Multiclass Gas Pipeline Leak Detection Using Multi-Domain Features and Genetic Algorithms: A Lightweight AI Framework for Industrial Safety

on

Gas pipeline leakage remains one of the most critical challenges in industrial infrastructure because undetected leaks can result in catastrophic accidents, economic losses, and environmental damage. Although numerous leak detection techniques have been developed, accurately distinguishing multiple leak types remains difficult, particularly when only limited labeled data and computational resources are available. This study proposes an efficient multiclass leak detection framework that integrates multi-domain acoustic feature extraction with genetic algorithm-based feature optimization and conventional machine learning classifiers. By combining time-domain and frequency-domain information, the proposed approach achieves high classification accuracy while maintaining computational efficiency. The research demonstrates that carefully engineered features and evolutionary optimization can provide a practical alternative to computationally intensive deep learning models for real-world industrial applications.

Bibliographic Information

Item Information
Article Title Multiclass gas pipeline leak detection using multi-domain signals and genetic algorithm-optimized classification models
Authors Wiwit Suprihatiningsih, Dedik Romahadi, Hadi Pranoto, Rikko Putra Youlia, Fajar Anggara, and Rizky Rahmatullah
Journal Teknomekanik
Volume & Issue Volume 9, Issue 1
Publication Date February 2026
Pages 25–41
DOI https://doi.org/10.24036/teknomekanik.v9i1.38372
Publisher Universitas Negeri Padang
License Creative Commons Attribution 4.0 International (CC BY 4.0)
e-ISSN 2621-8720
p-ISSN 2621-9980
Keywords gas pipeline leak detection; multiclass classification; acoustic signal; genetic algorithm

1. Research Background

  • Gas pipeline systems are indispensable for modern energy transportation. Pipelines remain one of the most efficient methods for transporting natural gas and other gaseous media because they provide high transport capacity, rapid delivery, and relatively low operational costs. Nevertheless, pipeline failures can rapidly escalate into major industrial accidents, making reliable leak detection an essential component of operational safety.
  • Leak detection has evolved beyond simply identifying leak occurrence. Traditional monitoring systems generally detect pressure, flow, vibration, or acoustic anomalies to indicate the presence of a leak. However, modern industrial environments increasingly require automatic identification of different leak types to enable faster diagnosis and more appropriate emergency responses.
  • Multiclass leak classification remains a challenging engineering problem. Acoustic signals generated by gas leaks vary according to pipeline material, operating conditions, sensor configuration, and environmental noise. These variations create complex classification problems that become even more difficult when only limited labeled datasets are available.
  • Deep learning is not always the most practical solution. Although convolutional neural networks have demonstrated excellent performance in many signal-processing applications, they generally require extensive labeled datasets, significant computational resources, and long training times. These requirements limit their applicability in many industrial environments where computational resources are constrained.
  • Conventional machine learning remains highly competitive for small datasets. Algorithms such as Support Vector Machines, Random Forests, Linear Discriminant Analysis, K-Nearest Neighbors, Decision Trees, Artificial Neural Networks, and Naïve Bayes continue to offer strong predictive performance when supported by carefully designed feature engineering strategies.
  • Feature engineering plays a critical role in signal classification. Rather than relying solely on raw acoustic signals, combining temporal information with frequency-domain characteristics can reveal complementary patterns associated with different leak conditions. Such multi-domain representations improve the ability of machine learning models to discriminate among multiple leak classes.
  • Genetic algorithms provide an efficient mechanism for feature optimization. Evolutionary optimization techniques are capable of identifying compact subsets of highly informative features while simultaneously reducing feature redundancy and computational complexity. This capability is particularly valuable when deploying machine learning models on industrial hardware with limited computational capacity.
  • The study addresses an important research gap in practical industrial AI. Instead of pursuing increasingly complex deep neural architectures, the research demonstrates that integrating multi-domain acoustic features with genetic algorithm-based feature selection and conventional classifiers can achieve high multiclass classification accuracy while maintaining simplicity, interpretability, and computational efficiency.

2. Research Objectives

  • To develop a multiclass gas pipeline leak detection framework capable of distinguishing twelve different leak conditions using acoustic signal analysis.
  • To integrate time-domain signal characteristics with Power Spectral Density (PSD)-based frequency-domain features into a unified multi-domain feature representation.
  • To employ a genetic algorithm as a wrapper-based feature selection method for identifying the most informative feature subset while reducing computational complexity.
  • To compare the performance of seven conventional machine learning classifiers under identical experimental conditions using the GPLA-12 benchmark dataset.
  • To optimize the best-performing classification model through simultaneous genetic algorithm-based feature selection and hyperparameter optimization.
  • To demonstrate that lightweight machine learning models can achieve competitive multiclass leak detection performance without requiring computationally intensive deep learning architectures.

