On the identifiability, parameter identification and fault diagnosis of induction machines

Abstract

Due to their reliability and low cost, induction machines have been widely utilized in a large variety of industrial applications. Although these machines are rugged and reliable, they are subjected to various stresses that might result in some unavoidable parameter changes and modes of failures. A common practice in induction machine parameter identification and fault diagnosis techniques is to employ a machine model and use the external measurements of voltage, current, speed, and/or torque in model solution. With this approach, it might be possible to get an infinite number of mathematical solutions representing the machine parameters, depending on the employed machine model. It is therefore crucial to investigate such possibility of obtaining incorrect parameter sets, i.e. to test the identifiability of the model before being used for parameter identification and fault diagnosis purposes. This project focuses on the identifiability of induction machine models and their use in parameter identification and fault diagnosis. Two commonly used steady-states induction machine models namely T-model and inverse Γ- model have been considered in this thesis. The classical transfer function and bond graph identifiability analysis approaches, which have been previously employed for the T-model, are applied in this thesis to investigate the identifiability of the inverse Γ-model. A novel algorithm, the Alternating Conditional Expectation, is employed here for the first time to study the identifiability of both the T- and inverse Γ-models of the induction machine. The results obtained from the proposed algorithm show that the parameters of the commonly utilised Tmodel are non-identifiable while those of the inverse Γ-model are uniquely identifiable when using external measurements. The identifiability analysis results are experimentally verified by the particle swarm optimization and Levenberg-Marquardt model-based parameter identification approaches developed in this thesis. To overcome the non-identifiability problem of the T-model, a new technique for induction machine parameter estimation from external measurements based on a combination of the induction machine’s T- and inverse Γ-models is proposed. Results for both supply-fed and inverter-fed operations show the success of the technique in identifying the parameters of the machine using only readily available measurements of steady-state machine current, voltage and speed, without the need for extra hardware. ii A diagnosis scheme to detect stator winding faults in induction machines is also proposed in this thesis. The scheme uses time domain features derived from 3-phase stator currents in conjunction with particle swarm optimization algorithm to check characteristic parameters of the machine and detect the fault accordingly. The validity and effectiveness of the proposed technique has been evaluated for different common faults including interturn short-circuit, stator winding asymmetry (increased resistance in one or more stator phases) and combined faults, i.e. a mixture of stator winding asymmetry and interturn short-circuit. Results show the accuracy of the proposed technique and it is ability to detect the presence of the fault and provide information about its type and location. Extensive simulations using Matlab/SIMULINK and experimental tests have been carried out to verify the identifiability analysis and show the effectiveness of the proposed parameter identification and fault diagnoses schemes. The constructed test rig includes a 1.1 kW threephase test induction machine coupled to a dynamometer loading unit and driven by a variable frequency inverter that allows operation at different speeds. All the experiment analyses provided in the thesis are based on terminal voltages, stator currents and rotor speed that are usually measured and used in machine control.