Contents

AI for manufacturing

Rafał Pytel

29 Oct 2024.6 minutes read

AI for manufacturing webp image

In our series, we previously covered agriculture but also mentioned possible applications of chatbots in LegalTech, and MedTech. We now move to another sector that can benefit greatly from AI: the manufacturing industry.

On the other hand, manufacturing companies already profit from automation, robotics, and a data-driven approach. Industry 4.0 is a specialised term for applying technological advancements to manufacturing jobs. It will be covered briefly soon, followed by a variety of examples of applications using AI to improve this sector.

What is Industry 4.0?

industry 4.0

Source: calsoft.com

Industry 4.0, otherwise called the 4th industrial revolution, began as a term in the 2010s and is generally defined by four types of technologies:

  • Increased connectivity: via IoT and cloud technology, sometimes also employing blockchain.
  • Advanced analytics and data-driven decisions via advanced data analysis and machine learning
  • Human-machine interaction: VR, AR, robotics, and automation.
  • Advanced engineering: 3D printing, renewable energy, or nanodroids.

It might sound extremely futuristic, but many of these things are already widely popularized (e.g.,the robotization of warehouses or cashier-less stores like Zabka Nano or Amazon GO). In the next section, I will review a few examples, mainly focusing on their AI aspect, as this is most relevant for this post.

AI for the manufacturing industry

In the following sections, we will go over a few main groups of applications where AI is improving work in the factories and warehouses. We will start with automatic quality inspection, document understanding and automatic counting, warehouse safety, and predictive maintenance.

Automatic Quality control in components and finished goods

This group of examples might not be so prominent, but in the large scale production when working in high-performance environments, faulty components need to be detected and either fixed or discarded. This is important as such tasks were performed by humans before and required much longer, while now with AI their manufacturing output can be greatly improved, while reducing the production and operation cost.

Challenges of the field

The main challenges of that field are:

  • Specialised datasets are not always easy to acquire.
  • Limited transfer learning capabilities as the problems are unique.
  • Rareness and variety of defects might require Generative solutions (i.e., GANs or Stable Diffusion)

Defects in steel sheets

Defects in steel sheets

Steel defects examples from the Severstal dataset.

So now we start with detecting defects in steel sheets (but the problem is similar in other raw materials). This is a critical problem considering that faulty steel sheets might be later used for some other product and impact its durability or physical characteristics. Defects can range from various porosities to oxidations and heat-related hotspots.

The problem became more popular in Deep Learning after the publication of the challenge on Kaggle, which later span related papers in the field (i.e., a paper from IEEE). The main architectures that were successful for that problem were ensembles of UNets and FPN networks trained explicitly on the steel defects datasets.

Defects in complex structures (including 3D printing)

Observed hotspots in complex steel structures

Observed hotspots in complex steel structures, source: Towards real-time in-situ monitoring of hot-spot defects in L-PBF

Another exciting area where defect detection can be used is complex structures, often produced using 3D printers. There, the defects might not be facing directly or even the surface but on welds or corners, making it difficult to detect via simple RGB cameras. Often, to mitigate that, IR/hyperspectral cameras are used, as they might be more resistant to environmental distractions, i.e., light reflections.

Predictive maintenance

Corrosion detection

Corrosion detection, Source: v7labs.com

With IoT infrastructure already surrounding us, we might use some of these sensors to predict when certain parts might need repair (to minimise losses or breakdowns) or simply monitor their health. To go more in the details:

  1. Vibration monitoring - detecting anomalous vibrations, which might indicate a need for maintenance before it is too late.
  2. Temperature tracking - By using thermometers or infrared cameras, it is possible to detect overheating.
  3. Oil quality inspection - via air pressure monitoring in pneumatic systems.
  4. Wear detection - same as in the example above, using machine vision to detect defects.

Background manufacturing process: Documents

Document AI from Google

Document AI from Google, source: Google Cloud

With large factories comes a problem of increased bureaucracy and, eventually, a larger amount of files being created. It might be impossible for a single person to monitor everything at a certain point, so here comes document understanding and OCR. I don't mean simple OCR, like reading car plates, but rather Relation Extraction (i.e., elastically finding who sold what to whom in what amount) i.e., offered by Google, Microsoft or even Snowflake. As the topic is quite complicated and has many domain-specific challenges, firms rarely decide on in-house specialised tools and accept “quite good” general tools from the aforementioned cloud providers. If you are interested in the further topic, we have covered it more deeply here.

Challenges of the field

The main challenges of that field are:

  • Layout-related problems include parsing tables and page breaks.
  • Limited availability of properly labeled data.
  • Limited support for multi-language documents.

Automatic counting in mass production

Automatic counting in mass production

Source:viso.ai

In manufacturing, you often create a lot of goods, and you might need a reliable and quick way to count them. This might be done via mechanical features or solutions that simply use cameras to calculate precise amounts of goods. The Computer Vision app can do this, and it can be later deployed in various setups (including mobile apps). These solutions mostly use correctly tuned object detection and are pretty simple to implement and deploy.

Challenges of the field

However, the main challenge of that field is a variety of environments in background, lighting, and possible occlusions.

Binging safety to manufacturing processes

Incidents prediction from Scanflow

Incidents prediction from Scanflow. Source: scanflow.ai

Warehouses and factories are not safe spaces, with the possibility of different injuries or other damages literally on every corner. Any incidents might impact the health of both humans and equipment, effectively impacting costs and disrupting the supply chain.

Additionally, warehouses are becoming more and more populated by automatic machines, so advanced solutions that predict and avoid such incidents are of great value (i.e., around production line or when moving pallet racks).

Good examples of such solutions can be Scanflow or Surveily. This can not only reduce incident problems, but also enforce certain safety policies in industrial environments.

Challenges of the field

Video analysis from Surveily

Video analysis from Surveily, source: suveily.com

The main challenges of that field are:

  • The temporal aspect of videos (i.e., the possibility of collision)
  • Limited data availability.
  • Unlimited list of possible incidents.

New product development

Nvidia H100 structure

Look into the Nvidia H100 structure, which was partly designed by AI, Source: Nvidia

Having vast data about usage, issues, demand, and market understanding makes it possible to design new and improved products.

An excellent example is Nvidia, which has experimented with using Generative AI and Reinforcement Learning for chip design.

Challenges of the field

The main challenges of the field are actually:

  • Unclear evaluation objectives.
  • The proposed solution must obey physics laws (which are often not straightforward with Generative AI).
  • Need for a complicated and detailed simulation environment.

Conclusions

In this blog post, we have covered Industry 4.0, its possible directions, and how AI enables such a development. The broader adaptation of AI is accelerating, not only in the field of chatbots but also thanks to the advances in applied Computer Vision (as we saw in multiple examples here).

If you are interested in exploring further opportunities in your industry, do not hesitate to contact us.

reviewed by: Kamil Rzechowski

Blog Comments powered by Disqus.