Research on Nondestructive Testing Technology for Drilling Risers Based on Magnetic Memory and Deep Learning

Abstract

Drilling risers play a crucial role in deepwater oil and gas development, and any compromise in their integrity can severely hinder the progress of drilling operations. In light of this, efficient and accurate nondestructive testing of drilling risers is paramount. However, existing inspection equipment falls short in both efficiency and accuracy, posing challenges to the sustainability of deepwater oil and gas exploration and development. To effectively assess the damage conditions of deepwater drilling risers, this study developed an inspection robot based on metal magnetic memory and researched intelligent defect recognition methods using computer vision. The robot can perform in situ inspections on drilling risers and has been successfully deployed for field application on a deepwater drilling platform. The application results demonstrate that this detection robot offers significant advantages regarding high reliability and detection efficiency. Utilizing data collected on-site, we constructed a dataset containing 1100 images that cover five typical types of defects in drilling risers, including pitting, groove corrosion, and wear. Based on this dataset, we proposed and trained a novel image classification model, SK-ConvNeXt-KAN. By deeply optimizing the ConvNeXt convolutional network incorporating the introduced SK attention module and replacing traditional linear classification layers with the KAN module, this model significantly enhanced its feature extraction capabilities and efficiency in handling complex nonlinear problems. Experimental results show that this model achieved an accuracy rate of 95.4% in identifying defects in drilling risers, which is significantly better than traditional methods. This achievement has dramatically improved the efficiency and accuracy of deepwater drilling riser inspections, providing robust technical support for deepwater oil and gas exploration and development sustainability.

1. Introduction

In deepwater drilling, the riser is the critical equipment that connects the drilling platform to the submarine wellhead. Unlike the embedded conduit commonly used in shallow water drilling, the riser system is more complex and is usually composed of dozens to hundreds of connected risers [1,2]. As shown in Figure 1, the bottom of the system is equipped with a blowout preventer, while the top is closely connected to the drilling platform through a chuck and tensioner. The riser operates in waters ranging from hundreds to thousands of meters deep, enduring alternating loads caused by waves, currents, and platform offsets while resisting seawater corrosion and uneven wear caused by drill pipes. The riser is prone to various defects such as cracks, corrosion, and wear during long-term service under such harsh conditions, making it a vulnerable component of the drilling system [3,4,5]. Once these defects cannot be detected promptly, they may lead to the rupture and failure of the drilling riser, which could result in a significant loss of drilling fluid and seriously hinder the progress of drilling operations. More seriously, this situation may also induce uncontrolled blowouts, causing severe pollution to the marine environment.

Numerous historical examples exist of such incidents [6,7,8,9,10]: In 1982, a semi-submersible drilling platform suffered a marine riser collapse accident in a 580 m deep sea area of the Gulf of Mexico, as shown in Figure 2a. In this accident, due to the massive leakage of drilling fluid, the bottom two sections of the drilling riser could not withstand the excessive external hydrostatic pressure and collapsed, resulting in significant economic losses and marine pollution. In 1999, during drilling operations conducted by BP at a depth of approximately 2200 m in the Gulf of Mexico, the main pipe above the flexible joint of the riser ruptured due to abrasion from the drill pipe, as shown in Figure 2b, which led to a substantial loss of drilling fluid, necessitating a forced shutdown for repairs. In 2003, BP’s Enterprise drilling vessel experienced excessive platform offset during operations at the Thunder Horse oil field, causing the high-speed rotating drill pipe to continuously wear down the inner wall of the marine riser, resulting in wall thickness reduction and ultimate rupture. This accident caused over 300 m of the marine riser to fall near the subsea wellhead, as shown in Figure 2c,d, leading to significant economic losses and severe pollution in the sea area. In 2010, the riser on Well DF1-1-14 in the Yinggehai Basin of the South China Sea developed a transverse crack of up to 48 cm in length, as shown in Figure 2f, resulting in a substantial loss of drilling fluid. This incident caused severe marine pollution and consumed significant labor hours for necessary repair work.

Therefore, strengthening research on nondestructive testing (NDT) of drilling risers to enhance their safety and reliability has profound significance for the protection of the marine environment and achieving the sustainable development of offshore oil and gas resources. However, given the harsh conditions of deepwater environments and the complexity of riser systems, there is currently a lack of mature in situ detection methods. Traditional detection methods can only be performed onshore at drilling bases, requiring the prior retrieval and transportation of risers from underwater to land, followed by inspections using flaw detection equipment such as ultrasonic, magnetic particle, and magnetic flux leakage tests [11,12]. This process is time-consuming and lacks intelligent data processing, making it difficult to meet actual demands.

This paper provides research on nondestructive testing and intelligent defect recognition technologies for deepwater drilling risers in this context. We successfully developed an in situ inspection robot for deepwater drilling risers based on magnetic memory technology that can accurately inspect in-service risers directly in deepwater environments, significantly reducing the inspection cycle and enhancing inspection efficiency. Additionally, leveraging deep learning technology, we propose an innovative defect image classification and recognition algorithm that enables the intelligent identification of riser defects. Through seamless integration of the proposed hardware and algorithm, we construct a comprehensive intelligent riser inspection system capable of automatically collecting, processing, and analyzing inspection data, providing timely and reliable inspection support for deepwater drilling operations.

