A gas turbine trip is an unplanned shutdown, of which the most relevant consequences are business interruption and a reduction of equipment remaining useful life. Thus, understanding the underlying causes of gas turbine trip would allow predicting its occurrence in order to maximize gas turbine profitability and improve its availability. In the ever competitive Oil & Gas sector, data mining and machine learning are increasingly being employed to support a deeper insight and improved operation of gas turbines. Among the various machine learning tools, Random Forests are an ensemble learning method consisting of an aggregation of decision tree classifiers. This paper presents a novel methodology aimed at exploiting information embedded in the data and develops Random Forest models, aimed at predicting gas turbine trip based on information gathered during a timeframe of historical data acquired from multiple sensors. The novel approach exploits time series segmentation to increase the amount of training data, thus reducing overfitting. First, data are transformed according to a feature engineering methodology developed in a separate work by the same authors. Then, Random Forest models are trained and tested on unseen observations to demonstrate the benefits of the novel approach. The superiority of the novel approach is proved by considering two real-word case-studies, involving filed data taken during three years of operation of two fleets of Siemens gas turbines located in different regions. The novel methodology allows values of Precision, Recall and Accuracy in the range 75–85 %, thus demonstrating the industrial feasibility of the predictive methodology.

A gas turbine trip is an unplanned shutdown, of which the consequences are business interruption and a reduction of equipment remaining useful life. Therefore, detection and identification of symptoms of trips would allow predicting its occurrence, thus avoiding damages and costs. The development of machine learning models able to predict gas turbine trip requires the definition of a set of target data and a procedure of feature engineering that improves machine learning generalization and effectiveness. This paper presents a methodology that focuses on the steps that precede the development of a machine learning model, i.e., data selection and feature engineering, which are the key for a successful predictive model. Data selection is performed by partitioning units into homogeneous groups according to different criteria, e.g., type, region of installation, and operation. A subsequent matching algorithm is applied to rotational speed data of multiple gas turbine units to identify start-ups and shutdowns so that the considered units can be partitioned according to their operation, i.e., base load or peak load. Feature engineering aims at creating features that improve machine learning model accuracy and reliability. First, the Discrete Fourier Transform is used to identify and remove from the time series the seasonal components, i.e., patterns that repeat with a given periodicity. Then, new features are created based on gas turbine domain knowledge in order to capture the complex interactions among system variables and trip occurrence. The outcomes of this paper are the definition of a set of target examples and the identification of a meaningful set of features suitable to develop a machine learning model aimed at predicting gas turbine trip.

With the increasing need for flexible operation (shorter startup time and higher load change rate etc.), clearance monitoring between the rotor and the stationary components in steam turbines is becoming more important. This is because as load change rate increases, minimum radial and axial clearances during operation tend to be smaller due to thermal deformation of steam turbines, and the risk of contact between the rotor and the stationary components becomes higher. This situation has accelerated development of clearance sensors. However, it is still difficult to monitor all possible points of contact only with physical sensors due to limited installation location and short lifetime in high temperature environment. From the above background, we have been developing a virtual clearance monitoring (VCM) technique based on a novel data fusion approach that utilizes both physical and data-driven models. Specifically, a reduced order model (ROM) is used as physical model in order to enable real-time prediction with an accuracy similar to that of finite element analysis (FEA). Then, the prediction error of the physical model is corrected by using a residual model built by machine learning from the past clearance sensor values and the corresponding physical model-based prediction results. As will be explained in this report, this technique has an advantage that the clearances can be predicted in real-time based only on operating data such as steam conditions at inlet and outlet, and some temperatures in the parts not modeled in the ROM. Therefore, the virtual sensor based on this technique can be used as a replacement for the physical sensor after it has failed. Furthermore, this technique can also be used to preliminarily study unsteady clearance behavior for inexperienced operating conditions. This paper describes how to build the ROM from a finite element model for thermal-structural analysis of an entire steam turbine by model order reduction (MOR), and the detail of the VCM technique, and a VCM system installed in a measurement room of a state-of-the-art GTCC power plant manufactured by Mitsubishi Power. In addition, the verification results of the VCM system are presented. In this research, the ROM and the residual model were built using the data obtained from four operations with different start-up modes each other. Then, VCM was performed for 12 operating cases. As a result, this survey revealed the followings: (1) This system is capable of real-time prediction with output intervals of roughly 2 seconds. (2) As for radial clearance prediction error during rotor rotating, the RMSEs and the absolute values of minimum value errors are less than or equal to 7.2 % and 7.0 % respectively relative to an initial radial clearance value during the steam turbine stopping. From the above results, we conclude that this VCM technique based on data fusion approach is effective in terms of computational speed and prediction accuracy. This means that if a physical clearance sensor fails, the radial clearance can be continuously monitored by a virtual clearance sensor with a residual model built using the data when the sensor was working normally. In the future, we plan to further improve the accuracy of this technique through improvement in physical modeling.

