The latest industrial revolution has introduced autonomous cyber-physical production systems, integrating machine learning into smart manufacturing to optimize production and resource management. However, this integration impacts trustworthiness due to less predictable and explainable behaviors. This paper presents a novel model-based methodology for evaluating the trustworthiness of such systems. A study was conducted to explore the potential and limitations of model-based assessment, categorizing limitations into structural, behavioral, and resource-related aspects. The findings highlight inadequate risk identification and assessment of ML components in these systems and the constraints of single modeling approaches. Based on these insights, we propose a new methodology to address these limitations and improve the risk assessment of ML components in autonomous production systems.