Introduction

Large-diameter gravity aqueducts serve as critical infrastructure for urban water supply, industrial transport, and irrigation systems, facilitating the efficient conveyance of substantial water volumes. However, these aqueducts experience complex and dynamic flow conditions, making them susceptible to hydraulic instabilities that can compromise operational performance and structural integrity1. In particular, unsteady flow phenomena such as air pocket effects, vortex formation, and water hammer may pose risks to long-term reliability.

Experimental studies have demonstrated that air-water interactions during filling and draining processes can induce water hammer effects, generating pressure fluctuations that threaten the structural stability of aqueducts2. Additionally, surface roughness within the aqueducts and the presence of solid particles can increase turbulence, degrading hydraulic efficiency and accelerating material wear under high-flow conditions3. Air pockets and cavitation, which frequently occur in large-diameter aqueducts, further destabilize flow, induce vibrations and noise, and exacerbate material fatigue and erosion4.

Given these challenges, continuous monitoring of transient hydraulic behaviors and flow pattern transitions is essential for ensuring the safety, reliability, and efficiency of water supply systems. Real-time identification of flow states is particularly critical for preventing infrastructure failures, optimizing operational efficiency, and reducing maintenance costs. Moreover, advanced monitoring techniques can provide essential data for the development of intelligent water management systems, promoting automated decision-making and predictive maintenance strategies.

Various flow monitoring techniques have been developed to address these challenges. Closed-circuit television (CCTV) inspection is widely used for pipeline monitoring, with cameras deployed either at manhole access points or mounted on robotic platforms for internal inspections5. While CCTV enables direct visual assessment, it is limited by range constraints and reduced effectiveness under high water levels6. Beyond CCTV, several indirect flow measurement techniques have been explored, including optical tomography3,4, ultrasonic sensing1,2, microwave tomography (MWT)7,8, electrical resistance tomography9,10, and electrical capacitance tomography (ECT)11,12. Supplementary Table 1 in Supplementary Note 1 provides a comparative summary of these techniques. While each method offers unique advantages in specific applications, their inability to provide continuous and real-time flow-state monitoring over large-scale pipeline networks highlights the need for more advanced sensing approaches.

DAS is an emerging fiber optic sensing technology that enables continuous, high-resolution acoustic monitoring along the entire length of an optical fiber13. Unlike conventional point-based sensors, DAS provides real-time, distributed detection of acoustic signals at any location along the fiber, making it a promising solution for large-scale flow monitoring applications. DAS technology operates by detecting Rayleigh backscattering variations in the optical fiber, allowing it to monitor acoustic wave propagation along an aqueduct’s structure14,15,16. While DAS has demonstrated remarkable success in pipeline leakage detection17, its application in flow-state monitoring within water conveyance aqueducts remains largely unexplored. The complex interaction between water flow, entrapped air, and aqueduct surface interactions generates acoustic signatures that propagate along the conduit, affecting the laser phase signals within the fiber. DAS technology, capable of detecting vibrations over distances of up to 40 km, offers the potential for precise localization and characterization of water flow dynamics in subterranean aqueducts. This ability to continuously monitor hydraulic conditions in real time represents an advancement over traditional methods.

However, despite the exceptional sensing capabilities of DAS technology, the interpretation of its extensive data remains a major challenge. Artificial intelligence (AI)-driven approaches, particularly deep learning algorithms, have shown substantial promise in deciphering complex acoustic patterns from DAS measurements. By training AI models on categorized DAS data, these methods enable accurate flow-state classification, facilitating the precise determination of water location, velocity, and phase transitions within underground aqueducts.

This study presents a real-time, distributed approach for monitoring flow dynamics in deep-buried, long-distance aqueducts, a type of large-scale pipeline system that has received limited attention in previous DAS applications. By combining distributed acoustic sensing (DAS) with a tailored machine learning classification framework, the method addresses the limited spatial coverage of traditional point-based flow monitoring and improves the identification of complex flow conditions, allowing for continuous and accurate detection of flow states in subsurface hydraulic environments. Compared with earlier studies, this work advances both the monitoring scope and analytical methods. A hierarchical learning model is developed to identify low-frequency acoustic features associated with flow regimes during the water-filling process. In addition, Shapley Additive exPlanations (SHAP) analysis is introduced to quantify the influence of individual frequency bands on model predictions, enhancing the interpretability and transparency of the classification results. The study further shows that low-frequency DAS signals can reliably reflect changes in flow regimes, providing new insights into hydraulic transitions. The method has been validated on a 6 km section of the Pearl River Delta water supply project, demonstrating its potential for broader application in intelligent aqueduct monitoring.

