Data-Based and Hybrid Methods in System Dynamics and Control for Fault Diagnosis

The increasing amount of sensors data in technical systems as well as larger computing capabilities provide many opportunities for data-based methods and machine learning algorithms in a wide range of system theory. In many cases, the interpretability and generalizability of data-based methods are limited compared to classical model-based methods. This fact motivates hybrid models, which combine model- and data-based methods to combine their benefits. This talk presents two exemplary applications of data-based and hybrid methods in system dynamics and control.

The first example deals with detection of anomalies and fault diagnosis in a paint shop by bipartite graphs. The structure of the bipartite graph is derived by mutual information using data from an active system excitation. The joint probability density function between the identified sets of stochastically depend variables is estimated with Gaussian Mixture Models.  The parametrized structure enables the evaluation of the likelihood for real-time data to detect and locate anomalies. Experimental results from an air supply unit illustrate the effectiveness of the approach.

The last example is the diagnosis of actuator and sensor faults in adaptive high-rise buildings using a hybrid approach. It combines the model-based method of parity equation with the data-based method of principal component analysis (PCA). PCA characterizes the unknown disturbance in the residual data derived by parity equations and determines orthogonal directions of decreasing variance in the residuals. These directions are decoupled to decrease the sensitivity of the residual to disturbances and maintain the diagnosability of faults.