A machine learning (ML) approach is developed to predict the effect of blend repairs on airfoil frequency, modal assurance criterion (MAC), and modal displacement vectors. The method is demonstrated on a transonic research rig compressor rotor airfoil. A parametric definition of blend geometry is developed and shown to be capable of encompassing a large range of blend geometry. This blend repair geometry is used to modify the airfoil surface definition and a mesh morphing process transforms the nominal finite element model (FEM) to the repaired configuration. A multi-level full factorial sampling of the blend repair design space provides training data to a Guassian stochastic process (GSP) regressor. The frequency and MAC results create a vector of training data for GSP calibration, but the airfoil mode shapes require further mathematical manipulation to avoid creating GSP models for each nodal displacement. This paper develops a method to significantly reduce blended airfoil mode shape emulation cost by transforming the mode shape training data into a reduced basis space using principal component analysis (PCA). The coefficients of this reduced basis are used to train a GSP that can then predict the values for new blended airfoils. The emulated coefficients are used with the reduced basis vectors in a reconstruction of blended airfoil mode shape. Validation data is computed at a full-factorial design that maximizes the distance from training points. It is found that large variations in modal properties from large blend repairs can be accurately emulated with a reasonable number of training points. The reduced basis approach of mode shape variation is shown to more accurately predict MAC variation when compared to direct MAC emulation. The added benefit of having the full modal displacement field also allows determination of other influences such as tip-timing limits and modal force values.

One of the critical components in Industrial Gas Turbines (IGT) is the turbine blade. Design of turbine blades needs to consider multiple aspects like aerodynamic efficiency, durability, safety and manufacturing, which make the design process sequential and iterative. The sequential nature of these iterations forces a long design cycle time, ranging from a several months to years. Due to the reactionary nature of these iterations, little effort has been made to accumulate data in a manner that allows for deep exploration and understanding of the total design space. This is exemplified in the process of designing the individual components of the IGT resulting in a potential unrealized efficiency. To overcome the aforementioned challenges, we demonstrate a probabilistic inverse design machine learning framework, namely Pro-ML IDeAS, to carry out an explicit inverse design. Pro-ML IDeAS calculates the design explicitly without costly iteration and overcomes the challenges associated with ill-posed inverse problems. In this work the framework will be demonstrated on inverse aerodynamic design of 2D airfoil of turbine blades.

Computational fluid dynamics (CFD) is increasingly used during the design phase of turbomachinery. Reducing the cost of such computations is one of the major challenges in the industrial field. The multi-physics phenomena and the multidisciplinary interactions needed for the design of the engine components are difficult to be faced with the classical design methods. In addition, the interest of manufacturers and operators of turbomachines in the increase in performance when degradation processes occur is growing. In the aeronautical sector degradation is among the most critical issues, as it can lead to the in-flight shutdown of the engine. In a bid to tackle these big problems, new design methods based on approximation techniques have been developed. These techniques are called surrogate models and are currently the most used design methods for the aerodynamic design of aircraft engines. In this work, the assessment of a surrogate surface built for the purpose of optimizing an HPT vane in degrading conditions is performed. Using machine learning and statistical techniques, a sensitivity analysis is conducted in order to reduce the problem dimensions. The results of the sensitivity analysis are used for a study of the surrogate, with the purpose of obtaining design guidelines when deterioration effects are considered in the design phase. The main outcome of this study is a map, that outlines the best design zone defined by the combination of the most influential parameters.

