关于汽车怠速及PID有关的翻译

关于汽车怠速及PID有关的翻译 毕业设计 () 英文翻译 院(系): 机电工程学院 专 业: 机械设计制造及其自动化 学生姓名: 学 号: 指导教师: A neural network model-based observer for idle speed control of ignition in SI engine , Jacek Czarnigowski , Lublin University of Technology, Machine Design, Nadbystrzycka 36, Lublin, Poland , Received 16 April 2008. Revised 19 October 2008. Accepted 11 September 2009. Available online 28 October 2009. , , How to Cite or Link Using DOI Permissions & Reprints Abstract The paper presents an algorithm of idle speed stabilization in the spark ignition automotive engine by means of spark advance control. The algorithm is based on a well-known approach of a model-based adaptive control and uses artificial neural networks. The control algorithm is based on a neural network model observer of the additional effective torque. The additional load is estimated as difference between effective torque, estimated by the neural network observer, and brake torque, estimated on the basis of a linear quadratic model. In that case the additional load is understood as the sum of the alternator brake torque (additional load form electric car equipments) and the momentary and/or permanent changes of the engine’s characteristics. On the basis of estimated values of the additional load, the required value of angular acceleration is determined to make the engine return to the specified speed. This acceleration is achieved by adjusting the spark advance. The required value of spark advance is estimated by means of a neural network model converse to that of the effective torque. The algorithm was experimentally compared with PID and adaptive algorithms in the same test bed. The tests were conducted under sudden change of external load. The proposed algorithm proved to be more effective in terms of control error. Keywords Internal combustion engine; Idle speed; Ignition control; Model-based observer; Spark advance control 1. Introduction Internal combustion engine control algorithms are required to provide stable work of the engine, regardless of the variability of loads, operating conditions and changes of the engine characteristics. In the same time, the market demands also a reduction of toxic emissions, noise and vibrations, and improvement of the systems’ reliability. All these requirements 第 - 1 - 页 impose implementation of new additional systems (e.g. exhaust gases recirculation valve, variable valve timing systems) and more flexible and accurate control algorithms. Energy produced by the engine is also used for supplementary onboard devices, the presence of which determines the quality of the car in terms of the user’s safety and comfort. It is estimated that in the next few years an average demand for electric power will exceed 2 kW (Alternators—Technical Instruction, 1997–1998; Nicastri and Huang, 2000). Such loads consume a considerable share of energy produced by the engine, especially during its work in idle speed, which takes from 15% to 20% of the total engine service life (Wendeker and Czarnigowski, 2000). Switching on (or off) individual electric devices is done discretely, which results in the engine being engaged (or disengaged) with a torque of several Nm. Especially in the case of the idle run, this can results in undesirable changes of the engine speed. Another problem is non-stationarity and non-uniformity of engine work. This is the effect of changes to the engine’s characteristics during its service life. The changes may be explained by changes of operating conditions that escape the control system’s sensors, such as humidity, position above sea level or air pollution, but also by natural wear of the engine, or variability of oil and fuel parameters. The purpose of controlling the engine in idle run is to stabilize the engine speed at a desirable level. Any oscillation in the speed of a crankshaft results in vibrations of the components of the vehicle’s body. These are important from the point of the user’s comfort. Moreover, they hinder control of air–fuel mixture composition. This is due to variability of mixture mass reaching the cylinder caused by changes of the filling time. Key control variables, such as the spark advance and the mixture mass (cylinder filling factor) are used to control the engine speed. In most cases, both of them are used in parallel. If the changes of the additional loads are considerable, the mixture mass to reach the cylinder is being adjusted, e.g. by means of the by-pass valve. If the changes are smaller and quick response is required, control is executed by adjusting spark