How Element AI's Visual Anomaly Detection Model Reduces Defect Rates
Jean-François Marcil Jean-François Marcil
October 21 5 min

How Element AI's Visual Anomaly Detection Model Reduces Defect Rates

Whether it's due to product defects that stem from subpar raw materials or a glitch in the production process, the total cost of poor quality (COPQ) in manufacturing can be steep. Manual visual inspection aided by machine vision has been a reliable workhorse in decreasing the number of defective products.

These traditional methods are not without challenges: set-up times can be lengthy, and the system needs regular maintenance and monitoring to ensure it's functioning as it should. Since manual inspection is still widely used, knowledge that comes from experience remains limited to a select number of operators. There's no centralized source of truth. The effectiveness of manual labor also varies widely depending on alertness and the person at the wheel.

Deep Learning Solutions Offer an Alternative

The growth of artificial intelligence (AI) offers an attractive alternative: supervised deep learning networks that train on vast databases of images so machine learning (ML) algorithms can help consistently and accurately identify defects. While human operators are still in charge, ML does the grunt work, flagging only problems for closer inspection.

ML algorithms operate using neural networks that learn every instance of what "normal" and "defective" looks like. When a product comes down the assembly line, the algorithm compares and contrasts a picture of the product taken by a camera against its database of existing pictures. Using a pattern-matching technique, the product then gets the thumbs-up if it looks like a normal one. If the product image matches that of a defect, the algorithm alerts the operator to take a closer look.

The machine "learns" by continuously adding these images to its growing database and by understanding what constitutes normal and abnormal. If, for example, the operator okays a defective-looking image as normal, the algorithm saves that information for future use, growing systematically more intelligent over time.

The problem with such traditional supervised deep learning models is that the initial set-up is resource-intensive. Since these algorithms work by learning what both good and bad look like, the databases have to account for every single instance of each no matter how improbable. This becomes especially challenging when defects occur due to a number of different circumstances: raw materials, differences in production, etc. Every single bottleneck has to be examined.

So, while traditional supervised deep learning models are valuable, the large upfront costs can be prohibitive.

A More Intelligent Visual Anomaly Detection Way

Element AI flips the traditional deep learning model on its head by working with semi-supervised and unsupervised equivalents. Unsupervised deep learning algorithms zero in exclusively on what normal looks like. The more sophisticated ML looks for unusual patterns in products instead of executing exact pattern-matching. Any outlier that doesn't conform or looks suspicious gets flagged.

While traditional ML needs thousands of images to get started, Element AI's visual anomaly detection (VAD) model can get started with as few as 200 images of good components for it to start recognizing outliers.

When quality specialists are especially concerned about a specific type of defect, the Element AI VAD model also allows them to upload a couple of images of the defect (about 10, compared to the 1000s needed with supervised deep learning solutions) to increase the accuracy at which the model can identify that particular anomaly faster.

Advantages of the Element AI VAD model

The Element AI VAD model requires a lower barrier of entry and is accessible to more manufacturing facilities. The versatile solution works with existing cameras and data-gathering infrastructure, so production plants don't have to reconfigure their operating systems to incorporate the VAD model into their inspection lineup.

Because the model doesn't need to train on images of anomalies, it can recognize defects supervised learning systems can miss. This leads to increased quality standards and customer satisfaction. The solution is also easily scalable, as the same model can support multiple production lines.

Beyond Product Defect Identification

Since the essential principle of the Element AI VAD model is that it is trained on images of what's normal, use cases can extend beyond identifying good and bad products on the manufacturing line.

In time, the model could learn to correlate defects to causes, providing root cause analysis. For example, it could determine that a specific type of defect is caused by a machine's component that is overheating. By identifying the root cause of the defect, operators can address the issue before it has the chance to whiplash and create further problems down the line. Such proactive monitoring could result in significant cost savings.

The Element AI VAD model is a smarter, leaner approach to deep learning models. Its low requirements and easy accessibility make it an attractive defect-detection engine for manufacturing companies across a variety of industries.

Interested in learning more about the Element AI VAD model and how it can help you reduce defect rates? Contact us to speak to one of our AI experts.