WeldiNet: An improved design of a Rigdelet neural network for welding defect detection based on an enhanced pufferfish optimization algorithm

1. Introduction

Welding is an important element in the production and manufacture of metal materials and is used in various types of industries. Among the main applications of welding, we can mention the construction of large metal structures such as bridges, buildings, and means of transportation, and forging, which must have high welding quality (Vasan et al., 2024). Also, in the automotive industry, welding is very important in the production and repair of automotive parts. In addition, welding is the main part of the production process in the marine industry, aircraft manufacturing, and the production of electrical and electronic devices. Welding defects cost a lot of money (Mustafa et al., 2024).

Non-destructive tests are one of the most important technical tools for inspecting various welded parts. In the welding process, it is very important to identify welding defects in the performance of parts and structures, and techniques such as radiography, which can determine their information without destroying or changing parts, are more important (Block et al., 2024). Due to the X-ray diffraction from the material and other electronic effects of the imaging system, usually the recorded images are accompanied by the effects, and in some cases, they do not have the required quality, or the image is distorted and making it difficult to detect defects (Cui et al., 2024).

Welding radiography is one of the types of NDT to find internal defects and discontinuities in small and large metal structures. Due to the high cost of this welding inspection method, the radiation welder or welding test expert must have the necessary knowledge to interpret the radiographic film and find all types of defects in it, so as not to waste money and time (Li et al., 2024). Image processing methods are used to solve these problems.

Due to the rapid development of machine vision technology, research on welding image defects using X-ray has been of interest to researchers. Here, welding joints, which are the main weaknesses in welding, and welding methods are examined using defect detection algorithms. In order to improve the efficiency of X-ray welding image detection, deep learning networks and double exposure algorithms are also used to identify welding defects, which improves the accuracy and speed of the inspection process (Zhang et al., 2024).

Therefore, different algorithms based on image processing, machine vision, deep learning, and other relevant observations were used to identify and classify welding defects from radiographic images.

For example, Madhav et al. (Madhav et al., 2023) aimed to assess and improve the welding operations’ accuracy by conducting and using Deep Convolutional Neural Networks (DCNN). Incomplete or missing weld procedures can be recognized in an accurate manner by the suggested DCNN. There were the number of 10000 Not-OK and OK pictures in the training set that assessed the DCNN. After that, the network was made, assessed, and optimized to identify flaws in welding. The output of the model was examined, and the findings represented that the DCNN could forecast the flaws with good accuracy. The accuracy could be enhanced by 99.01% by the use of data augmentation once the training procedure was accomplished with 9,600 pictures.

Block et al. (Block et al., 2024) represented LoHi-WELD, which was a public and original dataset to overcome the issues relevant to weld flaw classification and identification that were of 4 diverse types, including stains, discontinuities, deposits, and pores. There were the number 3022 actual weld bead pictures that were annotated in a manual manner for visual examination. They consisted of images with high- and low-resolution, which were obtained from a Metal Active Gas robotic welding procedure. A diversity of a baseline deep framework for the suggested database was explored on the basis of the YOLOv7 network; moreover, several case investigations were discussed. The mean average precision (mAP) value of a fine-grained flaw categorization was 0.69, and the value of a coarse categorization was 0.77, which were achieved by a lightweight framework.

Palma-Ramírez et al. (Palma-Ramírez et al., 2024) presented a kind of CNN on the basis of ResNet50 to categorize 4 diverse kinds of weld flaws within radiographic images, including no defect, non-penetration, pore, and crack. Regularization, data, augmentation, and stratified cross-validation were employed to avoid over-fitting and enhance generalization. The model was examined using three diverse databases, including a private dataset of low image quality, GDXray, and RIAWELC, which in turn, accomplished accuracy values of 75.83%, 90.25%, and 98.75%.

