Deep learning based gasket fault detection: a CNN approach | Scientific Reports

Mar 07, 2025

Scientific Reports volume 15, Article number: 4776 (2025) Cite this article

481 Accesses

Metrics details

Gasket inspection is a critical step in the quality control of a product. The proposed method automates the detection of misaligned or incorrectly fitting gaskets, ensuring timely repair action. The suggested method uses deep learning approaches to recognize and evaluate radiator images, with a focus on identifying misaligned or incorrectly installed gaskets. Deep learning algorithms are specific for feature extraction and classification together with a convolutional neural network (CNN) module that allows for seamless connection. A gasket inspection system based on a CNN architecture is developed in this work. The system consists of two sets of convolution layers, followed by two sets of batch normalization layer, two sets of RELU layer, max pooling layer and finally fully connected layer for classification of gasket images. The obtained results indicate that our system has great potential for practical applications in the manufacturing industry. Moreover, our system provides a reliable and efficient mechanism for quality control, which can help reduce the risk of defects and ensure product reliability.

The industrial sector is always evolving, with corporations looking for new ways to increase productivity. Quality control is critical in production, and product failures can cause significant losses1. The manufacture of the radiator, in particular, requires faultless gaskets for optimal performance. Thus, creating a strong quality control process to detect any misaligned or incorrectly fitting gaskets is critical. This research addresses the issue by creating an image processing based deep learning approach for gasket checking in radiators2,3.

Radiator.

A radiator is a heat exchanger device used in various applications as mentioned in Fig. 1, most commonly in vehicles, to transfer the heat in any one medium (usually a liquid coolant) to another medium (usually air) to cool the engine or other components. The radiator performs a critical position in keeping the premiere temperature of the engine. As the engine operates, it generates a substantial amount of heat, which wishes to be dissipated to save you overheating. The radiator is designed to facilitate this warmness transfer technique. It includes a network of thin tubes, often made of aluminum, that are connected to a chain of fins. The mixture of antifreeze and water, called coolant, flows through these tubes, absorbing the warmth generated by using the engine. As the coolant travels thru the radiator, cool air passes via the fins, inflicting the heat to be transferred from the coolant to the air. This process cools down the coolant, that’s then circulated back to the engine to soak up greater warmth. The radiator also contains a fan that facilitates in growing airflow through the fins, specifically while the vehicle is stationary or operating at low speeds. Overall, the radiator ensures that the engine operates inside the premier temperature range, stopping damage due to overheating and making an allowance for efficient combustion.

Gasket.

A gasket, depicted in Fig. 2, serves as a vital mechanical seal employed to bridge the space between two or more adjoining surfaces. Its primary function is to thwart the escape of fluids or gases when subjected to varying pressures or temperatures, thereby ensuring a tight and secure seal. It acts as a barrier, creating a tight and reliable seal between the joined components. Gaskets are usually made of soft materials such as rubber, silicone, and metal. It is designed to match the irregularities of the mating surface, providing a strong and reliable seal. The choice of materials depends on the specific application, with rubber gaskets for low-pressure applications, and metallic gaskets for high-pressure or high-temperature environments commonly used Gasket is widely used in a variety of industries such as automotive, aviation, plumbing, and construction They play an important role in preventing leaks, reducing noise and vibration, and ensuring proper operation of machinery and equipment.

Properly fixed radiator.

Furthermore, a well fixed gasket also allows to prevent any infection of the coolant through external elements including dust or particles. It plays a crucial position as shown in Fig. 3 in retaining the strain inside the radiator device, ensuring a consistent and green flow of coolant. Without a well constant gasket, there is a hazard of coolant leakage, that may cause overheating of the engine and capability harm to its additives. Additionally, a defective gasket can result in decreased cooling efficiency, causing the engine to perform at better temperatures and probably leading to reduced overall performance and elevated fuel intake. Therefore, it is essential to frequently look into and replace the gasket if any symptoms of wear and tear or damage are detected to make certain the clean and reliable operation of the cooling system.

Fault Image.

A faulty fixed gasket as shown in the Fig. 4 in a radiator can create significant issues and compromises the reliability and efficiency of the cooling system. When a gasket is improperly fixed or not correctly aligned within the radiator, it fails to create a secure and tight seal between the mating surfaces of the radiator components. This can result in several detrimental effects. Furthermore, a well fixed gasket also allows to prevent any infection of the coolant through external elements including dust or particles. It plays a crucial position in retaining the strain inside the radiator device, ensuring a consistent and green flow of coolant. Without a well constant gasket, there is a hazard of coolant leakage that may cause overheating of the engine and capability harm to its additives. Additionally, a defective gasket can result in decreased cooling efficiency, causing the engine to perform at better temperatures and probably leading to reduced overall performance and elevated fuel intake. Therefore, it is essential to frequently look into and replace the gasket if any symptoms of wear and tear or damage are detected to make certain the clean and reliable operation of the cooling system.

The visual simplicity of Figs. 3 and 4 does not adequately convey the difficulties present in real-world applications, despite the comparison suggesting that fault classification is an easy operation. Gasket defect identification entails managing a variety of subtle and varied variances across several defect categories, including overlapping layers, rotations, deformations, and lateral misalignments, which may be difficult to discern under different lighting circumstances, angles, or gasket textures. The ability of a CNN to automatically extract hierarchical features like edges, forms, and textures that are essential for spotting minute fault patterns makes it very useful in this situation. CNNs may generalize well over a variety of datasets, in contrast to conventional machine vision or rule-based approaches, which may have trouble with variability and necessitate manual feature engineering. CNNs can generalize effectively across diverse datasets, ensuring consistent and robust performance.