2. Related Work

2.1. Existing Nondestructive Testing Devices for Risers

In recent years, numerous scientific research institutions have conducted extensive research on riser field technology. The Dutch scholar Herman proposed a riser detection system called RTD-INCOTEST, based on a pulsed magnetic eddy current [13]. This detection method exhibits low requirements for the cleanliness of the pipe wall surface and boasts fast detection speed. However, its promotion and application in deep water are subject to certain limitations, primarily as the system is designed for a water depth of 150 m and necessitates the involvement of divers for the installation of probes and the deployment of detection cables. In 2013, the Federal University of Rio de Janeiro in Brazil developed a robotic system (SIRIS) for in situ inspection of risers [14]. This system is equipped with four cameras capable of performing visual inspections of pipelines and can be customized by attaching ultrasonic, eddy current, and radiographic detection systems. The primary design objective of this device is to facilitate the inspection of flexible risers located on production platforms, focusing on suitability for water depths of up to 250 m. TechnipFMC introduced a multi-functional offshore pipeline online inspection tool called IRIS at the Asian Offshore Technology Conference in 2018 [15]. This tool integrates three nondestructive testing methods—electromagnetic testing, ultrasonic testing, and X-ray testing—enabling it to effectively identify defects such as corrosion, fractures, and cracks. The primary application of IRIS targets the nondestructive testing of flexible risers and umbilical cables located on production platforms. Weatherford proposed an online inspection device for offshore risers based on ultrasonic technology, called ORIS, at the 2017 American Offshore Technology Conference [16]. The device can be directly lowered into the wellhead or vertically stacked on the drilling platform using a winch, with a detection speed of up to 4 m/s. The ultrasonic sensor transmits real-time data through a wired logging unit, enabling online detection of pipe wall cracks and thickness. However, due to the coupling issue between the probe and the pipe wall that has yet to be adequately resolved, the device is still in the prototype development stage and has not been implemented in practical engineering applications. The China University of Petroleum (Beijing, CUP) conducted a series of studies on the weak magnetic distortion characteristics during the fatigue damage process of marine risers [17,18]. The China University of Petroleum (East, UPC) introduced an alternating current field measurement (ACFM) technology for nondestructive testing of underwater structures associated with deepwater drilling platforms in 2022 [19]. The ACFM system developed by the university has been successfully used in multiple field tests. However, as the detection method is external, it was only used to inspect the risers in the splash zone and could not be applied to risers in deeper waters.

Table 1. Comparison of existing riser detection technologies.

Institution	Detector	Method	Type	Features
Dutch scholar	RTD	Pulsed Eddy Current	External inspection	Water depth of 150 m
Brazil	SIRIS	Visual, Ultrasonic, Radiographic	External inspection	Water depth of 250 m
TechnipFMC	IRIS	Electromagnetic, Ultrasonic, and X-ray Inspection	External inspection	Production platforms
Weatherford	ORIS	Ultrasonic Testing	Internal inspection	In situ inspection
UPC	ACFM	Alternating Current Field Measurement	External inspection	In situ inspection
CUP	MMM	Metal Magnetic Memory	Internal inspection	Offline inspection

presents a comparison of these detection methods.

In summary, the inspection of drilling risers faces significant challenges due to the deepwater characteristics of their service environment and complex structure. During operations, the interior of a riser is filled with drilling fluid, while the exterior is continuously battered by waves, posing significant risks for inspection operations. Consequently, existing inspection technologies are primarily limited to shallow water areas. This study proposes introducing magnetic memory detection technology, which has advantages such as not requiring magnetization and having low cleanliness requirements for the surface. However, introducing this technology must also address the pressure resistance of the device and the storage and transmission of signals to ensure its practical application in deepwater environments.

2.2. Research on Intelligent Identification of Pipeline Defects

Given the unique challenges inherent to deepwater drilling, ordinary inspectors must undergo various complicated reporting procedures, including exams and physical examinations, before they can board the platform to conduct inspections, which significantly impacts the efficiency of the inspection process. Therefore, there is an urgent need to develop methods for intelligent analysis of inspection data to eliminate the reliance on inspectors.

In recent years, image recognition technology based on deep learning has made significant progress and has been widely applied in defect detection. Researchers have conducted numerous innovative works in this field. Geng utilized an improved deep convolutional neural network to classify and identify magnetic flux leakage signal images of circumferential welds in long-distance pipelines [20]. Hou developed a model based on deep convolutional neural networks, achieving an accuracy of 97.2% for deep features [21]. This high accuracy enables the model to effectively assist workers in assessing weld X-ray images more intelligently, thereby enhancing detection efficiency and accuracy. Mao proposed an online detection method for early bearing faults based on deep transfer learning, providing new insights for the timely detection of bearing faults [22]. Abu explored the application of four transfer learning models (ResNet18, VGG, MobileNetv3, and DenseNet-121) in steel defect detection, offering more options for defect detection in the steel industry [23]. Ma proposed a novel CNN ensemble framework based on transfer learning to classify bearing surface defects [24]. With an average detection time of only 155 milliseconds per bearing, this framework fully meets the requirements of industrial online detection, providing strong support for the rapid detection of bearing surface defects. Wu proposed a two-step detection method that combines array ultrasound with deep learning to address the limitations of traditional ultrasonic detection technology in detecting internal defects in concrete structures [25,26]. This approach enables the reliable identification of internal defects and intelligent interpretation of abnormal features in ultrasonic images through an improved YOLOv5 model. Ding integrated current perturbation theory with gradient imaging algorithms and, through detailed indoor experiments, meticulously constructed a dataset covering three types of defects: corrosion, cracks, and irregular shapes [27]. They successfully applied convolutional neural networks to achieve efficient, intelligent recognition and precise defect classification assessment of the detection images.

