A factory approached us with the request to develop an intelligent system capable of detecting and separating defective PET Parisons before the inflation process. The solution needed to be highly efficient and compact, designed to operate on a small device directly integrated into the production line.
Our Approach:
- Data Collection and Augmentation: We started by collecting defective and non-defective PET Parison samples from the factory. To enhance the dataset’s robustness and variability, we applied data augmentation techniques, which allowed us to simulate a wider range of possible defects and ensure that the model would generalize well in production conditions.
- Deep Neural Network Training: Using a Convolutional Neural Network (CNN) architecture, we trained a deep learning model to classify and detect faulty PET Parisons with over 80% accuracy on the initial, limited dataset. As we gathered more samples from the production line and retrained the model, the accuracy continued to improve.
- Efficient Deployment on Small Devices: The model’s lightweight architecture allowed us to deploy it on a Raspberry Pi, enabling real-time defect detection directly on the production line. To further reduce the size of the device, we employed quantization techniques such as Quantization-Aware Training (QAT), allowing the model to run efficiently on a microcontroller using TinyML technology.
Impact and Results:
The deployment of this AI-driven quality control system resulted in significant improvements in the detection and separation of defective PET Parisons before the inflation process. The lightweight model, capable of running on small devices, provided real-time defect detection without compromising on accuracy or speed. As more data was collected and the model was retrained, the system’s performance continued to improve, offering the factory a reliable and scalable solution for quality control.