You are on page 1of 85

Fault Detection and

Diagnosis
SENAKA SAMARASEKERA

Scope
Fault detection : Indication that something has gone wrong in the system

Fault Diagnosis
Fault isolation : determination of the exact location of the fault (FDI)
Fault identification : Determination of the magnitude of the fault (FDD)

Fault Classes
Additive ( Unknown inputs) vs Multiplicative (Changes in the parameters)
Sensor faults, Plant faults, and Actuator faults

Faults, disturbances, and model errors


Faults: changes in the system we want to detect and act accordingly
Disturbances : we want to ignore (e.g. : sensor, plant and actuator noise)
Model errors : discrepancies between the mathematical model used in signal processing and the
actual system. This is a nuisance effect we want to ignore

Type of faults

Detection and Isolation


performance
Detection performance can be analyzed as
Fault sensitivity : ability to detect a fault of reasonably small size
Reaction speed : ability to detect a fault of reasonably small delay after their occurrence
Robustness : ability to detect fault in midst of noise, disturbances and modelling errors

Isolation performance is much more complex to analyze. E.g.: Localization,


Both detection and isolation will depend on

Properties of the plant,


Size of the fault,
Fault Multiplicity (How many faults occur simultaneously)
Noise and other disturbance levels, and model errors.

Some time one fault might not be distinguishable from another fault. These occur due to
limitations of observability of the system. Then for the given set of sensors, system model and
plant the best we can do is to provide a fault class (i.e. Classify the faults in to different classes.
In practice this is the most common method.)

Approaches to FDD
Model
free methods

Can be used in simple FDD scenarios such as detecting a directly observable fault. Cannot be used for
multiple FDD hidden state FDD.
E.g. :
Multiple sensors ( to identify sensor failure
Limit checking : If then fault
Limit checking can be extended to two techniques
Trend checking :
e.g. Time domain : if FOR then detect fault.
Frequency domain: if FOR then detect fault.
Time frequency plane: if FOR then detect fault.
Knowledge based checking : Enable to use coupling between symptoms and faults. This contains the expert
knowledge gained by experience.
e.g. IF (condition 1 and condition 2) then detect fault 1
if condition 3 then detect (fault 1 and fault 2)

Approaches to FDD
Model- based methods
Use an explicit mathematical model of the monitored plant.
The original plant model can be given as a continuous time differential
equation or a Laplace domain transfer function
Since the monitoring is mostly done via computers we need to convert this
to a discrete time representation (i.e. difference equation or z domain
transfer function)
Most of the time we are able to linearize the system model around a suitable
operating point.

Residuals
The main concept behind model-based methods is
Analytical redundancy : Is the comparison of the actual
behavior of the monitored plant to the behavior predicted by the
fault free plant model . The plant measurement ae checked for
consistency with the model.
Residual : is the difference between the actual and the
computed estimate and are the outcome of he consistency
checks. This is the main indicator function used in model based
FDD.

Residuals
Methods for generating residuals
Kalman filter
Innovation (prediction error) of the Kalman filter can be used as a fault detection residual
Since innovation is white relatively easy to construct statistical tests
Fault isolation : One needs to run a bank of matched filters (one for each suspected fault) and find
the maximum SNR

Diagnostic observers: When the state is not directly observable we can design a state
observer. Isolation is just like Kalmann
Parity relations: Rearrange the system (equation) in to subsystems (equations) via linear
transformation so that a decision is taken (parity bit is fired) when a fault occur. Looking at
the parity bit structure an error could be isolated
Parameter estimation: Fault free parameter values are specified for a parametric model of
the system. If the parameter values deviate too far from the value a fault will be declared.

