On-line current-based condition monitoring of three-phase induction machines

This work addresses the problem of failure prediction in line-operated three-phase induction machines using current-based spectral analysis. The predictable characteristic frequencies generated in the stator current spectrum by air gap eccentricities and broken rotor bars are analyzed. This predictability is extended to include stator current frequencies associated with rolling-element bearing defects. Comparison of experimental current and vibration spectra confirm the presence of these characteristic frequencies.; Existing evaluation schemes detect either broken rotor bars or air gap eccentricities by monitoring the characteristic components in the spectrum of a single phase of the stator current. These schemes, however, do not take into account the effects of load torque oscillations on the current spectrum. A theoretical development, confirmed by simulation and experimental results, shows that these load torque effects cannot be separated from fault induced effects when monitoring only a single phase of the stator current.; The current spectral characteristics described in this work are utilized in a practical system that predicts impending or incipient faults in three-phase induction machines operating under any arbitrary load condition. This sensorless detection scheme uses a rule-based frequency filter and a clustering neural network to account for the load effects by learning the current spectra for all "good" operating conditions and then monitoring for changes that indicate the development of a machine fault. The efficacy of this method is verified with experimental results. While this system allows for the immediate evaluation of the machine health without requiring the presence of an expert, false alarms can occur if the system does not ascertain all normal load conditions.; Removing the load effects from the monitored spectrum would improve the capability of any scheme to detect the presence of a machine fault regardless of the machine load. The proposed method utilizes model reference estimation to remove the load effects from the measured stator current. Computer simulations illustrate the advantages of the modified stator current spectrum. This representation increases the ability to detect an anomaly by emphasizing the spectral changes caused by the fault condition. Experimental test results demonstrate the feasibility of this method.