Self-learning Condition Monitoring

Early on, a predictive CMS approach furnishes meaningful information about the condition of wear on components like motors, pumps and bearings. It facilitates a reliable determination of maintenance and repair needs. The downside of such a CMS is the enormous amount of data it generates, consisting of sensor values and sometimes specific process data. This can complicate using such systems and make them rather difficult to control.

Our Solution

The key feature of the Fraunhofer approach is its significantly simplified operation, linking automatic data analysis with a self-learning classification. To do this, the Condition Monitoring System must establish settings, such as the limit values, automatically. This is achieved by means of mathematical algorithms, which evaluate known operating conditions of a plant. The CMS »learns« the data characteristics of various conditions. Any changes in this data fingerprint are automatically detected. To eliminate false alarms, outliers are filtered out. Once enough measurements are available, limit values can be set automatically. After completing the initial learning phase only minimum additional effort is required for operating the CMS.

Example of Various Clusters in a 3D Feature Space

Example of Various Clusters in a 3D Feature Space

The CMS executes regularly scheduled measurements of values which are relevant to wear monitoring. This results in a data pool, which is evaluated together with the respective process data. The ultimately relevant differentiators are obtained automatically by targeted selection from a wide range of statistical parameters. Each record represents a point in a feature space, which can then be captured visually. Multiple measurements of operating conditions form clusters. They initially represent different operating conditions. The first appearance of machine wear will change the form and location of the clusters. As a consequence, this can then be detected in the fingerprint and the diagnosed change is reported.

This is how the CMS expands its database step-by-step with measurements of various wear and operating conditions and monitors all previous classification features continuously. When they no longer represent the current situation, the differentiators of the data fingerprints will be recalculated and an adjusted set of specific features is established.