Research on a Novel Unsupervised-Learning-Based Pipeline Leak Detection Method Based on Temporal Kolmogorov–Arnold Network with Autoencoder Integration

Hengyu Wu, Jiang Zhu, Xianghua Zhang, Jian ChengSensors 2025 · DOI

TL;DR

This paper introduces a novel unsupervised pipeline leak detection system, the TKAN-AE, which combines a Temporal Kolmogorov-Arnold Network (TKAN) with an autoencoder (AE). This hybrid model, coupled with a new anomaly sequence labeling method, significantly improves leak detection accuracy and interpretability, offering a cost-effective solution for urban water management.

What Problem Does This Paper Address?

Water supply pipelines are prone to leaks due to aging infrastructure, installation issues, and environmental factors. These leaks lead to significant water wastage and pose risks to urban water supply systems. Traditional AI methods for leak detection struggle with effectively capturing the entire leak process and can be computationally expensive, often requiring powerful GPUs. Existing methods also face challenges with accurately identifying leak segments and can lead to false positives or negatives, especially with dynamic data. The paper aims to address these limitations by proposing a more accurate, interpretable, and cost-efficient unsupervised leak detection system.

Methods

The core of the proposed solution is the TKAN-AE model, a hybrid architecture that integrates a Temporal Kolmogorov-Arnold Network (TKAN) with an Autoencoder (AE).

  • Autoencoder (AE): This component is used for learning efficient data representations. The encoder compresses the input time series data into a latent representation, and the decoder reconstructs the original data from this representation. The reconstruction error between the original and reconstructed data is a key indicator for anomaly detection.
  • Temporal Kolmogorov-Arnold Network (TKAN): This novel neural network architecture is designed to capture complex temporal dependencies in time series data. It combines the strengths of Kolmogorov-Arnold Networks (KANs) for function approximation with the gating mechanisms of Long Short-Term Memory (LSTM) networks. This allows TKAN to effectively handle long-term dependencies, a common challenge in traditional Recurrent Neural Networks (RNNs).

In addition to the model, the paper introduces a novel unsupervised anomaly sequence labeling method. This method builds upon traditional reconstruction error theory but incorporates prior knowledge and a detailed analysis of the reconstruction error curve. Instead of relying on simple fixed thresholds, it identifies potential leak segments by analyzing key points on the reconstruction error curve, aiming to improve both accuracy and interpretability.

Datasets

The study utilized real-world pipeline leakage experimental data collected from urban water supply pipelines in Shanghai, China. This dataset was specifically constructed for the research and is referred to as a benchmark dataset. The experiments involved simulating leaks by manipulating valves at different openings and monitoring pressure fluctuations using high-frequency dynamic pressure sensors. The dataset consists of 12 leakage samples and 200 normal samples, carefully balanced to prevent bias during model training.

Key Insights and Ideas

  • KANs for Time Series: The paper highlights the potential of Kolmogorov-Arnold Networks (KANs), originally designed for general functions, to be adapted for time series analysis through the TKAN architecture. This offers a more interpretable and parameter-efficient alternative to traditional deep learning models.
  • Hybrid Model Synergy: The combination of TKAN and AE proves effective. TKAN excels at capturing temporal dynamics, while the AE's reconstruction capabilities are leveraged for anomaly detection, creating a synergistic approach.
  • Beyond Simple Thresholds: The proposed anomaly labeling method moves beyond static thresholds, which are prone to errors in dynamic systems. By analyzing the reconstruction error curve in conjunction with prior knowledge, it offers a more nuanced and robust way to identify leak events.
  • Real-World Validation: The use of actual field data from urban pipelines lends significant credibility to the proposed method, demonstrating its practical applicability in industrial settings.

Key Results

The proposed TKAN-AE model and its associated labeling method achieved impressive results:

  • Segment-wise Precision: The system achieved a high segment-wise precision of 93.1%.
  • Improved Precision: The novel labeling method improved segment-level precision by 54.2% compared to traditional methods.
  • Robustness: The model demonstrated robustness in complex environments with abrupt data changes.
  • Efficiency: The TKAN-AE model has a significantly reduced number of parameters compared to Transformer-based models, making it more suitable for cost-efficient industrial deployment.
  • Generalizability: The proposed labeling method showed strong generalization capabilities when applied across different models.

Limitations and Future Work

  • Data Scarcity: A significant challenge is the scarcity of anomaly samples due to the high cost and safety concerns of conducting leakage experiments. The unsupervised nature of the current approach doesn't fully leverage the limited anomaly data.
  • Future Work: The authors suggest transfer learning (combining lab and real-world data) and semi-supervised learning (using a small amount of labeled anomaly data with a large volume of unlabeled normal data) as potential solutions.
  • Model Adaptability: The current model is tailored to a specific urban pipeline network. Changes in the pipeline configuration would require retraining.
  • Future Work: Federated learning is proposed as a promising avenue to address adaptability by enabling real-time model updates and distributed learning while preserving data privacy.
  • Model Interpretability for Localization: While KANs offer improved interpretability, further work is needed to fully leverage this for precise leak point localization.
  • Future Work: Enhancing KAN interpretability could improve the visualization of processes like negative pressure wave analysis, speeding up and increasing the accuracy of leak localization.

Summary updated

Abstract

Artificial intelligence (AI) technologies have been widely applied to the automated detection of pipeline leaks. However, traditional AI methods still face significant challenges in effectively detecting the complete leak process. Furthermore, the deployment cost of such models has increased substantially due to the use of GPU-trained neural networks in recent years. In this study, we propose a novel leak detector, which includes a new model and a sequence labeling method that integrates prior knowledge with traditional reconstruction error theory. The proposed model combines the Kolmogorov-Arnold Network (KAN) with an autoencoder (AE). This model combines the Kolmogorov-Arnold Network (KAN) with an autoencoder (AE), forming a hybrid framework that effectively captures complex temporal dependencies in the data while exhibiting strong pattern modeling and reconstruction capabilities. To improve leak detection, we developed a novel unsupervised anomaly sequence labeling method based on traditional reconstruction error theory, which incorporates an in-depth analysis of the reconstruction error curve along with prior knowledge. This method significantly enhances the interpretability and accuracy of the detection process. Field experiments were conducted on real urban water supply pipelines, and a benchmark dataset was established to evaluate the proposed model and method against commonly used models and methods. The experimental results demonstrate that the proposed model and method achieved a high segment-wise precision of 93.1%. Overall, this study presents a transparent and robust solution for automated pipeline leak detection, facilitating the large-scale, cost-effective development of digital twin systems for urban pipeline leak emergency management.
BibTeX
@inproceedings{wu2025research,
  author = {Hengyu Wu and Jiang Zhu and Xianghua Zhang and Jian Cheng},
  title = {{Research on a Novel Unsupervised-Learning-Based Pipeline Leak Detection Method Based on Temporal Kolmogorov–Arnold Network with Autoencoder Integration}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25020384},
}
Figure