Data analytics plays one of the key roles in building intelligent systems, which bring automation to the new level of safety, reliability and efficiency, at the same time lowering the perceived complexity for the user. In this paper, we present the goals of data analytics in manufacturing and illustrate several application scenarios we have successfully worked on at Reifenhäuser REICOFIL GmbH & Co. KG. These include process monitoring and anomaly detection using virtual sensors, root cause analysis, plant simulation and optimization, assessing trade-offs between product quality criteria and extracting knowledge from data. Furthermore, we list a number of challenges that data analytics typically faces in manufacturing environments, demonstrating them on several concrete examples.

A tailored model of a system is the prerequisite for various analysis tasks, such as anomaly detection, fault identification, or quality assurance. This paper deals with the algorithmic learning of a system’s behavior model given a sample of observations. In particular, we consider real-world production plants where the learned model must capture timing behavior, dependencies between system variables, as well as mode switches—in short: hybrid system’s characteristics. Usually, such model formation tasks are solved by human engineers, entailing the well-known bunch of problems including knowledge acquisition, development cost, or lack of experience. Our contributions to the outlined field are as follows. (1) We present a taxonomy of learning problems related to model formation tasks. As a result, an important open learning problem for the domain of production system is identified: The learning of hybrid timed automata. (2) For this class of models, the learning algorithm HyBUTLA is presented. This algorithm is the first of its kind to solve the underlying model formation problem at scalable precision. (3) We present two case studies that illustrate the usability of this approach in realistic settings. (4) We give a proof for the learning and runtime properties of HyBUTLA.

This paper presents a novel model-based approach for the prediction of energy consumption in production plants in order to detect anomalies. A special Ethernet-based data acquisition approach is implemented that features real-time sampling of process and energy data. Hybrid timed automaton models of the supervised production plant are generated and executed in parallel to the system by using data samples as model input. According to comparisons of predicted energy consumption with the production plant observations, anomalies can be detected automatically. An evaluation within a small factory shows that anomalies of 10 % differences in energy consumption, wrong control sequences and wrong timings can be detected with a minimum accuracy of 98 %. With this approach, downtimes of production systems can be shortened and atypical energy consumptions can be detected and adjusted to optimal operation.

Today's hybrid technical systems, such as production facilities, evolve rapidly. Therefore the efforts and resources needed to obtain their trustworthy behavior models are significantly increased. This forms the modeling bottleneck, as the model creation mostly relies on the expensive and durable manual modeling. Behavior models are of the greatest importance in monitoring, anomaly detection, and diagnosis applications. To deal with this issue, the novel HyBUTLA algorithm was recently proposed for automated learning of behavior models from process data. Despite being the first hybrid automaton learning algorithm, it does not model the autonomous jumps (abrupt changes in process variables), and it suffers from long runtime due to the use of advanced machine learning methods. In this paper the HyBUTLA algorithm is improved to account for the autonomous jumps by introducing the splitting step based on discrete wavelet transform. This significantly improves the algorithm runtime, since high function approximation accuracy can be reached by using rather simple methods such as multiple linear regression. The benefits that splitting brings are formally proved and demonstrated in a real-world production system.

The increasing complexity of today's production systems and the variety of model-based approaches to their monitoring, diagnosis and testing emphasize the importance of the modeling step. Modeling is mostly done manually, in a costly and time-consuming way. In this paper, an alternative that comes from the learning theory is given: an automated procedure for identifying behavior models from recorded observations. Assuming the system's structure is known, the algorithm presented here is capable of learning behavior models for its components. The algorithm accounts for probabilistic, timing, discrete and continuous aspects of the given system, using the modeling formalism of hybrid automata. The practical usability of identified models is demonstrated using an anomaly detection application for a real production system.