Structural health monitoring: Closing the gap between research and industrial deployment

Introduction

There has been a huge and increasing volume of research on structural health monitoring since the 1970s with ∼17,000 papers published in the last decade alone (Scopus search on ‘structural health monitoring’ from 2008 to 2017 conducted in March 2017). However, this enormous research effort has yielded only a small number of routine industrial applications.

Structural health monitoring (SHM) and non-destructive evaluation (NDE) both aim to assess the integrity of a structure non-destructively and, depending on the definitions employed, the fields can overlap to a greater or lesser extent. In this article, I define NDE as the use of measurements taken using removable transducers and instrumentation to assess structural integrity. It is often carried out at plant shutdowns so measurements are usually infrequent and are generally interpreted by the test technician on a one-off basis without attempting to compare the signals obtained in the current inspection with those obtained previously. By contrast, SHM involves permanently attached transducers, often combined with instrumentation, and so enables frequent measurements during operation. The signals obtained are often interpreted by comparing them with previous measurements using a process commonly called baseline subtraction; the detection of damage from the sequence of signals may be automated.

The subject of SHM is very broad and, if defined widely, can include applications on very different structures with very different requirements. Table 1 splits the subject into four categories:

OPEN IN VIEWER

Table 1.

Classification of SHM techniques.

Type of SHM	Availability ofstandards	Type ofMeasurement	Applications onreal systems	Number of sensorsrequired for fullcoverage
Machine condition monitoring a	Many	Mainly passive	• Multiple • Routinely applied in industry	Increasing b
Global monitoring of large structures – in practice, Structural Identification (usually of large structures, for example, bridges)	Some	Mainly passive	• Increasingly common in some applications but not mature • Many trials
Large area monitoring for localised damage. Full coverage of a large structure would typically require multiple such systems.	Limited	Mainly active	• A few commercial applications • Many trials
Localised damage detection, for example, cracks and corrosion	Limited	Mainly active	• A few specialist commercial applications • Many trials

a

Machine condition monitoring is included here for completeness but is not strictly concerned with structural health.

b

There is increasing interest in the use of full field measurements, for example, extracting vibration signatures from a video camera image sequence or from a scanning vibrometer. This enables multi-point measurements from a single measurement device so this may enable large area monitoring for localised damage without a large increase in the number of sensors.

Machine condition monitoring employs mainly passive measurements, for example, vibration levels during operation or debris levels in lubricating oil. Similarly, structural identification can be done by monitoring the vibration response to wind or traffic loading, though it is also possible to use active excitation. By contrast, large area and local monitoring generally involve measuring the response to an active excitation input, though acoustic emission (AE) can sometimes be used. The number of sensors required for full structure coverage increases going down the table.

This article discusses each area in more detail, and then goes on to discuss the business motivation for installing an SHM system since, unless this is clear, the technology will not be widely adopted. In traditional NDT, data interpretation is typically done by the trained technician doing the inspection; in contrast, SHM systems can generate vast amount of data that has to be analysed. This problem has not received sufficient attention in the literature and is discussed in a separate section. This is followed by a discussion of the design of SHM systems and the validation of their performance as unless they can be shown to be reliable, the required investment will not be made. The article does not claim to be a comprehensive review of the state of research in the field – this has been covered by many review papers, some of which are referenced here; it concentrates on applications of the technology and does not pretend to discuss state-of-the-art research that has not reached commercial use. Rather, this article seeks to focus discussion on the barriers to adoption of the technology, and the work required to overcome them. Also, by splitting the field into different categories, it hopes to promote focused discussion around better defined topics.

Machine condition monitoring

As discussed above, machine condition monitoring is not strictly concerned with structural health; however, many authors treat it as a type of SHM (see, for example, Farrar and Worden ² ) so it is included here for completeness and because, as a relatively mature field, useful lessons can be read across to the rest of SHM. Randall ³ gives an excellent introduction to the subject; he identifies three maintenance strategies:

Condition monitoring requires information about the internal condition of the machine to be obtained externally. Vibration analysis is the most commonly applied method, and the analysis techniques have advanced greatly over the historic observation that if a machine is vibrating more than normal then it is likely to be faulty. Lubricant analysis can be used to identify wear particles and chemical contaminants and is also commonly applied, while the use of thermography to identify temperature anomalies is increasing.

