Subscribe Button 1

Quality Inspection and Big Data

The main benefit of deploying AI for quality inspection is significant improvements in defect detection. However, the data generated and stored by inspection systems has the potential to deliver additional benefits, including major improvements in yield. In this article Miron Shtiglitz, director of product management at quality inspection software specialist QualiSense, explores the value of this data in greater depth.

Increasing Yield Through Automating Defect Detection

Anyone working in the world of quality inspection will be aware of the limits of manual inspection and the potential benefits of greater automation, including systems that use artificial intelligence (AI) and deep learning algorithms. With recent advances in AI, the most sophisticated inspection systems available today can reduce the error rate to below one per cent. For manual inspectors in comparison, a host of factors such as fatigue and cognitive bias mean the error rate is usually closer to ten per cent. However, a significantly reduced error rate is not the only area where automation can make an impact.

The Limits of Manual Inspection

On high volume production lines where one hundred per cent inspection is required, manual inspectors do not have the time or capacity to record information about defects. In most cases, a human inspector can do little more than alert their supervisor if there was a higher number of defects on their shift.

If inspectors are required to inspect one or two parts per minute, the most that can be expected is a quick checkbox system. The variety of this data is very limited, which undermines its value. Many factories operate along these lines and in some cases hard copies, rather than digital records, are still the norm.

In some factories, manual inspectors are required to inspect parts at a rate of twenty per minute. In such cases, there is no time to log anything, or store any data pertaining to the type of defect or its frequency. When a defect is detected, the part is simply discarded onto a pile. While this could be reviewed later offline, the information it would yield is very limited as all the parts are mixed up.

At the opposite end of the spectrum, where one hundred per cent inspection is not required and sampling is used instead, any data generated is limited by the lower volume. With a much smaller sample size, the data cannot pinpoint where the problems really lie.

The Three Vs of Big Data

The concept of big data has been around since the 1990s and has gained increasing traction in recent years with the arrival of smart factories. The latter go hand in hand with the arrival of industry 4.0, which aims to provide greater traceability and derive additional data from manufacturing systems. When thinking about big data, it is useful to refer to the three Vs: volume, velocity and variety.

In terms of volume, the introduction of inspection systems that can replace manual inspectors has enormous implications. These systems automatically store data generated about defects on a database. The impact of this will be greatest in applications where there is currently less than one hundred per cent inspection. Once a system is installed, it makes no difference from the machine’s point of view whether it inspects one part every hour or one part every second, so the result will be a massive increase in data for applications where sampling was previously the norm.

From a velocity perspective, the data generated by these systems is vastly superior to that provided by manual inspection, as processing and categorisation is generally done in real-time. Managers can be automatically updated by email or SMS if, for example, there is a sudden increase in the volume of defects during a shift. This does not detract from the value of being able to retrospectively review the data, but the opportunity for real-time alerts, and therefore instantaneous action, is a significant step forward for production managers.

Perhaps most important is the variety of this data. Where manual inspectors can rarely log much more than whether a part if okay or not, quality inspection systems can record more detailed data pertaining to things like type of defect. Systems that utilise deep learning algorithms will be far superior in the variety of data, and the accuracy of that data, when compared to more simple systems that rely on rule-based algorithms. That is one of the benefits of augmented AI.

Unlocking The Potential

The greater the variety of data and the more this data can be correlated with data from other machines and their sensors, the greater the possibilities for optimising production processes. For example, maintenance schedules could be optimised using data that showed the correlations between defect frequency and length of intervals between maintenance activity; correlations between a category of defect and a specific machine or production line could help guide root cause analysis to find where defects were being introduced.

Once gathered, data is automatically stored in a database, where it awaits analysis. At this stage the correct query is key to unlocking the value contained in the data. Standard analytical tools available will allow users to create their own dashboard for this purpose. This data can be integrated with other systems, like a manufacturing execution system, to further increase its potential value.

At QualiSense, our primary mission is developing software that will automate the process of model building when using AI for visual inspection.  Working with leading manufacturers like Johnson Electric has given us access to vast quantities of proprietary data for model training. Our immediate aim is to build a system identifies defects. The data generated by our AI system in defect detection is automatically stored on a database and can be accessed using standard analytical tools. Our longer-term goal however, is to supplement this with the development of our own analytical tools which will help not only spot defects when they arise, but to avoid them in the first place during the design phase.

For more information: