Latest revision as of 13:15, 8 June 2020

Abstract

Many countries are implementing new dam safety regulations that often include more restrictive standards. This, together with the increasing average age of dams, results in a greater need for dam control and maintenance works. The advances in information and communications technologies improved the performance of dam monitoring systems, so a large amount of information on the dam behaviour can be collected. This has led to the use of more powerful tools for its analysis, many of which were first developed in the field of machine learning (e. g. neural networks). They offer some advantages over the conventional statistical methods. However, their capacity for early detection of anomalies has seldom been studied. As a result, they are far from being fully accepted by practitioners, whose analyses are often restricted to the interpretation of simple plots of time series data, together with basic statistical models. The present work describes a methodology for anomaly detection in dam behaviour, with the following features: a) The prediction model is based on boosted regression trees (BRTs). b) Causal and auto-regressive models are combined to detect different types of anomalies. c) It is checked whether the values of the external variables fall within the range of the training data. The performance of the proposed methodology was assessed through its application to a test case corresponding to an actual 100-m height arch dam, in operation since 1980. Artificial data were generated by means of a finite element model. Different anomalies were later added in order to test the anomaly detection capability. The method can be applied to other response variables and dam typologies, due to the great flexibility of BRTs, which automatically select the most relevant inputs.