Methods

Case study area and data collection

The 113.1-km Pearl River Delta Water Resources Allocation Project is a key infrastructure initiative in South China’s water management strategy, designed to address critical water distribution challenges in the Guangdong–Hong Kong–Macao Greater Bay Area. This large-scale hydraulic project diverts 1.7 billion cubic meters of water annually from the Xijiang River system to the eastern regions of the Pearl River Delta, significantly enhancing water supply security and improving emergency water transmission capabilities for major cities, including Shenzhen, Guangzhou, and Hong Kong.

To minimize disruption to urban development and subterranean infrastructure, the project employs deep-burial shield construction technology, situating aqueducts 40 to 60 meters underground. During operation, continuous monitoring of water flow dynamics is essential for maintaining infrastructure integrity and optimizing water resource management. The aqueduct filling process follows a full-course gravity flow method, gradually introducing water with a small flow rate from upstream to downstream. Automatic air release valves were installed at high-elevation points along the pipeline to discharge trapped air and ensure stable hydraulic operation.

In terms of flow rate, the system is designed to operate within a range of 20–80 m3/s, depending on the number and configuration of downstream delivery zones. Under restricted or partial supply conditions, the typical flowrate ranges from 20 to 40 m3/s. Importantly, the system undergoes an annual centralized maintenance period of ~1 month, during which all pump stations are shut down and the pipeline is completely emptied for inspection, flushing, and structural maintenance. As a result, large portions of the pipeline remain dry during this period and must be refilled before resuming operation.

To evaluate the feasibility of DAS-based flow monitoring in large-scale, long-distance, and deep-buried aqueducts, multiple field monitoring campaigns were conducted across different segments and time periods. This study focuses on data collected from a 6-kilometer aqueduct segment, recorded on December 2, 2023, to validate the proposed method. Figure 1 illustrates the layout of this monitored section. The selected aqueduct has an internal diameter of 2.9 meters and is encased in self-compacting concrete, enclosing a steel pipeline containing two 36-core optical cables. The DAS system utilizes single-mode fibers within these cables for monitoring. A DAS interrogator is installed at a monitoring station above the maintenance shaft, where it connects to the aqueduct fibers via single-mode optical fiber.

Fig. 1: Layout of the monitored aqueduct.
figure 1

a Map showing the entire route of the Pearl River Delta Water Resources Allocation Project. The monitored test site is highlighted in red. Basemap ©Mapbox, ©OpenStreetMap. b Enlarged view of the monitored test site and DAS monitoring area over Google satellite imagery. Imagery ©2025 Airbus, Imagery ©2025 Airbus, CNES / Airbus, Landsat/Copernicus, Maxar Technologies, Map data ©2025.

Full size image

For this study, flow monitoring was conducted using the MS-DAS2000 system (Wuhan Optical Valley Interlink Technology Co., Ltd, Wuhan, China), configured with a channel spacing of 5 meters. The system sampled vibrational data at 1000 Hz, with 1200 sensing points covering the entire 6-km aqueduct section. This dataset, while part of a broader monitoring initiative, serves as the primary basis for validating the DAS-based flow monitoring approach.

DAS-Hydro Hierarchynet: hydraulic flow state monitoring in aqueducts

When analyzing the acoustic characteristics of water flow within transmission mains, it is critical to recognize that flow-induced variations in sound waves often lead to power amplification within specific frequency ranges. This phenomenon arises because dynamic changes in flow velocity and direction alter the propagation characteristics of sound waves, thereby altering the spectral distribution of the signal18. To systematically capture and quantify these effects, this study employs the power spectral density (PSD) method to transform time-domain signals into the frequency domain for refined analysis19. By examining the PSD of DAS data across different time intervals and sensing channels, the spatiotemporal migration of water flow within the aqueduct can be effectively characterized.

Building upon this, we propose a hierarchical clustering model, termed DAS-Hydro HierarchyNet, designed for high-precision classification of water flow states. The analytical pipeline begins with computing the PSD of DAS signals, followed by a two-stage classification process: first, DAS-Hydro HierarchyNet detects the presence of water flow; if flow is identified, the model further discerns specific flow regimes. This systematic approach enables comprehensive, real-time monitoring of hydraulic dynamics within large-scale conveyance infrastructures. The complete data processing workflow is illustrated in Fig. 2.