The traditional image processing algorithm that is used in commonly in industrial sectors, are typically involves the following steps before starting the processing. First the image that needs to be processed must be obtained and saved in the appropriate directory. The algorithm then finds the white pixels in the image computes their non-zero (nnz) value and compares it with the given OK value to find any differences. Next the algorithm tackles the important job of figuring out where a gasket has to be placed in a radiator. The following methods can be employed to figure out the issues. Analysis of the images total number of white pixels is one method used for the work. The presence of the gasket is indicated by white pixels and the higher concentration of white pixels indicates a greater gasket coverage on the radiator.

For applications like document image binarization(DIB) or object detection in high contrast settings, white pixel analysis works well. For instance, it may separate text from the backdrop by examining the density of white pixels to extract text sections. This approach uses less computing power when the foreground-background contrast is distinct. But it has trouble with complicated backgrounds, necessitating sophisticated methods like machine learning or edge detection. Calculating non-zero elements which correspond to the white pixels of images is made possible by using Matlabs nnz function.

The calculated count is then cross-referenced with an initial value that represents the anticipated count for the radiator component that is correctly fitted this is referred to as the OK value. This comparison helps determine whether the gasket has been fixed correctly. The gasket is considered to be firmly in place if the amount of white pixels is within a reasonable range around the OK value. On the other hand the gasket is regarded as being incorrectly fixed if the count rises above the designated threshold. Lastly the program produces an output based on the results of Step 2 in the algorithms final step. A notice verifying the accuracy of the gasket is shown if the number of white pixels is within the permitted range. In contrast a notice indicating a problem with the gasket is generated if the number of white pixels differs from the predetermined range. This simple clear-cut method of processing images which is based on white pixel analysis is a basic but useful technique though it might not be sufficient for more complex tasks and situations.

Integrating deep learning into image processing4,5 with machine interfacing for gasket inspection dramatically improves detection accuracy by utilizing convolutional neural networks (CNNs). The combination mentioned above results in an efficient and fast gasket inspection system to detect the presence of gasket. The integration of the methods symbols an important development in exploring artificial intelligence to improve manufacturing processes. A number of different types of gasket misalignment faults can arise, such as lateral or rotational shifts, overlapping layers, incomplete seating, deformation, and obstruction by foreign materials. Each of these flaws presents subtle visual distinctions that make hand diagnosis laborious and prone to errors. Conventional machine vision systems find it difficult to handle these flaws because they are unable to handle subtle fluctuations, real-world irregularities like lighting or angles, and the requirement for reliable feature extraction. By automatically learning hierarchical features, identifying minute differences, and generalizing over complicated data with high variability, Convolutional Neural Networks (CNNs) overcome these difficulties. CNNs are efficient for automating the detection of various gasket misalignments in industrial applications because of their scalability and capacity to get better results with larger datasets.

The field of defect detection in engineering products has undergone a transformation thanks to Convolutional Neural Networks (CNNs) one of the key techniques in deep learning. These cutting-edge methods will undoubtedly surpass the conventional detection-based systems resulting in noticeably improved accuracy rates. One major benefit of deep learning for fault identification is its capacity to automatically identify complex patterns and features from large datasets without the need for human feature engineering.

CNNs are particularly good at removing hierarchical features from images which makes them a good choice for tasks involving visual inspection. Combining a feature extraction encoder and a dense prediction decoder the encoder-decoder structure is one of the most widely used deep learning architectures for defect detection. Because of this the model is able to precisely localize and segment defects by capturing both local and global data. Researchers have looked into strategies like feature fusion anchor optimization and attention mechanisms to boost the performance of deep learning models even more. Through effective integration of data from multiple network layers and focus on the most relevant aspects the techniques help the model function more efficiently. Although the results are promising there are challenges in applying deep learning to defect detection. One of the main challenges is the requirement for large annotated datasets for training which can be costly and time-consuming to obtain.

Higher Accuracy: Defect detection accuracy is much higher with deep learning models especially with Convolutional Neural Networks (CNNs). As an example research findings indicate that silicone rubber gaskets can be identified with up to good accuracy when compared to conventional based methods which generally yield lower accuracy rates.

Automated Feature Extraction: Deep learning algorithms automatically will identify and extract prominent features from images in contrast to traditional methods which frequently rely on human feature selection and predetermined rules. This capability can detect more and complex defects that traditional techniques might miss to analysis.

Scalability and Adaptability: Without requiring any significant reprogramming deep learning systems can be instantly scaled to handle massive volumes of data and can adapt to new kinds of defects. This adaptability is essential in dynamic manufacturing settings in which the variety in defect types and product projects is common.

Real-Time Processing: During the manufacturing process real-time inspection is made possible by the ability of deep learning models to process images rapidly. This speed is necessary to minimize downtime which is frequently a drawback of conventional inspection techniques and to maintain production efficiency.

Robustness to Variability: Deep learning techniques are more resilient to changes in illumination viewpoint and other environmental parameters making them suitable for a variety of inspection scenarios. This resilience aids in preserving steady performance over various production lines and batches.

Data augmentation is a common technique used in deep learning methods to improve the training dataset and increase the models capacity to identify flaws in new images and generalize. In comparison traditional approaches might not make good use of these strategies.

Deep learning is generally incorporated into gasket inspection procedures to produce more accurate dependable and efficient quality control greatly improving the manufacturing process.

Specialized deep learning models named convolutional neural networks (CNNs)6 are made for processing visual data, particularly images. Convolutional layers are employed, which employ filters to extract features such as textures and edges from the input data. Because of its architecture, CNNs can automatically learn about the spatial patterns, which makes them useful for varied choice of tasks like speech identification and natural language processing. Convolutional layers, integral to neural networks, excel at processing spatial data like images by employing filters for feature extraction.

