Approche de prédiction de présence d'amiante dans les bâtiments basée sur l'exploitation des descriptions temporelles incomplètes de produits commercialisés

Thamer Mecharnia, Nathalie Pernelle, Lydia Chibout Khelifa, Fayçal Hamdi

TL;DR

This paper proposes a novel approach to predict the presence of asbestos in buildings by leveraging incomplete temporal data on commercialized products. By building an ontology and using probabilistic methods, the authors aim to assist building inspectors in identifying potential asbestos risks, especially in older structures.

What Problem Does This Paper Address?

Asbestos, widely used for its fire-resistant properties, remains a significant health hazard in many buildings, even after its ban in France in 1997. Identifying asbestos-containing materials (ACMs) is crucial for renovation, demolition, and general safety. However, existing building records often lack detailed information about the specific products used, especially their exact composition and the precise timeframes of their asbestos content. This temporal incompleteness makes it challenging to accurately assess the risk of asbestos presence. The paper addresses this by developing a method to infer asbestos presence from fragmented historical data, bridging the gap between available building descriptions and the need for precise risk assessment.

Methods

The core of the proposed method involves several key steps:

  • Ontology Construction: An "Amiante" ontology was manually built to model knowledge about buildings, diagnostics, and asbestos-related information. This ontology serves as a structured representation of the domain.
  • Data Enrichment and Population: The ontology is enriched with information from external resources like ANDEVA (National Association for the Defense of Asbestos Victims) and INRS (National Institute for Research and Safety). These sources provide temporal data on commercialized asbestos-containing products.
  • Data Fusion: When conflicting information exists between sources (e.g., different dates for when a product was no longer asbestos-containing), a pessimistic approach is used. The method takes the maximum probability of asbestos presence across sources for a given time interval.
  • Probabilistic Prediction: The paper proposes a non-supervised method to calculate the probability of asbestos presence. This involves:
  • Calculating the probability of asbestos in a specific product class based on the fused temporal data of commercialized products.
  • Adjusting this probability using a parameter ($\alpha$) derived from comparing diagnostic data with external resource data, accounting for potentially missing non-asbestos products.
  • Propagating this probability to building elements (locations, structures) and finally to the entire building by taking the maximum probability among its components. This "pessimistic" approach mirrors expert reasoning in risk assessment.

Datasets

  • ANDEVA Data: A dataset describing 650 products, including product class, names, supplier, and the date after which products are no longer asbestos-containing.
  • INRS Data: A dataset describing 300 products, including product names, supplier, time intervals of asbestos presence with uncertainty, type of asbestos, and usage types.
  • Synthetic Building Data: The authors generated 120 synthetic building descriptions, each comprising approximately 1133 products. This dataset was created to simulate real-world building data with inherent incompleteness and was used for experimentation.

Key Insights and Ideas

  • Leveraging Temporal Incompleteness: The paper demonstrates that even with incomplete temporal data on product usage, a probabilistic prediction of asbestos presence is feasible.
  • Pessimistic Approach for Risk Assessment: The adoption of a pessimistic strategy (taking the maximum probability) aligns with expert judgment in risk assessment, ensuring that potential hazards are not underestimated.
  • Ontology as a Knowledge Hub: The ontology serves as a crucial tool for integrating and structuring diverse information sources, enabling a more comprehensive understanding of asbestos-related data.
  • The Importance of External Data: The study highlights the value of external databases (like ANDEVA and INRS) in enriching internal building records and providing crucial temporal context for product usage.

Key Results

  • Ontology Enrichment: The process successfully created 64 subclasses of products and 694 instances of commercialized products by integrating INRS and ANDEVA data.
  • Probability Evolution: The calculated probability of asbestos presence in commercialized products generally remained stable until around 1972, then decreased, reaching 0 by 1997 (the year of the ban).
  • Class-Specific Probabilities: The probability of asbestos presence varied significantly across different product classes, with adhesives showing a faster decline in asbestos content compared to other classes.
  • Synthetic Data Validation: Experiments on synthetic data showed that the probabilistic prediction method could propagate risk from products to building elements and the entire building. The results indicated that the probability of asbestos presence did not significantly decrease for older buildings in the synthetic dataset, suggesting that even buildings constructed near the ban date could still pose a risk if older materials were used.
  • Expert Validation: An expert validated the calculation process and the resulting asbestos classes, leading to a simplification of the probability classes into "low" and "high" due to data incompleteness.

Limitations and Future Work

  • Data Incompleteness: The primary limitation acknowledged is the inherent incompleteness of data, which prevents definitive conclusions about the presence or absence of asbestos. The threshold for "low" and "high" probability classes requires further determination.
  • Limited Diagnostic Data: The study relied on a small number of diagnostic samples for calculating the adjustment parameter ($\alpha$), which might limit the generalizability of the results.
  • Future Work:
    • Testing the solution on real-world data provided by CSTB.
  • Using the calculated probabilities, building descriptions, and ontology to learn more precisely which building characteristics influence asbestos presence.
  • Exploring more sophisticated methods for learning from incomplete data and potentially incorporating more diagnostic data.

Summary updated

Abstract

National audience
BibTeX
@inproceedings{mecharnia2019approche,
  author = {Thamer Mecharnia and Nathalie Pernelle and Lydia Chibout Khelifa and Fayçal Hamdi},
  title = {{Approche de prédiction de présence d'amiante dans les bâtiments basée sur l'exploitation des descriptions temporelles incomplètes de produits commercialisés}},
  booktitle = {IC},
  year = {2019},
  url = {https://openreview.net/forum?id=z9zrsKG8Kcd},
}
Figure