A method for detecting and locating leaks in pressurized pipelines

Abstract

This invention discloses a method for detecting and locating leaks in pressurized pipelines, belonging to the field of pressurized pipeline leak detection and location technology. The method includes: acquiring high-frequency pressure signals during the operation of the pressurized pipeline using multiple sensors deployed on the pipeline under test; arranging all high-frequency pressure signals according to the acquisition sequence to obtain a high-frequency pressure signal sequence; performing a convolution operation on the high-frequency pressure signal sequence to obtain local features; using a convolutional neural network to compress the local features to obtain leakage features; using a gated recurrent unit network to dynamically weight the leakage features to obtain fused features; and locating the leak based on the fused features to obtain the leak detection and location result of the pressurized pipeline under test. This invention can accurately distinguish between leakage signals and background noise, significantly improving the detection and location capabilities of low signal-to-noise ratio signals under low-pressure, small-leakage conditions, and reducing the missed detection rate.

Description

Technical Field

[0001] This invention relates to the field of pressurized pipeline leak detection and location technology, and in particular to a method for pressurized pipeline leak detection and location. Background Technology

[0002] Pressurized pipelines, as an efficient and economical means of fluid transportation, are widely used in many fields such as petrochemicals, water conservancy and hydropower, and municipal engineering. With the continuous advancement of pipeline transportation technology, their role in urban infrastructure construction is becoming increasingly prominent. However, pipeline leakage problems seriously restrict the safe and stable operation of pipeline transportation systems. Specifically, due to the complex environment in which pipelines are located, they are constantly affected by factors such as geological subsidence, soil corrosion, and aging stress, making them highly susceptible to structural damage and subsequent leakage accidents. Once a leakage accident occurs, it not only leads to a large waste of resources and economic losses but may also cause serious environmental pollution and safety accidents. Therefore, developing efficient and reliable leak detection and location technologies is of great significance for ensuring the safe operation of pipeline transportation systems.

[0003] In recent years, with the rapid development of artificial intelligence technology, machine learning methods based on neural networks have gradually become an emerging direction in leak detection. Among them, deep learning, due to its powerful feature extraction and pattern recognition capabilities, excels in processing complex, high-dimensional data. Recurrent Neural Networks (RNNs), especially Long Short-Term Memory Networks (LSTM) and Gated Recurrent Unit Networks (GRUs), have a natural advantage in capturing dynamic dependencies in time series and are widely used in prediction tasks. However, RNN models are not as capable as CNNs in extracting primary features of signals in shallow structures. Therefore, hybrid models that combine CNNs and RNNs have emerged in recent years. CNNs perform efficient feature extraction, while RNNs model the temporal evolution, improving the overall detection accuracy. However, these hybrid models often still rely on traditional two-dimensional CNN architectures, facing problems such as computational redundancy and structural complexity.

[0004] While deep learning has made some progress in pipeline leak detection, it still faces technical bottlenecks under low-pressure and small-leakage conditions. In particular, when the leakage flow rate is less than 5% of the total transport volume (small leak), the leakage signal is difficult to distinguish from background noise, resulting in a high rate of missed detection and large location errors. Summary of the Invention

[0005] Therefore, it is necessary to provide a method for detecting and locating leaks in pressurized pipelines to address the aforementioned technical problems.

[0006] This invention provides a method for detecting and locating leaks in pressurized pipelines, comprising: By deploying multiple sensors on the pressurized pipeline under test, high-frequency pressure signals during the operation of the pressurized pipeline are collected, and all high-frequency pressure signals are arranged according to the acquisition time sequence to obtain a high-frequency pressure signal sequence. Convolution operations are performed on high-frequency pressure signal sequences to jointly model and fuse spatiotemporal features of all high-frequency pressure signals, thereby obtaining local features; A convolutional neural network is used to compress local features to determine the time delay and amplitude difference of the leakage signal during propagation, thereby obtaining the leakage characteristics. A gated recurrent unit network is used to dynamically weight the leakage features in order to capture the temporal dependencies of the leakage features and obtain fused features. Leakage location is determined based on fusion characteristics, yielding the leak detection and location results for the pressurized pipeline under test.

