Authors

• B. Janakiramaiah, Prasad V. Potluri Siddhartha Institute of Technology, Vijayawada, Andhra Pradesh, India.
• K. Sai Sudheer, Tejas Networks, Bengaluru, Karnataka, India.
• L. V. Narasimha Prasad, Institute of Aeronautical Engineering, Hyderabad, Telangana, India.
• M. Krishna, Sir C. R. Reddy College of Engineering, Eluru, Andhra Pradesh, India.

Background

Product defect detection is a fundamental aspect of quality control in the manufacturing industry, traditionally conducted through human visual inspection. However, this approach is fraught with challenges, including being labor-intensive, prone to errors, and inconsistent. The advent of deep learning, particularly Convolutional Neural Networks (CNNs), has opened new avenues for automating defect detection processes, significantly improving accuracy and consistency. Despite these advancements, deploying complex CNN models in real-time manufacturing environments remains challenging due to the limited computing capabilities of edge devices typically used in these settings.

Challenges

The primary challenge addressed in this study is the deployment of CNN-based defect detection systems on edge devices with constrained resources. While deep learning models excel in accuracy, their computational demands often exceed the capabilities of edge devices, resulting in slow processing times, increased memory usage, and potential delays in real-time decision-making processes. This necessitates the development of lightweight CNN models that maintain high detection accuracy while optimizing for speed, memory efficiency, and overall computing performance.

Objective

The objective of this study is to develop and deploy lightweight CNN models for defect detection that are optimized for edge computing in manufacturing environments. The goal is to achieve a balance between detection accuracy, model training time, memory consumption, and computing efficiency to enable real-time, on-device defect detection.

Methodology

The research introduces a novel approach to defect detection by leveraging lightweight CNN models, which are specially designed to operate efficiently on edge devices. The methodology involved the following key steps:

Model Development: Lightweight CNN models were developed using state-of-the-art techniques and optimized for edge device deployment. Transfer learning was applied to pre-trained models, adapting them to specific manufacturing defect detection tasks across different datasets, including fabric, surface, and casting datasets.

Edge Device Deployment: The optimized CNN models were deployed on the NVIDIA Jetson Nano kit, a popular edge device in industrial applications known for its balance between computing power and energy efficiency.

Performance Evaluation: The performance of the deployed models was rigorously evaluated based on several metrics, including detection accuracy, sensitivity, specificity, F1 score, processing speed, and resource utilization. These metrics were chosen to reflect the real-world demands of manufacturing environments, where both speed and accuracy are crucial.

Results

The study yielded several important findings regarding the effectiveness of lightweight CNN models for product defect detection in manufacturing:

Improved Detection Accuracy: The lightweight CNN models achieved high accuracy rates across various defect detection tasks, demonstrating that edge device constraints did not significantly compromise detection performance.

Efficient Real-Time Processing: The deployment on the NVIDIA Jetson Nano kit showed that the models could process defect detection in real-time, with reduced latency and improved speed, making them suitable for integration into live manufacturing workflows.

Robust Performance Metrics: The models maintained strong performance across key metrics, including high sensitivity and specificity, which are critical for minimizing false positives and false negatives in defect detection.

Resource Optimization: The lightweight CNN models were optimized to use less memory and computing power, ensuring that they could operate efficiently on edge devices without overwhelming their limited resources.

Conclusion

This study demonstrates the potential of lightweight CNN models to revolutionize product defect detection in manufacturing industries by enabling real-time, on-device detection using edge computing. The successful deployment of these models on the NVIDIA Jetson Nano kit highlights their applicability in resource-constrained environments, where traditional, more complex CNN models would be impractical.

The research underscores the importance of optimizing deep learning models for edge computing to meet the specific needs of industrial applications. The lightweight CNN models developed in this study offer a viable solution for manufacturers seeking to enhance quality control processes through automation while maintaining operational efficiency.

Implications

The findings of this research have significant implications for the manufacturing industry, particularly in the context of Industry 4.0, where real-time data processing and decision-making are critical. The ability to deploy effective defect detection systems directly on the production line using edge devices can lead to improved product quality, reduced waste, and increased operational efficiency. Furthermore, the approach outlined in this study can be extended to other applications within the industry, paving the way for more widespread adoption of AI-driven solutions in manufacturing.

Future Work

Future research could explore the application of lightweight CNN models in other areas of manufacturing beyond defect detection, such as predictive maintenance and process optimization. Additionally, further optimization of these models for even more constrained devices could expand their applicability in various industrial contexts, making advanced AI tools accessible to a broader range of manufacturers.

Source: Article