This paper describes a selection of Baker Hughes, a GE company (BHGE) activities to support Gas Turbine (GT) design and operation from simple to more elaborate applications of Machine Learning (ML).

Modern dry low NOx combustors can target very low emissions levels by operating at a lean air/gas ratio. However, ultra-lean combustion is extremely susceptible to thermoacoustic combustion instabilities and Lean Blowout (LBO), which can lead to large pressure oscillations in the combustor and decreased durability of components. Conventional on-board diagnostics embedded in the Unit Control Panel (UCP) of a Gas Turbine, continuously check the health status of the combustion section at a high scan rate and raise alarms when abnormal conditions occur. While ensuring protection and control, UCP control logics may not provide precise indications on the nature of the issue and further troubleshooting, also using specific tools, is typically required. In a changing environment where Industrial Internet of Things (IIoT) is offering a chance to drive productivity and growth, online Monitoring and Diagnostic (M&D) software and services on connected units are becoming strategic to increase asset availability and reliability, as well as reducing maintenance costs. In this paper, we present a hybrid analytic, which combines physics-based and data-driven models, for the detection of Lean Blowout conditions on Gas Turbines equipped with Dry Low NOx multi-can combustion system. Regarding the data-driven model, we face a problem of classification and exploit dimensionality reduction to reduce the number of variables under consideration. During the development, different techniques are tested and benchmarked. The analytic is trained on real LBO events and finally is deployed in a production environment to process incoming on-line data acquired from monitored units. Obtained results are presented.

Several operating parameters for the control and protection of the units are acquired by the control and protection systems used in industrial applications. The use of these parameters in conjunction of physical models, empirical models and transfer functions (that represent digital replicas of the engine) allows for a broader scope of condition monitoring, taking into account the wing to wing process which spans from data acquisition to end user actionable insight. This paper describes 3 specific cases: 1) an algorithm based on the performance model of the overall GT used to monitor the axial compressor degradation and optimize the planned axial compressor water wash of an aero-derivative GT; 2) an analytic based on the flow function physic model used to monitor the clogging of the fuel nozzles in a heavy duty GT and to plan their maintenance; 3) an analytic based on a hybrid model used to monitor the axial thrust acting on a roller bearing of an aero-derivative GT and used to verify the status of the bearing and to plan its maintenance. Moreover, the paper provides details about the evaluation of the measurements, describes the model accuracy and explains how the results obtained are affected by these uncertainties and the methods used to mitigate these uncertainties. In addition, this paper shows a method to aggregate and weigh the monitoring of each single component and its own status into an overall health view.

The monitoring and diagnostics of Industrial systems is increasing in complexity with larger volume of data collected and with many methods and analytics able to correlate data and events. The setup and training of these methods and analytics are one of the impacting factors in the selection of the most appropriate solution to provide an efficient and effective service, that requires the selection of the most suitable data set for training of models with consequent need of time and knowledge. The study and the related experiences proposed in this paper describe a methodology for tracking features, detecting outliers and derive, in a probabilistic way, diagnostic thresholds to be applied by means of hierarchical models that simplify or remove the selection of the proper training dataset by a subject matter expert at any deployment. This method applies to Industrial systems employing a large number of similar machines connected to a remote data center, with the purpose to alert one or more operators when a feature exceeds the healthy distribution. Some relevant use cases are presented for an aeroderivative gas turbine covering also its auxiliary equipment, with deep dive on the hydraulic starting system. The results, in terms of early anomaly detection and reduced model training effort, are compared with traditional monitoring approaches like fixed threshold. Moreover, this study explains the advantages of this probabilistic approach in a business application like the fleet monitoring and diagnostic advanced services.

Given the critical nature of Gas Turbines in most industrial plants, it is a high priority to find ways of reducing maintenance costs and increasing the availability. Quickly detecting and identifying combustion anomalies enables the choice of an appropriate recovery strategy, potentially mitigating the consequences of unscheduled down time and increased maintenance costs. Monitoring the Exhaust Gas Temperature (EGT) profiles is a good means of detecting combustion problems: plugged nozzles and/or combustor and transition piece failures will always result in distorted exhaust gas temperature patterns. However the conventional monitoring systems do not allow robust discrimination between instrumental failures and real gas turbine issues; furthermore weak diagnostic methods can be source of numerous false alarms.In this paper, we investigate the problem of monitoring the combustion chambers of a gas turbine and we attempt to address this issue by introducing a strategy for automatic and efficient patterns recognition by using Machine Learning Classification algorithms. Some historical events have been firstly retrieved and analyzed to discover which features are useful for classification. Based on the observations, two multiclass classification algorithms, one based on logistic regression, the other on Artificial Neural Networks (ANN), have been developed. Finally, real-world datasets have been used to benchmark the performance of the proposed algorithms against a traditional physics-based approach.