Results and Discussions Compared with the traditional AdaBoost algorithm (Fig. 5), the proposed improved AdaBoost ensemble learning framework reduces the maximum compensation error by 9.28 pm and the standard deviation by 2.2 in the cooling-heating experiment. Compared with the common traditional machine learning-based algorithms, the improved AdaBoost ensemble learning framework also offers great advantages (Table 2). The results show that the improved AdaBoost model overcomes the low accuracy and instability of the traditional AdaBoost model in temperature compensation. In the iteration process of the improved AdaBoost, the weight coefficient of the current weak learner is reasonably redistributed according to the error rate difference between the current weak learner and the one generated by the last round of iteration by comparing the error rates of the two weak learners, so that a close correlation between each two adjacent weak learners can be achieved. In this case, the weight of a weak learner is no longer determined by its error rate alone. Instead, it is generated by the iteration rule of the traditional AdaBoost and then optimized according to the performance difference between the two adjacent weak learners. The performance of the final strong learner is thereby improved compared with that of the traditional AdaBoost. This point is also reflected in the wide-range temperature drift experiment (Fig. 9). The maximum error and the standard deviation of the basic AdaBoost are 15.83 pm and 4.83, respectively, while those of the improved AdaBoost are 4.99 pm and 1.40, respectively.

The random fluctuation of the fiber Fabry-Perot tunable filter (FFP-TF) is easily intensified by the variation of ambient temperature, ultimately reducing the accuracy of the fiber Bragg grating (FBG) demodulation system. At present, the common solutions are the demodulation method combining the Fabry-Perot (F-P) etalon with reference grating, the demodulation method based on composite wavelength reference with acetylene gas cell, and so on. Although these methods can improve the demodulation accuracy of the system to a certain extent, the added hardware greatly increases the cost of the demodulation system. In addition, these methods are susceptible to ambient temperature. This study proposes a novel software-supported FBG demodulation method based on an improved AdaBoost algorithm. Specifically, the AdaBoost ensemble learning framework is used to construct a temperature drift model of the tunable filter. In the iteration process of the traditional AdaBoost, the weight of the generated weak learner is directly determined by its error rate, with no direct correlation between each two adjacent weak learners. In other words, the performance of the current generated weak learner is not directly affected by the weak learner generated by the previous round of iteration, and it cannot directly affect the results of the next round of iteration either. Consequently, the performance of the generated weak learners is likely to be random, which is unfavorable for the performance of the ensemble model. To solve this problem, this study proposes a dynamic weight update strategy for weak learners based on their error rate differences to accurately compensate the F-P tunable filter.

In this study, the AdaBoost ensemble learning framework is utilized to compensate the demodulation system. Specifically, data on the temperature drift characteristics of the tunable filter in a variable temperature environment are obtained, and the characteristics and labels of the data are determined. Subsequently, the AdaBoost algorithm is used to model the data. The AdaBoost algorithm framework is improved, and weight update steps are added to the AdaBoost iteration process. After the weight update coefficient is calculated with the difference between the error rates of two adjacent weak learners, it is utilized to update the weight coefficient of the current weak learner and ultimately to obtain a close correlation between each two adjacent weak learners. Then, the temperature drift data are modeled in the improved AdaBoost algorithm framework, and the accuracy and stability of the improved model are verified in different variable temperature environments. Finally, the proposed improved algorithm is compared with the common machine learning-based algorithms in the same environment to verify the effectiveness of the proposed algorithm.

By modeling the temperature drift characteristics of the tunable F-P filter and improving the traditional AdaBoost ensemble learning framework, this study proposes a new dynamic weight update strategy based on the error rate differences among weak learners. Furthermore, experiments of temperature drift compensation are carried out in two environments: cooling-heating and cooling. The wavelength shift of the tunable F-P filter is accurately compensated in variable temperature environments. Experimental verification reveals that the improved ensemble model offers the advantages of high accuracy and favorable stability, and it significantly outperforms the traditional AdaBoost algorithm and other traditional machine learning-based algorithms in variable temperature environments. In addition, compared with the traditional temperature drift compensation method for tunable filters based on the etalon and gas cell, the proposed temperature drift compensation method, with no need to add additional hardware to the existing demodulation system, is readily portable and boasts high economic practicability.