A Damage Classification Approach for Structural Health Monitoring Using Machine Learning

Within this work, we present a novel approach for Structural Health Monitoring (SHM) of composite materials during the autoclave curing process. For this research, samples of Carbon Fiber Reinforced Polymer (CFRP) have been manufactured with optical fibers embedded, both in the middle of the stack and on its surface. The strain development of the CFRP during the curing process has been monitored employing the Optical Frequency Domain Reflectometry (OFDR) technique. This technique allows the conditions of the element to be tracked not only during the operational lifetime of the composite, as in most of the research in this field, but also throughout the manufacturing process. The reported measurements demonstrate the potential of distributed optical fiber sensors for monitoring CFRP structures while their manufacturing, providing a unique opportunity for early detection of damages that may result in critical failures (e.g. delamination and debonding) and for improving the quality of the manufacturing processes and the final strength of the product.

The importance of developing robust monitoring systems that can detect and locate progressive deterioration in structures or abrupt damage induced by extreme loading events is well recognized in the aerospace, mechanical, and civil engineering communities. In the case of civil structures such as buildings, bridges, off-shore platforms, or dams, the most commonly used approach for structural health monitoring ~SHM! relies on visual inspections that are labor intensive, often not timely, and may miss hidden damage. A good example of these limitations is found in the fractures induced by the 1994 Northridge earthquake in welded momentresisting beam-column connections in many steel buildings that were not discovered until several months after the earthquake because there was no distress apparent from visual inspections. An alternative to visual inspections and localized testing that is being aggressively investigated in various engineering disciplines is that of structural health monitoring based on vibration signals. A typical SHM system consists of a distributed set of motion sensors ~e.g., accelerometers, velocimeters, or fiber-optic strain gauges! connected through a data acquisition system to a central processing unit that may be in an operator room. For remote real-time monitoring of structures, the processed data must be on-line, possibly through a dedicated line or via an Internet link. The basic idea is well established, namely, the dynamic characteristics of a system ~usually natural frequencies and mode shapes! that can be identified from recorded motions are a function of the physical parameters ~mass and stiffness distributions!. Therefore, changes in physical properties resulting from damage may be inferred from changes in the identified dynamic characteristics using suitable algorithms. Then, in theory, damage may be detected, localized, and assessed through vibration monitoring. In practice, however, realization of the concept has proved to be a very challenging proposition because the inverse problem that must be solved, for typical operating conditions, is poorly conditioned, that is, results are sensitive to measurement noise and modeling errors. An important part of the current effort to affect positive progress in SHM technology in civil engineering structures is the development of well-defined benchmark problems that allow comparison of the performance of various techniques for realistic conditions. This special issue of the Journal of Engineering Mechanics contains eight papers that define and study a benchmark developed by the Task Group on Structural Health Monitoring that was formed in 1999 under the sponsorship of the Dynamics Committee of ASCE and the International Association of Structural Control ~IASC!. These papers expand on some partial and

Composite structures are generally more susceptible to impact damage than non-composite structures, and early identification of damage is the primary goal of structural health monitoring (SHM). If such damage remains undetected or reaches a critical size, it can lead to sudden collapse and catastrophic failure. Modern SHM methods aim to preserve the integrity of composite structures through continuous inspection, monitoring, and damage assessment, including detection, localization, quantification, classification, and prognosis. These methods use sensor-based technologies to assess vibration, extension, and acoustic and thermal emission. This paper provides a review of various computational methods including physics-based methods (signal processing techniques, modal analysis, and finite element model updating) and optimization methods (inverse problems, particle swarm optimization, topology optimization, genetic algorithms, time series analysis, and hybrid techniques), alongside machine learning methodologies employing neural networks as well as deep learning for damage identification in composite structures. These computational and learning-based techniques are widely applied in the development of algorithms, optimization strategies, and hybrid frameworks for SHM. The review further summarizes the applications, advantages, and limitations of each method according to structure type and damage characteristics. The key emphasis of this review is on integrating computational approaches, as well as machine learning, to enhance the efficiency of damage identification. The conclusion is drawn based on an overview of the literature, focusing on the contributions of different computational methods and machine learning for damage identification in composites.

