A novel framework for effective structural vulnerability assessment of tubular structures using machine learning algorithms (GA and ANN) for hybrid simulations

3.1 3D nonlinear modeling

Full-scale nonlinear 3D model is developed in CSI PERFORM-3D. This PERFORM-3D serves as a generic software package that can substantially implement the displacement-based assessments, without any specific limitation set by ASCE 41 [34], and other pertinent codes that cover the subjects of seismic assessments and retrofits.

PERFORM-3D comprises splendid tools to model nonlinear structural response using plastic hinges and fiber sections. It allows for the direct transformation of material nonlinearity to the strains in structural elements. In 3D analytical model developed in PERFORM-3D, fiber sections were used throughout the height of the structure to model and monitor strains in the shear wall. Categorically distinctive fiber sections were produced for reinforcing steel and concrete. Development of fiber sections enabled the incorporation of confinement effect in concrete, and hysteretic behavior of constituent materials could adequately be encapsulated. For analytical modeling of frame components, the nonlinearity was captured by developing plastic hinges in accordance with the specifications provided by ASCE Standard [34]. However, the slab elements were kept as linear elastic. Figure 3 shows the isometric view of developed 3D analytical model in PERFORM-3D, originally produced by Zain et al. [33].

Figure 3

Isometric view of 3D nonlinear analytical model, developed in CSI Perform 3D.

3.2 Lumped-parameter modeling using genetic algorithm

The actual structure in this work is converted into lumped parameter models using GA. A full 3D nonlinear high-rise tubular structure can exert a substantial computational load on the overall process of vulnerability assessment using IDA, which may not be practically possible, especially when the large number of ground motions are used. Thus, the presented work employs lumped modeling, discretized into two stages to develop a structural model that could efficiently predict the structural dynamic response against ground motions. The lumped-parameter modeling has been done in two phases. The first phase targeted the elimination of outer frame by inserting nonlinear spring elements at the intersection of wall and frame, i.e., wall–frame connection, at every floor level by employing the genetic algorithms to establish the optimized values for nonlinear springs. While in the second phase, the genetic algorithm was used to produce the optimized lumped elements to model the shear wall. For the first phase, ZEUS-NL has been used as the primary tool, while the second phase is implemented through 2D continuum analyses software package. Initially, an objective function was established that aimed to optimize the structural stress and stiffness of springs in relation with the original structural integrity as predicted by the 3D nonlinear model. Afterward, configuration of springs, representing outer frame, were encoded as chromosomes, characterizing the stiffness values of springs based on the established objective function. The fitness scores were then evaluated to check for the desirable configuration of chromosomes. GA involves several genetic operations including mutation and crossovers that permit the selection of chromosomes with higher fitness, thus introducing genetic diversity to produce new generation of solutions to process the structural optimization. With the subsequent iterations of genetic operations to produce offsprings by fittest chromosomes, optimal configurations were established, and the algorithm was terminated when the difference in the stiffness levels of springs in comparison with actual stiffness of structural members was less than 10%. The optimal configuration established by GA was then implemented in the simplified structural modeling of outer frames and core.

The further discreet details about each respective phase have been discussed under the proceeding subsections.

3.2.1 Phase I: Structural system simplification using GA

At lower floor levels, shear walls and core play obtrusive roles in high-rise buildings to counter seismic loadings. On the other hand, the frame of the tubular structure primarily works against the gravity loadings, while supplementing the lateral strength through structural interaction with core or shear walls. Thus, considering the conventional structural behavior of tubular structures, it is suitable to utilize nonlinear spring elements for representing the three directional reaction forces, including translational forces and moments, of frame elements. The schematic representation of the springs is shown in Figure 4. For analyzing the structure, opensource ZEUS-NL [35], developed by MAE Center and National Science Foundation, USA, has been used. ZEUS-NL has been discreetly developed for dynamic analysis against ground motions and is capable of incorporating rotational and axial springs, possessing nonlinear character. Furthermore, by the virtue of fiber-based modeling approach, ZEUS-NL allows for the nonlinear static and dynamic analysis of simplified models.