[0007] Optionally, the leak detection and localization results are obtained through a leak detection and localization model; the leak detection and localization model includes: a data processing module, a two-layer lightweight convolutional neural network, a two-layer gated recurrent unit network, a fully connected layer, and a Softmax layer connected in sequence; The data processing module includes: a one-dimensional convolutional layer, an activation function module, and a pooling layer connected in sequence; Each lightweight convolutional neural network layer includes: two Fire modules connected in sequence; each Fire module includes: a 1×1 convolutional kernel and a dimension expansion module connected in sequence, the dimension expansion module includes: a 1×1 convolutional branch and a 3×3 convolutional branch, the 1×1 convolutional branch includes multiple 1×1 convolutional kernels, and the 3×3 convolutional branch includes multiple 3×3 convolutional kernels; Each gated loop unit network consists of two gated loop units connected in sequence, and each gated loop unit includes a parallel update gate and a reset gate.

[0008] Optionally, a leak detection and localization model is trained, specifically including: Acquire high-frequency pressure signal data and corresponding leak detection and location results under single-point leakage conditions in pressurized pipelines; The high-frequency pressure signal data of all pressurized pipelines under single-point leakage conditions are standardized to obtain a high-frequency pressure signal dataset; The high-frequency pressure signal dataset is divided into training set, validation set and test set according to a set ratio; Input the training set into the leak detection and localization model to obtain the leak detection and localization prediction results; The cross-entropy loss function is determined based on the leak detection and location prediction results and the actual leak detection and location results. The hyperparameters of the leak detection and location model are updated through multiple iterations using the backpropagation algorithm. After each iteration, the model performance is evaluated using a validation set, and the hyperparameters of the leak detection and location model are adjusted. The iteration stops when the maximum number of iterations or the validation accuracy convergence condition is reached, resulting in the pre-trained leak detection and location model. The leak detection and localization model after initial training was independently validated using a test set, resulting in the final trained leak detection and localization model.

[0009] Optionally, a convolution operation is performed on the high-frequency pressure signal sequence to jointly model and fuse the spatiotemporal features of the high-frequency pressure signals from multiple sensors, obtaining local features, specifically including: A one-dimensional convolution operation is performed on the high-frequency pressure signal sequence to perform weighted operations on the local time windows of all high-frequency pressure signals, resulting in a multi-channel feature map. A nonlinear mapping is performed on the multi-channel feature map, and a one-dimensional average pooling operation is used to reduce the dimensionality and compress the feature map after the nonlinear mapping to obtain local features.

[0010] Optionally, a convolutional neural network is used to compress local features to determine the time delay and amplitude difference of the leakage signal during propagation, thereby obtaining leakage features, specifically including: The leakage features are obtained by performing multiple feature compressions on the local features; For each feature compression, a one-dimensional convolution operation is performed on the input feature to compress its dimension, resulting in the first intermediate feature. Then, one-dimensional and three-dimensional convolution operations are performed on the first intermediate feature to expand its dimension at different scales, resulting in the second and third intermediate features. Finally, the second and third intermediate features are concatenated and fused along the channel dimension to obtain the output feature. Specifically, local features are used as input features for the first feature compression, the output features of the previous feature compression are used as input features for the next feature compression, and the output features of the last feature compression are used as leakage features.

[0011] Optionally, a gated recurrent unit network is used to dynamically weight the leakage features to capture the temporal dependencies of the leakage features and obtain fused features, specifically including: The leakage characteristics are dynamically weighted multiple times to obtain the fused characteristics; For each dynamic weighting, the input features of the current dynamic weighting are weighted and fused with the hidden state of the previous dynamic weighting to obtain the temporal representation; the temporal representation is weighted and fused with the output features of the previous dynamic weighting to obtain the output features. Specifically, the leakage feature is used as the input feature for the first dynamic weighting, the output feature of the previous dynamic weighting and the initial input feature of the next dynamic weighting are used together as the input feature for the next dynamic weighting, and the output feature of the last dynamic weighting is used as the fusion feature.

[0012] Optionally, leak location is performed based on fusion characteristics to obtain the leak detection and location results for the pressurized pipeline under test, specifically including: A linear transformation is performed on the fused features to obtain multiple candidate leakage locations and their corresponding prediction scores; The predicted scores are normalized to obtain the probability distribution of each candidate leak location, and the candidate leak location corresponding to the highest probability distribution is taken as the leak detection and localization result.

