RETRACTED: Machine vision model for drip leakage detection of pipeline

Cong Xiao, Yiyuan Zhou, Y. WuPLoS 2025 · DOI · OpenAlex

This paper, titled "Machine vision model for drip leakage detection of pipeline," presents a framework for the early detection and quantification of minor pipeline leaks. By leveraging computer vision and mathematical modeling, the authors aim to catch "micro-leaks" before they escalate into catastrophic industrial failures.

Note:* This paper is marked as RETRACTED* by the publisher. The following summary describes the methods and claims as presented in the original text for informational purposes.

TL;DR

The researchers developed a machine vision system that uses YOLOv5 for droplet detection and polynomial curve fitting to reconstruct the 3D volume of a leak. By treating a falling droplet as a "solid of revolution," the system achieves over 95% accuracy in estimating leak volume from simple 2D industrial camera feeds.

What Problem Does This Paper Address?

Pipelines are the "blood vessels" of industrial infrastructure, but they are prone to leaks due to harsh operating conditions. While major bursts are easy to spot, minor "drips" are often missed by traditional sensors or manual inspections. These small leaks are early warning signs; catching them early allows for proactive maintenance.

The core challenge is not just detecting that a drip is happening, but quantifying it. Measuring the volume of a transient, falling liquid droplet using only a camera requires overcoming issues like image noise, reflections, and the high speed of the event.

Methods

The authors propose a multi-stage pipeline to move from raw video to a precise volume measurement:

  1. Detection & Capture: A CCD industrial camera captures the drip. The YOLOv5 (You Only Look Once v5) algorithm is used to identify and "crop" the droplet from the background in real-time.
  2. Image Preprocessing: To handle the "messiness" of industrial environments, the authors use wavelet transforms for denoising and morphological operations (like "filling" and "burring") to remove reflections and hollow spots within the droplet image.
  3. Contour Fitting: This is the mathematical heart of the paper. After extracting boundary points from the 2D image, the authors use the Least Squares Method and Multiple Linear Regression to fit polynomials to the droplet's profile. They tested 1st through 6th-degree polynomials, ultimately determining that a 4th-degree (quadric) fitting provided the best balance between accuracy and computational complexity.
  4. 3D Reconstruction: Because droplets are generally symmetrical, the authors treat the 2D profile as a "rotating body." By integrating the mathematical function of the profile around its central axis, they calculate the 3D volume.
  5. Physical Calibration: They use the convex lens imaging principle to calculate an adjustment coefficient ($k$), which converts "pixel volume" into actual milliliters (mL).

Datasets

The study did not use a standard public dataset. Instead, the authors built a custom experimental test bench designed to simulate an industrial cooling water circulation system.

  • Data Collection: They ran six distinct experiments (I-VI).
  • Variables: They varied the object distance (10cm vs. 15cm) and the drip frequency (ranging from one drop every 3 seconds to one every 10 seconds).
  • Ground Truth: To validate their model, they collected the leaked water and weighed it, using the density of water to determine the "true" volume.

Key Insights and Ideas

  • Symmetry as a Solution: Droplet images are often distorted on one side due to lighting or camera angles. The authors realized they could analyze only the "cleaner" left side of the droplet and use its physical symmetry to mirror the profile, creating a more accurate 3D model than if they had used the distorted raw edges.
  • The "Quadric" Sweet Spot: While higher-degree polynomials can fit a curve more tightly, they often introduce "wiggles" (overfitting). The authors found that the 4th-degree polynomial perfectly captured the specific physics of a falling drop—the concave top where it detaches and the convex bottom where gravity pulls it.
  • Dynamic Capture: By using YOLOv5, the system doesn't just look at a static image; it dynamically captures the "critical" moment of the drip, allowing for volume estimation even when the camera's frame rate is relatively low (30 fps).

Key Results

  • High Accuracy: At an object distance of 10cm, the model achieved a peak accuracy of 96.04%.
  • Distance Sensitivity: Accuracy slightly decreased as the camera moved further away, dropping to roughly 93.57% at 15cm.
  • Frequency Impact: The model performed better with lower drip frequencies. When drips occurred rapidly (every 3 seconds), the "noise" in the capture increased, slightly lowering the accuracy compared to slower drips (every 10 seconds).

Limitations and Future Work

  • Retraction Status: The most significant limitation is the paper's retraction, which suggests that the underlying data or methodology may have unresolved integrity issues.
  • Hardware Constraints: The authors note that standard 30fps cameras are a bottleneck. High-speed leaks would likely require much more expensive, high-frame-rate hardware to capture clear profiles.
  • Environmental Robustness: The experiments were conducted with water in a controlled lab. Real-world industrial fluids (like thick oils or corrosive chemicals) have different viscosities and surface tensions, which would require the mathematical model to be recalibrated.
  • Lighting: The system relies heavily on clear silhouettes; poor lighting or steam in a factory setting could significantly degrade the segmentation performance.

Summary updated

Abstract

The prevailing trend in industrial equipment development is integration, with pipelines as the lifeline connecting system components. Given the often harsh conditions of these industrial equipment pipelines, leakage is a common occurrence that can disrupt normal operations and, in severe cases, lead to safety accidents. Early detection of even minor drips at the onset of leakage can enable timely maintenance measures, preventing more significant leaks and halting the escalation of pipeline failures. In light of this, our study investigates a method for monitoring pipe drips in industrial equipment using machine vision technology. We propose a machine vision model specifically designed for pipe drip detection, aiming to facilitate monitoring of pipe system drips. The system designed to collect the image of the droplet side cross-section with a Charge charge-coupled device (CCD) industrial camera, is aided by the computer image processing system used to analyze and process the collected images. Image enhancement technology is applied to improve the visibility of the image and image filtering technology is applied to remove the noise of the image. With the help of image segmentation technology, target droplet identification and division are achieved. Morphological reconstruction and region-filling techniques are used to remove the noise caused by shooting in the side cross-section image, such as hollow, reflection, and irregular droplet edge, to upgrade the quality of the solution droplet edge. The mathematical model is established for boundary position points extracted from the droplet side cross-section image. Then, the fitting droplet image is drawn. The droplet volume is obtained by calculating the volume of the rotating body. The two-dimensional image of the target droplet is obtained dynamically through the camera capture technology. The droplet boundary extraction algorithm is proposed, and the three-dimensional model of the target droplet is established, so the volume calculation problem of the droplet is solved, which provides a way of thinking for drip leakage detection of the pipeline.
BibTeX
@inproceedings{xiao2025retracted,
  author = {Cong Xiao and Yiyuan Zhou and Y. Wu},
  title = {{RETRACTED: Machine vision model for drip leakage detection of pipeline}},
  booktitle = {PLoS},
  year = {2025},
  doi = {10.1371/journal.pone.0316951},
}

Loading related content...

Figure