It is well known that image recognition networks based on deep learning rely on large-scale datasets for training. However, in the specific field of nondestructive testing, the inspection data are often different from traditional natural images. Taking magnetic memory signals as an example, this type of data is acquired with multi-channel sensors from the magnetic field information on the surface of pipe walls. Therefore, before introducing recognition algorithms from the field of computer vision, we must first explore and determine suitable visualization algorithms to convert these particular data into two-dimensional images that accurately reflect the morphology of defects. Due to the inherent challenges in this process, defect datasets are particularly scarce in nondestructive testing. In particular, no publicly available dataset based on magnetic memory testing poses difficulties for related research.

3. Materials and Methods

3.1. Development and Application of Magnetic Memory-Based Riser Inspection Robots

3.1.1. Design Requirements

When delving into the complexity and challenges of drilling riser inspection technology, we must recognize a pivotal obstacle: the bulky buoyancy blocks equipped on the outer side of the riser. These structures, made of composite materials, not only escalate the difficulty of inspection but also hinder the direct acquisition of high-quality inspection signals from the exterior. Additionally, considering the fixed outer diameter of the riser, coupled with its varying wall thicknesses according to water depth, the inspection task becomes even more intricate and variable. Consequently, we shifted our focus to internal inspection solutions to achieve a more efficient and precise inspection process. To cater to diverse inspection scenarios—including onshore bases and offshore platforms—the design of our inspection robots emphasizes lightweightness and compactness, ensuring operational flexibility and convenience. Furthermore, given the unique environment of offshore riser inspection, the robots are equipped with exceptional sealing and pressure-bearing capabilities, ensuring their stable performance even in deep-sea, high-pressure environments.

In light of this context, this paper aims to develop a riser inspection robot based on Metal Magnetic Memory (MMM) technology. Compared to traditional nondestructive testing methods, MMM technology stands out due to its unique advantages: it does not require additional magnetization of the inspected object. Thus, the surface cleanliness requirements for the pipe walls are low [28,29]. These characteristics make MMM technology exceptionally effective for the early detection of metal fatigue damage, establishing it as an ideal choice for the on-site inspection of drilling risers.

3.1.2. Design Proposal

The proposed robot is designed to inspect external 21-inch (533 mm) diameter risers. The device is equipped with 20 arcuate probes arranged in an orderly circular fashion, where each probe is embedded with three susceptible giant magneto-resistive sensors. As illustrated in Figure 3, this circular array of 60 sensors enables a comprehensive overlapping scan of the pipeline’s inner wall, ensuring full coverage and high-precision inspection. The probes are ingeniously mounted on the robot’s main body and secured by a telescopic frame featuring three support rods and tension springs. Under the action of the tension springs, the probes can flexibly extend and retract, effectively adapting to the varying inner diameters of pipelines, thus guaranteeing flexibility and precision during the inspection process.

The core component of the detector is a sealed inner cylinder, as shown in Figure 4, which is equipped with batteries, a data acquisition card, a router, and a wireless switch, among other vital components, forming the detector’s data collection and control center. The power and signal cables of the sensors are safely encapsulated with epoxy resin and connected to the inner cylinder cover via special waterproof pins. This design ensures a stable power supply to the sensors and facilitates communication with the data acquisition card. The data acquisition card performs exceptionally, with a maximum sampling rate of 2000 Hz per channel and up to 32 available channels. The collected signals can be transmitted in real time via a wireless network or temporarily stored on a USB device for subsequent analysis. Additionally, the top cover of the inner cylinder uses an O-ring locking design capable of withstanding up to 30 MPa of water pressure, fully meeting the stringent requirements of in-service riser inspection.

The detector exhibits high flexibility, allowing it to inspect horizontally stacked risers at onshore bases using a winch. It can also be lowered from offshore platforms using a crane to perform in situ inspections of operational risers. Figure 5 shows the specific posture of the robot as it enters the riser and conducts inspections inside the pipeline.

3.1.3. Industrial Application

Under the coordination and organization of the project team, we successfully carried out over 100 rounds of drilling riser inspections at both an onshore drilling tool base and an offshore drilling platform. Figure 6a shows the scene of damage detection on horizontally stacked risers at the drilling tool base, and Figure 6b shows the scenario of in situ inspection of the riser conducted on an offshore drilling platform.

