自适应在线增量ELM的故障诊断模型研究

为满足现役装备根据故障样本数据集积累的特点进行自适应故障诊断的需求, 本文将极限学习机(extreme learning machine, ELM) 的数据增量学习、隐藏层增量学习和输出层增量学习(类增量学习)3种增量学习模式, 融合到一个统一的学习框架内, 提出一种凸最优自适应增量在线顺序ELM(convex optimal adaptive incremental online sequential ELM, COAIOS-ELM)。模型能够根据增量学习中误差的变化情况, 自适应地增加隐藏层神经元, 减小分类误差; 并可根据增量数据集中新出现的故障类别, 进行相应的类增量学习, 增加故障诊断的范围。有效解决了ELM增量学习过程中模型自适应动态选择最佳网络结构的问题, 提高模型的故障诊断的精度和故障诊断的范围。本文选择UCI数据集中公共数据集和Biquad低通滤波电路故障诊断数据集, 通过与类增量ELM (class incremental ELM, CI-ELM)模型对比实验, 验证了所提方法的有效性。

Research on an adaptive online incremental ELM fault diagnosis model

In order to meet the needs of active equipment for adaptive fault diagnosis based on the characteristics of the accumulation of fault sample data sets, this paper uses three types of incremental learning of extreme learning machine (ELM) data incremental learning, hidden layer incremental learning and output layer incremental learning (class incremental learning) The learning mode is integrated into a unified learning framework, and a convex optimal adaptive incremental online ELM (COAIOS-ELM)is proposed. The model can adaptively increase hidden layer neurons according to the change of the error in incremental learning to reduce the classification error; and can perform corresponding class incremental learning according to the newly appeared fault category in the incremental data set to increase fault diagnosis range. It effectively solves the problem of model adaptive and dynamic selection of the best network structure in the process of ELM incremental learning, and improves the accuracy and scope of fault diagnosis of the model. This paper selects the public data set in the UCI data set and the Biquad low-pass filter circuit fault diagnosis data set, and verifies the effectiveness of the proposed method by comparing experiments with the class incremental ELM (CI-ELM) model.