advance. Most of the currently used idle speed control algorithms are based on a PID controller. There are many publications discussing examples of synthesizing the parameters for such controllers (Badreddine et al., 2001; Hrovat and Sun, 1997; Howell and Best, 2000; Shim et al., 1995). However, these algorithms do not satisfy growing performance requirements, and better solutions are looked for. Algorithms such as PID or H-infinity base on simple linear models of the controlled phenomenon (Badreddine et al., 2001; Howell and Best, 2000; Shim et al., 1995). Therefore, the engine’s operation is brought to a linear function describing either the relationship between the engine speed in the idle run and the spark advance (PID controller), or between the engine speed and the spark advance plus state parameters (LQR and H-infinity algorithms (Shim et al., 1995). As is generally known, the engine’s characteristics are far from being linear (Wendeker and 第 - 2 - 页 Czarnigowski, 2003a). This is particularly visible in the case of the effective torque. To solve control problems arising from this fact, non-linear models are applied. It is argued that the most effective tools for that are artificial neural networks. Once the effective torque value is known, the additional loads can be easily estimated, and consequently control can be exercised. 2. Model-based observer algorithm An original control algorithm developed by the authors is based on the indirect adaptation, with one parameter identified by means of a neural-network observer. The parameter is the additional load on the engine in idle run . This parameter reflects not only the actual alternator load, resulting from engaging an energy-consuming onboard device, but also any momentary changes of the effective torque or braking torque. The parameter is then used by the regulator for calculating a new control value. A flowchart of this approach to control is presented in Fig. 1. Fig. 1. Model-based control algorithm flowchart. The additional load is estimated according to (1) In the above equation, the indicated torque is estimated by means of a neural network model, inputs of which are: engine speed n, intake-manifold pressure MAP and spark advance angle φi. (2) The braking torque is estimated by means of a model described by a quadratic function of engine speed (3) 第 - 3 - 页 The third summand in Eq. (1) is the braking torque due to the engine’s inertia, calculated on the basis of the moment of engine’s inertia J and the reduced engine acceleration . The latter is calculated from the difference between the current engine speed and the speed measured in the previous step. (4) In order to reduce the system’s vulnerability to measurement noise and non-repeatability of the engine’s operation, the acceleration is to be filtered, and the reduced acceleration is calculated using (5) where α(i) is the actual momentary engine acceleration calculated in the i step of calculations, γα the coefficient of the rate of the estimation procedure adaptation, its value 2.4 has been established experimentally. The additional load is used for calculating the value of the required effective torque—sufficient for achieving a required engine speed in the next step of calculations. The value of indicated torque to correct the engine speed is calculated using (6) It is assumed that the braking torque is to change from its value calculated in the previous step, which is due to the change of the engine speed. The authors adopted a simplified method of estimating on the basis of the average initial engine speed n(i) and the required engine speed n0. The method allows also for the braking torque () caused by acceleration from the current speed to the required speed during one step of calculations (i.e. during one control cycle) (7) where n0 is the required engine speed, n(i) the current (measured) engine speed of the i step of calculations, Δt the duration of one step of calculations. In the case of the tested engine, the duration of the calculation step (one control cycle) has been established to be a half of the duration of the crankshaft turn. The last stage of calculations is aimed at finding the value of the control parameter, i.e. the spark advance angle, to achieve the required indicated torque: (8) The function defining the required spark advance was calculated on the basis of successive approximations. A flowchart of these calculations is presented in Fig. 2. 