基于深度学习的DPF故障诊断及应用

DPF fault diagnosis and application based on deep learning

针对传统柴油颗粒过滤器(diesel particulate filters,DPF)故障诊断中人工提取特征过程繁杂且特征参数难以表征DPF故障状态的问题,提出基于深度学习的DPF故障诊断方法。通过车载传感器采集发动机转速、DPF压差等5种信号数据,经数据融合后建立压差类、温差类和压差+温差类三类样本数据;利用深度学习特征自提取的优势,提取车辆在不同行驶工况下样本数据中压差或温差表征的DPF故障特征;结合深度学习网络中Softmax多分类器实现端到端的DPF故障诊断。利用GT-Power仿真数据,验证了所提方法的可行性,开发了相应的应用软件。

To address the problem that the traditional diesel particulate filters(DPF) fault diagnosis method is tedious to extract features manually and the feature parameters are difficult to characterize the complex DPF fault states, a deep learning-based DPF fault diagnosis method is proposed. Firstly, five kinds of signal data such as engine speed and DPF differential pressure are collected by on-board sensors, and three types of sample data are established after data fusion: differential pressure class, differential temperature class and differential pressure plus differential temperature class;secondly, the DPF fault features characterized by differential pressure or differential temperature are extracted from the sample data under different driving conditions by using the advantage of deep learning feature self-extraction;finally, the end-to-end DPF fault diagnosis method is implemented by combining the Softmax multi-classifier in the deep learning network. Experiments using GT-Power simulation data verify the feasibility of the proposed method: for single or multiple on-board sensor signal data, one-dimensional convolutional neural networks or feature fusion convolutional neural networks are used to achieve the diagnosis of four types of DPF fault states, respectively. For engineering applications, the corresponding Application software is developed.