Fig. 2: Data processing workflow of DAS-Hydro HierarchyNet.
figure 2

a Schematic of data acquisition and preprocessing. A distributed acoustic sensing (DAS) system was used to monitor five flow states: dispersed bubble flow, plug flow, stratified wavy flow, stratified smooth flow and dry. The raw signals were processed using a Hann window and a bandpass filter (0–15 Hz), and power spectral density (PSD) features were computed to construct the input dataset. b Schematic architecture of DAS-Hydro HierarchyNet for multi-class classification of flow states across monitored sections.

Full size image

DAS-Hydro HierarchyNet is a multi-layer one-dimensional convolutional neural network (CNN) integrated with attention mechanisms. The network architecture, as illustrated in Fig. 2b, consists of nine convolutional layers, five enhanced channel spatial attention (ECSA) modules, three max-pooling layers, and two fully connected layers. These components are specifically designed to efficiently extract and learn subtle, intricate features within flow states, such as dispersed bubble flow, plug flow, stratified wavy flow, and stratified smooth flow20.

To enhance the network’s performance, the study employs an Inception module comprising four parallel pathways (Fig. 3a). Conventional convolutional layers typically utilize fixed kernel sizes, which limit the model’s ability to capture features at varying scales. Simply stacking additional layers to increase model capacity may lead to issues such as gradient explosion or vanishing, while increasing computational costs. In contrast, the Inception module concurrently uses multiple kernel sizes and pooling operations in the same layer, merging their outputs along the channel dimension21. This design enables multi-scale feature extraction from DAS flow data, improving both model adaptability and predictive accuracy.

Fig. 3: Key structural modules of DAS-Hydro HierarchyNet.
figure 3

a Inception module, b Residual connection, c Enhanced channel spatial attention module.

Full size image

Furthermore, the model integrates residual connections, a technique introduced in ResNet22, to enhance learning capacity and facilitate the training of deeper networks. Unlike traditional methods that apply pooling layers after every convolutional layer, this study incorporates residual connections after each group of convolutional layers (Fig. 3b). This prevents the depth limitations imposed by excessive pooling layers, allowing the model to retain richer hierarchical features and improve classification performance.

Attention mechanisms are crucial in improving the performance of CNN-based models by dynamically prioritizing relevant features. In traditional CNNs, all features are treated equally, which can lead to the extraction of irrelevant information and increased computational overhead. Attention mechanisms, in contrast, focus computational resources on critical features, enhancing both classification accuracy and model efficiency.

A widely adopted attention module, the convolutional block attention module (CBAM)23, combines channel attention, which identifies “what” is important in the input, and spatial attention, which determines “where” critical features are located. However, CBAM’s channel attention mechanism relies on max-pooling and average-pooling operations, which introduce computational redundancy and limit the learning of inter-channel relationships. To address these shortcomings, the Enhanced Channel Attention (ECA) module was proposed24, which utilizes a one-dimensional convolution with an adaptive kernel size, improving cross-channel interactions without dimensionality reduction.

This study implements the ECSA module, which is specifically tailored for one-dimensional DAS data processing. The ECSA framework, illustrated in Fig. 3c, replaces CBAM’s channel attention module with ECA, improving feature selection efficiency. The module first applies average pooling to the feature map to extract a global vector, which is then processed using an adaptive convolutional kernel to compute channel attention weights. These weights are applied to the original feature map, generating an enhanced representation. The enhanced feature map is subsequently processed by a spatial attention module, which extracts spatial features using both max-pooling and average-pooling operations before generating spatial attention weights. These weights are then applied to the feature map, optimizing its feature representation.

By combining ECA and spatial attention, the ECSA module effectively highlights critical flow-related features while suppressing irrelevant information. This mechanism enhances the model’s ability to detect the presence or absence of water flow and accurately classify flow states.

To further optimize feature extraction and model stability, batch normalization (BN) layers are integrated to reduce internal covariate shift, improving convergence rates during training. Additionally, a hyperbolic tangent (Tanh) activation function is employed to introduce non-linearity, allowing the network to capture complex relationships within DAS data. A detailed description of the model architecture, including convolutional kernel sizes, input/output dimensions, and padding schemes, is provided in Supplementary Table 2 in Supplementary Note 2.

Dataset establishment

The stability of flow regimes and their associated transitions has long been a critical area of research in fluid mechanics and multiphase flow studies25. In horizontal or near-horizontal conduits, stratified flow is commonly observed and can be further categorized into sub-regimes (e.g., stratified wavy flow) based on qualitative and quantitative analyses26,27. To describe transitions from smooth stratified flow to wavy flow, various theoretical models have been proposed, contributing to a deeper understanding of flow regime stability and evolution.