The important aids of the study are CNN layers were used to classify radiator images, Extracting features for discerning proper gasket alignment.

Deep Learning Techniques have absolutely shown various flexibility by seamlessly integrating into a wide array of disciplines, demonstrating their adaptability and potential to revolutionize various fields. The proposed system of image processing involving machine learning combination with a gasket find this adaptability, aiming to streamline and enhance the image recognition and analysis tasks by exploring the capabilities of various Machine Learning algorithms. This interactive system represents a cutting-edge method that grasps potential for automating complex processes and advancing the capabilities of image analysis tools. By delving into a comprehensive literature survey focusing on machines and machine learning algorithms, researchers appreciated visions into the latest developments and trends in this quickly developing field. Through an exploration of literature, researchers can stay well-informed regarding the most recent breakthroughs and methodologies, informing their own research endeavors and contributing to the continuous enhancement and refinement of machine learning applications in diverse settings. Such a thorough examination not only excavates considerate but also fuels innovation and fosters collaboration within the research community, catalyzing further advancements and discoveries. In this way, the interconnected nature of research and technological progress underscores the importance of staying informed and engaged with the latest findings and advancements to drive continuous improvement and innovation in the field of machine learning and its applications across various domains.

Machine Learning and Deep Learning Techniques, are more multipurpose in their application, can be flawlessly combined with virtually any field, transcending boundaries and revolutionizing traditional practices. An advanced approach within this realm is the proposed system of image processing with machine interfacing using a gasket, showing the control of mechanization and improvement in image recognition and analysis tasks through the utilization of sophisticated Machine Learning algorithms. This leading-edge system not only rationalizes processes but also meaningfully boosts accuracy, efficiency and reliability in handling complex data sets.

Henceforth, the innovative addition of gasket fault analysis and deep learning algorithms stands at the front of research endeavours in the particular domain, importance the interaction between progressive skills and fault detection methodologies. As the literature survey investigates into the shades of gasket fault account and the complicated workings of deep learning algorithms, a clearer understanding appears of the possible aids and experimentations related with such integration. By travelling the crossing tracks of fault detection and algorithmic intelligence, researchers are paving the way for revolutionary developments in fault diagnosis and predictive conservation strategies.

The interface among methods for image processing and machine learning frameworks not only underlines the increasing horizons of technical invention but also underlines the critical role of interdisciplinary collaboration in driving progress and fostering innovation. As researchers and practitioners similar dig deeper into the multi-layered applications of these techniques, the possible for transformative impact indirect various fields becomes increasingly deceiving, placing the substance for a forthcoming where smart systems flawlessly interact with their environments to tackle multifaceted trials and drive meaningful change. In the following area section, the literature survey related to gasket fault and deep learning algorithms is discussed.

The key findings of the paper is that it emphasizes that detecting defects is crucial for quality control in manufacturing, as defects can lead to increased costs and reduced product lifespan7. It categorizes various types of defects found in products, including electronic components, pipes, welded parts, and textiles, highlighting the diverse nature of manufacturing defects. The authors review traditional and methods of deep learning used for defect detection, discussing their characteristics, strengths, and weaknesses. Traditional methods include magnetic particle testing, ultrasonic testing and machine vision, while deep learning methods focus on image processing. The paper outlines the rapid development of deep learning technologies, particularly Convolutional Neural Networks (CNNs), which have shown significant success in object detection and defect identification. It identifies ongoing challenges in the field, in accordance with the need for high precision, rapid detection capabilities, and effective handling of complex backgrounds and occluded objects. The authors propose future research directions to address the challenges identified, aiming to enhance the effectiveness and efficiency of defect detection in manufacturing. Overall, the paper serves as a comprehensive resource for understanding the state-of-the-art in deep learning applications for defect detection in manufacturing while highlighting ongoing challenges and future research needs.

In order to solve the issue of plastic gaskets’ multiple surface flaws and challenges in feature extraction and classification, the GoogLeNet Inception-V2 deep convolutional neural network (DCNN) was used3. To prevent model overfitting in deep learning applications, a lot of training data is needed, however there aren’t many datasets with plastic gasket flaws. Our dataset underwent data augmentation in order to resolve this problem. Ultimately, the contrast of the three convolutional neural networks’ performances was conducted. The GoogLeNet Inception-V2 transfer learning model performed good in less time, according to the results. It implies that the study utilised here, it had better accuracy, dependability, and efficiency.

Electrical devices and drive systems are now indispensable in a variety of applications as discussed in this paper8. Over time, different types of failures arise from continuous, long-term operation. Owing to the growing impact of these technologies on various industries, industrial sectors, and everyday human existence, condition monitoring and prompt fault diagnoses have become reasonably significant. This review article examines various diagnostic procedures that can be applied to the training of algorithms and the implementation of predictive maintenance. The advantages and disadvantages of sophisticated diagnostic methods are emphasized. The most common electrical machine failures are covered, and methods for monitoring parameters are presented.

Nonetheless, identifying simultaneous failures remains difficult when distracting sounds are present or when several faults result in overlapping features9. Recently, multi-label classification has become more well-liked across a range of application areas as a productive technique for defect identification and system monitoring that yields encouraging outcomes. This work contributes by offering a novel multi-label classification algorithm for concurrently diagnosing numerous faults and assessing the problem severity in noisy environments. Both conventional vibration data and the Electrical Signature Analysis have been taken into consideration for modeling in this study. Additionally, a comparison is made between the performances of different multi-label categorization models. Signals related to vibration and current are recorded in both fault and normal circumstances. The applicability of the proposed method is experimentally validated under diverse fault conditions such as unbalance and misalignment.

