Fault detection and diagnosis of a 12-cylinder trainset diesel engine based on vibration signature analysis and neural network

Abstract

This paper presents a condition monitoring and combustion fault detection technique for a 12-cylinder 588 kW trainset diesel engine based on vibration signature analysis using fast Fourier transform, discrete wavelet transform, and artificial neural network. Most of the conventional fault diagnosis techniques in diesel engines are mainly based on analyzing the difference of vibration signals amplitude in the time domain or frequency spectrum. Unfortunately, for complex engines, the time- or frequency-domain approaches do not provide appropriate features solely. In the present study, vibration signals are captured from both intake manifold and cylinder heads of the engine and were analyzed in time-, frequency-, and time–frequency domains. In addition, experimental data of a 12-cylinder 588 kW diesel engine (of a trainset) are captured and the proposed method is verified via these data. Results show that power spectra of vibration signals in the low-frequency range reliably distinguish between normal and faulty conditions. However, they cannot identify the fault location. Hence, a feature extraction method based on discrete wavelet transform and energy spectrum is proposed. The extracted features from discrete wavelet transform are used as inputs in a neural network for classification purposes according to the location of sensors and faults. The experimental results verified that vibration signals acquired from intake manifold have more potential in fault detection. In addition, the capacity of discrete wavelet transform and artificial neural network in detection and diagnosis of faulty cylinders subjected to the abnormal fuel injection was revealed in a complex diesel engine. Beside condition monitoring of the engine, a two-step fault detection method is proposed, which is more reliable than other one-step methods for complex engines. The average condition monitoring performance is from 93.89% up to 99.17%, based on fault location and sensor placement, and the minimum classification performance is 98.34%.