Scholarlink

<p>To address the wavelength temperature drift of tunable F-P filters under dynamic thermal conditions, this study constructs a closed-loop system of “physical perception signal processing feature analysis intelligent inference error correction” and proposes a Dual-Attention LSTM compensation method: after extracting temporal features from LSTM, feature attention adaptively allocates the importance of temperature, temperature change rate, reference FBG spectral position, and driving voltage, while temporal attention focuses on key historical segments. Validation was conducted based on two operating conditions (heating cooling reheating for 1000 min and monotonic cooling for 950 min). The results showed that in the heating–cooling–reheating scenario, the MAE/RMSE of FBG1 decreased from 2.165/2.886 pm of LSTM to 0.726/0.939 pm, and the MAXE of FBG3 decreased from 28.344 pm to 2.661 pm. In the monotonic cooling scenario, the MAE of the three FBGs is between 0.707 and 0.818 pm, and the RMSE is about 0.88–1.02 pm. The ablation experiment further shows that compared with the attention free LSTM (MAE 2.12 pm, RMSE 2.88 pm), this method significantly converges; The hyperparameter sensitivity provides the optimal window length L = 8 and the number of layers is 2. The method achieves unified high-precision compensation for non steady state and gradual temperature drift.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Temperature drift compensation of tunable F-P filter based on dual-attention LSTM model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月3日

查看全文
  • Xiaojuan Zhang

摘要

To address the wavelength temperature drift of tunable F-P filters under dynamic thermal conditions, this study constructs a closed-loop system of “physical perception signal processing feature analysis intelligent inference error correction” and proposes a Dual-Attention LSTM compensation method: after extracting temporal features from LSTM, feature attention adaptively allocates the importance of temperature, temperature change rate, reference FBG spectral position, and driving voltage, while temporal attention focuses on key historical segments. Validation was conducted based on two operating conditions (heating cooling reheating for 1000 min and monotonic cooling for 950 min). The results showed that in the heating–cooling–reheating scenario, the MAE/RMSE of FBG1 decreased from 2.165/2.886 pm of LSTM to 0.726/0.939 pm, and the MAXE of FBG3 decreased from 28.344 pm to 2.661 pm. In the monotonic cooling scenario, the MAE of the three FBGs is between 0.707 and 0.818 pm, and the RMSE is about 0.88–1.02 pm. The ablation experiment further shows that compared with the attention free LSTM (MAE 2.12 pm, RMSE 2.88 pm), this method significantly converges; The hyperparameter sensitivity provides the optimal window length L = 8 and the number of layers is 2. The method achieves unified high-precision compensation for non steady state and gradual temperature drift.