Performance Evaluation of AI Algorithms on Heterogeneous Edge Devices for Manufacturing

Novel Artificial Intelligence (AI) approaches try to process an excessive amount of field-level data. However, challenges arise as network bandwidth is limited, and thus this data cannot be entirely transferred to the cloud for further processing. Edge computing tries to overcome that limitation by bringing the computational resources closer to the data generating sources. However, edge devices are also constraint by both CPU power and memory, and strict real-time requirements of the manufacturing domain have to be met. Thus, the selection of suitable devices for specific AI algorithms poses a severe challenge. Currently, the choice is often made by a trial-and-error approach or by selecting more powerful devices than needed. This paper tries to address those challenges by showing relevant aspects for algorithm benchmarking in the manufacturing domain. Selected algorithms, namely Grubbs Test, Butterworth Filter, DBSCAN, Random Forest, Support Vector Machine, Matrix Multiplication, and Matrix Inversion, are examined. Analysis of their theoretical time and space complexity sheds some light on the behaviour of the algorithms with respect to their input data points. In addition, relevant metrics for the manufacturing domain, such as execution time, memory consumption, and energy consumption, are identified. This paper furthermore examines the algorithm behaviour on various heterogeneous hardware devices, such as PLCs, an MCU, IPCs, a single-board computer, and a dedicated edge device. Altogether, this paper can guide selecting suitable algorithms and hardware to equip Cyber-Physical Production Systems (CPPS) with novel data processing solutions thoughtfully. Moreover, the presented metrics can support the creation of novel ML benchmarks for smart manufacturing (SM).