3. Why This Research Matters

  • Improves industrial safety. Accurate identification of gas pipeline leak types enables faster emergency response, reducing the likelihood of explosions, equipment damage, environmental contamination, and human casualties.
  • Supports practical artificial intelligence deployment. The proposed methodology demonstrates that highly accurate industrial AI systems do not necessarily require deep neural networks, making deployment feasible for facilities with limited computational resources.
  • Enhances predictive maintenance. Reliable multiclass leak identification allows maintenance personnel to diagnose pipeline conditions more effectively and prioritize corrective actions before failures become catastrophic.
  • Reduces computational cost. Genetic algorithm-based feature optimization minimizes unnecessary variables, enabling faster model training and inference while preserving high predictive performance.
  • Bridges signal processing and machine learning. By combining temporal signal descriptors with spectral analysis, the framework illustrates how domain-specific feature engineering can substantially improve machine learning performance.
  • Demonstrates the continuing relevance of classical machine learning. The findings challenge the assumption that deep learning is always the optimal solution by showing that carefully engineered conventional models remain highly competitive for limited-data engineering applications.
  • Provides a scalable framework for industrial monitoring. The proposed methodology can potentially be adapted for other acoustic monitoring applications involving rotating machinery, structural health monitoring, fault diagnosis, and industrial condition monitoring where computational efficiency is essential.

4. Research Methodology

  • Research Design

    The study employed a quantitative experimental approach to develop and evaluate a multiclass gas pipeline leak detection framework using acoustic signal processing, feature engineering, evolutionary optimization, and supervised machine learning. Rather than relying on end-to-end deep learning, the framework emphasized carefully designed feature extraction and feature selection to maximize classification performance while minimizing computational complexity.

  • Dataset

    The experiments utilized the publicly available GPLA-12 dataset, which contains acoustic recordings collected from twelve different gas pipeline leakage conditions. The dataset consists of 684 samples, with each class containing 57 balanced observations. Each sample contains 1,460 signal variables, accompanied by corresponding class labels. The balanced dataset enabled fair comparison among classification algorithms while representing multiple leak scenarios encountered in industrial environments.

  • Multi-Domain Feature Engineering

    Instead of relying solely on raw acoustic signals, the researchers combined two complementary feature domains. Time-domain information preserved waveform characteristics and transient signal behavior, while frequency-domain features were extracted using Power Spectral Density (PSD) estimation based on Welch's method. The integration of temporal and spectral information produced a richer feature representation capable of distinguishing subtle differences among leak categories.

  • Power Spectral Density Analysis

    Frequency-domain analysis was performed using Welch's PSD estimation technique. The acoustic signals were partitioned into overlapping segments, windowed using a Hann function, transformed into frequency spectra, and averaged to reduce spectral variance. This process generated robust spectral descriptors while minimizing the influence of measurement noise, providing stable inputs for subsequent machine learning classification.

  • Classification Models

    Seven conventional supervised machine learning algorithms were evaluated under identical experimental conditions:

    • Naïve Bayes (NB)
    • K-Nearest Neighbors (KNN)
    • Random Forest (RF)
    • Linear Discriminant Analysis (LDA)
    • Decision Tree (DT)
    • Support Vector Machine (SVM)
    • Artificial Neural Network (ANN)

    These classifiers were selected because they are widely used for engineering classification problems involving relatively small datasets and provide different learning characteristics for comparative evaluation.

  • Genetic Algorithm Feature Selection

    Feature optimization was performed using a wrapper-based genetic algorithm. Each candidate solution represented a binary feature selection vector indicating whether individual features were retained or discarded. Through iterative selection, crossover, and mutation operations over 300 generations, the genetic algorithm searched for feature subsets that maximized classification performance while minimizing model complexity.

  • Model Evaluation

    Model performance was evaluated using five-fold cross-validation to ensure reliable estimation of generalization capability. During each iteration, four folds were used for model training while one fold served as the testing dataset. Performance was quantified using multiple evaluation metrics including accuracy, precision, recall (sensitivity), specificity, and F-score.