Yadav et al. (Yadav et al., 2024) employed a CNN-SVM hybrid model and assessed it on the basis of several evaluation metrics with the purpose of welding flaw categorization. There were several flaw categories, including Slag Inclusion, Undercut, Porosity, Cracks, and Ideal Weld. For these categories, the suggested model had the values of more than 96% for F1-score, recall, and precision. The suggested model could classify the defects with an accuracy value of 96.92 using a database with 4880 images. The results demonstrated that the model could perform well and was superior to other models existing in the study. However, there might be some problems in the robustness and reliability of the current model.

Vasan et al. (Vasan et al., 2024) suggested a deep learning method on the basis of the ensemble for detecting and monitoring immersed weld flaws in adjacent regions and weld beads throughout Non-Destructive Testing (NDT). The approach made use of an open-source dataset to classify immersed arc weld classes into three categories of defects, including lack of penetration, cracks, no defects, and porosity. To improve the investigation of images, preprocessing and feature extraction techniques were implemented in frequency and spatial domains while utilizing segmentation methods, such as the Grey Level Co-occurrence Matrix (GLCM), Grey Level Difference Method (GLDM), texture analysis, Fast Fourier Transform (FFT), and Discrete Wavelet Transform (DWT). Once the preparation of images has been conducted, the preprocessed data undergoes testing and training in the suggested deep learning model on the basis of the ensemble. The efficacy of the model was generally evaluated by the use of several evaluation metrics. Moreover, it is worth noting that the model could accomplish an accuracy value of 93.12% for the categorization and detection of faults in welding.

Ridgelet neural networks (RNNs) are one of the deep learning techniques that can be used as an efficient tool for image-based defect detection. However, using the conventional methods for selecting the RNN structure is not so promising. Using metaheuristics in this task is a better way to resolve this issue.

The Pufferfish Optimization Algorithm (POA) is a newly introduced metaheuristic algorithm with lucky solutions for complicated optimization problems. In addition, based on the free lunch theorem, the original POA may not be so effective in optimizing the structure of the RNN. So, providing an enhanced version of this algorithm can be useful for welding defect detection.

The present study provides a new framework for detecting welding defects through the use of an RNN that is optimized by an enhanced variant of the POA. The proposed methodology will be tested on a dataset comprising welding images that exhibit a range of defects.

2. Data source benchmark

The present study uses the “GDXray” benchmark dataset for validating the proposed model. The GDXray consists of 19,407 x-ray images for weld defect detection, which is collected by Brewster Angle Microscopy (BAM) (Carrasco Zambrano 2023). The dataset is divided into five categories: welds, castings, natural entities, and luggage, each containing multiple series of images. For this study, we focus on the “welds” category, which comprises 88 images organized into three series: “W0001”, “W0002”, and “W0003”.

Series W0001 and W0002 contain 10 x-ray images each, while series W0003 consists of 68 digitized radiographs images (Mery et al., 2015) (Table 1). These images represent various types of weld defects, including porosity, cracks, and lack of fusion, making them suitable for evaluating the performance of the proposed weld defect detection method.

Fig. 1 shows some samples of the GDXray dataset.

Close

Fig. 1.

Some samples of the GDXray dataset for welding.

Export to PPT

The GDXray dataset can be considered a trustworthy and standardized reference for assessing the effectiveness of the proposed method, removing the necessity for costly X-ray machinery and guaranteeing the consistency of the results.

3. Methodology

Fig. 2 shows the graphical illustration of the proposed method. As can be observed, the images downloaded from the “GDXray” dataset have been improved by two distinct preprocessing stages. The first stage is to apply a CLAHE filter for contrast enhancement of the images. Then, some augmentations have been applied to increase the number of images (Hayati et al., 2023; Xu et al., 2023). This is done because the number of datasets for training a deep learning model is small.

Close

Fig. 2.

The graphical illustration of the proposed method.

Export to PPT

The images are then trained by an RNN for the diagnosis task. As can be observed from the image, an EPOA has been applied to the network to improve the efficiency of the RNN. Finally, the network has been analyzed based on some different metrics, and its results have also been compared with state-of-the-art methods to show its superiority for using in the diagnosis task for welding applications.

