|
In a spinning rotor gauge (SRG) pressure measurements are made indirectly by measuring decay of rotational frequency due to gas friction of a freely spinning magnetically levitated steel rotor. In addition to gas friction there exists a small pressure independent contribution due to residual magnetic friction in suspension system, called also residual drag. For precise pressure measurements residual drag is subtracted as offset correction. Additional pressure independent changes of rotor rotational frequency can come from thermal expansion effects by changing rotor's moment of inertia. It can not be resolved from residual magnetic friction and in practical work appears as temperature induced instabilities of offset correction. Main problem in accurate prediction of temperature influences comes from the fact that direct measurement of the rotor’s temperature is not possible. Our approach to partially solve the problem was to measure the ambient temperature in the vicinity of the SRG suspension head and use these data to predict the temperature-induced variations of offset correction. Typical daily variations of laboratory temperature and the corresponding SRG zero variations were modelled by neural networks. A new architecture of three separate neural networks performing completely different tasks has been introduced. Normally one neural network would be used to form the model, but the complexity of the problem dictated the decomposition of the problem to the three separate sub-problems. Each neural network performs a separate task. The first neural network models the temperature–time profile, which is needed to adequately smooth the measured data in order to calculate the time derivative of the temperature. The second neural network produces the pressure–time characteristic for the observed time interval. Special emphasis is given to the smoothening process produced by the two neural networks since the functioning of both networks is crucial for the proper operation of the third neural network. The third neural network finally combines the time derivative of the temperature and the SRG pressure readout in order to produce the correction for the SRG zero indication. |