This research proposes a novel artificial intelligence (AI) technique for the unsupervised defect diagnosis of a gear transmission chain using a deep belief network (DBN) and a genetic algorithm to optimize the structural parameters of the network10. Furthermore, it is more adept at modeling intricately structured data. The suggested method’s efficacy is verified through the utilization of rolling bearing and gearbox datasets. The performance of the suggested method is compared with two well-known classifiers, namely the support vector machine (SVM) and the back propagation neural network (BPNN), in order to demonstrate its superiority. With the suggested method, the fault classification accuracy is 100% for gearboxes and 99.26% for rolling bearings, which are significantly higher than the other two methods.

In order to assess the bearing flaws, three distinct classification algorithms—KNN, decision tree, and random forest—are finally trained and assessed utilizing these features11. This set of methods lowers computational complexity while improving accuracy. The findings of the experiment indicate that all three classifiers attain an accuracy rate over 97%. Furthermore, improved performance is indicated by the evaluation metrics, including specificity, sensitivity, F1-score, and accuracy. Lastly, the suggested model is contrasted with a few recently used methods to confirm its efficacy. The comparison results show how promising the recommended method is for IM bearing defect diagnostics.

This study reviews the diagnosis of problems in electric motors. With the use of long short-term memory neural networks, brushless direct-current (BLDC) motor stator fault diagnosis protocols were examined12. A highly accurate vibration data gathering approach utilizing cloud technologies was examined, along with features extracted using spectral entropy, instantaneous frequency standardization using mean and standard deviation, and feature extraction using spectral entropy. A comparison was also made between training models using raw and normalized data. A total of 97.10% of the model was found to be accurate. By using data from actual experiments, training the model, and testing it theoretically, the proposed methods could successfully identify the motor stator status from normal, to loss of stator winding imminent and arcing, and finally to open circuit in stator winding—motor needing to stop immediately.

This method improves the traditional convolutional neural network (CNN) model structure by introducing the global average pooling technology and SVM13. Firstly, the temporal and spatial multichannel raw data from multiple sensors is directly input into the improved CNN-Softmax model for the training of the CNN model. Secondly, the improved CNN are used for extracting representative features from the raw fault data. Finally, the extracted sparse representative feature vectors are input into SVM for fault classification. The proposed method is applied to the diagnosis multichannel vibration signal monitoring data of a rolling bearing. The results confirm that the proposed method is more effective than other existing intelligence diagnosis methods including SVM, K-nearest neighbour, back-propagation neural network, deep BP neural network, and traditional CNN.

First, continuous wavelet transform (CWT) is used to transform time domain vibration data into two-dimensional (2D) grayscale images with a wealth of fault information14. Second, a CNN model that is based on LeNet-5 is designed to automatically extract multi-level features from the images that are sensitive to defect detection. Lastly, the ensemble of multiple RF classifiers diagnoses bearing defects using the multi-level features that comprise both local and global information. Specifically, the diagnostic performance is enhanced by combining low-level data with correct details in the buried layers and local characteristics. The effectiveness of the proposed method is validated by two sets of bearing data collected from reliance electric motor and rolling mill, respectively. The experimental results indicate that the proposed method achieves high accuracy in bearing fault diagnosis under complex operational conditions and is superior to traditional methods and standard deep learning methods.

The study evaluates the suitability and performance of various classifiers for induction motor fault diagnosis using three classification algorithms: ensemble, K-nearest neighbours (KNN), and support vector machine (SVM), with a total of 17 different classifiers available in the MATLAB Classification Learner toolbox15. It is discovered that 12 classifiers perform poorly, whereas five classifiers—Fine Gaussian SVM, Fine KNN, Weighted KNN, Bagged trees, and Subspace KNN—can offer about 100% classification accuracy for all faults applied to each motor. To compute characteristics for motors when vibration signals or stator currents are not checked for a specific defect under certain loadings, a novel curve fitting technique is devised. The proposed fault diagnosis method can accurately detect single- or multi- electrical and mechanical faults in induction motors.

The cast-resin transformer’s normal and defective IRT images, which were obtained, using the suggested thermal camera monitoring method, are shown in this study16. The Wasserstein Autoencoder Reconstruction (WAR) and Differential Image Classification (DIC) models’ model training comes next. The differential picture can be obtained by computing the absolute difference, pixel by pixel, between the original and regenerated images. Ultimately, the trained WAR and DIC models are linked in series to create a defect diagnostic module during the testing phase. The experimental results show that the suggested model has significant advantages over the state-of-the-art deep learning methods. It is lightweight, requires less storage space, has a quick inference time, and has sufficient diagnostic accuracy.

To promote sustainable growth, an intelligent manufacturing process can be implemented to reduce human costs and eliminate faults17. In this research, we studied an autonomous defect detection model based on deep learning that can identify faults utilizing open sources and train product features to detect defects easily during the production process. This model’s performance was tested by applying it to the production of disposable gas lighters in order to identify the lighter’s liquefied gas volume defect. It was found that both the processing time and detection accuracy were enough for use in this process.

This research reviewed and classified recent computer vision works into the following categories: (a) analysis of seed quality; (b) analysis of soil; (c) management of irrigation water; (d) analysis of plant health; (e) management of weeds; (f) management of livestock; and (g) yield estimation18. Additionally, the study covers current developments in computer vision, including popular deep learning architectures, vision transformers (ViT), and generative adversarial networks (GAN). This study also identifies the difficulties in applying the answers in real-time in the farmer’s field. Overall, the results show that convolutional neural networks constitute the backbone of contemporary computer vision techniques, and that the different topologies of these networks offer high-quality, accurate, and precise solutions for a range of agricultural applications.

