Data-driven identification, classification and update of decision trees for monitoring and diagnostics of wind turbines

Abstract

This paper describes a conceptual framework for the online monitoring of wind turbines (WTs) relying on an object-oriented (OO), event-and-decision tree-driven platform for information processing and propagation. To this end, a WT is viewed as a multilayered system of objects (e.g. structure, controller, actuator, etc.) that are defined on the basis of abstract superclasses, attributed with specific properties and methods. The former generally provides insight about the current state of the respective object, while the latter communicates state information and determines the interaction among objects and events. The term state refers herein to a set of mutually exclusive "positions", which a specific object may reach (e.g. safe, critical, fail, etc.), while an unknown state is also included in order to take into account possible combinations of events that have not been registered during the design phase and would eventually be identified in the decision tree through the real-time telemetry. The envisioned platform is purely probabilistic, e.g. a set of probabilities is initially assigned to all events and updated accordingly, based on actual information extracted from the WT. This information may either be acquired using sensors (through corresponding sensor objects), or may be estimated using appropriate algorithms (through corresponding methods, such as Kalman-based filters). A paradigm of the proposed conceptual framework focuses on the tower substructure of the WT and indicates the potential of the proposed approach, especially in respect to the design of specialized software for monitoring and diagnostics of both new and existing WT installations.