Human–robot collaborative manufacturing, positioned at the core of the Industry 5.0 paradigm, increases the stochasticity of production systems through human variability, safety-related stops, and resource-sharing constraints. In such environments, rapid identification of dominant bottlenecks and their root causes is critical to operational resilience, production-management responsiveness, and preserving the human-centric character of operations. This study proposes an explainable bottleneck-diagnosis framework based on convolutional neural networks (CNNs) that leverages process heatmaps derived from a discrete-time flow-line simulation model inspired by the digital twin concept. Each simulation run is encoded as a three-channel heatmap (queue length, station busy state, and downtime), yielding a compact spatio-temporal “process fingerprint” suitable for image-based representation learning. We consider a six-class taxonomy: balanced operation, bottleneck caused by human fatigue, bottleneck caused by cobot performance degradation, bottleneck caused by safety stops in a human–cobot cell, bottleneck caused by robot failure, and supply delays (line starvation). On a balanced dataset of 1,500 synthetic samples, the CNN achieves a test accuracy of 0.938 (macro-F1 = 0.938; 95% Wilson confidence interval for accuracy: 0.898–0.963). An ablation study shows that multi-channel heatmaps improve accuracy by 0.10 compared with queue-only heatmaps. Explainability is operationalised using Gradient-weighted Class Activation Mapping (Grad-CAM), which consistently localises station–time regions associated with queue build-up and downtime patterns, providing process engineers with evidence that can be translated into corrective actions. The results indicate that simulation-generated heatmaps can support transparent and scalable bottleneck analytics for human–cobot manufacturing, enabling rapid prototyping prior to deployment on shop-floor data and providing an evidence-based decision-support instrument for operations and quality management. The framework also clarifies the shop-floor data required for deployment and proposes a staged implementation path for real-factory decision support.