AI ML Tool for Enhancing RDE Trip Validity G. Sathya Narayanan* 1, Rajesh Kibile 2, Milind Bhamare 1, Kedar Marathe 2 1Tata Motors Ltd, India 2Tata Technologies, India KEYWORDS – Real Driving Emissions, Trip validity, Trip Normality, Artificial Intelligence, Machine Learning. ABSTRACT - Objective: Real Driving Emission test, is the most important regulatory standard to be met for BS6 compliance. RDE tests verify compliance to emission limits during on-road trials. Government authorities have set ~30 boundary conditions on conducting these trials to ensure unbiased emission assessment. Failing to meet even 1 condition renders the trial invalid and needs to be repeated. The success rate of executing a valid RDE test was just 30% prior to the use of our innovation. To ensure high success rate, there needed to be an intelligent system that continuously directs the driver by analyzing and forecasting emission patterns. Methodology: To ensure high success rate, historical data of PASSED & FAILED trials is fed as input to an AI engine. The engine builds relationship between driving behavior, terrain patterns, traffic conditions and corresponding emission patterns. CO2 emissions during a RDE trial are highly retrospective. Hence, CO2 patterns during a trial may get affected by something that driver did during very early phase of a trial. To train a machine learning model which takes into account such long-term effects, a novel 3-step approach is developed. The trained ML model is deployed on a laptop connected to vehicle systems. During the trial, by accessing vehicle & emission data in real-time, the system characterizes vehicle’s behavior, predicts future emissions based on current driving style, and provides actionable instructions to driver. The system is adaptive to ever-changing parameters that impact emissions, and generates assistance that ensures valid trial under any environmental and traffic conditions. Results: Post tool development, RDE trip validity has increased from 35% to 75%, which has benefitted to expedite RDE development and meet program timelines with less number of RDE tests. On broad level, this tool helps significantly to optimize our vehicles to perfect trade-off point, where we achieve robust RDE emissions along with optimal fuel efficiency (CO2) & benefit on reduction of greenhouses gases. Limitations of the study: Currently the AI model trained on historical PASS/FAIL data, is unique to every vehicle. For any major change in the powertrain, or vehicle, the model needs to be retrained and fine-tuned based on 2-3 RDE trial data. Once the model is retrained, it can be used for further RDE trials. Based on historical data across various vehicle models augmented with numerical representation of vehicle aggregates and properties like weight, shape, etc. we are in a process of training a generic AI model which will work across various vehicles without the need of fine-tuning. Uniqueness • Going beyond real-time status: Apart from only showing the real-time status, the system predicts emission patterns in the future and assists the driver to avoid non-compliant behavior. The AI-algorithm models driver's never-before-seen driving behavior, effects of terrain, traffic, and engine calibration to achieve this. • Assistance not depending on route/recipe: Humidity, temperature, traffic conditions etc. keep on changing during trials, and hence it is impossible for a driver to follow a pre-defined recipe to achieve validity. Our system ensures valid trial under any environmental /traffic conditions. • The unique 3-step modelling approach is unavailable in any commercial AI-ML platform. Conclusion: The technical challenge of complying the CO2 based trip normality requirement on the test vehicle was successfully addressed with the AI-ML based tool by which the real-time monitoring and prediction of futuristic CO2 values was made possible, along with simplest voice commanding option was used for providing real-time guidance for the driver to align his driving style for meet the CO2 normality requirement during on-going trial itself. This has significantly improved the trip validity from 35% to 75%.
Mr. SATHYA NARAYANAN GOVINDARAJAN, Engineering Manager, Tata Motors Ltd