4. Preprocessing

The preprocessing is an important stage in our work to prepare the data for the analysis and processing steps. This part includes two key components: contrast enhancement (based on CLAHE) and data augmentation. CLAHE is a powerful technique for enhancing the contrast of images, especially in cases where the contrast is low or the dynamic range is limited. In addition to contrast enhancement, the data augmentation technique is used to artificially increase the size of our data, which involves applying a series of transformations to the original data to produce new samples similar to the original data but with changes that help. In the following, the two methods are explained in detail.

4.1 Contrast enhancement

Contrast enhancement is an important stage in image processing that aims to improve the visibility of features in an image by adjusting the contrast between different regions. In this study, contrast-limited adaptive histogram equalization (CLAHE) has been used to enhance the contrast of X-ray welding images. The AHE divides the image into small regions called tiles and applies histogram equalization to each tile separately. CLAHE, on the other hand, limits the stretch of contrast to avoid excessive amplification of noise in the image. The CLAHE algorithm can be summarized by the following stages (Musa et al., 2018):

• -

  Dividing the original image into non-overlapping tiles, usually 8×8 (in this paper) pixels.

• -

  Calculating the histogram of the tiles.

• -

  Trimming the histogram to the maximum value, which is a user-defined parameter.

• -

  Applying the histogram smoothing to the tiles based on the clipped histogram.

• -

  Interpolating the results from the tiles to create a unified image

Fig. 3 shows an example of clarification of the effect of CLAHE on the welding image, with the top image showing the original image and the bottom image the improved image after applying CLAHE, along with the histograms of both images.

Close

Fig. 3.

One example of clarification of the effect of CLAHE on the welding image.

Export to PPT

As can be observed from Fig. 3, the original image appears dark and lacks detail, with a narrow dynamic range, which makes it challenging to distinguish between different features, while the histogram shows a narrow peak with a low intensity value, indicating low contrast. Although the CLAHE-enhanced image shows a significant improvement in contrast and feature visibility, appearing brighter and more detailed, with a wider dynamic range, and a more uniform distribution of intensity values, with a higher peak value histogram, indicating contrast. Above, the CLAHE algorithm has effectively enhanced the contrast of the welding image, making it easier to distinguish between different features and increasing the contrast by stretching the intensity values. It improves the visibility of details and features and increases the dynamic range. This shows the effectiveness of CLAHE in enhancing the contrast and visibility of features in a welding image for image analysis and processing.

4.2 Augmentation

To enhance the size of the input dataset and strengthen the resilience of the model, various image augmentation techniques have been implemented on the original images. The augmentations included:

Flipping: Creating mirrored versions of the original images by flipping them horizontally and vertically.

Rotating: Generating new images by rotating the originals at angles of 180 degrees. Contrast Adjustment: Modifying the contrast levels of the images to produce variations in contrast (Xu et al., 2023).

These augmentations were applied randomly to the original images, resulting in a more extensive and diverse dataset. The augmented images were subsequently utilized for training and testing our model, thereby enhancing its capacity to generalize new, unseen data. Fig. 4 shows some examples of the augmentations performed on an original image.

Close

Fig. 4.

Some examples of the augmentations performed on an original image.

Export to PPT

Table 2 illustrates dataset statistics after augmentation.

Table 2. Dataset statistics after augmentation.

Category	Number of images	Series	Number of images per series
Welds	1760	W0001	200
		W0002	200
		W0003	1360

Through these augmentation techniques, we successfully expanded our dataset and bolstered the robustness of our model, enabling improved performance on new, unseen data.

5. RNN

RNN has been considered an architecture of neural network that employs Rigdelet as a function of activation within the hidden layer. Rigdelet is the base of directional data description in domains with high dimensions and provides a better function in finding characteristics that are arranged based on particular orientations or directions. Rigdelet originates from the procedure of mathematical signals that are used for analyzing functions with the help of singularities of a hyperplane. The RNN could find and display directional information inside the data through merging Rigdelet in the hidden layer’s function of activation (Wang et al., 2024).

