Experimental Results of HDV Longitudinal Control Using Compression Braking

1
Experimental Results of HDV Longitudinal Control
Using Compression Braking and Parameter
Estimation

Phil Farias Tsu-Chin Tsao University of
California Los Angeles
Ardalan Vahidi Anna Stefanopoulou The University
of Michigan Ann Arbor
October 24, 2002
2
Overview

• Project Goals
• Design and Implementation of a controller for
  coordinated compression and service braking
• Robustification of the controller by online
  model-based parameter estimation

Control
Mass and Grade Estimation
Model Development

• Evaluation of existing methods
• Our approach to the problem
• Demonstration using experimental data

• Obtaining the engine torque map
• Construction of compression braking map
• Construction of gear shift map
• Design of the controller

• Identification of vehicle parameters
• Obtaining road grade from the available plans
• Experimental validation

3
The Process for Model-based Control and Parameter
Estimation
Mathematical Model
Model Validation
Online Estimation
4
Obtaining and Understanding Signals

• In-Vehicle Computer Interface
• Thanks to the entire PATH Group
• Dan Empey, Sue Dickey, Xiao-Yun Lu, Ching-Yao
  Chan, Hong Bae, and all those who made the
  experiments possible and supported us before and
  after

SAE J1939 Sensor Specifications
5
Road Profile Digitization

• Necessary for identification of longitudinal
  dynamics
• Essential for validation of grade estimates
• Potentially useful for direct use in control

Thanks to Asfand Siddiqui from Caltrans for
sending the maps
6
Model Validation

• Recorded Signals
• Engine Torque
• Compression braking torque
• Transmission retarder torque
• Gear number
• Velocity
• Engine speed
• Also known
• Mass
• Road Grade
• Unknown Parameters
• Drag coefficient (0.7)
• Coefficient of rolling resistance (0.006)
• Engine inertia (2.82)

7
Compression Brake Model

• Engine RPM vs Torque Generated
• No high or low data signal available
• Find runs where only high or low mode is used
• July 27, 2002 Run 14 only high
• July 27, 2002 Run 6 only low
• Compare runs 6 14 to a run with both modes
  (19)
• Curve fit high data with 2nd order polynomial
• Curve fit low data with linear approximation

8
Compression Brake Model

• Compression Brake High
• 0.0003x2 - 0.0347x 162.84
• Compression Brake Low
• 0.2352x 1.8568
• Compare model with actual data acquired in run
  s 4 15
• Low model is very accurate
• High model is tuned to make sure that output does
  not exceed physical limitations

9
Transmission Map

• Model transmission shifting schedule as a
  function of output shaft speed and accelerator
  pedal position
• Compare model predictions with actual data run
• Results
• Characterized shifting schedule

10
System Block Diagram
11
Controller Scheme

• From TO 4202 Stefanopoulou, et. al

Tfuel
U gt 0
Fueling Mode
Vref
u

Vehicle Model
PI
Tbraking
-
Braking Mode
U lt 0
Vactual

• Controller output is Torque required

• Lets take a closer look at braking mode

12
Splitting Torques Technique

• Compare total braking force required (Tbraking)
  with braking force available from Compression
  brake (Thigh, Tlow)

Thigh
Engine Speed RPM
Compression Brake Model
Tlow

• If Tbraking lt Tlow,
• Tbraking Tservice

Tcomp

• Else, use algorithm to determine high or low
  mode.

Braking Mode
Tservice

• Tservice Tbraking T(high, low)

13
Previous Model Results
14
Velocity tracking through a grade

• Simulation
• Tracking a reference velocity through the
  digitized grade profile from I-15 HOV
• Run simulation with and without transmission map
• Results
• Including the transmission map leads to improved
  response in uphill portion of grade
• Physical limitations restrict a faster or more
  aggressive controller

15
Fueling and Braking Through the Grade

• Controller decreases fueling torque
• Compression brake decreases use of service brake
  by 72 for this simulation

16
Importance and Difficulties in Mass and Grade
Estimation

• Why estimation of mass and grade is important?
• Important in cruise control and vehicle following
• Critical for proper gear changing
• Control of anti-lock brake system
• Difficulties
• Mass can vary as much as 400
• Grade is time-varying
• Coupled in vehicle dynamics equation

17
State of the Art Research Papers

• Finding first the grade then the mass (Sensor
  based)
• Bae et al. Find grade using GPS antennas, then
  use least square on engine torque and vehicle
  speed to estimate mass
• Ohnishi et al. Find grade using accelerometer
  data, then use engine torque and vehicle speed to
  estimate mass
• Simultaneous mass-grade estimation (Model based)
• Druzhinina et al. Show convergence to piecewise
  constant mass and grade, under persistent
  excitations, within an adaptive control scheme

18
State of the Art Patents, Industry Interviews

• US Patents
• Zhu et al, Cummins, 1998 Recursive mass
  estimation based on selected qualified data
  portions of engine torque, speed, etc.
• Genise, Eaton Corp, 1994 Estimate mass based on
  the velocity drop during gear shift
• Hayakawa, et al., Aichi-Ken Japan, 2000 Propose
  using high pass filters on torque and
  acceleration signals can remove the influence of
  grade, then estimate mass using RLS
• Ohnishi et al, Hitachi, 1992 used an artificial
  neural network to estimate mass based on
  torque-acceleration relation
• Industry interview
• Contacted Freightliner One approach they used
  was to estimate mass based on velocity drop
  during gear shift and grade based on elevation
  signals from GPS. (communication with Thomas
  Connolly)

19
The Estimation Problem Formulation
y
f1
f2
tan(a)coefficient of rolling resistance
grade
20
The Solution

• Classical Least Square
• Least Square with Single Forgetting
• Least Square with Multiple forgetting

0lt?lt1 is the forgetting factor
0lt?1lt1 forgetting for first parameter 0lt?2lt1
forgetting for second parameter
21
Scenario I Constant Mass, Step Changes in Grade
Single Forgetting ?0.999
Multiple Forgetting ?11.0 ?20.8
22
Scenario II Constant Mass, Sinusoidal Grade
Single Forgetting ?0.999
Multiple Forgetting ?11.0 ?20.5
23
Comparison by Simulation
24
The Challenge with Experimental Data

• Lack of persistent excitation during normal
  cruise
• Model mismatch during gear change
• Signal noise

25
Experimental Results
Day1 Run 5 Northbound ?grade0.4
?mass0.95 Batch4 seconds Error in Grade RMS
0.21 degrees Error in Mass Max2.77 RMS340
kg
26
Summary

• A vehicle dynamics model was tuned and validated
  in experiments.
• Compression braking and transmission shift maps
  were characterized.
• A controller was designed and tested successfully
  in simulations for coordinated compression and
  service braking.
• An effective algorithm was proposed for online
  estimation of multiple time-varying parameters.
• The algorithm was successful in estimation of
  mass and time varying grade using simulated and
  experimental data.

27
Next Steps

• Implementation and Testing of Integrated
  Compression Brake/Service Brake Under Different
  Speed Control Strategies
• Incorporating the estimator in the control
  methodology
• Evaluation of braking under different automated
  scenarios