Internal combustion engine diagnostics using vibration simulation

Simulation was used as a viable way of generating data to train Artificial Neural Networks (ANN) to detect, diagnose and classify faults in Internal Combustion engines. Engine faults can be divided into combustion faults such as misfire, and mechanical faults such as piston slap fault and big end bearing knock fault. Experimental vibration signals with combustion faults and mechanical faults were studied first. The torsional vibration of the crankshaft and angular acceleration of the engine block were used to detect engine misfires. It has been observed that the amplitude and phase of the first few harmonics of the cycle frequency are suitable feature vector components when using either torsional vibration or angular acceleration of the engine block. As opposed to the misfires, the vibration signals from mechanical faults are typical second order cyclostationary signals and it has been demonstrated that an appropriate signal processing approach for their diagnosis is envelope analysis of the vibration signals. The kurtogram was applied to select the optimum demodulation band before the envelope analysis. Then, selected amplitude and phase features of the Fourier coefficients of the squared envelope were used for the diagnosis of piston slap and bearing knock faults.The simulation models are based on the thermodynamic and mechanical principles of Internal Combustion engines. As inputs to the simulation models, the inertial properties of the engine components and the parameters of the engine mounts were first measured and/or calculated. Simulation software was found suitable for modelling the effects of misfire, on both crankshaft torsional vibration and block angular accelerations. For mechanical faults such as piston slap fault and bearing knock fault, because the squared envelopes were used for their diagnosis, the simulations only had to produce the correct envelope signals, so the dynamic response did not have to be reproduced exactly; the results were not sensitive to small changes in resonance frequencies.The simulated signals were processed by the same signal processing techniques as the measurements and the selected features from the processed data were used to train the networks. The processed simulated data, with added random variations of the same order as in the measurements, provided sufficient data for the training of the ANNs. Three-stage network systems are proposed to diagnose faults: fault detection, fault localization and fault severity identification. The detection stage and severity/prognosis stage were enacted by MLP (Multi Layer Perceptron) networks and PNNs (Probabilistic Neural Networks) were used to identify which cylinder has faults. Finally, it was demonstrated that the networks trained on simulated data can efficiently detect both the combustion faults and mechanical faults in real tests and identify the location and severity of the faults as well