Sebastian Lotz, summer semester 2014


Artikel auf Deutsch


Structural Health Monitoring (SHM) is designed for the diagnosis of the current state of a structure, which consists of different components and materials. Structures undergo continuous change caused by ageing processes, environmental influences and also by unforeseen events such as earthquakes or (wind) buffeting. SHM methods, describe the process of the implementation of strategies and techniques for damage detection and damage characterisation for the infrastructure of the aerospace industry, civil engineering and mechanical engineering. [1] Over the last three decades, engineers and materials scientists have increasingly researched techniques for damage detection, as existing systems such as aircrafts, bridges and buildings draw closer to their anticipated design age. For the reason that many of those systems cannot be economically replaced, methods for damage detection have become/been established to assure secure operation beyond the constructively set lifespan. Regular dynamic response measurements are made to monitor the state of the structure or the mechanical system. The current state can be determined by extraction of damage-sensitive properties from these measurement data followed by appropriate statistical analysis. Long-term monitoring takes place through periodic use of the previously described process to guarantee that the structure can still meet its intended function. This is important, as almost every system is inevitablyaged by its operational environment and thus its properties deteriorate. SHM methods are used to assess reliable information in real-time about the function and state of a structure. [2]

Introduction

The first employment of SHM can be dated back to the 18th century: At that time, people known as “wheel-tappers” examined train wheels with a hammer to detect cracks in them. However, research and development of monitoring methods has only really significantly increased over the last 30 years. SHM techniques are designed to replace the existing rathere qualitative methods through more sensitive, quantifiable methods for damage detection. SHM is slowly starting to become established in practice and is increasingly used on numerous real, structural and mechanical systems. [2] These include (as described by Balageas [1]) for example, vibration based methods such as the implementation of smart materials e.g. piezoelectric sensors (PZT) and fibre-optic sensors which provide constant monitoring or stimulation of structures. In addition, the approaches of electromagnetic waves or utilization of the change in electric resistance of systems for damage detection in practical use are described. During the evolution of the field of damage detection, several subdivisions emerged which are employed as a matter of routine in practice. The interdisciplinary nature of SHM’ includes the following subdivisions:

  • Condition monitoring (CM)
  • Non-destructive evaluation (NDE) or also named non-destructive testing (NDT)
  • Health and usage monitoring systems (HUMS)
  • Statistical process control (SPC)
  • Damage prognosis (DP)

SHM is an enhancement of NDT and can be linked as follows: [3]

Fig. 1: Base components of SHM, following Chang [3]


Particularly highlighted are the applications of condition monitoring. These methods already have achieved the transition from theory to practical application in the field of rotating machinery. Complete systems such as buildings, aerospace structures, bridges and so forth can be partially monitored using SHM methods but cannot be fully monitored as a “system”. [2]

How engineers and scientists analyse damage

The exploration of damage in materials is predominantly carried out by materials scientists and engineers. The following questions provide approaches to solve the problem:

  1. What causes the damage?
  2. Which measures can be taken to prevent damage?
  3. Is there damage present?
  4. How fast will the damage spread and reach a critical level?
  5. How can the effects resulting from damage be reduced?

Engineers and material scientists will answer these questions from different perspectives. In SHM, the focus lies with the last two questions and differs for material scientists and engineers, in particular in the extent of the damage. [2] Material scientists examine the microscopic damage occurrence, i.e. the spread of damage (e.g. crack propagation at grain boundaries and many others) of the materials, whereas the engineer places emphasis on the operation, for example, to restrict overloading and thereby reduce damage. Damage is defined as intentional or unintentional changes of the material and/or geometrical properties of structural and mechanical systems, including changes of boundary conditions and system connectivity, which negatively influence the current or future performance of these systems [2]

Definition of structural health monitoring

SHM methods define the process of implementing strategies and techniques for damage detection in the infrastructure of the aerospace industry as well as mechanical engineering. This process includes:

  • A time-dependent monitoring of a system, which periodically measures the dynamic response of a sensor array.
  • the extraction of damage sensitive properties of these measurements and
  • a statistical analysis of these properties is made to determine the current state of the system.

Existing methods are on one hand microscopically based, which develops a fundamental understanding of different modes of material failure. On the other hand, the macroscopic damage level of the materials/components are, for example, examined with NDT or wave-based SHM methods. Examinations of the entire system are carried out directly with methods such as CM, HUMS and many others. In addition, methods have been for damage prognosis and remaining lifespan of systems have been developed that are employed together with SHM. [2]

Statistical pattern recognition paradigm for SHM

In order to define a damage unambiguously, the system’s state has to be compared at different points in time. The reference state, ideally the unimpaired state, is compared with the current state to monitor the changes in the system. When viewing for example, a collapsed building, it is at this moment in a damaged state although no exact information about the initial state of the building is available. [2] However, experiences are automatically recalled which activate rational thinking and say that the building is definitely damaged. In reference to SHM, “knowledge” can be interpreted in data and “experience” can be associated with “learning”. This natural way of damage detection can be described by several mathematical methods more or less precisely. When using SHM'’, the statistical type of pattern recognition is the most common seen and recorded under the term statistical pattern recognition'’ (SPR). The concept of “learning” with “training data” is carried out using SPR. The mathematical frame concept for this method delivers machine learning'’ by Cherkassky and Mulier. [4] SPR is defined by following points:

  1. Operational evaluation
  2. Data acquisition
  3. Feature extraction
  4. Statistical model development for feature discrimination

The method of operation and description of SPR will be explained in the following section.

