泄漏检测信号滤波技术比较

通过泄漏检测模型试验分析测量信号中的噪声来源，在对比研究传统小波去噪、改进神经网络去噪、最小二乘拟合去噪等方法在实测数据中去噪效果的基础上，借鉴神经网络反向传播学习算法的思路，提出了信号预滤波结合闽值自学习小波去噪的综合滤波方法。该方法通过对恒定状态下带噪压力信号阈值自学习使得重构信号与期望输出均方误差最小来获得单一工况下的最佳去噪阈值，再将此阈值用于同一工况下整个时间段的去噪，这样根据不同工况下得到的最佳阈值可以获得最优输出。数值计算结果比较表明该方法对噪声的抑制作用明显，比传统小波去噪、改进神经网络去噪等方法效果更好。

Comparison of filtering methods for leak detection signals

The sources of noise in leak detection signals were analyzed based on physical model tests. Different filtering methods, such as the conventional wavelet denoising algorithm, the improved neural network denoising method, and the least-square spline fitting method were comparatively studied. Based on the thought of neural network back-propagation learning algorithm, a synthetic filtering method with the pre-filtering of signals combined with threshold serf-learning wavelet algorithm was proposed. With the method, the optimal denoising threshold in the single operating case was obtained by threshold self-learning of pressure signals with noise in the steady flow state to make the mean square error between reconstruction signals and desirable outputs minimal, and then the optimal threshold was used for denoising in the whole period of time in the same operating case. Thus the optimal output could be obtained according to the optimal threshold in different cases. Numerical results and comparative studies show that the present approach has obvious effects on noise suppression, and is superior to the conventional wavelet denoising algorithm and improved neural network denoising method.