structural health monitoring using pattern recognitionmide.aalto.fi/en/attach/ismo/lecture_1.pdf ·...
TRANSCRIPT
New Trends in Vibration-Based SHM, CISM Udine, September 2008 1
Structural Health Monitoring using Pattern Recognition
Keith Worden and Graeme [email protected] Research Group
Department of Mechanical EngineeringUniversity of Sheffield
New Trends in Vibration-Based SHM, CISM Udine, September 2008 2
Lecture 1: Damage Identification and the Data to
Decision Process
3
Are These Systems Damaged?
Did you use pattern recognition?
4
Dog or Cat
5
Normal or Damaged
Uncracked Cracked
Wigner-Ville Transform
6
Background TextsRelevant literature – Structural Health Monitoring:
D.E.Adams, “Health Monitoring of Structural Materials and Components”, Wiley, 2007. W.J.Staszewski, C.Boller & G.R.Tomlinson, “Health Monitoring of Aerospace Strucutres”, Wiley, 2003.C.R.Farrar, H. Sohn & K. Worden, “Structural Health Monitoring using Statistical Pattern Recognition”, Wiley, 2009.
7
Lecture Overview
Damage Taxonomy
Damage Identification Hierarchy
Online vs. Offline Techniques
A Unified Framework for Intelligent Damage Identification
Data-driven vs. Model-driven Approaches
The Data-to-Decision Process
Data Fusion
8
Damage TaxonomyIn order to provide a unified approach to damage evaluation it is necessary to have an unambiguous definition of damage and associated terms.
Firstly, all materials contain defects at a nano/microstructural level.
Difficulty is in deciding when a structure is damaged.
9
Damage TaxonomyCracks and Fracture (Broberg, 1999) gives an excellent account of the inception and growth of microcracks and voids in the process region of a material.
Even more complicated for composite materials.
Loading and environmental conditions play major part in crack growth.
Corrosion and wear also lead to mechanical failure.
10
Damage Identification HierarchyRytter (1993) proposed the hierarchical structure for damage identification in his thesis:
Level One:Damage Detection
Level Two:Damage Location
Level Three:Damage Assessment
Level Four:Prediction
The method gives a qualitative indication thatdamage might be present in the structure.
The method gives information about theprobable position of the damage.
The method gives an estimate of the extentand/or type of damage.
The method offers information about thesafety of the structure (e.g. residual life).
11
Damage Identification HierarchyEach level requires that all lower-level information is available.
If possible, damage identification should be implemented on-line.
Hierarchy applies to Structural Health Monitoring (later will use as framework for Condition Monitoring and Statistical Process Control).
One addition to hierarchy will be required when multiple damage mechanisms are possible.
12
Damage Identification Hierarchy
Classification is important, if not vital, for identification at level 5 and possibly at level 4.
Level One:Damage Detection
Level Two:Damage Location
Level Three:Damage Classification
Level Four:Damage Assessment
The method gives a qualitative indication thatdamage might be present in the structure.
The method gives information about theprobable position of the damage.
The method gives information about thetype of damage.
The method gives an estimate of the extentof damage.
Level Five:Damage Prognosis
The method offers information about thesafety of the structure (e.g. residual life).
13
Damage Identification HierarchyLevel 5 is distinguished from others in that it cannot be done without understanding the physics of damage.
Level 1 is also distinguished from others – can be accomplished without knowledge of system behaviour when damaged.
Often a trade-off between the level of a diagnostic system and expense of training it adequately.
14
Online vs. Offline Techniques
SHM can be replaced by Condition Monitoring for machine applications.
This is a generalisation – not all NDT methods are offline and not all SHM techniques are online.
Non-destructiveEvaluation (NDE)
Non-destructive Testing(NDT)
Offline implementation of NDE
Generally implemented locally
Structural HealthMonitoring (SHM)
Online implementation of NDE
Generally implemented globally
15
Unified Framework for Intelligent Damage Identification
One of the aims of the course is to present a unified framework for damage identification, applicable over broad range of application areas.
The five level Rytter’s hierarchy can be modified for Conditional Monitoring of machines and Statistical Process Control – not pursued here.
16
Data-driven vs. Model-driven Approaches
There are two main approaches to the problem of damage identification.
Model-driven approaches.
Data-driven approaches.
17
Data-driven vs. Model-driven Approaches
Model-driven approaches treat damage identification as an inverse problem.
First a high-fidelity model of the undamaged structure or machine is constructed using physical laws based on first principles.
Changes in the measured data are then related to modifications in the physical parameters via system identification algorithms based on linear algebra or optimisation theory.
18
Data-driven vs. Model-driven Approaches
Data-driven approaches treat damage identification as a pattern recognition problem.
Measured data from the system of interest are assigned a damage class by a pattern recognition algorithm.
