In the context of data heterogeneity in the industrial internet of things (IIoT), existing federated learning schemes suffer from unstable communication and negative impacts caused by the aggregation of irrelevant models. This paper proposes a “cloud-edge-end” personalized federated learning hierarchical architecture for industrial internet of things (FEDI), and designs a personalized federated learning algorithm based on similarity clustering (PFedSA). In the model update mechanism, the cosine similarity is used to maintain the relationship matrix, and the personalized cloud model with high similarity parameters is adaptively selected and downloaded. In the model aggregation strategy, weights are dynamically calculated and regularization is introduced to aggregate and update local models. On three datasets of MNIST, FMNIST and CIFAR10, compared with 8 algorithms such as FedAvg, the experimental results show that: 1) In terms of accuracy, the PFedSA algorithm has the best accuracy (up to 99.78%) or close to the highest accuracy in three classic scenarios such as pathological non-independent identical distribution (Non-IID); 2) In terms of communication efficiency, the PFedSA algorithm leverages a similarity clustering mechanism to accelerate the model’s convergence to the target accuracy, and the single-round calculation time is reduced compared to the FedRep algorithm; 3) In terms of the impact of hyperparameters, the PFedSA algorithm has better robustness to device dropout rates, stronger adaptability to data heterogeneity, and can effectively improve the personalized performance of the model.