Objective Fiber Fabry-Perot tunable filters (FFP-TF) controlled by piezoelectric ceramics are prone to temperature drift in fiber Bragg grating (FBG) sensing systems. During the long-term measurement process, FFP-TF will cause continuous drift of the output wavelength, which will adversely damage the FBG sensing system's measurement accuracy. At the moment, FFP-TF temperature drift compensation primarily entails adding hardware calibration modules to the FBG sensing system, such as the reference grating method, F-P etalon method, gas absorption method, and composite wavelength reference method. Although these technologies can efficiently adjust for temperature drift, they greatly increase the system's cost and complexity. As a result, utilizing software approaches to compensate for temperature drift in FFP-TF is a practical and low-cost method. However, most contemporary temperature drift compensation approaches based on artificial intelligence technologies neglect temperature drift data's temporal features. In fact, the fresh sample has a higher impact on the prediction outcomes of the following data than the old sample. As a result, this work extensively addresses the impact of temporal features on temperature drift compensation when processing temperature drift and other highly time-dependent data. A tunable filter temperature drift compensation approach with time weight is suggested based on the AdaBoost-SVM algorithm and time weight. Methods We use FBG0 as the reference grating and the other three FBGs as sensing gratings, and each sensing grating is modeled individually. The temperature-related values of the experimental environment are chosen as the model's input features in this investigation. Furthermore, because the wavelength drift errors of each FBG in the FFP-TF output spectrum have a high correlation, we use the drift of the reference grating as an input feature of the dynamic compensation model to compensate for the lack of accurate temperature information in the F-P cavity. The significant link between the temperature drift sequence data before and after is taken into account in full by this investigation. The idea of time weight is introduced in the process of modeling the temperature drift of FFP-TF to assign various temporal attributes to each sample. After that, temperature drift samples are modeled using support vector machines (SVM) as weak learners, and several SVM learning models are integrated using the AdaBoost framework. In the integrated prediction process, the time attribute of samples has an impact on the update of sample weights in addition to the prediction performance of each model. Multiple temperature change modes have been used to validate the aforementioned procedure. Results and Discussions First, the temperature drift compensation results of the proposed algorithm are compared with the conventional AdaBoost-SVM algorithm for three transmission gratings in the 2. narrow changing temperature environment experiment of cooling and heating (Table 3). Secondly, in the 15. cooling amplitude experiment, the temperature drift compensation results of the proposed algorithm are compared with the traditional AdaBoost-SVM algorithm for three transmission gratings. The experimental results show that the maximum temperature drift compensation error of the traditional AdaBoost-SVM algorithm is 10. 83 pm, while the maximum temperature drift compensation error of the AdaBoost-SVM based on time weight is reduced to 7. 04 pm. The results show that the classic AdaBoost-SVM algorithm's maximum error is approximately 11. 57 pm, whereas the maximum error of the AdaBoost-SVM based on time weight is only approximately 4. 05 pm. The strategy suggested in this research, however, outperforms unoptimized machine learning methods in terms of superior stability, stronger reliability, and higher prediction accuracy (Table 4). The aforementioned findings show that the method suggested in this article may successfully determine samples' temporal properties, allowing for more reasonable sample weight allocation, a decrease in model performance fluctuations, and an increase in model accuracy. Conclusions First, the high link between the temperature drift sequence data before and after is thoroughly taken into account in this article. The ratio of new and old samples is altered by applying various new weights at various time points, which makes the distribution of sample weights more logical and enhances the model's performance. The experiment next establishes a nonlinear model between the filter surface temperature and output drift error using the spectral locations of three reference gratings as input features. Experiments are carried out on two datasets with different temperature change patterns, and the results reveal that the first dataset does not fully comply with the more important rule of closer samples in general time series proposed in this article due to the short-term fluctuation of temperature changes, so the performance improvement of the model is not significant; the temperature change in the second dataset demonstrates a monotonic cooling trend with apparent gradients, which is more consistent with the more important principles of closer samples, and the performance gain is more significant. Unlike typical hardware techniques, the method suggested in this paper does not require any additional hardware, resulting in a novel approach to temperature drift compensation of tunable filters.