Rigdelet have been used for activation function in the RNN’s hidden layer, because they can display functions with singularities of hyperplane. Such singularities are related to the occasions that functions represent unexpected changes or irregularities among specific hyperplanes. Rigdelet has been the best in demonstrating these functions, which makes it a great choice for the tasks that comprises finding hyperplane singularities.

The RNN and Rigdelet’s cooperation could be beneficial in an exceptional way for implementing data with a high dimension, like vision of the computer recognizing, the patterns and procedure of image. The directional information, which Rigdelet is able to find, could help the network recognize and remove relevant features from the data, which in the end leads to improved action within tasks that need a directional consciousness of characteristics.

It is possible to achieve a more extensive comprehension and to be deeper to the subject that prepares particularized perceptions within the relation between Ridgelets in the neural network structure and emphasizes their benefits in displaying information of direction and operations with the singularities of hyperplanes. Actually, the results of the study include more data or modified versions of the RNN.

The conversion of Ridgelet is a mathematical procedure, which has been used to analyze and manage directional data. Its base is located in a function of neurons that $\psi :R\to R$ and is acceptable when it assembles the recommended situation. Soc, if it does not meet the recommended condition, it is not allowed.

here, the Fourier conversion of the $\psi$ has been displayed by $\hat{\psi }\left(\theta \right)$.

The present situation guarantees that the function owns a limited energy and is square integrable. In fact, the transformation of Ridgelet uses the allowed function $\psi$ in the input data during various directions for extracting the directional data characteristics. The neural network prepares the resulting converted data for the task of regression or performing the intended categorization.

Through considering the word $\zeta$, a neuron area $\zeta =(a, u, b)$ could be obtained:

Here, the placement of the Ridgelet function is displayed by $b$, the Ridgelet function’s scale is represented by $a$, the orientation of the function of the Ridgelet is illustrated by $u$, and the $S^D$ illustrates a sphere in area $R^D$.

When transformation, rotation, and the scaling process of the allowed functional neuron $\psi$ is completed, a Ridgelet group could be extracted through utilizing the formula below:

where the Ridgelet direction is displayed by the variable $\theta$. The Ridgelet, which is a result of the transformation, rotation, and scaling of the original function $\psi$, is illustrated by ${\psi }_{\zeta }$.

The Ridgelets bank’s use of the input data in various directions is a method applied by the Ridgelet conversion for extracting the data’s directional characteristics.

Moreover, the transform of Ridgelet is an important device to estimate the singularities of hyperplane in a rapid way within a function group to be used for estimating function which are multivariate. Within present context, the function [$y|y=f(x):R^n\to R^m]$ is divided into $m$ computation ($R^n\to R)$, and the estimated multivariable function has been obtained employing the formula below:

here, the Ridgelet’s superposition coefficient has been shown via $c_{ij}$.

An RNN that has one hidden layer could be improved by using the conversion of Ridgelet as the function of activation in the hidden layer. The transformation of Ridgelet has been applied to eliminate the input data’s directional characteristics within different directions, after which it has been moved along the hidden layer. The Ridgelet coefficients achieved from the hidden layer are integrated in a linear manner to shape the RNN’s output layer.

The structure of the RNN has been depicted in Fig. 5. First, the input data was converted by employing the transformation of the Ridgelet, which removes the data’s directional characteristics. After that, the converted data was transported along the hidden layer where the transformation of Ridgelet was applied as the function of activation. Then, the coefficients of the Ridgelet produced by the hidden layer were integrated within the output layer for creating the network’s final output.

Close

Fig. 5.

The Ridgelet-NN’s structure.

Export to PPT

The structure of the RNN has many benefits over traditional neural networks which are traditional. First of all, it is more effective in preparing directional data since it concentrates on the directional characteristics of the data, decreasing the neural network’s calculation complexity. Then, it is less affected by the input direction, preparing it to be more resistant to changes within the data.

Lastly, it is displayed that traditional neural networks outperform in various applications, namely, seismic data analysis, medical image analysis, and remote sensing. The structure of the RNN is used in different areas within medical imaging, like X-ray. Within the analysis of medical images, RNN is utilized for segmenting, registering, classifying, and reconstructing the image. For instance, inside the recreation of the MRI image, the RNN is illustrated for decreasing the calculative complexity in comparison with the traditional neural network and enhancing the quality of the image.