There is a choice between continuous and intermittent monitoring; once a fault is present, the resultant vibration signature will always be present so intermittent monitoring does not risk missing the problem. This contrasts with crack monitoring using AE ⁴ where the emission is produced by the crack growth events, so continuous monitoring is essential. The choice between continuous and intermittent monitoring is analogous to that between NDT and SHM – in intermittent monitoring, the instrumentation costs are spread over many machines and the data analysis is likely to be done by the trained operator, while in continuous monitoring, the instrumentation is permanently installed on one machine and generates large volumes of data that have to be interpreted. Lubricant analysis looks for build-up of debris over time and so is inherently an intermittent method.

Mitchell ⁵ gives a fascinating account of the 70 years of progress in condition monitoring to 2007. He identifies a series of key advances:

While some early implementations were bedevilled by false calls, the analysis techniques have advanced greatly and the technology is now mature and widely applied; multiple case studies demonstrating the detection of out-of-balance, bearing faults, shaft cracks, fractured gear teeth and other problems are available on the websites of the large system suppliers, for example, GE Measurement, ¹⁰ Honeywell ¹¹ and National Instruments. ¹²

It is interesting that, in contrast to structural monitoring, many of the developments in machine monitoring identified by Mitchell ⁵ originated in industry rather than academia, and were applied immediately. Also in contrast to structural monitoring, only a small number of passive measurements is required to cover the whole machine, and it is relatively easy to link measurements to remnant life via fleet statistics incorporating readings of load, pressure, temperature and other process parameters, along with the specific condition monitoring measurands. The longer history of machine condition monitoring, the small number of sensors required for full coverage and the large installed base on similar machines facilitating fleet statistics are probably the main reasons for the much larger number of practical applications of machine condition monitoring compared to the rest of SHM.

Global monitoring of whole structures

Attempts to obtain local damage detection with full structure coverage

Condition monitoring aims to infer the health of a whole machine from measurements at a small number of locations, and in the 1970s, researchers started to investigate the possibility of assessing the integrity of structures from a similarly small number of measurements. The work concentrated on vibration measurements since vibration modes are properties of the whole structure, and the resonance frequencies are independent of measurement position provided there is motion in the mode at the chosen point. The chief motivation for the initial work was to assess the integrity of offshore oil platforms in the North Sea or Gulf of Mexico from measurements of the vibration response to wave excitation taken above the water line. However, it was found that the measured resonance frequencies were sensitive to environmental conditions such as marine growth that adds significant mass to the structure, and also to changes in the mass distribution on deck, and these changes effectively corrupted the data. The tests also identified uniqueness issues associated with the damage prediction if only resonance frequencies are used, and this lack of success led to the oil industry abandoning the technology in the mid 1980s. ¹³ There has been a great deal of further work since that time and many reviews, for example, Farrar and Doebling ¹⁴ and Fan and Qiao. ¹⁵

The resonance frequencies of the lower structural modes are affected by changes in boundary conditions, mass distribution and temperature-induced modulus changes much more than by the stiffness changes produced by localised damage unless the damage is very severe. Changes in boundary conditions due to scour around a pier, expansion joint failure or seizing of bearings designed to restrain motion in a particular direction can be very serious and are detectable from changes in the lower structural modes. However, localised damage produces much smaller changes, for example, Figure 1 shows the change in the resonance frequency of the first flexural mode for a crack-like defect at the root of a cantilever beam. A crack through 1% of the cross section produces a frequency change of ∼0.1% and this can also easily be produced by other effects, leading to high false call rates; the sensitivity is even lower at other locations on the beam, particularly towards the free end. ¹⁶ This problem with resonance frequency measurements has led to a great deal of work on mode shape measurements that are then often double-differentiated to obtain the curvature, and hence the flexural strain; a variety of processing methods can then be used to assess whether damage is present. Farrar and Jauregui^17,18 assessed six processing methods for the detection of severe simulated cracks in one girder of a 48.8-m-long bridge over the Rio Grande in Albuquerque, New Mexico. Data from the first six modes were used, and in the initial measurements using 26 accelerometer measurement locations over the bridge, the damage was not reliably located until it had removed half the thickness of the I-girder, including one flange. The performance improved considerably when measurements from 16 accelerometers distributed along the girder in which damage was introduced were included in the analysis; a 2-ft-long simulated crack in the web of the girder was then detectable. This indicates that useful damage detection performance might be obtainable if significant numbers of measurements are taken along all the members of interest. This moves away from the original goal of assessing structural integrity from only a small number of measurements, as is achieved with machine condition monitoring. However, there are increasing possibilities of measuring vibration at multiple locations on a structure using a remotely positioned scanning laser vibrometer ¹⁹ or by processing frame sequences from a video camera; ²⁰ these approaches greatly increase the attractiveness of multi-point measurement methods.

