Quality costs — the costs associated with product defects — can cost manufacturers as much as 20% of their entire annual sales revenue, according to experts. Recalls, lawsuits, regulatory fines and social media scorn and outrage are just some of the pain points a manufacturer may experience if a product they made turns out to be defective. So it is understandable that manufacturers dedicate a great deal of money — as much as 50% of sales revenue — to avoid these defects.
In order to minimize losses due to defects, manufacturers need precision during the production process. Identifying and specifying manufacturing defects on a continuous basis is paramount to achieving this. Defects can be caused by issues such as degradation of materials, impact and corrosion, weather, machine or human error, and other factors. Artificial Intelligence (AI)-based systems can be a great help here — but not all AI is created equal. Manufacturers need to understand the methods available and determine which one will help them keep defects to a minimum — or eliminate them altogether.
Visual inspection is one of the most common methods used by manufacturers to check for errors, anomalies and defects that cannot be detected cost-efficiently using other methods (such as spot-checks). Traditionally, visual inspection for quality control of products or machinery is performed by humans — but humans are often not up to the job, because of the large number of items that need to be inspected.
Computers, on the other hand, can process large numbers of inputs quickly and consistently without tiring. Where computers are involved, AI is not far behind — and AI-based defect identification is becoming more common in a large number of industries, reducing or replacing altogether the need for the monotonous, boring and repetitive inspection process that is an open invitation to human error and that could itself be responsible for a defective item being sent out to customers.
In AI inspection, manufacturers gather appropriate data and feed it into a training system, creating algorithms that can be used to inspect equipment, machinery or components rolling off production lines. Since they are computer-based, these inspection systems do not tire, get bored or miss defects like human inspectors can — and they operate around the clock. AI algorithms also self-improve as time goes on and more data is fed into them, providing them with feedback on how to further refine their knowledge.
But even with AI defect detection, there are several approaches — and the reliability of these systems depends to a great extent on the manufacturing operation. There are two primary approaches to identifying defects: Anomaly Detection and Object Detection.
Anomaly Detection aims to identify anomalies, or things that are out of place among the rest of the data. This means technically few-to-no examples of defects could be required for AI training, as they rely on building models based on supplied data of what is “normal” in order to then notice something “abnormal.” Using this method, the detection system simply checks the specifications of a produced item against a checklist of what it is supposed to look like, and approves or disapproves as appropriate. This is appropriate for manufacturing situations where the inspection environment is highly consistent and has low-variance throughout, but lacks robustness — meaning that it is less useful if something in the environment/background suddenly changes, and is identified as anomalous. This method also does not identify the type of defect detected — just that it may be present.
Object Detection, on the other hand, is designed to look for specifically defined objects — or in this case, defects — directly, by understanding and evaluating the specific features of a product. Thus it requires labeled defect examples for its AI training. As a result, it tends to be much more robust in detecting defects due to unrelated changes in the environment, and it can identify the type of defects as well — such as whether a defect is a scratch or a chip, a difference that could be important in terms of the quality, usability or safety of the product.
Truly sophisticated AI detection systems need to make use of deep-learning models, including Convolutional Neural Networks, among other proven state-of-the-art technologies. In addition, these systems should ideally be able to take into account factors beyond just images; data on things such as weather, location, time of day, etc., can make a significant difference in determining the nature and cause of a defect. And, these systems should be customizable, at least to some extent. The systems need to take into account the unique factors and characteristics associated with each specific problem, and ensure compatibility for more-efficient AI-based training, using tried-and-true expert knowledge.
While customization is important — a shoe is not a car, and obviously detection systems appropriate to one are not appropriate to the other — manufacturers do not necessarily need to build their own systems, either. While many off-the-shelf AI anomaly/object-detection solutions do not take into account the factors necessary for high-performance in industrial settings, there do exist services tailored for these types of problems.
By performing due diligence and thoroughly evaluating an AI service — seeking out a service with a record in helping customers, a user-friendly interface, and customizable options for knowledgeable customers, along with investigating the technology behind it, manufacturers can know they are getting a quality service. The success of a manufacturer preventing losses depends on their choosing the right error-remediation service. Such due diligence takes a little effort, but that effort pays off when manufacturers avoid detrimental losses.
Dr. Jonathan Masci is the Director of Deep Learning at NNAISENSE, a Swiss startup bringing AI to industrial process inspection, modeling, and control.