第 - 4 - 页 Fig. 2. Model-based control calculation flowchart. In terms of the control systems design theory, the first part of the algorithm, described by Eqs. (1) and (7), is a neural network-based disturbance observer. The second part – Eq. (8) uses the indicated torque model to determine control value; it is thus an element of a model-based control algorithm. 3. Test stand The experiments were conducted in the laboratory of the Chair of Combustion Engines of the Lublin University of Technology. The test stand (Fig. 3) was divided in two sections. The engine section comprised a brake, the engine’s systems of electric power, fuel supply, cooling water circulation and exhaust disposal, and the engine itself together with its sensors and actuators. The control section housed recording and control systems. The test stand allowed applying variable loads to the engine and stabilizing its speed by means of the SAK-N-760 brake (by VEB Elbtalwerk), equipped with an AMX 231 controller (by Automex). The cooling system, controlled by means of the ADAM 5510 system (by Advantech), stabilized the temperature of the coolant. The fuel consumption was measured by weight by means of AMX 212F scales (by Automex) (Fig. 4). For the measurement of 第 - 5 - 页 pressure in the combustion chamber a piezoelectric sensor (601A type) with 7055B spark plug adapter was used, both manufactured by Kristler. The measured signal was recorded by means of PCL 818HG data acquisition device by Advantech, connected to a PC. Fig. 3. Test bed diagram. Fig. 4. Engine of a POLONEZ 1.5 GLI car in the test bed. The engine was equipped with a camshaft position sensor. It was a MH420-6 optoelectronic converter by Megatron. Its 10-bit converter permitted determining the angle with the graduation of 1024 points per rotation. The object of tests was a standard engine of a POLONEZ 1.5 GLI car, equipped by the manufacturer with a MULTEC control system (one-point injection system with Distributorless Ignition System—DIS). (Table 1). 第 - 6 - 页 Table 1. POLONEZ 1.5 GLI—basic parameters. Engine capacity 1481 cm3 Cylinder diameter 77.0 mm Piston stroke 79.5 mm Compression ratio 9.2 Valves per cylinder 2 Power 57 kW—5300 RPM Torque 115 Nm—2800 RPM Injection system Single point MULTEC ACG TBI 700 Ignition system Distributorless ignition system—DIS The control unit was a universal engine controller AMX 200 CAN designed and constructed at the Lublin University of Technology for purposes of testing control algorithms. The system was designed to control single-point or multi-point injection system engines (up to 4 injector groups) with a DIS ignition system and a by-pass air control system with a stepper motor. The controller processed signals from eight onboard sensors. Installation of a CAN bus enabled the real-time control of the engine by means of a PC. TDC signal sent by DIS activated the measurement of signals from all available sensors. The data obtained that way were then transmitted through the CAN bus to the data transmission servicing module, then coded and transmitted to the PC, where the control algorithm was calculated and control parameters were produced (Fig. 5). On their receipt, the transmission servicing module was decoding and transferring them to the control module. Control parameters were realized by the control module in the next cycle of the engine’s operation. 第 - 7 - 页 Fig. 5. Control system software. The source of the additional load was a set of light bulbs connected in parallel to the engine wiring. A discrete change in the consumption of power resulted in an increased power generation in the alternator as a result of voltage controller operation, i.e. increasing the alternator’s load. The use of bulbs permitted attaching a stable load in a discrete manner, as the power consumed by bulbs did not change significantly while switching the system on and off. At the same time, activation of a full power consumption occurred in a sufficiently short time (shorter than a single computation cycle of the control system). It was determined that the consumption of energy at the level of 100 W corresponded to the increase in the external torque of the engine of approximately 2.5 Nm. 4. Algorithm identification and verification The algorithm required identifying the parameters of the model of the effective torque, brake torque and the moment of the engine’s inertia. The identification of the parameters covered exclusively the idle speed, as it was designed for the synthesis and verification of idle speed control algorithms. The scope of verification was therefore confined to • engine speed range from 600 to 1300 RPM, • manifold pressure range from 30 to 