基于改进类别嵌入发掘网络的液压泵零样本故障诊断

Zero-shot Fault Diagnosis of Hydraulic Pump Based on Improved Visually-Grounded Semantic Embeddings

当引入类别嵌入发掘(VGSE)对液压泵进行零样本故障诊断研究时,该模型不具备收敛性、泛化性和稳定性,且存在计算效率低等问题,因此提出了改进类别嵌入发掘网络(SCP-VGSE)。为了改善VGSE网络性能,将SE注意力机制嵌入到网络结构的特征提取功能模块中;为了进一步改善网络性能,引入CAME优化器替换Adam优化器,对网络的学习率、权重等超参数进行优化处理;最后,利用粒子群算法对网络的学习率进行优化。通过液压泵实测零样本故障实验验证分析可知,所提改进方法提升了模型的收敛性、泛化性和稳定性,实现了高达96%的收敛精度,且诊断效率提升了68.75%。

When visually-grounded semantic embeddings(VGSE)is introduced to zero-shot fault diagnosis of hydraulic pump,the net has some problems,such as lack of convergence,generalization and stability,and low computational efficiency.Therefore,an improved VGSE of SCP-VGSE was proposed.In order to improve the VGSE network performances,the squeeze-and-excitation(SE)mechanism was embedded in the feature extraction function module of the network.In order to further improve network performances,CAME optimizer was introduced to replace Adam optimizer to optimize the hyperparameters such as learning rate and weight of the network.Finally,the particle swarm optimization was used to optimize the learning rate of the network.Through the experimental verification and analysis of zero-shot fault of hydraulic pump,it can be seen that the improved method improves the convergence,generalization and stability,and the convergence accuracy achieves up to 96%,and the diagnostic efficiency is increased by 68.75%.