3.1.4. Data Visualization Analysis

In the context of electromagnetic detection, the gradient field of the magnetic field is more capable of reflecting the abrupt characteristics of the magnetic field signal. Therefore, by solving the gradient field of the magnetic field, the distortion location of the magnetic field can be effectively determined, and the defect can be accurately located accordingly. The gradient calculation formula is shown in Equation (1) [30].

$$G_x={\displaystyle \frac{\Delta H}{\Delta x}}$$

(1)

In the formula, $\Delta H$ represents the variation of the tangential component of the magnetic field within the span $\Delta x$, which is usually selected as 1 mm or 2 mm. Gradient processing is a data processing method applied to various sensors’ scanning paths. After noise reduction and gradient operations, we can generate waterfall charts to visually observe the trends of each sensor’s signal over time, particularly in areas that show anomalies or distortions, thereby determining whether there is damage at those locations. Furthermore, we map the signal intensity values to a color space (e.g., the copper color spectrum used in this paper) for color encoding and control the image’s aspect ratio according to the actual detection range to generate a two-dimensional color image. In the 2D image, the variations in color depth reflect differences in the magnetic field strength, while the rows and columns of the matrix define the image’s resolution in the vertical and horizontal directions, respectively. This data processing method enables inspectors to locate defect areas quickly visually and provides a foundation for introducing subsequent image classification networks.

Figure 7 illustrates that, at the 14 m position, notable fluctuations were detected in the sensor signals across multiple channels, suggesting the existence of a potential defect at this locale. When these findings were correlated with the outcomes of the video inspection conducted on the pipeline’s inner wall, it was confirmed that several instances of corrosion were present at this specific location. This discovery indicates that the magnetic memory testing method can effectively detect corrosion on the wall of the drilling riser.

As depicted in Figure 8, the two-dimensional gradient image presented below provides a clear visualization of a continuous ‘pit’ located approximately 4 m from the terminus of the marine riser. At this particular locale, notable fluctuations were observed in the gradient data obtained from all 60 sensors. It is worth noting that each riser is equipped with a girth weld positioned at a distance of 4 m from its pin end, which is intuitively illustrated in the accompanying photograph of the riser. This indicates that the magnetic memory testing method can effectively detect girth welds in risers.

3.2. Intelligent Classification of Defect Images Based on SK-ConvNeXt-KAN

As previously discussed, we can utilize gradient imaging algorithms to convert magnetic memory signals into pseudo-color images of pipe wall defects and introduce intelligent image classification methods based on deep learning to eliminate the reliance on manual inspections. However, although these pseudo-color images resemble real photographs, they are not natural images and suffer from inter-class ambiguity and data imbalance. Inter-class ambiguity refers to the potential similarity in signal characteristics among different defect types, leading to unclear class boundaries, which increases the difficulty for the model to distinguish between different defects, leading to misclassifications. The issue of data imbalance is particularly pronounced in this study, as we utilize real-world field detection data, where the data acquisition method directly results in a severe lack of samples for certain rare defect types, leading to a notable imbalance in data distribution across different categories within the dataset. This causes the model to be biased towards the more frequently occurring classes during training, affecting its accuracy and reliability. Additionally, a major challenge currently faced in magnetic memory testing is the need for publicly available datasets, while various factors showing significant constraints limit the existing sample data. This scarcity of data increases the risk of model overfitting and may severely hinder the effective convergence of deep learning models. Consequently, this situation can adversely affect classification models’ accuracy and generalization ability, thereby limiting their efficiency and reliability in practical applications.

In response to these difficulties, this study innovatively proposes a new convolutional neural network called SK-ConvNeXt-KAN. This model builds upon ConvNeXt and significantly enhances the classification capability for drilling riser defects by incorporating an SK attention module into the input layer and replacing the original linear classification layer with a KAN module.

3.2.1. ConvNeXt Network

The ConvNeXt network made its first public appearance in a paper published by the Facebook AI team in 2022 [31]. This innovative network design deeply incorporates advanced design strategies from the Swin Transformer, comprehensively improving and optimizing the architecture of ResNet. The ConvNeXt network has successfully achieved or even surpassed the performance of Vision Transformer across multiple vision tasks, including image classification, object detection, and semantic segmentation, thereby reigniting interest in convolutional neural networks within the research community. The detailed architectural specifications of ResNet-50 and ConvNeXt-T are presented in

Table 2. Detailed architecture specifications for ResNet-50 and ConvNeXt-T.

	Output Size	ResNet-50	ConvNeXt-T
stem	56 × 56	7 × 7, 64, stride 2 3 × 3 max pool, stride 2	4 × 4, 96, stride 4
Stage1	28 × 28	$\left[\begin{array}{l}1\times 1, 64\\ 3\times 3, 64\\ 1\times 1, 256\end{array}\right]\times 3$	$\left[\begin{array}{l}\mathrm{d}7\times 7, 96\\ 1\times 1, 384\\ 1\times 1, 96\end{array}\right]\times 3$
Stage2	14 × 14	$\left[\begin{array}{l}1\times 1, 128\\ 3\times 3, 128\\ 1\times 1, 512\end{array}\right]\times 4$	$\left[\begin{array}{l}\mathrm{d}7\times 1, 192\\ 1\times 1, 768\\ 1\times 1, 192\end{array}\right]\times 3$
Stage3	56 × 56	$\left[\begin{array}{l}1\times 1, 256\\ 3\times 3, 256\\ 1\times 1, 1024\end{array}\right]\times 6$	$\left[\begin{array}{l}\mathrm{d}7\times 7, 384\\ 1\times 1, 1536\\ 1\times 1, 384\end{array}\right]\times 9$
Stage4	7 × 7	$\left[\begin{array}{l}1\times 1, 512\\ 3\times 3, 512\\ 1\times 1, 2048\end{array}\right]\times 3$	$\left[\begin{array}{l}\mathrm{d}7\times 7, 768\\ 1\times 1, 3072\\ 1\times 1, 768\end{array}\right]\times 3$
FLOPS		$4.1\times {10}^9$	$4.5\times {10}^9$

.