Figure 4

Equivalent nonlinear springs at wall joint.

For selecting the parameters, GA has been used. The readers of this work are referred to the work of Goldberg [36] for a detailed theoretical background of GA. Initially, the primary values of each spring, i.e., stiffness, displacement, and rotations, have been defined at each floor level in the analytical software. Subsequently, the fitness functions and population have been defined to run the trials until a discreet model was identified that could adequately predict the behavior in comparison with the 3D nonlinear analytical model.

With the increase in height, the stiffness value decreases, and therefore, only selected spring parameters, i.e., S _j , were optimized. The considered building comprises first four floors as podium, while rest of the floors, i.e., from fifth story and above, are part of the tower structure, and thus, the stiffness disparities are different for podium and tower floors, and expression for S _j changes accordingly. Thus, the population’s protocol parameter vector is expressed as follows:

(1) ( S j ) T = ( S 1 ) T . i a T , T = 1 , 2 , 3 , … for i ≤ 4 ( S 1 ) T . 4 a T . ( i / 4 ) b T , T = 1 , 2 , 3 , … for i > 4 .

In Eq. (1), i represents the floor number, i.e., 1–55, and for each trial solution in GA, the parameter vector (T) of each joint is evaluated. The fitness function in the GA is defined as follows:

(2) f ( ∆ , i ) = ∆ i l − ∆ i 3 D i × ∆ i 3 D .

In Eq. (2), the fitness function, ∆ i l , represents the displacements of nodes at the i th story, obtained from the simplified model, while ∆ i 3 D represents the nodal displacements at the corresponding floor level, evaluated from the 3D nonlinear model. The development number for GA generations was kept at 100, and for evolution, crossover and mutation techniques were employed. In the presented study, crossover was used as the primary genetic operator that produced 80% of the population, while rest were produced using mutation technique.

Crossover amalgamates parent genetic material from two individuals to develop offspring. It is based on the concept of exchanging genetic material between chromosomes. For instance, two individuals with chromosomes as 010100110 and 011011010 can crossover in a random fashion at any appropriate position, i.e., 3rd, 4th, etc. i.e., 0101|00110 and 0110|11010 to reproduce the offspring 0101|11010 and 0110|00110. The produced offspring would now be included in the exiting population to produce subsequent offspring, and the operation/production remains continuous. Mutation, on the other hand, introduces new genetic material into the population through random alteration of genes in an individual’s chromosome. The current study utilizes the random mutation technique that randomly changes a structural property to a random value to be used in the analysis at ZEUS-NL. Figure 5 illustrates a schematic of the GA, using crossover and mutation technique.

Figure 5

Schematics of employed GA technique.

The established simplified model in ZEUS-NL is exhibited in Figure 6. The transformation of the 3D analytical model to a simplified model in ZEUS-NL was developed through the GA parametric study. The gray lines indicate the frame components that were replaced by springs. A comparison of modal analysis has been made between the two models, i.e., CSI Perform 3D model and simplified ZEUS NL model, and respective results are presented in a tabular form along with the percentage of errors observed in the values. It is pertinent to state that the observed error percentage is quite low (Table 2).

Figure 6

Transformation from CSI PERFORM-3D to ZEUS-NL.

Table 2

Modal properties comparison between original and GA-based simplified ZEUS-NL

Mode no.	Time period		% Error
	Original model	Zeus-NL model
1	4.67	4.725	1.2
2	1.12	1.10	−1.82
3	0.51	0.474	−7.59
4	0.30	0.26	−15.38

The comparison of modal analysis depicts that cumulative error is approximately 10% for the first three substantial modes, comprising around 90% of the modal mass participation ratio. Nonlinear static pushover curves were developed for 3D nonlinear analytical model and simplified lumped model to compare their respective responses. The comparison revealed the adequacy of the established simplified model as evident in Figure 7, which shows the pushover analysis (POA) curves obtained from the full nonlinear model and simplified model with springs. The disparity of results from the ZEUS-NL model, in comparison with the results from 3D model, is primarily attributed to the increased stiffness of the ZEUS-NL model due to nonnegative tangent stiffness of spring joints. However, the modal responses from both models are significantly on a par with each other with a maximum difference of 15.38% only.

