| description abstract | Building maintenance tasks to solve unpredictable faults typically start with written communications from end-users (e.g., emails). Technicians manually translate end-users’ requests in work-orders (WOs) assigning them a priority level and the needed staff typology. When the number of contemporary requests is too high, these actions can lead to the interruption of critical services and then possible safety issues. Machine Learning (ML) methods can be trained to automatize this process due to large databases of annotated requests. Nevertheless, natural language preprocessing is needed to apply ML methods because of the unstructured form of the requests. This work aims to verify how preprocessing impacts the ability of ML methods to properly assign priority to the requests. The research methodology combines four different text preprocessing approaches (e.g., symbols and numbers remotion, stop-words remotion, stemming, meaningful words selection) and five consolidated ML methods to classify WOs according to two different priority scales (binary, 4-classes). Accuracy, recall, precision, and F1 are calculated for each combination. Tests are performed on a database of about 12,000 end-users’ maintenance requests, generated for 34 months in 23 university buildings. Results show that strong preprocessing methods, usually performed to increase the effectiveness of ML, do not significantly improve the accuracy of the predictions. Moreover, they show that four of the five tested ML methods obtained a higher accuracy for binary classification and for high and mean priority classes of 4-classes classification. This means that ML methods are especially effective in a preliminary check of the most urgent requests. These results then encourage the use of ML methods in automatic priority assignment of building maintenance tasks, even if based on natural language unstructured requests. The ML can significantly speed up the interventions assignment process for the technical staff, thus improving the maintenance process especially in large and complex buildings organizations. Maintenance tasks imply significant effort, especially in large organizations and complex buildings, where several faults and contemporary end-users’ requests are managed by facility management (FM) technicians. Maintenance plans usually include procedures to prioritize corrective actions and establish the maximum intervention time, as economic penalties are due by FM contractors when delays occur with respect to the assigned priority. Most end-users’ requests are represented by unstructured textual messages stored in large databases, then relevant human and economic resources are provided to timely evaluate and assign priority to interventions, thus reducing the time needed to solve the issue and ensuring continuity of critical services. Automation approaches, based on machine learning (ML), can significantly improve the process, optimizing time and resources. Nevertheless, natural language preprocessing (NLP) is needed to apply ML, because of the unstructured form of requests. This work evaluates the influence of NLP on the ability of ML to automatically predict priority levels for corrective actions. Results demonstrate that ML is effective in detecting the most urgent requests, even without NLP (then directly applied to the unstructured requests). This encourages a wider use of ML in automatic priority assignment of building maintenance tasks to speed up the interventions assignment process for the FM technical staff. | |
