Research On Inter-turn Short Fault Detection Technology Of Permanent Magnet Wind Generator Based On CNN-LSTM

The permanent magnet wind turbine is becoming more and more popular as the mainstay of offshore wind generation.The Permanent Magnet Wind Generator（PMWG）,being the principal component of wind turbines,operates in severe settings for a long period and is prone to numerous problems.The Inter-turn short fault（ISF）,one of the most common faults of the stator winding of a PMWG,can prevent serious accidents if detected as soon as possible.By using traditional fault detection methods such as signal analysis and machine learning,fault characteristics must be identified from initial data.The feature extraction procedure is based on engineers’ and technicians’ theoretical knowledge and practical expertise,which may affect fault detection accuracy.With the use of a convolutional neural network and a long-short memory neural network（CNN-LSTM）,an early fault detection model for ISF in a PMWG is constructed in this thesis.The specific work contents are as follows:（1）The faulting mechanism of the ISF of the PMWG in the wind turbine is analyzed,and a finite element simulation model of the ISF of the PMWG is constructed.To enhance simulation speed,the necessary parameters of a PMWG are extracted from the finite element model,and a lumped parameter simulation model of a wind turbine is developed.The ISF data set is obtained on this basis.（2）Using the simulated data,the fault characteristics of the short-circuit current,phase current,and q-axis current are analyzed in a time domain and frequency domain.The frequency of the fault harmonic component occurring in the stator q-axis current is more sensitive to the initial fault of inter-turn short,according to study and comparison,which is considered the fault characteristic of detecting an ISF.The stator q-axis current of a generator during an ISF is investigated using wavelet packet transform and fast Fourier transform to identify the early defect of an ISF under constant speed conditions.The findings demonstrate that this approach can efficiently identify the fault characteristic amount in stator q-axis current under harmonic and noise interference.The stator q-axis current during ISF is investigated using the short-time Fourier analysis approach for variable speed situations.The findings indicate that the fault characteristic quantity is a change quantity that is positively connected to speed,which is inconvenient for fault identification.（3）To address the drawbacks of the preceding analytical approaches,a defect detection model of an ISF in a PMWG is built using a one-dimensional convolutional neural network（1DCNN）.After segmentation and normalization,the recorded q-axis current of the generator stator may be utilized directly as the model’s input.The 1D-CNN model can extract data characteristics automatically.Because the stator q-axis current signal is a time series signal,the long and shortterm memory neural network（LSTM）can optimize the model’s learning effect for time series attributes,and a fault detection model based on CNN-LSTM is established.Both the 1D-CNN model and the CNN-LSTM model have excellent recognition accuracy,but the CNN-LSTM model has stronger anti-noise performance and shorter training durations,making it more ideal for learning small sample sets.