We use artificial intelligence almost every day. Be it in self-driving cars, face, image, or voice recognition, personalised advertising, music suggestions, smart homes, weather, and sports reports, or shopping. These neural networks enable effective predictions based on efficient data analyses, which make human life much easier.
For this reason, numerous industries, such as medicine, sales, or the automotive industry, are now relying on artificial intelligence to improve their work processes. However, the construction industry has made comparatively little use of the potential of neural networks yet.
To pave the way for artificial intelligence on the construction site, this PhD thesis aims to present a holistic approach for the application of neural networks, which includes all necessary steps from data collection to model building. For this purpose, first, the theoretical background is explained and deepened through selected forecast examples, which emphasise the importance of the distribution analysis of training and test data. Subsequently, the focus is on use cases in construction management and economics, whereby the input parameter selection for predictions of the labour consumption rate and the productivity losses is examined in greater depth.
In the next step, the extensive literature review has shown that one major reason why artificial intelligence is not widely used in cost, duration, or productivity forecasting is data availability. The used data sets often have few cases. Thus they are not suitable for being analysed by artificial intelligence. For this reason, five concepts for the automatic collection of working hours with simultaneous area and activity allocation are developed. Subsequently, one of these concepts is selected and evaluated during reinforced concrete work on a construction site. The result is that global navigation satellite systems in combination with beacons based on Bluetooth Low Energy are very well suited for this task because they enable reliable data collection inside and outside the building.
Based on the collected data, the procedure for using neural networks is then illustrated. This example pursues the goal of recognising formwork, reinforcement, and concreting activities based on the workers’ movement patterns. During the iterative modelling process, it turns out that neural networks have great potential in the classification of selected activities based on duration, distance, number of ways, and warehouse visits as well as average and maximum speed.