  • Experimental Environment

    All experiments were implemented using MATLAB 2024b on a Windows 11 workstation equipped with a 14th-generation Intel Core i9 processor and an NVIDIA RTX 4080 Super graphics card. However, the graphics processor was intentionally not utilized during computation, demonstrating that the proposed framework can operate effectively using only CPU resources.

5. Key Findings

Multi-Domain Feature Fusion Significantly Improved Leak Classification

The combination of time-domain acoustic descriptors and frequency-domain PSD features produced a richer representation of pipeline leak characteristics than either domain alone. The complementary information captured transient waveform behavior alongside spectral energy distributions, enabling more effective discrimination among twelve leak categories.

The study demonstrates that feature engineering remains a decisive factor in engineering machine learning applications, particularly when datasets are relatively small and computational resources are limited.

Genetic Algorithm Successfully Reduced Feature Complexity

The genetic algorithm effectively identified compact subsets of informative features while removing redundant variables. Rather than increasing model complexity, the optimization process improved discriminative capability and reduced computational burden simultaneously.

This evolutionary feature selection strategy enabled the classifiers to focus on the most meaningful acoustic characteristics associated with different leak conditions, contributing directly to improved predictive performance.

Conventional Machine Learning Achieved Excellent Predictive Accuracy

The comparative evaluation of seven machine learning algorithms demonstrated that carefully optimized conventional classifiers remain highly competitive for industrial fault diagnosis. After genetic algorithm optimization, the best-performing classifier achieved a maximum multiclass classification accuracy of 96.35% on the GPLA-12 benchmark dataset.

The findings indicate that excellent predictive performance can be achieved without relying on computationally intensive deep convolutional neural networks.

The Framework Maintained Low Computational Requirements

An important outcome of the research is the successful balance between predictive accuracy and computational efficiency. All experiments were executed using CPU resources only, demonstrating that high-performance multiclass leak detection does not necessarily require specialized GPU hardware or extensive computing infrastructure.

This characteristic increases the practicality of deploying the proposed framework in industrial environments where computational resources may be limited.

The Proposed Framework Supports Resource-Constrained Industrial Applications

Compared with previously reported deep CNN approaches, the proposed methodology achieved comparable or superior classification performance while employing a considerably simpler model architecture. The reduced computational complexity makes the framework suitable for industrial monitoring systems requiring rapid decision-making, limited training data, and efficient implementation.

The Study Clearly Defines Its Operational Scope

The research specifically focuses on supervised multiclass classification of acoustic gas leak signals. The authors explicitly state that leak localization, leak severity estimation, multi-sensor fusion, and domain adaptation across different pipeline materials remain outside the scope of the current investigation.

6. Scientific Contribution

  • Introduces an integrated multi-domain acoustic feature representation that combines temporal waveform characteristics with Power Spectral Density features to improve multiclass gas leak discrimination.
  • Demonstrates the effectiveness of wrapper-based genetic algorithm optimization for simultaneously reducing feature dimensionality and improving classification performance.
  • Provides a systematic benchmark comparison of seven conventional machine learning algorithms under identical experimental conditions using the GPLA-12 dataset.
  • Shows that lightweight machine learning models can rival deep learning performance when supported by effective feature engineering and optimization strategies.
  • Contributes an efficient methodology for multiclass industrial fault diagnosis that balances predictive accuracy, computational efficiency, and implementation simplicity.
  • Expands knowledge in intelligent pipeline monitoring by demonstrating how evolutionary optimization and signal processing techniques can be integrated into practical industrial AI systems.

7. Industrial Implications

  • Enhances pipeline safety. Accurate multiclass leak identification enables operators to respond more rapidly to hazardous pipeline failures, reducing operational risks and improving public safety.
  • Supports predictive maintenance. Earlier and more reliable leak identification allows maintenance activities to be scheduled before failures develop into catastrophic events.
  • Reduces deployment cost. Because the framework operates effectively without GPU acceleration, implementation costs can be significantly reduced compared with deep learning solutions requiring specialized hardware.
  • Improves industrial decision support. High classification accuracy enables maintenance engineers to differentiate multiple leak conditions rather than merely detecting leak existence, supporting more informed maintenance planning.
  • Facilitates real-time monitoring. Reduced computational complexity makes the proposed methodology suitable for embedded monitoring systems and industrial edge computing environments.
  • Demonstrates scalable engineering AI. The overall methodology can potentially be adapted to other industrial acoustic monitoring applications such as rotating machinery diagnostics, structural health monitoring, compressor fault diagnosis, and predictive maintenance systems.