CFD-based design optimization of turbulent flow scenarios is usually computationally expensive due to requirement of high-fidelity simulations. Previous studies prove that one way to reduce computational resource usage is to employ Machine Learning/Surrogate Modeling approaches for intelligent sampling of data points in the design space and is also an active area of research, but lacks enough experimental validation. Such a method has been used to optimize the shape of a U-bend channel for the minimization of pressure drop. U-bends are an integral part of serpentine cooling channels inside gas turbine blades but also contribute to total pressure drop by more than 20%. Reducing this pressure loss can help in more efficient cooling at low pumping power. A ‘U-bend’ or 180-degree bend shape has been used from literature, and a 16-dimensional design space has been created using parametrized spline curves, which creates a variety of shapes inside a given bounding box. A Latin Hypercube Sampling (LHS) was carried out for populating the initial design space with output data from the CFD simulation. After training a surrogate model on this data set, Bayesian updates were used to search for an optimum using an exploration vs exploitation approach. The resulting optimum shape showed that pressure drop was lowered by almost 30%, when compared to the baseline. The aim of this study is to experimentally validate this method using 3D printed models of the baseline and optimum channels respectively. Pressure taps placed across stream-wise locations on these channels helped to create a pressure profile for turbulent flow at a Reynolds number of 17000, for comparison to CFD results.

The present paper presents an enhanced method for multi-disciplinary design and optimization of centrifugal compressors based on Machine Learning (ML) algorithms. The typical approach involves the preliminary design, the geometry parameterization, the generation of aero-mechanical databases and a surrogate-model based optimization. This procedure is able to provide excellent results, but it is time consuming and has to be repeated for each new design. The aim of the proposed procedure is to actively exploit the simulations performed in the past for subsequent designs thanks to the predictive capabilities of the ML surrogate model. A commercial 3D (three dimensional) computational fluid dynamics (CFD) solver for the aerodynamic computations and a commercial finite element code for the mechanical integrity calculations, coupled with scripting modules, have been adopted. Two different compressors, with different geometry and operating conditions, have been designed and two aero-mechanical databases have been developed. Then, these two databases have been joined and have been used for the training and validation of the surrogate model. To assess the performance of this approach, two new compressors have been designed, case 1 with operating conditions between those of the databases used for training and validation and case 2 with operating conditions far above. The use of an optimizer coupled to the prediction of the surrogate model has enabled to define the “best set” of model parameters, in compliance with aero-mechanical objectives and constraints. The accuracy of the ML algorithm forecast has been evaluated through CFD and FEM simulations carried out iteratively on the optimal samples, with new simulations added to the database for further training of the surrogate model. The results have been presented with reference to cases 1 and 2 and highlight all the benefits of the proposed approach.

Machine learning models are becoming an increasingly popular way to exploit data from fluid dynamics simulations. This project investigates how autoencoders and output consolidation can be used to increase the accuracy of machine learning models, by injecting knowledge of the full flow field in the predictions of Quantities of Interest (QoI) used in the optimisation of highly loaded transonic compressor blades. Results show that the accuracy of the predicted QoI can indeed be increased, by using both an appropriate autoencoder and output consolidation. Most significantly, the prediction accuracy is increased in the range of QoI values which is involved in optimisation problems. As a result, a more accurate and faster computational design approach driven by machine learning methods has been demonstrated.