For this study, dataset establishment focused on a 6-kilometer segment of the aqueduct, selected because it captures the complete transition from a dry state to active water flow. This section provides a comprehensive dataset representing changing flow characteristics along the conveyance system. Based on the research framework and identified acoustic signal features, the flow was classified into five distinct regimes: dispersed bubble flow, plug flow, stratified wavy flow, stratified smooth flow, and dry (Fig. 4). This classification aligns with well-established multiphase flow frameworks for horizontal or near-horizontal conduits, supported by theoretical and experimental investigations into flow regime stability and transitions.

Fig. 4: Acoustic characteristics during aqueduct water filling process.
figure 4

a PSD patterns during the aqueduct water filling process, b schematic diagram of the aqueduct water filling process, c PSD curve diagrams of the four states of the water filling process.

Full size image

All selected DAS-acquired data were subjected to PSD analysis, extracting frequency-domain features and energy distribution patterns from acoustic signals. This approach facilitates a detailed characterization of the spectral properties of different flow regimes, providing a robust foundation for subsequent model training and classification.

$$PSD( \, f)=\frac{1}{KNL} \sum\limits_{k=0}^{K-1}{\left| \sum\limits_{n=0}^{N-1}x(n+kD)\times w(n)\times {e}^{-i2\pi fn/N}\right|}^{2}$$
(1)

where \(x(n)\) represents the value of the signal at time point \(n\)；\(w(n)\) is the value of the window function;\(N\) is the length of each segment；\(K\) is the length of each segment; \(D\) is the number of overlapping data points between segments；\(L\) is the sampling rate；\(f\) is the frequency index.

The PSD values used for model training were extracted from DAS data within the 0–15 Hz frequency range, computed using Welch’s method. The rationale for selecting this frequency band is provided in Supplementary Note 3. Each training sample was derived from a 2-min data segment, ensuring sufficient spectral resolution for flow state classification. The PSD computation parameters included a Hann window, a segment length of 10,240 data points, an overlap of 5120 data points, and linear detrending to mitigate low-frequency bias. Each training sample comprised 154 PSD points, ensuring a detailed spectral representation of flow-induced acoustic variations.

Model training

To address the classification problem of water flow states, this study employs a binary cross-entropy loss function combined with logistic regression, incorporating the Sigmoid activation function to map predictions to probabilities. The mathematical expression for the loss function is defined as:

$${{{\rm{loss}}}}({{{\rm{x}}}},{{{\rm{class}}}}) = - \,\log \left(\frac{\exp (x[{{{\rm{class}}}}])}{{\sum }_{i=1}^{C}\exp (x[i])}\right) \\ = - x[{{{\rm{class}}}}]+\,\log \left(\sum\limits_{i=1}^{C}\exp (x[i])\right)$$
(2)

where \(x\) is a vector of unnormalized outputs from the model, with \(x(i)\) indicating the unnormalized score for the i-th class; class is the index of the true class; \(C\) is the total number of classes.

The Adam optimizer is employed to dynamically adjust the learning rate, starting from an initial value of 0.001. As an extension of momentum-based gradient descent, Adam combines momentum and adaptive learning rate strategies, improving both convergence speed and stability. Specifically, it maintains exponentially decaying averages of the first moment (mean of gradients) and the second moment (uncentered variance of gradients). The first moment estimate introduces momentum, effectively dampening oscillations and smoothing updates to enhance training stability. Meanwhile, the second moment estimate scales the learning rate based on gradient magnitudes, preventing extreme parameter updates and mitigating instability during training. This adaptive adjustment mechanism allows the model to converge efficiently and reliably, ultimately enhancing its ability to distinguish water flow states with high accuracy. A complete description of data annotation, model training, and validation procedures is provided in Supplementary Note 4.

Results

Acoustic characterization of waterflow dynamics in subterranean aqueducts

By implementing DAS monitoring in a water conveyance aqueduct, distinct acoustic signatures associated with different flow states are captured, revealing unique spectral characteristics in PSD analysis. Building on previous findings regarding sound propagation in water-filled aqueducts28, it is established that larger-diameter and more flexible aqueducts tend to attenuate higher-frequency sound waves. Our observations confirm that as water progressively fills the aqueduct, acoustic variations corresponding to different flow patterns emerge, as illustrated in Fig. 4.

