Machine-Learning-Based Carbon Footprint Management in the Frozen Vegetable Processing Industry

Abstract

In the paper, we present a method of automatic evaluation and optimization of production processes towards low-carbon-emissions products. The method supports the management of production lines and is based on unsupervised machine learning methods, i.e., canopy, k-means, and expectation-maximization clusterization algorithms. For different production processes, a different clustering method may be optimal. Hence, they are validated by classification methods (k-nearest neighbors (kNN), multilayer perceptron (MLP), binary tree C4.5, random forest (RF), and support vector machine (SVM)) that identify the optimal clusterization method. Using the proposed method with real-time production parameters for a given process, we can classify the process as optimal or non-optimal on an ongoing basis. The production manager can react appropriately to sub-optimal production processes. If the process is not optimal, then during the process the manager or production technologist may change the production parameters, e.g., speed up or slow down certain batches, so that the process returns to the optimal path. This path is determined by a model trained via the proposed method based on the selected clustering method. The method is verified on an onion production line with more than a hundred processes and then applied to production lines with a smaller number of cases. We use data from real-world measurements from a frozen food production plant. Our research demonstrates that proper process management using machine learning can result in a lower carbon footprint per ton of the final product.

Keywords

• management
• carbon footprint
• machine learning