With the ever-stringent vehicles exhaust emission standard and higher requirements on on-board diagnostic technology, the importance of misfire detection in vehicle emission control is emerging. The performance of a traditional misfire detection algorithm predominantly depends on the features and classifier selected. Fixed and handcrafted features require either a reliable dynamic model of an engine or a large number of experiment data to define the threshold, and then, form a map. Since convolutional neural networks (CNNs) have an inherent adaptive design and integrate the feature extraction with classification functions into a compact learning framework, the misfire fault-sensitive features can be auto-discovered from raw speed signals. Furthermore, CNNs can detect the fault features of the misfire through network training with fewer engine operating conditions. In this paper, the theory and method of the misfire diagnosis based on CNNs are presented. The experimental data for network training and testing are sampled on a six-cylinder inline diesel engine. The misfire patterns containing every one-cylinder and two-cylinder misfire are tested under the wide speed and load conditions of the engine. The results show that when the engine operates under steady-state conditions, one-cylinder or two-cylinder complete misfires can be detected accurately by CNNs. In addition, one-cylinder partial misfire is employed to examine the adaptability of trained 1-D CNN. It turns out that when the partial misfire reaches the same level as half amount of the normal fuel injection quantity, one-cylinder partial misfire can be detected with accuracy more than 96%. At last, the misfire detection under the non-stationary conditions, such as acceleration or deceleration, is conducted. The results show the 1-D CNN performed well in a limited acceleration range, and network failure occurs when the absolute acceleration of the engine speed is more than 100 r/min/s.
