Sensor-informed risk-oriented robust identification of composite adaptive dynamic models (ADM) of a digital twin of a production asset

The subject of the research is a sensor-informed risk-oriented approach to robust online identification of composite adaptive dynamic models used as the computational core of a digital twin of a production asset. The object of the research is a production asset observed through a stream of multidimensional sensor measurements under conditions of changing operating modes, non-stationarity, noise, and outliers. The author thoroughly examines the task of transforming heterogeneous signals into a degradation-resistant state model that supports reproducible management decisions on the advisability of intervention. Special attention is given to aligning identification procedures with risk criteria for operation so that diagnostic uncertainty translates into a quantitative assessment of the pre-critical state and further into an economically justified binary control rule. The compositional structure of the model is also considered, where local subsystems are formed from groups of sensors and combined into an integral risk taking into account the criticality of the elements. The methodology is based on signal standardization, grouping sensors by statistical relationships, constructing a health index using principal component analysis, and robust recursive identification of its local dynamics in the form of AR(1) with adaptive forgetting and a limitation of the influence of anomalies; then, risk is assessed through forecasting and logistic transformation. The main conclusions of the conducted research are the confirmation of the effectiveness of the proposed scheme for robust identification and the demonstration that a risk-oriented interpretation of the degradation forecast allows for a stable intervention rule to be obtained in cases of data quality degradation. A significant contribution of the author to the research topic is the integration of several levels of robustness, where stability is achieved simultaneously through the adaptive forgetting factor, which limits the influence of non-stationarity, and through the influence function of Huber, which suppresses outliers. The novelty of the research lies in the compositional construction of the digital twin, where local probabilities of the pre-critical state of subsystems are aggregated into integral risk with criticality weights and then translated into a binary decision through an asymmetric cost model that accounts for the difference in the cost of intervention and the cost of missing a hazardous state. The practical significance is confirmed by testing on separate training and testing datasets from the open NASA C-MAPSS set, where the pre-critical state is defined through residual resource and the threshold proximity to failure, and quality is evaluated using metrics of probabilistic ranking, binary classification, and average costs based on the loss function.