Analysis of LPG Engine PID Parameter Control of Transient Air-fuel Ration Based on Improved Elman Neural Network

Transcription

1 54 JOURNAL OF SOFWARE, VOL. 5, NO. 1, JANUARY 2009 Analysis of LPG Engine P Parameter Control of ransient Air-fuel Ration Based on mroved Elman Neural Network Yao Jubiao College of Environmental and Energy Engineering, Beijing University of echnology, Beijing100022, China; Beijing Vocational College of Electronic Science, Beijing100029, China yaojubiao@yahoo.com.cn Jiang Keyu et. of nformatics, Fort Hays State University, KS , USA keyujky@gmail.com Yao Chaoyi Southwest Jiaotong University, Chengdu , China Abstract his aer uts forward one kind of air-fuel ration control method which integrates imroved Elman neural network with normal P control. t constructs a model of the single LPG electro-lating engine and simulation latform for air-fuel ration controlling with G- Power software which seamless connect to Matlab/Simulink based on JL1P39FMB single cylinder engine and develo the electronic fuel-injected controller based on ntel MCS- 96. t also constructs mini LPG electro-lating engine exeriment system for air-fuel ration to the LPG njection System and adot the new inlet channel injection tye of duty ratio controlling injector. Hel the throttle ercentagerev- duty cycle ulse sectrum diagram after the calibration bench exeriment for the best duty cycle of mini electronic LPG-injected under the steady working conditions. t redicts the air-fuel ration signals of nontransmission delay through the Elman neural network. he normal P controller which deals with the redictive signals imlements the concurrent control of air-fuel ration under transient conditions. he bench test and the simulation result indicate that the control methods have the strong adatation, which can make the statism of air-fuel ration under transient conditions on ±5%. ndex erms P arameter Control, Air-fuel ration under transient, Elman Neural Network,Predictive control. NROUCON LPG (Liquefied Petroleum Gas) electro-lating engine control the air-fuel ration accurately when the engine in the stability condition. But automobile-used LPG electrolating engines are almost all in transient running among actual oerating mode. nflate-exhaust and gas backflow lead to the fluctuation range increasing of air-fuel ration when transient running. So most scholars devoted to he Youthful Scientific Fund of Beijing University of echnology ( ); he Scientific Fund of Beijing Education Commission (KM ) imrove recision of air-fuel ration and reduce the calibration rocess work. he model-based control strategy and the intellectual control strategy based on neural network are the new focus in air-fuel ration controlling field at resent. he technology that combines the electronic fuel injection and the three-way catalyst has become the most effectively method to decrease the exhaust ollution of the gasoline engine in vehicle. However, the efficiency is influenced by air fuel ratio greatly. f the high efficiency is required (above 80%), the air fuel ratio needs to be controlled about 14.7, and the error needs to be held within the range of ±3.5%[1].he engines of vehicle in cities oerating in the steady state and the transient state, which alter frequently. he measured air fuel ratio signal is delayed due to the case that the UEGO sensor is fixed in the exhaust ie of the engine. When the engine oerates in the steady state, the controlled results of the air fuel ratio can meet the requirement of the three-way catalyst; When oerating in the transient state, the throttle osition changes greatly, the fluctuating amlitude of the air fuel ratio in the cylinder increases, then the transortation delay of the air fuel ratio degrades the control quality. n order to avoid this deficiency, most of the researchers dedicated themselves to study on how to imrove the control quality and to reduce the workload of calibration rocess, thus the control strategy based on model or neural network become new research focus at resent [2,3]. First in the essay, the transortation characteristics of the air fuel ratio were analyzed. And then, a kind of air fuel ratio control strategy that combined the modified neural network and the traditional P controller was ut forward in this aer, in which the neural network was used to estimate the air fuel ratio signal without the transortation delay, while the traditional P controller doi: /jsw

2 JOURNAL OF SOFWARE, VOL. 5, NO. 1, JANUARY was used to control the transient air fuel ratio by using the estimated signal. he model of a single cylinder engine was set u by using G-Power; the model of the control system was set u by using Simulink. According to the characteristic that G-Power can be linked with Matlab/Simulink seamlessly, the simulation model of the air fuel ratio control of the gasoline engine was set u by using the couled technique between the two kinds of software[4]. On basis of this, air fuel ratio control strategy that combined the neural network and the traditional P controller was studied. Figure 1. he model of a single cylinder electronic controlled gasoline engine. ANALYSS ON HE RANSPORAON CHARACERSCS OF AR FUEL RAO A. he duration 1 of the mixture held inside the cylinders he transortation delay of the air fuel ratio mainly includes the following three arts: 1)he duration 1 of the mixture held inside the cylinder. 