基于深度迁移学习的柴油机故障诊断研究

Diesel engine fault diagnosis based on deep transfer learning

得益于大数据和人工智能的高速发展,数据驱动的智能故障诊断方法受到广泛关注。然而,在柴油机故障数据稀缺的情况下,传统神经网络训练容易出现过拟合且网络泛化能力差。为解决上述问题,提出一种基于深度迁移学习的小样本故障诊断方法。构建一种适用于柴油机原始振动信号的宽卷积核卷积长短期记忆神经网络,来提高故障数据特征提取和抗噪的能力,另外从原始数据自动提取特征,增强特征学习的智能性。进一步采用迁移学习方案,将大型标签源域数据的诊断知识迁移到目标域网络上,改进网络在目标域任务小样本条件下的学习和分类能力。在跨故障域和跨设备域迁移任务上进行算法评估,并与传统深度神经网络进行比较,验证了所提方法可有效改进小样本诊断性能。

Being profited from rapid development of big data and artificial intelligence,data-driven intelligent fault diagnosis methods receive wide attention.However,in cases of scarce diesel engine fault data,traditional neural network training is easy to overfit and has poor network generalization ability.Here,to solve the above problems,a small sample fault diagnosis method based on deep transfer learning was proposed.A wide convolution kernel convolution long short-term memory neural network suitable for diesel engine original vibration signals was constructed to improve the ability of fault data feature extraction and noise resistance.In addition,features were automatically extracted from original data to enhance the intelligence of feature learning.Further,transfer learning scheme was adopted to transfer diagnostic knowledge of large type label source domain data into target domain network,and improve network's learning and classification ability under the condition of target domain task's small sample.The algorithm of the proposed method was evaluated with transfer tasks in cross-fault domain and cross-device domain,and the results were compared with those obtained with traditional deep neural network.It was verified that the proposed method can effectively improve the diagnosis performance of small sample.