Manufacturing Quality Control Using Deep Learning & Computer Vision

1 0
  • 0 Collaborators

The project is a cost-effective computer vision system to monitor production and detect anomolies using convolutional neural networks & DL. ...learn more

Internet of Things, Artificial Intelligence

Groups
Movidius™ Neural Compute Group

Overview / Usage

Background:
Manufacturing processes operate continuously to achieve high production rates. This continuous operation increases potential risks for faults in the final product. Continuous quality control of the production is necessary to ensure high quality of the output. Generally, this quality control is performed either with human labor or with intelligent systems most notably computer vision systems using video feed from cameras installed on the production line.

For this project, I will consider a plastic bag production line to which I have easy access in order to implement this project.

Plastic bag production is a 2 step process:
1) Blowing film, surface treatment, and printing
2) Sealing, cutting, and punching
This factory uses flexography printing to print the desired design on the plastic bag. This involves aligning very precisely multiple colors. More on flexography below (at the end).
http://www.sy-pm.com/english/images/news/2011731172452414.jpg

Problem statement:

In old flexo printing systems (which are most common in this industry), quality control systems are very basic relying on an operator to visually inspect the printing quality.
However, during the printing phase of the plastic bags manufacturing process, several problems may arise:
• Colors misalignment
• Printed colors not respecting Pantone specifications (due to solvent use)
• General misalignment of the print,
• Roll dimensions changes due to blowing process, etc…

These issues mentioned above can cause the print on the plastic roll to be out-of-spec and result in faulty output that goes to waste, thus impacting the overall efficiency of the factory.
The faster the operator corrects the before mentioned problems, the less the number of bags goes to waste. Consequently, the operator must constantly and repetitively monitor the quality of the printing process. In small factories with limited number of operators, this limits the operators productivity by preventing him to focus on other tasks.

Solution:

The general idea of this project is to provide factories with low-cost, easy to install machine vision quality control systems that can be retrofitted on old systems to enable high-speed monitoring on production and assembly lines. The objective is use deep convolutional learning and computer vision for anomaly detection in the manufacturing process.

In this specific application of the plastic bag manufacturing line, the project aims to detect printing quality anomalies using deep convolutional neural networks and alert the operator through an alarm to fix the problem.

Another objective is to design the device to be cheaper than sophisticated machine vision systems by using the Movidius Neural Compute Stick installed on a Raspberry Pi and a standard high-quality USB camera, to enable edge inference and anomaly detection.

Technical details:

Training
• The training of the convolutional neural network will take place on the Intel Nervana DevCloud.
• The neural network architecture will probably be a customized GoogleLeNet v1 topology with a reduced fully-connected layer class from 1,000 to 2 binary classification as output.
• "This topology balances the training/inference time and testing accuracy, making it well suited for use as image classification." This was inspired by this Intel project published in August 2017: https://software.intel.com/en-us/articles/manufacturing-package-fault-detection-using-deep-learning
• This will be implemented using Caffe optimized for Intel architecture with the Intel MKL
• The training will use supervised learning using images from the real-time video feed labeled as normal, and several faulty prints labeled as abnormal.
• Another method could be by using a CONV-LSTM architecture for the anomaly detection without having to use binary classification.

Testing
The inference will take place on the Movidius NCS by processing the frames of the real-time video feed from the standard USB camera. The stick will be placed on a Raspberry Pi to send a warning signal to alert the operator incase of an anomaly detection.

N.B. The use of the Raspberry Pi is to be able to do edge feed-back control in future use-cases.
This would allow to completely rely on this device to ensure the quality of printing on old flexo printers with no modern feed-back control system.

More about Flexography:
Flexography is a form of relief printing where ink is applied to a rubber or polymer plate on which the printing image is raised above the rest of the surface as a 3d positive mirrored relief
Flexography prints from a flexible printing plate that is wrapped around a rotating cylinder. The plate is usually made of natural or synthetic rubber or a photosensitive plastic material called photopolymer. It is usually attached to the plate cylinder with double-sided sticky tape.
More: http://www.tappi.org/content/events/09PLACESY/Course_Papers/durling.pdf

Comments (0)