Structural Health Monitoring via Measured Ritz Vectors Utilizing Artificial Neural Networks

This article proposes a novel dynamic response reconstruction approach for structural health monitoring using densely connected convolutional networks. Skip connection and dense block techniques are carefully applied in the designed network architecture, which greatly facilitates the information flow, and increases the training efficiency and accuracy of feature extraction and propagation with fewer parameters in the network. Sub-pixel shuffling and dropout techniques are used in the designed network and applied to reduce the computational demand and improve training efficiency. The network is trained in a supervised manner, where the input and output are the measurements of the available channels at response available locations and desired channels at response unavailable locations. The proposed densely connected convolutional networks automatically extract the high-level features of the input data and construct the complicated nonlinear relationship between the responses of available and desired locations. Experimental studies are conducted using the measured acceleration responses from Guangzhou New Television Tower to investigate the effects of the locations of available responses, the numbers of available and unavailable channels, and measurement noise. The results demonstrate that the proposed approach can accurately reconstruct the responses in both time and frequency domains with strong noise immunity. The reconstructed response is further used for modal identification to demonstrate the usability and accuracy of the reconstructed responses. The applicability of the proposed approach for structural health monitoring is further proved by the highly consistent modal parameters identified from the reconstructed and true responses.

A wavelet-based approach is proposed for structural damage detection and health monitoring. Characteristics of representative vibration signals under the wavelet transformation are examined. The methodology is then applied to simulation data generated from a simple structural model subjected to a harmonic excitation. The model consists of multiple breakable springs, some of which may suffer irreversible damage when the response exceeds a threshold value or the number of cycles of motion is accumulated beyond their fatigue life. In cases of either abrupt or accumulative damages, occurrence of damage and the moment when it occurs can be clearly determined in the details of the wavelet decomposition of these data. Similar results are observed for the real acceleration data of the seismic response recorded on the roof of a building during the 1971 San Fernando earthquake. Effects of noise intensity and damage severity are investigated and presented by a detectability map. Results show the great promise of the wavelet approach for damage detection and structural health monitoring.

Ultrasonic guided waves are sensitive to many different types of defects and have been studied for defect recognition in rail. However, most fault recognition algorithms need to extract features from the time domain, frequency domain, or time-frequency domain based on experience or professional knowledge. This paper proposes a new method for identifying many different types of rail defects. The segment principal components analysis (S-PCA) is developed to extract characteristics from signals collected by sensors located at different positions. Then, the Support Vector Machine (SVM) model is used to identify different defects depending on the features extracted. Combining simulations and experiments of the rails with different kinds of defects are established to verify the effectiveness of the proposed defect identification techniques, such as crack, corrosion, and transverse crack under the shelling. There are nine channels of the excitation-reception to acquire guided wave detection signals. The results show that the defect classification accuracy rates are 96.29% and 96.15% for combining multiple signals, such as the method of single-point excitation and multi-point reception, or the method of multi-point excitation and reception at a single point.

Eddy current testing is a nondestructive testing method, which is used to detect discontinuities and defects in conductive materials. Using this technique, two different types of artificial defects in a railhead were evaluated in order to analyze the relationship between different types of defects and eddy current signals, and to obtain data on the size of the rail surface defects and crack location. The actually used rail sample was also studied. Surface cracks and defects were clearly observed as amplitude and phase changes of detected signals. This study succeeds in quantitatively analyzing and discriminating the damage types.

In the machining of precision components, a range of defects, including scratches, pits and uneven materials, will inevitably manifest on the surface of the components. Such defects not only result in a reduction in the quality and precision of the parts, but also have an impact on the overall quality, performance, and longevity of the equipment. It is therefore of particular importance to implement effective detection and repair procedures for surface defects in parts. The implementation of efficient detection methods enables the identification and rectification of defects in a prompt and effective manner, thereby ensuring the reliability and long-term stability of the equipment. In this work, we propose a deep neural network-based model that utilises deep convolutional neural networks (DCNNs) to extract features on the surface of parts, and gradually extracts the spatial and textural features of images through multiple convolutional layers and pooling layers. Subsequently, the aforementioned features are fed into the fully connected layer, where complex feature representations are further learned. Finally, the accurate classification of defect types is achieved through multilayer neural networks. The model is thus capable of effectively capturing and identifying different types of surface defects, thereby improving the accuracy and efficiency of identification. The experimental analysis demonstrates that the model exhibits high accuracy and robustness in the identification of different types of defects, thereby verifying its effectiveness and reliability in practical applications.

The present paper takes place in the context of identification problems: the aim is to propose a new approach for the detection of damages within large structures via the deployment of a set of moving sensors. The goal of the method is to optimally control the motion of the swarm driving the sensors in the optimal measurement regions, in order to perceive the field as clean as possible and to enhance the identification of any possible irregularity along the structure. The optimal control of the sensors motion and the consequent damage identification are achieved through a two-step strategy. The first one is based on: i) models for the dynamics of the structure, for the dynamics of the sensors; ii) LQR control logic to define the optimal trajectory of the swarm. The second step includes the identification of the damage module: the process of data analysis is investigated through use of the technique based on Empirical Mode Decomposition, combined with the Hilbert-Huang Transform.KeywordsStructural Health MonitoringMoving SensorsOptimal ControlEmpirical Mode DecompositionHilbert-Huang Transform