OPEN IN VIEWER

Figure 1.

First-mode flexural resonance frequency reduction predicted using finite element analysis as a function of the depth of narrow slot across the full width at root of a cantilever beam.

The cost effectiveness and reliability of modal measurements in the presence of environmental and loading changes would have to be assessed before it was implemented on an operational structure. Peeters and De Roeck ²¹ monitored a bridge in Switzerland over a period of a year and tracked the variation in natural frequencies with temperature. They showed that it was possible to correct the natural frequency measurements for temperature changes and so improve the detectability of severe damage to the bridge piers; they did not investigate the effect of load changes. To the author’s knowledge, a modal-based local damage detection system has not been implemented on civil engineering structures, but such a system was implemented on the NASA space shuttle ²² where the exterior surface was inaccessible due to the thermal protection system which made conventional NDT particularly difficult.

Monitoring for localised damage

Dealing with data

In most conventional inspection, the data interpretation is done directly by the trained technician operator so those responsible for the integrity management of the structure are only alerted when there is an abnormality. In contrast, SHM data are generated automatically; it is typically much more frequent than NDT inspections (e.g. daily rather than annual) and is transmitted directly to the structure operator. As the number of monitoring locations on a structure or plant increases, this data stream can become unmanageable unless some automatic pre-processing is applied. When the thickness monitoring system^80,81 was first deployed at 100s of locations on a plant, operators described the experience of frequent, multi-point data as being like ‘drinking from a hosepipe’. Professor Zhu Mao cites the design lead of a major Asian bridge saying that they do not know exactly what information to keep, neither do they know what intelligent feature(s) to extract, so they measure and save everything they can over many channels with perhaps unnecessarily high sampling rates. This results in terabytes of data every day, and several rooms filled with hard drives for historical data storage. Goulet and Smith ⁸⁸ note that with increasing availability of communication systems and decreasing cost of sensors, more and more structures are being measured, but our capacity to analyse large amount of data is only marginally increasing.

In a report to the Scottish Parliament investigating whether an SHM system could have detected a problem early enough to prevent the need to close the Forth Road Bridge for urgent repairs, an expert witness, Richard Fish, is quoted ‘The big issue with what is got out of structural health monitoring concerns the translation from data to information. People can get bombarded with millions and millions of bits of data, but they have to be interpreted’. ⁸⁹

The CSIC ⁹⁰ stresses the importance of data handling and interpretation and presents a pyramid model, a simplified version of which is shown in Figure 2. Here, data are processed to a digestible form before being passed to the decision-making level that can be more or less automated.

OPEN IN VIEWER

Figure 2.

Pyramid model of decreasing data volume and increasing data value (after Cambridge Centre for Smart Infrastructure and Construction (CSIC) ⁹⁰ ).

This issue was also recognised in the early years of machine condition monitoring.^5,91 An article in Aviation Maintenance ⁸ notes that without trend monitoring, many internal engine problems will not manifest themselves until a turbine blade is shed through the shroud or the engine shuts down. Operators could obtain trend monitoring data through the engine control unit (ECU) data, or by observing temperature and pressure gauges in the cockpit and manually recording the data during each flight. However, to obtain a trend, the operator would need to collect the data from a number of flights, plot the data in a spreadsheet and then analyse it. The amount of work involved with trending a fleet, coupled with the potential for saving substantial sums by catching small failures before they become inflight engine shutdowns, has led to a growth in the number of third-party trend monitoring programmes. Farrar and Worden ¹³ point out that the success of machine condition monitoring is due in part to

These factors have allowed condition monitoring to make the transition from a research topic to industry practice several decades ago, resulting in comprehensive, automated condition management systems that handle the huge data flows efficiently.

