STRUCTURAL HEALTH MONITORING
A MACHINE LEARNING PERSPECTIVE
Charles R. Farrar, Keith Worden
Structural Health Monitoring: A Machine Learning Perspective makes extensive use of the authors’ detailed surveys of the technical literature, the experience they have gained from teaching numerous courses on this subject, and the results of performing numerous analytical and experimental structural health monitoring studies.
- Considers structural health monitoring in a new manner by casting the problem in the context of a machine learning/statistical pattern recognition paradigm
- Emphasises an integrated approach to the development of structural health monitoring solutions by coupling the measurement hardware portion of the problem directly with the data interrogation algorithms
- Benefits from extensive use of the authors’ detailed surveys of 800 papers in the technical literature and the experience they have gained from teaching numerous short courses on this subject.
1.1 How Engineers and Scientists Study Damage
1.2 Motivation for Developing SHM Technology
1.3 Definition of Damage
1.4 A Statistical Pattern Recognition Paradigm for SHM
1.4.1 Operational Evaluation
1.4.2 Data Acquisition
1.4.3 Data Normalisation
1.4.4 Data Cleansing
1.4.5 Data Compression
1.4.6 Data Fusion
1.4.7 Feature Extraction
1.4.8 Statistical Modelling for Feature Discrimination
1.5 Local versus Global Damage Detection
1.6 Fundamental Axioms of Structural Health Monitoring
1.7 The Approach Taken in This Book
References
2 Historical Overview
2.1 Rotating Machinery Applications
2.1.1 Operational Evaluation for Rotating Machinery
2.1.2 Data Acquisition for Rotating Machinery
2.1.3 Feature Extraction for Rotating Machinery
2.1.4 Statistical Modelling for Damage Detection in Rotating Machinery
2.1.5 Concluding Comments about Condition Monitoring of Rotating Machinery
2.2 Offshore Oil Platforms
2.2.1 Operational Evaluation for Offshore Platforms
2.2.2 Data Acquisition for Offshore Platforms
2.2.3 Feature Extraction for Offshore Platforms
2.2.4 Statistical Modelling for Offshore Platforms
2.2.5 Lessons Learned from Offshore Oil Platform Structural Health Monitoring Studies
2.3 Aerospace Structures
2.3.1 Operational Evaluation for Aerospace Structures
2.3.2 Data Acquisition for Aerospace Structures
2.3.3 Feature Extraction and Statistical Modelling for Aerospace Structures
2.3.4 Statistical Models Used for Aerospace SHM Applications
2.3.5 Concluding Comments about Aerospace SHM Applications
2.4 Civil Engineering Infrastructure
2.4.1 Operational Evaluation for Bridge Structures
2.4.2 Data Acquisition for Bridge Structures
2.4.3 Features Based on Modal Properties
2.4.4 Statistical Classification of Features for Civil Engineering Infrastructure
2.4.5 Applications to Bridge Structures
2.5 Summary
References
3 Operational Evaluation
3.1 Economic and Life-Safety Justifications for Structural Health Monitoring
3.2 Defining the Damage to Be Detected
3.3 The Operational and Environmental Conditions
3.4 Data Acquisition Limitations
3.5 Operational Evaluation Example: Bridge Monitoring
3.6 Operational Evaluation Example: Wind Turbines
3.7 Concluding Comment on Operational Evaluation
References
4 Sensing and Data Acquisition
4.1 Introduction
4.2 Sensing and Data Acquisition Strategies for SHM
4.2.1 Strategy I
4.2.2 Strategy II
4.3 Conceptual Challenges for Sensing and Data Acquisition Systems
4.4 What Types of Data Should Be Acquired?
4.4.1 Dynamic Input and Response Quantities
4.4.2 Other Damage-Sensitive Physical Quantities
4.4.3 Environmental Quantities
4.4.4 Operational Quantities
4.5 Current SHM Sensing Systems
4.5.1 Wired Systems
4.5.2 Wireless Systems
4.6 Sensor Network Paradigms
4.6.1 Sensor Arrays Directly Connected to Central Processing Hardware
4.6.2 Decentralised Processing with Hopping Connection
4.6.3 Decentralised Processing with Hybrid Connection
4.7 Future Sensing Network Paradigms
4.8 Defining the Sensor System Properties
4.8.1 Required Sensitivity and Range
4.8.2 Required Bandwidth and Frequency Resolution
