基于深度主动学习的示功图诊断方法及应用

Diagnostic Method and Application of Indicator Diagram Based on Deep Active Learning

示功图是衡量抽油机作业工况的重要依据,基于深度学习开发的示功图诊断方法大幅提高了自动化检测的精度。由于深度学习的图像分类方法需要大量有标签图片进行训练,因此模型精度易受限于训练数据的数量与质量。基于此,提出了一种基于主动学习的示功图诊断方法,一方面基于迁移学习,通过预训练深度卷积神经网络的先验知识提高模型初始化性能;另一方面基于深度主动学习,有效地挖掘出新样本以扩充训练集。实验表明:该方法可以较好地提升示功图诊断模型的精度,相较于手工标注,大幅度降低了人工成本。

The indicator diagram is an important basis for measuring the working conditions of the oil pumping.The accuracy of automatic detection is improved greatly by using the indicator diagram diagnosis method based on deep learning.A large number of labeled images are in need for the image classification methods based on deep learning for training,so the accuracy of the model is easily limited by the quantity and quality of training data.An indicator diagram diagnosis method based on active learning is proposed,on one hand,based on transfer learning the model initialization performance is improved through the prior knowledge of pre-trained deep convolutional neural networks,on the other hand,based on deep active learning,new samples can be effectively mined to expand the training dataset.Experiments show that the method can better improve the accuracy of the indicator diagram diagnosis model,labor costs can be greatly reduced comparing with manual labeling.