Search Results

Knock control systems are fundamentally stochastic, regulating some aspect of the distribution from which observed knock intensities are drawn. Typically a simple threshold is applied, and the controller regulates the resultant knock event rate. Recent work suggests that the choice of threshold can have a significant impact on closed loop performance, but to date such studies have been performed only in simulation. Rigorous assessment of closed loop performance is also a challenging topic in its own right because response trajectories depend on the random arrival of knock events. The results therefore vary from one experiment to the next, even under identical operating conditions. To address this issue, stochastic simulation methods have been developed which aim to predict the expected statistics of the closed loop response, but again these have not been validated experimentally.

Knock control strategies attempt to optimize the tradeoff between improving torque output and engine efficiency while also regulating knock intensity and protecting the engine from damage. This tradeoff must be made in a stochastic framework since knocking combustion behaves as a random process. This paper therefore examines the marginal and joint statistical properties of both knock intensity, and IMEP under knock limited conditions. Autocorrelation and Pearson chi-squared tests are also used to validate the cyclic independence of the data, or to identify prior cycle effects. The results and joint distribution give insight into the tri-variate relationship between knock intensity, IMEP, and spark advance, providing a foundation for improved knock/IMEP simulation and optimized controller design.

A new knock threshold optimization method is presented based on minimization of the total misclassification error of knocking / non-knocking engine operating conditions. The procedure can be used in conjunction with any knock-event-based controller, but is illustrated on a classical knock control strategy. Initial simulations suggest that the method delivers significant performance improvements with no changes other than a retuning of the controller. However, it is not possible rigorously to evaluate controller performance based on any individual experiment or simulation time history due to the random nature of the knock process. A recently developed stochastic simulation technique is therefore used to compute and compare the statistical properties of the closed loop steady state and transient response characteristics.

This paper collates and summarizes recent advances in knock analysis, simulation and control. The statistical properties of knock intensity and knock events are reviewed showing in particular that knock intensity behaves as an independent random process, and that knock events conform to a binomial distribution. These properties have a significant impact on knock control and simulation. Traditional and recently proposed cumulative-summation-based and Likelihood-based knock control strategies are reviewed and illustrated in this context. Efficient tools for simulating both specific instances of the closed loop time response, and the evolution of the distribution of these responses based on a Markov-like approach, are also briefly reviewed. Finally, it is shown how an optimization of the knock threshold and an associated retuning of the controller parameters can result in significantly improved closed loop performance without any other modification of the control algorithm.

In this paper a new knock control algorithm is developed based on a stochastic interpretation of the knock signal and on a control objective specified as a certain percentage of knocking cycles. Unlike previous ‘stochastic’ knock controllers, the new algorithm does not average or low pass filter the knock intensity signal and the transient response of the controller is consequently much faster. The performance of the new controller is compared in detail with the response of a traditional deterministic controller using a simple but effective knock simulation tool. The results show that the new controller is able to operate at a more advanced mean spark angle and that there is much less cyclic variance about this mean. The transient response to excess knocking events is as fast, or faster, than the conventional controller, though the rate of recovery from overly retarded conditions is slower.

Increasingly, advanced engine management systems incorporate high speed Digital Signal Processing (DSP) units for analyzing high-bandwidth, in-cycle signals such as those obtained from cylinder pressure, or knock sensors. In order to develop, calibrate and test the robustness of these algorithms, it is helpful to work in a simulation environment capable of simulating high-speed in-cycle data and its interaction with the engine management and DSP control strategies. Typically, however, in-cycle simulation is both deterministic and highly computationally intensive so a realistic, cyclically-varying simulation of in-cycle data is hard to generate. In this paper an alternative approach is used, based on initially recording files of high-speed, in-cycle data at different engine conditions. This database is then used to simulate the engine response as the specified engine condition varies, by playing back data from the appropriate files at each time instant.

Look-up tables are widely used in engine management strategies to characterize nonlinear relationships between inputs such as speed and load, and the desired output. However, the calibration of such tables can be time consuming, and is prone to errors due to fluctuating engine measurements, or to small mismatches between the actual test operating condition and the desired operating point in the lookup table grid. In this paper a recursive least-squares identification technique is used to automate the calibration of the table values as the engine is operated over the desired range. The memory and computational requirements of the technique have been optimized so that it can run in real time on a typical engine management system, and the same technique may be used to adapt the table during normal operation if a feedback value is available. The adaptation rate can be adjusted depending on the noise in the available signals.

In this paper, we review previous approaches to oxygen-related OBD strategies and then discuss the use of a new model-based approach together with a distribution-free statistical testing strategy for fault detection. The method is illustrated using experimental pre- and post-catalyst data for which a simplified catalyst-plus-sensor model has been developed. By monitoring the distribution of prediction errors between the ‘healthy’ model output, and the actual catalyst response even small levels of oxygen storage degradation can be detected with a high degree of confidence.

Reversible catalyst deactivation dynamics can have a significant effect on both conversion efficiency and post-catalyst EGO sensor distortion, yet are often ignored in conventional oxygen storage modeling for on-board catalyst control and OBD systems. The aim of the present paper is to include these dynamics in an extended model which exploits the otherwise unfortunate effects of sensor distortion to provide a measure of catalyst deactivation, and hence obtain more accurate predictions of conversion efficiency. Furthermore, by fitting the combined oxygen storage and reversible deactivation model to the data, unbiased estimates of the true post-catalyst AFR can be obtained which are then available for improved catalyst control and diagnostic strategies.

The dynamic behavior of three-way catalysts has significant impact on tailpipe emissions levels, but remains one of the last unknowns in the overall vehicle emissions model. A simple empirical model (appropriate for use in real-time engine control and on-board diagnostic strategies) has therefore been identified using fast response input / output measurements of the actual process. The model is able to characterize the (significant) dynamic behavior which has recently been observed under rich conditions, as well as the more well known dynamics which arise from oxygen storage. The results therefore compare well with measured responses over a wide range of air / fuel ratio conditions.

A framework for a new approach to modelling the cyclic pressure variations which occur in spark ignition engines, based on stochastic process theory, is discussed. The aim is to of establish a stochastic process model for the entire pressure/time history of cycle-by-cycle variation throughout the combustion period. Three types of model are discussed. In the first two it is possible to incorporate correlation across cycles, arising from “prior cycle effects”. In the third, simpler version, the individual cycles are treated as statistically independent. Through a statistical analysis of some pressure data acquired under typical engine conditions the basic characteristics of the stochastic process representations are illustrated.

A stochastic model for the entire pressure-time history of cycle-by-cycle cylinder pressure variations is obtained by fitting simple parametric models of cylinder pressure development to 506 cycles of continuous experimental data taken at four operating conditions. The cyclic variation is therefore encapsulated in a sequence of cyclically varying model parameters whose statistical properties then complete the stochastic description. Different model forms, (including computationally efficient linearised models), are compared for their degree of fit, and for the ease with which the statistics of the identified parameters can be defined. This approach, which typically accounts for 80-90% of the rms cyclic pressure variation, provides a more complete quantification of the phenomena than previously available, and a basis for simulating statistically identical pressure traces.