Using detrending methods for intelligent processing of construction process monitoring data

Abstract

Data processing of monitoring systems of various processes at the stage of construction and operation of buildings requires the development of special tools that belong to the field of artificial intelligence. The trend removal method is one of the ways to preprocess data collected from various sensors and IoT devices that monitor the state of buildings, structures and soil masses during construction and operation. This article analyzes how different approaches to detrending time series affect the performance and accuracy of algorithms for CI computational intelligence models. The analysis compares three approaches: linear detrending, non-linear detrending, and first-order differentiation. Five representative methods are used as CI models: DENFIS dynamic evolving fuzzy neural network, GP Gaussian process, MLP multilayer perceptron, OP-ELM optimally trimmed extremal learning machine, and SVM support vector machine. There are three main conclusions from the experiments performed on the four datasets: 1) detrending does not improve overall performance, 2) the empirical mode decomposition method provides better performance than linear detrending, and 3) first-order differentiation in some cases can be effective and in some cases counterproductive for series with common patterns.