Samir Saraswati

Neural networks and fuzzy logic-based spark advance control of SI engines

Research highlights? A combined neural network and fuzzy logic-based control scheme is designed for SA control. ? The controller is designed to maintain LPP of SI engine close to 160 ATDC. ? The controller works in conjunction with RNN model for cylinder pressure identification. ? The controller gives desired performance and is found to be robust. In SI engines, spark advance (SA) needs to be controlled to get Maximum Brake Torque (MBT) timing. Spark advance can be controlled either by open loop or by closed loop controller. The open loop controller requires extensive testing and calibration of engine, to develop look up tables. In closed loop controller, empirical rules relating variables deduced from cylinder pressure are used. One of such empirical rules is to fix location of peak pressure (LPP) at a desired value of the crank angle. In the present work, a combined neural network and fuzzy logic-based control scheme is designed for SA control to get MBT timing. The fuzzy logic controller is designed to maintain LPP of SI engine close to 16? ATDC. The controller works in conjunction with Recurrent Neural Network model for cylinder pressure identification. LPP is estimated from cylinder pressure curve reconstructed using neural network model and is used as feedback signal to fuzzy logic controller. The simulations have been carried out to test the performance of the combined neural network and fuzzy logic-based control strategy. The simulation results show that the proposed strategy can quite satisfactorily control LPP to its desired value.

Reconstruction of cylinder pressure for SI engine using recurrent neural network

Cylinder pressure based engine control systems use variables deduced from cylinder pressure as a feedback input. Monitoring of cylinder pressure is possible through various intrusive and nonintrusive sensors but cost of these sensors limits their use in the engines of on-road vehicles. In the present work, a recurrent neural network (RNN) is proposed which can reconstruct cylinder pressure of spark ignition engine. The network uses instantaneous crankshaft speed and motored pressure as inputs. Initially, parameters of two-zone model are tuned at limited number of experimental points, so that cylinder pressure predicted by model matches to that of experimental results. Further, the tuned model is used to generate large number of training data. Validation has been carried out using experimental as well as simulated pressure trace. It has been found that RNN can reconstruct cylinder pressure with reasonably good accuracy.

Online linearization-based neural predictive control of air–fuel ratio in SI engines with PID feedback correction scheme

Model predictive control (MPC) frequently uses online identification to overcome model mismatch. However, repeated online identification does not suit the real-time controller, due to its heavy computational burden. This work presents a computationally efficient constrained MPC scheme using nonlinear prediction and online linearization based on neural models for controlling air–fuel ratio of spark ignition engine to its stoichiometric value. The neural model for AFR identification has been trained offline. The model mismatch is taken care of by incorporating a PID feedback correction scheme. Quadratic programming using active set method has been applied for nonlinear optimization. The control scheme has been tested on mean value engine model simulations. It has been shown that neural predictive control with online linearization using PID feedback correction gives satisfactory performance and also adapts to the change in engine systems very quickly.

Cycle-by-cycle estimation of cylinder pressure and indicated torque waveform using crankshaft speed fluctuations

In this work, frequency response functions (FRFs) are used for estimation of cycle-by-cycle cylinder pressure and indicated torque waveform, using crankshaft speed fluctuations. The FRFs are mapped as a function of the discrete Fourier transform of engine speed, mean speed and manifold pressure using a multilayer neural network. The accuracy of the model is analysed using some of the parameters derived from the cylinder pressure. These include the indicated mean effective pressure and peak pressure. The load torque on the engine is also estimated using a closed-loop observer. The model is tested on a test rig consisting of single-cylinder engine coupled with an eddy current dynamometer. The results show that the model is suitable for the estimation of cylinder pressure and other variables related to it at the operating points where the cyclic variations are within a driveability limit.

