The first steps for AI implementation on the plant floor - an example of Computer Vision in Quality Control

Every manufacturer aims to find fresh ways to save money and make money by reducing risks and improving overall production efficiency.

It is crucial for the survival and to ensure a thriving, sustainable future. The key lies within 4IR (the fourth industrial revolution) technologies, most recently AI-based and ML-powered innovations.

AI tools can process and interpret vast volumes of data from the production floor to spot patterns, analyze, predict, detect anomalies in production processes in real-time, and more.

These tools help manufacturers gain end-to-end visibility of all facility manufacturing operations across all geographies. AI-powered systems can learn, adapt, and improve continuously thanks to machine learning algorithms.

Such capabilities are crucial for manufacturers to thrive after the pandemic-induced rapid digitization.

According to McKinsey, companies using AI have witnessed cost savings and revenue growth. 16% of those surveyed noticed a 10-19% decrease in costs, whereas 18% saw a 6-10% increase in overall revenue.

Over 50% noticed AI adoption being measurably positive.

It tells us that many companies make mistakes during their AI adoption, be it not having a solid foundation, lacking data, or having limited talent pool access. On the other hand, the companies that manage to overcome initial difficulties and approach the problem wisely see real value that can be translated to cash flow.

Quality Control - the most promising starting point

The highest AI usage is seen in quality-related actions. Why?

First, digitized measurement systems and computer vision are very mature technologies.

They were battle-tested and proven to work even in harsh conditions, be it mining or raw material processing.

Secondly, it's a large chunk of a pie.

The American Society for Quality estimates that for many organizations, this cost of quality is as high as 15-20% of annual sales revenue.

If you could:

- lower the number of defective products significantly,

- avoid the hidden cost of inaction, as many companies are already starting to operationalize AI,

- detect faulty products as soon as the defect happens (lowering scrap and, in some instances, preventing snowballing costs, e.g., when the defect happened many stages before it was caught and the whole batch must be scrapped).

Why wouldn't you?

The path is relatively simple, although it requires commitment, critical thinking, and working with a technical prowess partner (inside or outside the organization).

Step-by-step guide:

If you haven't implemented AI solutions on your shop floor, starting with quality inspection might be the best choice.

You don't need a massive dataset and costly infrastructure for many use cases.

Especially if you bet on a safer approach -> starting small with a series of PoCs in identified zones and production stages and gradually increasing the AI budget with more promising results.

Ideally, you should turn each successful PoC into an internal MVP-oriented project and then a full-scale development.

As individual as the latter stages are, PoC will usually look the same.

Let's break down the process of building such PoC for a Computer Vision application to support Quality Control processes.

1. Choose one use case

Analyze your production process, looking for bottlenecks and stages where manual inspection is the least feasible or where defects have the highest chance of snowballing.

Think of how to automate that part, allowing for more accurate, timely, and consistent inspections.

2. Define KPIs

To assess the PoC phase, you must base on data. Define KPIs to indicate the success (or potential) of an AI application.

For Quality Control, these might be:

- Defect Detection Rate - the ability of an automated system to identify faulty parts. The higher, the better.

- False Positive Rate - the percentage of parts incorrectly identified as faulty. A low FPR indicates the system's high precision.

- Inspection Time - Tracked time needed to inspect each part. A decrease in time implies higher system efficiency.

3. Gather data

Having a high-quality, well-labeled dataset will allow your model to learn effectively, improving the accuracy and precision of its predictions.

In defect detection, you'll have to gather diverse data on defect-free and faulty components, possibly with some information about the production environment.

4. Train a simple model

Choose an off-the-shelf model or a relatively simple algorithm to learn on the provided dataset. You need just enough data to verify the initial hypotheses on AI's usability.

Of course, training is not the only step here.

The model must be then tested using a separate dataset to evaluate it's performance. This stage provides valuable information on fine-tuning the model and adjusting parameters to improve performance.

5. Evaluate results

After the first couple of iterations, you should be ready to evaluate the model performance based on the KPIs.

If the results are promising, develop a roadmap for incrementally building the AI system.

Remember that every deployment must bring value, be it process optimization, higher system flexibility, gains in efficiency, or cost savings.

Case Study - BMW

BMW is an excellent case of following such a path, starting small with relatively simple but well-thought-out internal projects and gradually increasing investment in AI to expand technology usage.

Let's look at some of the implementations that BMW's R&D team developed in cooperation with multiple technology experts.

Assembly line production monitoring

An AI application to evaluate component images in ongoing production based on hundreds of other images of the same sequence.

Looking for deviations from the standard in real-time, it checks whether all the required parts have been mounted and if they were mounted at the right place.

Final inspection automation

At one of BMW's plants that releases up to 1600 cars daily, the final inspection area was augmented with an AI application. The system compares the vehicle order data with a live image of the model. A final inspection team receives a notification if the data don't correspond.

A very clever and relatively simple build that might significantly increase the inspection speed while preventing costly rollbacks and brand damage if a client gets a car that doesn't comply with the order.

Pseudo-defect elimination

Built with heavy support from BMW's Data Analytics Team, this solution is an AI-based analysis software used in torque measurements in the engine cold tests.

Previously, irregularities in test output led to costly and complex manual inspections, although in most cases, they turned out to be insignificant.

The analysis software was trained on multiple recorded test runs to learn to distinguish actual and presumed errors.

Smart control application

Implemented at BMW's Steyr plant, this application speeds up logistic processes by preventing unnecessary transports of empties on conveyor belts.

Various containers pass through an AI-powered camera station that uses stored image data annotated by employees to recognize whether the container must be lashed onto a pallet to be assembled with other smaller boxes or is large enough not to need additional securing.

For the latter example, an AI application directs a container to the nearest removal station.

These containers that need to be additionally secured are guided directly to the conveyor section with the lashing system and only then to the removal station located behind.

Before implementing the intelligent camera system, all the containers went straight to the removal station, where they were manually assessed and conveyed once again, first to the lashing station and then to the correct removal station.

BMW now

A long-term AI implementation project eventually led to the building of the first 3D print-based production line with all the automated processes and on-site operations requiring minimal human supervision.

Around 50,000 components per year can be manufactured cost-effectively in common part production, and more than 10,000 individual and new parts, all through 3D printing.

Quality assurance of the finished parts takes place in-line with AI algorithms that correlate the data collected with actual component quality.

This means process deviations can already be identified during production and component quality evaluated.

Summary

Implementation on a plant floor is much different than in a startup environment.

The goal is to introduce minor process improvements (via a 1% approach).

It’s critical to identify places where ML is the most feasible by the following criteria:

• There’s a bottleneck in the process.
• It’s a mundane job with over-the-top man-hour investment.
• There’s easy access to data.
• It’s proven to work in other companies (e.g., via case studies).

The business is the primary driver of ML-based change. It needs to craft the right approach by:

• Identifying ML opportunities.
• Choosing multiple single, simple use cases.
• Validating each use case with a PoC/MVP prototype.
• Pursuing the most promising cases with more significant in-house projects with SMEs and business engagement.

--------------------

For more ML and AI insights, follow?Sparkbit?on LinkedIn.

If you're looking for a team to run your AI project end-to-end, DM us or write to [email protected]