Energy-based fault detection and isolation in an industrial steam turbine system

Abstract

System monitoring, especially with the Industry 4.0 push to automate the industry, is becoming more crucial for the successful operation of industrial systems. The effective and early detection and isolation of faults is an important part of both system monitoring and control. This can be a laborious process especially if the system consists of more than one domain of operation i.e. elec-trochemical or electromechanical etc. A method that can overcome this multi-domain complexity is the energy graph-based visualisation (EGBV) method that can be used as a hybrid fault detection and isolation (FDI) technique. This study aims to evaluate this energy-based approach to fault detection and fault isolation applied to a steam turbine system (STS). A static thermodynamic model of the STS situated in Jeanschwalde Germany has been developed in the software package engineering equation solver (EES). The model is based on mass and energy balance equations pertaining to individual components of the system. The modelling knowledge has then been applied to simulate a fault type (FT) using an existing validated model of the STS developed in the software package Ebsilon by engineering staff working at the plant. Four faults have been considered, namely: Solid particle erosion (SPE), leakage of the turbine’s overflow valve, overall ageing or wear of the turbines and pump cavitation. Physical data in the form of temperature, pressure and mass flow rates from the STS have also been obtained for a number of components in the system. The simulations and physical data have been combined, using statistical methods, in an attempt to create time series energy based data for the purposes of FDI. An energy characterisation of the STS has been done and an attributed graph of the STS is composed pertaining to modern graph theory techniques. The attributed graph has been used to construct node signature matrices containing the energy data from the constructed time series. These node signature matrices are utilised in three analytic approaches to FDI. Each of the three approaches uses different aspects of the information encapsulated within the node signature matrices. Ap-proach 1 and 2 are based on graph matching to compose cost matrices with the Heterogeneous Euclidean-Overlap Metric (HEOM) metric. Approach 1 uses a single distance parameter that is indicative of a fault being present in the system or not. Approach 2 utilises eigenvalues of the cost matrices as the mathematical analysis technique for FDI. Approach 3 uses a residual approach to determine whether a fault is present and to isolate one fault from another. The FDI results obtained from the 3 approaches show that approach 2 had the best perform-ance. The detection accuracy of approach 2 was 100% and an isolation accuracy ranging from 19%, for some variations of FT2, to 100% for for FT4. Approach 1 had the lowest accuracy in terms of fault isolation, attaining isolation accuracy of less than 20.4% for all FT’s except FT1 which was isolated with a 100% accuracy. Approach 3 had the worst detection accuracy with moderate isolation, attaining only 72% accuracy in terms of fault detection and an isolation accuracy ranging between 5.2% and 70%. Future research on this topic can entail the development of a general approach to energy-based FDI. The establishment of guidelines for composing and attributed graph of a system will be crucial for such future research. The application of alternative metrics for graph matching and the analysis of these matched graphs using machine learning techniques can also be considered. A comparative study of the effectiveness of these FDI methods compared to traditional methods like Principle Component Analysis (PCA) is also required.