AI Reshaping the New Era of Inspection Machines

The fully automatic intelligent inspection machine utilizes advanced machine vision technology, optical imaging technology, and computer image processing algorithms to achieve fully automated inspection of visible foreign particles (such as glass shards, metal debris, hair, fibers, etc.) inside pharmaceutical vials, as well as the fill volume and the external appearance of the bottles.

First, the product to be inspected is transported to the inspection area via a conveyor system. In the inspection area, specialized LED lighting illuminates the interior of the bottle. Next, a high-resolution industrial camera continuously captures images of the bottle's interior. These images are transmitted to a computer system, where image processing software compares and analyzes them to determine whether any foreign particles or other defects are present inside the bottle. Finally, based on the inspection results, the equipment automatically separates qualified products from defective ones, rejecting the defective items using a sorting device.

Currently, the mainstream method for detecting foreign particles in pharmaceutical liquids relies on manually identifying foreign particle characteristics, such as setting specific grayscale and size thresholds. The traditional process includes region selection, feature extraction, and classification using a classifier. For different pharmaceutical products, specialized personnel need to set different parameters based on varying characteristics. If the parameters are set too strictly, bubbles may interfere and cause false positives, reducing production efficiency. If the parameters are too lenient, defective products may go undetected. The parameter-setting process is time-consuming and labor-intensive, requiring continuous testing and adjustments by experts.

To address the false detection and missed detection issues that traditional inspection algorithms cannot resolve, Truking Technology?has introduced a new AI-based foreign particle detection method. This approach effectively mitigates missed detections and false positives in complex background conditions. The AI inspection method enables precise tracking of foreign particle trajectories through extensive sample training. This technique is entirely based on big data AI analysis and processing, requiring no predefined patterns or frameworks—only a large volume of sample data and appropriate annotations to enable self-learning and inference model generation.

?

The AI-based foreign particle detection system involves two key processes: model training?and model deployment. During the training phase, a large dataset of images containing foreign particle characteristics is required as input. The more comprehensive and accurate the input data, the better the performance of the resulting detection model. High-quality datasets significantly improve model training effectiveness and prediction accuracy. The vast and precise image database accumulated by Truking?serves as a crucial foundation for the success of its AI-driven foreign particle detection system.

With years of experience in the pharmaceutical equipment industry, Truking?has amassed an extensive collection of product images and defect samples, providing valuable support for applying deep learning technology in pharmaceutical inspection. The Truking inspection image database?contains images from over 100 pharmaceutical factories covering ampoules, vials, and oral liquid production lines—totaling over 100,000 images, including foreign particle defects, external defects, and images of good products. The foreign particle image database?includes 80,000 images?of various contaminants, such as hair, fibers, glass, metal, black spots, and white dots. Additionally, it features:

· 2,000 images of floating foreign particles

· 5,000 images of foreign particles adhering to the bottle bottom and body

· 8,000 images of bubbles not fully eliminated This comprehensive dataset meets the specialized inspection needs of water-injection pharmaceutical factories.

Compared to traditional algorithms, the AI detection model has improved the detection rate of foreign particles in complex background images by 80%, while reducing the false positive rate by 70%. The quantized model significantly shortens prediction time, greatly enhancing the inspection speed at each workstation. Additionally, the modular, plug-in algorithm framework?reduces coupling between detection modules, facilitating the coexistence and complementarity of deep learning and traditional detection algorithms. The detection model can effectively identify foreign particles with varying geometric features, accurately distinguish floating contaminants from those adhering to the bottle bottom and body, and differentiate between bubbles and dirt—significantly improving accuracy while reducing missed detections.

Advantages of the Deep Learning-Based Foreign Particle Detection System:

1. Superior recognition of irregularly shaped foreign particles?compared to traditional machine vision, with higher accuracy and reduced susceptibility to environmental factors.

2. Continuous iterative training?based on user-provided data, gradually improving model accuracy.

3. Intelligent detection of new defects, enabling defect control before products are shipped.

4. Synthetic defect generation, effectively addressing the challenge of insufficient image data for some defects.

5. Empowering end users, allowing them to complete the model training and prediction process independently, achieving a self-optimizing and upgrading model.