The machine learning method based on static structural features has poor detection results for gate-level Hardware Trojans (HT). A HT detection method based on cascaded structure features is proposed. The features are constructed by co-occurrence matrix and are recognized by a many-to-many stacked long short-term memory (LSTM) network. The experimental results show that this method obtains 93.1% of the average true positive rate (TPR), 99.0% of the average true negative rate (TNR) and 79.3% of F1-score in 15 benchmarks from TrustHub. The experimental results are better than the existing methods.