Study On Ultrasonic Signal Characteristics And Pattern Recognition Method Of Typical Defect Partial Discharge

Electrical equipment plays an important role in the power system.Whether it is the power transmission and transformation equipment like transformers and power cables,or vehicle high-voltage cables as energy transmission links for high-speed trains,partial discharge phenomenon occurs during operation because insulation defects cannot be completely avoided in production and operation.Partial discharge(PD),as an important reason and early manifestation of equipment insulation fault,is often used to reflect the insulation state of equipment.Partial discharge detection methods are various.Compared with traditional local discharge detection methods such as pulse current method and HFCT detection method,ultrasonic detection method has the advantages of not affecting the operation mode of equipment and strong anti-electromagnetic interference ability.Although the attenuation characteristics of the partial discharge ultrasonic signal make it difficult to correspond the amplitude of the signal with the discharge quantity,it has abundant discharge type characteristic information and can be applied to the recognition of partial discharge type.Therefore,the ultrasonic signal characteristics of partial discharge in typical defect models are analyzed,and the methods of feature extraction and pattern recognition are studied.In this paper,the generation mechanism and acoustic attenuation of ultrasonic signal generated by discharging in the power equipment are analyzed.The paper summarizes the typical defects in the operation of electrical equipment,such as burr of electrode structure,layer of insulation material,air gap or scratch of insulation layer,and conductive particles attached to insulation surface,and the four typical defects are equivalent to tip discharge,surface discharge,air gap discharge and suspension discharge.Four typical partial discharge(pd)models were designed and fabricated under laboratory conditions.An experiment platform which can generate discharging ultrasonic signal without halo discharge is designed and built.Using the wavelet threshold denoising method to deal with the noise of the collected ultrasonic signals,and explore the influence of the mother wavelet,decomposition layers,threshold selection rules,reset threshold value method and threshold function for denoising effect,optimal wavelet threshold parameters is used to study the discharge ultrasonic signal preprocessing,and achieved good results.Different types of discharge ultrasonic signals are analysised at time and frequency domain,the results show that the time domain waveform shapes of the four types of discharge ultrasonic signals are quite different,and the energy of different types discharge ultrasonic signals distribute in frequency bands is different.The rising time,duration,energy and amplitude of ultrasonic signals were used to form a 2/3d correlation map,and ultrasonic signals of different types of discharge were distributed in different clusters in the feature correlation map.Based on the time domain waveform differences of four types of discharge ultrasonic signals,18 primary waveform characteristic parameters,such as rise time,duration,event peak factor and event waveform factor,were extracted.The primary characteristic parameters were de-redundant by principal component analysis method,and finally four new independent principal component indexes were obtained.The fuzzy entropy parameters at multiple scales(10 scales)are also extracted from the same ultrasonic signal sample.The main component index of waveform feature parameter and fuzzy entropy value in 10 scales are used as input feature vectors for classification and recognition.A support vector machine(SVM)algorithm suitable for small-sample training was designed and optimized,two kinds of input feature vectors were used for classification and recognition training and testing respectively.The training and testing samples are 100.The results showed that the recognition rate of multiscale fuzzy entropy feature vectors was higher than characteristic parameters,reaching 92%.In summary,by extracting the waveform characteristic parameters and multi-scale fuzzy entropy parameters of typical defect PD ultrasonic signals,and combining with the optimal classification algorithm of support vector machine,the PD types can be identified effectively.