| 引用本文: | 陈淑梅,余建波.卷积神经网络多变量过程特征学习与故障诊断[J].哈尔滨工业大学学报,2020,52(7):59.DOI:10.11918/201906120 |
| CHEN Shumei,YU Jianbo.Feature learning and fault diagnosis in multivariate process with convolutional neural network[J].Journal of Harbin Institute of Technology,2020,52(7):59.DOI:10.11918/201906120 |
|
| 摘要: |
| 为提取复杂多变量过程的有效特征,提高故障诊断性能,提出一种基于卷积神经网络（convolutional neural network,CNN）特征学习的多变量过程故障诊断模型. 将高维过程信号归一化处理转为图像信号,多层卷积滤波器与子采样滤波器交替构成的轻量级CNN网络通过多个卷积核与图像进行卷积,采用本地连接和权重共享,滤除过程噪声和干扰信息,从而获得过程数据的高层抽象化表达. 通过Softmax层有监督的微调方式学习故障特征完成故障诊断. 利用以田纳西过程为代表的多变量非线性过程验证了模型的有效性,与经典分类器和近几年流行的深度神经网络进行对比, 结果表明:将高维过程信号转为图像信号输入CNN提高了多变量过程的故障诊断精度；通过t-SNE方法对模型提取的特征进行可视化分析,说明模型强大的特征提取能力；将模型提取的特征作为传统分类器的输入时,故障识别准确率显著提升,进一步说明有效的特征提取有利于提高故障诊断的准确度和可靠性；与无监督学习方式相比,模型通过标签能获取更有效、稳定和抽象化的数据特征. |
| 关键词: 多变量过程 故障诊断 卷积神经网络 特征学习 田纳西过程 |
| DOI：10.11918/201906120 |
| 分类号:TP277 |
| 文献标识码:A |
| 基金项目:国家自然科学基金(71777173) |
|
| Feature learning and fault diagnosis in multivariate process with convolutional neural network |
|
CHEN Shumei,YU Jianbo
|
|
(School of Mechanical and Energy Engineering, Tongji University, Shanghai 201804,China)
|
| Abstract: |
| A multivariate process fault diagnosis model is proposed based on convolutional neural network (CNN), aiming at extracting effective features from complex multivariate processes and improving fault diagnosis performance. First, the high-dimensional process signals are normalized and then converted into images. Second, a lightweight CNN network composing of multi-layer convolution filters and sub-sampling filters is convolved with images through multiple convolution kernels, using local connections and shared weights to remove noise and interference information to obtain the high-level abstract representations of process data. Finally, a Softmax layer is used in a supervised way to implement fault diagnosis. Tennessee Eastman Process is used to verify the effectiveness of proposed model and compare the performance between the proposed model with classical classifiers and deep neural networks. The results show that the fault diagnosis accuracy is improved by converting high-dimensional process signals into images. The t-SNE visualization analysis method is used to illustrate the powerful feature extraction ability of proposed model. The features extracted by the proposed model are sent to the traditional classifiers and the accuracy of fault identification is significantly improved, which further illustrates the benefit of effective feature extraction for improving the fault diagnosis accuracy and reliability. Compared to unsupervised learning, the proposed model with the guidance of label helps to extract more efficient, stable, and abstract feature representations. |
| Key words: multivariate processes fault diagnosis convolutional neural network feature learning Tennessee Eastman Process |
