This unprecedented blend of 5G and ML expertise creates a unique perspective on ML-supported industrial communications and their role in facilitating industrial automation. Their impact is expected to increase in the coming years, given the abundant data growth and the rapid availability of powerful edge computers in production facilities.
Depending on the application, ML has in the last years proven to be a powerful tool for classification of events, detection of anomalies, prediction, and tactical decision making. This post will explore the key advantages of ML-based industrial communications in data acquisition in real systems.
ML and 5G in manufacturing
Manufacturing is experiencing a trend from mass-production to mass-customization, with the time and cost of reconfiguring production lines posing major challenges. To achieve profitable operation with a batch-size of one, factory owners invest heavily in digitalization, machines become more versatile, human-machine collaboration is becoming more sophisticated, and flexible production stations replace fixed production lines.
The adoption of 5G and the application of ML techniques appear promising to solve several industrial challenges of modern production enterprises. ML enables synergies of production and communication worlds in industrial environments.
Given the nature and complexity of the dynamic factory environment, application requirements, and data traffic, automatization and optimization require adopting ML approaches. A challenging task before applying ML approaches is to find a proper classification of different system states and associated traffic patterns. Such classification can help predict bottlenecks and avoid them by properly load balancing, i.e., dynamically offloading computationally expensive tasks from the devices to the edge cloud. Having found such a classification scheme, ML approaches such as reinforcement learning (RL) can be applied to predict load changes or optimize the function placement and scheduling.
Advantages of ML and 5G
Here are the key advantages of ML and 5G wireless communication in industrial manufacturing.
- Scalability of industrial control systems and flexibility of the factory floor – ML-based prediction methods allow proactive adaptation of industrial applications based on predicted network and IT conditions and classification methods to improve dynamic offloading to the edge cloud based on classified system states and traffic patterns.
- The flexibility of production facilities and adaptability to changes with minimum downtimes – Object recognition and SLAM algorithms enable 3D reconstruction of the industrial environment for network planning and transfer learning for network reconfiguration.
- Dynamicity of heterogeneous services while providing reliability and isolation – RL algorithms for learning optimal slice parameters configuration and resource allocation, and unsupervised learning for anomaly detection.
- Mobility and management of robots and vehicles while guaranteeing safety and security – Pattern recognition algorithms and correlation analysis learn mobility patterns, support QoS prediction, resource management, and distributed or federated learning to balance computing and network resources.
Traditional communication approaches focus on maintaining QoS primarily with reactive methods, e.g., embedding pilot signals in the transmissions and conducting measurements on the receiver side. However, thanks to recent advances in data-driven ML techniques, there are now tools that can help enhance future communication systems with proactivity. It applies particularly well in a factory environment because AGVs (automated guided vehicles) and robots have designated pathways and parking areas, moves between a predefined number of stations, and have repetitive mobility patterns.
ML algorithms, namely correlation analysis, and pattern recognition algorithms can help find such correlations and patterns automated. Once the future location is known and the radio conditions in space and time can be predicted with a certain accuracy, the communication and network management systems can take advantage of this information. Benefits are expected in terms of enhanced latency and reliability of the communication link and an increased number of supported network devices.
The increased number of mobile production machines and AGVs transporting goods also needs to cope with large obstacles, which potentially require a periodic relocation in the factory. Since such relocations are not random but follow patterns well-defined by the production processes, ML algorithms can proactively use this knowledge to mitigate interference and link failures on the network side and optimize the paths of AGVs to avoid unfavorable radio conditions on the production side.
Finally, ML and especially deep learning algorithms can perform well only if a vast amount of data is available for training continuously. In this respect, mobile robots and vehicles can be the valuable data source. However, it would defeat the purpose of transmitting this data to the edge cloud (i.e., where it is processed), compromising the campus network’s radio resources dedicated to production applications. Distributed or federated ML can be exploited to balance the need for computing capabilities close to the data sources and the required bandwidth to transmit this data to a central location. It leads to an additional degree of complexity, e.g., due to the mobility of devices hence changing of the network topology, and at the same time interesting new optimization problems.
One driving factor behind the success of machine learning has been the ability of high-capacity learning methods to learn generalizable models from large amounts of data. The success of ML stands and falls on the availability and quality of datasets coming from the specific application domain. Future 5G systems follow a distributed multi-tier design concept in all layers to provide better load management, speed, and separation of computing and networking resources. Large amounts of critical operational and user data are generated at each tier, which constitutes the so-called data plane of 5G systems. Exploiting the full potential from this data requires appropriate strategies to be put in place and design choices to be made to support data collection, data storage, data transfer, data wrangling, ML model training, testing, and lastly, implementing those models in the production environment.
A current challenge that remains to be solved is the integration of analytics into the network, which is currently complicated due to the various non-standardized interfaces and inconsistent data collection techniques. Initial steps towards enabling such functionalities in 5G systems have been addressed by the network data analytics function (NWDAF). By enabling standard interfaces from the service-based architecture (SBA) to collect data on a subscription/request basis, analytics functionalities can be delivered across the network to enable automation and systematic reporting, thereby addressing the challenges related to custom interfaces.
Campus networks are, in general, deployed in environments characterized by highly correlated processes that are then reflected in the corresponding manufacturing-related data. Augmenting campus network data with such production data can greatly benefit ML methods by reducing training time, increasing data efficiency and accuracy, and introducing new insights useful to the learning process.
Production plans can provide typical manufacturing-related data sources, e.g., a list of ongoing tasks and their status, goods, and production stations from enterprise resource planning (ERP) systems, production planning tools, product lifecycle management software, and fleet management tools. Furthermore, location information of end devices, such as AGVs and mobile objects, can be obtained via dedicated localization systems, e.g., based on ultra-wideband (UWB) technology, lidar and/or video (SLAM), or directly from the 5G system.
Consolidating manufacturing data with campus network data into a joint ML pipeline poses several data acquisition and integration challenges. Problems arising from, e.g., different time synchronizations/granularities or heterogeneous data types/formats must be addressed to fully leverage potential synergies. Furthermore, access to real-world data may be difficult in some cases due to, e.g., regulatory policies or the need for costly intrusions into the production processes.
Data acquisition in such cases has to be planned and motivated by corresponding ML benefits. In this regard, network simulators/emulators and digital twin representations can serve as synthetic data sources to provide initial performance assessments and enable fast prototyping of ML solutions. Plausibility of synthesized data shall be ensured by calibrating the employed simulators/ emulators with accurate radio propagation models and traffic models, backed, when possible, with real-world data and expert knowledge for system parametrization.