A fault model extension for a geometric fault isolation methodology to detect leakages and sensor faults on engine test beds

Abstract

Analytical model-based methods for fault diagnosis on engine test bench systems remain the subject of intensive research. In some cases, such methods are also being used productively even though their application is mainly limited to the isolation of sensor faults. This paper presents a method for isolating process faults—and leakages in particular—that does not require any direct changes to the analytical models used for residual generation. Based on a geometric isolator, this method generates relations between fault causes and fault symptoms through the definition of fault models using latent variables. The application of the methodology is first explained in detail using a simple example. Taking data from a simulated engine test bench, we then show how different causes of faults, including sensor faults and leakages at different positions, can be correctly isolated.

Introduction

Process monitoring and fault detection have a long tradition. Earlier goals were primarily reliability and safety. Today the focus is on high data quality and better efficiency. For engine test beds, all these points are important. Faults, failures and malfunctions that may have a potentially dangerous effect must be avoided for safety reasons. Faults that do not directly affect the correct functioning of a test bed system, for example measurement errors, are often not detected until postprocessing, yet they should be previously detected online due to the high cost of repeated measurements. The quality of data from the test bed is critical because it forms the basis for hypotheses regarding the engine configuration especially in the area of research.

The topic of fault detection has been intensively studied by many researchers. Methods based on deep knowledge and analytical models have been analyzed (Li, Li, Gu, & Chen, 2020a), and data-driven approaches have also been the focus of interest (Ge, Song, & Gao, 2013). Knowledge of a fault in the data is important but insufficient. Once a fault is detected, information about what caused the fault or which variables are affected is also essential. Only with this information is the user able to take appropriate actions to correct the fault. After fault detection, therefore, a fault isolation step is necessary. Methods including both steps are referred to as fault detection and isolation (FDI) methods.

Model-based FDI approaches use residuals to compare the expected behavior with the actual behavior in order to detect faults. In general, residual generation methods are divided into parity-space, observer-based and parameter estimation-based methods (Frank, 1990). Such methods usually require a high degree of analytical system knowledge (Ceccarelli, 2012, Jahn et al., 2020) for modeling and often require a special residual/observer design adapted to the application (Salehi, Alasty, & Vossoughi, 2014). Engine test beds involve frequent hardware changes, and detailed system knowledge is only generated during the test. This is why a highly flexible method for residual generation is needed that enables fast adaption to hardware changes and allows model-based information to be processed with a minimum of parametrization. In addition, the method should be applicable to all types of models since system information may only be available in the form of either inequality relations for pressures and temperatures or simple, generally valid balance equations (Flohr, 2005, Fritz, 2008). This also results in corresponding requirements for the flexibility and generality of the fault isolation method.

A wide variety of fault isolation approaches exist. Kóscielny and Syfert (2014) pointed out the difficulty of such diagnostic systems and the specific requirements and thus solutions for each application. Application-specific solutions abound. In Keliris, Polycarpou, and Parisini (2015), each networked subsystem is monitored by its own fault detection agent, and further conclusions about the nature of the fault can be drawn from the agents’ decisions. A process fault that occurs in one subsystem can only be detected by its corresponding detection agent. In Zhang, Polycarpou, and Parisini (2008), the fault detection and isolation architecture is based on a bank of nonlinear adaptive estimators, one for each nonlinear process fault in the fault class and one for each sensor or output variable. The set of dynamic features most sensitive to different tool conditions induced by chamber cleaning or different faults was determined using linear discriminant analysis in Bleakie and Djurdjanovic (2013). Changes in these features were statistically quantified as the condition of the tool changed using Gaussian mixture models of the dynamic feature distributions.

Kóscielny and Syfert (2014) indicated that methods for fault isolation can be classified according to what knowledge is available. Isermann (2006) also classified methods according to their knowledge base. He distinguished between heuristic and analytical diagnostic knowledge. The former is expert knowledge gained through the operator’s experience of the process, which can be either a process represented by a graph (Wan et al., 2013) or unique information flow patterns between process variables that are used for fault isolation (Hajihosseini, Salahshoor, & Moshiri, 2014).