Residuals
Parameter, state and measurand estimates are never perfect match for
their actual counterparts. Therefore the residuals are never zero due to the
presence of noise, disturbances and modeling errors. But they will be
definitely large when a system fault occurs. (Why?)
Therefore we need to analyze the residuals and give them a threshold over
which we would claim a fault occurrence.
If we make this threshold too low : we would get many false alarms
If we make this threshold too high : we would miss a fault detection
How to decide what is the best threshold level?
We have to have an understanding on the noise statistic to answer this
question

Car engine exhaust emission control system


Most modern cars have emission control system that will keep the CO,
Hydrocarbons and the NOx below regulatory thresholds. The regulations
dictate that the OBD (Onboard diagnostics) should detect and isolate a fault
in this system if the emissions increase to 50% of this value.

Stoichiometric ratio that minimizes


emissions = 14.7

Excess oxygen sensor


characteristics

Car engine exhaust emission control system


EGR valve dilutes the input mix
with exhaust gas by lowering
the combustion temperature
leading to recued NO emissions.
Scope of the diagnostic task

Fault detection and isolation of the emission


control system using Parity relations

The models are build using system


identification methods. (grey box)

Thresholds are determined empirically


Thresholds are low pass filtered to reduce noise
As a second stage of filtering its only after the
fire counter reaches a certain threshold that the
fault is detected

Residuals and counters with no faults

Residuals and counters with a small EGR


fault

Analytical redundancy based


Fault detection
Contain two sub blocks
1. Residual generator
2. Decision maker
A single residual might be sufficient for fault detection. But for fault isolation
the residuals should be enhanced, that is multiple residuals which are
orthogonal to different faults should be designed.

Modelling additive faults


Let

Let be actual plant inputs and be the actual


plant output

Modelling additive faults


Let
the fault free model be
where and
Assumptions: System is time invariant. System is can be modelled using the shift
parameter
E can incorporate the faults as
Where is the plant fault transfer function matrix.
We can rewrite this combining all fault response in
Where and

Modelling additive disturbances


Let

Let be actual plant inputs and be the actual


plant output

Modelling additive faults and


disturbances
Using similar procedure as in faults we can incorporate the disturbances as
Where

Taking faults and disturbances together we can write

Example

Let
1. Find the fault free model.
2. Assume that is controlled and is a measurement. Assuming no plant
disturbances or noise, find the model with input and output faults and
noises.

Modelling multiplicative faults


and disturbances
These
can be modelled as changes in the system transfer function

These may reflect a


1. Parametric fault
2. Modelling error
To simplify analysis we will assume that these model changes are not time
varying (for the time period of analysis).
In most cases rather than defining individual deviations for the scalar terms in the
system matrix we can parametrize model using an underlying parameter vector .

Parameterizing model deviations


be the model transfer function.
If the nominal (fault free) values of
we can approximate the deviations using the Jacobian matric
Then we can write

Parameterizing model deviations


By grouping the model deviations to parametric faults and model errors,

This enable us to write the total equation as

Example (cont) :
Assume that the two poles and are the source of model uncertainty. Find and
and there by give an expression to the model uncertainty.

Parameter estimation for linear


models

Equation error method

Output error method

Residual generation
The
residual generator is a linear discrete dynamic algorithm acting on the observables.
where
In other words
which leads to
Since
We can write

Parameter estimation

Example : Air cabin pressure


valve control using BLDC

Example : Air cabin pressure


valve control using BLDC
Modelling of the BLDC
Electrical subsystem
In order to get out we recognize that only two out
of 3 windings are active at a given time which leads
to
e.g.
Similarly 5 more equations can be written.
Let

Example : Air cabin pressure


valve control using BLDC
Mechanical subsystem
Where

where gives the flap position.


Where is the motor position and is
the gear ratio.

Example : Air cabin pressure


valve control using BLDC
LS Parameter estimation
Electrical subsystem
Let ,
Let be the sliding window data matrix.

Example : Air cabin pressure


valve control using BLDC
LS parameter estimation
Mechanical subsystem
Neglecting dynamic friction and, assuming is known,
Let ,
Let be the sliding window data vector.

Here can be estimated using a state variable filter or simply an active


differentiator

Example : Air cabin pressure


valve control using BLDC
Generating residuals

Each of the residuals are decoupled from at least one variable.