4.8.3 Sensor Number and Locations
4.8.4 Sensor Calibration, Stability and Reliability
4.9 Define the Data Sampling Parameters
4.10 Define the Data Acquisition System
4.11 Active versus Passive Sensing
4.12 Multiscale Sensing
4.13 Powering the Sensing System
4.14 Signal Conditioning
4.15 Sensor and Actuator Optimisation
4.16 Sensor Fusion
4.17 Summary of Sensing and Data Acquisition Issues for Structural Health Monitoring
References
5 Case Studies
5.1 The I-40 Bridge
5.1.1 Preliminary Testing and Data Acquisition
5.1.2 Undamaged Ambient Vibration Tests
5.1.3 Forced Vibration Tests
5.2 The Concrete Column
5.2.1 Quasi-Static Loading
5.2.2 Dynamic Excitation
5.2.3 Data Acquisition
5.3 The 8-DOF System
5.3.1 Physical Parameters
5.3.2 Data Acquisition
5.4 Simulated Building Structure
5.4.1 Experimental Procedure and Data Acquisition
5.4.2 Measured Data
5.5 The Alamosa Canyon Bridge
5.5.1 Experimental Procedures and Data Acquisition
5.5.2 Environmental Measurements
5.5.3 Vibration Tests Performed to Study Variability of Modal Properties
5.6 The Gnat Aircraft
5.6.1 Simulating Damage with a Modified Inspection Panel
5.6.2 Simulating Damage by Panel Removal
References
6 Introduction to Probability and Statistics
6.1 Introduction
6.2 Probability: Basic Definitions
6.3 Random Variables and Distributions
6.4 Expected Values
6.5 The Gaussian Distribution (and Others)
6.6 Multivariate Statistics
6.7 The Multivariate Gaussian Distribution
6.8 Conditional Probability and the Bayes Theorem
6.9 Confidence Limits and Cumulative Distribution Functions
6.10 Outlier Analysis
6.10.1 Outliers in Univariate Data
6.10.2 Outliers in Multivariate Data
6.10.3 Calculation of Critical Values of Discordancy or Thresholds
6.11 Density Estimation
6.12 Extreme Value Statistics
6.12.1 Introduction
6.12.2 Basic Theory
6.12.3 Determination of Limit Distributions
6.13 Dimension Reduction – Principal Component Analysis
6.13.1 Simple Projection
6.13.2 Principal Component Analysis (PCA)
6.14 Conclusions
References
7 Damage-Sensitive Features
7.1 Common Waveforms and Spectral Functions Used in the Feature Extraction Process
7.1.1 Waveform Comparisons
7.1.2 Autocorrelation and Cross-Correlation Functions
7.1.3 The Power Spectral and Cross-Spectral Density Functions
7.1.4 The Impulse Response Function and the Frequency Response Function
7.1.5 The Coherence Function
7.1.6 Some Remarks Regarding Waveforms and Spectra
7.2 Basic Signal Statistics
7.3 Transient Signals: Temporal Moments
7.4 Transient Signals: Decay Measures
7.5 Acoustic Emission Features
7.6 Features Used with Guided-Wave Approaches to SHM
7.6.1 Preprocessing
7.6.2 Baseline Comparisons
7.6.3 Damage Localisation
7.7 Features Used with Impedance Measurements
7.8 Basic Modal Properties
7.8.1 Resonance Frequencies
7.8.2 Inverse versus Forward Modelling Approaches to Feature Extraction
7.8.3 Resonance Frequencies: The Forward Approach
7.8.4 Resonance Frequencies: Sensitivity Issues
7.8.5 Mode Shapes
7.8.6 Load-Dependent Ritz Vectors
7.9 Features Derived from Basic Modal Properties
7.9.1 Mode Shape Curvature
7.9.2 Modal Strain Energy
7.9.3 Modal Flexibility
7.10 Model Updating Approaches
7.10.1 Objective Functions and Constraints
7.10.2 Direct Solution for the Modal Force Error
7.10.3 Optimal Matrix Update Methods
7.10.4 Sensitivity-Based Update Methods
7.10.5 Eigenstructure Assignment Method
7.10.6 Hybrid Matrix Update Methods
7.10.7 Concluding Comment on Model Updating Approaches
7.11 Time Series Models
7.12 Feature Selection
7.12.1 Sensitivity Analysis
7.12.2 Information Content
7.12.3 Assessment of Robustness
7.12.4 Optimisation Procedures
7.13 Metrics
7.14 Concluding Comments
References
8 Features Based on Deviations from Linear Response
8.1 Types of Damage that Can Produce a Nonlinear System Response
8.2 Motivation for Exploring Nonlinear System Identification Methods for SHM
8.2.1 Coherence Function
8.2.2 Linearity and Reciprocity Checks
8.2.3 Harmonic Distortion
8.2.4 Frequency Response Function Distortions
