1. Introduction
The United Nations included Sustainable Development Goal (SDG) 9 within Agenda 2030, which emphasizes the importance of building resilient infrastructure, promoting inclusive and sustainable industrialization, and fostering innovation [
1]. Achieving SDG 9’s targets is crucial for sustainable economic growth and equitable development [
2]. In this context, advanced technologies are playing a pivotal role in enhancing the efficiency and safety of infrastructure systems. The integration of Industry 4.0 (I4.0) technologies [
3], including the Internet of Things (IoT), artificial intelligence (AI), machine learning (ML), big data analytics, and cloud computing, among others, into infrastructure management can enable real-time monitoring and enhance predictive maintenance processes [
4,
5]. Hence, I4.0 has become a required framework for Structural Health Monitoring (SHM), which has been used to ensure the longevity and safety of critical infrastructure by providing high-quality information for decision-making in risk management.
SHM is a well-known concept and practice [
6,
7,
8] and its main techniques that have been used are shown in
Figure 1. It is useful not only to prevent damage but also to extend the lifespan of elements, avoiding costs in maintenance, repair, and grounded or inoperable structures. The concept of using machine assistance for monitoring will lead to the next step of implementation of this methodology, with the continuous monitoring of elements.
SHM has become a topic of interest due to the large number of benefits it may provide to several industries, including reliability improvement and real-time damage detection, leading to lighter structures and maintenance cost reductions [
9]. By implementing a real-time damage detection strategy, safety factors can be reduced under the damage-tolerant philosophy, especially for structures with high uncertainty in their failure modes (e.g., structures made of composite materials) [
10].
One crucial aspect of SHM is the generation of vast amounts of data from sensors embedded in structures [
11]. Therefore, advanced ML and AI algorithms can help process and analyze these data to extract valuable insights, patterns, and trends that may not be apparent through traditional methods [
12]. This analytical capability enables the identification of subtle changes in structural behavior over time, providing early warnings of potential issues.
In this regard, ML and AI techniques are particularly valuable in damage detection and classification [
13]. By training models on labeled datasets, the system can learn to recognize specific damage signatures and distinguish them from normal operating conditions. This ability to detect and classify various types of damage, such as cracks, corrosion, or material degradation, contributes to effective and targeted maintenance efforts [
14], where a human could take time with the large amount of data to process or be subject to human error caused by fatigue or inexperience.
In parallel with the aforementioned intelligent processing techniques, sensor technologies have also advanced considerably during the last decades. For instance, with the advent of Fiber Optic Sensors (FOS), several SHM methodologies based on these sensors have been proposed in the scientific literature, being a suitable option for SHM of aerospace vehicles due to their small size, light weight, electromagnetic immunity, and multiplexing capability [
15]. Such FOS are usually classified into local or point sensors (e.g., extrinsic Fabry–Perot interferometer), multi-point sensors (e.g., fiber Bragg Gratings) and distributed sensors (e.g., Rayleigh and Brillouin distributed sensors) [
16].
The future of SHM within I4.0 is indeed bright. As these technologies continue to evolve and become more sophisticated, we can expect even more innovative applications that will shape the way we monitor, maintain, and ultimately safeguard our infrastructure for generations to come [
17]. This powerful combination has the potential to create a world where infrastructure is not just static and reactive, but intelligent, adaptable, and capable of anticipating and responding to the ever-changing demands of the modern world [
18].
SHM is set to revolutionize intelligent infrastructure development and management by transitioning from simple data collection to predictive maintenance systems [
19]. By leveraging vast amounts of data collected from a dense network of sensors, SHM systems can utilize advanced analytics to predict potential failures. ML algorithms play a key role in this process, analyzing data to identify patterns and anomalies that may indicate emerging problems [
20]. This predictive approach enables targeted interventions before issues escalate, saving time, money, and lives.
Modern sensors, combined with I4.0 technologies such as the IoT, enhance SHM systems by creating interconnected networks within a broader industrial ecosystem [
21]. These systems enable real-time data exchange, allowing for a coordinated response to potential infrastructure threats. For example, traffic management systems can reroute vehicles to prevent overloading a bridge [
22], while engineers receive notifications to plan necessary repairs. This real-time communication minimizes downtime, safeguards public safety, and supports proactive maintenance [
23].
