It is a visitor submit by Neslihan Erdogan, International Industrial IT Supervisor at HAYAT HOLDING.
With the continued digitization of the manufacturing processes and Business 4.0, there’s huge potential to make use of machine studying (ML) for high quality prediction. Course of manufacturing is a manufacturing technique that makes use of formulation or recipes to supply items by combining substances or uncooked supplies.
Predictive high quality includes the usage of ML strategies in manufacturing to estimate and classify product-related high quality primarily based on manufacturing course of knowledge with the next objectives[1]:
- High quality description – The identification of relationships between course of variables and product high quality. For example, how does the amount of an adhesive ingredient impact the standard parameters, akin to its power and elasticity.
- High quality prediction – The estimation of a top quality variable on the idea of course of variables for determination assist or for automation. For instance, how a lot kg/m3 adhesive ingredient shall be ingested to realize sure power and elasticity.
- High quality classification – Along with high quality prediction, this entails estimation of sure product high quality sorts.
On this submit, we share how HAYAT HOLDING—a world participant with 41 corporations working in numerous industries, together with HAYAT, the world’s fourth-largest branded diaper producer, and KEAS, the world’s fifth-largest wood-based panel producer—collaborated with AWS to construct an answer that makes use of Amazon SageMaker Mannequin Coaching, Amazon SageMaker Computerized Mannequin Tuning, and Amazon SageMaker Mannequin Deployment to constantly enhance operational efficiency, improve product high quality, and optimize manufacturing output of medium-density fiberboard (MDF) wooden panels.
Product high quality prediction and adhesive consumption advice outcomes may be noticed by discipline specialists by means of dashboards in near-real time, leading to a quicker suggestions loop. Laboratory outcomes point out a major influence equating to financial savings of $300,000 yearly, decreasing their carbon footprint in manufacturing by stopping pointless chemical waste.
ML-based predictive high quality in HAYAT HOLDING
HAYAT is the world’s fourth-largest branded child diapers producer and the most important paper tissue producer of the EMEA. KEAS (Kastamonu Entegre Ağaç Sanayi) is a subsidy of HAYAT HOLDING, for manufacturing within the wood-based panel {industry}, and is positioned because the fourth in Europe and the fifth on this planet.
Medium-density fiberboard (MDF) is an engineered wooden product made by breaking down wooden residuals into fibers, combining it with adhesives, and forming it into panels by making use of excessive temperature and stress. It has many software areas akin to furnishings, cabinetry, and flooring.
Manufacturing of MDF wooden panels requires in depth use of adhesives (double-digit tons consumed every year at HAYAT HOLDING).
In a typical manufacturing line, tons of of sensors are used. Product high quality is recognized by tens of parameters. Making use of the proper quantity of adhesives is a crucial price merchandise in addition to an vital high quality issue for the produced panel, akin to density, screw holding capability, tensile power, modulus elasticity, and bending power. Whereas extreme use of glue will increase manufacturing prices redundantly, poor utilization of glue raises high quality issues. Incorrect utilization causes as much as tens of hundreds of {dollars} in a single shift. The problem is that there’s a regressive dependency of high quality on the manufacturing course of.
Human operators determine on the quantity of glue for use primarily based on area experience. This know-how is solely empirical and takes years of experience to construct competence. To assist the decision-making for the human operator, laboratory exams are carried out on chosen samples to exactly measure high quality throughout manufacturing. The lab outcomes present suggestions to the operators revealing product high quality ranges. However, lab exams will not be in actual time and are utilized with a delay of as much as a number of hours. The human operator makes use of lab outcomes to step by step modify glue consumption to realize the required high quality threshold.
Overview of answer
High quality prediction utilizing ML is highly effective however requires effort and talent to design, combine with the manufacturing course of, and preserve. With the assist of AWS Prototyping specialists, and AWS Accomplice Deloitte, HAYAT HOLDING constructed an end-to-end pipeline as follows:
- Ingest sensor knowledge from manufacturing plant to AWS
- Carry out knowledge preparation and ML mannequin technology
- Deploy fashions on the edge
- Create operator dashboards
- Orchestrate the workflow
The next diagram illustrates the answer structure.
Information ingestion
HAYAT HOLDING has a state-of-the artwork infrastructure for buying, recording, analyzing, and processing measurement knowledge.
Two forms of knowledge sources exist for this use case. Course of parameters are set for the manufacturing of a specific product and are often not modified throughout manufacturing. Sensor knowledge is taken through the manufacturing course of and represents the precise situation of the machine.
Enter knowledge is streamed from the plant through OPC-UA by means of SiteWise Edge Gateway in AWS IoT Greengrass. In whole, 194 sensors have been imported and used to extend the accuracy of the predictions.
