Simple telematics devices known as OBD “dongles” are being used for a wide range of applications, including driver insurance programs, boundary and speed alerts for young drivers, and powertrain diagnostics. SGS has explored the potential for another application, using OBD dongle data to predict fuel consumption and tailpipe exhaust emissions. In this study, SGS accurately measured instantaneous fuel consumption and emissions in the laboratory and on the road using PEMS technology. We then employed an advanced analytical technique known as “machine learning” to discover the relationship between engine sensor data and exhaust emissions. The machine learning approach showed promise to predict fuel consumption and emissions more accurately, and could be used to augment government Remote OBD and emissions inventory modeling programs.
1. 1
On the Efficacy of Predicting Light Duty Vehicle
Fuel Consumption and Exhaust Emissions
using SAE J1979 CAN Data
I/M SOLUTIONS 2017
SGS Transportation
Keith Vertin and Brent Schuchmann
Email: keith.vertin@sgs.com
May 23, 2017
2. 2
Light Duty Vehicle
Fuel Consumption & Emissions Prediction
State I/M programs have adopted
OBD inspections and have reduced
tailpipe emissions testing
EPA recognizes Remote OBD as a
continuous monitoring approach
Low cost dongles provide a viable
means to transmit Diagnostic Trouble
Codes and I/M readiness
Telematics/dongles can also log and
transmit time-series vehicle data
(SAE J1979 Mode 01 data)
Can fuel consumption and exhaust
emissions be accurately estimated
using Mode 01 data?
3. 3
Benefits for
Predictive Fuel Consumption and Emissions
Traffic network impacts on energy and the environment
OBD provides granular data that enables seasonal and diurnal
emissions trends analysis
Real-time feedback for Connected Vehicle driver assistance
features
Individual vehicle data may have a role for future improvement
of emissions modeling and validation activities
– “Micro scale” data processed on big data platforms
– Potential data source for MOVES, to support emissions
inventories and State Implementation Plans
4. 4
Data Sources for Emissions Rate Models
MOVES 2014a Data Sources for
Emissions Rates (historical)
I/M Lanes
In Use Verification Program (IUVP)
Mobile Source Observation Database
Government sponsored studies
Potential Future Data Sources for Activity
and Emissions (real world micro scale)
Roadside Sensing
Remote OBD derived information
PEMS
MOVES emissions rates grouped by
Vehicle Specific Power operating modes
5. 5
Method Approach Issues
Gasoline vehicles with MAF
Fuel consumption only:
Mass Air Flow / AFR * Equivalence
Ratio
MAF sensor not first principles
measurement, transfer function is OEM
specific
Gasoline vehicles without MAF
Fuel consumption only:
Ideal Gas Law using stoichiometric
combustion and volumetric efficiency
Volumetric efficiency assumption
Regression
Multivariate linear regression
Nonlinear: a * (Eng Speed * MAP)^b
Form of equation may not fit data for
diverse vehicle model possibilities -
different forms for different vehicles
Vehicle simulation models
Powertain system models including
major components and control features
Component design and performance
information must be known and
specified
Machine Learning
Pattern recognition models using
engine sensor data
Larger computational requirements
Micro Scale Estimation of
Fuel Consumption and Emissions
This study explores another possibility for modeling -
*MAF = Mass Air Flow, AFR = Air Fuel Ratio, MAP = Manifold Absolute Pressure
6. 6
Previous Emissions Modeling Studies
Using Time Series OBD Data
Several studies have been published, but there is sparse information about
machine learning models using lab-grade PEMS data
Emissions prediction is challenging for low emission vehicles, as shown for
this exponential equation solution = a * (Eng_Speed * MAP)b
Source: “Comparison of Vehicle-Specific Fuel Use and Emissions Models Based on Externally and Internally Observable Activity Data”,
Hu, Frey, Washburn, 2015. Note R-squared values are for model fits, and not for blind prediction using new data.