By harnessing the power of I4.0, SHM promotes improved safety, optimized resource allocation, and the longevity and resilience of infrastructures. In alignment with SDG 9’s targets, this paper addresses a successful case focused on a novel SHM methodology for detecting and locating damages in metallic aircraft structures, employing dimensional reduction techniques. The proposed methodology aims to identify subtle changes in local strain distribution indicative of damage. The study explores the influence of various factors on damage detection, including sensor placement, noise levels, and damage size and type. The proposed methodology for detecting cracks and holes as small as 2 mm in length showcases the potential for early damage identification and targeted interventions in diverse sectors such as aerospace, civil engineering, and manufacturing.
This paper is organized as follows.
Section 2 addresses the important relation between SHM and I4.0.
Section 3 describes the ML methods used in this study, the virtual testing setup, and the damage detection methodology. Then,
Section 4 contains the results and the discussion, and finally
Section 5 includes the conclusions.
3. Materials and Methods
With the evolution of computer science, advancements in hardware, and the development of new software that facilitates the rapid implementation of ML algorithms, the need for constant human monitoring has decreased. Instead, tools from I4.0 and AI are increasingly used to prevent structural damage. These tools are based on fundamental ML algorithms and schemes [
81], as illustrated in
Figure 3. Combining these data-processing methods with various SHM monitoring systems represents the next step in maintenance for several engineering domains.
3.1. Descriptive Methods
Descriptive methods in ML focus on analyzing and summarizing data to understand its patterns and characteristics without making predictions [
82]. These methods aim to reveal the inherent structure of the data through statistical measures, visualization tools, and clustering algorithms, which help uncover trends, distributions, and relationships within the dataset. Descriptive techniques provide a comprehensive overview, serving as a crucial initial step in the ML process. They guide subsequent modeling decisions and establish a solid foundation for further analysis. Often referred to as unsupervised methods, they do not require labeled datasets and can identify patterns in the data. The dataset and model will have the following characteristics:
Recent historical data are crucial for this analysis as it forms the foundation of the model. The model aims to identify trends based on recent behavior or events. If the dataset includes outdated historical data, the model may struggle to accurately simulate and classify current events. Therefore, it is essential to purge older data and retain only recent information that reflects current trends.
Nonexistence of an objective variable: As previously mentioned, these models focus on understanding the inherent structure of the data rather than making predictions. The model aims to describe phenomena effectively rather than optimizing or predicting a specific numerical value.
Clustering, exemplified by the k-means technique, is an ML method that groups similar data points based on shared characteristics [
83]. k-means iteratively partitions the dataset into k clusters, each represented by its centroid. This technique is widely used for segmentation and pattern recognition, helping to identify distinct subgroups within a larger dataset.
Factor selection involves choosing relevant variables to improve the performance of the model. Regression trees use a tree-like structure to recursively split data based on variables, identifying key predictors for the target variable. Conversely, Principal Component Analysis (PCA) is a dimensionality reduction technique that transforms the original variables into a smaller set of uncorrelated components [
84]. Both regression trees and PCA assist in factor selection by highlighting the most influential variables, leading to more efficient and interpretable models.
3.2. Predictive Methods
Predictive methods in ML involve using algorithms to analyze historical data and identify patterns, which allows the model to make predictions or classifications for new, unseen data [
85]. These methods rely on mathematical models to learn relationships between input features and output labels, adjusting their parameters through training. Common predictive techniques include regression for continuous results and classification for discrete categories. The goal is to develop a model that generalizes well to new data, providing accurate predictions and insights based on learned patterns [
86]. Often referred to as supervised methods, these techniques require a labeled dataset to train the model and recognize results within a defined set of options. The dataset and model will have the following characteristics:
Historical Data: the dataset may include older historical data compared to descriptive methods, while the data describe the same event, and the age of the data—whether from 5 years ago or 20 years ago—does not matter as long as the dataset is labeled with the expected results.
Existence of an Objective Variable: The model must predict a specific result, which can be numerical or categorical. If categorical, it is typically represented by a binary dummy variable.
Relation Between Predictive and Result Variables: The dataset needs to be cleaned and analyzed to select variables that have a direct relationship with the objective variable. This requirement distinguishes predictive methods from descriptive methods, which do not need this level of variable selection.