60 kPa, • spark advance range from 0? to 30? before TDC. The experiments covered overall 108 test points. The tests were conducted in a steady state. The coolant temperature was 75?2 ?C, the oil temperature was 50?2 ?C, and the temperature of the surrounding was 25?1 ?C. The air humidity was 50?5% and the air pressure 1004?3 hPa. The indicated torque was calculated on the basis of the combustion chamber pressure as average from 200 cycles, which was assumed to eliminate the 第 - 8 - 页 interferences resulting from the adopted measurement and calculation methods. The data were used for constructing a neural network model of the indicated mean effective pressure as described in Section 4.1. In parallel with the in-cylinder pressure measurements, the braking torque was tested. Using a direct current device facilitated not only braking, but also propelling the engine during the tests. The results were approximated with a polynomial of second degree: (9) where n(i) is the engine speed in RPM; a0=3.80; a1=?0.0084; a2=0.000018. The last parameter to be identified was the engine’s moment of inertia. The test of free coasting was conducted and deceleration measured (Fig. 6). On its basis, the average value of moment of inertia was estimated to be J=0.16 kg m2. Fig. 6. Deceleration on free coasting of the engine. 4.1. Neural network models The effective torque model was constructed on the basis of a MLP BP neural network. The inputs were: the engine speed, the intake-manifold pressure and the injection start angle. The network type was a one-direction with one or two hidden layers and of bi-polar sigmoid activation function (multi-layer preceptor). As the literature on the subject (Kimura et al., 1998; Shi-Wei and Ding-Li, 2005; He and Rutland, 2004) argued that MLP networks were suitable for modeling internal combustion engines, the author decided to use the same network type. The choice of input signals resulted from the analysis of parameters affecting the indicated torque. In order to increase precision of the network, the inputs and output were standardized to take values from the range from 0 to 1. The process of teaching used the back propagation algorithm (Bromnick, 1999) and proceeded off-line on the basis of data obtained during the identification tests. The results of the identification tests used in the teaching process were divided in two sets: a 第 - 9 - 页 teaching one and a testing one, in such a way that they were representative for the whole range of engine operation examined (Table 3). The number of samples was defined at the stage of planning the experiment. The author developed his own C++ program for calculations and teaching the neural network. The reason was the need for compatibility with the software used by the controller. The structure of the network was based on preliminary calculations and a compromise between the accuracy and speed of data processing. A very important factor considered while choosing the network type was the duration of calculations. It had to be shorter than 1.4 ms, i.e. 10? of crank rotation at 1200 RPM. Only under this condition the algorithm was capable of processing data on-line. For the preliminary tests, a set of 10 one-hidden-layer and 15 two-hidden-layers neural networks was used. A set of 70 samples was used as a teaching set for all these networks. Table 2 presents some results of teaching process with their learning errors and calculation durations in ms. The learning factor and moment factor were β=0.015 and L=0.4 accordingly. These values were assumed as by the authors of (Bromnick, 1999; Kimura et al., 1998) and on the basis of preliminary tests. 第 - 10 - 页 Table 2. The neural network learning results for different ANN types. Artificial neural network Learning error for Duration of calculations type testing samples (ms) One hidden layer 3-4-1 0.120 0.73 3-7-1 0.102 0.82 3-10-1 0.091 0.91 3-14-1 0.072 0.99 3-18-1 0.064 1.08 3-19-1 0.058 1.12 3-20-1 0.062 1.16 Two hidden layers 3-4-3-1 0.092 1.21 3-4-4-1 0.083 1.24 3-5-5-1 0.048 1.35 3-5-4-1 0.052 1.31 3-5-8-1 0.042 1.43 3-7-4-1 0.083 1.52 MLP 3-5-4-1 BP was chosen to be the best network type for the main tests as its learning error was the lowest among the networks whose duration of calculations was below the required limit. Table 3 contains results obtained in 200 000 epochs for MLP 3-5-4-1 BP neural network. This corresponded in any case to the moment of completing the teaching process. The correctness of the model was verified on the basis of additional tests. Table 3. The neural network learning