A linear integration of the Ridgelets and weights ($w_i$) is use to enhance the Ridgelet-NN’s estimation operation. The initial level in creating the Ridgelet-NN is to appropriately choose the network parameters. Such parameters have been explained below:

where the quantity of Ridgelets, employed in the network, is displayed via $K$, and the variables of the Ridgelet scalar are illustrated by $a$, $b$, and $w$, and the orientation is represented by $u$ as follows:

A linear integration of the Ridgelts and weights were applied by the Ridgelet-NN in order to estimate the input data. The RNN computes the coefficients of the Ridgelet within the hidden layer as a linear integration of the Ridgelets and the input data. The RNN’s output layer is considered a linear integration of coefficients of the Ridgelet achieved from the hidden layer.

The scalar variables’ choice ($a$, $b$, and $w$), the weights ($w_i$), and the vectors of Ridgelet orientation ($u$) are important for the RNN’s operation. The existing parameters should be chosen in a proper way to guarantee the efficient extraction of the input data’s directional characteristics and estimation of the intended output. The choice of the present parameters could be completed with the help of a trial-and-error procedure or algorithms’ optimization and gradient descent. The network’s $k^{th}$ output neuron is computed employing this formula:

where,

here, $l=1,2,\dots ,m$.

For achieving a meticulous approximation of the issue, it is needed to decrease the cost function to its smallest quantity.

To obtain a precise estimation of the issue, the cost function needs to be reduced. This cost function is a sum of squared errors between the actual outputs and forecasted outputs in all of the instances of training, and it is divided by the training instances of total number ($T$). For minimizing the cost function, the parameters of the Ridgelet require to be chosen in an optimal way.

The primary goal of this search is to propose a novel change to an algorithm, which is metaheuristic, in order to obtain finer outcomes by choosing the parameters of the Ridgelet and reducing the cost function. The metaheuristic algorithm is a kind of optimizer that could be utilized to discover the optimum solution to a problem by exploring the solution space in an iterative way. The suggested change’s purpose is to enhance the algorithm’s effectiveness and efficiency within the optimum solution for the RNN.

6. EPOA

6.6 EPOA

The POA has been considered a novel metaheuristic inspired by the pufferfish’s natural manner and aimed at solving complex issues. In fact, its primary aim is to stimulate the pufferfish’s defensive mechanism of inflating its body to scare the predators. The POA has been arranged to address the limitations of these optimization approaches, including computational costs, insufficient convergence, and low accuracy. The improved POA or IPOA introduces several improvements, including a harmony between local and global search, managing adaptive parameters, and a new technique that all of them upgrade the capability of the algorithm’s accuracy and effectiveness.

A major improvement is made through incorporating an adaptive weight element, which is nonlinear. This gets expanded in a finer solution space based on the candidate solutions’ quality. This adjusting element is important because it avoids the population’s early convergence and promotes the distinction, resulting in a better function. It must be mentioned that the improved equation of $X$ might show fluctuations between the negative and positive values.

(22)

$y_{i,j}^{Q1}=y_{i,j}+w_i\times \left(S{P}_{i,j}-K_{i,j}\times y_{i,j}\right)$

The weight element of $i^{th}$ value has been illustrated via $w_i$, and it has been calculated utilizing the following equation:

(23)

$w_i=\frac{\overline{f}-f_i}{\overline{f}-\underset{\_}{f}},i=1,2,\dots ,N_P,$

where the cost function’s greatest value has been represented via $\overline{f}$, and its least value has been displayed via $\underset{\_}{f}$. The cost function of $i$ has been demonstrated via $f_i$. Candidates with a lower quantity of cost functions are capable of exploring wider solution spaces through using a weighting element, while the ones with greater cost function values are limited to constrained solution spaces. The present approach assists in maintaining a harmony between the capacity of local and global search within the algorithm. Another improvement uses a chaotic map $\gamma$ instead of random $r$ through the subsequent formula:

(24)

$y_{i,d}=l{b}_d+{\gamma }_t\times \left(u{b}_d-l{b}_d\right),$

The current search employs Bernoulli shift map through using the subsequent equation:

(25)

${\gamma }_t^{new}=\left(\begin{array}{c}\frac{{\gamma }_t}{1-\beta }, 0<{\gamma }_t\le 1-\beta \\ \frac{{\gamma }_t-\left(1-\beta \right)}{\beta }, 1-\beta <{\gamma }_t<1\end{array}\right)$

where, the quantity of iterations has been displayed via $t$, and $\beta =0.4$. The Bernoulli shift map is capable of enhancing the algorithm’s convergence.

This study uses the EPOA for computing the parameters of new fractional model to optimize the neural network for correction of writing. The present model has been represented in Fig. 6.

Close

Fig. 6.

The optimization of model on the basis of EPOA.

Export to PPT

The main aim of this study is to enhance the architecture of the RNN by employing the EPOA to develop an effective system for detecting welding defects.

7. Results and Discussion

This section outlines the simulation outcomes of the proposed approach for detecting welding defects through the use of an RNN that has been optimized by the EPOA. Given the stochastic nature of metaheuristic algorithms, the results can differ with each execution. Consequently, each experiment was conducted 25 times, adhering to the parameter configurations.

Various other algorithms were employed for comparative analysis. To ensure a fair comparison, all algorithms were executed with a limit of 200 iterations (equating to 5,000 function evaluations). The programming was carried out in MATLAB R2019b 64-bit, and the computations were performed on a system equipped with an Intel Core i7 CPU running at 2.00GHz, 2.5GHz, with 16GB of RAM and a 64-bit operating system.

7.4 Regression performance

In addition to analyzing the classification effectiveness of our proposed method, its regression performance is also examined by using the Mean Squared Error (MSE) metric. MSE quantifies the average squared deviation between the predicted values and the actual values of the welding defect severity. The results of the regression performance evaluation have been presented in Fig. 7.

Close

Fig. 7.

The results of the regression MSE performance evaluation.

Export to PPT

The regression performance evaluation shows that our proposed RNN/EPOA shows superior performance in terms of MSE. This finding shows that the proposed EPOA significantly increases the model's ability to accurately predict the severity of welding defects.

In contrast, the original type of RNN shows the highest MSE, which reflects the largest mean squared deviation between the predicted and actual values. This result shows that RNN alone is insufficient to achieve accurate regression performance, and optimization is necessary to increase its effectiveness.

The variant RNN/POA shows a decrease in MSE compared to the original RNN, which indicates that the original POA positively contributes to the regression performance of the model. Nevertheless, the RNN/EPOA variant achieves an even lower MSE, which indicates that the advanced optimization algorithm is more adept at improving the regression capabilities of the model.

The increase in regression performance observed in the RNN/EPOA variant can be attributed to the superior ability of the advanced optimization algorithm to fine-tune the model parameters. This leads to more accurate prediction of weld defect severity, which is very important in industrial fields where accurate defect detection and classification are critical

7.5 Computational time

Computational time is a key factor in the creation of machine learning models, especially in industrial frameworks where prompt decision-making is important. In this research, we evaluated the computational time associated with our proposed method, RNN/EPOA, and compared it with the original RNN and RNN/POA variants. The findings from the computational time assessment have been displayed in Fig. 8.

Close

Fig. 8.

The computational time assessment (seconds).

Export to PPT

The computational time evaluation results show that the RNN/EPOA variant requires the longest computational time, followed by the RNN/POA variant, and then the original RNN, which is expected to be increased due to the complexity added by the POA. The increased computational time of the RNN/EPOA type can be attributed to various factors, including optimization complexity, which is due to a more complex algorithm, model architecture, which is more complex with additional layers and connections, and more training data, which requires more calculations to optimize the model parameters. Despite the increased computational time, the RNN/EPOA variant achieves the best performance in terms of classification and regression accuracy, indicating that the improved optimization algorithm is effective in improving model performance, albeit at the cost of increased computational time, which represents a trade-off. There is a gap between model performance and computational efficiency

