Detection of the Onset of Trip Symptoms Embedded in Gas Turbine Operating Data

Abstract

Nowadays, the importance of gas turbine monitoring and diagnostics pushes OEM operators to exploit historical data to search for early indicators of incipient failures. One of the most disrupting events that affect GT operation is gas turbine trip, since its occurrence causes a reduction of equipment remaining useful life as well as revenue loss because of business interruption. Thus, early detection of incipient symptoms of gas turbine trip is crucial to ensure efficient operation and lower operation and maintenance costs. This paper presents a data-driven methodology aimed at investigating and disclosing the onset of trip symptoms. The goal of the methodology is the identification of the time point at which trip symptoms are triggered, by exploring multiple scenarios characterized by different trigger positions. For each scenario, a time window of the same length is considered before and after the trigger time point. For classification purposes, the former is supposed to be representative of normal operation and thus is labeled as “No trip”, whereas the latter is labeled as “Trip”. A Long Short-Term Memory (LSTM) neural network is employed as the classification model. A predictive model is first trained for each scenario and subsequently tested on new observations (i.e., trip events) by considering the whole available timeframe. Finally, trips are clustered into homogeneous groups according to their most likely trigger position, which identifies the time point of onset of trip symptoms. The methodology is applied to two real-world case studies composed of a collection of trips, of which the causes are different, taken from various fleets of Siemens gas turbines. Data collected from multiple sensors during three days of operation before trip occurrence are employed and analyzed. The methodology provides the most likely trigger position for four clusters of trips within the two days before trip occurrence with a confidence in the range 66%–97%.

Influence of the trigger time window on the detection of gas turbine trip

“…The learned models were used to extract statistical rules for predicting trip. Finally, in the study [15], the authors developed a data-driven methodology aimed at disclosing the onset of trip symptoms. A Long Short-Term Memory (LSTM) neural network was employed as the classification model.…”

“…The methodology provided the most likely trigger position for four clusters of trips within the two days before trip occurrence with a confidence in the range 66% -97%. The outcome of the study [15] is the starting point of the analyses conducted in the current paper.…”

“…As anticipated, the current paper stems from the study [15] in order to further investigate the time evolution of trip symptoms and identify the most likely time window within which the symptoms of an incoming trip arise, thus potentially allowing early actions for trip prevention. In this paper, symptoms are intended as perturbances in the data that cannot be detected by means of classical approaches (e.g., physics-based thermodynamic models or frequency-domain vibration analysis) and thus require a data-driven approach to grasp signs embedded in the measured values.…”

“…More in detail, the current paper applies the data-driven methodology presented in [15], which is composed of two steps. The first step makes use of a data labeling approach to model a variation in GT health state due to the trigger of trip symptoms; multiple scenarios characterized by different trigger positions are investigated.…”

“…The main novel contribution of this paper is represented by the systematic analysis of different lengths of the training and testing time window (namely "trigger time window"). First, the data-driven models developed in [15] by employing the longest available time window for each trigger scenario (up to 24 hours) are tested on unseen trips by reducing the trigger time window up to 1 hour in order to search for the most likely time window within which symptoms appear. Second, for each trigger scenario, the training of the data-driven models is also conducted by considering different trigger time windows, i.e., 24, 18, 12, 6 or 1 hour before and after the considered trigger position, while the test is always conducted by considering the longest available time window (up to 24 hours), in order to investigate the influence of the amount of training information on the prediction capability of the data-driven models.…”