Figure 4a presents the distance–frequency spectrogram, depicting the PSD distribution at different monitored positions during the water-filling process. The spectrogram is segmented into four distinct zones, with color transitions from red to blue, indicating a gradual reduction in PSD levels as flow conditions evolve. As shown in Fig. 4b, the water flow states during aqueduct filling are classified into dispersed bubble flow, plug flow, stratified wavy flow, and stratified smooth flow, each exhibiting distinct acoustic characteristics. These flow states are influenced by bubble collapse, turbulence, and water interactions with aqueduct walls29, generating unique vibrational signals. Figure 4c further illustrates the PSD diagrams corresponding to each flow state, capturing the acoustic waves generated by turbulence and bubble bursting at different stages of the water flow process. These results highlight the effectiveness of DAS-based spectral analysis in identifying and classifying flow regimes, reinforcing the utility of acoustic sensing for real-time hydraulic monitoring in subterranean aqueducts.

Performance evaluation of the DAS-Hydro HierarchyNet in dynamic waterflow monitoring

A comprehensive evaluation of the DAS-Hydro HierarchyNet’s performance was conducted over the entire monitoring period, with key insights depicted in Fig. 5, which showcases the model’s capability to accurately detect and classify various flow states and to track the dynamic progression of water within the aqueduct. As the aqueduct fills, an intensification of acoustic signals is observed, elucidating flow migration patterns at successive stages.

Fig. 5: Model-based water flow monitoring and PSD analysis.
figure 5

a–d Model-based water flow monitoring and recognition at four distinct time points: 13:00, 14:30, 16:00, and 17:19. Each panel combines a strip heatmap and a line graph to depict flow state distribution along the aqueduct’s length. The strip heatmap employs a color gradient to represent transitions between flow states: “dispersed bubble flow,” “plug flow,” “stratified wavy flow,” “stratified smooth flow,” and “dry.” The vertical axis of the line graph indicates the model-predicted probability of water presence, with the gray curve showing raw predictions and the orange curve presenting a smoothed version obtained by averaging adjacent predictions. A probability threshold of 50% (dashed line) delineates regions of active water flow, with the estimated water head position marked by a red circle; e–h Corresponding PSD cloud diagrams illustrating the spatial variation in signal energy as the aqueduct fills with water; i Model-estimated water head migration path (blue line) and velocity (read line) throughout the monitoring period.

Full size image

At 13:00 (Fig. 5a), field measurements from pressure transducers indicate that water has reached 3000 m within the aqueduct. The model estimates the water head position at 2940 m, resulting in a minimal discrepancy of 60 m (2%). During this phase, stratified smooth flow predominates, with localized stratified wavy flow near the inlet. However, the PSD signals (Fig. 5e, f) remain relatively weak, making it challenging to discern the water front based solely on spectral data. By 14:30 (Fig. 5b), the water advances further into the aqueduct, expanding the region of stratified smooth flow, while stratified wavy flow persists near the inlet. At 16:00 (Fig. 5c), the water reaches the terminal end of the aqueduct. In addition to the existing stratified flows, dispersed bubble flow and plug flow emerge, creating distinct flow boundary transitions. Correspondingly, the PSD signal (Fig. 5g) intensifies significantly, facilitating precise identification of the advancing flow front. By 17:19 (Fig. 5d), the aqueduct is nearly fully filled, with all four flow regimes coexisting and exhibiting spatial transitions likely resulting from turbulence and regime shifts within the structure. Figure 5i illustrates the model’s estimated water head positions throughout the filling process, along with the corresponding water head velocities derived from consecutive position estimates. For a more detailed and continuous depiction of flow state distribution and water head identification over the entire monitoring period, see Supplementary Note 5. The observed water migration pattern aligns with expected dynamics of the filling process, underscoring the robustness of DAS-Hydro HierarchyNet in handling spatiotemporal variations and accurately classifying flow states.

Interpretability analysis of DAS-Hydro HierarchyNet for waterflow classification

DAS-Hydro HierarchyNet exhibits outstanding performance in water flow regime detection. However, as an inherently black-box model, it is difficult to explicitly identify the key features driving its predictions. Nevertheless, the model implicitly captures critical spectral characteristics associated with different flow states. A deeper understanding of these features is essential for elucidating the spectral dynamics of flow regime transitions during aqueduct filling and bridging the gap between deep learning-based predictions and the underlying physical mechanisms governing flow behavior.

To address this, we employ SHAP, a powerful method for interpreting machine learning model predictions by quantifying the contribution of each feature to the decision-making process. By computing Shapley values, SHAP provides a systematic framework for understanding the role of different input variables in model outcomes30. In this study, we employ SHAP analysis to comprehensively investigate the DAS-Hydro HierarchyNet, elucidating the influence of distinct frequency components in DAS signals on water flow detection and flow regime classification.

