Transformer-based fiber temperature threshold prediction model

With the increasing demand for industrial equipment safety monitoring, Distributed Fiber Optic Temperature Sensing (DTS) systems have been widely applied in critical scenarios such as cables and pipelines due to their high precision and long-distance continuous monitoring capabilities. However, traditional temperature threshold prediction methods (e.g., control charts, LSTM) struggle to meet the dynamic prediction needs in complex environments due to their strong dependence on data distribution and limited ability to model long-term dependencies. To address this, this study proposes a temperature threshold prediction model based on an improved Transformer architecture. By designing a feature interaction module, the model integrates time-series temperature data with time scalar features and optimizes the multi-head self-attention mechanism to capture nonlinear relationships across time steps. The model adopts a hierarchical encoding-decoding structure, combining a vector encoder and threshold decoder, significantly improving prediction accuracy and robustness. Experimental results show that, on the same dataset, the proposed model outperforms LSTM, GRU, Bi-LSTM, and PSO-XGBoost models in MAE (10.54%), MAPE (6.08%), and RMSE (25.29%) metrics, demonstrating its superiority in high-noise and multi-variable scenarios. This study provides an efficient and reliable dynamic threshold prediction solution for industrial temperature monitoring, offering significant engineering application value.