An improved system and method for producing replacement sensor signals for failed sensors (150), and inferred sensor signals (160) for non-instrumented physical parameters, in processes and equipment having one or more sensors in place for monitoring physical parameters. A data history from a fully-instrumented prototype (110) provides a representative data set of anticipated operating parameters for forming an empirical model for generating 'virtual' signals for a process or machine in real-time. A memory (551) for storing the representative training set, or a transformation thereof, is coupled to a processor (545). The processor (545) receives from an input, data signals embodying real values from sensors on the process or machine. The processor (545) takes a set of readings of the actual sensors from the input, and generates an estimate of one or more desired inferred sensors, using a linear combination of the representative training set sensor data, as weighted by the result of a measure of similarity of the input data to the representative data.