Mannequin coaching and optimization with SageMaker computerized mannequin tuning
Previous to the mannequin coaching, a set of knowledge preparation actions are carried out. For example, an MDF panel plant produces a number of distinct merchandise on the identical manufacturing line (a number of sorts and sizes of wooden panels). Every batch is related to a unique product, with totally different uncooked supplies and totally different bodily traits. Though the gear and course of time collection are recorded constantly and may be seen as a single-flow time collection listed by time, they have to be segmented by the batch they’re related to. For example, in a shift, product panels could also be produced for various durations. A pattern of the produced MDF is shipped to the laboratory for high quality exams on occasion. Different characteristic engineering duties embrace characteristic discount, scaling, unsupervised dimensionality discount utilizing PCA (Principal Part Evaluation), characteristic significance, and outlier detection.
After the info preparation section, a two-stage method is used to construct the ML fashions. Lab check samples are performed by intermittent random product sampling from the conveyor belt. Samples are despatched to a laboratory for high quality exams. As a result of the lab outcomes can’t be introduced in actual time, the suggestions loop is comparatively sluggish. The primary mannequin is educated to foretell lab outcomes for product high quality parameters: density, elasticity, pulling resistance, swelling, absorbed water, floor sturdiness, moisture, floor suction, and bending resistance. The second mannequin is educated to suggest the quantity of glue for use in manufacturing, relying on the anticipated output high quality.
Establishing and managing customized ML environments may be time-consuming and cumbersome. Amazon SageMaker gives a set of built-in algorithms, pre-trained fashions, and pre-built answer templates to assist knowledge scientists and ML practitioners get began on coaching and deploying ML fashions shortly.
A number of ML fashions have been educated utilizing SageMaker built-in algorithms for the highest N most produced product sorts and for various high quality parameters. The standard prediction fashions determine the relationships between glue utilization and 9 high quality parameters. The advice fashions predict the minimal glue utilization to fulfill high quality necessities utilizing the next method: an algorithm begins from the best allowed glue quantity and reduces it step-by-step if all necessities are glad till the minimal quantity of glue allowed. If the max quantity of glue doesn’t fulfill all the necessities, it provides an error.
SageMaker computerized mannequin tuning, often known as hyperparameter tuning, finds the most effective model of a mannequin by operating many coaching jobs in your dataset utilizing the algorithm and ranges of hyperparameters that you simply specify. It then chooses the hyperparameter values that end in a mannequin that performs the most effective, as measured by a metric that you simply select.
With computerized mannequin tuning, the crew targeted on defining the suitable goal, scoping the hyperparameters and the search house. Computerized mannequin tuning takes care of the remainder, together with the infrastructure, operating and orchestrating coaching jobs in parallel, and enhancing hyperparameter choice. Computerized mannequin tuning gives a variety of coaching occasion sorts. The mannequin was fine-tuned on c5.x2large occasion sorts utilizing an clever model of hyperparameter tuning strategies that’s primarily based on the Bayesian search principle and is designed to search out the most effective mannequin within the shortest time.
Inference on the edge
A number of strategies can be found for deploying ML fashions to get predictions.
SageMaker real-time inference is right for workloads the place then are real-time, interactive, low-latency necessities. Throughout the prototyping section, HAYAT HOLDING deployed fashions to SageMaker internet hosting providers and bought endpoints which might be absolutely managed by AWS. SageMaker multi-model endpoints present a scalable and cost-effective answer for deploying massive numbers of fashions. They use the identical fleet of assets and a shared serving container to host all of your fashions. This reduces internet hosting prices by enhancing endpoint utilization in contrast with utilizing single-model endpoints. It additionally reduces deployment overhead as a result of SageMaker manages loading fashions in reminiscence and scaling them primarily based on the visitors patterns to your endpoint.
SageMaker real-time inference is used with multi-model endpoints for price optimization and for making all fashions accessible always throughout growth. Though utilizing an ML mannequin for every product sort ends in larger inference accuracy, the price of creating and testing these fashions will increase accordingly, and it additionally turns into tough to handle a number of fashions. SageMaker multi-model endpoints tackle these ache factors and provides the crew a speedy and cost-effective answer to deploy a number of ML fashions.
Amazon SageMaker Edge gives mannequin administration for edge units so you may optimize, safe, monitor, and preserve ML fashions on fleets of edge units. Working ML fashions on edge units is difficult, as a result of units, not like cloud cases, have restricted compute, reminiscence, and connectivity. After the mannequin is deployed, it’s good to constantly monitor the fashions, as a result of mannequin drift may cause the standard of mannequin to decay additional time. Monitoring fashions throughout your gadget fleets is tough as a result of it’s good to write customized code to gather knowledge samples out of your gadget and acknowledge skew in predictions.
For manufacturing, the SageMaker Edge Supervisor agent is used to make predictions with fashions loaded onto an AWS IoT Greengrass gadget.
Conclusion
HAYAT HOLDING was evaluating a complicated analytics platform as a part of their digital transformation technique and wished to convey AI to the group for high quality prediction in manufacturing.
With the assist of AWS Prototyping specialists and AWS Accomplice Deloitte, HAYAT HOLDING constructed a singular knowledge platform structure and an ML pipeline to handle long-term enterprise and technical wants.
