Fault diagnosis of industrial motor bearings is crucial for equipment performance and lifespan. Traditional diagnostic methods aggregate data from multiple factories, leading to issues with data privacy and high annotation costs. To address these problems, a fault diagnosis strategy based on adaptive local collaboration (ALC) federated learning was proposed. In this approach, bearing data under different working conditions was stored across multiple clients, with a central server collaborating with each client to build a federated learning diagnostic model. An improved ResNet-18 network was used as the classifier, which was trained within the personalized federated learning framework. The ALC federated learning method enables each client to effectively integrate global and local models, extracting global information to optimize local training results. Experiments demonstrate that this method enhances fault diagnosis accuracy while protecting data privacy, showing higher fault classification precision compared to other methods, especially in multi-factory environments.