Quantifying, understanding and verifying product quality is a critical step in any manufacturing process. Moving from a manual inspection to an automated inspection provides greater insight into the quality of the product and process by creating measurable baselines and enabling statistical process controls. This case study provides an overview and takeaways from an ATS project that involved designed flexibility, close collaboration with the customer on their evolving needs, and advanced technologies including Artificial Intelligence-based (AI) inspection.

THE PROBLEM

The customer came to ATS with the goal of developing an inspection system to measure the volume dispensed within their product. This inspection was previously performed with a manual go/no-go gauge, and the desire was to move to automated measurement to increase reliability and accuracy.

THE SOLUTION

The customer product contains multiple wells which are filled simultaneously with multi-up dispense tooling, which removed a weight check as a potential quality measurement solution. A machine vision-based inspection was agreed upon to automate the manual quality check. The machine vision system designed by ATS can be broken into the hardware for image acquisition, and the software for image processing.

Image Processing

ATS’ long history of designing and implementing automated vision systems led to the development of machine vision products to enable engineers to design and integrate the best solution to each unique problem. ATS SmartVision^TM is a rapid-development machine vision programming environment that has been developed for over 20 years with thousands of license deployments on hundreds of different machines.

Open architectures to leverage state-of-the-art technology:

Evolving Problem Description

After the initial order for the equipment, the customer’s internal use-case for the system pivoted, meaning they now required validation of the system. ATS engaged their Validation Engineering and Systems Engineering teams to plan and execute a capability analysis and system validation at FAT (Factory Acceptance Testing). Part of the validation process included implementing recipe control to allow a user to select from a number of different part configurations.

The system was then installed at site, and product straight off the manufacturing line was run through to collect initial data. At this point, it was observed there were air bubbles in the media when freshly filled which caused the meniscus to be artificially high. As the customer ran the inspection system with these new challenges, every image was saved to the vision PC, allowing for review of the data and for further inspection development.

Figure 2: Illustration to demonstrate user interface to the vision inspection, showing the initial implementation results where bubbles were not detected or considered in the volume measurement.

These images with real-world data were sent to ATS to evaluate the new challenge and potential solutions. Reviewing the hundreds to thousands of images collected there were two observations which guided the direction for the solution:

Those two observations, paired with an abundance of data, indicated a high potential for solving this vision challenge using an AI approach.

Extendable Technology to Address New Problems

The Deep Neural Network (DNN) tool within ATS SmartVision has been developed to be able to run deep learning models within any machine vision inspection. The DNN tool can run models trained with the majority of popular deep learning frameworks, including Tensorflow, Pytorch and others. This DNN tool acts like all other graphical tools within an ATS SmartVision application;  allowing for AI models to be easily added to an existing application to enhance the functionality as inspection requirements or product appearances evolve over time.  

ATS trained a deep learning model to detect the presence and location of air bubbles – a problem previously not possible to solve with traditional machine vision tools. One hundred images with different liquids and fill levels, from different production batches, were used to teach an object detection model to locate bubbles within the liquid.

Seventy percent of images were allocated to a training set, 15% to a validation set, and 15% to a test set. In the training set, there were 89 bubbles labeled of varying sizes and locations. Once all the images were collected, the labeling and training process took less than 8 hours. The model trained had a mean average precision of 0.966, indicating a very good overlap of the labeled bounding boxes versus the predicted bounding boxes.

The deep neural network trained to detect bubbles was then integrated into the existing vision inspection, consisting of traditional tools to take accurate measurements of the fill level. The new inspection calculated the volume in each tube, as it previously did but, if an air bubble was present, its volume would be subtracted from the tube measurement.

This code update was completed by ATS remotely and sent to the customer for deployment.

Figure 3: Illustration of ATS Deep Neural Network (DNN) tool capability, running in inference mode to output the location of detected bubbles.

Figure 4: Illustration of user interface to the vision inspection, showing the updated implementation results where bubbles are detected, and the calculated volume is subtracted from the part volume.

THE RESULTS

From the beginning of the project, ATS designed a benchtop system with the intent to be flexible in customer support and software. This was leveraged with the addition of validation prior to system delivery and the addition of AI vision code updates to address new challenges with product coming straight from the manufacturing line. Following the code updates to address the product seen onsite, the customer ran parts in parallel through the manual inspection process and the ATS machine vision inspection system for a period. This was done to prove the performance and reliability of the vision system as a quality system before making the final change over from the manual inspection to the validated machine vision system.

A further benefit to the customer was the ability to procure additional vision systems for other sites across their company.; this helps drive consistent and standardized quality monitoring regardless of the manufacturing site. Further, in a next generation manufacturing line this vision system can be integrated directly in-line with the automated manufacturing process.