Europe has an extensive transport infrastructure network where bridges play a vital role. Most of them were built as part of the post-World War II reconstruction effort, meaning that we, as society, are already facing the beginning of the end of their design life. This shows the necessity of efficient approaches, complementing visual inspections, for early detection of damage that might jeopardise structural integrity and ultimately might lead to loss of life. This work introduces a novel, pro-active structural health monitoring (SHM) approach to better identify and quantify representative damage types, on prestressed concrete bridges. Based on a numerical simulation of a comprehensive case study available in the literature: the Lezíria Bridge, the results show that damage can be identified with good accuracy for early stages of damaged bridges. Pier settlements and prestress losses are the damage types where the severity is quite accurately quantified even for lower damage severity levels. Once the damage type is identified, it is found that a pair of two vertical displacements reveals to be enough to quantify the damage extent. The results also show potential in the utilisation of the approach for a rational and efficient design of monitoring systems towards damage identification.

Despite the advances in Structural Health Monitoring (SHM) which provides actionable information on the current and future states of infrastructures, it is still challenging to fuse data properly from heterogeneous sources for robust damage identification. To address this challenge, the sensor data fusion in SHM is formulated as an incremental tensor learning problem in this paper. A novel method for online data fusion from heterogeneous sources based on incrementally-coupled tensor learning has been proposed. When new data are available, decomposed component matrices from multiple tensors are updated collectively and incrementally. A case study in SHM has been developed for sensor data fusion and online damage identification, where the SHM data are formed as multiple tensors to which the proposed data fusion method is applied, followed by a one-class support vector machine for damage detection. The effectiveness of the proposed method has been validated through experiments using synthetic data and data obtained from a real-life bridge. The results have demonstrated that the proposed fusion method is more robust to noise, and able to detect, assess and localize damage better than the use of individual data sources.

Applications of Machine Learning (ML) algorithms in Structural Health Monitoring (SHM) have become of great interest in recent years owing to their superior ability to detect damage and deficiencies in civil engineering structures. With the advent of the Internet of Things, big data and the colossal and complex backlog of aging civil infrastructure assets, such applications will increase very rapidly. ML can efficiently perform several analyses of clustering, regression and classification of damage in diverse structures, including bridges, buildings, dams, tunnels, wind turbines, etc. In this systematic review, the diverse ML algorithms used in this domain have been classified into two major subfields: vibration-based SHM and image-based SHM. The efficacy of deploying ML algorithms in SHM has been discussed and detailed critical analysis of ML applications in SHM has been provided. Accordingly, practical recommendations have been made and current knowledge gaps and future research needs have been outlined.

The authors have been developing a new tiny (~12 mm cube and <4 g) Bluetooth wireless accelerometer unit based upon the technology that is prevalent in modern Smart Phones and Smart Watches. The primary intended application of this unit is for training/teaching purposes in vibration-based Experimental Modal Analysis (EMA) approaches for Structural System Identification (SSI) and Structural Health Monitoring (SHM). Digitised tri-axial acceleration data from a single unit can be streamed using purpose-specific software on a notebook or tablet at up to 400 Hz where the longest clear distance of this unit from the streaming device, (smart phone, tablet, laptop computer), is less than approx. 15 m. It is anticipated that tri-axial data synchronously captured from up to 8 accelerometer units at a sweep rate of 100 Hz per set, can be achieved on completion of the second stage of development of this system. This will certainly be suitable for use on model structures for teaching purposes, but also for industrial use (machinery; on-board vehicles; smaller structures) where the clear distance from the computer (or tablet) to these units is less than approx. 15 m. This paper focuses on the performance testing of a collection of commercially available accelerometers and compares these against the capabilities being aimed for in our new miniature accelerometer units.

The guided wave technique is commonly used in structural health monitoring as the guided waves can propagate far in the structures without much energy loss. The guided waves are conventionally generated by the surface-mounted piezoelectric wafer active sensor (PWAS). However, there is still lack of understanding of the wave propagation in layered structures, especially in structures made of anisotropic materials such as carbon fiber reinforced polymer (CFRP) composites. In this paper, the Rayleigh-Lamb wave strain tuning curves in a PWAS-mounted unidirectional CFRP plate are analytically derived using the normal mode expansion (NME) method. The excitation frequency spectrum is then multiplied by the tuning curves to calculate the frequency response spectrum. The corresponding time domain responses are obtained through the inverse Fourier transform. The theoretical calculations are validated through finite element analysis and an experimental study. The PWAS responses under the free, debonded and bonded CFRP conditions are investigated and compared. The results demonstrate that the amplitude and travelling time of wave packet can be used to evaluate the CFRP bonding conditions. The method can work on a baseline-free manner.