Classification involves assigning predefined labels to input data based on their features [
87]. Neural networks, with their interconnected nodes organized in layers, are effective at learning complex patterns and relationships for accurate classification. Decision trees recursively split the dataset based on feature conditions to create a tree-like structure, facilitating efficient classification. Support Vector Machines (SVMs) are skilled at finding optimal hyperplanes to separate distinct classes in high-dimensional space, making them versatile for both linear and non-linear classification tasks. K-nearest neighbors (KNN) classifies data points by evaluating the majority class among their k-nearest neighbors, providing a straightforward yet effective approach. These diverse classification techniques address various data structures and problem complexities, demonstrating the versatility of ML in solving real-world classification challenges across different domains.
Regression, on the other hand, involves predicting continuous values based on input features [
88]. Neural networks excel in capturing complex patterns for accurate predictions. Time series regression models are designed for temporal data, capturing trends over time. An SVM for regression identifies optimal hyperplanes to model non-linear relationships in the data. These techniques, including neural networks, time series models, and SVMs, provide adaptable solutions for predicting numerical values across various domains.
3.3. Virtual Testing Setup
To perform a virtual testing methodology within the context of SHM, the wing structure of an MQ1 Predator aircraft [
89], made of aluminum alloy 2024-T3, was chosen for this study. The structure comprises two beams, 18 ribs, and skins [
90]. Rivets and other structural joints were replaced with perfect contacts between components, and fuel tanks and other systems were omitted to reduce computational cost.
Figure 4 illustrates the drawing views of the wing structure.
The Static Structural Simulation solver in ANSYS 2020 R1 was employed, with non-linear effects being neglected. The boundary conditions for the model included fixed supports for each beam at the root and a distributed load applied to the lower skin, representing the total lift generated by the aircraft during steady cruise flight. This load was calculated as 3753 N per semi-wing, based on the aircraft’s weight of 766 kg.
The model’s mesh used hexahedral elements with a multi-zone method applied to the beams, and a finer mesh refinement was applied to the main beam, which was the primary focus of this study. The damage was positioned at various locations (25%, 50%, and 75% of the wingspan), as shown in
Figure 5. This approach helped minimize significant mesh changes when damage was introduced, as depicted in
Figure 6. Additionally, beams, ribs, and skins were modeled with shell elements to reduce overall model complexity, resulting in a total of 50,144 elements with a minimum element quality of 0.22 (where a quality index of 1 indicates maximum reliability).
The primary component studied in this structure was the leading edge beam, also referred to as the main beam, of the aircraft wing. A total of 408 virtual strain sensors were defined along this beam: 102 sensors were placed on the top face, 102 on the lower face, 102 on the front face, and 102 on the back face. These virtual sensors were configured to measure the normal strains along the longitudinal axis of the beam, emulating the functionality of FBG sensors; see
Figure 7.
Mesh convergence was assessed by comparing the degrees of freedom (DOF) in the mesh with the maximum deflection of the wing, as shown in
Figure 8. The analysis indicated a mesh convergence around 440,000 DOF. This configuration was achieved with a general element size of 15 mm in the geometry and a special refinement of 5 mm in areas near the damage. However, given the need to analyze the strain field around very small damages relative to the overall structure size, the maximum number of elements possible within the available computational capacity was utilized. This involved applying a special mesh refinement specifically around the damaged areas to ensure accurate analysis.
Two types of damage were analyzed: a circular hole and a high-aspect ratio rhombus, simulating a crack. For each damage type, three different sizes were defined: 20 mm, 10 mm, and 2 mm in diameter for the holes, with equivalent lengths for the rhombus’s largest diagonal (maintaining an aspect ratio of 10). The damages were positioned at three different locations along the wingspan: 25%, 50%, and 75% (
Figure 5).
Table 1 provides the details of the damage locations and sizes.
To emulate a real sensing scheme using FBGs, different experimental trials were artificially defined. The trials were based on the strains measured for each sensor and replicated I times using a random number generated from a Gaussian distribution, with a mean equal to the initial strain value and a standard deviation of . This deviation was chosen considering that the sensitivity of the available FBG sensing techniques in the market is approximately . Thus, the dimensions for each data matrix were 1000 × 600 for baseline , 700 × 600 for validation data , and 900 × 600 for damages ( to ).
