Structural health monitoring or short health check is a multidisciplinary branch of engineering that focuses on assuring a component’s structural integrity and operating safety. This is accomplished by making it easier to detect and characterize damage to a structure that may impair its capacity to fulfill its intended function fully and safely. As a result, the challenge of damage detection is at the center of SHM, intending to detect damage as soon as feasible. This is critical for guaranteeing prompt remedial action, which reduces system downtime, total operating and maintenance costs, and the danger of total collapse.

For automatic damage identification in SHM, two techniques have historically been used: 1) physics-based and  2) data-driven.


To extract meaningful information about the damage and its progression from the observed sensor data, physics-based approaches rely on the physical principles regulating structure behavior in some form or another.

However, in the case of environmental variations, complex structures, material nonlinearity, the difficulty in modeling complex real-world structures, and boundary conditions, are some of the factors that make this exclusive reliance on system physics impractical. As a result, such approaches can only be used to monitor the health of very basic structures with well-defined boundary conditions and well-controlled surroundings. As the complexity of the underlying system grows, such an approach becomes increasingly unreliable. Because of recent advancements in information and sensor technology, it is now possible to continuously or sporadic monitor a large number of parameters in-situ in big/complex real-world structures. This promotes the adoption of a data-driven strategy for SHM, in which damage assessment is treated as a form of statistical pattern identification issue, at least at a lower level, thereby avoiding some of the key problems associated with the physics-based approach.

  • A physics-based technique for measuring structural health becomes very inaccurate when the structures being monitored are considerably complex or when fluctuations in ambient variables cannot be ignored.


Machine learning (ML) approaches have been widely used by researchers in the last several decades for both vibration-based and ultrasonic-guided wave-based damage detection. In SHM, machine learning is used to create models or representations for mapping input patterns in sensor data to output targets for a damage assessment at various levels.

Traditional machine learning algorithms, on the other hand, are restricted in their capacity to interpret vast volumes of raw sensor data. As a result, rigorous engineering and extensive domain expertise are required to extract damage-sensitive characteristics from raw data, which are then put into an appropriate machine learning model. Excluding the algorithm, the damage detection system’s overall performance is determined by the damage-sensitive characteristics used. The trouble with hand-crafted features is that, although they may be sub-optimal for the structure in question, there’s no assurance that they’ll work for other structures. Deep learning (DL) approaches have gotten a lot of attention in recent years as a way to solve this issue. These allow raw data to be utilized, and features are automatically learned from data using a general-purpose learning technique.

Deep learning identifies subtle characteristics in high-dimensional data automatically, which has led to its broad use across a wide range of applications. The advancement of processing capacity through the development of high-tech CPUs and GPUs, availability of vast volumes of big data, and development of new learning algorithms will all ensure that DL thrives in the future.

While supervised learning is the most popular and well-developed kind of learning, damage detection in many SHM applications must be done unsupervised. This is because data matching to various damage situations is extremely unlikely to exist for real-world buildings.

The lowest degree of damage assessment has already been achieved by unsupervised learning using what is known as novelty detection. Higher degrees of damage assessment in real-world structures necessitate either the development of more comprehensive data gathering systems or a method for augmenting insufficient training data by introducing some type of past knowledge into the learning/training process.