Figure 7

Comparison of nonlinear static analysis results at global level.

Dynamic response history analysis was executed on both models to establish the effectiveness of the ZEUS-NL model. The same ground motion was applied to both models for attaining their global deformation history, and a graphical comparison had been developed accordingly. Figure 8 shows the global displacement histories of both models against the same ground motion. The obtained displacement history from simplified model was observed to be very close to the actual response of the 3D nonlinear model. The maximum global drifts from the actual structure and established GA-optimized model were observed to be 1.721 and 1.826, respectively, exhibiting an overall error of 6.4% only, substantiating the adequacy of GA optimization. Thus, the developed simplified model enabled an effective and adequate assessment of structural dynamic response and efficiently imitated the results gained from full nonlinear model.

Figure 8

Comparison of global time histories from actual model and GA-based simplified model.

3.2.2 Phase II: core-wall system simplification using GA

In the presented study, the core wall system has also been simplified as merely employing the simplistic frame model will not be adequate for modeling the wholesome structural performance, as lumped model would not be able to incorporate the localized effects such as the rebar yielding or crack propagation within the structural walls. For that purpose, commercially available LS-DYNA package has been utilized to develop a microscopic model of shear wall, and composite layered shell elements, based on Belytschko and Tsay [37]. Consequently, the frame centerline representations have been mixed with the 2D wall panels. The elements have been characterized into seven layers comprising four layers for reinforcements, i.e., two layers for each direction, on distinctive layer for the concrete, and two discrete layers for the covers. A fixed smear crack approach for the constitutive material relationship has been used. From several concrete formulation models, this study utilizes the curves established by the study by Mander et al. [38]. Following the same procedure as described in the preceding subsection, the parametric study was conducted by utilizing the crossover and mutation techniques for GA-based structural optimization, and respective values were used in the ZEUS-NL model. It is noted that central nodal displacements of wall panels had essentially been required for every floor to establish overall structural fragility information. Therefore, the fitness function was defined as follows:

(3) f ( ∆ w ) = ∆ w l − ∆ w d ∆ w d .

In Eq. (3), ∆ w l represents the central nodal displacements of the lumped model in ZEUS-NL, while ∆ w d denotes the LS-DYNA’s nodal displacements. By the virtue of GA-based optimizations, it was amply possible to simplify the full nonlinear analytical model into a lumped-parameter model, which projected reasonable accurate results. These two distinctive models were combined using the UI-SIMCOR [39] that possesses the ability to execute the analysis involving concurrent components, while producing a constitutive network for amalgamating action–deformation characteristics to develop a collective structural response. Figure 9 presents the schematics of the whole process, incorporating the hybrid simulation involving ZEUS-NL, LS-DYNA, and UI-SIMCOR.

Figure 9

Hybrid simulation schematics for reference building in UI-SIMCOR.

The developed lumped model has adequately predicted the nodal displacements and forces in the structure while taking into account the convoluted frame–wall interaction, and simultaneously reduced the computational effort for executing a dynamic analysis [40]. Thus, dynamic analysis was conducted in the ZEUS-NL environment on the simplified lumped with the reduced computational effort.

