How to improve the recognition speed and recognition accuracy in the voiceprint recognition of SF6 high-voltage circuit breaker fault judgment
High-voltage circuit breaker as an important power transmission and transformation equipment, its operating mechanism can operate correctly, directly related to the safe operation of the system, according to statistics, high-voltage circuit breaker operating mechanism mechanical failure accounts for 70% ~ 80% of all failures, the development of high-voltage circuit breaker operating mechanism mechanical state assessment and fault diagnosis technology is of great significance.
At present, the research on the mechanical fault diagnosis of circuit breakers is usually based on vibration or voiceprint signals, with artificial neural networks, dynamic time regularization and support vector machines and other algorithms for pattern recognition, vibration signal detection: based on the characteristics of the occurrence time of vibration signals when circuit breakers fail, the time parameters of vibration signals are taken as fault characteristics, and FCM is used to identify the fault type of circuit breakers.
This time, the application of single-class support vector machine in circuit breaker diagnosis was studied, and the vibration signal of circuit breaker under normal state was used as a training sample, and the Gaussian kernel support vector machine was used for training to determine whether the circuit breaker failed.
Usually in real-world operation, MCC-CNN fault acoustic signal recognition methods. The overall process of circuit breaker acoustic signal recognition is shown in Figure 1, which first converts the circuit breaker action acoustic signal collected by the microphone into a time spectrogram with two dimensions of time domain frequency domain.
Then, the original time spectrum is extracted and dimensionally reduced by calculating three cepstral spectra, including MFCCs, GFCCs and PNCCs, and finally the convolutional neural network is used as a classifier for fault type identification.
The whole process can be roughly divided into three parts: acoustic signal acquisition, preprocessing and pattern recognition, of which the preprocessing method of acoustic signal is the most important, its main function is to perform feature extraction and data compression on the original time domain signal of the circuit breaker, so as to reduce the amount of calculation of the subsequent recognition model and improve the recognition effect.
The time domain signal of the interrupter acoustic signal is a one-dimensional pulse signal, and its characteristic information is not obvious enough, and it can be converted into a two-dimensional time spectrum by using short-time discrete Fourier transform, which is conducive to improving the recognition speed and recognition accuracy of the deep learning model.
The general characteristics of the latent fault of the circuit breaker are not obvious, so when the circuit breaker voiceprint diagnosis is carried out, the voiceprint features should be extracted under the premise of ensuring the recognition speed of the acoustic signal, so as to improve the recognition accuracy, and the cepstral coefficient calculation method widely used in the field of speech recognition can compress the data of the sample while retaining the key voiceprint information of the closure.
Therefore, the compression and feature extraction of the acoustic signal of the circuit breaker are realized, which is helpful to improve the diagnosis speed and diagnostic accuracy of classifiers such as convolutional neural networks connected later.
In the calculation of convolutional neural network, since the cepstral feature matrix is a three-dimensional map superimposed on multiple two-dimensional maps of the same size, a representative convolutional neural network (CNN) in the field of image recognition can be introduced as a classifier for the cepstral feature matrix of acoustic signals.
In the field of image recognition, CNNs usually split color images into three color layers of red, green and blue (RGB) as the input layer of the network, so as to learn and perceive the characteristics of different color changes.
A MixedCepstralCoefficient-Convolutional Contemporary Network (MCC-CNN) recognition model was constructed for sound classification recognition.
Compared with the manually designed cepstral mixing method, the fusion of the three cepstral spectra through the learning mechanism of the deep neural network can make the fusion method of the hybrid cepstrum adaptive, and the circuit breaker closing voiceprint data has been dimensionally reduced and compressed.
Therefore, classification can be achieved by a lightweight CNN network like VGG [17], as shown in Figure 3.
Among them, the input layer of the network structure parameters is set according to the size of the input sample, and the convolution kernel, pool kernel size and corresponding moving step size are selected from the widely used parameters, and finally the overall network structure containing 3 convolution-pooling layers and 4 fully connected layers is determined after structural parameter debugging.
In the network, batch normalization and Dropout operation with probability of 0.5 are added to prevent overfitting and gradient disappearance, and the detailed structural parameters are shown in Table 1.
In mechanical fault simulation and data acquisition, the main components of high-voltage circuit breaker include three parts: energy storage unit, transmission unit and control unit, LW30-252 SF6 high-voltage circuit breaker and its internal CT26 spring operating mechanism are shown in Figure 4.
The schematic diagram of the structure of CT26 spring operating mechanism is shown in Figure 5, which has compact structure and high integration, but it also faces common problems such as mechanical failure during use.
Therefore, based on the structural principle of CT26 spring operating mechanism, artificial fault setting and simulation experiments were carried out around five typical latent mechanical faults of high-voltage circuit breakers: oil leakage of oil buffer, fatigue of closing spring, wear of transmission shaft pins, jamming of spindles, and loose anchor bolts, and data sets were constructed for subsequent voiceprint classification examples.
The fault setting method adopted in the institute can simulate different latent mechanical fault states of circuit breakers, and the faint difference of acoustic signals can be reflected in the spectrogram of the acoustic signal of the closing action, which can reflect the early characteristics of the acoustic signal when the latent fault occurs, and can provide effective support for the subsequent MCC-CNN recognition model.
By extracting all kinds of cepstral information of acoustic signals and combining them with convolutional neural networks, the latent mechanical fault state can be identified through the acoustic signal of circuit breaker closing action, compared with the method of directly entering the time spectrum into the convolutional neural network, the recognition speed and recognition accuracy of the method studied here are improved.
In the diagnosis of latent mechanical fault voiceprint, the hybrid cepstral method has stronger advantages in recognition accuracy than the single cepstral extraction method, but the calculation speed will be slightly reduced.
