Application of a data-driven monitoring technique to diagnose air leaks in an automotive diesel engine: a case study
David Antory Electrical Test for Advanced Architectures, International Automotive Research Centre, Warwick Manufacturing Group, University of Warwick, Coventry, CV4 7AL, U.K. (E-mail:
[email protected]; Tel: +44-24-76575441; Fax: +44-24-76575403)
Abstract
This paper presents a case study of the application of a data-driven monitoring technique to diagnose air leaks in an automotive diesel engine. Using measurement signals taken from the sensors/actuators which are present in a modern automotive vehicle, a data-driven diagnostic model is built for condition monitoring purposes. Detailed investigations have shown that measured signals taken from the experimental test-bed often contain redundant information and noise due to the nature of the process. In order to deliver a clear interpretation of these measured signals, they therefore need to undergo a ‘compression’ and an ‘extraction’ stage in the modelling process. It is at this stage that the proposed data-driven monitoring technique plays a significant role by taking only the important information of the original measured signals for fault diagnosis purposes. The status of the engine’s performance is then monitored using this diagnostic model. This condition monitoring process involves two separate stages of fault detection and root-cause diagnosis.
-1-
The effectiveness of this diagnostic model was validated using an experimental automotive 1.9L 4-cylinder diesel engine embedded in a chassis dynamometer in an engine test-bed. Two joint diagnostics plots were used to provide an accurate and sensitive fault detection process. Using the proposed model, small air leaks in the inlet manifold plenum chamber with a diameter size of 2 to 6 mm were accurately detected. Further analysis using contribution to T2 and Q statistics show the effect of these air leaks on fuel consumption. It was later discovered that these air leaks may contribute to emissions fault. In comparison to the existing model-based approaches, the proposed method has several benefits: (i) it makes no simplifying assumptions, as the model is built entirely from the measured signals; (ii) it is simple and straight-forward, (iii) there is no additional hardware required for modelling, (iv) it is a time and cost-efficient way to deliver condition monitoring (i.e. fault diagnosis application), (v) it is capable of pinpointing the root-cause and the effect of the problem, and (vi) it is feasible to be implemented in practice.
Keywords: application, data-driven technique, condition monitoring, diagnosis, air leaks, automotive diesel engine
1. Introduction
Stringent emission regulations have led automotive manufacturers to develop systems which can detect and diagnose any fault which may cause tailpipe emissions to rise above a prescribed threshold. This can be achieved by continuously monitoring the
-2-
automotive data characteristics for any abnormal behaviour. Recently, Mills [1] discussed a way to perform automated analysis of automotive data to oversee vehicle system operations, to automate data capture and analysis, and also to improve the diagnostic process. Such an approach can be viewed as a method for improving the reliability, safety and efficiency of the processes as discussed by Isermann [2] and Gertler [3]. This can also be used as a way to conduct fault detection and identification. Previous work by the author [4] investigated faults in an automotive engine using measurement signals which were available in production engines and excluded the remaining signals which can only be measured in a test bed environment. The work reported in this paper extends previous investigations by using all measurement signals taken from an engine during tests conducted in a laboratory. This additional step allows a complete analysis of the experimental data which may be beneficial in the design, development, manufacturing and service stages of the vehicle lifecycle. A detailed analysis is then performed to demonstrate the detection and diagnosis processes. This paper showed that fault caused by various small leaks (of 2mm, 4mm and 6 mm diameters) in the intake manifold plenum chamber of an TDI 1.9 litre diesel engine can be well detected and diagnosed. The model, built using a data-driven technique named principal component analysis (PCA), performed more accurate condition monitoring of this fault than that achieved by using a conventional physical model (Section 4). The improved performance is especially apparent for the smallest air leak (with a diameter of 2 mm). This paper is organised as follows: the next section describes the data-driven technique, the PCA method, which is followed by a discussion of the experimental data in Section 3. Section 4 discusses the condition monitoring process where the detection
-3-
and diagnosis of various sizes of air leak is explained in detail. Finally Section 5 concludes this paper and discusses the future work.
2. Data-driven technique for condition monitoring
This section discusses the data-driven technique known as principal component analysis (PCA). PCA has gained considerable attention mostly in the field of industrial chemical and semiconductor processes for condition monitoring [5-8]. The technique can be successfully applied to automotive applications [4].
2.1. The PCA method
The different types of signals collected from the process are recorded in a range of different unit scales. Therefore, in PCA, normalisation is an essential first stage to make the variance between one process variable comparable to that of any other [9]. Normalisation can be done by mean-centring or auto-scaling the raw data, the latter is done by dividing the mean-centred data by its standard deviation. The normalised data are then stored in column vectors that form a matrix. PCA identifies a combination of variables that describe major trends in the data set. It relies on an eigenvector decomposition of the covariance or correlation matrix of the process variables [10]. The most important information can then be described using a small number of principal components (PC). PCA is a powerful tool in this respect, for analysing multivariate data sets [11].
-4-
For a given data matrix, X ∈ ℜm× n , in which m samples, which are stored as row vectors, of n (n