No abstract

Statistical parametric methodologies are widely employed in the analysis of time series of gas turbine (GT) sensor readings. These methodologies identify outliers as a consequence of excessive deviation from a statistical-based model, derived from available observations. Among parametric techniques, the k–σ methodology demonstrates its effectiveness in the analysis of stationary time series. Furthermore, the simplicity and the clarity of this approach justify its direct application to industry. On the other hand, the k–σ methodology usually proves to be unable to adapt to dynamic time series since it identifies observations in a transient as outliers. As this limitation is caused by the nature of the methodology itself, two improved approaches are considered in this paper in addition to the standard k–σ methodology. The two proposed methodologies maintain the same rejection rule of the standard k–σ methodology, but differ in the portions of the time series from which statistical parameters (mean and standard deviation) are inferred. The first approach performs statistical inference by considering all observations prior to the current one, which are assumed reliable, plus a forward window containing a specified number of future observations. The second approach proposed in this paper is based on a moving window scheme. Simulated data are used to tune the parameters of the proposed improved methodologies and to prove their effectiveness in adapting to dynamic time series. The moving window approach is found to be the best on simulated data in terms of true positive rate (TPR), false negative rate (FNR), and false positive rate (FPR). Therefore, the performance of the moving window approach is further assessed toward both different simulated scenarios and field data taken on a GT.

The reliability of gas turbine 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 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 bi-weight methodology implements bi-weight mean and bi-weight standard deviation 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 data sets taken on selected Siemens gas turbines. 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.

Statistical parametric methodologies are widely employed in the analysis of time series of gas turbine sensor readings. These methodologies identify outliers as a consequence of excessive deviation from a statistically-based model, derived from available observations. Among parametric techniques, the k-σ methodology demonstrates its effectiveness in the analysis of stationary time series. Furthermore, the simplicity and the clarity of this approach justify its direct application to industry. On the other hand, the k-σ methodology usually proves to be unable to adapt to dynamic time series, since it identifies observations in a transient as outliers. As this limitation is caused by the nature of the methodology itself, two improved approaches are considered in this paper in addition to the standard k-σ methodology. The two proposed methodologies maintain the same rejection rule of the standard k-σ methodology, but differ in the portions of the time series from which statistical parameters (mean and standard deviation) are inferred. The first approach performs statistical inference by considering all observations prior to the current one, which are assumed reliable, plus a forward window containing a specified number of future observations. The second approach proposed in this paper is based on a moving window scheme. Simulated data are used to tune the parameters of the proposed improved methodologies and to prove their effectiveness in adapting to dynamic time series. The moving window approach is found to be the best on simulated data in terms of True Positive Rate (TPR), False Negative Rate (FNR) and False Positive Rate (FPR). Therefore, the performance of the moving window approach is further assessed towards both different simulated scenarios and field data taken on a gas turbine.

Anomaly detection in sensor time series is a crucial aspect for raw data cleaning in gas turbine industry. In addition to efficiency, a successful methodology for industrial applications should be also characterized by ease of implementation and operation. To this purpose, a comprehensive and straightforward approach for Detection, Classification and Integrated Diagnostics of Gas Turbine Sensors (named DCIDS) is proposed in this paper. The tool consists of two main algorithms, i.e. the Anomaly Detection Algorithm (ADA) and the Anomaly Classification Algorithm (ACA). The ADA identifies anomalies according to three different levels of filtering based on gross physics threshold application, inter-sensor statistical analysis (sensor voting) and single-sensor statistical analysis. Anomalies in the time series are identified by the ADA, together with their characteristics, which are analyzed by the ACA to perform their classification. Fault classes discriminate among anomalies according to their time correlation, magnitude and number of sensors in which an anomaly is contemporarily identified. Results of anomaly identification and classification can subsequently be used for sensor diagnostic purposes. The performance of the tool is assessed in this paper by analyzing two temperature time series with redundant sensors taken on a Siemens gas turbine in operation. The results show that the DICDS is able to identify and classify different types of anomalies. In particular, in the first dataset, two severely incoherent sensors are identified and their anomalies are correctly classified. In the second dataset, the DCIDS tool proves to be capable of identifying and classifying clustered spikes of different magnitudes.