To enhance interpretability, we visualize SHAP results using bar plots and swarm plots. The bar plots depict the mean Shapley values for each frequency band, ranking them in descending order to highlight their global importance (Fig. 6). Meanwhile, swarm plots illustrate the distribution of Shapley values across individual samples, where color intensity represents feature magnitude, and the position of each point indicates the feature’s contribution—positive values support the target class prediction, whereas negative values suggest a greater likelihood of an alternative classification.

Fig. 6: SHAP-based interpretability results for water detection and flow regime classification.
figure 6

a SHAP summary plot for water presence. b SHAP summary plot for stratified smooth flow. c SHAP summary plot for stratified wavy flow. d SHAP summary plot for plug flow. e SHAP summary plot for dispersed bubble flow.

Full size image

SHAP analysis enables both global and local interpretability of DAS-based flow state classification. At the global scale, it quantifies the contribution of specific frequency bands to model predictions, while at the local scale, it provides insights into sample-specific decision weight distributions. Our findings indicate that the presence of water is primarily characterized by the 7–8 Hz and 6–7 Hz frequency bands, which are strongly linked to the turbulent acoustic signals generated by water flow within aqueducts. In these bands, samples with high Shapley values (red points) are concentrated in the positive region, indicating their decisive role in classifying the “water present” state. Conversely, high-frequency bands (e.g., 10–15 Hz) contribute less to the overall prediction but still provide auxiliary information for certain individual cases.

For flow regime classification, DAS-Hydro HierarchyNet exhibits distinct frequency dependencies across various flow patterns, aligning with established hydrodynamic principles. Stratified Smooth Flow is primarily governed by the 9–10 Hz band, while the 7–9 Hz bands exhibit negative contributions, highlighting the contrasting spectral features between different stratified flow types. Stratified Wavy Flow is predominantly influenced by the 2–3 Hz band, emphasizing the significance of low-frequency components in capturing wave-induced oscillations. Plug Flow is closely associated with the 6–7 Hz band, likely reflecting low-frequency vibrations generated during plug formation. In contrast, Dispersed Bubble Flow relies heavily on the 13–14 Hz band, underlining the importance of higher-frequency signals in characterizing complex gas-liquid interactions. Notably, prior studies have identified dominant frequency ranges of 2–10 Hz in stratified wavy flows31, further validating the physical consistency and reliability of the SHAP-based analysis in this study. These findings confirm the effectiveness of the proposed approach in describing and interpreting the spectral characteristics of different flow states.

Discussion

PSD signatures and flow regime interpretation based on DAS monitoring

A detailed analysis of the PSD derived from DAS monitoring of the 6-km aqueduct section during the water-filling process reveals distinct and repeatable spectral patterns that are temporally aligned with different stages of flow evolution. Although direct visual observations are not available due to the buried depth of the system (40–60 m), the interpretation of these patterns is grounded in known two-phase flow dynamics and acoustic propagation principles, supported by the temporal and spatial consistency of the DAS signals.

In the early stage of filling, the system exhibits stratified smooth flow, characterized by a low water level gradually climbing along the pipe wall and forming a nearly planar gas–liquid interface. The DAS signal during this phase displays a single dominant peak in the 4–9 Hz range, which is interpreted as resulting from low-frequency vibrations induced by the advancing water front and mild shear at the water–pipe boundary. This spectral feature aligns with the gravity-dominated, weakly turbulent behavior of stratified smooth flow reported in the literature, where interfacial shear stress is low and velocity profiles remain stable32,33.

As the water level continues to rise, the flow transitions into the stratified wavy regime. The gas–liquid interface becomes increasingly unstable and exhibits noticeable oscillations. The corresponding DAS spectra show enhanced energy in the 2–4 Hz and 10–15 Hz bands, which likely reflect interfacial instabilities. Although these specific frequencies have not been directly validated in large-scale systems, prior laboratory studies have documented similar dynamics in stratified wavy flows, including periodic surface undulations, roll wave propagation, and shear layer disturbance amplification34,35, providing useful physical context for interpreting our observations.

As the aqueduct becomes further filled and large gas slugs begin to propagate intermittently, the system enters the plug flow regime. During this stage, the DAS signal exhibits a broadband spectral profile, with an overall energy level of approximately –50 to –40 dB—higher than in the stratified wavy flow phase. This spectral variation is interpreted as an indication of increased turbulence and acoustic activity, reflecting a transition from intermittent, gas-dominated transport to more intense interfacial disturbances. Although short-duration shear or shock events may still occur, the overall signal energy remains relatively elevated, consistent with the quasi-discontinuous hydraulic characteristics of plug flow. These observations align with existing descriptions of plug flow behavior and energy distribution in the literature36.

