Fault Detection, Identification, and Handling for Automated Vehicles

1
Fault Detection, Identification, and Handling
for Automated Vehicles
Adam Howell Prof. Karl Hedrick Department of
Mechanical Engineering University of California,
Berkeley MOU 312, 373
Jingang Yi Prof. Roberto Horowitz
2
Overview

• Fault Detection and Identification
• Dedicated Observers
• Detection and Identification Algorithms
• SmartAHS simulation
• Fault Management
• FDIM architecture
• Simulation Results
• Conclusions Future Work

3
Fault Detection and Identification System
Architecture
Observers Detection Filters
Pattern Classification System
Communication, Sensor Measurements, Command Inputs
Change Detector
Residual Generator
Parity Equations
Fault Status
4
Motivation for Dedicated Observers

• Previous design (Rajamani,98) based on intuitive
  knowledge about system
• Detection and ID of all single faults
• Few classes of multiple faults identifiable
• Dedicated observers provide
• Detection of multiple sensor faults
• Simplified detection and ID
• Redundancy for controller reconfiguration

5
Dedicated Observer
a
Dedicated Observer 1
Automated Vehicle
u
ww
Dedicated Observer 2
u
Residual Generator
Dn
Dedicated Observer 3
u
d
Dedicated Observer 4
u
we
Dedicated Observer 5
u

• Residual vector element ri created by comparing
  estimate to all others, i.e.

6
Feedback Linearized Vehicle Model

• Multiple Sliding Surface controller for internal
  dynamics results in overall linear vehicle model
• Input usyn chosen to provide string stability
• Convert to state space model and design linear
  observers to estimate using input u and a
  single output (a, d, v, Dn, we)

Model
Observer
7
Relationship between Faults and Residuals

• Dedicated Observers create orthogonal fault modes
  for all sensors
• All single faults identifiable
• Actuator fault modes and sensor fault modes are
  linearly dependent

8
Change detection using Significance Test

• Assume residual vector r is a random vector with
  a Gaussian distribution characterized by m and C
• Mahalanobis Distance
• Two Hypothesis
• Normal
• Fault

P(r)
Normal
Fault
9
Fault Identification via Pattern Classification

• Each fault causes to grow in a unique
  direction of the residual space, termed fault
  mode
• Compare known fault mode,
  
  
  
  
  , to residual vector after change
  detected
• Minimum Distance Classifier

Residual 1
f1
r
q1
q2
Residual 2
f2
10
SmartAHS Simulation Environment

• SHIFT programming language developed at PATH
• Natural modeling of hybrid systems in state
  machine formalism
• Object oriented programming
• SmartAHS libraries contain
• Vehicle Models 3D, 2D, Simple, Double Integrator
• Controllers Automated, Human Driver Model
• Highway Lanes, Segments, Curves

11
PATHVehicle type

States
Data Acquistion
Vehicle Model
Command Inputs
Sampled Measurements
Control System
Communications
Communicated Measurements
12
ControlSystem type
Command Messages
Coordination Layer
Regulation Layer
Measurements
Maneuver
Desired Acceleration and Lateral Position
Fault Status
Actuator Commands
Actuator Commands
Lower Level Controller
FDI System
Measurements
13
Simulation Results for Wheel Speed Sensor Fault

• Wheel Speed Sensor faults of bias 3 and 6 m/s
  at 5 and 6 seconds, respectively

14
Simulation Results for Magnetometer Fault

• Magnetometer faults of bias 2 and 4 counts at 5
  and 6 seconds, respectively

15
Simulation Results for Magnetometer and Wheel
Speed Sensor Faults

• Wheel Speed Sensor faults at 5 and 6 seconds,
  followed by Magnetometer faults at 5.5 and 6.5
  seconds

16
Fault Tolerant AHS Control Structure
For fault tolerant AHS we need1

• Capability structure monitor hard faults
• Performance structure monitor soft faults
• Sensor structure collect information

and

• Control structure execute maneuvers

1 Lygeros et al., 1996
17
Extended AHS Architecture
Link Layer
Coordination Layer
Coordination Supervisor
Communications
Capa. Perf. Structures
Neighboring Vehicles
Fault Detection and Handling
Protocols
Controller Reconfiguration
Regulation Layer
Regulation Supervisor
Regulation Control Laws
Fault Detection
Physical Layer
Lower-Level Controllers
Sensors
Vehicle Dynamics
18
Hierarchical Fault Management Design
All faults can be handled using one of two
strategies

• Regulation level control reconfiguration
• for soft faults and sensor faults for which we
  can use redundant sensor/observer information
• Coordinated degraded mode maneuvers
• for severe faults that cannot be recovered by
  controller reconfigurations (e.g. actuator
  faults)

19
Design and Implementation in SmartAHS
Assumption Only ONE fault happens at a time

• Use normal mode and degraded mode regulation
  control laws based on Li et al., 1997 Swaroop,
  1994 Chen et al., 1997
• Extend coordination maneuver protocols Hsu et
  al., 1994 Eskafi, 1996 including a simplified
  communication model
• Extend coordination supervisor to handle
  different faults when faults are detected

20
Design and Implementation in SmartAHS (cont)

• Implement and test the hierarchical fault
  handling architecture for all longitudinal
  control faults in AHS
• SmartAHS Micro-simulator
• hybrid automata simulation language SHIFT
• flexible and easily expandable
• compatible with SmartPATH animator

21
Design and Implementation (cont) - example
capability structure design for normal mode
22
Simulation Results - scenario 1 radar fault
23
Simulation Results - radar fault (cont)
(a) Relative distance
(b) Relative velocity
24
Simulation Results (cont)
(c) Absolute velocity
(d) Absolute acceleration
25
Simulation Results - scenario 2 throttle
actuator fault
26
Simulation Results - throttle act. fault (cont)
(a) Relative distance
(b) Absolute velocity
27
Simulation Results - throttle act. fault (cont)
(c) Absolute acceleration
(d) Road lateral position
28
Conclusions

• Current fault detection and identification scheme
  can identify all single sensor and actuator
  faults, and up to 6 concurrent multiple sensor
  faults
• Current fault management scheme can handle all
  faults currently being detected on vehicles for
  longitudinal control system
• Different kinds of control strategies are needed
  for different faults
• Redundant information for fault detection can be
  used to reconfigure the controllers for most
  sensor failures
• Degraded mode maneuvers have been initiated for
  actuator faults

29
Future Work

• Extend current FDI system for multiple failures
  in both sensors and actuators
• Improve robustness of FDI system to speed
• Integrate the performance structure into the
  fault tolerant control architecture
• Integrate the lateral fault detection and
  handling scheme into the extended fault tolerant
  architecture
• Implement and test the fault handling scheme on
  the real automated vehicles