posted on 2018-06-04, 08:37authored byJian Chen, Robert Bond Randall, Bart Peeters
Artificial Neural Networks (ANNs) have the potential to solve the problem of automated diagnostics of piston slap faults, but the critical issue for the successful application of ANN is the training of the network by a large amount of data in various engine conditions (different speed/load conditions in normal condition, and with different locations/levels of faults). On the other hand, the latest simulation technology provides a useful alternative in that the effect of clearance changes may readily be explored without recourse to cutting metal, in order to create enough training data for the ANNs. In this paper, based on some existing simplified models of piston slap, an advanced multi-body dynamic simulation software was used to simulate piston slap faults with different speeds/loads and clearance conditions. Meanwhile, the simulation models were validated and updated by a series of experiments. Three-stage network systems are proposed to diagnose piston faults: fault detection, fault localisation and fault severity identification. Multi Layer Perceptron (MLP) networks were used in the detection stage and severity/prognosis stage and a Probabilistic Neural Network (PNN) was used to identify which cylinder has faults. Finally, it was demonstrated that the networks trained purely on simulated data can efficiently detect piston slap faults in real tests and identify the location and severity of the faults as well.
Funding
The authors would like to convey special gratitude to the Australian Research Council and LMS International (now Siemens PLM Software) for sponsoring this research under Linkage Project LP0883486.
History
Citation
Mechanical Systems and Signal Processing, 2016, 75, pp. 434-454
Author affiliation
/Organisation/COLLEGE OF SCIENCE AND ENGINEERING/Department of Engineering