4.1 Uncertainty considerations

Uncertainties are conventionally categorized into two major categories, i.e., epistemic and aleatory [41]. The epistemic uncertainties characterize the ambiguities and precariousness related to the scantiness of the available knowledge. In a typical sense of infrastructural vulnerability process, epistemic uncertainties portray the variability involved in the material characteristics, construction process, and other pertinent information for which the experimental evidences exist and are still implementing for subsequent performance improvements. On the other hand, aleatory uncertainties portray the unpredictability involved in natural processes that are beyond the control of humans. Thus, in the sense of vulnerability assessment, aleatory uncertainties characterize the intrinsic variability involved in the properties of seismic ground motions, as each ground motion is statistically independent of the other in terms of their frequency contents, peak ground accelerations (PGAs), and other pertinent properties. The quantification of aleatory uncertainties can be established using probabilistic seismic hazard analysis, but since that is not the scope of the presented study, stochastic selection of actual ground motions was made considering various geological and geo-seismic settings.

Taking into account the epistemic uncertainties, the material strengths can be taken as random variables, and a statistical technique, i.e., Monte Carlo simulation, can be used for producing a substantial number of randomly selected pairs of constitutive materials’ strengths to use in the analysis; however, the literature indicates that even after the consideration of material variability and other pertinent factors under epistemic uncertainties, the analysis results do not change substantially in comparison with those developed without considering the epistemic uncertainties. Celik and Ellingwood [42] conducted several research studies for evaluating the discreet effects of considering both types of uncertainties for fragility assessment processes and concluded that the fragility curves developed with and without considering the epistemic uncertainties did not differ to a considerable extent. However, the aleatory uncertainties, i.e., the ground motions’ inherent characteristic variations played a vital role in inducing changes in the fragility results, and a similar inference was drawn by the study by Zain et al. [33]. Considering the available literature, the current study discreetly considers the aleatory uncertainties, i.e., ground motions record-to-record connate variability, as the sole source of uncertainty into consideration for developing the fragility curves for the case study at hand. Since ground motions serve as the major source of uncertainty, it is vital to consider various factors, i.e., attenuations, source-to-site distances, and producing source features, while selecting the ground motions to adequately cover the possible spectrum of variations in overall seismic energy content and respective structural responses. This work considers 15 meticulously selected ground motions that were incrementally scaled and applied in the ZEUS-NL environment to obtain deformation-based response and for the subsequent development of fragility curves.

For selecting the ground motions, criteria established in the study by Ji et al. [43] have been employed, and resultantly, source-to-site-distance (S-T-S distance), respective magnitudes, soil conditions, and varying source mechanisms were considered herein. Earthquakes having magnitude between 4.5 and 8.0 only have been considered, and source-to-site distance has been kept between 1 km and 90 km. The selected ground motions are presented in Table 3 that presents the recorded PGA of each ground motion. For selecting the ground motion records, PEER ground motion database, NGA-West2 [44], was used. The database contains thousands of natural ground motions and allows the selection as per the set criterion along with the options for selection based on the target spectrum.

Since high-rise buildings can experience the higher structural modes, the selected ground motions with varying inherent characteristics are intended to cover both low-frequency and high-frequency structural responses. For the implementation of IDA, as discussed, the weaker directional frame has been selected. By employing various ground motions and their iterative scaling, the selected records are considered to cover a wider range of structural response, thus covering the associated aleatory uncertainties effectively.

4.2 Incremental dynamic analysis (IDA)

This study employs IDA for nonlinear dynamic assessment. IDA is considered to be one of the most efficacious methods for dynamic analysis, and despite being a computationally highly intensive method, IDA has wide applicability [45,46]. During implementation, the intensities of considered ground motions are incrementally increased for each subsequent iteration of analysis. The ground motions are incrementally scaled to gain a discreet increase of 0.1 g for each next iteration of analysis. For performing IDA, it is necessary to define or select an appropriate IM and corresponding engineering demand parameter (EDP) for the vulnerability assessment process. The following subsections address the selection of IMs and EDPs respectively.