2)he duration 2 of the waste gas flowing from the exhaust valve to the osition of the UEGO sensor. 3)he resonse time 3 of the UEGO sensor. he duration 1 is related to the rotate seed of the crankshaft, the crank angle from the closing of the intake valve to the oening of the exhaust valve, exressed as the following formula. 6n (1) 1 Herein, denotes the crank angle while the mixture is held inside the cylinder in a working cycle, =360 in this aer; n denotes the rotate seed of the crankshaft. B. he duration 2 of the waste gas flowing from the exhaust valve to the osition of UEGO sensor he time 2 can be confirmed by the following formula [5]: V (2) 2 v eg Herein, V denotes the cubage of the exhaust ie, which is related to the structure of the exhaust ie and the osition of UEGO sensor; v eg denotes the volume velocity of the exhaust flow. v ( R P ) m (3) eg eg eg eg Herein, R denotes the gas constant; denotes the temerature of the exhaust gas, which is related to the load of the engine; denotes the ressure of the exhaust ie; denotes the mass velocity of the flow. C. Set u the G-Power model of enginer Knowing from the working rincile of the UEGO sensor, the value of the air fuel ratio is acquired by insecting the electrical current of the umed oxygen, so a certain resonse delay is in existence. t was suosed that the UEGO sensor belongs to a kind of aaratus with the first order resonse characteristic [6], the time constant was suosed at 100ms in this simulation.o the above analysis, the transortation delay of the air fuel ratio is related to the structural arameters of engines, oerating condition of engines and the resonse characteristic of the UEGO sensor V Figure 2. rigger of magneto-ignition electrical fire Figure 3. Calibration results on transmission delay of Airfuel ration Wave, Boost and G-Power are oular onedimensional engine simulation software at resent. G- Power is used most widely therein, which was adoted in this aer. he model of a single cylinder electronic controlled gasoline engine was set u in G-Power, and the detailed structure is illustrated as Figure 1. he adjustable throttle valve was included in the model, whose adjustable range was 0~40mm. here was no intake ressure regulated cavity in the model, so the mass flow of the intake air was measured according to the throttle osition and the rotate seed of the crankshaft. he inut signals to the ECU included throttle osition signal, crankshaft rotate seed signal and the feedback air fuel ratio signal, the outut from the ECU was the CA

3 56 JOURNAL OF SOFWARE, VOL. 5, NO. 1, JANUARY 2009 injection duration signal, the injection angle was set to 90 before the to dead-center, when the injection ended in a working cycle.. Aaccurate Comarison and Exeriment Measurements o rovide an accurate comarison, measurements in all working condition should be consistent with the hase of engine working rocess. n this exeriment, C is calibrated by C trigger signal which roduced by engine magneto, and it will be regarded as the hase of the signal of recording measurements.he resolution of engine crank angle is 1 CA, and the maximum allowable seed is 12000r/min (instantaneous). he crank angle signal is used as the trigger signal, which is used to samling once er degree of crank angle. he cylinder ressure signal and the C synchronization signal are recorded with the samling frequency of 400 khz.n order to ensure the integrality of information in samling, there should be a sufficiently large samling frequency in the system. Assuming that the maximum engine seed is n, and er degree crank angle can we get m samling oints, then the samling frequency is: n f 360 m 6mnHz 60 ata acquisition board uses the order samling or synchronous acquisition time-transfer structure. f the maximum engine seed n = 9000r/min, m = 1, channel = 2, the samling frequency is required for the 108k. he system has the greatest samling frequency 400k, fully meet the collection requirements.angle encoder generates 3 signals, A, B, and Z. A, B signals mean that they can give a square wave signal for each degree of crank angle, and the hase distance is 1 / 4 degree of crank angle. A, B signals can be used to judge turning direction of crankshaft and raise the angle resolution to 0.25 degrees of the crank angle. Z signal is that it roduces a square wave signal when the crank changes 360 degrees, which is used as the calibration C osition synchronous signal in the acquisition system. n the measurement rocess, the angle encoder A signal is used as a trigger signal. t generates a square wave signal while crankshaft changes er angle, and the trigger samling haens in the ositive rising. Seed ABLE 1 CONROL MOEL OF UY RAO idling Small duty med. duty med. duty full load full load % 75% 55% 30% 20% 10% % 40% 35% 9% % 35% 25% 5% We can record the cylinder ressure and the C synchronization signal when the system samling frequency comes to 400 khz. he time difference is 2.5μs between the two signals, because it uses the order of samling. We use JL1P39FMB in the exeriment. he maximum engine seed is 7200r/min, and the shortest time when the crankshaft changes er angle is 23μs, that is, synchronous signal C is behind the ressure signal at the time of maximum of about 0.1 crank angles. his error value is not heavy, and it can be eliminated by amend thermodynamics of C osition. ue to the big