Industrial production plants, especially plants in special machine construction, are characterized by increasing complexity in terms of configuration and operation. As a result, more and more intelligent assistance systems are emerging to help plant operators detect errors and anomalies in industrial processes at an early stage, optimize process flows and improve resource efficiency. This is of considerable financial importance to operators: for example, the failure of an injection molding machine costs between €10,000 and €40,000 per hour on average. If a defective machine is part of a production chain, e.g. in automotive manufacturing, the costs of a breakdown can even quickly run into the millions. These costs can be significantly reduced by using predictive maintenance and condition monitoring systems (CMS). The challenges in the use of CMS are that available systems are far from being able to assess the plant condition in every operating situation with sufficient accuracy. In addition, in industrial practice, there are major hurdles in the implementation and commissioning of a CMS, as complex programming or configuration is usually required to adapt the CMS to the respective plant. This problem is particularly pronounced in special plant engineering, where plants are developed for different branches of industry and each plant and thus also each CMS must be implemented and adapted individually for the customer.
In the CLArA project, a self-learning CMS is being developed for more complex production plants and, in particular, also for plants in special mechanical engineering. Such a system shall automatically adapt to existing plants by using machine learning methods and detect anomalies and error causes based on a learned representation of the system behavior. This is to be achieved by learning analysis and diagnostic models from historical process data in a training phase and using them to detect anomalies and diagnose fault causes while the CMS is in operation. In this way, the user should be able to use the CMS in different plants without extensive programming and configuration. This should significantly reduce the development and commissioning costs for CMS and, in particular, enable economical use in customer-specific plants that are only produced in small quantities. Scientific questions arise with regard to the scaling of existing self-learning anomaly detection methods to complex production plants as well as with regard to the development of self-learning diagnostic methods for the identification of error causes.