,

FDI indicators for the cabin


pressure valve system

Detection of faults using


Residuals
How to be robust against
Additive disturbances: Slow varying in time, compared to the processing
window. No quantitative information is available most of the time. Therefore
the residuals should be orthogonal to these disturbances.
Multiplicative disturbances: Slow varying in time, the residual generator
should adapt online for parameter changes
Noise : Fast varying in time. Statistical description of these are available.
Filtering and appropriate thresholding (to balance false alarms and missed
detections)

Fault sensitivity
Fault
sensitivity

Triggering limit : The value of the particular fault j which brings a particular
residual i to
its detection threshold , provided no other faults and nuisance inputs are present.
This can be analyzed using
Assuming step fault (applying final value theorem)
Therefore

Fault sensitivity

For
a nominal (typically occurring, minimum) fault size for successful
detection
When analyzing the detection quality of a we would look at the spread

Modelling error robustness


Since the residual response for a model error only scenario is
we can define the limit error as
A higher limit error signifies lower sensitivity to model errors hence, better
robustness.

Finding thresholds for residuals


Considerations
1. Magnitude (distribution) of the Fault
2. Magnitude (distribution) of noise and other disturbances
If these information is available we can formulate the problem as a problem in
statistical dicision theory.
Noise zero mean with known covariance
Fault with a deterministic value (mean)
Fault detection -> Testing for zero mean hypothesis
Fault Isolation -> Testing for multiple hypothesis

The Neyman- Pearson detector

Isolation properties of the


residuals

Directional
residuals : We can define a vector for each fault j such that the
residual vectors are of the form
Where the is the scalar transfer function and is the fault magnitude.

Isolation properties of the


residuals
Structured residuals Each residual responds o a different subset of faults.
This enables the definition of a fault signature or a fault code, by thresholding
each fault.

Residuals from output error


method

Residuals from equation error


method

and are known as primary residuals. The two are


related as

SISO residuals
In the SISO case nothing much can be done with the single residual.
E.g. : Let Find and under different fault conditions. Comment on the
diagnosis.

SISO residuals :First order


system

SISO residuals : First order


system

Continuous time state space


parity equations
The derivatives of y can be found to
be

Continuous time state space parity


equations

Continuous time state space


parity equations

Continuous time state space


parity equations
Using the Laplace transform

Discrete time state space parity


equations
Similar to the continuous time case we
create time difference equations

Discrete time state space parity


equations

Discrete time state space parity


equations

Discrete time state space parity


equations

Properties of residuals

Generation of Enhanced
residuals
Structured residuals vs directional residuals

Errors in residuals : Strong


isolation /Weak isolation

Generation of structured
residuals
Generating good isolating patters of the residual vector
The equation error residual for p inputs and r outputs MIMO system can be
given as

To generate structured residuals this is multiplied by matrix W

Generation of structured
residuals

Let

Generation of structured
residuals
Example : Let the system be SIMO with 2outputs

Rearranging the process equations and vectorising them

Generation of structured
residuals

Generation of structured
residuals

Fault detection with state


observers
Let the system be

With the assumption that the model parameters are known the state observer estimates the
state using the measurands as

The underline assumption is that the system is observable.

State observers

Asymptotic state error

iff is stable.
So the trick is to build observers that are stable.

Additive fault detection with


state observers

Additive fault detection with


state observers

Additive fault detection with


state observers
Using the Laplace transform of the nofault system

Using the Laplace transform of the fault system (no


disturbance)
Using Final value theorem, assuming step fault and

Multiplicative fault detection


with state observers

Fault isolation with state


observers

Fault isolation with state


observers

Fault isolation with state


observers
Similar to the structured residuals. Use different observers that are
insensitive to deviations of a sub set of measurands. Let

An observer bank is obtained using one output at a time


Leading to an output error residual

Fault isolation with state


observers

Kalman Filter based FDI

Kalman Filter based FDI

Kalman Filter based FDI

FDI based on Output observers


Do not care about observing states. Just want to observe the faults.

FDI based on Output observers

FDI based on Output observers

FDI based on Output observers

FDI based on Output observers

FDI based on Output observers

You might also like