HAYAT KIMYA built-in the ML answer in one among its vegetation. Laboratory outcomes point out a major influence equating to financial savings of $300,000 yearly, decreasing their carbon footprint in manufacturing by stopping pointless chemical waste. The answer gives a quicker suggestions loop to the human operators by presenting product high quality predictions and adhesive consumption advice outcomes by means of dashboards in near-real time. The answer will finally be deployed throughout HAYAT HOLDING’s different wooden panel vegetation.
ML is a extremely iterative course of; over the course of a single venture, knowledge scientists practice tons of of various fashions, datasets, and parameters searching for most accuracy. SageMaker presents probably the most full set of instruments to harness the ability of ML. It enables you to arrange, monitor, examine, and consider ML experiments at scale. You’ll be able to enhance the bottom-line influence of your ML groups to realize important productiveness enhancements utilizing SageMaker built-in algorithms, computerized mannequin tuning, real-time inference, and multi-model endpoints.
Speed up time to outcomes and optimize operations by modernizing your enterprise method from edge to cloud utilizing Machine Learning on AWS. Benefit from industry-specific improvements and options utilizing AWS for Industrial.
Share your suggestions and questions within the feedback.
About HAYAT HOLDING
HAYAT HOLDING, whose foundations have been laid in 1937, is a world participant right this moment, with 41 corporations working in numerous industries, together with HAYAT within the fast-moving shopper items sector, KEAS (Kastamonu Entegre Ağaç Sanayi) within the wood-based panel sector, and LIMAS within the port administration sector, with a workforce of over 17,000 individuals. HAYAT HOLDING delivers 49 manufacturers produced with superior applied sciences in 36 manufacturing services in 13 nations to thousands and thousands of shoppers worldwide.
Working within the fast-moving shopper items sector, Hayat was based in 1987. At this time, quickly advancing on the trail of globalization with 21 manufacturing services in 8 nations around the globe, Hayat is the world’s fourth-largest branded diaper producer and the most important tissue producer within the Center East, Jap Europe, and Africa, and a significant participant within the fast-moving shopper items sector. With its 16 highly effective manufacturers, together with Molfix, Bebem, Molped, Joly, Bingo, Check, Has, Papia, Familia, Teno, Focus, Nelex, Goodcare, and Evony within the hygiene, house care, tissue, and private well being classes, Hayat brings HAYAT* to thousands and thousands of properties in additional than 100 nations.
Kastamonu Entegre Ağaç Sanayi (KEAS), the primary funding of HAYAT HOLDING in its industrialization transfer, was based in 1969. Persevering with its uninterrupted development in the direction of changing into a world energy in its sector, it ranks fourth in Europe and fifth on this planet. KEAS ranks first within the {industry} with its roughly 7,000 staff and exports to greater than 100 nations.
*“Hayat” means “life” in Turkish.
References
- Tercan H, “Machine studying and deep studying primarily based predictive high quality in manufacturing: a scientific evaluation”, Journal of Clever Manufacturing, 2022.
In regards to the authors
Neslihan Erdoğan, (BSc and MSc in Electrical Engineering), held varied technical & enterprise roles as a specialist, architect and supervisor in Data Applied sciences. She has been working in HAYAT because the International Industrial IT Supervisor and led Business 4.0, Digital Transformation, OT Safety and Information & AI tasks.
Çağrı Yurtseven (BSc in Electrical-Electronics Engineering, Bogazici College) is the Enterprise Account Supervisor at Amazon Internet Companies. He’s main Sustainability and Industrial IOT initiatives in Turkey whereas serving to clients notice their full potential by displaying the artwork of the potential on AWS.
Cenk Sezgin (PhD – Electrical Electronics Engineering) is a Principal Supervisor at AWS EMEA Prototyping Labs. He helps clients with exploration, ideation, engineering and growth of state-of-the-art options utilizing rising applied sciences akin to IoT, Analytics, AI/ML & Serverless.
Hasan-Basri AKIRMAK (BSc and MSc in Pc Engineering and Government MBA in Graduate College of Enterprise) is a Principal Options Architect at Amazon Internet Companies. He’s a enterprise technologist advising enterprise section shoppers. His space of specialty is designing architectures and enterprise circumstances on massive scale knowledge processing programs and Machine Studying options. Hasan has delivered Enterprise growth, Methods Integration, Program Administration for shoppers in Europe, Center East and Africa. Since 2016 he mentored tons of of entrepreneurs at startup incubation packages pro-bono.
Mustafa Aldemir (BSc in Electrical-Electronics Engineering, MSc in Mechatronics and PhD-candidate in Pc Science) is the Robotics Prototyping Lead at Amazon Internet Companies. He has been designing and creating Web of Issues and Machine Studying options for among the largest clients throughout EMEA and main their groups in implementing them. In the meantime, he has been delivering AI programs at Amazon Machine Studying College and Oxford College.