• Macro design

Based on the Swin Transformer model, ConvNeXt revises the proportion of block numbers in the four stages of ResNet-50 from 3:4:6:3 to 3:3:9:3. This adjustment aims to better adapt to the feature extraction requirements of different stages and enhance the model’s performance. Additionally, when down-sampling the feature map in the Stem layer, ConvNeXt adopts the same strategy as the Swin Transformer, using a convolution kernel with a stride of 4 and a size of 4 × 4. This operation is similar to the patch partitioning in the Transformer and helps the model process image data better. Table 2 provides a detailed comparison of the structures between the two models.

• Micro design

ConvNeXt innovatively adopts the GELU activation function, replacing the traditional ReLU, and reduces the usage of activation functions and regularization layers, thereby decreasing computational load. Inspired by the Transformer model, it substitutes batch normalization (BN) with layer normalization (LN), ensuring that the model remains stable during training across different batch sizes and enhancing its generalization ability. Furthermore, ConvNeXt separates the down-sampling from the bottleneck in ResNet, employing a dedicated down-sampling layer. This design facilitates a stable training process and improves the model’s generalization ability. The specific details are presented in Figure 9.

Additionally, ConvNeXt introduces parallel network layers and grouped convolutions inspired by the ResNeXt network to balance model size and performance well. The network employs a reverse bottleneck layer structure, similar to that of MobileNet-V2, where convolution operations are performed in the order of dimension increase, depth-wise convolution, and then dimension reduction. It also utilizes large 7 × 7 convolution kernels to increase the model’s receptive field, enabling each convolution operation to capture a broader range of contextual information. These strategies enable ConvNeXt to achieve excellent performance in various vision tasks.

3.2.2. Selective Kernel Attention

Selective Kernel Attention (SK Attention) is an advanced attention mechanism applied in the field of deep learning, particularly within Convolutional Neural Networks (CNNs), which was designed to enhance model performance significantly [32]. Its core design philosophy is deeply inspired by the neurons in the visual cortex, revolving around the dynamic selection of different sizes of convolution kernels. This enables the network to adjust its receptive field adaptively according to varying inputs. Such a unique adaptive mechanism allows the model to capture and utilize multi-scale features more effectively. Advanced models such as YOLOv5 and YOLOv7 have incorporated the SK Attention mechanism to augment their detection capabilities. The specific architecture of the SK network, as illustrated in Figure 10, primarily consists of three vital operational steps: Split, Fuse, and Select.

• Split

This operation involves convolving the input features with multiple convolution kernels of different sizes (each convolution operation corresponds to a set of CBR). This process generates feature representations at multiple scales, which are then concatenated to capture multi-scale feature information.

• Fuse

This step integrates fully connected layers, global average pooling, and the ReLU function. It generates K-scale channel descriptors through summation, compression, and dimension reduction or expansion transformations. After reshaping and stacking, softmax normalization is applied to the weights of each scale, enabling the efficient fusion of multi-scale features. Based on global information, this process guides the selection operation in the next step.

• Select

Based on the scale weights generated by the Fuse operation, this module dynamically selects and fuses features at different scales. This procedure ensures that the network can focus on the most critical feature scales for the current processing task, optimizing overall effectiveness and performance.

Existing research has demonstrated that embedding the SK module can effectively improve the accuracy of the original model; for instance, models such as YOLOv5 and YOLOv7 presented significantly enhanced detection capabilities when integrating the SK Attention mechanism.

3.2.3. Kolmogorov–Arnold Network

The Kolmogorov–Arnold Network (KAN) is a neural network architecture deeply inspired by the Kolmogorov–Arnold Representation Theorem [33]. This theorem indicates that any multivariate continuous function can be expressed as a composition of a finite number of univariate continuous functions. Based on this principle, KANs approximate multivariate functions through a two- or multi-layer network structure, thereby enhancing the accuracy and interpretability of the network. The formula is shown in Equation (2).

$$f(x)=f(x_1,\cdots ,x_n)={\displaystyle \sum _{q=1}^{2n+1}{\Phi }_q\left({\displaystyle \sum _{p=1}^n{\varphi }_{q,p}(x_p)}\right)}$$

(2)

KANs innovatively apply learnable activation functions to the edges of the network rather than the traditional nodes. This design greatly enhances the flexibility and expressive power of the model, enabling it to adapt more effectively to complex patterns and relationships in data, thereby significantly improving the model’s expressive ability and accuracy. Compared with traditional Multi-Layer Perceptions (MLPs), KANs can achieve similar performance levels with fewer parameters. Figure 11 illustrates the architecture of a shallow KAN network.

