To progress from sensor measurements to equipment failure prediction, an adaptive technique combining physics-based digital twin executable along side the sensor telemetry was explored in this paper. Sensor telemetry in real conditions of operations is always prone to signal errors which can be misinterpreted as outliers or anomalies. First, a layered approach recognizes that no single combination of sensor data cleaning will be optimal in all circumstances. A first layer of data cleaning was devised as stored rules to detect basic point, contextual and collective anomalies. With more common anomalies identified and set aside, additional layers of more advanced outlier detection using isolation forest, autoencoder or GAN based methods may be used. In the presence of more common anomalies, techniques to identify rare anomalies are likely to fail. One way around this common problem is explored in our paper to combine sensors and a physics-based digital twin of the equipment in operations to further divide the problem. Without more advanced machine learning methods to find remaining anomalies, data will not be reliable enough for AI predictive models and digital twins may not have high fidelity on their own to cover all circumstances, but the combination of both is where we found promising results. Next, flexible AI training recognizes that the best model or even the set of inputs cannot be predetermined. It must adjust according to the quantity, quality and diversity of data available after cleaning. Available model types will depend on data quantity and quality. For instance, training and testing of prospective models may progress from linear and polynomial fitting to random forest regression, but only proceed to deep learning with credible quantities of data. Finally, the system will be adaptive, anticipating changes in available sensors, data cleaning and model choice over time. The layered approach is not a fixed data pipeline but an adjustable scaffolding from sensors and digital twins towards failure predictions. The industrial AI marketplace is transitioning from an academic, prototyping phase to a more healthy focus on successful real-world implementations. Our paper is a step towards applicability of latest techniques combining the best of data-driven approaches infused with physics-based understanding of what is truly happening before an industrial equipment fails.