7.6 Comparative analysis

To comprehensively evaluate the performance of the current RNN/EPOA model, a 5-fold comparative analysis was performed with five existing models: Convolutional Neural Network (CNN) (Alhijawi and Awajan 2024), DCNN (Madhav et al., 2023), YoloV7 (Block et al., 2024), CNN-SVM (Yadav et al., 2024), and deep learning (Vasan et al., 2024). Comparison analysis was performed using the following criteria: precision, sensitivity, specificity, accuracy, F1 score, and MCC. The results of the 5-fold comparison analysis have been presented in Fig. 9.

Close

Fig. 9.

The results of 5-fold comparison analysis.

Export to PPT

The findings from the comparative analysis indicate that the RNN/EPOA model surpasses the current models regarding classification performance. The CNN and DCNN models exhibit lower values in precision, sensitivity, specificity, accuracy, F1-score, and MCC when compared to the RNN/EPOA model. Although the YoloV7 model performs slightly better than both CNN and DCNN, it still falls short of the performance demonstrated by the RNN/EPOA model. Furthermore, while the CNN-SVM and DL models show improved performance relative to CNN and DCNN, they do not reach the performance level of NasNet/AWCA.

8. Conclusion

Welding is connecting the metal parts in general and steel parts using different methods. This can be done based on heat, pressure, or a combination of both. Welding defects production of healthy and defect-free welds, along with reducing costs, is one of the desired goals in the welding operation of metal structure joints. The main welding defects that cause the welding connection to break are among the things that every welder should be aware of. One of the most important reasons to be aware of these issues is the production of quality welds and the reduction of welding inspection time. The materials to be welded together (sheet or pipe) should be checked for weld surface defects. Because pipe welding defects are one of the most common defects. Image processing and machine learning as one useful Non-destructive techniques for this purpose. This study proposed a new method based on deep learning and a metaheuristic algorithm for the detection of welding defects. Here, a RNN was used as a classifier of the defects. The RNN efficiency was improved by optimizing its parameters to get optimal results. The optimization was performed by an EPOA. The proposed RNN/EPOA model was assessed using a benchmark dataset, called GDXray and its outcomes were compared with several leading techniques, including Convolutional Neural Network (CNN), DCNN, YoloV7, CNN-SVM, and Deep learning (DL) to demonstrate the superiority of the method. The results reveal that the proposed RNN/EPOA model can effectively identifying a range of welding defects. Future works will concentrate on enhancing the efficacy of the proposed detection system based on using different deep learning frameworks and optimization techniques. A capable direction is to use transfer learning and domain adaptation methods to modify the system for this purpose. Moreover, incorporating additional NDT methods, such as ultrasonic testing and radiography, alongside the proposed system could yield a more holistic approach to defect detection. Also, the practical application in real-time or industrial settings may be hindered due to the high computational time, which can primarily be attributed to the sophisticated nature of EPOA. But this extra computational cost may be an impediment to implementing it in time-critical applications. Indeed, future studies can be also focus on improving model computational efficiency, either by simplifying the optimization algorithm applied, adapting it to the use of conventional methods such as parallelization or hardware revolution, preserving the performance achieved by the model. Sources of defects include more focused data, infinities; new methods of coding deeps. These would be the basis of making the higher-performance RNN/EPOA model more malleable and, more importantly, more useful in real-time applications where quick, accurate prediction of defect production is necessary. We will also focus on increasing the performance of the proposed detection system using other deep learning frameworks and optimization methods. Moreover, the study is not considering transferring learning to the proposed model, where the model, successfully trained on one dataset, can be used for a different dataset. Additive techniques such as transfer learning would enable the model to build on currently available networks and retrain on new datasets with less data loss. Meanwhile, domain adaption techniques could make the model adaptable to environmental or material-specific noise. Additionally, the addition of supplementary NDT techniques, including ultrasonic testing and radiography, harmonizing with the proposed system offered, would present a more comprehensive means of detecting defects covering wider industrial challenges.