It seems likely that as the number of SHM systems deployed increases, there will be more interest from third parties in developing the required analysis and decision-making tools. There will probably be a gradually increasing degree of automation; when only a small number of sensors are deployed, it is feasible for an operator to look at the output from each, while as the number increases, displays highlighting those with anomalous signals are useful so that these can be followed up. When a very large number are deployed, automatic decision-making may become feasible, underpinned by validation on the large volume of historic data available from many sites.

Rytter ⁹² identifies four levels of sophistication in SHM information: existence, location and severity of damage, followed by prognostics – assessment of whether the structure is still fit for service and what its remnant life is. Identification of the type of damage should be added to the second level since severity estimation requires the type of defect to be known. Ultimately, it may be possible to do automatic diagnosis and prognostics, but this requires a very large installed base so that the software can be tested very thoroughly to ensure reliability.

Business case for monitoring

One reason for the slow take-up of SHM technology is lack of attention to the business case for monitoring. Even in the well-established machine condition monitoring field, Mitchell ⁵ notes that, despite over 50 years of demonstrated success and benefits, condition assessment programmes are still not fully accepted within an industrial operating culture as a permanent, essential business activity. Many successful programmes are reduced or terminated altogether as a ‘cost saving’ measure because there are few failures. He suggests that for many technical specialists, analysis and details such as determining that a bearing is failing due to a defect on the outer race are of far greater interest than demonstrating the business value they contribute to their companies. He concludes that ‘the key to success is not technology but awareness of value, resources and time’. The value identified should also include avoidance of the costs of consequential damage following catastrophic failure. ⁸ However, machine condition monitoring is increasingly accepted, partly driven by the possibility of online data gathering and central processing (see, for example, Solar Turbines ⁹³ ).

If SHM is to be more widely applied, it is similarly crucial to be clear about the business case for employing the new technology. Both NDT and SHM always compete with a planned replacement strategy, and sophisticated methods compete with, for example, complete strip-down and visual inspection; the added information from SHM must be sufficiently valuable to be worth the cost. SHM will never totally remove the need for intervention, but should enable it to be better targeted and less frequent. It may also be most cost effective to mix SHM and traditional inspection. For example, Professor Zhu Mao says that the bridge monitoring strategy in China can involve a permanently installed approach for global parameters or large changes due to, for example, scour at a pier, and periodic visual inspection for local defects since the number of sensors required to monitor all the critical areas for local defects would be uneconomic. The SAE standard ⁴¹ lists the main economic motivations for the use of SHM systems as to

Another significant motivation is that monitoring may reduce the probability of catastrophic failure and consequent litigation/liability issues, and may also help to satisfy regulators. For example, the boiler spine monitoring ⁶³ was introduced following the detection of a large, unexpected defect on a routine inspection and helped to satisfy the regulator. The rail break detection system ⁶⁴ was developed in response to past failures and concern that future incidents would lead to very severe consequences for the operators. However, even in these cases, SHM must still compete with periodic, possibly very frequent, NDT inspection or planned replacement. It is therefore important that the costs of an SHM system are properly accounted for; it is easy to concentrate on the initial capital cost, but the system will need maintenance (e.g. battery replacement) and may have a limited life. As discussed above, it is also vital to include the recurrent cost of dealing with the data.

