Automation systems play a critical role in modern industry, enabling efficient, precise, and large-scale operations across sectors such as manufacturing and smart infrastructure. These systems rely heavily on AI to process data, optimize decision-making, and enhance system reliability. As automation increasingly depends on distributed data from numerous devices, federated learning (FL) has emerged as an attractive solution. However, in real-world deployments of FL, client dropout attacks are commonly encountered, which significantly degrade the performance of automation systems employing FL. To mitigate the impact of dropouts on the training process, we propose a blockchain-integrated FL architecture with an adaptive fallback mechanism (BFL-AF). By leveraging the transparency and immutability of the blockchain, the system can effectively verify client participation and securely record model updates during each training round. Furthermore, an adaptive fallback mechanism is introduced, which utilizes clients’ historical model weights to enhance training stability and recovery capability. Experimental results demonstrate that the proposed method significantly improves convergence speed and robustness under various abnormal dropout scenarios, offering strong resilience and a practical solution for a building privacy-preserving and security-enhanced FL-based automation system.