Testing and Evaluation of Robust Fault Detection and Identification

1
Testing and Evaluation of Robust Fault Detection
and Identification

• Robert Chen, Hok Ng, and Jason Speyer
• Mechanical and Aerospace Engineering Department
• University of California, Los Angeles

2
Fault Detection and Identification

• Objective
• Detect and identify actuator and sensor failures
  on Buick LeSabre.
• Approach
• Residual generation
• Robust fault detection filter
• Parity equation
• Residual processing
• Multiple-hypothesis Shiryayev sequential
  probability test (SPT)
• Testing and evaluation
• Crows Landing
• Magnetic curve track
• Real-time environment on Buick LeSabre

3
Fault-Tolerant System Current and Future Efforts
Actuator Faults
Sensor Faults
Actuators
Vehicle
Sensors
Residual Generator
Fault Probabilities
Residuals
Residual Processor
Control Commands
Fault- Tolerant Controller
Fault Magnitudes
Fault Reconstruction Process
4
Characterization of Our Approach

• Robust fault detection filter
• Uses control commands and measurements to
  generate residuals.
• Each residual is only sensitive to one fault, but
  insensitive to the other faults and disturbances.
• Nonlinear parity equation
• Detects faults that have strong nonlinearities.
• Multiple-hypothesis Shiryayev SPT
• Assigns a probability to each fault hypothesis
  based on the residuals generated by the fault
  detection filter and parity equation.
• Decreases the time that it takes to detect and
  identify a fault.
• The fault detection filter can be designed to
  generate residuals that respond faster, but are
  noisier.

5
Residual Generator Design

• Fault detection filters and parity equations are
  developed for the longitudinal and lateral modes
  separately.
• There are 3 actuators and 12 sensors on Buick
  LeSabre.
• Longitudinal mode
• Throttle actuator and brake actuator
• Manifold pressure sensor, engine speed sensor,
  longitudinal accelerometer, sum of front wheel
  speed sensors, sum of rear wheel speed sensors,
  throttle sensor and brake sensor
• Lateral mode
• Steering actuator
• Lateral accelerometer, yaw rate sensor,
  difference of front wheel speed sensors,
  difference of rear wheel speed sensors and
  steering sensor

6
Residual Generation for Longitudinal Mode
Fault Detection Filter 1 - Engine speed sensor -
Longitudinal accelerometer Fault Detection
Filter 2 - Front wheel speed sensors - Rear
wheel speed sensors Fault Detection Filter 3 -
Brake actuator - Rear wheel speed sensors Parity
Equation 1 - Throttle actuator - Manifold
pressure sensor - Engine speed sensor Parity
Equation 2 - Throttle actuator - Throttle
sensor Parity Equation 3 - Brake actuator -
Brake sensor
Residuals
Residuals
Control commands
Residuals
Measurements
Vehicle
Residual
Residual
Residual
7
Residual Generation for Lateral Mode (Work in
Progress)
Fault Detection Filter 1 - Steering actuator -
Lateral accelerometer - Front wheel speed
sensors - Rear wheel speed sensors Fault
Detection Filter 2 - Yaw rate sensor - Lateral
accelerometer - Front wheel speed sensors - Rear
wheel speed sensors Parity Equation - Steering
actuator - Steering sensor
Residuals
Control commands
Measurements
Vehicle
Residuals
Residual

• A new sensor fault model is used which allows
  more faults to be detected in each fault
  detection filter.

8
Actuator and Sensor Fault Condition

• Assumption
• One fault occurs at a time.
• Generation of actuator and sensor faults
• Faults were imposed by our computer to maintain
  the integrity of the instruments.
• Faults will be imposed by PATH computer so that
  the vehicle is under the effect of the faults.
• Performance evaluation of fault detection and
  identification
• The smallest faults that can be detected depend
  on
• The acceleration of the vehicle because it
  induces nonlinearities.
• Accuracy of the instruments (e.g., bias and
  noise).
• The smallest faults that need to be detected
  depend on the robust performance of the
  controller.
• The time that it takes to detect a fault depends
  on the size of the fault.
• Larger faults require less detection time.

9
Vehicle Operating Condition

• Assumption
• The vehicle is in third gear.
• Limitation
• Engine speed sensor fault cannot be detected when
  the throttle angle is smaller than 5 degrees.
• Experiments were conducted at Crows Landing when
  the Buick LeSabre was going straight and on the
  magnetic curve track.
• Constant speed
• 18, 20, 22, 24, 26 and 28 m/s (40.5 to 63 mph)
• Increasing speed
• Between 20 and 30 m/s (Largest acceleration is
  0.4 m/s2.)
• Decreasing speed
• Between 28 and 18 m/s

10
Experiment Setup
Faults
Faults
Faults Control commands
Control commands Measurements
Residuals
Measurements

• UCLA laptop
• Perform fault detection and identification tasks.
• Linux operating system

• PATH PC
• Perform control tasks.
• QNX operating system

• Faults will be imposed by PATH PC so that the
  vehicle is under the effect of the faults.

11
Real-Time Evaluation on Magnetic Curve Track
Fault initiation
66 mph
45 mph

• Each data point is 21 ms.

12
Evaluation of Longitudinal Fault Detection Filter
1
Each data point is 21 ms.
13
Evaluation of Longitudinal Fault Detection Filter
2
Each data point is 21 ms.
14
Evaluation of Longitudinal Fault Detection Filter
3
Each data point is 21 ms.
15
Evaluation of Part of Lateral Fault Detection
Filter 1 Using Experimental Data
Each data point is 21 ms.
16
Residual Processing Long. Accelerometer Fault
Fault size 0.4 m/s2 0.3 m/s2 0.2
m/s2 Noise level (RMS) 0.12 m/s2
17
Residual Processing Long. Accelerometer Fault
(Contd)
Fault size 0.4 m/s2 0.3 m/s2 0.2
m/s2 Noise level (RMS) 0.12 m/s2
18
Residual Processing Brake Actuator Fault
Fault size 100 psi 75 psi 50
psi
19
Residual Processing Brake Actuator Fault (Contd)
Fault size 100 psi 75 psi 50
psi
20
Conclusion

• Longitudinal fault detection filters were
  designed and evaluated on the curve track at
  Crows Landing in real-time on a Buick LeSabre.
• Fault detection filters work well for a wide
  range of car speed.
• Part of lateral fault detection filters were
  designed and evaluated using experimental data.
• Multiple-hypothesis Shiryayev SPT was designed
  and evaluated using experimental data.
• The fault can be announced on the basis of the
  probabilities of the faults.

21
Future Work

• Fault reconstruction process
• Generates the magnitudes of the sensor and
  actuator faults based on the residuals generated
  by the fault detection filter and parity
  equation.
• For a sensor fault, the correct measurement can
  be obtained with a small time delay.
• For an actuator fault, the condition of the
  actuator can be assessed (e.g., bias or stuck).
• Fault-tolerant controller
• Is designed such that it is robust even when
  using the reconstructed measurement.
• Uses the actuator information to determine a
  safety action.