Data driven technologies are revolutionising the engineering sector by providing new ways of performing day to day tasks through the life cycle of a product as it progresses through manufacture, to build, qualification test, field operation and maintenance. Significant increase in data transfer speeds combined with cost effective data storage, and ever-increasing computational power provide the building blocks that enable companies to adopt data driven technologies such as data analytics, IOT and machine learning. Improved business operational efficiency and more responsive customer support provide the incentives for business investment. Digital twins, that leverages these technologies in their various forms to converge physics and data driven models, are therefore being widely adopted. A high-fidelity multi-physics digital twin, HFDT, that digitally replicates a gas turbine as it is built based on part and build data using advanced component and assembly models is introduced. The HFDT, among other benefits enables data driven assessments to be carried out during manufacture and assembly for each turbine allowing these processes to be optimised and the impact of variability or process change to be readily evaluated. On delivery of the turbine and its associated HFDT to the service support team the HFDT supports the evaluation of in-service performance deteriorations, the impact of field interventions and repair and the changes in operating characteristics resulting from overhaul and turbine upgrade. Thus, creating a cradle to grave physics and data driven twin of the gas turbine asset. In this paper, one branch of HFDT using a power turbine module is firstly presented. This involves simultaneous modelling of gas path and solid using high fidelity CFD and FEA which converts the cold geometry to hot running conditions to assess the impact of various manufacturing and build variabilities. It is shown this process can be executed within reasonable time frames enabling creation of HFDT for each turbine during manufacture and assembly and for this to be transferred to the service team for deployment during field operations. Following this, it is shown how data driven technologies are used in conjunction with the HFDT to improve predictions of engine performance from early build information. The example shown, shows how a higher degree of confidence is achieved through the development of an artificial neural network of the compressor tip gap feature and its effect on overall compressor efficiency.

Machine learning and automatized routines for parameter optimization have experienced a surge in development in the past years, mostly caused by the increasing availability of computing capacity. Gradient-free optimization can avoid cumbersome theoretical studies as input parameters are purely adapted based on output data. As no knowledge about the objective function is provided to the algorithms, this approach might reveal unconventional solutions to complex problems that were out of scope of classical solution strategies. In this study, the potential of these optimization methods on thermoacoustic problems is examined. The optimization algorithms are applied to a generic low-order thermoacoustic can-combustor model with several fuel injectors at different locations. We use three optimization algorithms — the well established Downhill Simplex Method, the recently proposed Explorative Gradient Method, and an evolutionary algorithm — to find optimal fuel distributions across the fuel lines while maintaining the amount of consumed fuel constant. The objective is to have minimal pulsation amplitudes. We compare the results and efficiency of the gradient-free algorithms. Additionally, we employ model-based linear stability analysis to calculate the growth rates of the dominant thermoacoustic modes. This allows us to highlight general and thermoacoustic-specific features of the optimization methods and results. The findings of this study show the potential of gradient-free optimization methods on combustor design for tackling thermoacoustic problems, and motivate further research in this direction.

Rain erosion of wind turbine blades represents an interesting topic of study due to its non-negligible impact on annual energy production of the wind farms installed in rainy sites. A considerable amount of recent research works has been oriented to this subject, proposing rain erosion modelling, performance losses prediction, structural issues studies, etc. This work aims to present a new method to predict the damage on a wind turbine blade. The method is applied here to study the effect of different rain conditions and blade coating materials, on the damage produced by the rain over a representative section of a reference 5MW turbine blade operating in normal turbulence wind conditions.

One of the key issues in turbomachinery design is the identification of loss mechanisms and their quantification, both during preliminary design and in all subsequent optimization loops. Over the years, many correlations have been proposed, accounting for different dissipative mechanisms that occur in blade-to-blade passages, such as the development of boundary layers, turbulent wake mixing, shockwaves, and secondary flows or off-design incidence. In recent years, the fan industry started the production of more complex rotor geometries, characterized by sinusoidal leading and trailing edges, mostly to extend stall margin and to reduce noise emissions. Literature still lacks a quantification of the losses introduced by the secondary motions released by serrated leading-edges. In this paper we investigate a design of experiments that entails 76 cases of a 3D flow cascade with NACA 4digit profiles with sinusoidal leading edges to measure losses according to the Lieblein’s approach. The flow field simulated with RANS strategy was investigated using an unsupervised machine learning strategy to classify and isolate the turbulent wake downstream of the cascade with a combination of Principal Component Analysis and Gaussian Mixture clustering. Then a gradient boosting regressor was used to derive the correlation between input parameters and cascade deflection.