The primary rationale for this is because the caliber of input data used in neural network models can have a significant impact on the analysis’s findings19. According to Lou, if critical input data are distorted or missing, it can have a substantial impact on the neural network’s performance. Properly prepared data are generally easy to work with, which simplifies the effort of data analysis. However, badly produced data might make data analysis challenging, if not impossible. Data preparation has also become a time-consuming activity due to the growing volume of data produced by modern data collecting techniques and the data coming from various sources. It has been stated that data preparation accounts for between 50 and 70% of the time and effort spent on data analysis initiatives.

To increase their generalization performance, deep learning models require the efficient training of a large number of free parameters on an extensive amount of training data20. Actually, data collection and labelling are costly. One way to mitigate this issue is by data augmentation. In this research, we perform a preliminary investigation into the effects of three factors on deep learning accuracy for picture classification: augmentation method, augmentation rate, and quantity of basic dataset per label. The report offers several recommendations: (1) It is preferable to employ transformations that change the images’ geometry as opposed to those that only change the lighting and color. (2) For training, two to three times the augmentation rate is sufficient. (3) The more evident a contribution could be, the less data there is. Methods, Strengths, weaknesses and Applications of related works are shown in Table 1.

The proposed method employs a Convolutional Neural Network (CNN) to efficiently classify gasket images into two categories: OK and Not OK. This section details the various stages involved in the methodology, including data collection, pre-processing, data augmentation, CNN architecture, training procedures, and performance evaluation.

The proposed method employs a Convolutional Neural Network (CNN) as its primary tool to effectively classify gasket images into two distinct categories: OK and Not OK. Throughout this section, detailed insights will be provided into the numerous stages integral to the methodology. These stages encompass data collection, pre-processing methodologies, data augmentation techniques, the intricacies of CNN architecture utilized in the process, elaborate training procedures applied, as well as comprehensive performance evaluation procedures.

Flow Diagram for the proposed methodology.

Notably, the workflow of the proposed work, as depicted in Fig. 5, is an important visual aid providing details of the relevant processes clear and well organized. This visual aid helps facilitate a full understanding of the seamless operation of the system. It not only clarifies the difficult steps that have been taken but also describes the logical progression from one stage to the next. Moreover, the longitudinal separation given in the figure plays an important role in the interpretation of the distribution pattern.

The careful steps shown in the workflow underscore the rigorous approach taken to ensure accuracy and efficiency in the gasket inspection process. Combining state-of-the-art technologies such as Convolutional Neural Networks (CNNs) with sophisticated techniques such as data enhancement and comprehensive performance analysis, the proposed method offers innovative and reliable into the process costs. Each element of the manufacturing process is carefully designed and strategically placed to ensure the accuracy of distribution and to meet the required standards of quality.

Specifically, Fig. 5 provides the overall underlying theory supporting the proposed approach, which implies a harmonious integration of advanced technologies and processes. This integration aims to redefine the classification model for gasket models, establishing new parameters of accuracy and reliability.

To build an effective CNN model for gasket inspection, a comprehensive dataset of radiator images with both correctly and incorrectly installed gaskets was collected. The dataset gaining process involved thorough planning has to be done and then execution, focusing on obtaining high-quality images essential for training the model. By using the high-resolution industrial cameras placed deliberately to cover the gasket area widely while minimizing uncooperative elements, a diverse set of images was captured. It was vital to capture multiple images from variable angles and lighting conditions to ensure the model’s robustness and flexibility in real-world scenarios.

Subsequent data acquisition, the pre-processing steps were used to make ready the dataset for training the CNN model effectively. An essential step in this preparation was image normalization, which complicated resizing all images to a standardized dimension, such as 224 × 224 pixels. This resizing not only ensured steadiness in input size but also optimized the training process by providing uniformity in the dataset. Normalizing pixel values to a specific range (0 to 1) by dividing by 255 played a vital role in quickening model convergence during training, enhancing the competence and efficiency of the learning process.

Furthermore, the dataset underwent careful image annotation, where each image was exactly labelled as either ‘OK’ or ‘Not OK’ based on the gasket’s presence and alignment. This physical labelling process, carried out by industry specialists, was decisive in establishing accurate ground truth data, contributing significantly to the model’s precision and reliability in identifying correctly and incorrectly installed gaskets in subsequent inspection tasks.

Data augmentation techniques are critical for boosting the performance and generalization capabilities of models trained on image data. Given the limited number of OK and Not OK images, various transformations were applied to artificially increase the training dataset21. These transformations included:

- Rotation: Rotating images by up to 40 degrees to simulate different orientations of the gasket.

- Scaling: Adjusting the width and height by up to 20% to account for variations in the size of the gasket in different images.

- Shearing: Applying shearing transformations within a 20% range to simulate distortions.

- Zooming: Zooming in and out within a specified range to account for variations in distance from the camera.

- Flipping: Horizontally flipping the images to increase variability in the dataset.

- Noise Addition: Introducing salt and pepper noise to simulate real-world imperfections and sensor noise.

- Blurring: Applying Gaussian blur operations to mimic out-of-focus images and test the model’s robustness to image quality degradation.

The dataset became more diverse and resilient, allowing the model to handle a wider range of real-world scenarios while also improving its overall performance and generalization capabilities during training.

By incorporating these augmentation techniques, such as flipping, data rotation, and scaling, the dataset was enriched with a variety of new illustrations and differences. This improved diversity and flexibility not only enabled the model to effectively address a broader spectrum of real-world situations but also significantly boosted its adaptability and robustness. Consequently, during the training phase, the classical model was able to learn from a more complete set of data instances, thereby purifying its presentation across different situations and improving its ability to simplify to hidden cases.

The bigger dataset helped the model a lot. It was like a treasure chest full of different examples. This meant the model could understand complex patterns just like in real data. When it trained with this big dataset, the model got better at noticing small details and patterns, which helped it not get tricked by seeing the same things too many times. Plus, it got really good at spotting the connections in the data. This careful training made the model much better at guessing and sorting information correctly. By learning from lots of different data situations, the model became more flexible and tough. This way, it can make good guesses and sort things right, even when the situations are tough and unexpected.

