Federated Learning (FL) is vulnerable to the introduction of poisoned attack, resulting in inaccurate predictions for specific inputs. Some existing defense aim to identify and remove potentially poisoned model updates. However, this can also incorrectly exclude a fraction of benign client updates. So we propose a framework based on learning behavior for secure personalised blockchain federated learning. Clients build a learning process model containing learning behavior and small number of samples before uploading the local model, the miner performs global aggregation based on process consistency and knowledge quality and builds a knowledge repository for each client. Each client’s knowledge repository contains samples provided by clients with consistent behavior. The samples in the knowledge repository are examined in conjunction with the local knowledge of the client to generate a personalised knowledge repository. Even if the global model is poisoned, due to the differences in the behavior of normal and malicious clients and the differences in the features of normal and poisoned samples, the normal repository does not contain the poisoned features, and the poisoned samples are regarded as anomalies by the normal clients by comparing them with the personalised knowledge repository. Experimental results show that our method has better defense results compared to BVFB, RLR, MultiKrum and Median algorithms, with attack success rate below 3% under multiple attack scenarios.