How to Predict Defects with Machine Learning - Acting Instead of Reacting, Part 2
Even with the highest requirements for quality control, up to 4% of production in the metal industry has to be scrapped. Most defects occur in the early stages of the production process, but are often only discovered during finishing. But what if the software could automatically predict which defects will occur and adjust production processes accordingly in order to avoid them? Machine Learning-based prediction quality can predict up to 75% of defects, most of which can then be avoided. Learn how ML is used in order to predict production-related defects.
The Dilemma of Unplanned Events
Every unplanned event causes a reaction and finally an action. The path to action, however, is often a long and difficult one:
- Insight phase: time to actually realize that the event occurred.
- Analysis phase: time to understand what happened.
- Decision phase: time to decide what we want to do about it.
- Action phase: time to take measures to address the event.
The more severe the event and the longer the time between occurrence and action, the more value is potentially lost. On the other hand, shortening this time bears great potential.
Industry 4.0 shows its real value by improving all the phases of addressing an unplanned event, and shortening the time between the occurrence and the action.
Further, it also allows the shift from reactive decision-making (focused on correctly addressing the situation that has already occurred) to proactive decision-making (based on expected future events).
Big Data in Metals Production
The digitalization of the manufacturing process delivers a huge amount of digital data from sensors, production control, quality control, material genealogy, material logs, process logs, equipment wear, maintenance data, and so on.
Most people think that they already collect more data than they can handle, and this amount is only expected to grow.
The truth is that most industrial companies do not know how to correctly make use of all the data.
They collect the data as offered and store it to a database where it is left unexamined. This data is then occasionally used for incident resolution and some high-level aggregated analysis, at best.
At the same time, the production logs can already become a valuable source of so-called “training data” for ML algorithms. While many industrial internet of things (IIoT) initiatives focus on installing new sensors to collect and transmit more data about ongoing production processes, it often makes sense to start the digitalization process by looking at the data already collected by existing process control systems. This way, industrial companies can make more out of previous wave of automation and, by applying novel machine learning technologies, receive tangible business improvements without major capital costs or changes to the processes.
Quality Prediction as Early as Possible
Zero defect manufacturing is still impossible for most industries nowadays. Defect rates of 5% and more are no exception for metals production even in our digital world of manufacturing with highly automated and quality monitored production processes. Test procedures and quality control are very expensive and form up a significant part of the total production cost, thus cannot be applied with scrutiny to every material.
The data availability and advanced analysis techniques open new opportunities to further reduce the defect rate significantly. This data should allow quality management to predict defects at an early stage of the production process.
Most defects are related to the early stages of the production process (casting and milling) but often only discovered at the finishing stages of the product. Defects that are detected late (so-called “hidden defects”) lead to significant cost due to unnecessary production steps, as well as to delivery delays as these orders may have to be fulfilled by starting a new production sequence with a new material.
If defects can be predicted early, most of them could be avoided by adapting the remaining production route or process, called defect handling. In case certain defects cannot be treated, a material can be downgraded to a different order, and a new material started early on to fulfil the current order. A pro-active approach to defect handling based on predictions has the potential to improve the yield of the overall production drastically.
Quality prediction as early as possible in the production stage allows production management to avoid defects for the end product.
Defect Prediction With Galton Board
Let's compare the steel production process with a pinball machine or a Galton board which brings to life the statistical concept of normal distribution. When rotating the Galton Board around its axis, a stream of steel beads that are equally likely to jump left or right through several rows of pegs are set in motion. As the beads accumulate in the bins, they approach the bell curve, as shown by the yellow line on the front of the Galton board which allows visualizing the order embedded in the chaos of randomness.
This could be compared to a production order following a production process from iron ore to the final product. Errors can occur on the various routes during the production process.
- But how can we predict defects that will only appear in the future?
- Can we learn to predict the future from the past?
- Can we learn to foresee future defects by looking at past defects?
Possibilities to Avoid Predicted Errors
- Selecting a different production route
- Revising the remaining production steps and their processing parameters, e.g. by slowing down the coating process for this material in order to avoid a problem with the coating surface
- Adding an additional rework step to prevent or correct the defect
- Reassigning the material unit that is likely to lead to a defect to another less demanding production order, e.g. some less visible car parts require less surface perfection than other car parts
Machine Learning for Quality Prediction
A promising prediction technique is using ML models, which comprehensively examine all possible combinations of effects of production factors on quality indicators and defect types. The digitalization of the manufacturing process delivers a huge amount of (unlabeled) digital data from sensors, production control, quality control, material genealogy and others. By using new AI-concepts for ML the availability and combination of this data opens a new opportunity to reach zero defect production. Based on historical defects and all its related process and production data, the data are qualitatively labeled and defect prediction model can be extracted in order to predict future quality defects as early as possible. The reliability of such automatic prediction reaches up to 75%. ML algorithms find correlations between data and label them in a way that experts are not even aware of.
In addition to avoiding defects, the solution even offers the chance for yet unknown improvements.
What's your opinion on this topic?
Director Marketing PSI Metals GmbH
After taking over the marketing department of PSI Metals in 2015 Raffael Binder immediately positioned the company within the frame of Industry 4.0. So it is no wonder that in our blog he covers such topics as digitalization, KPIs and Artificial Intelligence (AI). Raffael’s interests range from science (fiction) and history to sports and all facets of communication.
+43 732 670 670-61