vibration and electromagnetic interference in the working environment of measure, the cylinder ressure signal and the angle coder signal will be imacted. n order to reduce this imact, we make use of shielded cable for all signal transmission in this system, and the entire system is commonly grounded to imrove the anti-interference ability. o ead Center (C) determination: t is very imortant to calibrate C accurately in the analysis. Usually when the maximum outburst ressure oint in the cylinder reaches at 12~15 Crank Angle (CA) after C, the engine gets the best ower and economy erformance [7,8]. n this system, C is determined by the method of thermodynamic amendment on the maximum dragging comression ressure. herefore, as long as the fixed difference between the osition of angle coder synchronization signal and that of maximum dragging comression ressure is determined, then the osition of C will be determined according to the location of sync signal. Considering the angle coder may not be able to be installed, and in order to determine the aroximate location of C, we have measured the trigger signal, which is emanated by engine magneto, of C noncontact ignition system in exeriment, which is shown in Figure 2. When trigger device rotates over trigger coil, magneto inducts the ositive and negative ulse signal (which is shown in figure 2) to control the ignition time. he minimum value of negative ulse trigger signal corresonds to the osition of C. n this way, C still can be calibrated even without angle coder. Samling rocess: it needs large amounts of high-seedily and continuously samled data in one measure. At the beginning, the acquisition card works according to the set method and rate of samling. ata are accessed to the on-board cache of 4K FF0 in time and channel order; when on-board FF0 has read into 2K data(half full), the interrution is generated, then 2048 data are written into the 2MB data buffer of comuter. As the riority of interrution is higher than that of other oerations, the function of saving data and dislaying can be comleted at the same time without affecting the integrity of the date that are read into the memory buffer.. MEHOS OF CALBRAON EXPERMEN Calibration exeriment is carried out in a LPG singlecylinder motorcycle engine with EF. As a result of the adotion of LPG fuel suly mode with gaseity injection in gas inlet ort, fuel film effect is not taken account of in calibration. n order to reduce the accumulation errors,

4 JOURNAL OF SOFWARE, VOL. 5, NO. 1, JANUARY when calibrating transmission delay time of the air-fuel ratio, we consider the three stages of transmission delay of air-fuel ratio as a whole. Calibration exeriment is carried out in steady-state condition; the secific exerimental system is described as following. Engine fuel injection ulse width signal is a suerosition of steady-state injection ulse width signal and disturbance signal. he steady-state ulse-width signal is exorted by the ECU of LPG EF engine according to the actual condition and the exected air-fuel ratio, while the disturbance signal is exorted from /A outut channels which are embedded in AC6111 high-seed data acquisition card. We introduce AFRt as AFR disturbance and AFRe as AFR disturbance in exhaust, and then we record the changes over time of these two AFR disturbances. By comaring the hase difference between them can we measure transmission time of AFR. he oututs of engine intake manifold ressure sensor and airfuel ratio sensor are connected resectively to the A/ channel embedded in AC6111 high-seed data acquisition card % 50.00% Figure 4. uty-ratio control diagram % uring the calibration exeriments, we regulate C ower dynamometer and engine throttle. When the throttle osition and seed are adjusted to working conditions that we want to calibrate, PC roduces a ositive salutation disturbance signal, which adds to steady injection ulse width signal, through the /A channel embedded in AC6111. At the same time, PC makes a high-seed samling to the intake manifold ressure sensor (MAP) and air-fuel ratio sensor (UEGO). As a result of the seed-lock function of exerimental C ower dynamometer, the disturbed seed can still be maintained at the calibrated seed oint, while the airfuel ratio sensor outut will have a salutation. After analyzing the relationshi between amlitude and hase of intake manifold ressure signal and those of air-fuel ratio signal, we can get the transmission time from intake valve closing, combustion and gas exhausting to the outut of air-fuel ratio sensor in the calibrated working condition oint [9]. A. Construction of mathematical model and arameters otimized for Calibration Rresults he oerating condition variety scoe of calibration exeriment shows as follows: load from 20% to 80%, by the ste 10%; Seed from 2000rm to 6000rm, by 500rm ste. he changes in the ulse sectrum of calibration results is illustrated in Figure 3, which shows that the air-fuel ratio transmission delay time is the function of the load and seed, with inverse roortion[10]. he exerimental data under the low fuel consumtion, low HC and CO emissions, and constrained NOX emissions are otimized on condition that the engine oerates in security and stability. We choose the best results as the MAP basic raw data, which consider as the basic control data in different seed and throttle oening. he mathematical model is