Artificial Intelligence (AI) is becoming necessary today automotive world. The talk (presented in Azuga on 1st July in the AI & ML meetup) includes the major areas where AI is applied and the challenges faced in applying AI.
3. SAE INTERNATIONAL
Dr. Vivek Venkobarao
Paper # (if applicable) 3
Education:
Ph.D in Electrical Engineering
Innovation and Entrepreneurship Certificate, Stanford University
Energy Innovation and Emerging Technologies, ‘Stanford University
Data science and big data analytics: Making data driven decisions MIT University
Publications:
Has 35 papers in international conferences and journals published
IEEE Senior Member- Execom member of CT and TEMS
Co-Author “Handbook of Research on Emerging Technologies for Electrical Power Planning,
Analysis, and Optimization“ from leading international publisher
10 Patents granted in US,Germany and India
4. SAE INTERNATIONAL
Non Linear system identification
Mathematical model of a system from
measurements of the inputs and outputs.
Models are developed - data gathering,
parameter identification, model
development and validation
Paper # (if applicable) 4
6. SAE INTERNATIONAL
What to check in a measurement
Paper # (if applicable) 6
Measurements
Majority under
sampling
Minority
Oversampling
7. SAE INTERNATIONAL
Where AI/ML in Embedded systems
Paper # (if applicable) 7
Learning Techniques for embedded system
Automated Calibration System Controllers
Advanced non linear
Digital twins
Action
=
c
State
=
Measurement
Reward
=
-
F_c(m)
Action
=
pv_av
State
=
(vs,vs’,setpoint)
Reward
=
-
(vs-setpoint)
Action
=
pedalVector
State
=
SeepLimitCurve
Reward
=
-
8. SAE INTERNATIONAL
Non linear optimizers – Fmincon (Automated Calibration)
Model predictive control - Speed advisor
Paper # (if applicable) 8
Goal :
The idea is to suggest energy optimal vehicle speed
trajectories with constraints on vehicle dynamics on the one
hand and the upcoming speed limits on the other.
Solution Space:
Data Generation: The synthetic data for training the MPC is
generated via non linear optimizer.
Optimiser:
VSn+1 = f (pedal value, envn cdn, VSn )
Subject to constraints
f (pedal value, envn cdn, VSn ) < Speed limits
Pedal value min < predicted pedal value < Pedal value max
Vehicle Model
(Plant)
Vs
Environment
Recommendations
9. SAE INTERNATIONAL
Non linear optimizers - Fmincon
Model predictive control - Speed advisor
Paper # (if applicable) 9
AI is not used directly
Non Linear optimizers -> fmincon
Ant colony Optimization
Particle swam optimization
Goal : To find the global minimum for a constrained
nonlinear multivariable function
Hybridization of algorithm
10. SAE INTERNATIONAL
How to model missing control Bio Inspired Computing
Paper # (if applicable) 10
Adaptive Hill Climbing
PSO
ANT colony optimisation
11. SAE INTERNATIONAL
Deep Reinforcement Learning - System, Controllers
Mathematical Model
Paper # (if applicable) 11
Goal :
To predict post injected fuel quantity for reaching the
temperature setpoint.
Solution Space:
Observations = f(current temperature, error, integral error)
Reward = MSE < Threshold → Positive Reward
MSE > Threshold → Negative Reward
MSE Grad > 0 → Negative Reward
MSE Grad < 0 → Positive Reward
Stop Creterion = Min T > current T
• Max T < current T
• MSE < Threshold
• Action > Threshold
12. SAE INTERNATIONAL
What we do
Paper # (if applicable) 12
RL usually used in gaming
GO and chess are best examples
Typical fuel systems are stochastic processes
RL used as
Very limited information about the world
13. SAE INTERNATIONAL
Example for AI based Intellengent BMS
Paper # (if applicable) 13
• Smart battery usage for traveling A to B
• When the charge is less then can go to nearest charging station
• Optimize the battery usage in the route by having better charging and
discharging profile
• Intelligent Battery Management System for various stops in the drive.
14. SAE INTERNATIONAL
Predictive Battery Management System
Paper # (if applicable) 14
Without Predictive SOC
• No way to check the SOC thresholds
• No way to control the total charging
• SOC at charging not a function of
distance to be travelled.
15. SAE INTERNATIONAL
Predictive Battery Management System
Paper # (if applicable) 15
SoC Predictor
Time
Current
Voltage
SOC
Distance to
Destination
Decision Engine
(Fuzzy/SVM)
Charging
Station
Driving
16. SAE INTERNATIONAL
Predictive Battery Management System - SOC Prediction
Paper # (if applicable) 16
Neural Network
Based
SoC Predictor
Time
Current
Voltage
SOC
17. SAE INTERNATIONAL
Predictive Battery Management System
Paper # (if applicable) 17
Prediction of
Range
(SVM/Fuzzy)
Distance
SOC
Classification
Classification via SVM
Classification via Fuzzy
Based on Classification and visual inspection the
rider can decide on charging station
18. SAE INTERNATIONAL
Predictive Battery Management System
Paper # (if applicable) 18
With Predictive SOC
• SOC thresholds are monitored always
• Total charging control is based on the
operating conditions
• SOC is a function of distance to be
travelled.
19. SAE INTERNATIONAL
Conclusion
Paper # (if applicable) 19
Accurate faster models for embedded system
Fast transient response can always be achieved by having
encapsulation of numerical methods and AI
AI can be effectively used to model missing physics during
transients
Usage of AI in all stages of development improves the
accuracy and performance.
20. SAE INTERNATIONAL
References
Paper # (if applicable) 20
1. Orchestrating Infrastructure for sustainable Smart Cities: http://www.iec.ch/
whitepaper/pdf /iecWP-smartcities-LR-en.pdf
2. Rohith Kamath, Vivek Venkobarao, “RT nonlinear models and model reduction
techniques for engine management systems - airpath dynamics”, FISITA
World Automotive Congress 2018, F2018/F2018-PTE-089
3. Rohit Kamath, Vivek Venkobarao, Prof Subramaniam, Simulation and Design
of Decentralized PI Observer Based Controller for Nonlinear Interconnected
Systems of the Diesel Engine Airpath DOI 10.1016/j.egypro.2017.05.103
4.Rohit Kamath, Vivek Venkobarao, Prof Subramaniam, “An analytical model of
diesel engine intake system for performance prediction”, CMC congress pune
May 10 2016
5.2008E19407 IN Vivek Venkobarao - Hybridizing Genetic Algorithms with
simulated annealing and Dynamic adaptive methods for global optimization –
Patent Application published