基于时域特征与LSTM-Attention的IGBT退化预测方法

IGBT degradation prediction method based on time domain characteristics and LSTM-Attention

绝缘栅双极晶体管(IGBT)在可靠性分析任务中时间信息难以充分利用,导致预测精度不高。文中提出一种基于多维时域特征和注意力机制的深度学习方法,该方法结合主成分分析(PCA)技术、长短时记忆网络(LSTM)和注意力(Attention)机制。首先,采用时域分析来手动提取原始数据中的多维时间特征,并利用PCA技术对其进行特征融合处理;然后,利用LSTM网络从样本数据中自动学习序列特征,引入的Attention机制能够对更重要的特征和时间步长赋予更大的权值。最后,使用NASA Ames实验室加速老化数据库进行实验,结果表明所提方法优于最新方法。手动提取的时间特征在经过特征融合后,可以作为序列数据预测任务中的有效退化特征,并结合Attention机制大大提高预测精度。

Aiming at the problem that it is difficult to make full use of time information in reliability analysis task of insulated gate bipolar transistor(IGBT),resulting in low prediction accuracy,a deep learning method based on multi-dimensional features and attention mechanism is proposed.This method combines principal component analysis(PCA),long and short-term memory network(LSTM)and attention mechanism.Firstly,time domain analysis is used to manually extract multi-dimensional time features from the original data,and PCA technology is used for feature fusion.Then,the LSTM network is used to automatically learn sequence features from sample data.The introduced attention mechanism can learn the importance of features and time steps,and give greater weights to more important features.Finally,the prediction accuracy of the model is improved by combining the manually extracted features with the automatically learned features.Experiments are carried out using NASA Ames laboratory accelerated aging database,and the results show that the proposed method is better than the latest method.