wth the same data shown n hstogram form n Fgure 7. Note that ths data was used for evaluaton of the controller only and was not avalable to the control algorthm. Over the course of the test the mean λ was wth a standard devaton of For comparson, the OEM controller produced a mean of and a standard devaton of over the same drvng cycle. A wde-range oygen sensor was nstalled, but was only used for evaluatng the performance of the controller, not as part of the control scheme. A 200 ksamples/s data acquston card was used to record the manfold pressure, wde-range oygen sensor output, the two precatalyst ehaust gas oygen sensors, as well as the pulses sent to the njectors.

5 λ 0.9 HO2S (V) k (njectons) 0 4 Fgure 6: The measured relatve fuel-ar rato whle drvng the profle shown n Fgure had a mean of and a standard devaton of 0.040, compared to values of and for the OEM controller on the same drvng cycle. Note that the wde-range oygen sensor was used for montorng the performance of the control scheme only, and s not used for feedback λ Fgure 7: The data set n Fgure 6 s shown here n hstogram form. Fgure 8 shows the HO2S feedback readngs for the same test, demonstratng the epected lmt-cycle behavour requred for proper catalyst operaton k (njectons) 0 4 Fgure 8: The Heated Oygen Sensor (HO2S) was used for both feedback to the PI controller as well as for tranng the neural network. Notce the lmt cycle behavour. Ehaust emsson data was also recorded usng an ECOM AC portable emsson analyzer and sample dryer. Note that the calbraton of ths nstrument was not certfed, so the measurements should be used for comparson purposes only. The two pollutants measured were carbon monode (CO) and ntrogen odes (NO), as shown n Fgures 9 and 0, along wth measurements from the OEM controller for comparson. The mean carbon monode concentraton was 29. ppm wth a mamum of 788 ppm for the neural-based controller, compared to the OEM controller wth a mean of 06.4 ppm and a peak of 98 ppm. The mean and mamum NO concentratons were. ppm and 207 ppm for the neural controller, whch may be compared to values from the OEM controller of 4. ppm and 260 ppm. These prelmnary results ndcate that the controller presented n ths paper has performance comparable to the OEM controller (f not eceedng) wth regards to ehaust emssons. Furthermore, no msfres were detected, and the engne seems to run smoothly, after the frst few seconds of rough dle. Wth regard to computatonal performance, the MPC CPU could handle the calculaton of the pulse wdth wthn the RPM lmts of the engne. The pulse wdth calculaton wll take more processng power than a smple look-up table; measurements ndcate that the neural pulse wdth calculaton takes appromately 0. ms, the learnng phase takes 2.0 ms, and the complete loop takes 2.4 ms per njecton. For an eght cylnder engne, ths corresponds to a mamum engne speed of 600 RPM. It should be noted that the code used was relatvely unoptmzed, and one can epect the possblty of sgnfcant mprovements through optmzaton, especally n the learnng secton. The memory requrements of the algorthm are comparable

6 to a lookup table; a network wth N neurons requres (N- 4)*(N+)/2+ weghts. The -neuron network used n ths eperment requred 89 weghts, less than a 0 0 lookup table. Addtonally, the software requres an N element vector to store the neuron outputs, a temporary (N-4)*(N+)/2+ tem vector to store the gradents of the weghts, and four crcular buffers wth lengths equal to the engne s mamum delay. CO Concentraton (ppm) OEM Neural Control tme (s) Fgure 9: Ehaust gas carbon monode concentraton measurements were recorded durng the same test, shown here wth OEM results for comparson. These concentratons are for dry ehaust. NO Concentraton (ppm) OEM Neural Control tme (s) Fgure 0: Ehaust gas NO concentraton measurements were recorded durng the same test, shown here wth OEM results for comparson. These concentratons are for dry ehaust. CONCLUSION Ths paper presents prelmnary epermental verfcaton of a new type of adaptve fuel-ar controller. These prelmnary results demonstrate that the scheme has ehaust emssons performance that s comparable to, or eceedng that of the OEM controller, whle elmnatng the costly and tme-consumng process of calbratng the fuel map. It should be emphaszed that these results are prelmnary and apply to only one vehcle; however, t s beleved that the concept s very sound and can be appled to many other stuatons. Further, t s epected that these unoptmzed results can be mproved through tunng the algorthm. Ths can be acheved by mprovng the learnng rate by usng more advanced tranng schemes than the smple back propagaton algorthm, and also mprovng the computatonal speed so that more neurons may be used n the same calculaton perod. Further work may also nclude applyng the same controller to varous engnes to verfy the unversal nature of the controller, and contnung ongong long-term tests to evaluate the controller s performance as the engne wears. ACKNOWLEDGMENTS The authors would lke to acknowledge the support of NSERC, n the form of a PGS D Scholarshp; the Saskatchewan Research Councl for provdng equpment and epertse; and General Motors Alternatve Fuels for the loan of the test vehcle. Thanks are also gven to Natural Resources Canada and Precarn Inc who funded ntal work n ths area. REFERENCES. Isermann, R. and Muller, N. Desgn of Computer Controlled Combuston Engnes. Mechatroncs, : , Taylor, Charles Fayette. The Internal-Combuston Engne n Theory and Practce, volume I. M.I.T. Press, Cambrdge, second edton, 98.. Heywood, J.B. Internal Combuston Engne Fundamentals. McGraw Hll, New York, Wens, T., Burton R., Schoenau, G. and Sulatsky, M. Intellgent Fuel Ar Rato Control of Gaseous Fuel SI engnes. Techncal Report, Saskatchewan Research Councl, Saskatoon, Werbos, P. Backpropagaton Through Tme: What It Does and How to Do t. Proceedngs of the IEEE, volume 78, pages 0 60, Haykn, S. Neural Networks: A Comprehensve Foundaton. Prentce Hall, New Jersey, 2nd edton, Saad, D., edtor. On-lne Learnng n Neural Networks. Cambrdge Unversty Press, Cambrdge, Nguyen, D. and Wdrow, B. Improvng the Learnng Speed of 2-Layer Neural Networks by Choosng Intal Values for the Adaptve Weghts.

7 Proceedngs of the Internatonal Jont Conference on Neural Networks, pages 2 26, San Dego, Wens, T., Burton, R., Schoenau, G., Sulatsky, M., Hll, S., and Lung, B. Epermental Determnaton of Transport Delay n a Spark Ignton Engne. Techncal Report, Saskatchewan Research Councl, Saskatoon, 2006.