Ceramic matrix composites can offer clear potential for a variety of engineering applications where the temperature capabilities of conventional metals are exceeded. Continued mechanical characterisation is essential to gain an understanding of their associated damage and failure mechanisms across a wide range of representative temperatures. The present paper will report ongoing research to characterize the initiation of matrix cracking at room temperature under tensile stress and subsequent damage development under fatigue loading in a SiC f /SiC composite. Imaging and mechanical property data were obtained via in-situ loading within a scanning electron microscope. The temporal nature of damage development was also recorded through the selective employment of acoustic emission. Metrics to describe the spatial distribution of cracks, crack lengths and crack opening displacement under load will be presented. The inspections also provided detailed evidence of the associated crack closure phenomena. The understanding of matrix crack saturation and matrix/fibre interfacial mechanics will be explored, together with the implications for the use of X-ray tomographic inspection of engineering components during service. The potential for these emergent techniques as a basis for future CMC characterization, via automated image recognition and machine learning, will be highlighted.

This work presents an on-design component-level multiple-objective optimization of a small-scaled uncooled cavity-stabilized combustor. Optimization is performed at the maximum power condition of the engine thermodynamic cycle. The CFD simulations are managed by a supervised machine learning algorithm to divide a continuous and deterministic design space into non-dominated Pareto frontier and dominated design points. Steady, compressible three-dimensional simulations are performed using a multi-phase Realizable k-ε RANS and non-adiabatic FPV combustion model. Conjugate heat transfer through the combustor liner is also considered. There are fifteen geometrical input parameters and four objective functions viz. , maximization of combustion efficiency, and minimization of total pressure losses, pattern factor, and critical liner area factor. The baseline combustor design is based on engineering guidelines developed over the past two decades. The small-scale baseline design performs remarkably well. Direct optimization calculations are performed on this baseline design. In terms of Pareto optimality, the baseline design remains in the Pareto frontier throughout the optimization. However, the optimization calculations show improvement from an initial design point population to later iteration design points. The optimization calculations report other non-dominated designs in the Pareto frontier. The Euclidean distance from design points to the utopic point is used to select a “best” and “worst” design point for future fabrication and experimentation. The methodology to perform CFD optimization calculations of a small-scale uncooled combustor is expected to be useful for guiding the design and development of future gas turbine combustors.

Machine learning based predictive models are being extensively applied for predicting combustion properties like derived cetane number (DCN), which is the measure of a fuels ignition quality. In the present work, a comprehensive model was developed using artificial neural networks (ANN) that can predict the DCN of fuels containing a large number of chemical classes like paraffins, iso-paraffins, olefins, naphthenes, aromatics, alcohols, ethers, aldehydes, ketones and esters. Experimental DCN’s of 275 fuels was used as a dataset and the composition of the fuels expressed in the form of twelve functional groups and two structural parameters namely, branching index (BI) and molecular weight were used as the input features for the model. A feed forward neural network with two hidden layers with 40 neurons in each layer was developed using Levenberg-Marquardt algorithm. The developed model was validated with 15% of the data points that were randomly generated and kept aside for validation. A regression coefficient (R 2 ) of 0.99 was observed between the predicted and the experimental values along with an average absolute error of 1.1. The results showed that the developed model was successful in predicting the DCN of fuels and can be applied to pure compounds, blends and real fuels containing diverse chemical functionalities.

At the Turbo Expo 2018: Turbomachinery Conference & Expedition, in Oslo, Norway, an innovative approach for assessing operating and near real-time data from power generating assets with meaningful predictive analytics was presented and discussed. GT2018-75030, entitled; Energy Innovation: A Focus on Power Generation Data Capture & Analytics in a Competitive Market established a challenging objective for the industry: “To advance the notion that the fusion of total plant data, from three primary sources, with the ability to transform, analyze, and act based on integrating subject matter expertise is essential for effectively managing assets for optimum performance and profitability; executing and delivering on the promise of “Big Data” and advanced analytics.” Throughout 2019 and 2020, a team comprised of members from Strategic Power Systems, Inc. ® (SPS), Turbine Logic (TL), and two National Labs; National Energy Technology Laboratory (NETL) and Oak Ridge National Laboratory (ORNL), collaborated on the paper’s hypothesis. The team worked with the support of funding from DOE’s Fossil Energy Program through its HPC4 Materials Program, which provided access to the High-Performance Computing assets at both laboratories. The team brought unique skills, strengths, and capabilities that would serve as the basis for an effective, open, and challenging collaboration. The engineering and data science disciplines that converged on this project provided the back-bone for the unbiased analysis and model building that took place; relying on a unique and up-to-date source of plant operating and design data essential for performing the engineering scope of work. A key objective was to use the data and the modeling to be predictive; to characterize remaining life, expended life, and to determine the “next failure” for critical systems and components. Proof-of-concepts were tested for longer term, data-driven reliability prediction for fleets of power generating assets, near real-time prediction of power plant faults which could lead to imminent failure, and physics-based model prediction of life consumption of critical parts. Each of these pilot scale projects is summarized with key results presented.

