基于声纹压缩和代价敏感的变压器状态检测评估方法

声纹检测技术可以助力巡检人员对变压器状态进行检测和评估。该文提出了一种基于声纹压缩和代价敏感的变压器状态检测和评估方法。该方法首先提取变压器音频的声纹特征，然后对声纹特征在频率维度上进行筛选和压缩，最后使用卷积神经网络评估变压器状态，并引入代价敏感损失函数提高对难检出样本的关注度。以某35kV变压器为对象，通过收集现场音频、模拟实验和样本扩充得到变压器音频数据集。测试结果表明，该文所提的方法将声纹维度从1025维降低到80维，计算量和显存分别降低到8.1%和7.7%。同时，声纹识别准确率达到83.5%，并将最难检出的短路电流异常状态的召回率从48.2%提升至63.6%。

Evaluation method for transformer condition detection based on acoustic compression and cost sensitivity

The acoustic pattern detection technique can assist inspectors in detecting and evaluating the transformer condition. The paper proposes a transformer condition detection and assessment method based on acoustic pattern compression and cost sensitivity. The method first extracts the acoustic features of transformer audio, then filters and compresses the acoustic features in the frequency dimension, and finally evaluates the transformer status using a convolutional neural network, and introduces a cost-sensitive loss function to improve the attention to the difficult-to-detect samples. Taking a 35kV transformer as an object, the transformer audio dataset is obtained by collecting field audio, simulation experiments and sample expansion. The test results show that the method proposed in this paper reduces the voiceprint dimension from 1025 to 80 dimensions, and the computation and memory are reduced to 8.1% and 7.7%, respectively. Meanwhile, the accuracy of sound pattern recognition reaches 83.5% and the recall rate of the most difficult-to-detect short-circuit current anomaly state is improved from 48.2% to 63.6%.