8.2.5 Probability Density Function
8.2.6 Correlation Tests
8.2.7 The Holder Exponent
8.2.8 Linear Time Series Prediction Errors
8.2.9 Nonlinear Time Series Models
8.2.10 Hilbert Transform
8.2.11 Nonlinear Acoustics Methods
8.3 Applications of Nonlinear Dynamical Systems Theory
8.3.1 Modelling a Cracked Beam as a Bilinear System
8.3.2 Chaotic Interrogation of a Damaged Beam
8.3.3 Local Attractor Variance
8.3.4 Detection of Damage Using the Local Attractor Variance
8.4 Nonlinear System Identification Approaches
8.4.1 Restoring Force Surface Model
8.5 Concluding Comments Regarding Feature Extraction Based on Nonlinear System Response
References
9 Machine Learning and Statistical Pattern Recognition
9.1 Introduction
9.2 Intelligent Damage Detection
9.3 Data Processing and Fusion for Damage Identification
9.4 Statistical Pattern Recognition: Hypothesis Testing
9.5 Statistical Pattern Recognition: General Frameworks
9.6 Discriminant Functions and Decision Boundaries
9.7 Decision Trees
9.8 Training – Maximum Likelihood
9.9 Nearest Neighbour Classification
9.10 Case Study: An Acoustic Emission Experiment
9.10.1 Analysis and Classification of the AE Data
9.11 Summary
References
10 Unsupervised Learning – Novelty Detection
10.1 Introduction
10.2 A Gaussian-Distributed Normal Condition – Outlier Analysis 322
10.3 A Non-Gaussian Normal Condition – A Neural Network Approach 325
10.4 Nonparametric Density Estimation – A Case Study
10.4.1 The Experimental Structure and Data Capture
10.4.2 Preprocessing of Data and Features
10.4.3 Novelty Detection
10.5 Statistical Process Control
10.5.1 Feature Extraction Based on Autoregressive Modelling
10.5.2 The X-Bar Control Chart: An Experimental Case Study
10.6 Other Control Charts and Multivariate SPC
10.6.1 The S Control Chart
10.6.2 The CUSUM Chart
10.6.3 The EWMA Chart
10.6.4 The Hotelling or Shewhart T2 Chart
10.6.5 The Multivariate CUSUM Chart
10.6.6 The Multivariate EWMA Chart
10.7 Thresholds for Novelty Detection
10.7.1 Extreme Value Statistics
10.7.2 Type I and Type II Errors: The ROC Curve
10.8 Summary
References
11 Supervised Learning – Classification and Regression
11.1 Introduction
11.2 Artificial Neural Networks
11.2.1 Biological Motivation
11.2.2 The Parallel Processing Paradigm
11.2.3 The Artificial Neuron
11.2.4 The Perceptron
11.2.5 The Multilayer Perceptron
11.3 A Neural Network Case Study: A Classification Problem
11.4 Other Neural Network Structures
11.4.1 Feedforward Networks
11.4.2 Recurrent Networks
11.4.3 Cellular Networks
11.5 Statistical Learning Theory and Kernel Methods
11.5.1 Structural Risk Minimisation
11.5.2 Support Vector Machines
11.5.3 Kernels
11.6 Case Study II: Support Vector Classification
11.7 Support Vector Regression
11.8 Case Study III: Support Vector Regression
11.9 Feature Selection for Classification Using Genetic Algorithms
11.9.1 Feature Selection Using Engineering Judgement
11.9.2 Genetic Feature Selection
11.9.3 Issues of Network Generalisation
11.9.4 Discussion and Conclusions
11.10 Discussion and Conclusions
References
12 Data Normalisation
12.1 Introduction
12.2 An Example Where Data Normalisation Was Neglected
12.3 Sources of Environmental and Operational Variability
12.4 Sensor System Design
12.5 Modelling Operational and Environmental Variability
12.6 Look-Up Tables
12.7 Machine Learning Approaches to Data Normalisation
12.7.1 Auto-Associative Neural Networks
12.7.2 Factor Analysis
12.7.3 Mahalanobis Squared-Distance (MSD)
12.7.4 Singular Value Decomposition
12.7.5 Application to the Simulated Building Structure Data
12.8 Intelligent Feature Selection: A Projection Method
12.9 Cointegration
12.9.1 Theory
12.9.2 Illustration
12.10 Summary
References
13 Fundamental Axioms of Structural Health Monitoring
13.1 Introduction
13.2 Axiom I. All Materials Have Inherent Flaws or Defects
13.3 Axiom II. Damage Assessment Requires a Comparison between Two System States
13.4 Axiom III. Identifying the Existence and Location of Damage Can Be Done in an Unsupervised Learning Mode, but Identifying the Type of Damage Present and the Damage Severity Can Generally Only Be Done in a Supervised Learning Mode
