Active sensor configuration validation for refrigeration systems

Abstract

SummaryThis master's thesis presents an approach to validation of the sensor configuration in control systems. Major faults in the commissioning phase of refrigeration systems are caused by defects related to sensors. With a number of similar sensors that differ only by spatial location and their use in the control system, fault-finding is not trivial and it often results in expensive delays. Validation and handling of faults in the sensor configuration are therefore essential to cut costs during commissioning. With passive faultdiagnosis methods falling short on this problem, this thesis describes an active diagnosis procedure to isolate sensor faults at the commissioning stage, before normal operation has started. Using statistical methods, residuals are evaluated versus multiple hypothesis models in a minimization process to uniquely identify the sensor configuration. The method as such is generic and is shown in the sequel to work convincingly on refrigeration systems with both nonlinear behaviors and significant mutual differences.The thesis includes sections concerning the target system with respect to building a dynamical model and analysis (both general and structural) of system properties and of the types of faults to be dealt with. Furthermore a complete solution strategy is presented in which a test signal is applied to one of the system inputs and to different system models resulting in residual signals for each of the faulty cases to be diagnosed. It is discussed how the dependency on accurate system models can be minimized by appropriate filtering of the residuals and a statistical detection method is applied, identifying the model best describing the system. Furthermore it is demonstrated how only very limited information of the system need to be contained in the models for the solution to be reliable. Thorough explanation of the utilized theory is given along with references to previous works in the area.Using both simulations and real data from commercial refrigeration systems validation of the method is offered. In the simulations it is illustrated how the proposed solution works as intended in the ideal cases with completely known system models and also how it is able to handle rather severe noise and model uncertainties when up to 25% variations are added randomly to both parameters and input of the model. Data collected in the refrigeration lab at Danfoss A/S reveals that parameter variations from one refrigeration system to another, depending on construction and dimensions, might exceed the amount of uncertainty that was successfully tested in the simulation environment. Hence, it is concluded that the solution is generic in the sense that it can be tuned to a certain class of systems for which it is very robust. With respect to measurement noise or low frequent disturbances e.g. the day/night or season changes the method is also proved to be very robust.In addition the work has resulted in a conference paper of the same title which has been submitted to the American Control Conference 2010...

Awareness and its use in Incremental Data Driven Modelling for Plug and Play Process Control

“…As a consequence, it is often a non-trivial problem to select the optimal sets of measurements for a given purpose [1], and it may sometimes happen that a different combination of measurements turn out to be better suited for control purposes than the originally chosen one [5].…”