针对SF6局部放电下分解组分传统检测方法存在的消耗被测气体多、检测时间长、不适用于在线监测等不足，利用光声光谱技术具有检测气体灵敏度高、不消耗被测气体等特点，研究了用于SF6局部放电分解组分检测技术，给出了局部放电下SF6分解特征组分SO2、CO2、CF4的特征频谱，利用研制的光声光谱实验平台获得了气体的光声信号与气体体积分数关系，得到了SO2、CO2、CF4的最低检测限分别为3.8×10 -6、3.1×10 -6、4.7×10 -6，建立了用于降低SO2、CO2、CF4混合气体的光声信号交叉响应的RBF神经网络算法，使3种气体平均检测误差分别降为5.6%、1.6%、3.3%，提高了检测准确度，并用气相色谱法和检测管法的对比测量验证了其可信性，为解决交叉响应的影响问题提供了一种技术手段。

The traditional detection methods of SF6 decomposition components under partial discharge have some shortages，including consuming much detected gas，long detecting time，and unsuitable for on-line monitoring. While，photoacoustic spectroscopy has some advantages，including high sensitivity on detecting gas，and without consuming detected gas，etc. According to these reasons，the detection technology used in SF6 decomposition components under partial discharge is studied，and the feature spectrum of SO2，CO2 and CF4by SF6 decomposing is given. Through the photoacoustic spectroscopy device，the relation lines between photoacoustic spectroscopy(PAS) signal and concentration of gas components are obtained. The minimum detection limits of SO2，CO2 and CF4are about 3.8×10 -6，3.1×10 -6 and 4.7×10 -6 respectively. A method of RBF neural network is set up to reduce the crossover response of PAS signals of SO2，CO2 and CF4 mixed gas. It makes the average examination error of three gases reduce to 5.6%，1.6% and 3.3% respectively. Its reliability is checked by comparative testing of gas ehromatography and detector tube. The results indicate that the RBF neural network is useful in improving detection precision and provides a kind of technology to crossover response problem.

受国家重点基础研究发展计划资助项目（2009CB724506）；优秀博士论文基金资助（200749）