A new hybrid inference model for human performance reliability prediction: a case study of construction workers

Researchers from Nanyang Technological University have developed a novel framework that integrates worker self-reportsaZ with objective expert evaluations to address the challenge of validating human reliability models. They also proposed a new hybrid inference model combining Self-Organizing Maps (SOM) and Bayesian Networks (BN) to more accurately predict worker performance and identify potential failures. Published in Smart Construction, this research transcends the limitations of traditional expert-driven methods, offering a validated, data-driven approach to enhance workplace safety and operational efficiency in the high-risk construction industry.

The construction industry is recognized as one of the most hazardous sectors, with human error being a primary cause of most workplace accidents. To assess and mitigate these risks, Human Reliability Analysis (HRA) methods are employed, with second-generation techniques like the Cognitive Reliability and Error Analysis Method (CREAM) being prominent for their focus on cognitive and contextual factors. However, traditional CREAM models suffer from significant limitations: they rely heavily on subjective expert judgment, lack independent ground-truth for validation, and often oversimplify the complex, non-linear interactions between performance-shaping factors .

To overcome these challenges, Yamo Cao, Zeren Jin, and Assistant Professor Yuguang Fu from Nanyang Technological University proposed a comprehensive framework for data collection and a new hybrid model for prediction. The proposed framework consists of two main components: A Unified Data Collection and Processing Framework and A Hybrid SOM-BN Inference Engine.

The Unified Data Collection and Processing Framework brings together subjective worker self-assessments and objective expert evaluations. To process subjective data, the researchers introduced the Contextual Human Reliability Score (CHRS), which applies fuzzy logic and rank-standardization to worker self-reports, effectively reducing bias and creating a robust input metric. The objective data, consisting of expert observations of worker performance, serves as the independent ground truth for model validation.

The Hybrid SOM-BN Inference Engine uses a two-stage process. First, a Self-Organizing Map (SOM), an unsupervised clustering technique, analyzes the CHRS data to identify complex and non-linear patterns among the various performance conditions. Second, these empirically derived clusters are fed into a Bayesian Network (BN), which performs transparent probabilistic inference to predict the worker's performance control mode (e.g., effective or ineffective).

To validate the framework, the researchers conducted a case study on a construction site, collecting data from 35 workers. The results were compelling: the proposed SOM-BN model demonstrated high accuracy (88.57%) and specificity in its predictions. In stark contrast, traditional CREAM models, when applied to the same dataset, were unable to effectively distinguish between safe and unsafe performance, highlighting their limitations in this real-world context. Furthermore, the use of explainability tools like SHAP provided clear insights into individual worker performance, making the model's predictions actionable.

In the future, the research team plans to expand the dataset to include a more diverse worker population and apply the framework to other industries, further demonstrating its broad applicability in improving human reliability and safety.

This paper ”A new hybrid inference model for human performance reliability prediction: a case study of construction workers” was published in Smart Construction.

Cao Y, Jin Z, Fu Y. A new hybrid inference model for human performance reliability prediction: a case study of construction workers. Smart Constr. 2025(3):0016, https://doi.org/10.55092/sc20250016.