4.2.2 Selection of engineering demand parameter (EDP)

EDP correlates the structural damage with the seismic loading, i.e., IM. In this regard, inter-story drift was considered as the EDP by Soleimani et al. [45] to materialize the characterization of structural damage. FEMA 356 [49] established limits for structural damages using global drifts as the EDPs. Local-level EDPs such as localized stresses, strains, and hinge rotations can also be considered. However, the presented work specifically relates to the global-level fragility assessment, and hence, a global-level EDP shall be a prerequisite. Therefore, the present study employs global deformational response of the considered building as the EDP to correlate with the discreetly seismic intensities.

The results from the IDA, conducted in ZEUS-NL environment, are presented in Figure 10 against all three considered IMs, i.e., PGA, S _a at 0.2 s, and S _a at 1.0 s, respectively. The IDA yielded the variation in structural responses against broader range of seismic intensities for all IMs. The higher intensities of seismic motions in Figure 10 demonstrated the enhanced ductility demands as the global drift reaches to more than 5%, indicating the structural capability to transit into inelastic range under higher seismic demands. Figure 10 further demonstrates the varying influence of intrinsic ground motion features, producing different structural responses against equally scaled ground motion intensities. Thus, distinct structural drift values are observed in Figure 10 at a single intensity level due to the variance in the spectral properties and frequency contents of ground motion histories. IDA results have substantiated that ground motion uncertainty plays the most discernible role as every considered ground motion, scaled at a similar level, induces different results. Conclusively, Figure 10 depicts the significance of incorporating diverse seismic motions in structural analysis against dynamic seismic demands.

Figure 10

IDA results for all considered IMs: (a) IM | PGA; (b) IM | S _a@0.2 s; and (c) IM | S _a@1.0 s.

4.3 Characterization of limit states

Fragility curves serve as the graphical portrayal of probabilities, exceeding specific values of structural limit states through their qualitative and quantitative thresholds. Thus, at this point, it becomes necessary to define the structural limit states (also known as damage states) for establishing the fragility relationships, representing the probabilities for each considered limit state discreetly. The limit states can be defined qualitatively, and their qualitative definitions correspond with the selected EDP for their mathematical threshold limits. For instance, ASCE41-17 [50] has defined three limit states, i.e., immediate occupancy (IO), life safety (LS), and collapse prevention (CP) for classifying the structural limit states and performance objectives and defined corresponding drift values to correlate the structural damage with respective qualitative definitions. Similarly, Chaulagain et al. [51] considered three damage states, i.e., slight, moderate, and complete collapse for buildings in Nepal, and established empirical fragility curves for each limit state, respectively. The present study defines three distinctive limit states, i.e., IO, first limit state (LS1); life safety, second limit state (LS2); and collapse prevention, third limit state (LS3). However, the qualitative and quantitative definitions are different [50]. LS1 characterizes the formation of first plastic hinge in the structure and thus controlled by force-controlled action; LS2 is dependent on the mix of force-controlled and deformation-based action, whereas the LS3 is based on a deformation-controlled action, characterized by 75% of ultimate deformation capability of primary lateral load resisting system at global level, as suggested by Altug [52]. The mathematical threshold values for LS1, LS2, and LS3 are 0.65, 2.4, and 3.1%, respectively, in terms of the established EDP.

4.4 Fragility curves development

Once the results from IDA are available, the constitutive definitions of limit states were applied to develop the fragility relationships for each of the established limit state discreetly. As discussed in the preceding section, fragility curves provide the probability of being in a specific limit state against the discreet level of seismic intensity. Mathematically, Eq. (5) is widely used to develop the fragility curves:

In Eq. (5), P _fragility denotes the probability of being in a specific limit state, against a specific seismic intensity, characterized by IM. Φ shows the standard cumulative distribution function (CDF), while λ c and β c are the controlling parameters of the fragility curve and exhibit the standard deviation and median for a discreet fragility curve, respectively. After application of prescribed limit states on the results of IDA, the probabilities of exceeding each LS were evaluated separately, and eventually the plots were developed. Figure 11 shows the established curves against all considered IMs and limit states. It is pertinent to mention that maximum likelihood method (MLM) was utilized to determine the optimized values of CDF’s controlling parameters.

