Serverless Edge Computing für industrielle Prozesse

Industry 4.0 aims to enhance industrial processes by introducing smart edge systems, leveraged by data collection and processing of sensor readings. This improves efficiency, communication and flexibility in industrial processes. Especially, real-time processing of sensor data benefits factories as it enables the use of data-based optimization techniques, such as deploying machine learning models for detecting anomalies in manufacturing. Serverless computing has the potential to play an important role as an enabling technology in industrial automation, as it can facilitate the flexible deployment of service-oriented software architectures (potentially also facilitating customization of product lines) with a fine granularity and the ability to scale resources appropriately following compute demand that may vary over time. Especially in resource-limited edge scenarios, serverless can act as a crucial enabler. In case of sudden and unpredictable spikes in computation
load, having the option to use different providers helps to reduce computing stress on the overall system and ensures frictionless operation. In order to maximize performance of serverless in industrial scenarios, we leverage a hybrid edge-cloud setting and propose an intelligent serverless workload scheduler based on Deep Reinforcement Learning (DRL), which decides on the best-fitting serverless instance based on specific requirements. This scheduler is an integral component of an end-to-end system architecture that we design and implement, with the aim of supporting Industry 4.0 application scenarios. Our architecture focuses on efficient data ingestion and interoperability, supporting the widely
adopted OPC UA standard for sensor data collection from industrial equipment and building on the NGSI-LD standard to facilitate application integration and data exchange. We demonstrate the feasibility and industrial relevance of our approach by applying it to a realistic use case from the domain of lightweight metal manufacturing. Furthermore, we show that our DRL-based scheduler improves overall serverless function serving performance significantly depending on the defined priorities, such as function invocation reliability, response time, cold start mitigation and/or avoiding CPU overload.