In convolutional classification networks, a linear classification layer is commonly employed to map the features extracted by the convolutional layers onto the final category labels. However, this linear mapping may need to adequately capture the intricate relationships among features, potentially affecting the classification performance. We propose using a KAN to replace the traditional linear classification layer to surmount this limitation. With its unique nonlinear activation functions and excellent approximation capabilities, KANs can more effectively capture and utilize the complex relationships between features, ultimately enhancing the accuracy and robustness of classification.

3.2.4. SK-ConvNeXt-KAN Network

As shown in Figure 12, the network proposed in this study is based on the ConvNeXt model, with the innovative integration of the SK (Selective Kernel) attention module in its initial input layer. This module can adaptively select different scale convolution kernels to extract and utilize critical information in images more effectively. By applying the SK Attention mechanism, the model can capture multi-scale features in the early stages of image processing, which facilitates a more comprehensive understanding of the input image content. Furthermore, the traditional Linear Layer in the network’s last layer is replaced with the KAN as the classification layer. Considering the KAN’s learning activation function and flexible representation, it is expected to achieve more precise and robust image classification results.

4. Experiment Results and Discussion

4.1. Experimental Environment

In this article, we constructed a deep learning framework based on PyTorch and used Python as the programming language to perform the training and testing processes. The hardware platform utilized for these tasks was a Dell laptop equipped with an Intel^® Core™ i5-13500H CPU, NVIDIA GeForce RTX 3060 GPU, and 32 GB of RAM. The detailed configuration of the hardware platform and training options are presented in

Table 3. Experimental configuration and training options.

Experimental	Configuration	Training Options	Value
Deep Learning Frameworks	PyTorch 1.10.0	Optimizer	AdamW
Programming Language	Python 3.8	Loss Function	Cross-Entropy
CPU	Intel® Core (TM) I5-13500H	Epoch	100
GPU	NVIDIA GeForce RTX 3060 (6 GB)	Batch size	16
RAM	32 GB	Initial Learning Rate	0.0005

.

4.2. Dataset and Evaluation Metrics

In this study, we conducted two-dimensional visualization processing on the on-site inspection data (see Section 3.1.4 for specific details). Through manual comparison and verification, we carefully selected 1100 representative defect images to construct an image dataset. To facilitate subsequent data processing, all images in the dataset were resized to 224 × 224 pixels. Figure 12 showcases partial samples of various types of defect images from this dataset. During training, the dataset was split into training and validation sets at a ratio of 8:2, with a fixed random seed to ensure reproducibility.

As shown in Figure 13, the dataset was divided into five categories, representing common defect types of risers: pitting corrosion, groove loss, longitudinal wear, irregular flaw, and girth weld. Details of the sample labels and the number of samples in each category are listed in

Table 4. Details of image set.

No.	Label	Defect Types	Number	Proportion
1	Pit	Pitting	270	24.3%
2	Gls	Groove Loss	148	13.3%
3	Lgw	Longitudinal Wear	252	22.7%
4	Ire	Irregular Flaw	141	12.7%
5	Gwd	Girth Weld	299	26.9%
Total			1110	100%

.

To evaluate the effectiveness of the network in classifying defect images, metrics including accuracy, precision, recall, and F1 score were employed to assess the model’s performance. The corresponding calculation formulas are provided in Equations (3)–(6), respectively [34].

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(3)

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(4)

$$Accuracy={\displaystyle \frac{TP+FN}{TP+FN+FP+FN}}$$

(5)

$$\mathrm{F}1-\mathrm{Score}=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}}$$

(6)

It should be noted that the formulas discussed above apply to binary classification problems, whereas the research in this paper involves a multi-classification problem. Therefore, when comparing the performance metrics of models, we calculated the precision, recall, and F1 score for each class separately and then averaged these metrics to assess the model’s overall performance.

4.3. Comparative Experiment

To successfully introduce deep learning-based image classification technology into riser defect detection, we initially adopted several mainstream classification models for training and testing, including GoogLeNet, ResNet-50, and ViT. We maintained the same hyperparameter settings during training to ensure fair evaluation and selection of the best-performing model for further optimization. Ultimately, we chose the ConvNeXt network as the basis and proposed the SK-ConvNeXt-KAN model in this paper. The section will present the results of systematic comparative experiments using the classification algorithm proposed in this paper and several advanced deep learning models.

Table 5. Comparative experimental results of different models.

Model	Training Accuracy (%)	Validation Accuracy (%)
GoogLeNet	97.6	83.4
ResNet-50	94.6	90.7
ViT	90.4	90.4
Swin-T	84.1	92.3
ConvNeXt	91.2	91.8
SK-ConvNeXt-KAN	94.5	95.4

summarizes the final accuracies achieved by each model on the training and validation sets.

