Performance Indicators. A performance prediction method for moisture safety design
The aims of this research project is to develop and gain knowledge about a different approach to moisture safety design based on AI in order to attain healthy buildings and relate this new approach to current practice and prospective users.
Secondary data (data produced for some other purpose) was used to train Artificial Neural Networks (ANN) to predict the performance of outdoor ventilated crawl-spaces regarding microbiological smell, mould and rot. The best performing ANN managed to predict smell 100%, mould 76%, and rot 92% correctly on 38 validation cases not used in the training process. A reliability test was performed designed as a parameter study. The results highlighted some uncertainties in the trained ANN which are likely to be due to a high level of missing values and skewed data. In addition, the parameter study goes far outside what the ranges of the retrieved training data, forcing the ANN to extrapolate predictions.
The interview study with engineering consultants indicated that experience is considered to be a decision support in moisture safety design even though feedback from past projects rarely is available. In addition the general opinion was that available tools are too demanding. Through a questionnaire a performance prediction comparison was set up to test the competitiveness of the trained ANN. The average prediction result for the respondents (engineering consultants, moisture damage consultants, moisture experts) was 50% correct predictions whereas the ANN had a 93% correct prediction level. There was no notable indication of a correlation of the prediction results with the respondents’ background. The same study also revealed that a system to capture experience is highly requested by the respondents.
The results so far are promising but ANN, based on real life experience, must be tested further with better training data, preferably with data designed for this purpose. The method has a potential to capture real life experience in a structured and systematic manner. Moreover, it may be helpful in the decision process during the early stages of design.
Artificial Neural Network