Search

The reliability of gas turbine (GT) health state monitoring and forecasting depends on the quality of sensor measurements directly taken from the unit. Outlier detection techniques have acquired a major importance, as they are capable of removing anomalous measurements and improve data quality. To this purpose, statistical parametric methodologies are widely employed thanks to the limited knowledge of the specific unit required to perform the analysis. The backward and forward moving window (BFMW) k-σ methodology proved its effectiveness in a previous study performed by the authors, to also manage dynamic time series, i.e., during a transient. However, the estimators used by the k-σ methodology are usually characterized by low statistical robustness and resistance. This paper aims at evaluating the benefits of implementing robust statistical estimators for the BFMW framework. Three different approaches are considered in this paper. The first methodology, k-MAD, replaces mean and standard deviation (SD) of the k-σ methodology with median and mean absolute deviation (MAD), respectively. The second methodology, σ-MAD, is a novel hybrid scheme combining the k-σ and the k-MAD methodologies for the backward and the forward windows, respectively. Finally, the biweight methodology implements biweight mean and biweight SD as location and dispersion estimators. First, the parameters of these methodologies are tuned and the respective performance is compared by means of simulated data. Different scenarios are considered to evaluate statistical efficiency, robustness, and resistance. Subsequently, the performance of these methodologies is further investigated by injecting outliers in field datasets taken on selected Siemens GTs. Results prove that all the investigated methodologies are suitable for outlier identification. Advantages and drawbacks of each methodology allow the identification of different scenarios in which their application can be most effective. [ABSTRACT FROM AUTHOR]

Sensor fault detection and classification is a key challenge for machine monitoring and diagnostics, since raw data cleaning represents a key process in the gas turbine industry. To this end, this paper presents a comprehensive approach for detection, classification, and integrated diagnostics of gas turbine sensors (named DCIDS), which was previously developed by the authors and has been substantially improved and validated by means of field data. For a single sensor or redundant/correlated sensors, the improved diagnostic tool, called improved-DCIDS (I-DCIDS), can identify seven classes of faults, i.e., out of range, stuck signal, dithering, standard deviation, trend coherence, spike, and bias. First, this paper details the I-DCIDS methodology for sensor fault detection and classification. The methodology uses basic mathematical laws that require some user-defined configuration parameters, i.e., acceptability thresholds and windows of observation. Second, a sensitivity analysis is carried out on I-DCIDS parameters to derive some rules of thumb about their optimal setting. The sensitivity analysis is performed on four heterogeneous and challenging datasets with redundant sensors acquired from Siemens gas turbines (GTs). The results demonstrate the diagnostic capability of the I-DCIDS approach in a real-world scenario. Moreover, the methodology proves to be suitable for all types of datasets and physical quantities and, thanks to its optimal tuning, can also identify the exact time point of fault onset. [ABSTRACT FROM AUTHOR]

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 field data taken during three years of operation of two fleets of 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. [ABSTRACT FROM AUTHOR]

At present, the challenges related to energy market force gas turbine owners to improve the reliability and availability of gas turbine engines, especially in the ever competitive Oil and Gas sector. Gas turbine trip leads to business interruption and also reduces equipment remaining useful life. Thus, the identification of symptoms of trips allows the prediction of their occurrence and avoids further damages and costs. Gas turbine transients are tracked by gas turbine operators while they occur, but a database including the complete details of past events for many fleets of engines is not always available. Therefore, a methodology aimed at classifying transients into clusters that identify the type of event (e.g., normal shutdown or trip) is required. Clustering is a data mining technique that addresses the scope of partitioning multivariate time series (MTS) into a given number of homogeneous and separated groups. Thus, the multivariate time series belonging to the same cluster are expected to be very similar to each other. This paper presents a structured methodology composed of a subsequent matching algorithm, a featured-based clustering approach exploiting the unsupervised fuzzy C-means algorithm and a procedure that assigns a label to each cluster for classification purposes. The methodology is applied to a real-word case-study that includes transients acquired from a fleet of Siemens gas turbines in operation during 3 years. The results obtained by using heterogeneous datasets including six measured variables allowed values of Precision, Recall and Accuracy higher than 90% in almost all cases. [ABSTRACT FROM AUTHOR]

Nowadays, gas turbines (GTs) are equipped with an increasing number of sensors, of which the acquired data are used for monitoring and diagnostic purposes. Therefore, anomaly detection in sensor time series is a crucial aspect for raw data cleaning, in order to identify accurate and reliable data. To this purpose, a novel methodology based on Bayesian hierarchical models (BHMs) is proposed in this paper. The final aim is the exploitation of information held by a pool of observations from redundant sensors as knowledge base to generate statistically consistent measurements according to input data. In this manner, it is possible to simulate a "virtual" healthy sensor, also known as digital twin, to be used for sensor fault identification. The capability of the novel methodology based on BHM is assessed by using field data with two types of implanted faults, i.e., spikes and bias faults. The analyses consider different numbers of faulty sensors within the pool and different fault magnitudes. In this manner, different levels of fault severity are investigated. The results demonstrate that the new approach is successful in most fault scenarios for both spike and bias faults and provide guidelines to tune the detection criterion based on the morphology of the available data. [ABSTRACT FROM AUTHOR]

This paper addresses the challenge of forecasting the future values of gas turbine measureable quantities. The final aim is the simulation of "virtual sensors" capable of producing statistically coherent measurements aimed at replacing anomalous observations discarded from the time series. Among the different available approaches, the Bayesian forecasting method (BFM) adopted in this paper uses the information held by a pool of observations as knowledge base to forecast the values at a future state. The BFM algorithm is applied in this paper to Siemens field data to assess its prediction capability, by considering two different approaches, i.e., single-step prediction (SSP) and multistep prediction (MSP). While SSP predicts the next observation by using true data as base of knowledge, MSP uses previously predicted data as base of knowledge to perform the prediction of future time steps. The results show that BFM single-step average prediction error can be very low, when filtered field data are analyzed. On the contrary, the average prediction error achieved in case of BFM multistep prediction is remarkably higher. To overcome this issue, the BFM single-step prediction scheme is also applied to clusters of time-wise averaged data. In this manner, an acceptable average prediction error can be achieved by considering clusters composed of 60 observations. [ABSTRACT FROM AUTHOR]