Research on fault diagnosis and signal reconstruction technology of diesel engine NOx sensor based on deep learning algorithm

Abstract

Modern diesel engines are commonly equipped with NOx sensors for precise and efficient selective catalytic reduction (SCR) reductant injection control, but due to the limitations of operating characteristics and environmental conditions, the output signal of NOx sensors may show abnormalities such as drift, deviation, or even distortion. Therefore, the evaluation and adjustment of the NOx sensor signal is essential to achieve intelligent control of the engine SCR system. In this study, an engine NOx calculation model is designed using deep neural network (DNN) to predict the upstream NOx emission concentration of engine SCR and provide reference values for sensor fault diagnosis, evaluate the sensor fault status by calculating the relative integration error index (IE), and reconstruct the sensor signal using the model prediction signal after the fault is confirmed. In addition, this paper uses the firefly algorithm (FA) to find the appropriate diagnostic threshold (IEthre) and integration period ( T) to meet the requirements of diagnosis response speed. Finally, based on the data collected from World Harmonized Transient Cycle (WHTC) bench test of a commercial high-pressure common rail diesel engine, the IEthre = 12.76% and the T = 7.57 s. The root mean square error (RSME) of the NOx concentration signal predicted by the engine NOx calculation model was 37.35 ppm, and the normalized cumulative NOx emission was 98% of the actual signal. The performance of the designed algorithm was tested using the synthesized bias fault and drift fault signals. The sensor signal RSME was reduced from 108.46 to 29.66 ppm for the bias fault, and from 106.92 to 31.30 ppm for the drift fault.