基于TARO模型的设备剩余寿命预测方法

A Remaining Useful Life Prediction Method Based on the TARO Model

摘要:

随着工业设备服役周期延长，剩余使用寿命预测精度不足严重威胁运维安全与经济性；为了提高预测准确度，提出一种基于Transformer自编码器融合双损失策略的剩余寿命预测模型—TARO；使用改进后融合时间卷积网络的Time2Vec嵌入层充分提取特征信息；利用Transformer作为数据重构与预测骨干网络，设计将编码器隐藏层特征同时引入到重构编码器与回归解码器，进而融合重构误差损失与回归损失，提升复杂工况下预测精度；基于CMAPSS航空发动机数据集的对比实验表明，提出的方法在上述指标上具有最优的表现，有效提高了在复杂工况下和复杂故障下的剩余寿命预测精度。

Abstract:

As industrial equipment ages, the limited accuracy of remaining useful life prediction poses a significant threat to the safety and economic efficiency of operation and maintenance. To address this issue, this paper presents TARO, a novel RUL prediction model that incorporates a Transformer auto-encoder with a dual-loss strategy. An enhanced Time2Vec embedding layer, integrated with a temporal convolutional network, is employed to comprehensively extract feature information. The Transformer serves as the core network for data reconstruction and prediction. Specifically, the features of the encoder's hidden layer are fed into both the reconstruction encoder and the regression decoder. By fusing the reconstruction error loss and the regression loss, the model's prediction accuracy is enhanced under complex working conditions. Comparative experiments conducted on the CMAPSS aero-engine dataset demonstrate that the proposed method outperforms others in relevant metrics, effectively improving the RUL prediction accuracy under complex working conditions and in the presence of complex faults.