Application Of Machine Learning In The SCR On Board Diagnosis Of6K12Diesel Engine

As the environmental issues are increasingly concerned and the implementationof extremely stringent emission regulations, it is currently becoming one of theheated topics to improve the performance of diesel engine exhaust emissionafter-treatment system. The after-treatment control and OBD system(On-Board-Diagnostic) of the diesel engines have great significance to the exhaustperformance. The traditional OBD needs to monitor and diagnose the state of thevehicle, however, OBDII is required to measure the exhaust emission meanwhile,which is certainly a new challenge to the traditional OBD.In this paper, the author is attempting to obtain the characteristics of the NOxreduction reaction under the complex conditions in the methods of machine learningaccording to the6K12diesel engine SCR system and then apply the efficiency ofchemical reaction of NOx to diagnose algorithm in the OBDII system, which tendsto be more accurate and efficient, compared to the traditional calibration method.This paper clarifies specifically the requirements of the diagnosis algorithmby conducting a thorough analysis of the emission regulations of heavy-duty dieselengine and the OBD diagnostic processes and requirements.According to the regulations, if the emissions of NOx is more than3.5g/(kw.h) or less than7.0g/(kw.h),OBD should light the Malfunction Indicator Lamp(MIL); if the emissions of NOx is greater than7.0g/(kw.h), OBD should start thetorque limiter. The main difficulty of the diagnosis by OBDII is the uncertainty of the start and end point signal of the load condition while the engine is running.Both emission limit values are high transient emissions in ETC cycle, instead of theOBD cycle, which is chosen as the test cycle in practice. In addition, the conditionof reaction efficiency, which is an influential factor in the calculation of efficiency, isnot uniform in OBD. Accordingly, it is impossible to diagnose the reaction efficiencyaccurately by conducting direct efficiency diagnosis.Based on the above requirements, this paper constructs the basic structure andmethod of diagnosis algorithm. Its basic ideas are as follows:1. The runningcondition can be divided into a plurality of time slices. The correlation of these timefragments and the corresponding ETC circulation limit value could be constructedand those fragments with higher relevance with ETC circulation are required to befiltered by establishment of certain filtering algorithm.2. Multiple filtered timefragments could be used to obtain the eventual results of diagnosis classification.As to the segmentation of running condition and the analysis of emission valueof ETC circulation, the off-line trained neural network is adopted and theresults of the correlation of the efficiency between time slices and emissionvalue of ETC circulation are demonstrated in mapping network, which is appliedin the online diagnosis of filtering time slices. The SVM algorithm is employed inthe whole process of obtaining final classification diagnosis. The simulation resultsshow that the proposed diagnosis algorithm has a good diagnostic performance withstrong adaptability. Being not in need of complex manual calibration, this algorithmcan be realized by automatic machine calibration, which has a really excellentprospect in engineering due to its potential portability and scalability.