Figure 11

Seismic fragility curves using GA-based model: (a) IM | PGA; (b) IM | S _a at 0.2 s; and (c) IM | S _a at 1.0 s.

The dynamic analysis time for the GA-optimized model including postprocessing of results was observed to be approximately 2 h on a core i7 computer (8th generation processor) with 8 GB of Ram for 1,000 time-steps. The time required against each IM for solving 1,000 time-steps, the 300 ground motion histories, would have necessitated for 650  (approximately 27 days) for analyses. Thus, with the presented structural optimization methodology, the lumped parameter model consumed substantially lesser analyses time, providing significant reduction in computational burden for developing fragility relationships.

In the subsequent section of this article, an ANN was proposed for establishing the fragility information of considered high-rise tubular structures. Initially, the ANN was developed using the same ground motions, employed for conducting IDA, and afterwards, the fragility relationships were developed accordingly. Subsequently, a comparison was established in the form of fragility curves, obtained through the ZEUS-NL model and the proposed neural network.

5.1 Input and training of ANN model

For the production of training data, an input layer was established containing designated IMs. Some other Ims, i.e., S _d, PGV, and velocity spectrum intensity (VSI), could also be considered by interpreting the response spectra of the ground motions record; however, since the IDA on analytical model has utilized the PGA and S _a as Ims at respective periods, only these two have been used as the inputs in the proposed ANN so that a rational comparison of the fragility relationships, established through analytical model and ANN, could be made. The input layer may contain as much information as required, i.e., the seismic hazard parameters, structures’ geometric and material properties, and possible soil characteristics. Table 4 presents the enlistment of selected Ims for ANN along with their formulae. To establish the number of neurons in the hidden layer, a trial-and-hit methodology was adopted, as there exists no discreet methodology that could be used to decide the number of neurons in the respective hidden layer, and eventually, ADAM was utilized for optimization.

With the trial-and-error method, eight nodes were found to be sufficiently acceptable in the hidden layer of the proposed ANN. Figure 13 illustrates the schematics of the proposed ANN that can be utilized to produce fragility relationships of tubular structures.

For producing the datasets for training purposes, the same ground motions and scaling levels were implemented, and for the activation function, rectified linear unit (ReLU) was utilized. Table 5 shows the hyperparameters for ANN, involving employed activation function, optimizer, and loss function, indicating essential information for reproducibility. In the context of fragility production, employed activation functions should be apposite for regression problems to establish continuous output values.

For the optimization, ADAM has been utilized. ADAM has three primary hyperparameters, i.e., learning rate (α), decay rate for first moment estimates (β ₁), and decay rates for second moment estimates (β ₂). Initially, the gradient of the loss function was estimated; afterward, the second-moment estimate (mean, m) was updated as per the following equation:

Eventually, the uncentered variance vector, v, i.e., the second-moment estimate, was updated using the following equation:

In the subsequent step, corrected first moment, m _c, was estimated as per following equation:

Similarly, the corrected second moment, v _c, was estimated as follows:

Eventually, the requisite parameters were updated in accordance with the following equation:

In Eq. (10), θ denotes the parameters of the neural network, i.e., weights and biases, while gradient is the gradient of the loss function. The initial hyperparametric values of α, β ₁, and β ₂ were 0.001, 0.9, and 0.999, respectively. The values were tuned afterward for mean squared error (MSE). For training purpose, 70% of the data was utilized and 20% data was employed for validation, including the last 10% quartile, specifically for testing using established ANN. Figure 14 shows the correlation between the predicted and actual values against each IM, while Table 6 presents the model prediction evaluations, i.e., correlation coefficient, MSE, R ², mean absolute error (MAE), and root-mean-square error (RMSE). The obtained correlation coefficient of 0.9972 indicated a substantially positive linear relationship between the values obtained from conventional IDA and the established ANN. Thus, it can be inferred that the developed model adequately identifies the underlying data patterns, as indicated by the MAE of 0.1093 that further substantiates the existence of only the small differences between actual and predicted values. Conclusively, the established ANN model adequately performs and predicts the values, substantially closer to the actual ones, indicating its efficacy to identify the fundamental pattern of data.