results MLP 3-5-4-1 BP. Correlation No. of samples coefficient Teaching Testing 0.094 70 10 Learning error Teaching Testing 0.043 0.052 During the teaching process there was no interchange of samples between the teaching and the testing sets. The learning samples were fed into the network in a predefined order. After feeding each sample into the network, the weights of the network were adjusted. In the case 第 - 11 - 页 of the teaching set, the assessment of the network’s approximation error took place after a whole epoch of calculations. As for the testing set, the error assessment was conducted after a fixed number of epochs (1000). For both teaching and testing variables, the teaching process was continued until a constant value of the defined mean error between the consecutive teaching iterations was obtained (Bromnick, 1999). The results are shown in Table 3 and in Fig. 7 and Fig. 8. Fig. 7. Comparison between the calculated and the measured values of the indicated torque for MLP 3-5-4-1 BP. Fig. 8. Relationship between indicated torque and spark advance at constant pressure in manifold and engine speed. 第 - 12 - 页 5. Application of the algorithm Tests were performed at the required speed of 800 rpm, in the steady state of a warm engine (the temperature of the coolant was 90 ?C, the temperature of the lubricating oil was 92 ? C). The initial value of the spark advance was 10? before TDC. The recording of data started after 40 s from the experiment’s start (Fig. 9). After the next 20 s, there was a discrete increase in the engagement of the crankshaft of approximately 2.5 Nm, and after another several seconds, the additional load was disengaged. After about 30 s, the recording of data was terminated. The first 40 s of the test were aimed at stabilizing the parameters of the adaptive algorithm. To allow for the impact of the non-repeatability of the tests, the experiment was performed three times, and the average of three measurements was calculated. Fig. 9. The course of spark advance control by means of the model-based control algorithm. To assess the performance of the developed algorithm, the authors compared it with a PID and an adaptive algorithm (Wendeker and Czarnigowski, 2003b; Czarnigowski et al., 2005). The adaptive algorithm was a linear model-based control algorithm with continuous 第 - 13 - 页 identification of parameters. The following aggregate indicator of the quality of control w was calculated: (10) where ni is the engine speed in an ith performance, n0 the desirable engine speed, N the number of measurement points taken into account in an analysis. The changes of the load, the control value and the rotational speed of the crankshaft are presented in Fig. 9. The quality of control was experimentally tested under steady and suddenly changing loads. It is worth considering, that the proposed algorithm reacted promptly to the change of the actual additional load. The control error in the case of changing load was minimal—it did not exceed 20 RPM. As Fig. 10 shows, the estimated additional load covered not only the actual additional load (average value 2.5 Nm), but also momentary changes of the engine’s characteristics and other disturbance. 第 - 14 - 页 Fig. 10. Stabilization of engine speed by sudden change of load—comparison of three control algorithms. The engine speed oscillations around the required value were significantly lower than in the case of both PID and adaptive control. This is due to the instant reaction of the controller: the controller uses the correction based not only on the control error but also on the estimated 第 - 15 - 页 value necessary to recover the required speed within the shorter possible time, i.e. one cycle. This particular property makes the algorithm perform so efficiently. However, the proposed algorithm is very sensitive to changes of the additional load and reacts to both real changes of engine load and any momentary disturbance. This results in strong oscillations of estimated additional load values (Fig. 9) and consequently oscillations of spark advance. In the experiment, the cycle to cycle variability of the latter ranged from 0? to 25? before TDC. The sensitivity described above cannot be reduced due to the construction of the algorithm. A reduction of the cycle to cycle variability of spark advance is possible by means of an additional filter at the algorithm’s output. A filter described by Eq. (5) could be applied, but it would slow down the system’s reaction to the real changes of additional load, and thus increase the control