Optimization of By-product Rendering Parameters using Machine Learning
Jedari Heidarzadeh, Ali (2025)
2025
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2025-11-25
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2025112410854
https://urn.fi/URN:NBN:fi:tuni-2025112410854
Tiivistelmä
Animal by-product rendering is an essential yet underutilized domain in food manufacturing, transforming such biomasses into input materials for high-value products, such as animal feed, fertilizers, and biofuels. However, rendering these by-products is challenging due to their time-varying quality, as well as strict EU regulations, which increase operational complexity. As a result, process engineers typically follow a set of previously certified parameter settings that have proven to yield acceptable, but not top-grade, quality.
This research addresses this limitation by first designing and implementing a comprehensive 4-step feature selection pipeline to identify the most influential attributes affecting the product quality. Building on these insights, machine learning models are developed to predict optimal process parameters based on production data and external data sources. Nevertheless, the effectiveness of model-predicted parameters cannot be validated directly through before-and-after quality comparisons, as the raw material mixture cannot be estimated, and batches of materials cannot be tracked throughout the process in continuous processes. Moreover, untested process parameters risk operation disruptions to the production line.
To overcome these challenges, a framework is proposed consisting of two primary stages: offline and online validation. The offline validation stage of this framework incorporates quality prediction techniques to estimate product quality in hypothetical cases created from historical data combined with the predicted optimal parameters, calculating the expected improvements. Subsequently, the online validation applies the predicted process parameters to the production line and collects data during operation. The capability of the quality prediction models in accurately estimating the product quality based on the collected data confirms the validity of these models, as well as the process parameter optimization models and the improvements achieved in the offline stage.
Finally, an animal by-product rendering plant operated by a leading Finnish company is considered for a case study, demonstrating the effectiveness of the proposed methodology. Feature selection is executed through the mentioned 4-step pipeline, and machine learning models are built for both process parameter optimization and quality prediction. Data collected during operation under the model-predicted parameters are analyzed and validated through the proposed framework, yielding a 33.7% improvement in the final product's protein consistency with preliminary evidence highlighting promising positive environmental impacts in greenhouse gas emissions.
This research addresses this limitation by first designing and implementing a comprehensive 4-step feature selection pipeline to identify the most influential attributes affecting the product quality. Building on these insights, machine learning models are developed to predict optimal process parameters based on production data and external data sources. Nevertheless, the effectiveness of model-predicted parameters cannot be validated directly through before-and-after quality comparisons, as the raw material mixture cannot be estimated, and batches of materials cannot be tracked throughout the process in continuous processes. Moreover, untested process parameters risk operation disruptions to the production line.
To overcome these challenges, a framework is proposed consisting of two primary stages: offline and online validation. The offline validation stage of this framework incorporates quality prediction techniques to estimate product quality in hypothetical cases created from historical data combined with the predicted optimal parameters, calculating the expected improvements. Subsequently, the online validation applies the predicted process parameters to the production line and collects data during operation. The capability of the quality prediction models in accurately estimating the product quality based on the collected data confirms the validity of these models, as well as the process parameter optimization models and the improvements achieved in the offline stage.
Finally, an animal by-product rendering plant operated by a leading Finnish company is considered for a case study, demonstrating the effectiveness of the proposed methodology. Feature selection is executed through the mentioned 4-step pipeline, and machine learning models are built for both process parameter optimization and quality prediction. Data collected during operation under the model-predicted parameters are analyzed and validated through the proposed framework, yielding a 33.7% improvement in the final product's protein consistency with preliminary evidence highlighting promising positive environmental impacts in greenhouse gas emissions.