In essence, the addition of these augmentation techniques into the dataset meaningfully donated to the improvement of the model’s capabilities. By incorporating diverse examples through these techniques, the model’s contact was not only broadened but also exactly fine-tuned, thereby optimizing its learning process. This optimization was essential as it eventually led to a substantial enhancement in the model’s overall performance metrics and effectively elevated its simplification capabilities.

Furthermore, this complete addition approach played a crucial role in stimulating the model’s capacity to skilfully navigate through complex tasks with increased competence and accuracy. The thorough incorporation of these augmentation techniques not only strengthened the model’s resilience in handling complicated challenges but also solidified its reliability and effectiveness in practical, real-world applications. This seamless fusion of augmentation methods into the dataset acted as a catalyst in improving the model’s problem-solving skills and preparing it with the necessary tools to excel across various scenarios and domains.

As a result, the model showcased extraordinary advancements in its flexibility and heftiness, showcasing an extraordinary ability to adeptly tackle a wide array of tasks with precision and consistency. The methodical integration of augmentation techniques underlined the model’s evolution towards becoming a versatile and dependable tool, capable of delivering impactful results in real-world settings and driving innovation across diverse fields.

The proposed CNN architecture is designed to extract features and classify gasket images effectively. The architecture consists of the layered view as shown in Fig. 6. Table 2 shows the overview of Layers in the Deep CNN Architecture.

Layered view of the Deep CNN architecture.

1. First Set of Layers:

Convolutional Layers:

Two convolutional layers with 32 filters each and a kernel size of 3 × 3. These layers use ReLU activation functions to introduce non-linearity and extract initial features such as edges and textures from the input images.

Batch Normalization:

Applied after each convolutional layer to normalize the output and accelerate training.

Max Pooling Layer:

A max pooling layer with a pool size of 2 × 2 to reduce the spatial dimensions and computational complexity, thereby focusing on the most prominent features.

Dropout Layer:

A dropout layer with a dropout rate of 0.25 to prevent overfitting by randomly setting a fraction of input units to zero during training.

2. Second Set of Layers:

Convolutional Layer:

One convolutional layer with 64 filters and a kernel size of 3 × 3 to extract deeper features from the feature maps produced by the previous layers.

Batch Normalization:

Applied after the convolutional layer to normalize the output.

Max Pooling Layer:

A max pooling layer with a pool size of 2 × 2.

Dropout Layer:

A dropout layer with a dropout rate of 0.25.

3. Third Set of Layers:

One convolutional layer with 128 filters and a kernel size of 3 × 3.

Applied after the convolutional layer.

A max pooling layer with a pool size of 2 × 2.

A dropout layer with a dropout rate of 0.25.

4. Flatten Layer:

Converts the 2D feature maps into a 1D vector to prepare the data for the fully connected layers.

5. Fully Connected Layers:

Dense Layer:

A dense layer with 512 units and ReLU activation to learn high-level features from the flattened input.

Dropout Layer:

A dropout layer with a dropout rate of 0.5 to prevent overfitting.

Output Layer:

A final dense layer with a single unit and a sigmoid activation function for binary classification (OK/Not OK).

The CNN model was trained using the augmented dataset. The training procedure involved:

1. Loss Function: Binary cross-entropy was used as the loss function to handle the binary classification task. This loss function measures the difference between the predicted probabilities and the actual labels, guiding the model to minimize this difference during training.

2. Optimizer: Adam optimizer was selected for its efficiency in handling large datasets and adaptive learning rate capabilities. The learning rate was initially set to 0.001 and adjusted during training using a learning rate scheduler to ensure stable convergence.

3. Batch Size and Epochs: The model was trained with a batch size of 32 and for 50 epochs. The number of epochs was determined based on the model’s performance on the validation set, ensuring that training stopped when the model started to overfit.

4. Early Stopping and Model Checkpointing: Early stopping was implemented to monitor the validation loss and halt training if there was no improvement for a specified number of epochs (e.g., 10 epochs). Model checkpointing was also used to save the best-performing model based on validation accuracy.

During training, the model’s performance was evaluated on a separate validation set to monitor overfitting and ensure generalization. The training and validation loss, as well as accuracy metrics, were recorded at each epoch to track the model’s progress.

The performance of the proposed system was evaluated on a test dataset of 10,000 images. The dataset is split into 70% for training and 30% for testing, resulting in 7,000 images for the training set and 3,000 images for the testing set. For the training set, there are 3,889 normal gasket images (70% of 5,556) and 3,111 defective gasket images (70% of 4,444). Similarly, for the testing set, there are 1,667 normal gasket images (30% of 5,556) and 1,333 defective gasket images (30% of 4,444). This distribution ensures a proportional representation of both normal and defective classes in the training and testing datasets. The system’s ability to detect misaligned or improperly fitted gaskets was measured using the following metrics. In this research work, we evaluated the model using accuracy, precision, recall, and F1 score. The proposed framework’s classification of False Negative (FN), True Positive (TP), False Positive (FP), and True Negative (TN) uses the binary-class confusion matrix. The confusion matrix provides the following outcomes:

True Positive (TP): The number of correctly identified “OK” gasket images.

False Positive (FP): The number of “Not OK” gasket images incorrectly identified as “OK”.

True Negative (TN): The number of correctly identified “Not OK” gasket images.

False Negative (FN): The number of “OK” gasket images incorrectly identified as “Not OK”.

Accuracy: This equation calculates the proportion of correctly classified gasket images (both “OK” and “Not OK”) out of the total number of images in the dataset.