This study proposes machine learning models to predict the performance of a multi-stage ammonia-water radial turbine using variable nozzle operation under different operating conditions. A 1.2 MW four-stage ammonia-water radial turbine is firstly designed. Then, the one-dimensional off-design simulation model is developed based on the geometric parameters to mainly evaluate the effects of different nozzle outlet angles and turbine inlet temperatures on the turbine performance. A set of data, which consists of 10,000 training points based on one-dimensional model, is used to train the proposed two high-dimensional model representation (HDMR) methods. The forward HDMR model predicts the mass flow rate, turbine outlet temperature, turbine power and turbine efficiency for any combination of turbine nozzle outlet angle and turbine inlet temperature, while the reverse HDMR model predicts the mass flow rate, turbine outlet temperature, turbine efficiency and turbine nozzle outlet angle for any combination of turbine power and turbine inlet temperature. The two HDMR models are validated using 238 sets of separated test data. The results show that the minimum coefficients of determination (R 2 ) of forward HDMR model and reverse HDMR model are 0.9837 and 0.9953, respectively. The maximum relative errors of two HDMR models are below 1.6822%, so the quality of the proposed machine learning methods is high. The overall performance maps of multi-stage ammonia-water radial turbine under the variable nozzle operation method are constructed based on the reverse HDMR model. The reverse HDMR model is helpful in monitoring the healthy operation state of turbine.

Several image processing methods have been implemented over recent years to assist and partially replace on-site technician visual inspection of both manufactured parts and operational equipments. Convolutional neural networks (CNNs) have seen great success in their ability to both identify and classify anomalies within images, in some cases they do this to a higher degree of accuracy than an expert human. Several parts that are manufactured for various aspects of turbomachinery operation must undergo a visual inspection prior to qualification. Machine learning techniques can streamline these visual inspection processes and increase both efficiency and accuracy of defect detection and classification. The adoption of CNNs to manufactured part inspection can also help to improve manufacturing methods by rapidly retrieving data for overall system improvement. In this work a dataset of images with a variety of surface defects and some without defects will be fed through varying CNN set-ups for the rapid identification and classification of the flaws within the images. This work will examine the techniques used to create CNNs and how they can best be applied to part surface image data, and determine the most accurate and efficient techniques that should be implemented. By combining machine learning with non-destructive evaluation methods component health can be rapidly determined and create a more robust system for manufactured parts and operational equipment evaluation.

This paper covers three contemporary topics in the development and deployment of machine learning based diagnostics for large fleets of industrial machines. First, we address the philosophy of monitoring as to whether anomaly detection versus specific failure classification should be pursued, utilizing published statistics of reliability of industrial machines. Second we address the question of unsupervised versus supervised methods using a simulated example of a typical industrial machine fault, where we apply a number of popular unsupervised and supervised algorithms and directly compare their alerting ability. Lastly, model development and deployment at global scale is discussed, with application to a global fleet of gas turbines. The application includes a framework of neural network models that have been trained to find anomalous behavior for a system of the gas turbine package. The remainder of the paper includes a discussion of the results from the fleet application. Specifically, we discuss the fleet training procedure and hardships incurred in moving from proof of concept designs to full deployment on global production asset monitoring. Selected training models that failed to be of production quality are examined and the source of training error is identified. Throughout, the paper provides lessons learned, broad insights gained, and productionization issues that still need improvement as it relates to development and deployment of machine learning models at the scale of global industrial machine monitoring.