This study adopts a new algorithm that utilizes deep learning technology for data processing. Therefore, this project intends to use a multi-layer attention mechanism, combined with LSTM (Long Short Term Memory) network and CNN (Convolutional Neural Network) algorithm. This method can efficiently detect various abnormal phenomena. Experimental results have shown that the algorithm proposed in this paper achieves a prediction accuracy of 0.92 on the Area Under the Curve (AUC), with improvements to 0.89 and 0.84, respectively. Based on the above research, this project will also investigate the impact of parameter settings and sample sets on model performance. The experimental results show that the CNN LSTM model based on multi-level attention can better improve the detection rate of transaction anomalies in blockchain.