Strategies for data-based diesel engine fault diagnostics

Alok A Joshi, Purdue University

Abstract

The objective of this work is to develop simple algorithms for fault detection in diesel engines embedded with many sensors using data-based methods. The motivation to develop an on-board diagnostic tool is the reduction in the costs of troubleshooting, maintenance, and false alarms. On-board diagnostics also detects faults eventually resulting in excessive emissions so that the corrective action can be taken in advance. It helps in the manufacturing environmentally friendly automobiles that comply with government regulations. Two different data-based approaches are proposed to do fault detection through symptom detection. First approach focuses on healthy-faulty discrimination of engine states using classification rules trained with data from the minimum number of engine monitoring sensors. Besides inspecting a single sensor reading most relevant to the symptom detection, a collection of signals can aid in detecting a set of symptoms with more confidence. A new information-theoretic sensor subset selector is designed to select the minimum number of sensors that can classify eight different engine states with acceptable accuracy. It is found that at least eight sensors are required to classify eight different healthy and fault states of the engine to maintain the classification accuracy of at least 85%. The specific eight sensor subset selected by the proposed feature selector combined with a probabilistic neural network classifier achieved the best fault classification performance among a set of feature selectors and classifiers studied in this research. Second approach is based on statistical modeling of nominal behavior of a signal directly related to symptom and developing confidence bounds on the engine signal health for symptom detection. Simple regression models, estimated probability density functions, and nonlinear time series analysis tools are successfully used to model nominal behavior of oil pressure, crankcase pressure, oil filter differential pressure, and intake manifold temperature. These models can achieve a much tighter bound on nominal signal behavior than the existing fixed thresholds used for symptom detection. As a result, health monitors with improved performance for oil circuit, charge-air cooler, and exhaust gas return cooler are developed successfully. There are four significant contributions of this work: (i) design of combinatorial and stepwise information-theoretic sensor subset selectors (ii) development of novel histogram-based sparse matrix approach to estimate information to overcome issues with constraints on memory available for computation (iii) development of overall methodology to do data-based fault detection in systems (iv) application of this methodology to do fault detection in diesel engines and the investigation of such faults causing excessive emissions.

Degree

Ph.D.

Advisors

Meckl, Purdue University.

Subject Area

Statistics|Mechanical engineering

COinS