随机卷积神经网络的内燃机健康监测方法研究

Health monitoring method of the internal combustion engine based on the Random Convolutional Neural Networks

提出了一种深度学习网络即结构⁃随机卷积神经网络,用于实现内燃机的智能健康监测。集成多个单一的卷积神经网络构成新的网络构架,通过卷积计算和池化操作对振动信号进行自动特征提取。在随机卷积神经网络结构中,应用Adabound优化器使学习率自适应变化,从而加快模型的权值更新。同时通过Dropout技术使结构中神经元随机失活,防止对训练数据进行复杂的协同响应,通过Dempster合成规则融合各个网络模型的诊断结果。所提出的健康监测方案用于对内燃机工作时的振动信号进行分析。实验证明,该方法克服了传统的数据驱动和卷积神经网络健康监测方法的局限性,摆脱了对人工特征设计的依赖,并且有较好的诊断性能,能够有效地对内燃机进行健康监测。

Automatic and accurate identification on health condition of internal combustion engine system is still a major challenge in modern industry because of its complex mechanical system.In this paper,an innovative deep learning model called Random Con⁃volutional Neural Networks(RCNNs)is proposed for intelligent health monitoring of internal combustion engine.This novel net⁃work framework is constructed with several individual convolutional neural networks,which can automatically extract the feature of vibration signals by convolutional calculation and pooling operation.An improved optimizer Adabound and the technique of Drop⁃out are applied in this framework of RCNNs.The Adabound optimizer uses adaptive learning rates to adjust the network’s weight.Meanwhile,the Dropout technique makes neurons drop out with a probability in order to preventing complex co-adaptations on training data.The Dempster’s combinational rule is used to obtain the fusion diagnosis results from several individual networks.The proposed health monitoring scheme is used to analyze the experimental vibration signals acquired from engine.Experiments prove that the proposed RCNNs can overcome the limitations of health monitoring method based on traditional data-driven or con⁃volutional neural network,which gets rid of dependence on the manual feature design and delivers state-of-the-art performances.Therefore,the proposed RCNNs method is suitable for machine health monitoring.