Automated Glue Inspection System (Engine Parts)

PyTorch-based real-time image classification for quality control
Code not public due to intellectual property restrictions.

---

Background

In the production process, adhesive is applied to a certain component of an engine. However, in some cases, the adhesive may be incompletely applied. To ensure quality control, each part needs to be visually inspected and classified as either:
OK: Adhesive has been properly applied.
NG: Adhesive application is incomplete or defective.

Workflow Overview

Here’s how the image classification and processing pipeline works:

• Scanning Program: The scanning program captures an image of the part and saves it as a PNG file in the designated scanning directory.
• Monitoring Program:
  The monitoring program continuously watches the scanning directory for new PNG files.
  Once a new PNG is detected, it retrieves the image name and waits for one second(configurable) to avoid read-write conflicts — this is safe given that the station cycle time is 150 seconds, leaving ample processing time.
  After the delay, the monitoring program starts the image classification process (e.g., using a trained model to determine OK/NG).
• Classification & File Management:
  After classification, the monitoring program moves the image from the scanning directory to a target (output) directory.
  It then renames the image by appending either _OK or _NG to the filename, indicating the classification result.
• Downstream Processing: Other programs or systems further process the images located in the target directory, based on their OK/NG labels.

---

Some instructions:

1.The server is a Windows Server.
2.The scanned images are associated with the engine number (uniqueness), for example, 41235672_TJ.png.
3.The number of images per year is 200,000. To prevent too many files in the target directory, subdirectories will be created, with 1,000 files in each subdirectory. The subdirectories are created based on the image names, for example, Output\41\41235\41235672_TJ.png.
4.Classification failures or invalid images will be marked as NG. After image recognition, they will be renamed to Output\41\41235\41235672_TJ_OK.png or Output\41\41235\41235672_TJ_NG.png.

Workflow Design

graph TD
    A[Scanner program saves<br>image as *.png<br>in watched directory] 
    B[Monitor watches directory]
    B --> C{New PNG<br>detected?}
    C -->|Yes| D[Wait 1s :configurable:<br>Classification Model PyTorch]
    C -->|No| F
    D --> E[Move & rename:<br>672_TJ_OK.png or<br>672_TJ_NG.png]
    E --> F[Result logged]
    F --> G[Wait 5s :configurable]
    G --> B

Loading