set by tabular method, zoned by the throttle osition and seed. We comare the seed characteristics in a variety of throttle oening via the engine bench test to determine the otimum electric control arameters, which are ut into the comuter memory. he more detailed form, the smaller region formed by throttle and seed, and the higher control accuracy. However, because of restrictions on the sensor characteristics, throttle osition is divided into six sections, which results discontinuous of the control volume in different regions, and makes the air-fuel ratio control a ste change. hus, interolation table should be used in the actual control. he control model of duty ratio is shown in able 1 B. Calibration results of duty-ratio control values uty-ratio control values are showed on able2, and. ABLE 2 UY-RAO CONROL VALUES F n % % % % % % % %

5 58 JOURNAL OF SOFWARE, VOL. 5, NO. 1, JANUARY 2009 the Figure 4 is the duty-ratio control diagram above the duty-ratio control values. V. GAL MPLEMENAON ON ALGORHM OF P P control based on the deviation calculates of the imortation according to the roortion, integral and differential. he results of oeration are exorted as a control. P control system for the basic structure rogram consists of the feedback rincile, which is shown in Figure 5. A. Analysis of the results of oeration Outut c( is fed back to the inut, then digital P control is added by the deviation resulted in the comarison between the inut signal and the feedback signal to regulate the volume u (, which is exorted to the controlled lant. P control adjusts the control volume according to the changes of deviation, and it is also known as digital P regulator. For examle, the deviation increases, and the control volume of u have also increased, which is roortional control. As the deviation E exists all the time, we should be integral, and the Figure 5. Elements of P Control volume is increasing to eliminate deviation E. his is called integral control. he differential control is exected to lay the role, that is, when E> 0, indicating that the deviation is increasing, the control volume should be increased to reduce the deviation; When E <0 indicating that the deviation in reduction, the control volume should be reduced to avoid oscillation which is caused by deviation E develoing in oosite direction when it trends to zero. n ractical alications, the structure of P can be changed flexibly in accordance with the characteristics of the object and control requirements. Some arts of P, such as roortion (P) regulation, roortion-integral (P) regulation, roortion-integral-differential (P), etc., can be taken to comose control rules. Esecially in the comuter control system, they can be more flexibly used, in order to give full lay to the role of the comuter. n stimulation system, the formula of P can be written as follow: 1 de( P( K e( e( dt (4) dt n this formula, P(-outut signal of regulator; e(-the deviation signal of regulator, which is equal to the error between the measured value and the given value; KP-the roortion coefficient of regulator; -the integral time of regulator; -the differential time of regulator. Because comuter control is based on samling control, which can only calculate the control volume by the deviation of the samling time, firstly the formula (4) must be disersed in comuter control system. he differential equations for the system are relaced by the differential equations in digital form, in this way the integral arts can be exressed in the form of sum and increment: 1 de( P( K e( e( dt dt de ( k 1) k 1) (6) dt t After utting the formula (5) and the formula (6) into the formula (4), we can get discrete P exression: P( K j0 n this formula, t=-samling eriod, must be small enough to ensure a certain degree of accuracy; -the error at the time of k-samling; k-1)-the error at the time of (k-1)-samling; K-samling number, k = 0,1,2...; P(-the outut of the regulator at the time of k- samling. he outut value of formula (7) corresonds with the oening osition of the injector valve, therefore, usually formula (7) is called as the location formula for locationbased P control. t can be seen from formula(7), that if we want to calculate P(, we not only need the last error signal and current error signal k-1), but also need to add the revious error signals j) for integral arts, that is, j) j0 n this way, not only the amount of calculation is huge, but also a large number of memory will be used to reserve j). hus, it is not convenient to control directly by formula(7). n order to solve this roblem, we make following changes. According to the rincile of recursion, the exression for outut of P at level (k-1) can be written as follow: P( k 1) K k 1) k j) Formula(7)subtracted by formula(8),then get K K k k 1 j0 j) k 1) P( P( k 1) K [ k 1)] k 1) k 2) [ 2 k 1) k 2)] K K (5) (7) (8) (9)

