基于堆栈降噪自编码网络的机械设备磨损状态识别

Wear state recognition of mechanical equipment based on stacked denoised auto-encoding network

  • 摘要: 通过磨粒铁谱图像识别可实现机械设备磨损状态识别,但基于机器学习的磨粒铁谱图像识别存在较多人工干预、普适性差。为解决上述问题,提出了一种基于堆栈降噪自编码网络的机械设备磨损状态识别方法。将多个降噪自编码网络进行堆叠,即上一个降噪自编码网络隐含层的输出作为下一个降噪自编码网络的输入,并在最后一个降噪自编码网络隐含层后添加Softmax分类器,从而构建堆栈降噪自编码网络;利用磨粒铁谱图像对堆栈降噪自编码网络进行无监督预训练,并通过有监督微调优化网络参数,对磨粒铁谱图像进行分类,实现机械设备磨损状态智能识别。实验结果表明:当堆栈降噪自编码网络的激活函数为Relu、优化器为Adam、学习率为0.001时,网络性能最佳,识别准确率达98.43%。

     

    Abstract: Wear state recognition of mechanical equipment can be realized by image recognition of ferrography image of wear particle, but ferrography image of wear particle recognition based on machine learning has more manual intervention and poor universality. In order to solve the above problems, a wear state recognition method of mechanical equipment based on stacked denoised auto-encoding network was proposed. Multiple denoised auto-encoding networks are stacked, that is, the output of hidden layer of upper-level denoised auto-encoding network is taken as the input of the next-level denoised auto-encoding network, and Softmax classifier is added after hidden layer of the last level denoised auto-encoding network, so as to construct the stacked denoised auto-encoding network. The ferrography images of wear particle are used for unsupervised pre-training of stacked denoised auto-encoding network, network parameters are optimized by supervised fine-tuning, and ferrography images of wear particle are classified to achieve intelligent wear state recognition of mechanical equipment. The experimental results show that the stacked denoised auto-encoding network achieves the best performance when activation function is Relu, optimizer is Adam and learning rate is 0.001, and recognition accuracy is 98.43%.

     

/

返回文章
返回