基于重复模式提取与高斯混合模型的变压器故障声纹识别

Transformer Fault Voiceprint Recognition Based on Repeating Pattern Extraction and Gaussian Mixture Model

变电站存在大量干扰声音,高噪声环境会降低声纹识别的准确率。为此,提出一种基于重复模式提取(repeating pattern extraction technique,REPET)和高斯混合模型(Gaussian mixture model,GMM)的变压器故障声纹识别方法。首先,在真实的变压器油箱中模拟不同类型放电和机械故障声音。其次,针对变电站内非平稳性干扰声音,采用基于REPET的盲源分离算法将非稳定干扰声音从混合声音中分离。最后,针对主要由变压器冷却风扇声造成的持续性干扰声音,采用梅尔频率倒谱系数和基于GMM的声纹识别算法来降低风扇噪声对声纹识别系统的影响,并通过实验数据对该方法进行验证,实验室条件下对含噪声音的识别率可达100%。另外,GMM训练所需声音数据量少,针对实际变压器故障声音难以获取的情况,具有一定实用价值。

There are a lot of interference sounds in the substation,and the high noise environment will reduce the accuracy of voiceprint recognition.Accordingly,a transformer fault voiceprint recognition method based on repeating pattern extraction and Gaussian mixture model is proposed.Firstly,different types of discharge and mechanical faults are simulated in a real transformer oil tank.Secondly,aiming at the non-stationary interference sound in the substation,the blind source separation algorithm based on repeating pattern extraction is used to separate the non-stationary interference sound from the mixed sound.Finally,in view of the continuous interference sound mainly caused by the sound of the transformer cooling fan,the Mel frequency cepstrum coefficient features and the voiceprint recognition algorithm based on the Gaussian mixture model are selected to reduce the influence of the fan noise on the voiceprint recognition system,and the method is verified that the recognition rate of noise-containing sounds can reach 100%under laboratory conditions using experimental data.In addition,the amount of sound data required for the training of the Gaussian mixture model is small,and it has certain practical value for the situation that the actual transformer fault sound is difficult to obtain.