[0013] Optionally, after acquiring the high-frequency pressure signal during the operation of the pressurized pipeline, the high-frequency pressure signal is normalized based on the following formula: ; in, x i For the first i One set of raw data, y i For the first i A standardized set of data, The average of all the original data. σ ( x ) represents the standard deviation of all the original data.

[0014] The pressurized pipeline leakage detection and location method provided in this embodiment of the invention has the following advantages compared with the prior art: This invention synchronously acquires high-frequency pressure signals from multiple sensors deployed on a pipeline to form an original sequence. Then, convolutional operations are used to jointly model and fuse spatiotemporal features of the multi-source signals, effectively suppressing random noise and extracting local features. The lightweight convolutional network SqueezeNet further compresses and enhances the features, highlighting the subtle fluctuation patterns caused by leaks. A gated recurrent unit (GRU) captures the temporal propagation patterns of leak features and the delay relationships between sensors, achieving dynamic weighting and contextual fusion. The entire process progressively deepens feature extraction, accurately distinguishing leak signals from background noise, significantly improving the detection and localization capabilities of low signal-to-noise ratio signals under low-pressure, small-leakage conditions, and reducing the false negative rate. Attached Figure Description

[0015] Figure 1 This is a typical structural diagram of a gated loop unit for a pressurized pipeline leak detection and location method provided in one embodiment; Figure 2This is a typical structural diagram of the Fire module in a pressurized pipeline leak detection and location method provided in one embodiment; Figure 3 The flowchart shows the implementation of the SqueezeNet-GRU model for a pressurized pipeline leak detection and location method provided in one embodiment. Figure 4 This is a schematic diagram of the experimental circuit in Embodiment 1 of a pressurized pipeline leak detection and location method provided in one embodiment; Figure 5 This is the confusion matrix in Embodiment 1 of a pressurized pipeline leak detection and location method provided in one embodiment.

[0016] Figure 4 In the middle: 1. Water storage tank; 2. Centrifugal pump; 3. Check valve; 4. First ball valve; 5. First butterfly valve; 6. First electromagnetic flowmeter; 7. Third high-frequency pressure sensor; 8. Second high-frequency pressure sensor; 9. First high-frequency pressure sensor; 10. Air valve; 11. Second electromagnetic flowmeter; 12. Second butterfly valve; 13. Second ball valve; 14. First leak point; 15. Second leak point; 16. Third leak point; 17. Fourth leak point; 18. Fifth leak point; 19. Sixth leak point; 20. Seventh leak point; 21. Data acquisition system. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] In practical applications, especially in urban water supply systems, low-pressure conditions and minute leaks are frequently encountered. In these situations, the leak signal energy is weak and the characteristics are not obvious, making it difficult for traditional detection methods to effectively identify the leak. Currently widely used leak detection methods mainly include the mass balance method, the negative pressure wave method, and fiber optic sensing technology. However, these methods generally suffer from low sensitivity and high false alarm rates, and their performance is limited, especially under conditions of small leaks, low pressure, or high noise, making it difficult to meet engineering requirements.

[0019] Studies have shown that Convolutional Neural Networks (CNNs) have achieved good results in pipeline leak detection. However, traditional CNNs are mainly designed for two-dimensional image data and are difficult to process one-dimensional time-series signals, such as pressure or vibration sequences, which often require dimensionality transformation, increasing redundant computation and potentially losing key information.

[0020] Scholars have proposed one-dimensional convolutional neural networks (1D-CNNs), which can be directly applied to time-series data to extract features while preserving the original structure. Although 1D-CNNs possess efficient feature extraction capabilities, they have limitations in modeling the long-term dependencies of time series data. These long-term dependencies are precisely one of the key sequence features in leak detection; ignoring this may affect the accuracy of localization.

[0021] This invention provides a method for detecting and locating leaks in pressurized pipelines. It employs a model architecture that integrates a lightweight convolutional neural network (SqueezeNet) and a gated recurrent unit (GRU). By processing high-frequency pressure signals acquired under actual operating conditions, it achieves efficient detection and accurate location of small leaks. The method specifically includes the following steps:

[0022] By deploying multiple sensors on the pressurized pipeline under test, high-frequency pressure signals are collected during the operation of the pressurized pipeline. All high-frequency pressure signals are then arranged according to the acquisition time sequence to obtain a high-frequency pressure signal sequence.