Several sensitivity analyses were performed, including an examination of how changes in the F1 Score were influenced by the number of operative sensors. The F1 Score is a measure of a model’s accuracy in classification tasks [
91], especially when dealing with imbalanced datasets. By running the detection algorithm multiple times and artificially removing some sensors each time, the change in the F1 Score in response to variations in artificial noise was evaluated.
3.4. Validation model
To validate the model, an experiment involving a double-supported steel beam was replicated using cross-validation with ANSYS 2020 R1 [
77]. In this model, two supports were placed at both ends of the beam, one fixed and one with displacement, with two vertical loads applied at the midpoint, as shown in
Figure 9. The entire geometry was meshed with 2 mm shell elements. The model convergence was evaluated on the basis of the maximum bending moment. Various loads were applied and compared to the minimum safety factor associated with each load. Finally, a linear regression analysis was performed to determine the load corresponding to a safety factor of one.
The maximum calculated moment corresponding to a safety factor of one was 15.032 kN·m, as shown in
Figure 10. In comparison, the maximum moment reported by [
77] was 15.227 kN·m, resulting in an absolute error of 1.82%.
3.5. Damage Detection Methodology
For the experiments presented in this work, which involve simulations of the damaged structure with cracks and holes and their respective variations, each dataset was organized into a matrix , where I represents the trial number and J represents the sensor number. As mentioned above, the main objective of PCA is to reduce J to several principal components while retaining a significant portion of total variability.
Recently, other dimensionality reduction techniques such as UMAP (Uniform Manifold Approximation and Projection) and t-SNE (t-Distributed Stochastic Neighbor Embedding) have emerged in the field of SHM, offering advantages and limitations compared to PCA [
92]. A significant advantage of these techniques is their ability to capture highly nonlinear relationships, unlike traditional PCA, which only captures linear relationships. However, previous work by the authors demonstrated that for metallic structures and even some made of composite materials, the structural behavior is predominantly linear. Therefore, the use of nonlinear dimensionality reduction and modeling techniques does not significantly enhance detection sensitivity relative to sensor density, while the computational cost increases substantially, especially for large datasets. In the context of damage detection based on strain measurements, traditional PCA has shown a superior “quality-price” ratio compared to other nonlinear dimensionality reduction techniques in predominantly linear structures [
93].
For this study, 90% of the variability was retained by selecting three principal components, as suggested by Jolliffe [
94]. To achieve this, the matrix
, which represents the covariance eigenvalues of
A, was calculated for the baseline (BL) dataset. The remaining datasets, corresponding to the damage conditions, were then projected using the
P matrix obtained from the BL data. The principal component matrix
T is given by
where the first three columns correspond to the three first principal components of the new dataset [
95].
Then, to calculate the
Q statistic, the residual error matrix
E is computed as follows:
where
is given by
With these values, the
Q statistic array is given by
where
denotes the
ith row of
E. Then,
is calculated as follows:
where
is the
ith row of matrix
A, and
is a diagonal matrix composed of the eigenvalues of
A.
To determine if there is a damage or not, a damage threshold is defined based on BL’s
inverse distribution. To calculate this threshold
, mean
and variance
of
Q and
for BL dataset is computed as follows [
96]:
where
Given that
represents the squared inverse statistical distribution for a specified confidence level
; previous experiments reported by Sierra-Pérez et al. [
97] indicated that using
values between 95% and 99% yields acceptable results. It is important to note that
D0 corresponds to a validation dataset (pristine state); therefore, the statistics for
D0 should fall below the threshold, while the data for damaged conditions should be above the threshold.
According to the threshold, each datum is classified as true when it exceeds the threshold and false when it falls below it, with true data considered as having detected damage. With this in mind, ROC analysis was performed, classifying data into true positive (
), true negative (
), false positive (
), and false negative (
) categories. Finally, the F1 Score, which measures the experiment’s accuracy, can be computed as follows:
This score, which ranges between 0 and 1, measures the accuracy of data classification. In the context of SHM methodology, a higher F1 Score indicates more successful damage detection.
3.6. Damage Localization Methodology
Once the
Q and
statistics have been calculated and a dataset with successfully detected damage is obtained, a contribution analysis is performed. This analysis assigns a numerical value to each column (or data measured by each sensor) of the original dataset, quantifying the influence of the data of each sensor on the increase in statistics
Q and
. In the context of SHM, damage is identified in sensors (or rows of the original dataset) with higher contribution magnitudes. The contribution analysis for the
Q statistic is calculated as follows:
where
is
row of the identity matrix and
denotes the contribution of each row based on the
Q statistic. Similarly, the contribution analysis for the
statistic is given by
Once the contribution values for each sensor have been obtained, the damage location is estimated based on the sensors with the highest contribution.