Precision: This equation calculates the proportion of correctly identified “OK” gasket images out of all the images classified as “OK”.

Recall: This equation calculates the proportion of correctly identified “OK” gasket images out of all actual “OK” gasket images in the dataset.

F1 Score: This equation provides a harmonic balance between precision and recall, giving an overall measure of the model’s performance.

The results showed that the system achieved an accuracy of 97.32% and a precision of 96%, indicating its effectiveness in detecting gasket alignment issues.

In our proposed work, the performance of the system is evaluated using a dataset comprising 10,000 images of radiators and their gaskets. The dataset includes 5,556 normal images, consisting of 1,000 original images and 4,556 augmented images, and 4,444 defective images, which include 800 original images and 3,644 augmented images. This dataset ensures a balanced representation of normal and defective cases, providing a robust basis for assessing the system’s effectiveness. The dataset is split into 70% for training and 30% for testing, resulting in 7,000 images for the training set and 3,000 images for the testing set. The developed system was able to detect any misaligned or improperly fitted gaskets, and alert the operator to take corrective action. Table 3 shows the accuracy across different epochs during training.

The results in Table 3 demonstrate a consistent increase in accuracy as the number of epochs increases. Starting from 70.19% accuracy at 10 epochs, the model shows significant improvement, reaching 97.32% accuracy at 150 epochs. This indicates that the model continues to learn effectively with more training iterations, leading to higher performance.

Training and Validation Accuracy vs. Epochs.

Training and Validation loss vs. Epochs.

In Fig. 7, the plot compares the accuracy of the training and validation sets over 150 epochs. the results displayed affirm that our system excelled by achieving an impressive accuracy rate of 97.32% .These impressive numbers show how well the system can find and sort how gaskets line up inside radiators. Looking closely at these results, it’s clear the system could really change how things are made. It’s great at spotting mistakes, which means fewer errors and better-quality final products. With this system, we can check quality in a quick and reliable way, making things faster and cutting down on the need for people to check things by hand. This could save both time and money. We did tests to see how strong our system is when the light changes a lot. We gathered many pictures taken in different kinds of lighting and made the system go through tough checks. What we found out shows that our system can adjust well and keep working great, no matter what kind of light there is. This proves it’s flexible and you can count on it. Our study looked at how well a system that checks gaskets in radiators by combining image processing and machine interfacing works. This system is really useful for making things in factories. It can help make sure products are good quality and have fewer problems. We think there’s more to learn about making the system work even better and use it in new ways to improve how it operates. Figure 8 shows the Training and Validation loss vs. Epochs.

Comparison of performance metrics of proposed model with other models.

Figure 9 shows the performance of the various models on the validation set using four key metrics: accuracy, precision, recall, and F1 score. The model achieves an accuracy of approximately 98%, with precision slightly below at 97%. The recall is also close to 96%, indicating good sensitivity. The F1 score, which balances precision and recall, reflects strong overall performance at just under 97%, suggesting that the model is well-tuned for this classification task.

Table 4 summarizes the performance metrics of the proposed driver drowsiness detection model. The model achieves an accuracy of 0.9732, with a precision of 0.9695, indicating a high rate of correct predictions. The recall of 0.9600 shows good sensitivity, while the F1 score of 0.9656 highlights a strong balance between precision and recall.

Looking closely at these results, it’s clear this system could change how we make things. It’s great at finding mistakes, which means fewer errors and better quality stuff. Plus, it makes checking things faster and less costly because it cuts down on the need for people to check everything by hand.

Investigating deep into the system’s robustness, experiments were conducted aimed at assessing its resilience to variations in the lighting conditions. By amassing a varied dataset including images captured under varying lighting environments, the system was involved through rigorous testing. The results of this evaluation underscored the system’s adaptability and steadfast performance across different lighting scenarios, underscoring its versatility and reliability.

The consequence of our study on the image processing system, flawlessly integrated with machine interfacing for gasket inspection in radiators, brightens its practical implications within the manufacturing sector. Its inherent capacity to serve as a dependable quality control tool not only boosts product quality but also reduces the occurrence of defects. The road for further examination lies in enhancing the system’s efficiency and broadening its scope to unveil untapped potential and enhance its operational repertoire.

In conclusion, we have developed a deep learning model for gasket inspection in radiators. The primary objective of this work was to ensure that gaskets are properly aligned, which is crucial for the efficient and reliable operation of radiators. The system effectively detects misaligned or improperly fitted gaskets and alerts operators to take corrective action. Our evaluation, using a dataset of 10,000 images, demonstrated that the model achieved an accuracy of 97.32% and a precision of 97%, indicating its capability to accurately detect and classify gasket alignment in radiators. The results suggest that this system has significant potential to improve the manufacturing process by reducing defects and enhancing overall product quality. Further improvements and optimizations can be explored in the future to enhance the system’s performance and broaden its application. Overall, the system provides a reliable and efficient mechanism for quality control, helping to reduce defects and improve product quality in radiator manufacturing.

All data generated or analyzed during this study are included in this published article.

Mrugalska, B. & Tytyk, E. Quality Control methods for product reliability and safety. Procedia Manuf. 3, 2730–2737 (2015).

Article MATH Google Scholar

Ho, C. C. et al. Machine Vision and deep learning based Rubber Gasket defect detection. Adv. Technol. Innov. 5, 76–83 (2020).

Article MATH Google Scholar

Chen, X., Wang, D., Shao, J. & Fan, J. Plastic Gasket Defect Detection Based on Transfer Learning. Scientific Programming. 2021. 1–11. (2021).

Sharmila, V. J. & Florinabel, J. Deep Learning Algorithm for COVID-19 classification using chest X-Ray images’, Computational and Mathematical methods in Medicine. Article ID. 9269173, 10 (2021).