As GoogLeNet failed to converge fully within 100 training epochs, we adopted a transfer learning approach by importing pre-trained weights to facilitate the training process. As shown in Figure 14a, due to the introduction of pre-trained weights, GoogLeNet experienced a sharp drop in training loss at the initial stage of training and achieved high accuracy on the training set. However, its accuracy on the validation set remained significantly lower than on the training set, ultimately reaching only 83.4%. This indicates that the model suffers from severe overfitting on our dataset, potentially as the training weights for transfer learning were obtained through testing on natural image datasets, which are unsuitable for the pseudo-color images generated in this study. Similarly, the ResNet-50 network also encountered overfitting issues during training, as depicted in Figure 14b. Although it achieved a high accuracy of 94.6% on the training set, its accuracy on the validation set was only 90.7%. Furthermore, the training loss value fluctuated significantly during the last 20 epochs, suggesting that its training stability needs to be improved. The training process of the ViT network is shown in Figure 14c. Although the model trained was relatively stable, the training and validation accuracy plateaued after 60 epochs, reaching 90.4%. Figure 14d displays the training curve of the Swin-T model, from which it can be observed that, after 40 epochs of training, the model’s training loss had difficulty further reducing. The validation set accuracy (92.3%) was significantly higher than the training set accuracy (84.1%), indicating some underfitting.

In contrast, the ConvNeXt network demonstrated outstanding performance during training. As shown in Figure 14e, its loss steadily decreased while its accuracy on both the training and validation sets increased synchronously. Ultimately, the ConvNeXt network achieved accuracies of 91.2% on the training set and 91.8% on the validation set. This performance validated the ConvNeXt network’s good adaptability to the dataset used in this study, which is why it was chosen as the foundation for further optimization. Furthermore, the SK-ConvNeXt-KAN model proposed in this paper not only retained the stability advantages of the original network in training but also enhanced the accuracies on the training and validation sets to 94.5% and 95.4%, respectively, as shown in Figure 14f. This indicates that the model possesses exceptional generalization capabilities for magnetic memory defect images of offshore risers, highlighting its significant potential for application in classifying defect images in deep-sea drilling riser magnetic memory inspections.

4.4. Ablation Study

To verify the effectiveness of the SK attention and the KAN module, we selected ConvNeXt-t as the base architecture and set it as the baseline model. On this basis, we constructed three variant networks: SK-ConvNeXt, ConvNeXt-KAN, and SK-ConvNeXt-KAN. Using the magnetic memory detection dataset proposed in this paper, we conducted a series of detailed ablation experiments to deeply explore and verify the specific impacts of these modules on network performance.

• SK-ConvNeXt

This network embeds the SK attention module into the input layer of the model while keeping the rest of the structure unchanged, aiming to enable the network to capture and process richer multi-scale feature information at the early stage, thereby enhancing the model’s ability to process more complex data.

• ConvNeXt-KAN

In this variant, we replace the Linear Layer of the base network’s output layer with a KAN. This strategy aims to leverage the nonlinear processing capabilities of the KAN, enabling the model to better adapt to different data distributions and task scenarios, thereby enhancing the generalization ability of the entire model.

• SK-ConvNeXt-KAN

The network combines the above two strategies: embedding the SK attention module into the input layer while replacing the Linear Layer of the output layer with the KAN. This combination aims to further improve the network’s overall performance by capturing richer feature information and enhancing the nonlinear processing capabilities of the model, thereby achieving higher accuracy and more vital generalization ability.

Table 6. Ablation study results.

Model	Accuracy (%)	Precision (%)	Recall (%)	F1 Score
Baseline	91.8	88.43	88.55	0.884
SK-ConvNeXt	93.6	90.86	90.57	0.907
ConvNeXt-KAN	92.7	90.03	89.13	0.895
SK-ConvNeXt-KAN	95.4	91.68	91.88	0.917

presents the training outcomes for the four distinct models. Compared to the baseline model, integrating the SK attention module led to a notable increase in accuracy to 93.6%. Incorporating the KAN module also elevated the accuracy to 92.7%. These findings indicate that both modules effectively bolstered the network’s capability to extract image features and learn classification rules. Notably, the proposed SK-ConvNeXt-KAN model achieved a remarkable 95.4% accuracy, significantly improving the F1 score from 0.884 to 0.917, underscoring its superior performance.

As shown in Figure 15, introducing the SK attention and KAN modules reduced the training loss of the ConvNeXt network and further enhanced its accuracy. When these two modules were seamlessly integrated into the ConvNeXt network, resulting in the SK-ConvNeXt-KAN model proposed in this study, their synergistic effects at the input and output layers remarkably bolstered the model’s classification performance. Figure 15a depicts that the SK-ConvNeXt-KAN model distinctly surpassed the other models, achieving a consistently higher classification accuracy from 20 training iterations.

In the context of our study, Figure 16 provides a compelling visualization of the model’s discrimination prowess across different defect categories. Specifically, the proposed model demonstrates remarkable accuracy, with a mere five misclassifications observed for the “Pitting” defect and three for the “Longitudinal Wear” defect. Furthermore, it achieves a near-flawless performance for all remaining defect types, underscoring its robustness and precision in defect recognition. This outcome underscores the efficacy of our approach and its potential for practical applications.

4.5. Generalization Ability Test

4.5.1. Fivefold Cross-Validation

In machine learning, an inherent issue with small-scale datasets is that the evaluation results may be influenced by the data partitioning method, as a single random split of the training/testing sets may not fully reflect the model’s performance [35]. Given the relatively small size of the self-constructed dataset in this paper, we adopted the fivefold cross-validation method to validate the proposed algorithm’s stability and reliability.