4. Results and Discussion
After applying the previously proposed methodology, the F1 Score based on the
Q statistic was calculated for each damage type and location.
Table 2 and
Table 3 present the obtained F1 Score values for hole crack and crack damages, respectively. Additionally,
Figure 11 illustrates the distribution of the
Q statistic for each measurement and its location relative to the defined threshold.
The results show the F1 Scores for detecting hole and crack damages at various positions and sizes. For hole damages (
Table 2), the F1 Scores indicate high accuracy for 10 mm and 20 mm sizes across all positions, particularly at 25% and 50%, with scores near 0.9852. However, the detection accuracy drops significantly for 2 mm holes, especially at the 50% position (0.0976). For crack damages (
Table 3), the F1 Scores are similarly high for 10 mm and 20 mm sizes, with the 25% position reaching nearly perfect scores of 0.9936. The scores decrease for the 75% position, particularly for smaller crack sizes, where the 2 mm crack has an F1 Score of only 0.1584. These results indicate that larger damages and certain positions are more reliably detected.
According to these results, the location of the damage appears to influence the detection capability. For both types of damage, a decrease in the F1 Score can be observed when the damage is located in the central position of the wing (i.e., approximately 50% of the wingspan), especially for smaller damages. This could be attributed to the distribution of stresses and strains in the wing structure, where the central region might experience less variation compared to areas near the root and tip, making it harder to identify subtle damages in these areas.
Furthermore, comparing the results between the two damage types, there is a slight advantage in detecting holes compared to cracks, particularly for smaller damages. This suggests that damage geometry can influence the sensitivity of the methodology. Holes create an abrupt discontinuity in the structure and therefore might produce more distinctive strain patterns than cracks, which tend to propagate more slowly and subtly. Consequently, the proposed methodology might require specific adjustments or improvements for the early detection of cracks, especially in critical locations of structures.
Once damage was identified, a contribution analysis was performed to determine the most contributing sensors. The results of this analysis are detailed in
Table 4 and
Table 5.
In most cases, the contributing sensor was located on the lower face of the beam and close to the actual damage. It should be noted that the lower face typically experiences maximum tensile stresses and strains. Additionally, there are instances of undetected damages where the most contributing sensor is affected by system noise rather than damage locations. This occurs because the contribution analysis is only valid when damage is detected; thus, several contribution results are rejected if they do not correspond to actual damage.
Figure 12 and
Figure 13 present representative heat maps showing the contribution levels around the beam for different types of damage. In these maps, the damage location is clearly identified when the F1 Score is high. Conversely, when the F1 Score is lower, the highest contribution is not concentrated in a single, clearly marked spot.
In the sensitivity analysis based on the number of operative sensors, one of the largest damages (20 mm length at 25% of the wingspan) was selected. Previous calculations were performed by iteratively removing one sensor at a time nearest to the damage. The trends observed, shown in
Figure 14, indicate that up to 15 sensors closest to the damage can be inoperative, and the damage can still be detected normally for both types of damage (hole and crack).
For the sensitivity analysis based on noise, a previously detected damage (10 mm length at 75% of the wingspan) was selected. The F1 Score was calculated for varying noise levels, ranging from
to
. As shown in
Figure 15, the damage can be detected in both cases when the noise level is
. However, detection becomes unreliable at higher noise levels.
According to
Table 4 and
Table 5, the damages with the lowest performance of the technique correspond to
for the hole and
for the crack. These damages were analyzed at different noise levels to determine an adequate level for achieving acceptable F1 Score results. As shown in
Figure 16, it is possible to detect smaller damages as long as the noise level is sufficiently low. Compared to the minimum noise in previous analyses, this noise decreases to levels around
.
