特高压直流换流阀饱和电抗器振动声纹特性与松动程度声纹检测方法

Vibration Voiceprint Characteristics and Looseness Detection Method of UHVDC Converter Valve Saturable Reactor

饱和电抗器作为特高压直流换流阀的核心装备,运行中产生的振动声音包含大量的信息,其状态评估对换流阀的安全运行具有重要意义。该文提出一种基于优化S变换和改进深度残差收缩网络的饱和电抗器铁心松动程度声纹识别模型。首先开展了高频脉冲激励下的饱和电抗器振动试验,并测量了不同铁心松动程度下的声纹信号;其次在声信号频谱主值区间内,根据能量聚集性优化高斯窗参数来提高声纹图谱的时频分辨率;然后对松动后的声纹特性进行分析,发现高低频比和低频分量主频占比两个特征指标仅能对松动程度较高的状态做出预警;最后采用五个不同方位测点的铁心松动数据代入基于自适应参数修正线性单元的改进深度残差收缩网络中进行训练,来消除声纹图中的冗余信息,并对不同松动程度下的特征进行独立映射,从而增强共同特征的学习能力。研究结果表明,该文模型对电抗器不同铁心松动程度的平均识别准确率达到95.93%,优于传统深度学习算法,可为饱和电抗器在线监测提供重要依据。

As the key equipment in converter valves,the saturable reactor is used to protect the normal opening and closing of the thyristor.However,in the process of microsecond conduction of the thyristor,the high harmonics will produce high frequency mechanical vibration and thermal shock vibration on the saturable reactor,increasing the probability of the fault of the saturable reactor.Therefore,it is of great significance to evaluate its status.Recently,some methods were proposed to monitor the state of the saturable reactor,but most of them can only monitor the abnormal temperature rise of the core,and cannot identify the mechanical failure of the core.This paper presents a voiceprint identification model of saturable reactor core looseness based on optimized Stransform(OST)and improved deep residual shrinkage network(IDRSN).The saturable reactor vibration sound is used to identify the core looseness accurately.Firstly,the vibration test of saturable reactor at high frequency pulse excitation was carried out,and the voiceprint signals in different core looseness were measured.Secondly,in the principal value range of acoustic signal spectrum,the Gaussian window parameters were optimized by using the energy concentration formula to improve the time-frequency resolution of voiceprint spectrum.Thirdly,the voiceprint characteristics after loosening were analyzed,it was found that two characteristic indexes of high-low frequency ratio and dominant frequency ratio in low-frequency component can only give early warning to the state with high degree of loosening.Finally,the core loosening data of five different azimuth measuring points were brought into improved deep residual shrinkage network based on adaptively parametric rectifier linear unit for training,to eliminate the redundant information in the voiceprint spectrum and map the features in different degrees of looseness independently,so as to enhance the learning ability of common features.The results of saturable reactor vibration voiceprint test show that the vibration period of saturable reactor under high-frequency pulse excitation is the same as the electrical signal,which is 0.02 s.After the core is loosened,the frequency spectrum complexity and odd even subharmonic ratio do not change significantly,and the basic frequency changes greatly when the looseness is high,but it cannot give early warning in case of slight looseness.The proportion of dominant frequency of low frequency component and the ratio of high frequency to low frequency can sensitively reflect different loose states of iron core,but the threshold is difficult to set.The basic frequency proportion of low frequency component and the ratio of high frequency to low frequency can reflect different core loosening degrees sensitively,but the threshold is difficult to set.The recognition results of test data show that the convergence time of the proposed model is 262.13 s,and the average recognition accuracy of core looseness is 95.93%,realizing the unified recognition of five different measuring points.Compared with short time Fourier transform(STFT),the accuracy of OST is improved by 5.56%.The comparison between IDRSN and other neural networks such as residual network50 and deep residual shrink network shows that the calculation accuracy of IDRSN is improved by 2.82%and 4.67%respectively,and there is no false or missing judgment.The results show that:(1)Saturable reactor vibration voiceprint spectrum contains a large number of high order harmonics,and the ratio of 50 Hz odd and even frequency multiplication is equal.After the core clamp is loosened,the proportion of 50~1200 Hz component decreases,and the proportion of 1250~18000 Hz component increases.(2)The optimized S-transform method is proposed,which solves the problems of frequency band aliasing and spectrum leakage by adjusting the frequency domain window parameters with energy concentration.Compared with STFT,it has higher time-frequency resolution.(3)An improved deep residual shrinkage network based on adaptive parameter modified linear unit is built.The uniform recognition of voiceprint signals collected in different directions is realized by independent feature mapping,which has higher recognition accuracy compared with other traditional models.