Neural network models for multi-step ahead prediction of air-fuel ratio in SI engines

The non-linear dynamics present in SI engine combined with transport delay, limits the performance of the engine controller. Identifying the air-fuel ratio, few steps in advance can help the engine controller to take care of these. In the present work, various neural network models are evaluated for multi-step ahead prediction of air-fuel ratio. Neural network models are trained and validated using uncorrelated data generated from engine simulations in Matlab/Simulink® environment. It is shown that a neural network autoregressive model with exogenous inputs (NNARX) and a neural network autoregressive moving average model with exogenous input (NNARMAX) are able to predict engine simulations with reasonably good accuracy.

Identification of one-zone heat release parameters for SI engine

Heat release analysis is used for predicting the gross heat release characteristic from the pressure data inside the cylinder. In present work, simulation of one-zone heat release model, which involves large number of physically important parameters, is done in Matlab/Simulink environment. The objective here is to find out that how many of these parameters can be reliably determined using in-cylinder pressure measurements. Levenberg-Marquardt optimisation method is used to minimise a least-square objective function for determining the parameters. The smaller value of the objective function is not enough to validate the quality of parameters estimated; hence, their reliability is determined by studying the logarithmic P-V diagram. A methodology is also proposed to find out the parameters, which cannot be identified reliably. The present approach helps in identifying one-zone heat release parameters systematically and prevents the use of arbitrary values of parameters.

Air path identification of turbocharged diesel engine using RNN

This paper discusses an identification scheme for air path of turbocharged diesel engine equipped with exhaust gas recirculation (EGR) and variable geometry turbocharger (VGT). Recurrent neural networks (RNNs) are used to estimate the mass flow rate of air and the intake manifold pressure of a turbocharged diesel engine. Training and generalization of neural network is performed using data generated from a virtual engine model developed in Matlab/Simulink© environment. The implementation of RNN gives satisfactory result for both steady state and transient conditions. The identification models may found its application in model predictive control scheme and can replace costly physical sensors used to measure air flow rate and intake manifold pressure for air path control of a turbocharged diesel engine.

Identification of flame development and rapid burning angle for SI engine

In this work, a method is suggested to identify flame development and rapid burning angle for SI engine using one zone combustion model. Models of burn angles are developed as a function of engine speed, manifold pressure, spark advance and equivalence ratio. The models are tested on a test rig consisting of single-cylinder engine coupled with eddy current dynamometer. It is shown that models have capability to estimate burning angles with a reasonable accuracy.

Cycle-by-cycle estimation of IMEP and peak pressure using crankshaft speed measurements

Cylinder pressure based control of internal combustion (IC) engine uses variables derived from cylinder pressure trace as a feedback input to the engine control and diagnostic systems. Direct measurement of such variables using cylinder pressure sensor is quite expensive. This paper proposes an indirect method of estimating two of such variables namely peak pressure (PP) and indicated mean effective pressure (IMEP) using crankshaft speed measurements. Discrete Fourier Transformation (DFT) is used to transform crankshaft speed fluctuation in time domain to frequency domain. Real and imaginary parts of frequency domain signal, at different harmonics of engine firing frequency, are used as input to multilayer perceptron (MLP). The output being PP and IMEP. Various combinations of inputs starting from signals at first harmonic to signals at first five harmonics are tested. Training and validation of MLP is done using data generated on a test rig consisting of single cylinder engine with eddy current dynamometer. The results show that the MLP is suitable for estimation of cycle-by-cycle values of PP and IMEP for most of the operating points where cyclic variations are within driveability limits.

Application of neural network for air-fuel ratio identification in spark ignition engine

In the present work, Recurrent Neural Network (RNN) is used for Air-Fuel Ratio (AFR) identification in Spark Ignition (SI) engine. AFR identification is difficult due to nonlinear and dynamic behaviour of SI engines. Delays present in the engine dynamics limits the performance of engine controller. Identifying AFR few steps in advance can help engine controller to take care of these. RNN is trained using data from engine simulations in MATLAB/SIMULINK© environment. Uncorrelated signals were generated for training and generalisation and it has been shown that RNN can predict engine simulations with reasonably good accuracy. RNN discussed can also work as a virtual AFR sensor and it can very well replace costly AFR sensor used in SI engines.