8. Research Limitations

  • The experimental evaluation was conducted exclusively using the GPLA-12 benchmark dataset. Although this dataset contains twelve balanced leak categories, broader validation using additional datasets, different pipeline materials, operating conditions, and sensor configurations is necessary to evaluate the generalizability of the proposed framework.
  • The proposed methodology focuses solely on supervised multiclass classification of acoustic leak signals. Other important diagnostic tasks, including leak localization, leak size estimation, leak severity assessment, and remaining useful life prediction, were beyond the scope of the present study.
  • Only acoustic signal information was utilized for model development. Other sensing modalities such as pressure measurements, flow monitoring, vibration fusion, thermal sensing, or distributed optical fiber sensing were not incorporated into the proposed framework.
  • The comparative analysis considered seven conventional machine learning algorithms. Although these models represent widely adopted classification techniques, more recent approaches such as transformer-based architectures, graph neural networks, or hybrid deep learning models were not investigated.
  • The framework assumes that all leak categories are represented within the training dataset. Open-set recognition involving previously unseen leak mechanisms, changing environmental conditions, or domain shifts between different pipeline systems remains an important challenge for future investigation.
  • Real-time deployment and long-term industrial validation were not included in the experimental evaluation. Consequently, additional studies are required to investigate system robustness under continuous industrial operation and varying environmental conditions.

9. Future Research Opportunities

  • Investigate multimodal sensor fusion by combining acoustic signals with pressure, flow rate, vibration, thermal imaging, or fiber optic sensing to further improve leak detection reliability under diverse operating conditions.
  • Extend the framework from leak classification to comprehensive pipeline diagnostics by incorporating leak localization, leak severity estimation, and predictive maintenance capabilities within a unified intelligent monitoring system.
  • Evaluate the proposed methodology using larger industrial datasets collected from different pipeline materials, operating pressures, environmental conditions, and geographical locations to improve model robustness and generalizability.
  • Investigate hybrid machine learning architectures that combine engineered acoustic features with lightweight deep learning models to balance interpretability, computational efficiency, and predictive performance.
  • Develop adaptive learning strategies capable of recognizing previously unseen leak patterns, handling concept drift, and supporting continuous model updating during long-term industrial operation.
  • Explore deployment on embedded industrial hardware, edge computing platforms, and Industrial Internet of Things (IIoT) systems to enable real-time pipeline condition monitoring with minimal computational overhead.
  • Investigate explainable artificial intelligence (XAI) techniques to improve model transparency, allowing engineers to better understand which acoustic characteristics contribute most significantly to multiclass leak identification.

10. Potential for Public Policy Citation

The proposed framework offers practical value for government agencies, pipeline operators, environmental regulators, and industrial safety authorities responsible for critical energy infrastructure. By demonstrating that highly accurate multiclass leak detection can be achieved using computationally efficient machine learning techniques, the research supports the development of evidence-based monitoring strategies that improve operational safety while reducing implementation costs.

The methodology aligns with broader initiatives promoting industrial digitalization, infrastructure resilience, environmental protection, and sustainable energy transportation. Policymakers responsible for pipeline integrity management may find the framework useful when developing technical guidelines for intelligent monitoring systems, preventive maintenance strategies, and risk-based inspection programs.

11. Who Should Read This Paper?

  • Mechanical engineers working in pipeline integrity and condition monitoring.
  • Researchers in machine learning, signal processing, and industrial artificial intelligence.
  • Pipeline operators and maintenance engineers responsible for oil and gas transportation systems.
  • Industrial automation specialists developing predictive maintenance solutions.
  • Researchers in structural health monitoring and intelligent sensing technologies.
  • Electrical and instrumentation engineers involved in industrial monitoring systems.
  • Graduate students studying machine learning applications in mechanical and industrial engineering.
  • Researchers interested in lightweight artificial intelligence for resource-constrained engineering applications.

12. Final Thoughts

This research demonstrates that effective industrial artificial intelligence does not necessarily require increasingly complex deep learning architectures. Through the thoughtful integration of multi-domain acoustic feature engineering, genetic algorithm-based feature optimization, and conventional supervised machine learning, the proposed framework achieves excellent multiclass leak classification performance while maintaining computational simplicity.

