Research on pipeline flange leakage detection method based on random forest and Pearson correlation coefficient

Abstract

Pipeline flange leakage is influenced by multiple factors, including discharge area, temperature, pressure, flow velocity, and installation position. Conventional leakage prediction models suffer from computational inefficiency and reduced generalizability due to high-dimensional feature spaces and redundant parameters. To overcome these limitations, this study introduces an interpretable hybrid framework integrating Random Forest (RF)-driven feature selection with Pearson correlation analysis for high-precision leakage diagnostics. Vibration signals were systematically acquired from paired upper and lower flanges at leakage-critical regions, followed by extraction of multicomponent features including total vibration energy, spectral kurtosis, centroid frequency, Shannon entropy, and bidirectional cross-correlation metrics. Through rigorous RF-based recursive feature elimination, four dominant sensitivity indicators were identified: inter-flange cross-correlation (CC), upper-flange centroid frequency (CF), total vibration energy (TVE) of both flanges, and downstream entropy (DE). Subsequent Pearson correlation mapping revealed minimal collinearity (|r| < 0.4) between CC, CF, and lower-flange TVE, establishing these orthogonal features as prediction model inputs. A Multilayer Perceptron (MLP) network is trained to establish the leakage prediction model. Experimental results demonstrate high accuracy of the proposed method, achieving a mean squared error (MSE) of approximately 6.8 and a coefficient of determination (R²) of 0.97. The feature dimensionality reduction not only improves prediction accuracy but also reduces iteration cycles.