Structural dam safety control activities are commonly based on visual inspections and monitoring data recorded through different sensors. The sensor data collected is used to create statistical and predictive models. However, this data must be processed and validated beforehand. Outlier detection and treatment is a costly and slow process that can be improved using data mining and machine learning techniques. In a Big Data centered world, outliers appear more often, and without an automated way to detect them, engineers cannot anticipate and act on time. The presented work proposes an approach to identify and treat outliers from sensor data retrieved from an Automated Data Acquisition System (using real datasets from a dam), aiming to improve current baseline methods. Since sensor data is unlabeled, unsupervised methods, such as clustering, must be used to group data and understand which points should be classified as an outlier. A novel approach is presented and evaluated, taking advantage of already validated Manual Data Acquisition System measurements, a clustering algorithm (DBSCAN) and baseline methods. This method can identify and remove most outliers in the datasets used for demonstration.

--

Outlier detection,Sensor data,Machine learning,Data mining

• Computer and Information Sciences - Natural Sciences
• Civil Engineering - Engineering and Technology
• Electrical Engineering, Electronic Engineering, Information Engineering - Engineering and Technology