One of the study's greatest strengths lies in its engineering practicality. Rather than pursuing computational complexity, the researchers focus on designing informative feature representations and efficient optimization strategies that can operate successfully using standard CPU resources. Such characteristics make the proposed methodology highly attractive for industrial environments where computational efficiency, implementation cost, and model interpretability remain important considerations.

Beyond gas pipeline monitoring, the methodological framework presented in this study provides valuable insights into how signal processing, feature engineering, evolutionary optimization, and classical machine learning can be combined to address a wide range of industrial condition monitoring problems. As industries continue to adopt intelligent monitoring technologies, this work illustrates that carefully engineered lightweight AI solutions remain an important and highly competitive direction for future engineering research.

13. Suggested Citations

Teknomekanik (UNP)

Suprihatiningsih, W., Romahadi, D., Pranoto, H., Youlia, R. P., Anggara, F., & Rahmatullah, R. (2026). Multiclass gas pipeline leak detection using multi-domain signals and genetic algorithm-optimized classification models. Teknomekanik, 9(1), 25–41. https://doi.org/10.24036/teknomekanik.v9i1.38372

APA (7th Edition)

Suprihatiningsih, W., Romahadi, D., Pranoto, H., Youlia, R. P., Anggara, F., & Rahmatullah, R. (2026). Multiclass gas pipeline leak detection using multi-domain signals and genetic algorithm-optimized classification models. Teknomekanik, 9(1), 25–41. https://doi.org/10.24036/teknomekanik.v9i1.38372

IEEE Style

W. Suprihatiningsih, D. Romahadi, H. Pranoto, R. P. Youlia, F. Anggara, and R. Rahmatullah, "Multiclass gas pipeline leak detection using multi-domain signals and genetic algorithm-optimized classification models," Teknomekanik, vol. 9, no. 1, pp. 25–41, Feb. 2026, doi:10.24036/teknomekanik.v9i1.38372.

Harvard Style

Suprihatiningsih, W., Romahadi, D., Pranoto, H., Youlia, R.P., Anggara, F. and Rahmatullah, R., 2026. Multiclass gas pipeline leak detection using multi-domain signals and genetic algorithm-optimized classification models. Teknomekanik, 9(1), pp.25–41. Available at: https://doi.org/10.24036/teknomekanik.v9i1.38372.

Vancouver Style

Suprihatiningsih W, Romahadi D, Pranoto H, Youlia RP, Anggara F, Rahmatullah R. Multiclass gas pipeline leak detection using multi-domain signals and genetic algorithm-optimized classification models. Teknomekanik. 2026;9(1):25–41. doi:10.24036/teknomekanik.v9i1.38372.

Chicago (Author–Date)

Suprihatiningsih, Wiwit, Dedik Romahadi, Hadi Pranoto, Rikko Putra Youlia, Fajar Anggara, and Rizky Rahmatullah. 2026. "Multiclass Gas Pipeline Leak Detection Using Multi-Domain Signals and Genetic Algorithm-Optimized Classification Models." Teknomekanik 9 (1): 25–41. https://doi.org/10.24036/teknomekanik.v9i1.38372.

MLA (9th Edition)

Suprihatiningsih, Wiwit, et al. "Multiclass Gas Pipeline Leak Detection Using Multi-Domain Signals and Genetic Algorithm-Optimized Classification Models." Teknomekanik, vol. 9, no. 1, 2026, pp. 25–41. Crossref, https://doi.org/10.24036/teknomekanik.v9i1.38372.

14. Editorial Note

This review was prepared for Engineering Research Insights based exclusively on the published research article and is intended to provide an independent scholarly overview of the study. The review summarizes the research background, objectives, methodology, principal findings, scientific significance, industrial relevance, and future research opportunities while preserving the original scientific contributions of the authors. Readers are encouraged to consult the original publication for complete experimental details, mathematical formulations, and supplementary analyses.

15. SEO Meta Description

Discover how multi-domain acoustic features and genetic algorithm optimization achieve 96.35% accuracy for multiclass gas pipeline leak detection using lightweight machine learning models suitable for industrial monitoring.

16. SEO Keywords

gas pipeline leak detection, multiclass classification, genetic algorithm, acoustic signal processing, machine learning, feature selection, power spectral density, predictive maintenance, industrial artificial intelligence, pipeline monitoring, condition monitoring, signal processing, fault diagnosis, lightweight AI, engineering research

17. Recommended URL Slug

multiclass-gas-pipeline-leak-detection-genetic-algorithm-machine-learning-review

Comments

Post a Comment