With the rapid development of the smart warehouse domain, there is a growing need for data privacy protection and accurate decision-making. This urgent need promotes the wide application of federated learning architectures in this field. The federated learning architecture has received attention for its significant advances in protecting users’ private data. However, while the introduction of federated learning in smart warehouses mitigates privacy concerns, the unique global model in federated learning may not apply to all sensors in smart warehouses. Therefore, our breakthrough introduces a personalized federated learning framework, catering to unique regional smart warehouse needs and crafting task-specific models for each sensor, ensuring privacy and security. We propose a novel personalized federated learning framework, SACPFL, whose core goal is to improve decision-making accuracy and accomplish personalization while safeguarding user privacy and security. Specifically, SACPFL employs adaptive dynamic clustering to form discrete clusters based on the average soft prediction values of client models. In this way, our framework can process and analyze heterogeneous data more accurately, dramatically improving the model’s ability to handle heterogeneous data. The experimental results indicate that the performance of SACPFL has improved by 29.35%, outperforming the eight state-of-the-art methods.