Online Failure Detection and Identification for IFCS through Statistical Learning
Yan Liu, Srikanth Gururajan, Bojan Cukic
NASA OSMA SAS '04 July 2004
6/8/2009 1
Outline
Introduction A Validation Framework Failure Detection
SVDD
A Fast
Learning Algorithm Online detection for different flight sections
Failure Identification Summary and Future Work
6/8/2009 2
NASA IFCS Architecture
Failure Detection and Identification
S en so rs
analog
B aselin e N eu ral N etw o rk
b a s e line d e riva tive s
O n lin e N eu ral N etw o rk
c o n tro l com m ands d e riva tive c o rre c tio n s p ilo t in p u ts
R eal-T im e P ID
e s tim a te d d e riva tive s d e riva tive e rro rs
C o n tro ller
New V&V Techniques
6/8/2009 3
A Schematic of an Aircraft
Ailerons Elevators Rudder
6/8/2009
Primary control surfaces on F-15 Aircraft
4
Actuator Failures
•Primary Control Surfaces
•Elevators, Ailerons, Rudders
•Two kinds of actuator failures
•Locked surface
•Control surface locked at current or predefined deflection
•Results in coupling of lateral and longitudinal dynamics
•Loss of surface
•Part of control surface is lost
•Results in loss of efficiency on the surface
6/8/2009 5
A Validation Framework
Desired Response
Reference Model
Error
Actual Response
Command
+
-
Physical Process
3.
Estimate Trustworthiness of Outputs
Feedback Control Adaptive NN
2.
Monitor Stability of Learning
Learning Rule 1. Failure Detection and
Failure Identification
6
6/8/2009
Flight Section Division for Failure Detection and Identification
4
Altitude (0 ft –70,000 ft )
8
7
12
11
16
15
3
2
1
6
5
10
9
14
13
Mach (0.2 – 1.6)
6/8/2009 7
Failure Detection Using Support Vector Data Description (SVDD)
6/8/2009
8
The TOOL - Support Vector Data Description
• Based on SVM, Developed by Tax et. al., • Finds a sphere with the minimal volume that contains all data points. • Basically a one class classifier.
6/8/2009
9
Highlights
Misclassification error and the function complexity bound generalization error. Maximizing “volume” minimizes complexity – typical quadratic programming. “Eliminates” over-fitting. Solution depends only on Support Vectors, not the number of attributes. Evaluation and implementation are fast and simple.
10
6/8/2009
Previously…
We demonstrate that
SVDD can be used as an effective tool for novelty detection.
SVDD can provide novelty measures for online monitoring.
6/8/2009
11
Limitations of SVDD
Time is of essence in real-time (flight) control! Optimization takes time! - O(n3).
space complexity is high due to matrix operation.
Running on 1.6Ghz, 256M RAM, How much time does it take?
Data size 100 200 400
6/8/2009
Time (sec) 2 240 1257
12
Two Lemmas - 1
A simple sampling lemma
Let S be a set of size n and a function that maps any subset of S, denoted by R to some value f(R). The violators of R is defined as V(R):={ s in S\R| (R U {s}) (R) }. The extreme elements in R is defined as X(R):={ s in R | ( R\{s}) (R)}.
For a random sample R of size r, the expected number of violators and extremes of R,denoted by vr and xr respectively, has the following relationship:
vr / (n-r) = xr+1 / (r+1).
6/8/2009
13
Support Vectors = Extreme Points
Data Description - (S)
Outliers = Violators
Based on the sampling lemma, it has been proven that for an LP-type problem, SVM in particular, a fast random sampling working set selection algorithm can achieve running time complexity O(m logn).
6/8/2009 14
Data of A Violators of A SVs of A
Data of B Violators of B SVs of B
A
B
AUB
6/8/2009
15
Two Lemmas - 2
A
lemma of combining
For two sets A and B, where both A and B are subset of S, let C = A U B. Let X(R) and V(R) denote the extremes and violators of a set R respectively. (a) X(C) is a subset of X(A U B), where
X(A U B) = (X(A)^X(B)) U (X(A)^V(B)) U (V(A)^X(B));
and (b) V(C) is a subset of V(A U B), where
V(AUB) = V(A)^V(B) .
6/8/2009 16
One Important Observation
Only SV’s are relevant for the final form of the classifier. ( by Vapnik) - This means if we were given only the SV’s, we would obtain EXACTLY the identical classifier as if we dispose all other data points.
6/8/2009
17
A Fast Algorithm – Decompose and Combine
1. Decompose:
Sequentially decompose the training set (flight data) into small working sets of fixed size. Apply SVDD for the subsets.
2. Combine: Combine the SVDDs of current and previously learned data subsets to obtain the global solution.
6/8/2009
18
A decomposing example
6/8/2009
19
Time Advantage (n=100)
O(n log n ) << O(n3)
6/8/2009
20
Fast SVDD - on nominal flight condition simulations
6/8/2009
22
Compare results with/without decomposition
Normalized data of parameters (alpha , Cz_alpha), nominal flight condition, 20hz, running for 40 secs, n=800.
SVDD for whole dataset
6/8/2009
SVDD using decomposition
23
Failure Detection Tests
Normalized data of parameters (alpha , Cz_alpha). Control failure flight condition, 20hz, running for 40 secs, 800 data frames collected. Failure occurs at 600th data frame. Examine data every second = 20 data frames. (Online Detection)
25
6/8/2009
6/8/2009
26
Failure Identification by Cross-Correlation Analysis
6/8/2009
27
Nominal vs. Off-nominal
• Under nominal flight conditions -
No significant interaction between longitudinal and lateral dynamics. • Under off-nominal flight conditions Failure results in loss of symmetry and thus significant couplings between longitudinal and lateral dynamics becomes highly probable.
6/8/2009
28
Failure Identification by Correlation Analysis
Research results suggest that failures can be identified by studying the correlation between certain longitudinal and lateral dynamics parameters. p (RollRate) vs. q (PitchRate)
Longitudinal Failure
DeltaE vs. p
p (RollRate) vs. q (PitchRate)
Lateral Failure
DeltaA vs. q
6/8/2009 30
Experimental Results –
Failure identification
6/8/2009
31
Summary
With the fast SVDD algorithm, possible failures can be detected efficiently and effectively.
The correlation analysis provides us with accurate results and thus can be implemented as an online failure identification tool.
33
6/8/2009
Future Work
Explore flight sections and build a SVDD database of vectors for online failure detection. Embed the SVDD tools and cross-correlation analysis into the IFCS simulation environment for future testing. Continue building tools. Inclusion into VV of NN guidebook.
6/8/2009
34