FEATURE Industrial AI
Evie is your AI for manufacturing work instructions
LEARNING FROM DEFECTS
Garth Coleman, CEO, Canvas Envision, explains what happens after the vision system finds the fault, and says the cheapest defect is the one that never happens
A
modern assembly line is extraordinarily well instrumented for noticing something has gone wrong.
A vision system catches a misaligned component in milliseconds. Torque data flags a fastener outside tolerance before the operator has moved on. The PLC logs it, the manufacturing execution system records it, and the quality system has a trend before the shift ends. The detection is instant, and nothing that follows it is. Someone has to work out what changed, decide what the line should do differently, and get it to the people building it. Until then they follow the instructions already at the station, written when the process was designed and describing how the job should go rather than how it is going wrong. The machine half of that loop closes inside a shift. The human half takes weeks. Vanson Bourne surveyed UK, US, French and German manufacturers and found 23 percent of unplanned downtime is caused by user error, against as little as 9 percent elsewhere. The instinct is to read that as a workforce problem. A 2021 study in Applied Sciences locates the cause more precisely. Across three assembly tasks, the authors catalogued 59 ways the work could go wrong and named the instructions as a major cause. Workers were not getting the spatial and functional information the job needs in a form they could absorb. The remedy they rated highest was making
instructions visual, which is not new. In 2008, researchers at Queen’s University Belfast compared text, static diagrams and animation on a mechanical assembly task. First-build times
16 September 2026 | Automation
using animation were 37 percent quicker than text and 16 percent quicker
than static diagrams, the advantage greatest on the first attempt. That work is eighteen years old, and yet most factories still issue text and static diagrams. The reason is not indifference. Updating an instruction properly means re-sequencing steps, re- rendering views, re-shooting reference imagery, and routing it through approval- historically, days of engineering time. When a vision system flags a recurring alignment issue on a Tuesday, nobody spends
three days rebuilding the
instruction that would prevent it. The finding stays in the quality system, the operator carries on with the document they have, and the defect returns two weeks later. That calculation is what AI is changing,
upstream of the operator rather than in front of them. In our own platform, Evie, the AI built into Canvas Envision, works inside the authoring environment. Point it at the 3D assembly and its 3D Assembly Instruction agent returns steps with camera angles set, parts exploded, and callouts labelled. An engineer fixing the step where the line keeps going wrong can reset the view, isolate the part, add the callout, and animate the sequence in less time than finding the right screenshot once taken. The animation Belfast measured becomes a by-product of the work,
not a project of its own. Because the AI is trained on the platform’s Create SDK, it works through the same functions an author would use rather than merely generating text about them. That is the difference between a system that can explain an exploded view and one that can actually build it. The engineer validates and refines what comes back, and on safety-critical work that review is not a step to be optimised away. None of this displaces the vision system. It
produces the signal; the instruction produces the action. When the corrected instruction reaches the operator that shifts, and their confirmation flows back to the systems that raised the alarm, the finding becomes an input to the work, not a record of it.
Every defect a vision system catches has
already consumed the material, machine time, and labour behind it, so catching it limits the loss without avoiding it. The cheapest defect is the one that never happens, and that is decided upstream, in how well the instruction was written and how fast it can be rewritten when the line says it is wrong. Manufacturers have spent two decades learning to detect where work goes wrong. Turning what they find into better instructions fast enough to stop it happening again stayed out of reach for just as long. It is not out of reach any longer. References: 1. Vanson Bourne for ServiceMax from GE Digital, After the Fall: The Costs, Causes and Consequences of Unplanned Downtime, 2017. Survey of more than 100 manufacturers and 350 companies across the UK, US, France and Germany. 2. Torres, Y., Nadeau, S. and Landau, K., “Classification and Quantification of Human Error in Manufacturing: A Case Study in Complex Manual Assembly”, Applied Sciences, 2021, 11(2), 749.
https://doi.org/10.3390/ app11020749 3. Watson, G., Curran, R., Butterfield, J. and Craig, C., “The Effect of Using Animated Work Instructions Over Text and Static Graphics When Performing a Small Scale Engineering Assembly”, in Collaborative Product and Service Life Cycle Management for a Sustainable World: Proceedings of the 15th ISPE International Conference on Concurrent Engineering (CE2008), Springer, 2008. https://
doi.org/10.1007/978-1-84800-972-1_51
Canvas Envision
www.canvasenvision.com
automationmagazine.co.uk
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
