CARL Source is CARL Software’s main software application for managing maintenance processes within the company, as well as the historical monitoring of the various materials and equipment, buildings, and necessary interventions.
We conducted a study involving Natural Language Processing techniques for document understanding and qualification. Basically, we aim at interpreting real requests for intervention, to determine whereas some elements can be automated: classification by discipline and qualification (e.g. electrician), management of the lexicon to simplify their writing, or even to find other sources of automation (writing, the importance of the request for intervention, for example, characterization of breakdowns and emergencies), etc.
The techniques used on Python :
- Natural Language Processing with Spacy
- PyEnchant spell checker
- Qualification of n-grams with Scikit Learn
- Classification of requests by LDA (Latent Dirichlet Allocation)
These techniques have made it possible to build a fairly broad lexicon corresponding to many types of intervention and to classify them partially. Unfortunately, the final result remains constrained by a number of difficulties. However, this study revealed two types of requests for intervention:
- Requests for descriptive interventions that simply describe the problem and possibly the requester
- Requests for interventions that both specify the problem and indicate the operational course of action
It made it possible to understand that the themes of classification are limited: Interventions :
- Plumbing / Heating / Ventilation that can be grouped together
- On infrastructure: elevator, roof, green spaces, door, furniture
- On material/equipment machines specific to the company
- Computers, computer hardware, and multimedia
- Detection systems: alarms, fire
- Miscellaneous and administrative operations: Purchasing – Technical files – Cleaning
It helped to understand the possible structures for writing EOIs to make them more automated, and to highlight a fundamental writing vocabulary that would allow the automation of writing and save time. For example, by writing the first 3 leakage letters, it would be possible to write the whole word and to automatically propose several possibilities: in order: water, hot water tank, washbasin, siphon, air, oil. And to automatically set this element as a symptom.
Finally, it made it possible to understand that there is not one single classification possible based on this structure, but several: according to the type of equipment, according to the type of intervention (repair or replacement, maintenance, diagnosis, …), by the degree of urgency, by type of symptom and even by location, and that these elements of vocabulary can significantly influence the automated classification.
The problems identified in this study will be taken up by an intern in 2021 in order to find a solution to make the project viable.