Predictive diagnostics became a mature independent scientific trend during last several decades. The main idea of this methodology is to create a degradation model of a technical asset that can be utilized for prognosing its abnormal behavior and failures in advance. Historically this scientific sphere inherited two main approaches in creating the model itself. First approach grounds on traditional understanding of a mathematical model as an analytical summarization of fundamental laws of nature. From technical perspective this can be substituted by expert knowledge of experienced engineer which can be formalized as a formula dependent on certain measured parameters. The other direction, on the contrary, bases on ‘pure’ mathematical approximation of observed equipment behavior. Both approaches have their pros and cons in certain circumstances that are in details described in this paper. In our research we show one simple method of bringing advantages of both approaches into one common diagnostic algorithm. As a reference example we use a real pump which online parameters are stored in real-time database of a SCADA system. We use historical trends for training the adaptation rules and test how the resulting algorithm is applied to failure detection problem.