Operational evaluation

The process of operational evaluation'’ should provide answers related to the implementation of damage identifications. Basic questions for this examination:

  1. What are the reasons for safety and/or efficiency for the use of structural health monitoring?
  2. How is “damage” defined for the system examination and (for the possible different cases of damage) which cases are the most critical?
  3. What are the operational and environmental states in which the monitored system operates?
  4. Where are the limits of data acquisition during operation of the system/structure?

This first step sets limits regarding the sections of the structure that should be monitored and sets the methods to be employed. It would be ideal to determine damage sensitive features of the structure in an inertial state. A continuous evaluation of this data enables the early recognition of damage. [2] Further information about the method of operation and the answers to the questions can be found in Farrar. [2]

Data acquisition

Data acquisition Includes the selection of suitable stimulation and measurement methods as well as the software and hardware to evaluate and store measurement data. The selection is specific and is strongly dependent on the individual case of application. Up until now there has been no general measurement method to monitor every structure equally. The cost factor is at the foremost consideration when choosing data acquisition equipment for constant monitoring. Additionally, the interval when data is to be collected must be set or rather adjusted after every case of application. In the event of an earthquake, the data should be recorded and evaluated immediately before and at periodic intervals after the event, whereas in the case of a machine this should occur during entire operation. [2]

Feature extraction

The identification of 'data features'’ is the area that probably receives the most attention in the field of SHM. In many cases it allows for the distinction between damaged and undamaged system states. To extract features effectively, signal processing methods are used e.g. conversion of a measured time-series into a frequency spectrum. The line spectra contain information about the system’s condition. This is used for example in gear boxes. Optimal features possess a small dimension and are very sensitive to changes in the systems condition. Another option is to create a real or measurement data based parametric model of a system. These parameters or the predictable error of these models become damage-sensitive properties. In particular, standardization of the data is of great importance here as every measurement takes place under different influences and therefore comparability is warranted. [2]

Statistical model development for feature discrimination

The development of statistical models derives from the machine learning“-techniques which are divided in two categories.

  1. Supervised learning: Data from the damaged and undamaged structure are available and can be used for evaluation.
  2. Unsupervised learning: The undamaged state is available as data. [2]

The most commonly used models in SHM can, depending on the type of available data, be allocated to the previously described machine learning groups. An algorithm of supervised learning is group classification, which gives information on a known number of discrete states and most accurately describes the condition of the structure. Regression analysis shows the current state by continuous qualification of the system. One example for unsupervised learning is the method of identification of outliers. [5] This is a five step process to organize the identification of the damage’s state according to Rytter [6]. This can be led by the following five questions.

  1. Is the system damaged? (Existence)
  2. Where in the system is the damage? (Place)
  3. Which type of damage is present/displayed? (Type)
  4. How severe is the damage? (Extent)
  5. How much (secure) lifespan remains? (Prognosis)

For the implementation of two types of SHM these statistical models are used. The first model is that of protective monitoring. It uses damage-sensitive features to monitor the system. As soon as the features exceed previously set limits, the system is, for example, switched off before (additional) damage occurs. The second model is that of predictive monitoring and monitors trends in data features and uses these to predict when vital damage is likely to occur (keyword: maintenance plan creation) [2] The methods data acquisition, feature extraction and statistical model development for feature discrimination of SPR need important accessory information for data processing. These are defined as follows:

  1. Data normalization: Measurement data are normalized to achieve comparability of a system, e.g. in the case of different system states and environmental influences.
  2. Data cleansing: Data is chosen or discarded for the features. This process is often empirical. In signal processing, for example, filters and sampling rate are used.
  3. Data compression: Dimension reduction of measurement data without eliminating sensitive features can provide a conclusion about the state.
  4. Data fusion: Combination of measurement data from multiple sources (e.g. from a sensor array) should enhance the reliability of the damage identifying process. [2]

