Suveillance of large dams aided by automated monitoring systems and machine learning techniques. Contribution from the Portuguese experience

Abstract

One concern of the dam surveillance activities is the assessment of the real structural behaviour, to detect possible malfunctions early. For real-time decision be effective, there must be confidence in the measured data, and it must be possible to interpret the data (through adequate data-based methods) in order to properly assess the structure's behaviour and condition (based on reliable numerical models). In the actual context, the main threats are the inability to detect in a timely manner some scenarios of abnormal structural behaviour that may originate an accident. Therefore, the achievement of the objective of real time structural safety control through the use of automated monitoring systems requires the improvement of the management, validation, archiving and exploitation of information, and the implementation of a quality control process for the measured data.In Portugal, the Concrete Dam Department of the National Laboratory for Civil Engineering (LNEC) has been involved on the safety control of the Portuguese large dams presenting more significant risks since their design and construction stage. One of the more important objectives of LNEC’s activity is the development of a management information systems and several methodologies, aiming the improvement of structural dam safety control in real time that, using automatic data acquisition systems data, allows: i) the evaluation of the quality of instrument readings, considering the redundancy of measurement systems and using adequate outlier’s identification tools; ii) the analysis and interpretation of the structural behaviour using machine learning techniques, like, among others, neural networks, short time Fourier transform analysis and long short-term memory models; iii) the support to decisions resulting from the structural safety assessment, namely the emission of early warning messages due to real-time classification of measurements based on thresholds previously defined and the early detection of patterns related to dam failure scenarios. It should be noted that the ongoing improvements in new automated monitoring systems combined with new data analysis approaches, supported in a solid dam behaviour knowledge, are important to achieve the main goal of safety assessment of dams, preferably in real-time.