Introduction and Objectives:



The increasing complexity of industrial supply chains and the interdependencies among inbound logistics, production/processing operations, and outbound logistics have made equipment reliability management a strategic challenge for capital-intensive industries, particularly the mineral processing sector. Within such a structure, equipment failure is not merely a technical event confined to a single asset; rather, it can generate cascading effects throughout the entire supply chain, resulting in unexpected shutdowns, increased operating costs, reduced productivity, and weakened operational sustainability. At the same time, the accumulation of historical maintenance data in organizational databases has created a valuable opportunity to employ data-driven approaches and machine learning algorithms to predict equipment failure behavior. The main objective of this study is to propose a data-driven framework for predicting the probability of equipment failure in a three-echelon mineral processing supply chain and to compare the performance of machine learning algorithms in order to identify the best-performing model in terms of accuracy and stability.



Methods:



This study is applied in terms of its objective and quantitative and data-driven in nature. The dataset comprised 400 cleaned weekly equipment-level records collected over an eight-year period from the organizational maintenance database of the Central Alborz Coal Company. The study was conducted within a three-echelon supply chain consisting of inbound logistics, processing operations, and outbound logistics. Based on the research framework, ten input variables associated with equipment maintenance indicators, performance characteristics, and operating conditions were defined, along with one output variable representing the “probability of equipment failure in the subsequent period.”



Before model development, several data-preprocessing procedures were performed, including the Shapiro–Wilk normality test, the identification and treatment of outliers, correlation analysis, and the Kruskal–Wallis test to examine differences among the supply chain echelons. During the modeling stage, two algorithms—Artificial Neural Network (ANN) and Random Forest (RF)—were implemented, and their performance was assessed using standard evaluation metrics. The models were trained using five-fold cross-validation and hyperparameter tuning. After selecting the optimal model configurations, the final performance of the models was evaluated on an independent test set.



Findings:



The results of the data analysis indicated non-normal distributions and the presence of significant outliers in the operational indicators, reflecting the occurrence of critical operational events. Statistical tests also confirmed significant differences in the indicators across the three supply chain echelons, demonstrating the multi-echelon nature of equipment failure.



During the modeling stage, the Artificial Neural Network algorithm outperformed the Random Forest model in terms of error and accuracy metrics and demonstrated greater stability and better convergence in failure prediction. The confusion matrix also confirmed the ANN model’s optimal balance in detecting equipment failures. Furthermore, the model explainability analysis, based on the mean absolute SHAP values, revealed that “total unplanned downtime hours,” “number of registered corrective maintenance work orders,” and “operating hours since the last preventive maintenance (PM) activity” made the greatest contributions to equipment failure prediction.



In addition to confirming the effectiveness of nonlinear models in identifying complex and cross-echelon patterns, these findings provide a data-driven framework for predicting equipment failures in multi-echelon industrial settings.



Conclusion:



The findings of this study demonstrate that equipment failure in the mineral processing industry is a multidimensional phenomenon that depends on the specific echelon of the supply chain. Its analysis therefore requires an integrated, data-driven approach. By employing machine learning techniques, the proposed framework enables reliable prediction of the probability of equipment failure in the subsequent period and can serve as a foundation for implementing predictive maintenance systems in the mining and mineral processing industries. The application of such a model can reduce the risk of unexpected shutdowns, improve the allocation of maintenance resources, and enhance the reliability of the entire supply chain.