Analysis and Diagnosis

Intelligent analysis and diagnosis is an essential part in the fields of cyber-physical production systems and Industrie 4.0: Intelligent assistant systems are used to ease the development and maintenance of automation systems for the human expert.


This competence area focuses on the intelligent analysis of technical processes in production. The scientific focus is aimed at the application of methods of artificial intelligence for analysis and diagnosis of production systems and other related topics. Based on a sequence of the necessary steps standard procedures for analysis and diagnosis are established:

Data Acquisition: At first, all relevant data of a production line are acquired. In this connection research focuses on challenges like time synchronization, epistemic uncertainty and handling of heterogeneous systems.

Based on the recorded data, information fusion then acquires a consistent image of the current status of the system. It is observed that it is only possible to achieve a consistent image of the production lines and its symptoms by multisensory data analysis. Important application areas which are researched at inIT are on the one hand related to the research of evidence-theoretical concepts for a sensor fusion and are on the other hand examined regarding their plausibility of information by means of new accesses in the field of the degree of belief theory (cf. Fig. 1).

Context based information fusion process
Context based information fusion process

In the next step, non-normal situations, i.e. anomalies, must be detected. In this connection, we can distinguish between two approaches: on the one hand phenomenological methods, which directly extrapolate from measurements to anomalies (Fig. 2); on the other hand model-based approaches which compare the observed system behavior with model prognoses (Fig. 3).

Manual modeling of the knowledge which is indispensable for both approaches nowadays is hardly possible: production lines are too complex, people are very busy and many contexts are not even known to experts. One way out is machine learning of models based on system observations. At present, methods of model learning related to time automata, hybrid models and Ensemble-classifiers are in the focus of attention.


A major working topic in the field of model-based approaches is related to the machine and process monitoring (condition monitoring) as well as to the analysis and diagnosis of attack scenarios on Automated Teller Machines. By data comparison of a current system model status with the defined system model based on information fusion, it is possible to detect error symptoms even in complex systems reliably.
Moreover, another prominent working topic of model-based anomaly detection is the recognition of suboptimal time behavior and suboptimal energy consumption in production lines: assistance systems support humans to analyze complex systems and thus to take corrective measures at an early stage.
As humans are in the centre of modern automation systems a key issue must be handled by intelligent analysis and diagnosis tools: Semantics. In various projects semantical communication between intelligent technical systems and humans are researched. The main focus is on the knowledge formalisms for the description of industrial automation systems and text analysis (cf. Fig. 4).

Phenomenological analysis
Phenomenological analysis
Model-based analysis
Model-based analysis
Semantical patent analysis
Semantical patent analysis

Intelligent Optimization

So far, models have been used to identify anomalies. But models can also be used for system optimization. They allow typical Industry 4.0 scenarios – such as automatic energy and throughput optimization – to be implemented.


Figure 5 shows the basic principle: The optimization starts with a (learned) model of the system, including optimization goals such as energy consumption or throughput. From the available domain knowledge a better configuration, i.e. automation algorithm, is determined. Knowledge about the causal relationships between parameters and optimization goals is used for this purpose, for example in the form of equations. This is now assessed with a what-if analysis, i.e. a modified model is generated and the new configuration is analyzed with respect to the optimization goal. This process is repeated until a good new configuration is found, which is then used to reconfigure the automation system. This approach has been used in the Lemgo Smart Factory to optimize a high-bay warehouse. Figure 6 shows the result: After optimization, objects could be stored more rapidly and energy-efficiently in a high-bay warehouse.

Professor

Prof. Dr. Volker Lohweg
E-Mail: volker.lohweg(at)hs-owl.de
Phone: +49 (0) 5261 - 702 2408
Fax: +49 (0) 5261 - 702 2409

Prof. Dr. Oliver Niggemann
E-Mail: oliver.niggemann(at)hs-owl.de
Phone: +49 (0) 5261 - 702 2403
Fax: +49 (0) 5261 - 702 2409