Privacy-preserving federated learning (PPFL) is an emerging secure distributed learning paradigm that aggregates user-trained local gradients into a federated model via cryptographic protocols. Unfortunately, PPFL is vulnerable to model poisoning attacks launched by Byzantine adversaries. Most Byzantine defense strategies in existing PPFL schemes have problems such as incomplete data privacy protection, non-adaptive to highly heterogeneous data scenarios, and some are facing Single Point of Failure (SPOF). To address the above problems, this paper designs a blockchain-based robust clustered privacy-preserving federated learning system (BR-CPPFL), which can achieve resistance to model poisoning without leaking privacy through privacy-preserving detection of malicious models in both independently and identically distributed (IID) data setting and non-independently and identically distributed (Non-IID) scenario. The weighted global model robust aggregation strategy which combines the intra-cluster average aggregation and cross-cluster weight aggregation makes BR-CPPFL realize the Byzantine robustness. Extensive evaluation on benchmark datasets shows BR-CPPFL outperforms the existing defense strategies. In addition, the overhead is linear except for the quadratic communication and computational overhead caused by clustering in the first round. Overall BR-CPPFL overhead is relatively low.