[0023] A convolution operation is performed on the high-frequency pressure signal sequence to jointly model and fuse the spatiotemporal features of all high-frequency pressure signals, thereby obtaining local features.

[0024] A convolutional neural network is used to compress local features to determine the time delay and amplitude difference of the leakage signal during propagation, thus obtaining the leakage characteristics.

[0025] A gated recurrent unit network is used to dynamically weight the leakage features in order to capture the temporal dependencies of the leakage features and obtain fused features.

[0026] Leakage location is determined based on fusion characteristics, yielding the leak detection and location results for the pressurized pipeline under test.

[0027] The leak detection and localization results are obtained through a leak detection and localization model. This model comprises: a data processing module, a two-layer lightweight convolutional neural network, a two-layer gated recurrent unit network, a fully connected layer, and a Softmax layer, connected sequentially.

[0028] The data processing module includes: a one-dimensional convolutional layer, an activation function module, and a pooling layer connected in sequence. Each lightweight convolutional neural network layer includes: two Fire modules connected in sequence. Each Fire module includes: a 1×1 convolutional kernel and a dimension expansion module connected in sequence. The dimension expansion module includes: a 1×1 convolutional branch and a 3×3 convolutional branch. The 1×1 convolutional branch includes multiple 1×1 convolutional kernels, and the 3×3 convolutional branch includes multiple 3×3 convolutional kernels. Each gated recurrent unit network includes: two gated recurrent units connected in sequence. Each gated recurrent unit includes: a parallel update gate and a reset gate.

[0029] The training of the leak detection and localization model specifically includes: Acquire high-frequency pressure signal data and corresponding leak detection and location results for single-point leaks in pressurized pipelines. Standardize the high-frequency pressure signal data for all single-point leaks in pressurized pipelines to obtain a high-frequency pressure signal dataset.

[0030] The high-frequency pressure signal dataset is divided into training, validation, and test sets according to a predetermined ratio. The training set is input into the leak detection and localization model to obtain leak detection and localization prediction results. The cross-entropy loss function is determined based on the leak detection and localization prediction results and the actual leak detection and localization results. Multiple iterations are performed using the backpropagation algorithm to update the hyperparameters of the leak detection and localization model. After each iteration, the model performance is evaluated using the validation set, and the hyperparameters of the leak detection and localization model are adjusted. Iteration stops when the maximum number of iterations or the validation accuracy convergence condition is met, resulting in the pre-trained leak detection and localization model.

[0031] The leak detection and localization model after initial training was independently validated using a test set, resulting in the final trained leak detection and localization model.

[0032] Specifically, a convolutional neural network is used to compress local features to determine the time delay and amplitude difference of the leakage signal during propagation, thus obtaining leakage features, including: Multiple feature compression operations are performed on local features to obtain leaked features. For each feature compression, a 1×1 convolution kernel is used to perform a one-dimensional convolution operation on the input features to compress the dimensionality of the input features, thus obtaining the first intermediate feature.

[0033] One-dimensional and three-dimensional convolution operations are performed on the first intermediate feature using 1×1 and 3×3 convolution branches, respectively, to expand the dimensionality of the first intermediate feature at different scales, resulting in the second and third intermediate features. The second and third intermediate features are then concatenated and fused along the channel dimension to obtain the output feature.

[0034] Local features are used as input features for the first feature compression, and the output features of the previous feature compression are used as input features for the next feature compression. The output features of the last feature compression are used as leakage features.

[0035] Specifically, a gated recurrent unit network is used to dynamically weight the leakage features to capture their temporal dependencies, resulting in fused features, including: The leakage characteristics are dynamically weighted multiple times to obtain the fused characteristics.

[0036] The leakage features are input into the gated recurrent unit network in chronological order. In each dynamic weighting, a reset gate is used to perform weighted fusion (tanh function) of the input features of the current dynamic weighting with the hidden state of the previous dynamic weighting to obtain the temporal representation.

[0037] We use the updated goalkeeper time-series representation and the previous dynamically weighted output features for weighted fusion to model the temporal dependency between the output features of each dynamically weighted iteration, thus obtaining the output features.

[0038] The leakage feature is used as the input feature for the first dynamic weighting. The output feature of the previous dynamic weighting and the initial input feature of the next dynamic weighting are used together as the input feature for the next dynamic weighting. The output feature of the last dynamic weighting is used as the fusion feature.