Most of the analytical knowledge-based methods are classification schemes. One example of such a scheme is a support vector machine. Jung, Ng, Frisk, and Krysander (2018) combines model-based residuals with incremental anomaly classifiers. Vidal-Puig, Vitale, and Ferrer (2019) showed that besides classification, there are two other types of supervised methods: reconstruction and fault signatures. In the first, process faults and sensor faults are characterized by a direction vector which describes the behavior of the fault. Fault reconstruction is then accomplished by sliding the sample vector as close as possible to the principal component subspace as in Dunia and Joe Qin (1998). Yoon and MacGregor (2001) analyzes angles between the vector of the current fault and those of the known faults, the so-called fault signatures. The approach that determines which of the known failure directions is the closest to the observed residual vector is also referred to as key of fixed direction residuals.

Gertler (1991) investigated fault isolation based on residuals in analytical redundancy approaches. He found that there are two main ways to use residuals for fault isolation: Fixed direction residuals and structured residuals. The former has already been explained. The latter is a bank or hierarchy of detection filters where each filter in the hierarchy is insensitive to a particular fault but sensitive to a different set of faults (Naderi & Khorasani, 2017). Therefore, each fault has a signature that permits its isolation if isolation logic is implemented. Li, Raghavan, and Shah (2003) uses the typical incidence matrix to characterize the isolation logic. Spreitzer and Ballé (2000) used structured residuals with a directed incidence matrix to isolate process faults.

Since unsupervised methods do not use any knowledge at all, the main idea is to find critical variables. One solution is to find the variable that distinguishes between the error-free and the faulty data set, similar to the selection of variables in discriminant analysis. Due to the additional relationship between discriminant analysis and regression analysis, a non-negative garrote-based method for fault isolation is proposed that ranks process variables according to how critical they are to the detected fault (Wang et al., 2020). Another solution is to eliminate each variable backwards (Stork, Veltkamp, & Kowalski, 1997) or to treat it as if it were missing (Liu, Chen, & Yao, 2014) and re-estimate the monitoring statistic. Consequently, the variable that reduces the monitoring statistic the most is considered the cause of the fault since the correct reconstruction removes the effect of the fault.

Another option is to measure the impact of each variable on the current statistic. A common method is contribution plots, whose advanced methods have already been studied. One idea is to combine contribution and reconstruction methods to obtain a reconstruction-based contribution (Alcala & Qin, 2009). In nonlinear processes, fault directions using the gradient of the kernel are an option (Zhang, Zhang, & Lu, 2013). Such deviation contribution plots can also use a reference point under normal operation (Tan & Cao, 2019). For process faults in particular, there is a generalized form of reconstruction-based contribution that identifies the faulty variables and transfers entropy-based causality analysis to diagnose the cause of quality-related faults (Ma, Dong, Peng, & Zhang, 2017).

A method for fault isolation on engine test beds has been presented in Wohlthan, D., Priker, and Wimmer (2020). It is able to process any model-based information regardless of whether the models are formulated in equation or inequality form. However, the algorithm is limited to isolating of sensor faults. Two further typical fault types are actuator faults and process faults (Li, Wang, Wang, Wang, & Duan, 2020b). Special methods have been developed for actuator faults (Zhang, Dahhou, Cabassud, & Li, 2015).

The focus of this paper is on process faults and leakages in particular. Various methods for diagnosing leakages have been developed for concrete applications such as heat exchangers (Habbi, Kinnaert, & Zelmat, 2009), fuel cell vehicles (Tian, Zou, Jin, & Lin, 2021) or turbocharged gasoline engines (Salehi et al., 2014). In general, it is desirable to extend a method suitable for engine test beds so that not only sensor faults but also leakages can be isolated. Therefore, the main contribution of this paper is to modify and extend the geometric sensor fault isolation method presented in Wohlthan et al. (2020) to isolate leakages as well.

The fault isolation scheme we used is briefly summarized in Section 2. The adaptation of the method to process faults is the subject of Section 3. The challenge and objectives of such an extension for concrete use on engine test beds are discussed before the new procedure is presented. While Section 4 provides a simple example of the isolation process, Section 5 presents its implementation in the actual application area, namely engine test beds.