面向海量机械故障数据的胶囊网络算法研究

Research on Capsule Network Algorithm for Massive Mechanical Fault Data

针对卷积神经网络算法在大规模故障数据集检测中出现的故障敏感度低、部分特征丢失等问题,提出一种基于优化胶囊网络算法的机械故障检测方案。胶囊网络算法采用多神经元封装的胶囊体结构设计,且包含多个胶囊层,具有更强的故障数据处理能力和泛化能力;经过squash函数挤压后的胶囊矢量可以更准确地提取和描述故障特征;升维胶囊矢量,基于特征编码和归一化的处理方式,可得到更准确的故障分类结果。实验结果显示:优化胶囊网络算法具有更强的故障特征聚类性能和迭代运算性能,故障集检测精度值高于经典卷积神经网络算法。

Aiming at the problems of low fault sensitivity and easy loss of some features in large⁃scale fault data set detection based on convolutional neural network algorithm,a fault set detection scheme based on optimized capsule network algorithm was proposed.In the capsule network algorithm,multi neuron encapsulated capsule structure design was used,and multiple capsule layers were con⁃tained.So the algorithm had stronger fault data processing ability and generalization ability.The capsule vector extruded by squash func⁃tion could be used to extract and describe fault features more accurately.More accurate fault score result could be gotten by improving capsule vector and based on feature coding and normalization class.The experimental results show that the optimized capsule network algorithm has stronger fault feature clustering performance and iterative operation performance,and the fault set detection accuracy is higher than that of convolution neural network algorithm.