Aikido: Structuring data point identifiers of technical building equipment by machine learning

Abstract

Building automation and control systems (BACS) generate large amounts of data. Manual interpretation of the data, e.g. for identifying energy efficiency measures, is in many cases either too costly or too complex. In order to reduce the manual effort of building data analysis, it is necessary that the labeling of data points is human- and machine-readable. As there is currently no widely accepted standard for labeling, these labeling texts differ greatly. We developed an algorithm based on Natural Language Processing (NLP) that automates the translation of BACS metadata. In previous works, we developed a schema based on 40 other schemas for the designation of data points (BUDO Schema) that we use for the translation. The user labels data points suggested by our algorithm Aikido and checks the prediction in a browser-based GUI (key interface). In our case study, we use a dataset that comprises five different projects, each using its own schema for structuring the metadata of data points in BACS. After ten training labels, our algorithm achieved an accuracy of 0.87 across all data sets. This increased to 0.95 with 50 training labels. We show that the chosen approach is able to normalize the metadata of different schemas to transfer to a previously defined schema. Thus, the manual effort is decreased significantly.