Can be unsupervised or supervised learning. In supervised learning, examples of all damage classes are required.
We mainly use data-driven techniques.
19
Data-driven vs. Model-driven Approaches
Model-Driven Approaches
• High-fidelity physical model of structure is established.
Can potentially work without validated damage model.
Noise and environmental effects are difficult to incorporate.
Data-Driven Approaches
• Statistical model of systemestablished.
Data required from all classes of damage.
Noise levels and environmental variation are established naturally.
20
Data to Decision ProcessThere is no sensor which can measure damage. Damage can however be inferred through the processing of acquired data.
Several stages of processing may be required before a decision can be formulated.
This decision may be a control decision (SPC) or a health monitoring decision (CM or SHM).
21
Single Sensor Processing ChainSensor
Pre-Processing
Feature Extraction
Post-Processing
Pattern Recognition
Decision
22
Sensor
Pre-Processing
Feature Extraction
Post-Processing
Pattern Recognition
Decision
Provides an quantitative electrical signal proportional to structural or environmental variable of interest.
Sample rate will depend upon data being measured.
Issues such as sensor placement and sensor validation will be discussed later.
Single Sensor Processing Chain
23
Sensor
Pre-Processing
Feature Extraction
Post-Processing
Pattern Recognition
Decision
Can consist of two tasks to prepare data for feature extraction:
Cleansing of raw data.
Dimension reduction.
Data cleansing includes noise removal, spike removal, outlier removal and treatment of missing data values.
Dimension reduction tries to eliminate redundancy in data.
Carried out on basis of experience and Engineering judgement.
Single Sensor Processing Chain
24
Sensor
Pre-Processing
Feature Extraction
Post-Processing
Pattern Recognition
Decision
Feature is short for distinguishing feature from pattern recognition literature.
Aim is to magnify the characteristics of the various damage classes and suppress normal background behaviour.
Can be based on statistics or Engineering judgement.
Dimension of features should be in the tens.
Single Sensor Processing Chain
25
Sensor
Pre-Processing
Feature Extraction
Post-Processing
Pattern Recognition
Decision
Post-processing is final preparation for pattern recognition.
Is often subsumed into feature extraction.
May be normalisation of feature vectors as required by the pattern recognition algorithm.
May be more advanced such as nonlinearly transforming the data to produce some probability distribution.
Single Sensor Processing Chain
26
Sensor
Pre-Processing
Feature Extraction
Post-Processing
Pattern Recognition
Decision
Arguably, the most critical stage of the process.
Feature vectors passed to an algorithm which can classify the data.
Three types of algorithm can be distinguished depending on desired diagnosis:
Novelty detection
Classification
Regression
Single Sensor Processing Chain
27
Sensor
Pre-Processing
Feature Extraction
Post-Processing
Pattern Recognition
Decision
Decision and action based upon pattern diagnosis.
May be automated or may require human intervention.
Single Sensor Processing Chain
28
Data FusionIn many monitoring systems there will be a number of sensors recording different variables.
Numerous reasons why multi-sensor systems are desirable:
Higher signal-to-noise ratio.
Robustness and reliability.
Different sensors may be sensitive to different damage states.
The data will possibly require different pre-processing, feature extraction, post-processing and pattern recognition before a decision can be made.
29
Data FusionMany ways of implementing sensor/data fusion – largely due to various defence organisations attempting to formalise procedures for integrating information from disparate sources.
There are many fusion strategies. Common to all is replication of single sensor processing chain with chains being fused together at various stages.
30
Feature Extraction
Data FusionSensor n
Pre-Processing
Decision
Pattern Recognition
Post-Processing
Sensor 2
Pre-Processing
Sensor 1
Pre-Processing
CentralisedFusion
Fusion Centre
31
Data Fusion
Post-Processing
Feature Extraction
Sensor n
Pre-Processing
Decision
Post-Processing
Feature Extraction
Sensor 2
Pre-Processing
Post-Processing
Feature Extraction
Sensor 1
Pre-Processing
Pattern Recognition
Pattern-levelFusion
32
Data FusionActual architecture should be problem specific.
Information flows through graph and fusion is necessary when a module has more than one inputs – different techniques are adopted depending upon type of fusion (raw data fusion, feature-level fusion, pattern-level fusion, decision-level fusion).
Many different fusion models proposed over years but the current best for health monitoring purposes seems to be the Omnibus Model.
33
Data Fusion
DecisionMaking
ContextProcessing
SignalProcessing
Sensing
Control
ResourceTasking
PatternProcessing
FeatureExtraction
DECIDE
ACT
OBSERVE
ORIENTATE
Soft DecisionFusion
Hard DecisionFusion
SensorData Fusion
SensorManagement
OmnibusModel
Includes possibility of action (repair or control).
Action does not necessarily have to interrupt monitoring process, but may enhance it.