MATH Google Scholar

Arumai Shiney, S. S., Geetha, R., Seetharaman, R. & Shanmugam, M. Leveraging Deep Learning Models for Targeted Aboveground Biomass Estimation in Specific Regions of Interest. Sustainability 2024, 16, 4864. https://doi.org/10.3390/su16114864

Purwono, P., Karim, H. I., Frisky, Aufaclav & Haq Ma’arif, AlfianRahmaniar, WahyuImam, HarisFathurrahman, Qazi Mazhar Ul. Understanding of Convolutional Neural Network (CNN): A Review. 2. 739–748. (2023).

Yang, J. et al. Using Deep Learning to detect defects in Manufacturing: a Comprehensive Survey and Current challenges. Materials 13 (24), 5755. https://doi.org/10.3390/ma13245755 (2020).

Article ADS CAS PubMed PubMed Central MATH Google Scholar

Kudelina, K. et al. Methods of Condition Monitoring and Fault Detection for Electrical Machines. Energies 14, 7459 (2021).

Article MATH Google Scholar

Adrienn Dineva, A., Mosavi, M., Gyimesi, I. & Vajda Narjes Nabipour, and Timon Rabczuk. Fault diagnosis of rotating electrical machines using multi-label classification. Appl. Sci., 9(23), (2019).

He, J., Yang, S. & Gan, C. Unsupervised Fault diagnosis of a Gear Transmission Chain using a deep Belief Network. Sensors 17 (7), 1564. https://doi.org/10.3390/s17071564 (2017).

Article ADS PubMed PubMed Central MATH Google Scholar

Toma, R. N., Prosvirin, A. E. & Kim, J-M. Bearing Fault Diagnosis of Induction Motors using a genetic algorithm and machine learning classifiers. Sensors 20 (7), 1884. https://doi.org/10.3390/s20071884 (2020).

Article ADS PubMed PubMed Central MATH Google Scholar

Tomas Zimnickas, J. et al. Application of advanced vibration monitoring systems and long short-term memory networks for brushless Dc motor stator fault monitoring and classification. Energies, 13(4), (2020).

Gong, W. et al. A Novel Deep Learning Method for Intelligent Fault diagnosis of rotating Machinery based on improved CNN-SVM and Multichannel Data Fusion. Sensors 19 (7), 1693. https://doi.org/10.3390/s19071693 (2019).

Article ADS PubMed PubMed Central MATH Google Scholar

Xu, G., Liu, M., Jiang, Z., Söffker, D. & Shen, W. Bearing Fault diagnosis Method based on deep convolutional neural network and Random Forest Ensemble Learning. Sensors 19 (5), 1088. https://doi.org/10.3390/s19051088 (2019).

Article ADS PubMed PubMed Central MATH Google Scholar

Zawad Ali, M., Shabbir, Liang, X., Zhang, Y. & Hu, T. Md. Nasmus Sakib Khan Machine Learning-Based Fault Diagnosis for Single- and Multi-Faults in Induction Motors Using Measured Stator Currents and Vibration Signals. IEEE Transactions on Industry Applications. PP. 1–1. (2019).

Fanchiang, K-H., Huang, Y-C. & Kuo, C-C. Power Electric Transformer Fault Diagnosis Based on infrared thermal images using Wasserstein Generative Adversarial Networks and Deep Learning Classifier. Electronics 10 (10), 1161. https://doi.org/10.3390/electronics10101161 (2021).

Article Google Scholar

Park, S-H., Lee, K-H., Park, J-S. & Shin, Y-S. Deep learning-based defect detection for sustainable Smart Manufacturing. Sustainability 14 (5), 2697. https://doi.org/10.3390/su14052697 (2022).

Article MATH Google Scholar

Dhanya, V. et al. Deep learning based computer vision approaches for smart agricultural applications. Artif. Intell. Agric. 6, 211–229 (2022).

MATH Google Scholar

Yu, L., Wang, S., Lai & Kin Keung. Data Preparation in Neural Network Data Analysis. (2007).

Cheng, L., Wang, B. H. D., Zhang, S. & Chen, Z. A preliminary study on data augmentation of deep learning for image classification. pages 1–6, 10 (2019).

Shorten, C. & Khoshgoftaar, T. M. A survey on Image Data Augmentation for Deep Learning. J. Big Data. 6, 60. https://doi.org/10.1186/s40537-019-0197-0 (2019).

Article MATH Google Scholar

Download references

S. Arumai Shiney and R. Seetharaman contributed equally to this work.

Department of Computer Science and Engineering, S.A. Engineering College, Chennai, India

S. Arumai Shiney

Department of Electronics and Communication Engineering, College of Engineering Guindy Campus, Anna University, Chennai, India

R. Seetharaman

Department of Computer Science and Engineering, Loyola-ICAM College of Engineering and Technology, Chennai, India

V. J. Sharmila

Department of Electrical and Electronics Engineering, Loyola-ICAM College of Engineering and Technology, Chennai, India

S. Prathiba

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

S.A. and R.S. conceived the project. S.A., R.S., V.J.S. and S.P. carried out the data acquisition, design and analysis. R.S. wrote the manuscript and did the overall supervision. All authors reviewed the manuscript.

Correspondence to S. Arumai Shiney or R. Seetharaman.

The authors declare no competing interests.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

Shiney, S.A., Seetharaman, R., Sharmila, V.J. et al. Deep learning based gasket fault detection: a CNN approach. Sci Rep 15, 4776 (2025). https://doi.org/10.1038/s41598-025-85223-8

Download citation

Received: 30 September 2024

Accepted: 01 January 2025

Published: 08 February 2025

DOI: https://doi.org/10.1038/s41598-025-85223-8

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative