‘KOSMoS’ project

Networked Machines for Intelligent Production and Predictive Maintenance

The networking of machines, intelligent computers, and other devices in industrial environments is creating an ever-increasing amount of production-relevant data which can be collected, processed, and made available. In addition to sensors and smart devices, standardised protocols such as OPC UA and MQTT also play an extremely important role in facilitating the exchange of data between the various platforms. While the data generated is often very high-frequency, high-dimensional, and as such constitutes big data, only a few samples of it exist. In order to use the data collected as a foundation for intelligent processes such as condition and defect monitoring or predictive maintenance – and thus to realise the vision of intelligent production – specialised methods and algorithms are required. In both data engineering and data science, inovex has contributed its expertise to the KOSMoS project.

Data-Driven Business Models in IIoT

KOSMoS provides industry, as a user of data analysis and machine learning, with the opportunity to make its production and processes more efficient, flexible, and faster – and thus more competitive. It also enables new offerings and data-based business models to be designed using the information obtained. This, in turn, creates a value-added network: transparent maintenance concepts, dynamic leasing, and proof of quality for finished products mean that customers also benefit from this data. The advantages are discernible, for example, in lower and more effective maintenance efforts and processes, more cost-effective and efficient usage-based leasing rates, and optimised production processes. inovex provided particular support for the KOSMoS project’s industry partners during the requirements analysis and translated the results into data analysis questions.

Blockchain: Building Mutual Trust

A cohesive value-added network in which cooperating companies exchange production and process data can generate added value for each partner involved – provided that a sufficient level of trust exists. Part of the central KOSMoS instance therefore consists of a blockchain-based solution that serves as the “single source of truth” in KOSMoS. A blockchain is a data storage method which is not managed by a single company, but which is decentralised and simultaneously transparent for all business partners, and which thus (from a technical point of view) represents a network of distributed, independent servers. The concatenation of the data sets, as well as the use of what is known as a consensus algorithm to ensure common agreement among all network nodes regarding the present state of the data, prevents subsequent reversion to previous states or manipulation of the data.

“Smart contracts” can be implemented within the blockchain. Activated by a data threshold or another action, these trigger a process which uses tamper-proof stored data to automatically perform additional processes within a value chain. This enables contracts to be modelled, processed, and checked, while simultaneously reducing transaction costs and increasing contract certainty.

The Analysis Platform: Machine Learning in Productive Use

In addition to data storage, the KOSMoS system also includes an analysis platform that can be executed both locally and in a cloud infrastructure. It is linked by a connector to the local KOSMoS system and is, among other its other capabilities, able to recognise possible patterns in the data generated which help companies to better understand their machines and derive measures accordingly. Maintenance cycles, for example, could be demand-based rather than time-based, enabling operators to use predictive maintenance principles to replace components at the optimal point.

Due to the multitude of different components, especially machine learning libraries and tools, even tech giants hold differing opinions about the best approach for pipelines in productive use. As part of the KOSMoS project, inovex has been researching architectural concepts in which modular data processing and machine learning components can be operated securely and reliably not only in the cloud, but also “on the edge” or “in the fog”, i.e. much closer to the “shop floor”.

Smaller-scale analyses and evaluations, for example to detect outliers or anomalies in data generated during operation which indicate equipment malfunctions, can be carried out directly “on the edge”. This means that the data does not necessarily leave the shop floor. Interactive, exploratory offline analyses and complex forecasts require transferring data captured over longer periods of time to the cloud for storage purposes and in order to generate the complex machine learning models required. In order to do this, analysts need a way to centrally access and work with the data.

Federated Learning: Collaborative Training Using Private Data

For cases where certain information must not leave the shop floor under any circumstances, i.e. it must not be transferred to the cloud, KOSMoS’s integrated federated learning functions provide a solution. In federated learning, several companies make their private data available for the collaborative training of a model. During this process, the data remains with the participants providing it, rather than being collected centrally in the cloud or on a common server, and only the local models or model parameters are exchanged. This guarantees that the data remains protected – while still allowing everyone to benefit from the collaboratively trained model.

KOSMoS already fulfils the prerequisites for using federated learning. The KOSMoS Edge, located on the operator’s shop floor, collects failure data and then participates in the federated training process. The analytics platform acts as a central administrator, admitting only participants with appropriate failure data to the training process. During the training process, only model parameters representing the patterns and regularities from the private failure data are shared, while the private data from the participants remains on the KOSMoS Edge. The central model is aggregated using the model parameters of the locally trained models and made available for download on the analysis platform. The federated component in KOSMoS was implemented using Flower, an agnostic framework for building federated learning systems.