ENHANCING THE ROBUSTNESS OF ACOUSTIC LEAK DETECTION METHOD BY APPLYING ENTROPY AND RELATIONAL FEATURES

Abstract

Addressing the challenge of reliable transportation of energy carriers via pipelines is a priority for ensuring the state's environmental and economic sustainability. The reliability of existing pipeline condition monitoring methods, particularly acoustic leak detection techniques, is significantly affected by variations in operational conditions. This issue necessitates the development of advanced approaches for generating condition features that are robust to operational uncertainty. Objective. The objective of this work is to improve the diagnostic signal processing algorithm within the acoustic pipeline leak detection method by developing a state classification method robust to changes in operational parameters. The tasks include justifying the selection of informative acoustic signal features that are robust to variations in operating pressure and characteristics of the test acoustic signal, selecting the optimal classifier architecture capable of effective generalization to new data by comparing linear, non-linear, and ensemble machine learning methods. Methodology. The research methodology is based on processing acoustic signals reflected from pipeline wall inhomogeneities. These signals were recorded within the pipeline medium because of the propagation of generated test signals (active diagnostic method). Experimental data were acquired for five simulated pipeline states (closed pipeline and pipeline with leaks with diameters of 1, 3, 5, and 10 mm) under varying pressure conditions and using different test signal characteristics. In contrast to traditional approaches that rely on absolute spectral-energy parameters for feature vector formation, this study proposes augmenting the feature vector with entropy components and relational features derived from the ratios between estimates obtained for each recording channel. This approach enabled the compensation of common-mode noise caused by variations in operating conditions. A comparison was conducted between linear (Perceptron), non-linear (MLP), and ensemble (Random Forest) classifier architectures. Model performance was evaluated using stratified cross-validation on the training set and by testing on a new dataset containing pressure variations and test signal parameters unseen by the models. Scientific Novelty. The scientific novelty of the research lies in: developing a combined feature vector using entropy and relational characteristics of the acoustic signal, which ensured classifier robustness to variations in operating pressure and test signal carrier frequency; experimentally confirming the non-linearity of the acoustic leak signal classification task within the proposed feature space; establishing the superior generalization capability of the ensemble classifier compared to neural network-based approaches, which achieved a classification accuracy of 91% on an independent dataset. Conclusion. The effectiveness of applying the proposed feature vector and the Random Forest ensemble model within the acoustic pipeline state detection algorithm has been proven. The method demonstrated robustness to variations in operating pressure and test signal parameters. The research results can be applied in the development of algorithmic support for pipeline condition monitoring systems that are resilient to changes in technological transportation conditions.