6 JOURNAL OF SOFWARE, VOL. 5, NO. 1, JANUARY is the integral action coefficient; K K is the differential coefficient. From the formula(9), we can see that there is no need but to get the value of P(k-1),, k-1) and k-2) to calculate the outut value of P( at level k, so it is much simler than the use of formula(7). n many control system, as actuating devices are controlled by ste motor or multi-coil otentiometer, an incremental signal is enough. herefore, by the formula (7) subtracted from formula (8), we can get: P( P( P( k 1) K [ k 1)] K K [ 2 k 1) k 2)] (10) but also any failure of comuter will lead to a big change of P(. Although the modifications of the incremental control are not so big, it has brought a lot of advantages: the imact of malfunction is small because of the incremental outut in comuter, if necessary, it can be removed by logic; in the osition controlling algorithm, in order to achieve manual/automatic switch without disturbance, when the manual mode is being switched to the automatic mode, the outut value of comuter must be first of all equivalent to the original valve oening, that is, P ( k 1),which would creates difficulties for rogram design. n the incremental design, the outut only is related to the error, not the original osition of the injector valve. hus the incremental algorithm is easy to achieve manual/automatic switch without disturbance, and as it doesn t create any uncontrolled integral, good regulating effects can be achieved easily. However, this control also has its deficiencies: large integral truncation effects, which result in static error; large effects of overflow. herefore, the right controlling algorithm should be chosen based on the actual situation of the object. Generally seaking, location-based algorithm should be adoted in the system that uses servo motor or thyristor as imlementation device, or that has high requirement for controlling recision, while incremental algorithm should be adoted in the system that uses the ste motor or multi-coil otentiometer as imlementation device. n addition to these two controlling algorithms, there is also another control algorithm which is called seed controlling algorithm. hat is, P( V ( K k 1) [ 2 k 1) k 2)] (11) iscrete rogramming exression for the incremental P algorithm from formula (10) can get the incremental P algorithm as following: P( K K [ k 1)] K [ 2 k 1) k 2)] So P ( K [ k 1)] P ( K Figure 6. Flow chart of P he formula (10) shows that the increment of outut at level k ( P( ) is equivalent to the error between the outut of regulator at level k and that at level k-1, that is, the increased (or decreased) value at level k-1. herefore, formula (10) is called as the incremental P controlling formula. n the osition controlling formula, because of the total outut, each outut is relative to the original location. herefore, not only E (j) needs to be cumulated, P ( K [ 2 k 1) k 2)] And get P( P ( P ( P ( (12) Formula (12) is the discrete rogramming exression for the incremental P algorithm. t will be stored at the designated unit of RAM searately when get KP.K.K.

7 60 JOURNAL OF SOFWARE, VOL. 5, NO. 1, JANUARY 2009 B. Eliminates for P controls n the general control of the P regulator, because the linear range of the actuator is limited, when the deviation E become larger, such as system starting, stoing or significantly changing, because of the role of integral term, there will be a great overshoot leading to the system continual oscillation. his henomenon affects more deely the objects that change slow, such as temerature, liquid level -conditioning systems, and it also exists in the general simulation-conditioning systems. n order to eliminate this henomenon in the comuter control system, we can use the method of integration searation. hat is, when the deviation exceeds a given range, integration does not work until the volume is adjusted to the given value. For the given value of R (, measured value of M (, the maximum allowable deviation value of A. he formula of searation control is shown as follows: R( M ( (13) By the use of integral action searation, the overshoot of controlled variable and the transition time can be reduced significantly, which imroves the regulating erformance. P control flow chart for the system with integral action searation, which is shown in Figure 6. n the initialization rogram, E (k-1), E (k-2), P (k-1) should be unit reset. V. GAL CONROL SYSEM OF P PARAMEERS uning arameters is very imortant in the digital control system, which imacts on the quality of regulation directly. P control arameters include the samling eriod, the ratio of coefficient KP, integration time and the differential time. A reasonable choice of these arameters is one of the key issues in digital control system. A, P A, P etermination of samling eriod : according to the Shannon samling theorem, fs stands the uer limit of f s 2 f the samling frequency. f max, the system can be truly restored to the original continuous signal. A. he factors of imacting of samling eriod By theory, the higher the samling frequency, the smaller the distortion. However, from the controller itself, it adjust rely on deviation signal E (. When the samling eriod is too small, the deviation signal E ( will be too small, while the regulator will lose the function of adjustment. But samling eriod is too long, will cause the error. herefore, the samling eriod must be considered synthetically. he factors of imacting of samling eriod is described as follows: the higher disturbances frequency adding to the object, the higher samling frequency, that is, samling eriod is shorten; he dynamic characteristics of objects mainly connect with the object for the delay time θ and time constant τ, and when a significant delay time comes u, the samling eriod is basically the same with the ure delay time θ; f the digital control uses ahlin algorithm and eninsula actuator, samling