13.5 Axiom IVa. Sensors Cannot Measure Damage. Feature Extraction through Signal Processing and Statistical Classification Are Necessary to Convert Sensor Data into Damage Information
13.6 Axiom IVb. Without Intelligent Feature Extraction, the More Sensitive a Measurement is to Damage, the More Sensitive it is to Changing Operational and Environmental Conditions
13.7 Axiom V. The Length and Time Scales Associated with Damage Initiation and Evolution Dictate the Required Properties of the SHM Sensing System
13.8 Axiom VI. There is a Trade-off between the Sensitivity to Damage of an Algorithm and Its Noise Rejection Capability
13.9 Axiom VII. The Size of Damage that Can Be Detected from Changes in System Dynamics is Inversely Proportional to the Frequency Range of Excitation
13.10 Axiom VIII. Damage Increases the Complexity of a Structure
13.11 Summary
References
14 Damage Prognosis
14.1 Introduction
14.2 Motivation for Damage Prognosis
14.3 The Current State of Damage Prognosis
14.4 Defining the Damage Prognosis Problem
14.5 The Damage Prognosis Process
14.6 Emerging Technologies Impacting the Damage Prognosis Process
14.6.1 Damage Sensing Systems
14.6.2 Prediction Modelling for Future Loading Estimates
14.6.3 Model Verification and Validation
14.6.4 Reliability Analysis for Damage Prognosis Decision Making
14.7 A Prognosis Case Study: Crack Propagation in a Titanium Plate
14.7.1 The Computational Model
14.7.2 Monte Carlo Simulation
14.7.3 Issues
14.8 Damage Prognosis of UAV Structural Components
14.9 Concluding Comments on Damage Prognosis
14.10 Cradle-to-Grave System State Awareness
References
Appendix A Signal Processing for SHM
A.1 Deterministic and Random Signals
A.1.1 Basic Definitions
A.1.2 Transducers, Sensors and Calibration
A.1.3 Classification of Deterministic Signals
A.1.4 Classification of Random Signals
A.2 Fourier Analysis and Spectra
A.2.1 Fourier Series
A.2.2 The Square Wave Revisited
A.2.3 A First Look at Spectra
A.2.4 The Exponential Form of the Fourier Series
A.3 The Fourier Transform
A.3.1 Basic Transform Theory
A.3.2 An Interesting Function that is not a Function
A.3.3 The Fourier Transform of a Periodic Function
A.3.4 The Fourier Transform of a Pulse/Impulse
A.3.5 The Convolution Theorem
A.3.6 Parseval’s Theorem
A.3.7 The Effect of a Finite Time Window
A.3.8 The Effect of Differentiation and Integration
A.4 Frequency Response Functions and the Impulse Response
A.4.1 Basic Definitions
A.4.2 Harmonic Probing
A.5 The Discrete Fourier Transform
A.5.1 Basic Definitions
A.5.2 More About Sampling
A.5.3 The Fast Fourier Transform
A.5.4 The DFT of a Sinusoid
A.6 Practical Matters: Windows and Averaging
A.6.1 Windows
A.6.2 The Harris Test
A.6.3 Averaging and Power Spectral Density
A.7 Correlations and Spectra
A.8 FRF Estimation and Coherence
A.8.1 FRF Estimation I
A.8.2 The Coherence Function
A.8.3 FRF Estimators II
A.9 Wavelets
A.9.1 Introduction and Continuous Wavelets
A.9.2 Discrete and Orthogonal Wavelets
A.10 Filters
A.10.1 Introduction to Filters
A.10.2 A Digital Low-Pass Filter
A.10.3 A High-Pass Filter
A.10.4 A Simple Classification of Filters
A.10.5 Filter Design
A.10.6 The Bilinear Transformation
A.10.7 An Example of Digital Filter Design
A.10.8 Combining Filters
A.10.9 General Butterworth Filters
A.11 System Identification
A.11.1 Introduction
A.11.2 Discrete-Time Models in the Frequency Domain
A.11.3 Least-Squares Parameter Estimation
A.11.4 Parameter Uncertainty
A.11.5 A Case Study
A.12 Summary
References
Appendix B EssentialLinear StructuralDynamics
B.1 Continuous-Time Systems: The Time Domain
B.2 Continuous-Time Systems: The Frequency Domain
B.3 The Impulse Response
B.4 Discrete-Time Models: Time Domain
B.5 Multi-Degree-of-Freedom (MDOF) Systems
B.6 Modal Analysis
B.6.1 Free, Undamped Motion
B.6.2 Free, Damped Motion B.6.3 Forced, Damped Motion
References
Index
Observaciones 2012
Paginas 656
Medida 17x24
Precio 145,00 €