[0039] Among them, leak location is performed based on fusion characteristics to obtain the leak detection and location results of the pressurized pipeline under test, specifically including: The fused features are input into a fully connected layer, and the hyperparameters obtained during model training are used to perform a linear transformation on the fused features to obtain multiple candidate leak locations and their corresponding prediction scores. The prediction scores are then input into a Softmax layer for normalization to obtain the probability distribution of each candidate leak location. The candidate leak location corresponding to the highest probability distribution is taken as the leak detection and localization result (i.e., the leak location is determined based on the category corresponding to the highest probability in the probability distribution, and the leak detection and localization result is output).

[0040] A specific embodiment of the present invention is provided, including the following: Step 1: Experimental system setup and data acquisition preparation.

[0041] In pressurized water pipeline systems, an experimental loop is constructed and a data acquisition system is configured by installing necessary sensing and control equipment. The configured system is used to collect high-frequency pressure signals during pipeline operation, serving as a data source for subsequent model training and testing. The experimental system includes a water pump, water tank, pressure sensor, leakage control components, flow measurement equipment, and data acquisition unit, ensuring the representativeness and engineering applicability of the collected data. Specific equipment installation requirements: one high-frequency pressure sensor is installed every 1–300 m, with a sampling frequency of 100–5000 Hz; 2–10 leakage test points are set between every two high-frequency pressure sensors, and a pneumatic ball valve is installed at each leakage test point to simulate pipeline leakage accidents by opening and closing the ball valve.

[0042] Step 2: High-frequency pressure signal acquisition and preprocessing.

[0043] Pipeline pressure signals are collected in real time under normal and leakage conditions using pressure sensors, and the raw signals are normalized (as shown in Equation 1) to ensure data quality and subsequent modeling accuracy, and to reduce the influence of dimensions. Normal operating conditions refer to the operation of the water pipeline at a constant flow rate without valve braking, leakage, or other events. Leakage conditions refer to the situation where a leakage event occurs during the normal operation of the pipeline.

[0044] (1) in, x i For the first i One set of raw data, y i For the first i A standardized set of data, The average of all the original data. σ ( x ) represents the standard deviation of all the original data.

[0045] Step 3: Dataset partitioning.

[0046] The preprocessed dataset is divided into training, validation, and test sets in a specific ratio (8:1:1 to 9:0.5:0.5). The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and preventing overfitting, and the test set is used to evaluate the model's generalization ability. To ensure the comprehensiveness of model performance, the test set is not involved in any training process.

[0047] Step 4: SqueezeNet-GRU model construction and training.

[0048] Construct a deep learning model that integrates SqueezeNet and GRU, with the GRU structure as follows: Figure 1 As shown, the Fire module in SqueezeNet is as follows: Figure 2 As shown, the SqueezeNet-GRU model structure is as follows: Figure 3 As shown.

[0049] The model first uses a one-dimensional convolutional layer to extract local features from the preprocessed high-frequency pressure signal, and then performs efficient compression and feature extraction through multiple Fire modules. Subsequently, the features are input into a two-layer GRU network to capture long-range dependencies in the time series. The model will continuously iterate and train until it reaches the preset maximum number of iterations or the expected accuracy target, and save the optimal result. To prevent overfitting, dropout layers are set at multiple stages to enhance the model's robustness and generalization ability. Finally, the model's generalization ability is verified using test set data, and the final result is output.

[0050] The SqueezeNet-GRU model structure constructed in this invention includes: One-dimensional convolutional layers (Conv1D) are used for initial feature extraction; a SqueezeNet sub-network consisting of two sets of Fire modules (four modules in total) is used to compress and refine local features; a two-layer GRU network is used to model the temporal dependence of the leaked signal; a Dropout layer is used to prevent overfitting; and fully connected layers and Softmax layers are used for the final output. The model uses the cross-entropy loss function, the Adam optimizer for parameter updates, a maximum training epoch of 500 epochs, and a target accuracy of 99%.

[0051] The model training process uses the cross-entropy loss function and updates the parameters of each layer through the backpropagation algorithm. Training is complete after reaching the maximum number of iterations or the validation accuracy convergence condition, and the best-performing model is retained.

[0052] Step 5: Model testing and performance evaluation.

