Abstract: A parallel Transformer-LSTM( Long Short-Term Memory) hybrid model is proposed for complex Io T time series data.The model achieves real-time anomaly detection by optimizing the fusion of Transformer and LSTM networks. The model uses Shapley additive explanations to evaluate feature importance to identify key features,and uses sliding window segmentation to reconstruct time series data to enhance the extraction of temporal features. In addition,the model introduces a fusion layer to automatically fuse different features and adjust their weights to optimize anomaly detection performance. In the experiment,the mean square error and binary cross entropy are used as loss functions to train the LSTM model,serial Transformer-LSTM model,and parallel Transformer-LSTM model,respectively,and the final anomaly score is obtained by weighted aggregation. The experimental results show that the model achieves excellent performance on real datasets. Compared with existing methods,key indicators such as accuracy,precision,and recall are significantly improved. The effectiveness of each component is verified through ablation studies. Even on the Fridge dataset containing only 1% anomalies,the detection accuracy can be maintained at 99.93%,demonstrating the excellent ability of the model to monitor anomalies in complex time series.