eriod becomes longer; f digital control use the fastest non-corrugated systems steer motors, samling eriod become shorter; he larger the number of control loo, the longer the, otherwise, we have smaller ; n general, the higher the control accuracy, the shorter the samling eriod to reduce the time delay system. Figure 8. he controlled results of air fuel ratio based on traditional P controller Figure 7. he simulation model of air fuel ratio control with traditional P controller here are two tyes of selection of samling eriod, one is the calculation method, and the other is the exerience method. Calculation is less used, because it is more comlex, esecially difficult to determine the constant time in every sectors of the controlled system. So-called exerience is a multi-try method. hat is, in accordance to the ractice of working eole, as well as the characteristics of the objects and arameters, we choose an aroximate samling eriod firstly, utting it into the comuter control system to test, and then we

8 JOURNAL OF SOFWARE, VOL. 5, NO. 1, JANUARY reeat amendments to until satisfied. aking into account of the above factors, combined with the actual situation of the system, the samling frequency take 10Hz, that is, samling eriod = 0.1s. he selection of KP,, : Because of the rocess of actual control comlexity and the changing arameters, it is difficult to determine the dynamic characteristics of lants. Sometimes, maybe you can find it, but in this way, not only it has the trouble in calculation and heavy workload, but also it has the big difference between the results and the actual value[11]. herefore, the exerience method is the most widely used. According to the secific Regulating and the different characteristics of the object, after the closed-loo testing, we try it over and over to find the best adjustments. For the secific circumstances of this system, because of the general adjustment of the injection of no more than the basic injection amount of 5% duty cycle, the adjustment of P control of the duty ratio should not exceed 1%. hrough the calibration of oxygen sensor, we consider the air-fuel ratio which corresond oxygen sensor outut voltage of 0.5V as the theoretic air-fuel ratio. So at last get KP =0.005=1/200, =3.5,=1s and K K 1/ 7000 K K / 20 After acking the model in G-Power, which is shown in Figure 1, it could be used for Simulink. he simulation model of air fuel ratio control with traditional P controller is illustrated as Figure7, the injection duration signal inut to the ECU was the sum of two arts: one was the look-u value according to the value of the throttle osition and the crankshaft rotate seed, which needs to be calibrated in advance; the other was the calculated value of the P controller according to the feedback from the UEGO sensor. he target value of the air fuel ratio was set to Figure 9. he simulation model of air fuel ratio control based on neural network When the engine works in steady state, the air-fuel ratio can be controlled stably around the target value shown in Figure 8. When the engine works in transition condition, as a result of the "Charge Reflux" in intake manifold and the transmission delay of air-fuel ratio, the air-fuel ratio will be down when throttle angle increasing, that is, the concentration of mixture will become higher when throttle angle decreasing. C. Prediction method of the air fuel ratio without the transortation delay For the forecast mode of air-fuel ratio without transmission delay, lease refer to Figure 9. he essence of the air fuel ratio is to control the mixture s concentration inside the cylinder, the erfect mode is to measure the mixture concentration inside the cylinder directly, and however the UEGO sensor can only be fixed in the exhaust ie at resent. n order to avoid the maladjustment of the air fuel ratio because of the transortation delay measured by the UEGO sensor, the assumtion, which utilizes the air fuel ratio signal without the transortation delay to close the P feedback controller, was ut forward in this aer. he air fuel ratio without the transortation delay must be redicted, the rediction method is illustrated as Figure 9. Figure10. he structure of the modified Elman neural network From the fore mentioned analysis on the transortation delay of the air fuel ratio, 1, 2 and 3 can cause a bad quality of the air fuel ratio control. he air fuel ratio without the transortation delay can be acquired by training the neural network with the teacher s suervision on-line. llustrated as Figure 9, the inut to the neural network consists of throttle osition signal, crankshaft rotate seed signal and the injection duration signal, the outut from it was comared with the outut of the UEGO sensor after being delayed by the 1, 2 and 3, then the training algorithm regulated the weight of each neuron in the neural network according to the comared error. he outut from the neural network can be regarded as the air fuel ratio signal without the transortation delay when the error was limited within the set range. he transient air fuel ratio would be controlled erfectly when the trained signal acted as the feedback signal of the P controller [12]. he air fuel ratio signal was taken from the exhaust valve in this simulation study, for sake of simlifying the model. So 2=0, only 1 and 3 were considered while the neural network is being trained.

