Abstract:

Aiming at the shortcomings of existing methods for predicting remaining life of small-sample high-reliability equipment， such as low accuracy and poor prediction generalization， a new assessment method is proposed. This method uses a jointly trained high-order neighborhood aggregation graph convolutional neural network and a bidirectional gated recurrent unit （GRU）. This method constructs an attribute graph from large public datasets combined with small sub-sample information， such as equipment’s key electrical potentials. The HoNA-GCN model integrates different orders of neighbor information to capture the high-order associative features between equipment samples. Life prediction is then carried out using the bidirectional GRU， with a pretraining-finetuning strategy employed to enhance the model's generalization capabilities. The precision of equipment residual life prediction has been improved， along with generalization across different scenarios. Simulations and ablation studies validate the necessity of each module of the method. Results indicate a significant enhancement in the accuracy and robustness of predictions compared to those of other classic methods. By effectively utilizing the correlation information between public datasets and small sub-sample data of equipment， a new solution is offered for the state assessment of complex equipment systems.

Key words: remaining life prediction and evaluation, graph convolution neural networks, high order neighborhood aggregation, bidirectional gated recurrent unit, joint training

摘要：

针对现有小子样高可靠性装备剩余寿命预测方法精度较低、预测泛化性较差等不足，提出了一种基于高阶邻域聚合图卷积神经网络和双向门控单元联合训练的装备剩余寿命评估方法。该方法将公开的大样本数据集和装备关键部件测试数据等小子样本信息构建为属性图，整合不同阶邻居信息，捕获装备采样信息间的高阶关联特征，再使用双向门控单元进行寿命预测，并通过预训练-微调的联合训练策略提升模型泛化能力。提升了装备剩余寿命预测的精度，提升了不同场景下的寿命预测的泛化性，并通过仿真实验和消融实验证明了方法各个模块的必要性。与其他经典方法相比，该方法预测的准确性和稳健性均有显著提升。有效利用了公开数据集和装备小子样数据之间的关联信息，为复杂装备系统的剩余寿命预测评估提供了一种新的解决方案。

关键词: 剩余寿命预测评估, 图卷积神经网络, 高阶邻域聚合, 双向门控循环单元, 联合训练