[0053] The trained model was independently validated using all data and the test set data to evaluate its predictive ability on unseen data. Performance metrics included accuracy. ε acc (Accuracy) ε pre (Precision) ε recall Recall and F1 score ε FS Furthermore, a confusion matrix was used for quantitative and qualitative analysis to verify the effectiveness and stability of the model in the task of identifying and classifying leakage events.

[0054] (2) (3) (4) (5) In the formula, n TP This means that the predicted result is a positive sample, and the actual result is also a positive sample; n FP This indicates that the predicted result was a positive sample, but the actual result was a negative sample. n TN This means that the predicted result is a negative sample, and the actual result is also a negative sample; n FN This means that the predicted result was a negative sample, but the actual result was a positive sample.

[0055] The beneficial effects that this invention can achieve include, but are not limited to: (1) The model is constructed using gated cyclic units (GRUs) to give full play to their advantages in modeling long-range dependencies in time series. Since high-frequency pressure signals have typical long-sequence characteristics, the GRU structure can effectively capture the potential dynamic evolution patterns and improve the timeliness and accuracy of leakage signal identification.

[0056] (2) To address the relatively limited ability of GRU in extracting high-dimensional, multi-channel signal features, a lightweight convolutional neural network framework, SqueezeNet, was introduced to enhance the ability to extract local signal features. Especially under low-pressure and small-leakage conditions, leakage features are not obvious and the signal-to-noise ratio is low, making it difficult to identify using traditional methods. The introduction of the SqueezeNet framework significantly improves the sensitivity and recognition ability to weak signal changes.

[0057] (3) The constructed SqueezeNet-GRU model (leakage detection and localization model) combines the efficient feature compression capability of the lightweight convolutional neural network SqueezeNet with the time series modeling advantages of GRU. While ensuring the accuracy of the model, it significantly reduces the number of parameters and computational complexity, making the model suitable for deployment on resource-constrained terminal devices or edge computing platforms, and meeting the application requirements of real-time monitoring and online detection.

[0058] (4) The proposed leak detection and location model can not only determine whether a leak has occurred, but also has good leak location capabilities. By jointly modeling and fusing spatiotemporal features of multi-channel sensor data, the model can accurately identify the time delay and amplitude difference in the propagation process of the leak signal, thereby effectively realizing the accurate identification and precise location of low-pressure, small leak accidents with weak signals and indistinct features.

[0059] (5) It has good adaptability and engineering versatility. Since the leakage signals under different working conditions have certain commonalities in timing patterns, it can be applied to pipeline systems with various fluid media such as water, gas, and oil, and can be applied to leakage detection tasks under different materials, pipe diameters, and operating pressures.

[0060] (6) The model has good scalability. As the dimensions of the input data increase and the number of sensors increases, the collaborative features between multi-source information can be further explored to improve the robustness and generalization ability of leak detection. At the same time, it can be flexibly connected to existing pipeline monitoring systems, and has good integration and engineering promotion prospects.

[0061] Furthermore, a specific embodiment 1 of the present invention is provided: To verify the feasibility and effectiveness of the proposed SqueezeNet-GRU leak detection method in a real engineering environment, a pressurized water pipeline leak detection experimental platform was built, and experimental data under low-pressure and small-leakage conditions were obtained. The leak detection method proposed in this invention was then used for leak localization. The specific modeling and model verification process is as follows: I. Experimental Platform Setup and Data Acquisition

[0062] The experimental platform built in this invention is as follows: Figure 4 As shown, the circuit is 144m long and includes, in sequence, a water storage tank 1, a centrifugal pump 2, a check valve 3, a first ball valve 4, a first butterfly valve 5, a first electromagnetic flowmeter 6, a third high-frequency pressure sensor 7, a second high-frequency pressure sensor 8, a first high-frequency pressure sensor 9, an air valve 10, a second electromagnetic flowmeter 11, a second butterfly valve 12, and a second ball valve 13. The second ball valve 13 is connected to the water storage tank 1 via a pipe. It also includes a data acquisition system 21, with the first electromagnetic flowmeter 6, the second electromagnetic flowmeter 11, the first high-frequency pressure sensor 9, and the second high-frequency pressure sensor 8 all connected to it. The distances between the various devices are shown in the diagram. Figure 4 . Seven leak test points were set up on the experimental pipeline, each with a diameter of 13 mm, and equipped with ball valves for control of opening and closing to simulate sudden leak events. Leak points 14, 15, 16, and 17 were set up on the pipeline between the first high-frequency pressure sensor 9 and the second high-frequency pressure sensor 8, arranged along the water flow direction, with each leak point spaced 2 m apart. Leak points 18, 19, and 20 were set up on the pipeline between the second high-frequency pressure sensor 8 and the first high-frequency pressure sensor 7, arranged along the water flow direction, with each leak point spaced 2 m apart. The fourth leak point 17 and the seventh leak point 20 were spaced 20 m apart. The first, second, third, and fourth leak points constituted group A, and the fifth, sixth, and seventh leak points constituted group B.

