Internal pipe corrosion significantly impacts water quality in distribution systems. Cast iron pipes, commonly used worldwide, are susceptible to aging and corrosion, leading to the formation of porous iron oxide deposits. These deposits release iron into the water, creating "red water" and providing environments for harmful bacteria. This contamination can cause gastrointestinal infections, dermatological problems, and lymph node complications. Current methods for corrosion assessment, such as chemical analysis, coupon testing, and visual inspection, are destructive and disruptive. Therefore, continuous and non-destructive testing (NDT) methods are crucial for maintaining water quality and system integrity. This study addresses this need by developing a novel dual-mode approach that combines advanced ultrasound technology with convolutional neural networks (CNNs) to provide a comprehensive and non-destructive assessment of both pipe corrosion and water contamination.

Existing NDT methods for pipe defect detection include magnetic flux leakage (MFL) inspection and eddy current (EC) techniques. While MFL is efficient for ferromagnetic materials, quantifying defect size requires significant effort. EC, suitable for conductive materials, suffers from sensitivity to the lift-off effect. Ultrasound methods offer non-destructive and continuous measurement advantages. Array transducers are commonly used, but high-frequency transducers (above 15 MHz) are difficult to manufacture. This study utilizes scanning acoustic microscopy (SAM), which employs a single high-frequency transducer, overcoming the limitations of array transducers and enabling higher resolution measurements. Previous studies have demonstrated the effectiveness of CNNs in extracting physical properties from acoustic reflection signals, including cell properties. This study extends this approach to analyze acoustic signals from pipes to determine iron oxide concentration in the water.

The researchers developed a dual-mode system for assessing pipe corrosion. A syringe pump maintained a constant water flow (containing iron oxide particles) at 0.3 m/s through brass pipes (simulating water distribution pipes with varying thicknesses). SAM, equipped with 20 MHz (for thickness measurement) and 5 MHz (for A-scan data collection) transducers, was used to collect data. The 20 MHz transducer facilitated B-mode analysis to measure pipe thickness and assess corrosion extent. The 5 MHz transducer collected A-scan data used for iron oxide concentration classification using CNNs. Finite element analysis (FEA) using Ansys was performed to simulate the relationship between pipe thickness and stress under internal pressure, demonstrating the importance of accurate thickness measurements. Six brass pipes with different initial thicknesses were used, and some were subjected to accelerated corrosion testing by immersion in nitric acid and oven heating. Four different CNN architectures (VGGNet, InceptionNet, ResNet, and EfficientNet) were trained and tested on 60,000 acoustic reflection signals (10,000 per each of six iron oxide concentration levels: 1, 5, 15, 20, 100, and 300 mg/L). Model performance was evaluated using accuracy, precision, recall, and F1-score. Short-time Fourier Transform (STFT) was applied to the A-scan data before feeding it to the CNNs for analysis of both time and frequency domain characteristics. The detailed architecture of each CNN model, including the use of convolutional layers, pooling layers, batch normalization, dropout, and activation functions, is described in detail. The objective function used for training was cross-entropy loss.

The SAM system accurately measured pipe thickness, with errors within 10%. The CNN models demonstrated high accuracy in classifying iron oxide concentration levels. ResNet achieved the highest average accuracy (0.9972), significantly outperforming VGGNet (0.9735), EfficientNet (0.9716), and InceptionNet (0.9423). All models exhibited F1-scores above 0.92. The study found that ResNet's superior performance stems from its ability to capture correlations between distant frequency components due to its deep architecture and large receptive fields. In contrast, InceptionNet, with its multi-scale analysis, showed lower accuracy, especially for low concentrations. VGGNet showed robustness against noisy outliers. The study demonstrated that sufficient receptive fields, often achieved with sufficient network depth, are crucial for extracting physical properties from acoustic reflection signals. Lower particle concentrations resulted in less discernible features in the acoustic reflection signals. The high accuracy of the ResNet model in classifying iron oxide concentration levels, even with small concentration differences (1, 5, 15 mg/L), validates the potential of this method for precise water quality measurement inside pipes.

This study successfully integrated SAM and CNNs for non-destructive, continuous assessment of both pipe corrosion and water quality. The high accuracy of the CNN models in classifying iron oxide concentration expands the application of SAM beyond traditional image analysis. The dual-mode approach enables simultaneous monitoring of pipe integrity and water contamination, providing crucial information for proactive maintenance and improved water management. The ability to accurately measure pipe thickness in real-time allows for integration with analytical equations or FEA for real-time stress estimation, potentially leading to predictive maintenance strategies for water distribution systems. The findings significantly advance the field of water infrastructure monitoring by providing a precise, non-invasive, and continuous method for assessing pipe corrosion and water quality.

This study presents a novel dual-mode approach using SAM and CNNs for the non-destructive assessment of pipe corrosion and water contamination in water distribution systems. The high accuracy achieved using ResNet demonstrates the feasibility and potential of this integrated system for real-time monitoring and predictive maintenance of water infrastructure. Future research could focus on validating the method using cast iron pipes and exploring its application in diverse water distribution networks and with different types of contaminants. Further development could involve integrating this system with real-time data analysis and predictive modeling tools to optimize maintenance scheduling and minimize disruptions to water supply.

The study used brass pipes for the experiments, while real-world water distribution systems typically employ cast iron pipes. Although brass and cast iron exhibit similar acoustic properties, future studies should validate the method using cast iron pipes to improve the generalizability of the findings. The corrosion simulations were accelerated, which might not perfectly replicate the natural corrosion process over extended periods. The relatively small number of concentration levels evaluated in this study could be expanded in future investigations to improve the robustness and accuracy of the model in a wider range of conditions.