Learning from the Past and Applying It to the Future When monitoring refrigeration systems, a vast amount of sensor data is generated from measurement points, which is often collected and stored as part of a higher-level monitoring system. To derive more than just information about limit value exceedances from this data, complex software architectures are usually required. A software package developed by ILK Dresden is now set to make it easier to detect impending system failures. In the operation of refrigeration systems, beyond maintaining system availability and efficiency, the optimization of system operation to save operating costs and reduce CO2 emissions, as well as the timely detection of even subtle changes in operating parameters and system components, also play a role. When monitoring refrigeration systems, a vast amount of sensor data is generated from measurement points, which is typically collected and stored as part of a higher-level monitoring system. To derive more than just limit value exceedances from the data, complex software architectures or specialized mathematical operations are usually required. Depending on the monitoring task, different sensor data is usually required to provide the necessary evidence.

As part of the EnBeKa II[1] collaborative project, these topics were addressed and simple methods were developed to facilitate traceability. Within the scope of this project, ILK Dresden investigated methods for fault detection and diagnosis (FED) of plant failures based on a state-based analysis of plant data sets. By using a self-organizing map, a plant’s data sets can be separated based on the characteristics of their process variables and structured into state clusters. These state clusters can be overlaid with events occurring in the plant, enabling an analysis of the causes of these events. The plant operator can evaluate the causes and, if necessary, adjust the operating parameters of the plant. Once events have occurred, they are stored by the software and can be compared with future operating data. Should data sets with similar characteristics occur during ongoing operation, the software is capable of recognizing this and issuing a warning if necessary. [1] 03ET1449B Energy Efficiency and Optimized Operation of Commercial Refrigeration Systems, Phase II, funded by the Federal Ministry for Economic Affairs and Energy.

Data records from a plant are evaluated using a trained neural network and assigned a status. If a data record contains a plant event (fault or abnormal operating characteristic), it is stored in an event library. Depending on the software configuration, additional historical data is also linked to the plant event. This makes it possible to interpret several consecutive data records as symptoms of a future fault event and to use them for early fault detection (EFD). To prevent the generation of warnings of low relevance when linking plant events from the event library (Figure 1) with current plant data records, freely configurable criteria for suppressing such “weak information” have been implemented.

Individual symptoms can be monitored hierarchically, and warnings and alarms can be configured with appropriate priority levels.

The software’s functionality makes it possible to set up monitoring for multiple plants (Figure 2) in addition to monitoring a single plant. This requires identical process variables, similar applications, and the same structure for the plants’ data sets. Data sets from different plants can then be analyzed within a single project. Combining these individual plants creates a virtual integrated plant.

This functionality makes it possible to set up fault monitoring based on System A and gradually expand it to include new systems. This allows the system events from System A to be used to assess the condition of System B. The “knowledge” learned from System A is thus transferred to other systems. Conversely, events from System B or C can also be used to assess System A. Since the self-organizing map is based on a neural network, training processes are required at least once. If the application areas or operating parameters (e.g., setpoints) of the system(s) change significantly over time, it is recommended to retrain the database to ensure the optimal map resolution is always utilized. The stored events or faults are retained despite the training.

The microtools developed in the project are based on various calculation functions that are grouped together in a library consisting of several .dll files. The ILK Library (.dll) developed by ILK Dresden contains all the functions required for creating a new project, preparing and evaluating data, and feeding it into a database. Furthermore, the database can be trained, and an automatic evaluation can be performed. As part of online monitoring, a data set is evaluated via a function call (by a higher-level software system), and a result is returned that contains information about any anomalies in the data set. For this purpose, the .dll can be integrated into the corresponding development environment. Additionally, there is the option of script-based execution of the individual functions. For this purpose, small programs with specific parameters can be started cyclically to enable data import, the evaluation of data sets, or recurring training. This approach was successfully implemented as part of the project. The required modules are downloaded as a compressed folder (.zip file) and can be stored locally on a computer or server without the need for installation, bypassing the need for an external provider. Comprehensive documentation on how to implement the microtools, as well as a download option for interested users, has also been provided.

Link zum Abschlussbericht EnBeKa – Phase II