7. 7
Our Study:
Vehicle Testing On-Dyno and On-Road
MY 2013 Jeep Wrangler, 3.6L V6, PFI, EPA Tier 2 Bin 4, no MAF
Chassis dyno emissions testing at SGS in Aurora, CO using standard
emissions cycles (FTP, HWFET, US06, SRC) and one real-world cycle
AVL MOVES 483 Portable Emissions Measurement System
(emissions certification grade PEMS)
8. 8
Time Series Data (1 Hz)
Split into Micro Trips for Analysis
On Dynamometer
122 micro trips
3.1 hours of operation
On Road
93 micro trips
3.8 hours of operation
9. 9
Dynamometer and On-Road Vehicle Operation
On-road testing in Denver
metro area included modes of
operation similar to the dyno
tests
– Cold Start
– City
– Highway
– Rapid accelerations
On-road testing also included
mountain drives not simulated
in the dyno lab
The on-road testing had
greater variation in fuel
consumption as expected
10. 10
Machine Learning Approach
[TRAIN] Train model using dyno laboratory time-series data only
[TEST] Predict on-road vehicle operation
Make predictions using blinded OBD test data
Independent OBD-II Parameters*
Engine Speed
Intake Manifold Absolute Pressure
Ambient Air Temperature
Intake Air Temperature
Long Term Fuel Trim
Equivalence Ratio
Spark Timing
Exhaust Gas Temp (Catalyst Inlet)
Barometric Pressure
Coolant Outlet Temperature
*Correlated predictors such as load, torque and pedal
position removed
Dependent Parameters
Fuel Consumption
CO
NOx
THC
Diagonal Line
and R2 = 1.0
indicate perfect fit
Fuel Consumption
Model Fit to Training Data
Time Series R2 = 0.931
11. 11
Fuel Consumption Predictions –
For Each Time Series Data Point
Results shown are blind predictions (not the model fits to the training data)
On-Road fuel consumption could be predicted using dyno data alone
Predictions improved by including some randomly selected on-road microtrips
Machine Learning
with 30% On-Road Data
Time Series R2 = 0.834
Machine Learning
with Dyno Data Only
Time Series R2 = 0.785
Under Predict
Over Predict
12. 12
Fuel Consumption Predictions –
Micro Trips
Machine Learning
with 30% On-Road Data
Micro Trip R2 = 0.972
Time Series R2 = 0.834
Machine Learning
with Dyno Data Only
OBD Data
Ideal Gas Law
Micro Trip R2 = 0.945
Time Series R2 = 0.774
Micro Trip R2 = 0.962
Time Series R2 = 0.785
The estimated fuel consumption from the OBD dongle for this vehicle
correlated well with measurements, but underestimated at mid to high loads
13. 13
Exhaust Emissions Prediction for Micro Trips
Carbon Monoxide
Carbon monoxide had fewer non-detects compared to other species
Feature importance: equiv. ratio, long term fuel trim, MAP, engine speed
The predicted average CO emissions rate distribution by VSP mode had a
similar trend compared to measured values
Micro Trip R2 = 0.986
Time Series R2 = 0.841
Largest discrepancy at
Mode 14 attributed to very little
data at this highest power condition
14. 14
Exhaust Emissions Prediction for Micro Trips
NOx
21% of the NOx emissions data collected for the micro trips were below the
detection limit (treated as zero for training)
The model did not accurately predict NOx emissions, suggesting there was
not sufficient explanatory data. Considerations: additional CAN parameters,
physics, catalyst sensitivity to fuel sulfur, model training, algorithms.
Micro Trip R2 = 0.201
Time Series R2 = 0.061
Model discrepancy at
higher loads
15. 15
Exhaust Emissions Prediction for Micro Trips
Total Hydrocarbons
53% of the hydrocarbon emissions data collected for the micro trips were
below the detection limit (treated as zero for training)
The predicted average THC emissions rate distribution by VSP mode had
a similar trend compared to measured values
Micro Trip R2 = 0.941
Time Series R2 = 0.561
16. 16
Summary and Conclusions
Machine Learning (pattern recognition) was employed to predict fuel
consumption and emissions using only OBD parameters
Based on testing one vehicle and approximately 7 hours of data:
– On-road vehicle fuel consumption was predicted using only the dyno
laboratory data source for model training
– Good predictions of CO and THC emissions were possible but
required the use of some on-road data for model training
– NOx emissions were not accurately predicted, suggesting a lack of
explanatory information
– The predictions showed potential to faithfully represent the real-world
emissions rate distribution by Vehicle Specific Power mode
More on-road training data would further improve prediction accuracy