Data acquisition
t is now over a decade since the largest recall in the history of the automotive industry. In 2013, several automakers began significant recalls due to defect Takata airbags. In 2017, the Japanese parts manufacturer was forced to file for bankruptcy when the amount of money due in compensation was more than was required for its survival. Despite the costs of recalls and the strenuous efforts of quality managers and other production professionals, manufacturing defects are still wreaking havoc a decade on. In May 2023, for example, more than 30 million cars were included in an airbag recall issued by the US National Highway Traffic Safety Administration. Given the costs of defects and the efforts made by manufacturers to detect them, how are so many defects able to avoid detection?
I
THE LIMITS OF MANUAL INSPECTION Defects can begin at any stage of the manufacturing process due to a wide range of factors. It is generally unusual for a defect to be detected at source, instead defective parts are only spotted at the end of the production line where in many cases, they are extremely challenging to spot.
Despite modern developments in machine vision and AI, many manufacturers still rely on human inspectors. Manual inspectors have to work at incredible speed and under continuous pressure. The human eye is remarkable, one of the most complex and sophisticated organs evolved however Human beings are not immune from tiredness, lack of concentration or cognitive bias and a proportion of defective parts inevitably escapes capture at this critical stage. Estimates vary, but most research places the mistake rate anywhere between ten to twenty per cent when using manual inspectors. Depending on the industry, different inspection methods are adopted. If you are only using sampling, rather than inspecting every part individually, it is easy to see how a rarer defect would evade capture. In safety-critical industries though, every part will be inspected and given the shortcomings of manual inspection, many companies have invested in technology to improve the odds of detection.
INVESTING IN AI
AI based quality inspection systems are vastly superior in detecting defects, but most of the existing technology in this area has its limitations. Firstly, the AI system relies on hardware like imaging and other sensors. Relatively subtle changes in environmental conditions, like lighting for example, can create problems here.
The more basic algorithms are only effective at catching simple defects. Whether an AI can spot a defective part before it leaves the plant depends on the sophistication of the algorithm and the use case. Basic rule-based algorithms will only suffice for inspection challenges that are deterministic and the environment is repetitive in nature, but most modern production processes, particularly in safety-critical industries like
60 THE SECRET LIVES OF
MANUFACTURING DEFECTS 2023 saw a number of high-profile recalls linked to manufacturing defects. Despite technological improvements in the field of quality inspection, many safety-critical industries witnessed an increase in the number of defects and recalls. Here Zohar Kantor, chief revenue officer at quality inspection pioneer QualiSense, traces how defects enter the manufacturing process and avoid detection.
aerospace, automotive and med-tech, require deep learning algorithms. For training these AI models, the process is often lengthy and time-consuming. The end result might be a model that is good at detecting common defects, as the model will have seen lots of examples of these. However, we run up against the problem referred to by AI technologists a ‘‘class imbalance,’’ where common defects routinely show up in the training, but rare defects do not. Not surprisingly, the end result is a model therefore fails to catch the outliers.
CONTINUING TO AVOID CAPTURE Due to errors routinely made by manual inspectors and the shortcomings of existing AI, components with minor defects leave the plant. End of line detection is generally effective for single components, but when you go to assembly you have another layer. When multiple components are assembled together, defects that went undetected at an earlier stage of production can often continue to avoid detection. Imagine a simple single gear. A minor defect is not spotted on the surface, as sheet inspection of metal surfaces is notoriously challenging. This gear then becomes part of a transmission, which in turn becomes part of a car. Functional testing of the product may expose a defect in some instances, but in many cases the defective component, buried inside the assembled product, will continue to escape
attention at this stage. The defect can now pass through multiple stages of the production process undetected, as it is no longer visible even with the most thorough inspection process and the most sophisticated AI.
The point at which the defect comes to the attention of the manufacturer is when the product fails within its warranty period or worst still, is shown to violate regulatory requirements. The longer the defects evades detection, the bigger the cost to the manufacturer, as you now have to recall the entire product rather than simply addressing the defective part in isolation.
THE NEXT STEP IN AI
As already observed, although AI has assisted in improving defect detection, existing systems still have their limitations. Thankfully, that is now beginning to change. The hardware that inspection systems rely on is both improving and becoming affordable to more companies. At the same time, we are overcoming the challenges of training AI models for inspection, by automating the process of model training which makes the process faster and more accurate. It is the quality manager that transfers his knowledge in the model creation process, and with augmented AI, the quality manger can easily continue to update the model on the production line, therefore preparing for the most challenging inspection tasks.
QualiSense
qualisense.ai March 2024 Instrumentation Monthly
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55 |
Page 56 |
Page 57 |
Page 58 |
Page 59 |
Page 60 |
Page 61 |
Page 62 |
Page 63 |
Page 64 |
Page 65 |
Page 66 |
Page 67 |
Page 68 |
Page 69 |
Page 70 |
Page 71 |
Page 72 |
Page 73 |
Page 74 |
Page 75 |
Page 76 |
Page 77 |
Page 78 |
Page 79 |
Page 80