error. Analyses of the course of engine speed stabilization, based on the smoothed average from a number of cycles (Figs. 11, 12), confirm the advantage of the control algorithm based on the neural network observer. In the case of increased load (Fig. 11), as well as decreased load (Fig. 12), the engine speed did not vary from the required value by more than 20 RPM. There occurred no overshoot, characteristic for both the PID and adaptive algorithms. Moreover, the time of stabilizing the engine speed was considerably shorter: in both analyzed cases it was only 1.5 s, compared with 5 s required by the remaining algorithms. Fig. 11. Mean time course of stabilization of engine speed by sudden increase of load— comparison of three control algorithms. 第 - 16 - 页 Fig. 12. Mean time course of stabilization of engine speed by sudden decrease of load— comparison of three control algorithms. Fig. 13 compares the performance of the PID, the adaptation (Wedeneker and Czarnigowski, 2003b; Czarnigowski et al., 2005) and the model-based controllers during the same test. The indicator of control quality was calculated on the basis of unsmoothed values, mean of three consecutive tests. The model-based control algorithm proved to be the most precise in terms of keeping the engine speed at the desired level. Fig. 13. The aggregate indicator of control quality w according to the type of algorithm. 第 - 17 - 页 To sum up, the model-based control algorithm demonstrated the lowest deviation from the required value and was the quickest in stabilizing the engine speed without oscillation. This was confirmed by comparing control errors, described by the aggregate indicator of quality w. 6. Conclusions It has been proved that combining a neural network observer with an indirect adaptation based on a precise non-linear model of the engine, reduces the error of the idle speed control significantly. Exploiting the knowledge on the engine operation characteristics to develop a control algorithm facilitates more accurate estimation of changes of the engine operation in terms of variability of additional load and engine parameters. Therefore, the controller can react quicker and more accurate to the changes of the object of control. The deeper the understanding of the engine’s principles of operation, the more precise the regulation. Both PID and adaptive control algorithms determine corrections of the control parameter (spark advance), while the model-based control algorithm offers a direct value of the control parameter that comprises allowances not only for the reaction to the detected change, but also for returning to the required engine speed. However, this property of the model-based algorithm results in considerable fluctuation of the spark advance. Its prompt changes facilitate keeping the engine speed at a required level, but are likely to increase toxic emissions. Thus, the rate of changes should be reasonably reduced. This cannot be done without raising control error. Another drawback of the model-based control algorithm is the necessity of precise identification of the model. 第 - 18 - 页 References Badreddine et al. 2001 Badreddine, B., Zaremba, A., Sun, J., Lin, F., 2001. Active damping of engine idle speed oscillation by applying adaptive PID control. SAE Technical Paper No. 2001-01-0261. Bromnick, 1999 Bromnick, P., 1999. Development of a model predictive controller for engine idle speed using CPower. SAE Technical Paper No. 1999-01-117 Czarnigowski et al. 2005 Czarnigowski, J., Wendeker, M., Jakliński, P., 2005. Idle speed stabilization using competitive adaptation control of by-pass valve in SI engine. SAENA 2005-24-059. He and Rutland, 2004 Y. He, C.J. Rutland Application of artificial neural networks in engine modeling International Journal of Engine Research, 5 (4) (2004), pp. 281–296 Howell and Best, 2000 M.N. Howell, M.C Best On-line PID tuning for engine idle-speed control using continuous action reinforcement learning automata 154 Control Engineering Practice, 8 (2000), pp. 147– Hrovat and Sun, 1997 D. Hrovat, J. Sun Models and control methodologies for IC engine idle speed control design Control Engineering Practice, 5 (8) (1997), pp. 1093–1100 Kimura et al. 1998 R. Kimura, N. Nakai, T. Kishimoto Abnormal sound detection by neural network in the diesel engine Bulletin of the M.E.S.J., 26 (1) (1998) Nicastri and Huang, 2000 Nicastri P.R., Huang H. 42v powernet: providing the vehicle electrical power for the 21st century. SAE Technical Paper No. 2000-01-3050, 2000. 第 - 19 - 页 Shim et al. 1995 D. Shim, P.P. Khargonekar, W. Ribbens Idle speed control for automotive engine Proceedings of the 1995 American Control Conference, Seattle, USA (1995) Shi-Wei and Ding-Li, 2005 Wang Shi-Wei, Yu Ding-Li Adaptive air–fuel ratio control with MLP network International Journal of Automation and Computing, 2 (2) (2005) Wendeker and Czarnigowski, 2000 Wendeker, M., Czarnigowski, J., 2000. Hybrid air/fuel control using the adaptive estimation and neural network. SAE Technical Paper No. 2000-01-1248. Wedeneker and Czarnigowski, 2003a Wedeneker Miros?aw, Czarnigowski Jacek, 2003a. Adaptive control of the idle conditions of spark ignition engine. SAE_NA Technical Paper SAE_NA 2003-01-15. Wendeker and Czarnigowski, 2003b Wendeker Miros?aw, Czarnigowski Jacek, 2003b. Adaptive control of the idle speed. In: ICES2003-646 Proceedings of ICES03 2003 Spring Technical Conference of the ASME Internal Combustion Engine Division Salzburg, Austria, May 11–14 Tel.: +48 815384204; fax: +48 815384200. Copyright ? 2009 Elsevier Ltd. All rights reserved. 第 - 20 - 页 一个用于点火式发动机怠速控制的基于模型的神经网络 观测器 , Jacek Czarnigowski , 卢布林理工大学，机械设计学院，Nadbystrzycka36，卢布林，波兰 , 2008年4月16收到。 2008年10月19日订正。 2009年9月11日接收。 2009 年10月28日在线提供。 , 权限和重印 摘要 本文提出了一种通过提前点火控制的方法，火花点火式汽车发动机怠速稳定的算法。该控制算法是基于附加有效扭矩的神经网络观测器。额外的负荷用来估计由神经网络观测器估计的有效扭矩和在二次线性模型的基础上估计得到的制动力矩之间的区别。在这种情况下，额外负荷被理解为发电机制动转矩(汽车电器设备的附加荷载形式)和发动机性能的瞬时和/或永久改变的总和。 在附加载荷估计值的基础上，角加速度要求值的确定使发动机恢复到指定的速度。这个加速度是通过调整提前点火来实现。所需的提前点火值是通过一个有效转矩的逆神经网络模型的方式来估计。 在相同的测试床上，该算法实验性的和PID和自适应算法做了比较。测试在突然变化的外部负载下进行。该算法被证明能更有效地控制误差。 关键词 内燃机 怠速 点火控制 模型观测器 点火提前控制 1、介绍 内燃机的控制算法需要提供稳定的发动机工作，不论负荷的变异，还是发动机性能的运行状况和变化。同时，市场需求也希望减少有毒气体排放，减少噪音和振动，提高系统的可靠性。所有这些要求强制实施新的额外的系统(例如废气再循环阀，可变气门正时系统)和更灵活的和准确的控制算法。 第 - 21 - 页 发动机生产的能源也被用来辅助设备，它的存在决定了汽车在安全性和用户的舒适性方面的质量。据估计，在未来几年，平均用电需求将超过2千瓦(交流发动机，技术说明书，1997–1998;Nicastri 和Huang，2000)。这种负载消耗由发动机产生的相当份额的能量，尤其是在其怠速工作中，这需要花费15%至20%的总发动机使用寿命(Wendeker和Szarnigowski，2000)。打开(或关闭)个人电子设备是离散的，这导致发动机占用 (或未占用)几个纳米的扭矩。特别是在怠速运行情形下，这可能导致发动机转速的不良变化。 另一个问题是发动机工作的非平稳和非均匀性。这受到在其使用寿命中，发动机性能变化的影响。这些变化可能被解释为摆脱控制系统的传感器的操作条件的改变，如湿度，高海拔或空气污染，而且还受到发动机的自然磨损，或是油和燃料参数变化的影响。 控制发动机怠速运转的目的是使发动机转速稳定在一个理想的水平。任何一个曲轴速度的振荡都会导致车身部分的振动。从用户的舒适出发，这些都是很重要的。此外，这些方面还阻碍了对空气—燃料混合物的组成的控制。这是由于混合料质量达到了气瓶灌装时间变化引起的变异。 关键控制变量，如点火提前角和混合质量(气瓶充装因子)，是用来控制发动机的转速。在大多数情况下，他们都在并行使用。如果是相当可观的额外负荷的变化，达到了气缸正在调整的混合质量，通过采用如旁通阀的方式。如果变化较小并且需要快速反应，通过调整点火提前执行进行控制。 大多数目前使用的怠速控制算法都是基于比例积分微分控制器。有许多出版物讨论合成参数控制器的例子(Badreddine等人，Hrovat 和 Sun，1997;Howell and Best, 2000;Shim等人，1995)。然而，这些算法并不满足日益增长的性能要求，人们正在寻找更好的解决。 像PID或H—无穷的算法，都是基于简单的控制现象的线性模型(Badreddine等人，Hrovat 和 Sun，1997;Howell and Best, 2000;Shim等人，1995)。因此，发动机的运行带来了一个描述要么是发动机的怠速运行和的提前点火(PID控制器)之间的关系的，或是发动机速度和提前点火加状态参数(LQR和H —无穷大算法，Shim等人，1995年)之间关系的线性函数。 众所周知，发动机的特点是远从线性(Wendeker和Czarnigowski，2003年)。这在有有效转矩作用的情况下是特别明显的。为了解决由这一事实所产生的控制问题，非线性模型得到了应用。有人认为，解决这一问题最有效的工具是人工神经网络。一旦有效扭矩值已知，附加负荷可以很容易地得到估计，从而可以进行控制。 2、基于观测器的算法 作者原来开发的控制算法是基于间接自适应，带有一个通过神经网络观测器方式确定的参数。该参数是发动机怠速运转的额外负荷。此参数不仅反 第 - 22 - 页 映了由车载设备耗能造成的实际发电机负荷，而且反映了有效力矩或是制动力矩的任何瞬时的变化。该稳压器的参数用于调节计算新的控制值。这一控制方法，如图1所示。 图1 基于模型的控制算法流程图 附加载荷的估计是根据: (1) MiMiMiMi()()()(),,,addibinertia M上述方程中，指示转矩通过一个神经网络模型进行估计，它的输入i 是:发动机转速n，进气歧管压力MAP和提前点火角φi。 (2) MiANNniMAPii()[(),(),(1)],,,ii 制动转矩M通过一个描述发动机速度的二次函数的模型进行估计。 b2(3)Mianiania()()(),,，,， b210 式(1)的第三个加数是由发动机的惯性产生的制动力矩，在转动惯量J和发动机的减加速度状态的基础上计算得到。后者则是从当前的发动机转a 速和在上一步中测量得到的速度之间的差异计算得到。 MiJai()(),,(4) inertia 为了降低系统噪声测量和发动机运行非重复性的脆弱性，加速度被过滤掉，用减加速度计算。 ,,,，，，,，,，ii12(5)()i, ,1，,2 其中，,()i是在第i步计算中发动机的实际的瞬间加速度，是程序适,2应的估计率，其值2.4已经由实验建立。 M附加的负载用来计算所需的有效扭矩值—在下一步计算中实现所add 需的足够的发动机转速。这个值用于指示纠正发动机转速的扭矩的计算。 第 - 23 - 页 MiMiMiMi(1)()()(1)，,，，，(6) iaddbertiain 是从上一步的计算值变化而来，这是由于发动机转速是假定制动力矩Mb 变化的。作者采用了一种在发动机平均初始转速n(i)及所需的发动机转速n0基础上估算的简化方法。该方法同样适用于，在一个计算步骤中(即在Mb 一个控制周期内)的，从当前速度到所需速度过程中引起的制动力矩()Minertia的计算。 nni,(),0MiJ(1)，,(7) inertia30,t 其中，n0是所需的发动机转速，n(i)是在第i步计算中，当前(或是测量)的发动机的转速，Δt 是一步计算的持续时间。 在测试发动机的情况下，计算步骤的持续时间(一个控制周期)已经被估计为曲轴转一圈时间的一半。。 最后阶段计算的目的是寻找控制参数的值，即点火提前角，达到所需的指示扭矩。 ,1,()|[(1)]ifMi,，(8) mii,i 这个定义了所需的点火提前角的函数是在逐次逼近基础上计算得到。这些计算的流程图如图2所示。 第 - 24 - 页 图2 基于模型的控制计算流程图 在控制系统设计理论方面，算法的第一部分，条款(1)到(7)，是一个基于干扰观测器的神经网络。第二部分——式(8)使用了制定的转矩模型来确定控制值;因此，它是一种基于模型的控制算法中的一个元素。 3、试验台 这个实验是在卢布林大学主内燃机实验室中进行的。该试验台(图3)分为两部分。发动机部分包括制动，发动机的电力系统，燃料供应，循环冷却水和废气处理，发动机本身连同它的传感器和执行器。控制部分布置了记录与控制系统。该试验台允许采用可变负荷的发动机和通过SAK-N-760(由VEB Elbtalwerk公司制造)制动的方式稳定速度，配备了AMX231控制器(由Automex公司制造)。冷却系统，通过ADAM 5510系统(由Advantech公司制造)控制，稳定着冷却液的温度。燃料消耗用AMX 212F(由Automex公司制造)(图4)测量重量的方式得到。对于燃烧室压力测量时使用的7055B火花 第 - 25 - 页 塞适配器(601A型)与压电式传感器，都是由Kristler生产。被测信号由Advantech 公司生产的PCL818HG数据采集设备，连接到电脑。 图3 测试图 图四 测试中的POLONEZ牌1.5升排量豪华轿车发动机 该发动机配备有凸轮轴位置传感器。这是一个由Megatron 生产的MH420-6型的光电转换器。其10位转换器允许确定1024度每转的角度的确定。 