Initially, we thoroughly shuffled all the images in the dataset to ensure a uniform distribution of samples. Then, the dataset was divided into five equal subsets, each mutually exclusive and containing a balanced number of images from each category, as shown in Figure 17. In each experiment, we select a specific subset as an independent validation set to evaluate the model’s performance immediately while the remaining subsets are combined into a training set. This process is repeated five times, with a different subset chosen as the validation set each time, ensuring that every image in the dataset can be fully learned during the training phase and tested in the validation phase. We can deeply analyze the algorithm’s performance under different data distributions through the fivefold cross-validation method, leading to a more balanced and comprehensive evaluation of the model’s performance. The results of the cross-validation are compared in

Table 7. Results of fivefold cross-validation.

Experiment	Accuracy (%)	Precision (%)	Recall (%)	F1 Score
Fold-1	94.55	91.55	91.53	0.915
Fold-2	95.45	93.66	92.50	0.930
Fold-3	95.00	92.70	92.54	0.926
Fold-4	94.09	91.01	91.13	0.910
Fold-5	95.91	93.93	92.61	0.932
Average	95.00	92.57	92.06	0.923

.

The experimental results show that the accuracy of the proposed model remained consistently around 95% across different folds, aligning with the test results presented in the previous section. This consistency validates the model’s stability and highlights its robust generalization capability across varying data distributions. Moreover, the model presented a good balance between precision and recall, suggesting that it has achieved an equilibrium in accurately identifying positive instances and comprehensively covering them.

4.5.2. Cross-Dataset Validation

To further validate the generalization capability of the algorithm presented in this study, we chose the publicly available NEU DET dataset for cross-dataset validation experiments. The NEU DET dataset, collected by Northeastern University specifically for steel surface defect detection, comprises images of six common types of steel surface defects totaling 1800 [36]. It has been widely utilized in industrial computer vision-based inspections [37]. Importantly, this dataset is highly compatible with our custom-built dataset in terms of both application scenarios and image characteristics, and its larger scale makes it an excellent benchmark for assessing the performance of our algorithm.

The NEU DET dataset contains images with a resolution of 300 × 300 pixels, spanning six categories, with the specific number of images per category and detailed descriptions outlined in

Table 8. Details of NEU DET.

No.	Label	Defect Types	Train_Num	Validate_Num
1	In	Inclusions	240	60
2	Pa	Patches	240	60
3	Cr	Cracks	240	60
4	PS	Pitted Surfaces	240	60
5	RS	Rolled-in Scale	240	60
6	Sc	Scratches	240	60
Total			1800

. To train and test the SK-ConvNeXt-KAN model proposed in this paper using the NEU DET dataset, we resized the input images to 224 × 224 pixels and modified the output layer to include six categories corresponding to the six types of defects present in the dataset. Additionally, we preserved the model’s original architecture and hyperparameter settings to ensure the validation process’s fairness and accuracy. As depicted in Figure 18a, the model showcased excellent learning efficiency throughout the training process, with a consistent decrease in training loss and a concurrent increase in accuracy on both the training and validation sets. The model stabilized after approximately 80 iterations. At the end of the training period, the accuracy of the training set reached 97.71%, while an even higher accuracy of 97.5% was achieved on the validation set. Moreover, the training loss was significantly reduced to 0.071, illustrating the model’s robust fit to the data.

Furthermore, we can intuitively assess the model’s performance through the confusion matrix, as shown in Figure 18b. In the validation set containing 360 test images, the model misclassified or incorrectly judged only a negligible number of images (9), indicating its extremely high accuracy and robustness in identifying various steel surface defects. This result is consistent with our experimental findings on the self-constructed dataset, underscoring the remarkable cross-dataset generalization capability of the proposed SK-ConvNeXt-KAN model.

5. Conclusions

This study focuses on the field of nondestructive testing of deepwater drilling risers, achieving the following results and conclusions: First, we innovatively developed a riser inspection robot based on magnetic memory technology, which accomplishes precise in situ detection of deepwater drilling risers, fundamentally overcoming the inefficiencies and complexities of traditional detection methods, marking a significant leap in deepwater drilling inspection technology. Second, the inspection robot demonstrated high stability and reliability in field applications, completing inspection tasks and collecting a large amount of detection data. Utilizing these data, we established an image database of typical riser defects, laying a solid foundation for subsequent defect identification and classification research. Finally, we proposed a novel image classification model, SK-ConvNeXt-KAN, which significantly enhances the model’s feature extraction capabilities and its ability to handle nonlinear problems by integrating SK attention modules and KAN networks into the ConvNeXt convolutional neural network. Experimental results show that this model achieves an accuracy rate of 95.4% in identifying defects such as corrosion, cracks, and wear on the inner walls of the riser.

In summary, through the deep integration of hardware and software innovations, this study has successfully applied artificial intelligence and automation technologies to detect deepwater drilling risers, significantly enhancing industrial detection efficiency and markedly improving the accuracy of defect identification. Looking forward, we plan to introduce object detection models further to achieve real-time detection and precise localization of defects in risers, thereby providing more robust technical support and assurance for the in-depth implementation of deepwater oil and gas exploration and the industry’s sustainable development.