9 62 JOURNAL OF SOFWARE, VOL. 5, NO. 1, JANUARY 2009 he essential of air-fuel ratio control is controlling the concentration of air and fuel mixture, which is involved in the combustion. he ideal control mode is monitoring the concentration of in-cylinder mixed gas directly, but current air-fuel ratio sensor can only be installed in the exhaust manifold. n order to avoid the roblem of maladjusted control to air-fuel ratio in transition condition caused by transmission delay, in this aer, the idea of air-fuel ratio rediction with no delay in transmission is roosed, which is shown in Figure 9. From foregoing analysis of transmission delay of air-fuel ratio, we can see that 1, the in-cylinder residence time of gas mixture after the closure of intake valve, and 2, the time that emission gas moves from exhaust valve to sensor, and 3, the resonse time of air-fuel ratio sensor, are the key effects on real-time control of air-fuel ratio. herefore, we can realize the forecast of air-fuel ratio signal with no transmission delay by the use of suervised neural network learning. he inut of neural network is the signal of throttle angle, seed signal and fuel injection ulse width value. After delayed by 1, 2 and 3, the outut is comared to that of air-fuel ratio sensor. he error from the comare will be regarded as the basis of regulating neurons weight vector in neural network. When the error is controlled within the set range, the neural network outut signal is the air-fuel ratio signal with no transmission delay. By using this signal as P controller feedback signal, the real-time control of the airfuel ratio in transition condition can be realized. n this study, in order to simlify the model, air-fuel ratio signal is taken directly from the exhaust valve. So 2 =0. Only 1 and 3 are taken account of when we delay on the outut signal of neural networks. V. HE SRUCURE OF HE MOFE ELMAN NEURAL NEWORK he modified Elman network adds the self-feedback connection in the context layer. t has been roven that the modified Elman network is efficient to identify the higher order system dynamically [5]. he self-feedback connections at the context layer simlify the dimensions of the network, and then the nonlinear model with higher recision could be acquired, using fewer neurons. So the modified Elman network is fit to the embed control system. A. Algorithm of the Modified Elman network he modified Elman network can be regarded as a BP network with local mnemon and local feedback connection; its basic structure is illustrated as Figure 10. he modified Elman network consists of the inut layer, the hidden layer, the context layer and the outut layer. Each node in the hidden layer links with a corresonding node in the context layer, resectively. he nodes between the inut layer and the hidden layer link with each other through adjustable weights, so do the nods in the hidden layer and the outut layer, the context layer and the hidden layer. here are self-feedback connections with the fixed gain β at the neurons in the context layer. Suose i denotes the nodes in the inut layer, i = 1, j = 1, layer, 1,M, j denotes the nodes in the hidden layer,,l, and denotes the nodes in the context =,,L, then the connection weight from i to j can be denoted with w 1,ij from j to outut layer can be denoted with, the connection weight w 2, j, and the connection weight from to j can be denoted with w 3,j ; net_h j denotes the inut to node j in the hidden layer, o_h j denotes the outut of node j in the hidden layer, net_c denotes the inut to node in the context layer, layer, o_c denotes the outut of node in the hidden hidden layer, f denotes the function of node j in the f usually adots the Sigmoid function: 1 1 -x f x = / ( +e ). ( xi denotes the inut to the modified Elman network, i = 1,,M. he inut consists of three nodes in this work, which corresonds to ( the throttle osition duration signal, resectively; y, denotes the outut from the network, is the exected air fuel ratio signal without the transortation delay. x inut datum and ( i y ( outut datum according to the samling sequence. M, N, L denotes the number of the nodes in the inut layer, the outut layer and the hidden layer, resectively. he number of the nodes in the context layer is equal to that of the hidden layer. Mathematically, the network can be exressed with the following equations: o_c = β o_c +net_c j = 1,,L ( (k-1) ( j j j M L ( ( ( j 1,ij j 3,ij i i=1 i=1 1 net_h = w x + w o_c ( ( o _ h f ( net _ h ) j = 1,,L j y ( L i1 w j 2, i o _ h ( k ) i j =,,L (16) (14) (15) (17) Herein, the inut to the context layer at samling time is the corresonding node outut of the hidden layer at samling time, ( (k-1) net_c i = o_hi (18) n this work, the modified Elman network was trained using the gradient descent method. he objective function of the network can be exressed as follows, 1 ( ( 2 E = (Y - y ) 2 (19)

10 JOURNAL OF SOFWARE, VOL. 5, NO. 1, JANUARY is the exected outut of the outut layer at samling time, it comes from the outut signal of the UEGO sensor. transortation delay. he target value of the air fuel ratio can be ket within the range of ±3% in the transient. Figure 12. he controlled results of air fuel ratio based on neural network Figure11. he simulation model of air fuel ratio control based on neural network he weight is amended every samling, and what we have used is the gradient of instantaneous error, which is different from the total gradient of multi-samling study, so the learning rate of networks should be smaller. n this way their effects will be not so different [13,14]. B. Air fuel ratio control based on neural network he simulation model of air fuel ratio control based on the neural network is illustrated as Figure 11. he outut of the UEGO sensor was used to