Risk Assessment of Infrastructure Using a Modified Adaptive Neurofuzzy System: Theoretical Application to Sewer Mains

Risk assessment lies at the core of decision making practices, relying on the likelihood of failure incidents and their consequential impacts integrated through a set of decision rules. The incorporation of fuzzy inference systems (FIS) into risk assessment has been proven to elevate the management and sustainability of infrastructures by reducing uncertainties in decision making and mimicking human decisions. Its parameters, primarily the fuzzy sets, are conventionally defined heuristically to align with the visions and objectives of decision makers or through clustering methods if data are available. However, utilizing these methods is insufficient for the comprehensive adoption or finetuning of input-output relationships, and it does not adequately address the incorporation of future changes in decision makers’ opinions due to its inability to self train. To overcome this challenge, a Mamdani-based neurofuzzy system is proposed and modified for best practice in risk assessment, and a framework is developed for compiling the required training data set. Additionally, modifications in the training process include adding Lagrange equality constraints to the objective function and testing different optimizers to maintain the fuzzy logic and reduce the computational cost, respectively. The validation of the model’s ability to learn the knowledge base and parameters of fuzzy set membership functions was conducted through training simulation on generic data developed from a predefined knowledge base, as explained in a case study. The data serve as an idealization or simulation of any risk data set that can be generated using the proposed framework. The model has confirmed its ability to learn the knowledge base from which the data are generated and the corresponding parameters of membership functions toward the comprehensive adoption of input-output relationships. This emphasizes the model’s capability to learn the knowledge base from any risk data set developed using the proposed framework. Also, results have shown that the added constraints control the training process efficiently by maintaining proper interactions among fuzzy sets membership functions satisfying fuzzy logic; additionally, the Adam (adaptive momentum estimation) optimizer has reduced the training computational cost by more than 99%. Finally, to explain and validate the flexibility of the neurofuzzy model in risk assessment in adopting changes in the FIS hyperparameters and rules, the model was retrained after inducing three independent possible changes to the decision rules and the fuzzy sets. The model has proved its ability to grasp diverse decision maker requirements, giving more flexibility and broader implementation of FIS in risk assessment.