Detecting equipment activities by using machine learning algorithms

Discrete-event simulation can serve as a tool for using equipment data to control processes and calculate alternative scenarios. For this purpose, the simulation requires knowledge of the process states on the construction site. One way using these process states automatically in the simulation is to interpret sensor data using machine learning. This work shows the procedure and the results for the application of machine learning to a practical example in the field of special civil engineering. Sensor data (features) and activity data (target values) are used for data acquisition. Data preprocessing is performed using the moving average method to suppress data noise; data outliers are filtered using the box method. Feature selection is based on statistical considerations and the SHAP values. A total of five supervised machine learning algorithms are used to classify the existing data: (1) Decision Tree, (2) Logistic Regression, (3) Support Vector Machine, (4) Naive Bayes, (5) Artificial Neural Networks. Confusion matrices, cross-validation, and learning curves are used to evaluate the algorithms. Overall, the paper shows that machine learning is very well suited to supporting the integration of current data into the simulation.