No-Load Transformers: Vibration Spectra Analysis by Deep Learning Methods for Loose Windings Detection

Vibration data analysis is an effective method for assessing internal fault in transformers. Recently, neural network classifiers have been successfully proposed as a tool for the detection of loss of clamping pressure of the winding pack in transformers under load operating condition. In this paper, we extend this approach to the case of unloaded transformer, where, unlike load operating conditions, vibrations are mainly driven by the core. This investigation has been carried on with vibration data experimentally collected with a lab equipment constituted by an oil insulated transformer, either under no fault (tight winding pack) or fault (loose windings) conditions. The analysis proves that the fault can be still reliably detected with a high accuracy, robustly with respect to a possible misplacement in the positioning of the sensor.