5.2 Fragility assessment using ANN

Using the network proposed by ANN, the fragility curves have been developed in this section, and compared with the fragility relationships, developed through IDA. Furthermore, the accuracy of the proposed ANN framework was validated and discussed. Fragility relationships established through ANN are presented in Figure 15. The produced fragility relationships using ANN exhibited a good match with the curves, produced using IDA. Table 7 shows a brief comparison of the median and standard deviation of the obtained fragility curves from IDA and ANN.

Figure 15

Comparison of fragility curves developed using IDA and ANN: (a) IM | PGA; (b) IM | S _a at 0.2 s; and (c) IM | S _a at 1.0 s.

It should be noted that fragility curves developed using both IDA and ANN are dependent on the ML algorithms, as IDA was conducted on an ML-based analytical model.

However, ANN substantially reduces the computational burden by identifying the structural response patterns. The efficiency stems from the ANN’s capability to process and analyze larger datasets by identifying the behavior of data and the relationship between input and output parameters, while the traditional analytical analyses method primarily relies on solving the time steps using iterative methods, i.e., time-step integration method, that consume larger computational resources. In the present work, the GA-based analytical model consumed 2 h on a core i7 computer (8th generation processor) with 8 GB of Ram for solving 1,000 time-steps, whereas the ANN learned the data patterns for structural responses in less than 5 minutes, and after training, the ANN yielded commendable results by approximately consuming 2 min of computational time, as indicated by Figure 15 that shows the comparison of fragility curves developed by means of the analytical model and the established ANN model. The statistical indices for the developed ANN model are provided in Table 6, indicating its adequacy in predicting the structural responses. The obtained results by implementing ANN effectively suggest that ANN algorithm can be trained quite efficiently using ADAM in PyTorch, and thus, fragility assessment can be efficiently executed in comparison with the IDA, given the building stocks at hand comprise uniform and homogenous configurations.

5.3 Performance metrics of proposed ANN and comparison of results using elementary ML and deep learning

The accuracy of the proposed ANN was established in this section, and a comparison was made of the results obtained only from Weka. The proper implementation of the proposed framework entailing data preprocessing in Weka and establishment of ANN in PyTorch. For determining the accuracy of proposed ANN, the classification-specific metrics were employed as designated by the studies by Sabery et al. and [59] and Zain et al. [60]. Their work indicated accuracy, precision, recall, and F1-score as the metrics targeted to establish ANN’s performance. These metrics were dependent on the capability of trained model to separate a specific class of interest (positive category) from the residual data (negative category). Since the primary objective of this article is to discuss the performance metrics instead of their theoretical background, only the general description and their respective quantitative ranges were provided herewith.

The efficacy of learning classifiers is indicated by the accuracy, and its quantitative range varied between 0 and 1, where higher value indicates better performance. Precision represents the reliability of ANN to identify the actual damage and its mathematical range also ranges from 0 to 1, with 1 representing the highest precision of an ANN. Recall and F1-score also range from 0 to 1, with 1 indicating the highest level of ANN’s performance for both metrics. Recall and F1-score indicate the ANN’s capability to capture all instances of actual damage and the harmonic mean of recall and precision, respectively. Table 8 summarizes the values obtained using deep learning conducted with ADAM optimization and ML by using Weka only.

From the comparative evaluation, results indicated that the inclusion of deep learning substantially enhanced the overall performance of the established ANN, resulting in better performance metrics. The use of deep learning algorithm remained advantageous over the use of basic ML that had been implemented using Weka alone.