Therefore, for a mature SHM system, cost–benefit analysis would be required to justify replacing conventional NDT tasks with advanced SHM tasks. The cost–benefit analysis should highlight the hidden costs of NDT and quantify the savings of SHM. ⁴¹ Hidden costs include access costs that are very high when NDT personnel need to enter a hazardous or inaccessible area. For example, in 2006, Boeing indicated that for a typical fleet of aircraft, 70% of all structural maintenance expenses are incurred inspecting airframes during periodic maintenance tasks; the majority of the inspection expenses are associated with effort spent to access the inspection areas. In 2004, Boeing calculated that for a 747–400 commercial aircraft, 25,000 h are spent inspecting for corrosion, but over 21,000 of these hours are spent gaining access to hard-to-inspect areas, and only 4000 h are spent doing the actual inspection. Because of the initial costs required to develop, produce and install the SHM systems in each aircraft, conclusive evidence confirming the SHM functionality would be required to encourage the procurement of an SHM system. The conclusive evidence would be, for example, structural fault detection capabilities observed during the operation of the aircraft. The occurrences of structural faults such as cracks are infrequent, and hence, years of flight tests might be required to collect validation evidence; small number of flights would only be sufficient to prove the system ‘fitness for flight’ and would be insufficient to prove ‘fitness for purpose’. As a result, SHM systems are yet to be applied routinely on civil aircraft, though the manufacturers continue to invest in development and have ambitious future plans. ⁹⁴

Aktan and Brownjohn ²⁴ note that it is important that the owner/operator is convinced of the value of SHM (St-Id) in decision-making and that the technical team is competent and experienced; if these requirements are not met, it is best not to expect much from it. They suggest that SHM (St-Id) adds most value:

Similarly, the CSIC stresses the importance of setting out the business case for digital investment in infrastructure, noting that it is a hard fight to get backing to switch cash from capital projects to invest in digital tools for existing infrastructure. They suggest that shared case studies backed by strong financial data setting out the benefits of a whole-life approach in terms of longer lifespan and greater utilisation of existing infrastructure are required. ⁹⁰

The benefits identified for SHM tend to concentrate on cost reduction via, for example, only incurring access costs once for the installation of the system, rather than repeatedly for periodic inspection. Randall ³ also points out that maintenance is often regarded as a cost centre, but that condition-based maintenance via machine condition monitoring can convert it to a profit centre; this idea has been promoted by Al-Najjar and Alsyouf. ⁹⁵ Similarly, permanently installed corrosion monitoring giving sufficiently frequent readings to enable rapid estimation of corrosion rate enables the effect of changing feedstock on the plant to be evaluated. This makes it possible to use cheaper, higher total acid number (TAN) crude oil while still being confident that the corrosion rate is within safe limits. ⁸¹ Monitoring therefore enhances the profitability of the plant while giving the operators and regulators confidence that the more corrosive feedstock will not lead to a catastrophic failure. Similarly, increased availability of plant is a significant benefit in helicopter operations. ⁹⁶

SHM system design, performance assessment and validation

In formulating an NDT procedure for the inspection of a critical structure, it is routine to specify a required POD for a critical defect size and this must be demonstrated (see, for example, Department of Defense). ⁴² Goulet and Smith ⁸⁸ identify the importance of understanding the potential effectiveness of an SHM system and the SAE standard ⁴¹ stresses that environmental and operational conditions should be incorporated in the analysis of the system. However, the standard also recognises that POD methodologies developed for NDT may not always be directly applicable to SHM systems, so new approaches may be required.

When considering implementation of an SHM system rather than a traditional periodic inspection approach, it is vital to consider the design of the whole system not just, for example, the transduction. Todd and Flynn ⁹⁷ propose a structured design approach for SHM systems, starting with key questions:

Todd and Flynn ⁹⁷ then go on to use a Bayesian approach to the design of the system, further details on an optimal sensor placement study being given by Todd and colleagues.^98,99

Farrar and Doebling ¹⁴ note that the ability to discriminate changes in the modal properties as a result of damage from those due to variations resulting from changing environmental conditions and from the repeatability of the tests has received very little attention; they point out that a high level of uncertainty in the measurements will prevent the detection of low levels of damage. Peeters and De Roeck ²¹ present one of the very few modal-based damage detection studies to report statistical variations associated with the measured modal parameters, and there are even fewer reports of false positive studies. Additionally, research should be focused more on testing of real structures in their operating environment such as that done in Doebling et al., ¹⁰⁰ rather than laboratory tests of representative structures; ¹⁴ attention to these issues is crucial if SHM is to avoid the false call problems that plagued some early implementations of machine condition monitoring.