In the final filling stage, the system transitions into the dispersed bubble flow regime, where gas bubbles are randomly distributed in the upper water column. The DAS signal at this point exhibits a broadband, high-amplitude PSD with no dominant frequency, typically around –30 to –40 dB in energy. This spectral profile is commonly attributed to random acoustic scattering caused by unstable gas–liquid interfaces and the irregular motion of bubbles in space and time37. Additionally, we note that this stage typically corresponds to the period of greatest water volume and highest hydraulic continuity. The increase in overall water mass and sustained flow may further contribute to broader frequency excitation and elevated acoustic energy. Thus, the spectral features observed during this phase likely result from a combination of micro-scale bubble disturbances and macro-scale hydrodynamic enhancement, leading to a more complex evolution than typically observed in small-scale experiments.

These findings establish a physics-based interpretive framework for analyzing the PSD characteristics associated with different flow regimes, enabling continuous and full-scale observation of the tunnel flow field. This allows for the identification of localized flow states as well as real-time tracking of the spatiotemporal evolution of flow regimes. In the filling process of large-diameter tunnels, flow patterns such as stratified flow, bubble flow, plug flow, and dry segments are common and exhibit typical hydraulic features. The identification of these regimes using DAS helps establish baseline patterns of system operation, which in turn supports the detection of potential anomalies—such as persistent dry zones, flow interruptions, or localized disturbances—thus enhancing the interpretability and early-warning capability of system monitoring.

For example, long-lasting dry segments may indicate delayed upstream inflow, resulting in slower system filling. Frequent or sustained plug flow may reflect insufficient air exhausting capacity, such as air valve malfunction38; additionally, transitions from plug to slug flow can induce unsteady pressure fluctuations and pipeline vibrations, potentially leading to structural fatigue or long-term integrity risks39. Irregular transitions between flow regimes in space or time may indicate geometric irregularities, unstable hydraulic inputs, or other structural anomalies40. Therefore, the spatiotemporal features derived from DAS observations provide key insights into abnormal dynamic changes in water conveyance systems and support the safe monitoring of long-distance tunnel filling processes.

In future applications, traditional point-based sensing techniques such as ultrasonic sensors, pressure transducers, and microwave imaging can serve as valuable complements to the DAS system. These methods offer high spatial resolution at specific locations and are capable of capturing key hydraulic parameters such as water front position, flow state, and local pressure fluctuations. They are particularly suitable for validating features extracted from DAS data and supporting the physical interpretation of flow dynamics. When selectively deployed at critical sites—such as inlets, outlet structures, or known risk zones—they can enhance the overall reliability and responsiveness of the monitoring system.

At the same time, DAS provides continuous coverage and strong sensitivity to dynamic processes along the entire pipeline. It enables real-time tracking of water front propagation, flow regime transitions, and system disturbances across both spatial and temporal dimensions. Integrating these two types of sensing technologies can leverage the large-scale coverage of DAS and the localized accuracy of point-based sensors, forming a more robust and comprehensive multi-source monitoring framework. This hybrid architecture not only improves resilience to environmental noise interference but also enhances the model’s ability to capture and interpret underlying hydraulic mechanisms.

Capabilities, interpretability, and extension pathways of DAS-Hydro HierarchyNet

A key challenge in flow monitoring arises during the early filling phase, where distinguishing between water-bearing and dry sections based solely on PSD variations proves inefficient for water-front detection, primarily due to a low signal-to-noise ratio (SNR). To address this limitation, the proposed DAS-Hydro HierarchyNet employs a holistic analysis of acoustic patterns, enabling precise estimation of water states and locations. The framework demonstrates particular effectiveness in detecting subtle boundary transitions and identifying the water head under weak acoustic signals. By leveraging continuous spatiotemporal acoustic data, it captures both macro-scale flow migration patterns and micro-scale variations across different flow regimes, offering a comprehensive understanding of filling dynamics.

