A data-driven approximate causal inference model using the evidential reasoning rule

This paper aims to develop a data-driven approximate causal inference model using the newly-proposed evidential reasoning (ER) rule. The ER rule constitutes a generic conjunctive probabilistic reasoning process and generalises Dempster’s rule and Bayesian inference. The belief rule based (BRB) methodology was developed to model complicated nonlinear causal relationships between antecedent attributes and consequents on the basis of the ER algorithm and traditional IF-THEN rule-based systems, and in essence it keeps methodological consistency with Bayesian Network (BN). In this paper, we firstly introduce the ER rule and then analyse its inference patterns with respect to the bounded sum of individual support and the orthogonal sum of collective support from multiple pieces of independent evidence. Furthermore, we propose an approximate causal inference model with the kernel mechanism of data-based approximate causal modelling and optimal learning. The exploratory approximate causal inference model inherits the main strengths of BN, BRB and relevant techniques, and can potentially extend the boundaries of applying approximate causal inference to complex decision and risk analysis, system identification, fault diagnosis, etc. A numerical study on the practical pipeline leak detection problem demonstrates the applicability and capability of the proposed data-driven approximate causal inference model.