Federated learning has been shown to be vulnerable to poisoning attacks, particularly from Byzantine attackers who can manipulate the global model by injecting carefully crafted fake local model updates. This vulnerability is exacerbated by the reliance on server aggregation in the federated learning framework, making it challenging to defend against such attacks, especially in the presence of unreliable servers. To address these problems, we propose a blockchain-based Byzantine-robust decentralised federated learning framework BRFL. Specifically, the blockchain records the local models uploaded by the training clients and scores them according to the distances between the models. A new aggregation rule, reward and punishment mechanism and verification mechanism are designed to effectively mitigate the influence of malicious clients and bolster the robustness of the system. Extensive experiments show that our approach is robust against poisoning attacks and outperforms existing approaches even in scenarios with a substantial proportion of malicious clients.