ICACT20220141 Slide.17
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
thank you!
|
|
ICACT20220141 Slide.16
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
this is the reference
|
|
ICACT20220141 Slide.15
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
this is the reference
|
|
ICACT20220141 Slide.14
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
A partial discharge recognition method was introduced in this paper. Four kinds of PD signals were collected through a high-voltage partial discharge device and high-frequency current sensors. We use the wavelet transform to remove interference of the time-domain signal, and CNN was used to extract features and recognition. The accuracy of recognition is 91.8%. Compared with BP neural network[19] and OS-ELM algorithms, which the accuracy are 87.8% and 91.4% respectively, our methods get higher accuracy.
|
|
ICACT20220141 Slide.13
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Through the above processing and settings, the entire network can be trained to a suitable state on the training set, and feature extraction and classification recognition can be realized. There are four types of partial discharge to recognize, oil gap defect, floating defect, tip defect and surface defect. For each type, 200 samples are sampled to form the training set, and 50 samples are sampled to form the test set. Therefore, the total training set has 800 samples and the test set has 200 samples. Network training is carried out through the training set, and the results are verified on the test set.
|
|
ICACT20220141 Slide.12
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
In order to increase the generalization ability of the network and prevent over-fitting, on the basis of formula (9), an L2 regularization term is added
The stochastic gradient descent method is adopted as training method. Each iteration of the stochastic gradient descent method uses a set of samples, and then updates the parameters according to the gradient. The batch size of each training set here is 128 samples.
|
|
ICACT20220141 Slide.11
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
After the construction of the entire network, a cost function needs to be defined to measure the difference between the predicted classification and the true classification. Cross entropy is used as a method to measure this difference
|
|
ICACT20220141 Slide.10
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
The soft-max layer uses a multi-class logistic regression method, which can convert the input to a probability of 0-1, that is, to achieve the conversion of features and probabilities. The calculation formula is expressed as:
|
|
ICACT20220141 Slide.09
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Through these convolutional layers and related layers, the features in the original data can be extracted. The two fully connected layers are then connected to achieve a linear transformation from one feature space to another feature space.
|
|
ICACT20220141 Slide.08
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Then the next connect layers is the second set of convolutional layer, pooling layer and local normalization layer
|
|
ICACT20220141 Slide.07
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
The second layer is pooling layer, using the maximum pooling method. The pooling window size is 1×3, and the step size is 2
The third layer is the local response normalization layer (local response normalization), using the concept of lateral inhibition in neurobiology.
|
|
ICACT20220141 Slide.06
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
After the signal pre-processing, a deep learning network is built for PD recognition. Using supervised learning methods, using convolutional neural networks to extract signal features.
The first layer is convolutional layer, and the convolutional kernel is a 1×3 array, step length is 1, channels is 64. The input is pre-processed and enhanced signal
|
|
ICACT20220141 Slide.05
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
and the structures of four types of PD are shown in left Figure . The signal collected by the oscilloscope is shown in right Figure. It can be seen from the figure that the noise has a great influence. In order to reduce the noise influence, the wavelet transform is used. Wavelet transform is an effective transform analysis method, which can refine the high frequency and low frequency of the signal through the expansion and translation operation. The Daubechies5 wavelet with orthogonality and tight support is used here. As a sparse base, this wavelet is not easy to produce large smoothing errors, that is, the reconstructed signal is smoother.
|
|
ICACT20220141 Slide.04
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Transformer PD signal is collected using the system shown in the Figure . The PD devices are deployed inside the transformer, The high-frequency sensors are employed at the transformers ground wire to couple the high-frequency pulse current signal. The real-time waveform and synchronous voltage signal are collected by the oscilloscope. The high-frequency partial discharge sensor adopts TWHCT-7850 with the 3MHz-30MHz frequency band, and the oscilloscope is R&S RTO-2044.
|
|
ICACT20220141 Slide.03
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
The types of transformer partial discharge monitoring mainly include: ultrasonic, ultra-high frequency (UHF), Electrochemical, and high frequency (HF), and so on.
Researchers have studied in the field of partial discharge recognition for many years, and have proposed many methods [6], including Support vector machine (SVM), Artificial neural network (ANN), Random forest (RF), Deep learning, and so on.
|
|
ICACT20220141 Slide.02
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Power transformer is the most basic and core equipment of the power system, the transformers insulation degradation will rapidly develop into a fault, cause a large-scale blackout and a serious damage of the production and life.Therefore, it is very important to monitor the insulation status of the transformer, and partial discharge is an important reference for the insulation performance of the transformer. The recognition of the partial discharge types helps to assess the form and degree of transformer insulation deterioration, find hidden dangers in time, and avoid transformer failures.
|
|
ICACT20220141 Slide.01
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
ladies and gentlemen, I am very glad to share our work in few minutes. I am Meng Liu ,an engineer in state grid Shandong electric power research institute . The main work of this paper focus on the Recognition of Transformer High Frequency Partial Discharge Based on Time Domain Feature
|
