This article is part of our series of content about AI within Manufacturing. We invite you to visit our DarwinAI & Manufacturing landing page, where you can access our newest insights:
- Our white paper — Leading Artificial Intelligence Use Cases In Manufacturing
- Our case study — A Global Aerospace & Defense Manufacturer uses Artificial Intelligence
- The Aspire Food Group story, about optimized cricket harvesting using DarwinAI’s deep learning.
Industry 4.0, or the “Fourth Industrial Revolution,’’ was first introduced in 2011 and refers to the ongoing automation of traditional manufacturing practices. In the last ten years many VC firms have been attracted to the problem space, alongside the increasing use of industrial-IOT and resultant big data, an increase in computer power and the need and ability to automate as much as possible. Surely this should mean this problem is being addressed adequately… But in fact, our work with enterprises and our conversations with many Fortune 500 manufacturing companies keen to adopt parts inspection in manufacturing reveal a different picture.
Although the business value is massive to the sector, adoption lags
According to a report cited in Markets and Markets, “The AI in Manufacturing market is expected to be valued at USD 1.1 billion in 2020 and is likely to reach USD 16.7 billion by 2026; it is expected to grow at a CAGR of 57.2% during the forecast period.”
So, why is adoption slow when the business value and impact is clearly substantial?
Before we explore why adoption is slow and discuss how to improve your organization’s chances for success, let’s visit some specific examples to illustrate why leveraging AI for parts inspection can provide a huge benefit.
AI helps reduce scrap, rework, risk of litigation and additional costs
In manufacturing, the supply chain can be complex, making final inspection of a finished good challenging. Suppliers are contractually bound to deliver parts as per a specification, but defects often go undetected. The impacts of defective parts entering a supply chain can result in everything from costly recalls or missing project deadlines, to catastrophic failures that can result in lawsuits or loss of life. Clearly there is much to gain when leveraging AI for smoother more efficient parts inspection applications.
AI improves efficiencies
There is a large impact if a defective part is missed, and so often for high volume parts we find shop floors with round-the-clock shifts full of workers manually inspecting parts. This does not scale well with the business and, as much as we like to think we are, humans are not perfect: defects can go unchecked with large downstream ramifications. AI reduces errors and lets humans thrive more fully and effectively in the less repetitive parts of their jobs.
AI helps to handle the multi-dimensional complexity of manufacturing
Whether it’s multi-layer PCBs or organic shapes generated via additive manufacturing processes, production processes are complicated and minor variations can result in defects. Complicated production processes also require complicated inspection processes with many options to find defects, whether they’re using X-ray, penetrative dye inspection, or magnetic eddy current detection. However, these processes still rely on human interpretation, which also requires extensive training and error handling. AI helps humans do their most complex jobs better.
There are other reasons why parts inspection is critical to the modern manufacturing enterprise, but these are the primary ones when it comes to where AI solutions can provide huge gains.
While AI solutions are possible, adoption lags
A lack of talent and training
Most software developers lack the specific skills and knowledge required to implement Machine Learning, let alone Deep Learning, at the sophistication that is required to develop and implement AI solutions for large enterprises. When it comes to sourcing talent, there is a labour shortage for these specific skill sets, which often comes with high salary expectations, and the ability to live and work in an urban tech hub. Companies newer to adopting AI will have to invest substantially to implement AI effectively. While legacy technologies and software systems are prevalent, the cost to upgrade will be even higher, as retraining developers and hiring adds additional cost. Meanwhile, the infrastructure required for AI training and inference include a big investment in computational and storage systems, even while 5G helps with computing at the edge.
Data issues abound
Studies have shown that many companies still lack a strategy and the infrastructure required to access the data that is necessary to implement AI. As the old additive goes “garbage in, garbage out.” Many companies still simply lack clean data, and oftentimes data is in silos and not easily accessible by different departments in the enterprise. When the most simple data strategies and infrastructures are not working, engaging in AI seems like a distant future dream.
Camera-based computer vision isn’t perfect
The use of machine learning in parts inspection is not a new concept. In these systems, cameras capture images of incoming parts and the AI System compares these to a model trained previously with thousands of images previously collected. We’ve learned a few things about why this approach alone is insufficient, and where many enterprises fail to implement AI at scale:
- Building models with enough training data can often be a challenge. Creating a classification system that performs well and identifies essential features requires a representative dataset, which can often be difficult to acquire as there may not be enough balanced data to identify all of the possible categories (i.e., there simply aren’t enough photographed defects).
- AI Model performance can be a “black box”. It’s difficult to identify and correct data errors, to eliminate false positives and reduce false negatives. There’s simply not enough feedback to understand where models go awry.
- AI Models can become outdated as soon as they are deployed. For example, near-perfect lighting in a production environment is difficult to guarantee; minor variations can seriously impact performance, if left unchecked.
Here are DarwinAI’s best practices to implement AI at scale
Optimize the entire AI lifecycle
Here at DarwinAI, when we provide AI products and solutions to our enterprise customers, we solve these common problems by thinking through the entire process:
These are the steps we recommend:
- Collect and audit your data:
- Use Explainability to identify any problems in the dataset, whether it’s image noise or manual label errors.
- Too often we have seen artifacts in training data that can affect the model’s performance. Often, the data scientists aren’t even aware that minor things such as light-reflection or surfaces themselves, can have dramatic effects. With Explainability, you can learn about the quality of the training data and can help make adjustments to collect better data.
2. Calibrate your AI to a custom dataset with Explainability:
- Use Data Augmentation techniques to work with a small data set to account for different operating conditions. Use training algorithms to ensure a balance dataset.
- With our Explainability product, the ‘Black Box’ of AI disappears. Manufacturers will understand exactly how and why their AI models make the right predictions. This enables enterprises to responsibly adopt AI with greater confidence throughout all departments and levels of the enterprise.
3. Ensure that your AI is continuously learning in production:
- Monitor machine learning models in production, providing real time feedback to improve performance automatically.
- This enables a manufacturing enterprise to help their human operators and team members to reskill and perform higher value activities that leverage their unique capabilities, while reducing repetitive strain and workplace injuries. At the same time, parts inspection is greatly improved, and the enterprise realizes cost reductions and additional revenue.
Enable change throughout the enterprise
In order to adopt AI Transformation, many different business units and teams need to be involved. One of the most important groups (if not the most important) are the business leaders at the top, because they will need to shepherd change and adoption with the right policies and governance in place. The mission and vision of the company must embrace all aspects of digital transformation at scale, with the right resourcing, teams, and talent.
Find a proven AI partner that has practical solutions at scale
Manufacturing companies, whether they have internal data science teams or not, should seek out a trusted partner to help them achieve their parts inspection outcomes. These are difficult problems which require the expertise from a partner that can help shoulder the burden of modernizing manufacturing processes.
To learn more about implementing AI within manufacturing, visit DarwinAI’s landing page and sign-up to receive our new white paper on AI use cases in manufacturing and our case study on a smart inspection system, plus get access to our latest video — A panel discussion with DarwinAI’s CEO about the Aspire Food Group project.
DarwinAI, the explainable AI company, enables enterprises to build AI they can trust. DarwinAI’s solutions have been leveraged in a variety of enterprise contexts, including in advanced manufacturing and industrial automation. Within healthcare, DarwinAI’s technology resulted in the development of COVID-Net, an open source system to diagnose Covid-19 via chest x-rays.
To learn more, visit darwinai.com or follow @DarwinAI on Twitter.
If you liked this blog post, click the 👏 below so other people will see this on Medium. For more insights from our team, follow our publication and @DarwinAI. (Plus, subscribe to our letters if you’d like to hear from us!)