该测试对象是一个POLONEZ牌1.5升排量豪华轿车的发动机，制造 第 - 26 - 页 商配备了Multec控制系统(单点喷射系统，无分点火系统的DIS)。(1)。 表1 POLONEZ牌1.5升排量豪华轿车的基本参数 发动机容量 1481立方厘米 气缸直径 77.0毫米 活塞冲程 79.5毫米 压缩比 9.2 每汽缸阀门 2 功率 57千瓦，每分钟5300转 扭矩 115 NM，每分钟2800 转 喷射系统 单点MULTEC ACG TBI 700 点火系统 直接点火系统——DIS 数字化信息系统发送的时间数字转换器信号激活了从所有可用的传感器测量得到的信号。通过这样的方式得到的信号，经过控制器局域网总线的数据传输服务模块，然后进行编码和传输到电脑，在这里，控制算法进行 )。在数据接受的过程中，传输服务模块进了计算，产生了控制参数。(图5 行解码并将他们转移到控制模块。控制参数，在发动机运行的下一个周期中得到了实现。 图5 控制系统软件 附加载荷的来源是一组并行连接到发动机接线的灯泡。一个功耗的离散变化会导致发电机发电量的增加，作为电压控制器操作的结果，即增加发电机的负荷。灯泡的使用允许将负载稳定在一个离散的方式，在系统的切换关闭中，灯泡消耗的功率并不明显。同时，全功率消耗的激活发生在足够短的时间内(小于一个控制系统的计算周期)。经确定，能源消费水平在100瓦，相当于增加了约2.5牛?米的扭矩。 第 - 27 - 页 4、算法的识别与验证 该算法需要确定有限扭矩，制动力矩和发动机的惯性矩的参数模型。参数的确定包括专门的空转速度，因为它是专为怠速的合成与验证而设计的怠速控制算法。因此，验证的范围仅限于 •发动机转速范围从600到1300转， •歧管压力范围从30到60， •点火提前角的范围从上止点前0?到30?。 实验整体覆盖108个测试点。本试验是在一个稳定状态进行。冷却液温度为75?2?C时，油的温度为50?2?C，周围的温度为25?1?C。空气湿度为50?5,，气压为1004?3百帕。指定扭矩是在平均200周期的燃烧室压力的基础上计算的，这是假定消除了通过测量和计算方法而造成的干扰。这些数据被用来构建如4.1节所描述的表示平均有效压力的神经网络模型。结果与一个二次多项式近似: 2 (9)Mianiania()()(),,，,，b210 其中，,(,)是发动机每分钟的转速;a0=3.80; a1=?0.0084; a2=0.000018。 最后一个要确定的参数是发动机的转动惯量。该试验进行了自由滑行和减速的测量(图6)。在此基础上，惯性矩的平均值估计为,=0.16公斤平方米。 图6 发动机自由滑行的减速阶段 4.1 神经网络模型 有效转矩模型是在一个多层感知器神经网络基础上构造的。输入是:发动机转速，进气歧管压力和注射的起始角度。网络类型有着一个或两个隐藏 第 - 28 - 页 层，是多层变参数,函数(多层指导)。因为关于这个主题的文献(Kimura等人， 1998; Shi-Wei and Ding-Li, 2005; He and Rutland, 2004)认为，神经网络建模适合内燃机，作者决定使用相同的网络类型。 影响指示扭矩参数的分析导致了输入信号的选择。为了提高网络精度，输入和输出标准化值的范围从0到1。教学过程采用反向传播算法(Bromnick，1999)并在鉴定试验数据的基础上进行离线试验。 在教学过程中所使用的鉴定试验的结果被分为两类:教学和测试，以这样一种方式，这些结果代表了整个系列发动机的运行状况审查范围(见表3)。样本数的定义是在规划实验阶段进行。 笔者开发了自己的C + +程序用来进行神经网络的计算和教学。原因是需要兼容所使用控制器的软件。 网络的结构是根据初步计算，并且兼顾了数据处理的准确性和速度。在选择网络类型的过程中，一个非常重要的考虑因素是计算的期限。它必须是短于1.4毫秒，即曲轴转速在每分钟1200转时，转过10?的时间。只有在这种情况下，该算法才能够处理实时数据。 在初步测试中，使用了一组有10个一层隐藏层和15两层隐藏层的神经网络。一组70个样本被用来作为所有这些网络的教学组。表2列出了一些成果，这些成果是教学过程的学习错误，并以毫秒为单位计算工期。学习系数和力矩系数分别对应是β= 0.015和L = 0.4。这些值均是由作者(Bromnick, 1999年; Kimura 等人, 1998年)假定，并是在初步测试的基础上得到。 表2 针对不同网络类型的神经网络的学习结果 人工神经网络 测试样本的学习错误 计算的期限时间(毫秒) 一个隐藏层 3-4-1 0.120 0.73 3-7-1 0.102 0.82 3-10-1 0.091 0.91 3-14-1 0.072 0.99 3-18-1 0.064 1.08 3-19-1 0.058 1.12 3-20-1 0.062 1.16 两个隐藏层 3-4-3-1 0.092 1.21 3-4-4-1 0.083 1.24 3-5-5-1 0.048 1.35 3-5-4-1 0.052 1.31 3-5-8-1 0.042 1.43 3-7-4-1 0.038 1.52 第 - 29 - 页 多层神经网络3-5-4-1被选为作为主要测试的网络类型，因为在所有的计算期限低于所需期限的网络中，它的学习误差是最低的。 表3包含从200 000代的多层神经网络3-5-4-1中取得的成果。这相当于在任何情况下的时候，完成教学过程。该模型的正确性在额外的测试的基础上得到了验证。 表3多层神经网络3-5-4-1的学习结果 样品号 相关系数 教学 测试 70 10 0.994 学习错误 教学 测试 0.043 0.052 在教学过程中没有教学和测试集之间的样本交换。学习样本输入网络按照预定义的顺序。每个样品送入网络后，网络的权重进行了调整。在教学集的情况下，网络的逼近误差评估发生在整个计算期之后。对于测试集，误差的评估在固定的时期(1000)之后进行。对于教学和测试变量，教学过程一直持续到获得一个连续教学迭代定义的平均误差的恒定值(Bromnick，1999)。结果见表3和图7和图8。 图7 多层神经网络3-5-4-1的指示扭矩的计算和实测值的比较 第 - 30 - 页 图8 在持续压力下，在歧管和发动机速度方面，指示扭矩和提前点火角的关 系 5、该算法的应用 实验在所需的每分钟800转的速度和热发动机的稳定状态(冷却剂的温度是90 ?C，润滑油温度是92 ?C)下进行。提前点火角的初始值是在上止点前10?。数据记录是从实验开始后40秒开始的。(图9)。在接下来的20秒内，曲轴接触离散的增加了约2.5Nm，接下来的数秒后，额外的负载脱离。约30秒后，数据记录终止。实验的第一个40秒钟时间旨在稳定自适应算法的测试。为了消除非重复性试验的影响，实验进行三次，三次测量的平均用于计算。 第 - 31 - 页 图9 通过基于模型的控制算法进行点火提前控制 为了评估发展后的算法的性能，作者用它与一个PID和自适应算法做了比较(Wendeker和Czarnigowski，2003年。Czarnigowski等人，2005年)。自适应算法是一种基于线性模型的有参数连续识别的控制算法。 以下总指标的质量控制计算: N12(10) wnn,,(),0iN,0i 其中，ni是在第i步的发动机转速，n0是理想的发动机转速，n是在一个分析中需要考虑的测量点的数量。 负载的变化，控制值和曲轴转速示于图9。控制质量是在稳态和突然变化的负荷下测试的。 这是值得考虑的，该算法对实际的附加负载的变化作出了迅速反应。在负 第 - 32 - 页 荷变化的情况下控制误差是最小的——不超过每分钟20转。正如图 10所示，估计的附加负载不仅包括实际的附加负荷(平均值为2.5Nm)，而且还包括发动机性能和其他干扰的瞬间变化。 图10 在突变负荷变化和三种控制算法下发动机速度的稳定 发动机转速震荡范围所需的值比在PID和自适应控制的情况下明显低了许多。这是由于控制器的即时反应:控制器的使用的校正方法，不仅基于控制误差，还基于在较短的时间内，即一个周期内，需要恢复所需速度的估计值。这种特殊属性，使得算法以便有效地执行。 然而，上述这个算法对附加负荷的变化和对发动机负荷的实际变化和任何瞬时扰动的反应非常敏感。这样的结果在估计附加的负载值(图9)和提前点火角上有很大的振荡范围。在实验中，后者的变化范围从上指角前0? 第 - 33 - 页 到25?不等。 由于该算法的构建，以上所述的灵敏度不能减少。提前点火角的循环周期变异的的减少有可能通过一个额外的过滤器算法的输出。描述式过滤器(5)可以适用，但它会减慢系统的额外负载的实时变化的反应，从而提高了控制误差。 发动机转速稳定的基础上，从周期数(图11，12)的平滑平均过程的分析，确认基于神经网络观测控制算法的优势。在增加负载(图11)，以及降低负载(图12)的情况下，发动机转速变化没有超过所需的20 转每分钟的速度值。在此过程中，没有发生超调，特点是PID和自适应算法。此外，发动机转速稳定时间大大缩短:在两个分析过的情况下，相比其余算法所需的5秒，它只需要1.5秒。 图11 与三个控制算法相比，突然增加的负载下发动机转速稳定的时间 第 - 34 - 页 图12 与三个控制算法相比，突然减少的负载下发动机转速稳定的时间 图13比较了PID的性能，以及在相同测试中的附加(Wedeneker和Czarnigowski，2003年; Czarnigowski等人，2005年)和基于模型的控制器的性能。质量控制的指标是在非光滑值的基础上计算的，这意味着有三次连续的测试。基于模型的控制算法被证明是在发动机转速保持在理想水平方面最精确的。 图13 根据算法类型的控制质量W的总指标 综上所述，基于模型的控制算法表明偏离所需值的最低值，该算法在使发动机速度稳定无震荡方面是最快的。通过比较控制误差证实了这一点，通过质量W的总指标描述这一点。 第 - 35 - 页 6、结论 已经证明，把神经网络观测器与一个基于精确的非线性模型的发动机的间接适应相结合，这个方法可以有效地降低了怠速控制的错误。在附加负载和发动机参数的变化方面，利用发动机工作特点的知识，制定一个控制算法，有利于更准确的估计发动机运行的变化。因此，该控制器对控制对象的变化可以反应更快，更准确。对发动机的工作原理了解的越全面，调节越精确。PID和自适应控制算法确定了控制参数(提前点火)的改正，而基于模型的控制算法提供了一个直接的控制参数，不仅包括检测到了直接变化的反应，也包括返回到所需的发动机转速。 然而，基于模型算法的属性导致提前点火角相当大的波动。其迅速变化有利于发动机转速保持在所需的水平，但有可能增加有毒物质排放。因此，变动率应合理地降低。这不可能在不增加控制误差情况下完成。基于模型的控制算法的另一个缺点是模型的精确识别的必要性。 第 - 36 - 页