Leveraging programmable logic controllers for machine learning applications in industrial setups

Machine Learning (ML) has become a powerful tool for addressing complex classification and regression tasks in industrial settings. Despite the widespread use of programmable logic controllers (PLCs) in these environments, ML models are typically executed on external computing devices due to inherent PLC limitations: restricted memory capacity, long cycle times, and limited computational power. These constraints hinder the direct deployment of ML algorithms on PLCs and often require additional hardware, increasing system complexity and deployment costs. This paper addresses this challenge by demonstrating the direct implementation of four ML algorithms on a Siemens S7-1516 PLC: Linear Regression, Logistic Regression, k-Nearest Neighbors (kNN), and a Neural Network. In addition, a real-world laboratory prototype modeling the thermal behavior of an industrial 3D printer is presented to illustrate the practical applicability of our research findings. To overcome PLC resource constraints, two optimization strategies are proposed: (1) algorithmic adaptations to reduce execution cycle time, and (2) a custom, high-performance matrix multiplication library designed to replace the naive implementation common in standard tools. We evaluate the performance of the proposed implementations across datasets of varying sizes, comparing the standard TIA Portal functions to our optimized library. The results demonstrate that that the optimized implementations achieve acceptable cycle times and consistently outperform the baseline, confirming the feasibility of efficient, native ML execution on PLCs. These findings open new avenues for embedding ML capabilities directly into PLC-based automation systems, enabling smarter and more autonomous industrial control.