Manufacturing Enterprises: An Explainable Machine Learning Framework for Defect Risk and Process Performance Optimization

Abstract

Nepal’s manufacturing sector is a structural cripple. Contributing a mere 5.72 percent to GDP — a figure that would embarrass any comparable developing economy — it grew at a sluggish 2.83 percent in 2025/26 while the tertiary sector ballooned to 61.8 percent and remittances swallowed 33 percent of national income. Beneath these macroeconomic indignities lies a more insidious failure: the vast majority of Nepal’s manufacturing enterprises, particularly the small and medium enterprises that constitute 99.7 percent of registered businesses, manage quality through intuition, post-hoc inspection, and the occasional ISO 9001 certificate that gathers dust on a factory wall. Defects are discovered after they are produced, scrap rates are absorbed as “normal loss,” and process capability remains a foreign concept to floor supervisors who have never seen a control chart.

This paper constructs a rigorous theoretical framework for predictive quality management in Nepalese manufacturing enterprises through explainable machine learning. We derive the complete mathematical architecture of quality loss — from Taguchi’s quadratic loss function through process capability indices (Cp, Cpk, Cpm) and Shewhart statistical process control — and integrate it with ensemble learning methodologies (XGBoost, LightGBM, Random Forest) and anomaly detection via Isolation Forest. Post-hoc interpretability is grounded in SHAP values from cooperative game theory, enabling factory managers to translate black-box predictions into actionable process adjustments. We prove the expected-loss decomposition, derive the relationship between Cpk and Cpm, establish the distribution of Isolation Forest path lengths, derive the Average Run Length (ARL) of the EWMA control chart, prove consistency of the mutual-information feature selector, establish the bias-variance decomposition for gradient-boosted ensembles, and verify the four Shapley axioms for the TreeSHAP algorithm. Drawing upon Nepal’s Economic Survey 2025/26, the ADB Asia SME Monitor 2024, and the Industrial Enterprises Act 2076 (2020), we situate the framework within Nepal’s institutional reality: 923,000 registered businesses, 90 percent of them SMEs, manufacturing concentrated in agro-processing, textiles, and construction materials, and an industrial policy architecture that incentivizes cottage industries but rarely demands statistical rigor. The paper argues that without predictive quality management, Nepal’s manufacturing sector will remain permanently trapped in low-value assembly and import substitution — incapable of competing with Indian and Chinese imports, incapable of generating the productive employment that 500,000 annual labor market entrants desperately need, and incapable of justifying the infrastructure investments that cheaper hydropower now makes possible.