suervise the trained results of the network, so that the air fuel ratio signal without transortation delay can be redicted online. he modified Elman algorithm for redicting the signal without transfer delay was rogrammed with S-functions. When the throttle changed suddenly, the maladjustment of the air fuel ratio decreased greatly, the controlled error can be ket within the range of ±3%, illustrated as Figure 12. V. CONCLUSONS he model of a single cylinder electronic controlled engine was set u in this aer by using G-Power, the simulation model in air fuel ratio control using G- Power/Simulink was set u as well. On basis of this, the air fuel ratio controlled strategy that combined the modified neural network and the traditional P algorithm was studied. he results of simulation exeriments showed that it is feasible to validate the control strategy by using the simulation model of air fuel ratio control using G-Power/Simulink.he air fuel ratio transortation delay existing in the gasoline engine causes the bad controlled quality in the transient state.c combined control strategy can achieve the real-time rediction of the air fuel ratio signal without the Bench test for the best duty-cycle calibration of LPG EF under steady condition is conducted in bench testing environment. According to the result can we get the cycle ulse sectrum figure on the relationshi among throttle angle- seed-duty-cycle. Also the bench test for the calibration, which tests transmission delay time of airfuel ratio, is conducted in the transition condition, and from this test can we get the ulse sectrum figure on the relationshi among seed-load-air-fuel ratio. n the simulation and bench test latform these exeriments are carried out as follow: air-fuel ratio oenloo exeriment of the engine injected by electronically controlled LPG under transition condition, the controlling exeriment of P closed-loo control algorithm for airfuel ratio under transition condition combined with the ulse sectrum, and the exeriment of the closed-loo neurological network algorithm and P control algorithm on air-fuel ratio under transition condition combined with ulse sectrum. REFERENCES [1] Y. H. an, S. Mehrdad. Neural-networks-based nonlinear dynamic modeling for automotive engines [J]. Neurocomuting. 2000, 30(2): [2] Shinichiro Horiuchi,oshiyuki amatsukuri, Shinya Nohtomi. An automotive lateral controller based on generalized redictive control theory, JSAE Review.2000,21:53-59 [3] Chris Manzie, Marimuthu Palaniswami and Harry Watson. A novel aroach to fuel njection using a radial basis network. EEE. 1998: [4] Yao Jubiao. Fault diagnosis of Electric-hydraulic servo valves base on road roller[j]. Chinese hydraulics & Pneumatics. 2008,206(10): [5] CERCÓS Javier Nevot,CASN Joseba Quevedo. Air-fuel ratio control of a gasoline engine by means of a recurrent neural network with differentiated inut[j]. ntegrated Comuter-Aided Engineering, 2001(8):

11 64 JOURNAL OF SOFWARE, VOL. 5, NO. 1, JANUARY 2009 [6] Yao Jubiao, Zhou asen, Wu Bin. Alication on the road test of UA302/H ata Card based on auxiliary [J]. China: Microcomuter nformation. 2008, 31: [7] Liu Yabin, Liu Xuetian, et al. Forecast of gas line status at real time based on modified Elman neural network[j]. China: Joumalof Systems Engineering,2003,18(5): [8] Yao Jubiao. Research on transient Air Fuel Ratio control of gasoline engines [C].2009 nternational Forum on nformation echnology and Alications.2009,Volume1: [9] Yao Jubiao, Wu Bin, Zhou asen.simulation Control on transient Air Fuel Ratio of electronic controlled LPG engine Using G-Power/Simulink[J]. Journal of Beijing University of echnology(social Sciences Edition).2009,35(03) [10] Richard K Stobart,Bernard J Challen and Rob Bowyer. Electronic Control-Breedig New Engines.SAE Paer No [11] Huang Guanli, Wang Hui,et al. Research on drifting of ositioning based on temoral series[j].china: Comuter Engineering and Alications, 2008,44(31): [12] MANZE C,PALANSWAM M,RALPH, et al. Model redictive control of a fuel injection system with a radial basis function network observer [J]. Journal of ynamic Systems Measurement and Control. 2002, 124,(4): [13] akeshi akiyamaa, Eisuke Shiomib, Shigeyuki Moritaa. Air fuel ratio control system using ulse width and amlitude modulation at transient state. JSAE Review.2001,22: [14] Andrew Mills. Controlling the sensitivity of otical oxygen sensors. Sensors and Actuators B.1998,51:60-68 Yao Jubiao,.E. in thermal engineering, Beijing University of echnology, China,2009. r. Yao is currently an Associate Professor at Beijing Vocational College of Electronic Science. He leads the teaching deartments as dean. He ublished widely. He has ublished academic articles in Journal of Beijing University of echnology(social Sciences Edition),Microcomuter nformation and Chinese Hydraulics & Pneumatics etc. His research interests include thermal engineering,electronic auto. echnology and ollution control. Jiang Keyu, Ph.. in Comuter Science, Arizona State University, USA, r. Jiang is currently an Associate Professor at Fort Hays State University where he leads the nformation Assurance rogram within the nformatics eartment and National Center of Academic Excellence in nformation Assurance (NCAEA). He ublished widely. He has ublished academic articles in Journal of Comuting Sciences in Colleges, Journal of Business and Leadershi, and Academic Persective, etc. His research interests include information assurance, distributed system comuting and organizational management. Yao Chaoyi,B.E. in civil engineering, Southwest Jiaotong University, China,2009. He is currently a rofessional student with master's degree of Southwest Jiaotong University. His research interests include civil engineering,comuting and information management.