The proposed damage detection and localization methodology is aligned with the advances highlighted in the state-of-the-art and the targets proposed in SDG 9. Our approach leverages the potential of SHM systems to transition from mere data collection to predictive maintenance by employing data analytics, including PCA and contribution analysis, to identify and localize structural damage. This methodology not only detects damage but also identifies its location with high accuracy, thereby confirming the system’s effectiveness in a real-world context. Furthermore, our sensitivity analysis underlines the robustness of the proposed technique, showing that even with a significant number of inoperative sensors or the presence of noise, the system can still reliably detect damages. This capability is crucial in an I4.0 context, where interconnected systems and real-time data processing are essential for timely interventions in aerospace systems. The integration of such advanced SHM systems into broader industrial ecosystems, as discussed, underscores the importance of this work for enhanced infrastructure safety and management through predictive and proactive maintenance strategies.
The proposed damage detection and localization methodology, while effective, has some limitations. It relies on sensor placement quality and density, and is sensitive to noise in the data, which can affect detection accuracy. The computational demands are significant, potentially limiting real-time applications. Additionally, the methodology’s effectiveness was tested on specific damage types and sizes, which may not generalize to all structural configurations. Further refinement and validation in diverse settings are needed to ensure consistent performance. Another limitation of this study is the reduced sensitivity of the methodology in detecting cracks compared to holes, particularly for smaller damages. This suggests that the geometry of the damage can affect detection accuracy, indicating a need for adjustments to improve the early detection of cracks in critical structural locations.
5. Conclusions
SHM is poised for a revolutionary transformation as it integrates with the core principles of I4.0. This paper has explored various areas where this synergy will have the most significant impact, fundamentally reshaping our approach to infrastructure management in diverse sectors.
Big data analytics and ML algorithms are set to unlock the hidden potential within the vast amount of SHM data. These data will no longer merely serve as a record of the past; it will become a powerful tool for predicting future issues. This predictive capability will extend the lifespan of structures in various domains, from the delicate wings of aircraft to the buildings that support cities. Early detection of anomalies will enable targeted interventions before problems escalate, leading to significant cost savings and improved safety across all fields.
Real-time communication will enable proactive responses to potential issues before they snowball into major disruptions. Data acquisition is instantly transmitted to a central hub, triggering a cascade of automated responses within interconnected ecosystems, optimizing efficiency, safety, and cost-effectiveness in unprecedented ways.
The sensitivity of FBGs is a crucial factor in detecting small damage, as it relates to the strain field generated by the damage itself; if the sensitivity is not high enough, the damage will not be sensed. However, the strain field is associated with stress concentrations, which depend on geometry and size. Thus, the damage type is also important, as each type has a different geometry and therefore a different stress concentration.
In this study, several sensitivity analyses were performed, concluding that sensor noise and the number of sensors are key factors for achieving adequate accuracy in this methodology. This study is based on currently available FBGs; however, the sensitivity analysis results presented in this paper show that improving FBG performance can enhance SHM accuracy, potentially surpassing traditional NDT techniques.
Furthermore, the placement of FBGs plays a crucial role in damage detection; they must be located at points where strains are highest and near potential damage sites. Otherwise, they will only sense noise or very low-strain fields. As a result, some sensors near the damage may not contribute significantly to statistical indexes, even though they are close to the damage.
The main advantage of the methodology presented in this paper, compared to other methodologies previously developed by the authors, is that it enables damage detection and localization with good precision using sparse sensor networks. Unlike other techniques that require highly dense sensor networks, this approach is more efficient and cost-effective. Additionally, the developed methodology exhibits greater sensitivity, allowing the detection of even smaller damages than those detectable by previously developed methods. This results in a better relationship between damage size and sensor density, enhancing the overall effectiveness and applicability of the SHM system [
97,
98].
The proposed methodology is promising but also has some limitations. It is highly dependent on sensor placement quality and density, and sensitive to data noise, which can impact detection accuracy. The computational demands are substantial, which may hinder real-time applications. The methodology was tested on specific damage types and sizes, which may not be representative of all structural configurations. Additionally, the methodology demonstrated reduced sensitivity in detecting cracks compared to holes, particularly for smaller damages, highlighting the need for further adjustments and validation in diverse settings to improve accuracy and performance.
The future of SHM within I4.0 is promising. As these technologies continue to evolve and become more sophisticated, they will give rise to innovative applications that will reshape how we monitor, maintain, and ultimately safeguard our infrastructure. This powerful combination has the potential to create a world where infrastructure is not just static and dynamic, but also intelligent, adaptable, and capable of anticipating and responding to the changing demands of the modern world.