小训练样本下齿轮箱故障诊断:一种基于改进深度森林的方法

Gearbox fault diagnosis with small training samples:An improved deep forest based method

作为航空装备的重要传动部件,齿轮箱的故障诊断对保障装备可靠持续适航具有至关重要的作用。随着人工智能技术的不断发展,基于深度学习的方法成为了领域内的研究热点。然而,深度神经网络对超参数设置和训练数据量有严格的要求,难以满足实际工业中快速、准确与稳定的诊断需求。针对此问题,提出了一种基于改进深度森林的诊断方法,实现小训练样本下齿轮箱的多种类混合故障的高效诊断。针对旋转机械振动信号单样本数据的长特性与深度森林模型数据处理成本高的矛盾,设计了基于主成分分析特征提取的深度森林模型,解决原始模型中的数据计算冗余问题。同时,改进的深度森林模型提高了多粒度扫描与级联森林中的数据传递与处理能力,在保障数据多样性的同时,增强模型内的特征代表性,从而提高算法的运行效率和诊断性能。最后,通过控制数据集与训练样本比例变量,开展小训练样本下齿轮箱故障诊断实验研究。结果表明,在训练-总数据比例为50%和10%条件下,所提方法平均诊断精度高达97.3%和82.8%,验证了所提方法的有效性。同时,通过对比研究,所提方法诊断性能优于现有的齿轮箱智能故障诊断方法。

The gearbox is an important transmission part of complex aero-equipment,and fault diagnosis of gearbox plays a vital role in ensuring reliable and continued airworthiness.With the continuous development of artificial intelligence technology,deep learning-based methods have become a research hotspot in this field.However,deep neural networks have strict requirements on hyperparameter settings and training data volume,and are difficult to meet the requirements of fast,accurate and stable diagnosis in the actual industry.To solve these problems,a diagnostic method is proposed based on the improved deep forest to realize the efficient diagnosis of the multiple and mixed faults of gearbox with small training samples.For the contradiction between the long characteristic of the single-sample data in the rotating machinery vibration signal and the high cost of data processing in the deep forest model,a deep forest model is designed based on principal component analysis(PCA)feature extraction to solve redundant calculation of data in the original model.The improved deep forest enhances data transmission and processing capability in multi-grained scanning and cascade forest.While ensuring diversity of data,it enhances the representativeness of the internal features in the model,so as to improve the operating efficiency and diagnostic performance of the algorithm.Experiments on gearbox fault diagnosis under small training samples are carried out by controlling the variable of the ratio of total datasets to training samples.The results show that average diagnostic accuracy of the proposed method reaches 97.3%and 82.8%under the condition of 50%and 10%training-total dataset ratio,respectively,which verifies the effectiveness of the proposed method.The diagnostic performance of the proposed method is found to outperform the existing intelligent diagnostic methods for gearbox.