The interpretability analysis of DAS-Hydro HierarchyNet using SHAP provides valuable insights into the factors influencing classification decisions for different flow regimes. The results demonstrate that low-frequency components primarily govern water detection and stratified flow states, while higher-frequency bands capture the acoustic signatures associated with dispersed bubble flow. By correlating these frequency-domain features with well-established fluid-mechanical principles, the framework enhances the transparency of model predictions, mitigating the “black-box” nature of deep learning and reinforcing its consistency with hydraulic physics. More importantly, this frequency–flow regime mapping has practical implications: now that the characteristic frequency bands of each flow state are known, engineers can use this knowledge during ongoing monitoring to infer the likely flow condition from spectral content alone—even without relying on full model inference. This enhances situational awareness and provides a valuable basis for informed decision-making in system operation and maintenance. Rather than serving as a direct anomaly detection tool, SHAP enables us to understand what types of spectral features the model associates with typical versus abnormal flow patterns—thereby helping engineers contextualize unexpected spectral signatures as possible signs of delay, blockage, or air entrapment. This integration of DAS sensing, deep-learning-based classification, and SHAP-based interpretability establishes a robust framework for the monitoring, analysis, and optimization of large-scale water transmission systems.

Although the proposed framework demonstrates strong performance in identifying flow states, it relies heavily on spectral features extracted from DAS measurements. In scenarios where power spectral densities overlap among different flow regimes or where external acoustic noise is present, this dependence may lead to misclassification. In this study, robustness under such low-SNR conditions was partially addressed by conservatively selecting training segments with clear spectral and temporal characteristics, thereby reducing ambiguity during early-phase detection. While no explicit denoising algorithms were applied, we acknowledge that incorporating signal enhancement techniques—such as adaptive filtering, empirical mode decomposition (EMD), or learning-based denoising models—could provide further improvements in signal clarity and classification reliability, particularly in challenging environments.

In addition, it is important to recognize that the generalizability of this method may be constrained by differences in aqueduct geometry, hydraulic conditions, and acoustic propagation characteristics across systems. Although the DAS-Hydro HierarchyNet architecture is structurally transferable and its classification logic remains theoretically applicable, deploying it in a new system requires the collection of system-specific DAS data, manual annotation of representative flow segments, and subsequent model retraining. In data-scarce scenarios, future work may consider incorporating generative models—such as generative adversarial networks—to augment labeled datasets and enhance the model’s adaptability.

At the same time, incorporating physically complementary sensing strategies can enhance the robustness and adaptability of DAS in complex environments. For instance, pressure disturbances caused by water front advancement or flow transitions can be captured by distributed pressure sensors, offering reference signals that are immune to acoustic interference. When a significant thermal contrast exists between the water and the surrounding tunnel environment, localized temperature variations induced by water movement can be detected using distributed temperature sensing systems to indicate water presence. In addition, changes in electromagnetic properties, such as permittivity shifts due to water intrusion, may serve as indirect indicators of water distribution and can be detected via MWT or capacitive sensing techniques.

These methods hold notable potential for localized integration, system-level redundancy, and the development of multimodal sensing frameworks. Furthermore, coupling DAS data with physics-based models or computational fluid dynamics simulations may substantially enhance the physical interpretability of acoustic signals by establishing quantitative relationships between spectral features and specific flow phenomena—such as turbulence intensity, bubble interactions, and stratified flow behavior. This approach can not only improve the accuracy of flow regime classification but also deepen the understanding of system dynamics from a physical perspective.

Conclusions

This study introduces DAS-Hydro HierarchyNet, a novel framework that integrates DAS with deep learning to enable real-time hydraulic state monitoring in large-diameter subterranean aqueducts. By fully leveraging DAS’s extensive spatial coverage and high sensitivity to subtle acoustic variations, this approach addresses the limitations of conventional flow-state detection methods, which rely on sparse point measurements and low-resolution sampling. The framework enables continuous, high-fidelity monitoring of both macroscale water migration and microscale flow dynamics, improving the accuracy and efficiency of hydraulic state analysis. Its successful deployment in the Pearl River Delta Water Resources Allocation Project further validates its feasibility and effectiveness in ensuring the operational security of large-scale water transmission systems.

This study advances the integration of DAS and deep learning in fluid mechanics and water resource engineering, establishing a new paradigm for hydraulic monitoring. By effectively capturing complex acoustic signals from multiphase hydrodynamic processes, the framework enables precise flow-state delineation, even under low-SNR conditions. Additionally, the incorporation of SHAP-based interpretability enhances model transparency by identifying key frequency components that characterize different flow regimes, reinforcing its alignment with hydraulic principles. These contributions position DAS-Hydro HierarchyNet as a robust and scalable solution for real-time aqueduct monitoring.

By demonstrating the feasibility of DAS technology in subterranean aqueduct surveillance, this study contributes to the advancement of intelligent monitoring for large-scale water infrastructure. These findings provide new insights and technological innovations, offering a foundation for future research and practical implementations in water resource management and infrastructure resilience.