基于PSO-SVR的温度传感器寿命预测方法研究

Life Prediction Method for Temperature Sensors Based on PSO-SVR

摘要:

针对温度传感器实际工况寿命试验周期长、失效数据获取困难以及小样本退化数据下寿命预测精度不足的问题，提出一种基于加速老化试验与粒子群优化支持向量回归(PSO-SVR)的退化轨迹建模及实际工况寿命预测方法。首先，依据温度传感器目标服役工况，综合考虑温度循环范围、循环频率和电应力，设计源加速老化试验工况，并以测温偏差作为性能退化表征参数，获取温度传感器加速老化退化数据。其次，针对普通SVR模型参数依赖性较强、后续退化趋势预测稳定性不足的问题，引入粒子群优化算法对SVR惩罚系数、核函数参数和不敏感损失参数进行联合寻优，建立PSO-SVR退化轨迹预测模型。为验证模型对后续退化过程的预测能力，按照时间顺序将源工况数据划分为模型训练区间和时序留出检验区间，并与普通SVR模型进行对比。最后，基于加速工况退化轨迹获得目标工况下的寿命预测结果，并结合四级修、五级修样品阶段性实测数据进行验证。结果表明，PSO-SVR模型在加速工况时序留出检验区间的决定系数达到0.9919，优于普通SVR模型；目标工况下，普通SVR预测寿命约为11.59年，PSO-SVR预测寿命约为14.08年，且PSO-SVR预测的寿命曲线与阶段性实测数据分布更加一致。研究结果表明，所提方法能够有效提高小样本加速老化数据条件下温度传感器退化轨迹建模精度和实际工况寿命预测可靠性，可为长寿命温度传感器的寿命评估与服役状态判定提供参考。

Abstract:

To address the problems of long test duration, difficulty in obtaining failure data, and insufficient prediction accuracy under small-sample degradation data for temperature sensors under actual operating conditions, a degradation trajectory modeling and life prediction method based on accelerated aging test and particle swarm optimization-support vector regression(PSO-SVR)is proposed. First, according to the target operating condition of the temperature sensor, a source accelerated aging condition is designed by considering temperature cycling amplitude, cycling frequency, and electrical stress. The temperature measurement deviation is selected as the performance degradation indicator, and accelerated aging degradation data are obtained. Then, to overcome the strong parameter dependence and insufficient stability of conventional SVR in predicting subsequent degradation trends, PSO is introduced to jointly optimize the penalty factor, kernel parameter, and insensitive loss parameter of SVR, thereby establishing a PSO-SVR degradation trajectory prediction model. To evaluate the prediction ability of the model for future degradation trends, the source-condition data are divided into a model calibration interval and a time-ordered holdout test interval according to the aging time sequence, and the proposed model is compared with the conventional SVR model. Finally, based on the degradation trajectory under the source condition and the equivalent time-axis mapping method, the degradation trajectory under the target operating condition is extrapolated and verified using staged measured data from fourth-level and fifth-level maintenance samples. The results show that the coefficient of determination of the PSO-SVR model in the time-ordered holdout test interval reaches 0.9919, which is higher than that of the conventional SVR model. Under the target operating condition, the predicted life of the conventional SVR model is about 11.59 years, while that of the PSO-SVR model is about 14.08 years. Moreover, the extrapolated degradation trajectory obtained by PSO-SVR is more consistent with the distribution of staged measured data. The results indicate that the proposed method can effectively improve the accuracy of degradation trajectory modeling and the reliability of life prediction for temperature sensors under small-sample accelerated aging data, providing a reference for life assessment and service condition evaluation of long-life temperature sensors.