A spherical detector (SD) is capable of closely approaching a leak point and collecting leak sounds from the inside of a long pipeline, thereby enabling an extremely high leak detection sensitivity. However, acoustic noises arise from collision and friction while the SD is rolling forward, hindering the identification of leak acoustic signals. To address this challenge, this work presents a pipeline leak identification method for an SD based on combining variational mode decomposition (VMD) and a support vector machine (SVM). A leak generation system is set up where the pipe is water-filled, pressurized, and tiltable, and the SD can stand still or roll to collect a sufficient variety of leak sound samples. By decomposing the noisy signals into different modes and selecting the modes with high correlations to reconstruct the signals, the VMD can significantly decrease the collision noise. Additionally, the Mel frequency cepstral coefficients (MFCCs) are extracted and used to constitute a characteristic vector for SVM-based leak recognition. The trained neural network effectively identifies the occurrence of a leak; the recognition accuracy can reach up to 93 %, with a satisfactory specificity of 89.6 %.

球形探测器（SD）能够近距离接近泄漏点并收集来自长管道内部的泄漏声音，从而实现极高的泄漏检测灵敏度。然而，当 SD 向前滚动时，碰撞和摩擦会产生声学噪声，阻碍了泄漏声学信号的识别。为了应对这一挑战，这项工作提出了一种基于变分模式分解 (VMD) 和支持向量机 (SVM) 相结合的 SD 管道泄漏识别方法。在管道充满水、加压和可倾斜的情况下设置泄漏发生系统，SD 可以静止或滚动以收集足够种类的泄漏声音样本。通过将噪声信号分解为不同的模式，并选择相关性高的模式来重构信号，VMD 可以显着降低碰撞噪声。此外，Mel 频率倒谱系数 (MFCC) 被提取并用于构成基于 SVM 的泄漏识别的特征向量。经过训练的神经网络有效识别泄漏的发生；识别准确率可达93%，特异性达89.6%。