Farrar and Worden^2,13 view SHM in a statistical pattern recognition or machine learning paradigm, and Stull et al. ¹⁰¹ discuss how the robustness of an SHM method might be assessed. They identify the receiver operating characteristic (ROC) curve as a particularly valuable metric as it plots the POD against the probability of false alarm (PFA). Unfortunately, producing the ROC curve for an SHM system is even more difficult than for an NDT inspection as the effect of environmental variations and potential transducer/instrumentation drift over a long period, as well as the changes due to damage of different types and severity at different locations must be evaluated. This would involve an impractical number of test structures in the operational environment and an extensive programme of damage introduction over a long period. Another approach is to simulate the received signal; modern computational resources mean that it is relatively straightforward to reliably predict the signature produced by damage in vibration, guided wave, ultrasonic or magnetic measurements, even when it has complex shape. However, reliable prediction of signal changes due to environmental and other variability is much more difficult, whereas obtaining experimental data with environmental variation on an undamaged structure is easy. Liu et al. ⁶² proposed a methodology of measuring data over multiple environmental cycles on an undamaged structure and synthetically adding damage to the resulting signals as shown schematically in Figure 3. This approach enables the addition of damage at different locations with different growth patterns, and the investigation of other practical parameters such as the extent of environmental changes, damage severity and frequency of readings. They illustrated the approach on the guided wave monitoring of a pipe system; measurements were taken on an undamaged pipe over multiple cycles of heating and cooling and the effect of damage was modelled using FE analysis, the damage signature being added to the real measurements on the undamaged structure. Figure 4 shows estimated damage growth and ROC curves for one particular case of damage growing from 0% to 1% cross section loss at a particular location over the course of 100 readings with a 10°C random temperature variation over the period of the measurements. The same data were processed using baseline subtraction (Figure 4(a) and (b)), singular value decomposition (Figure 4(c) and (d)) and ICA (Figure 4(e) and (f)). The results show that ICA provides a virtually perfect detector with almost unity POD even at very low PFA (note that the PFA is plotted on a log scale in Figure 4(f) and the corresponding graphs for the other processing methods); the estimated damage amplitude of Figure 4(e) also tracks the true amplitude very closely. Singular value decomposition also performs well, whereas baseline subtraction gives much poorer performance with the lower 95% confidence bound giving very low POD at a false call rate of over 1%. The framework has therefore enabled robust identification of the best data processing methodology and this has since been demonstrated in a blind trial. ⁶⁰

OPEN IN VIEWER

Figure 3.

Schematic diagram of the SHM system evaluation framework proposed in Liu et al., ⁶² with dashed boxes indicating parameters, dark-shaded boxes indicating raw/processed data and white boxes indicating data operations. Synthetic datasets are generated using undamaged experimental signals and predicted damage signals. The ROC curve is then generated after damage detection is performed on the synthetic datasets.

OPEN IN VIEWER

Figure 4.

Results from guided wave monitoring of a pipe system. Case shown is for 100 readings taken while damage at a particular location grew to 1% cross section loss and the temperature varied randomly over a 10°C range. Data processed using (a, b) baseline subtraction algorithm; (c, d) singular value decomposition; and (e, f) independent component analysis. (a) Amplitudes of baseline-subtracted residual at damage location as a function of time (blue) and true damage reflection amplitude (black). Reflection amplitude is not a linear function of cross section loss so linear ramp cross section loss does not produce linear ramp of reflection amplitude; (c, e) weights from SVD and ICA (blue) over time, along with true damage amplitude (black). (b, d, f) Receiver operating characteristic (ROC) curves showing probability of detection (POD) versus probability of false alarm (PFA). Solid lines show mean ROC curve and dashed lines show lower 95% confidence bound (adapted from Liu et al. ⁶² )

The framework of Figure 3 can take data from any installed monitoring system to predict its damage detection performance. This makes it possible, for example, to assess whether a defect of a given size at a particular location would be detected reliably at a given data collection frequency. This enables the system to be tuned to meet particular requirements and it could be used in a safety case to provide assurance that had a defect of a given size and location been present, it would have been detected.

Conclusion

There are multiple reasons for the slow transfer of SHM research into routine industrial application; key areas that should be considered to improve the adoption rate are as follows.