Comparison of local and global damage detection

Local damage detection methods are limited to a relatively small environment on/in the component. Methods are used that are already known in non-destructive testing such as eddy current, propagation of ultrasound waves or magnetic fields. The use of ultrasound waves is very often used in conjunction with "guided waves" that are introduced into the structure via sensors/actuators. With this sort of application in the near field of the positioned sensor, reflections of the waves at component boundaries or other discontinuities (changes of thickness, holes, ...) make damage detection more difficult. Local methods are very sensitive and make it possible to identify even small defects. Damage inside the material can, for example, reduce local rigidity and thus change the global behaviour of the structure in time and space. These effects are also used in global methods but are not as sensitive as local methods. They are based on low frequency oscillations with which the entire system can be monitored. In this event, for example, changes in the resonance frequencies, dampening or oscillation modes would be looked for/considered, which are then extracted as (if possible, damage-sensitive!) a feature. As a result, it should be possible to distinguish between a damaged and undamaged state. For complete system monitoring, combinations of local and global methods are often deployed. [1]

Fundamental axioms of SHM

Over the last 30 years of research in the discipline of SHM many very good findings have been delivered. From this it is possible to formulate fundamental axioms, however, these are not comparable with those from mathematics. They show much more generally accepted principles and represent fundamental approaches of each SHM methodology. Axioms can be formulated as follows:

  • Axiom I: Each material has inherent cracks or defects.
  • Axiom II: Damage evaluation needs comparison between two system states.
  • Axiom III: The identification as well as localisation of a defect can be achieved through unsupervised learning. The identification of the type and severity of a defect can only be achieved with the supervised learning method.
  • Axiom IV a) Sensors cannot measure damage. The extraction of the feature using signal processing and static classification are necessary to convert sensor data into information about the damage.
  • Axiom IV b) Without intelligent extraction of the features, the following applies: The more sensitive a measurement regarding damage is, the more sensitive it is towards operational changes and environmental influences.
  • Axiom V: The length of time and timescales in connection with the onset of damage or damage evolution demonstrate the necessary features of the SHM-sensing system.
  • Axiom VI: There is a compromise between the sensitivity of a damage regarding the algorithm and its ability to reduce noise.
  • Axiom VII: The detectable size of a damage caused by changes in the system dynamic is inversely proportional to the frequency range of the stimulation.
  • Axiom VIII: Damages increase the complexity of the structure.

These axioms are defined and described in detail in Structural Health Monitoring - A Machine Learning Perspective. [2]

Components of SHM

The following basic SHM-components hardly differ for different applications. They serve as an overview of the basic components of a system for constant monitoring:

  • System/Structure
  • Sensor and sensor systems (wired and Wireless Systems [2])
  • Systems for data acquisition
  • Storage and transfer of measurement data
  • Management of measurements
  • Interpretation and diagnosis of measurements

Sensor systems are installed for constant monitoring on structures such as satellites, bridges, aircrafts etc. Recorded measurements are evaluated and processed according to 'statistical pattern recognition’ to carry out real-time'’ structure monitoring

Examples of application of SHM methods

  • Canton Tower, China: More than 800 sensors were positioned on the outer structure of the tower to monitor the construction and its structural ageing.
  • Golden Gate Bridge, USA: The implementation of 64 wireless sensor nodes distributed over the Golden Gate Bridge is the largest wireless network of SHM.
  • International Space Station: A successful constant monitoring of the ISS structure via SHM assures also future space missions.
  • Windenergieanlagen: Constant online monitoring of wind power stations should reduce maintenance intervals or also cause switching off of the station in critical system states.

Literature

Books

  1. Balageas, D. ; Fritzen, C.-P.; Güemes, A.: Structral Health Monitoring. ISTE Ltd, 2006.
  2. Farrar, C.; Worden, K.: Structural Health Monitoring - A Machine Learning Perspective. John Wiley & Sons, Ltd, 2013.
  3. Chang, F. - H.: Structural Health Monitoring: A Summary Report on the First International Workshop on Structural Health Monitoring, September 18-20, 1997. In Structural Health Monitoring 2000, Proceedings of the Second International Workshop on Structural Health Monitoring, Stanford, CA, September 8 - 10, 1999, Lancaster - Basel, Technomic Publishing Con, Inc, pp. xix-xxiv, 1999.
  4. Cherkassky, V.; Mulier, F.: Learning from Data: Concepts, Theory and Methods. Wiley-Blackwell, 2007.
  5. Overbey, L. A.: Time Series Analysis and Feature Extraction Techniques for Structural Health Monitoring Applications. ProQuest, UMI Dissertation Publishing, 2011.
  6. Rytter, A.; Brincker, R.; Pilegaard Hansen, R. L.: Vibration Based Inspection of Civil Engineering Structures. Aalborg Universitet (DK). Inst. for Bygningsteknik, 1993.
  7. Stepinski, T.; Uhl, T.; Staszewski, W.: Advanced Structural Damage Detection - From Theory to Engineering Applications. John Wiley & Sons, Ltd, 2013.

Conferences

Journals

Der Inhalt ist verfügbar unter der Lizenz a Creative Commons Namensnennung-Nicht kommerziell 4.0 International Lizenz.