[0063] Three high-frequency pressure sensors were installed on the pipeline for real-time signal acquisition, with a sampling frequency set to 2000 Hz. During the experiment, the relative distances between each leak point and the pressure sensors were carefully designed to ensure effective capture of changes in the leakage signal propagation process.

[0064] II. Data Acquisition

[0065] Before the experiment, water from the storage tank was injected into the pipeline system using a centrifugal pump, and different volumetric flow rates (40~80 m³ / h) were set for operation. During the experiment, the pressure range was 130~220 kPa, and the leakage volumetric flow rate was 0.65~1.33 m³ / h. The leakage volumetric flow rate was less than 5% of the total volumetric flow rate, which is a typical low-pressure, small-leakage condition. First, signal acquisition was performed under a leak-free state. Then, the ball valves at each leakage point were opened sequentially to simulate transient leakage, and the corresponding pressure changes were recorded.

[0066] The acquisition time for each experimental signal was 20 seconds, and the number of data points in a single group was 40,000. In each round of experiments, high-frequency pressure signal data were collected under normal operating conditions (no leakage) and single-point leakage conditions. A total of 962 experimental data were obtained, including 416 data under no leakage conditions, 312 data under Group A leakage conditions, and 234 data under Group B leakage conditions.

[0067] III. Data Preprocessing and Model Building.

[0068] Following step 2, all high-frequency pressure signal data were standardized, and then, according to step 3, the dataset was divided into a training set (769 sets), a validation set (97 sets), and a test set (96 sets) in an 8:1:1 ratio. The training and validation sets were used for model training and parameter tuning, while the test set was used to independently evaluate model performance.

[0069] IV. Model Evaluation and Result Analysis.

[0070] After training, the model performance was evaluated using all data and the test set. Evaluation metrics included accuracy, precision, recall, and F1 score. The test results were as follows: For the full dataset: accuracy 97.40%, precision 96.02%, recall 95.99%, F1 score 96.00%; For the test set: accuracy 96.00%, precision 95.57%, recall 92.65%, F1 score 93.590%. The confusion matrix for the full dataset is as follows: Figure 5 As shown in the figure. The verification results show that the SqueezeNet-GRU model constructed in this invention can accurately identify leakage conditions, has good classification performance and strong generalization ability, and is especially suitable for leakage location tasks under low pressure and small leakage conditions.

[0071] This invention constructs a leak detection and localization model by combining a lightweight convolutional neural network with a gated recurrent unit. This model integrates the efficient feature compression capability of a lightweight convolutional neural network with the time series modeling advantage of a gated recurrent unit, significantly reducing the number of model parameters and computational complexity while ensuring model accuracy.

[0072] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method of leak detection and location for a pressurized pipeline, the method comprising: include: By deploying multiple sensors on the pressurized pipeline under test, high-frequency pressure signals during the operation of the pressurized pipeline are collected, and all high-frequency pressure signals are arranged according to the acquisition time sequence to obtain a high-frequency pressure signal sequence. The high-frequency pressure signal sequence is convolved to jointly model and fuse spatiotemporal features of all high-frequency pressure signals to obtain local features; A convolutional neural network is used to compress local features to determine the time delay and amplitude difference of the leakage signal during propagation, thereby obtaining the leakage characteristics. A gated recurrent unit network is used to dynamically weight the leakage features in order to capture the temporal dependencies of the leakage features and obtain fused features. Leakage location is determined based on fusion characteristics, resulting in the leak detection and location of the pressurized pipeline under test. Leak detection and localization results are obtained through a leak detection and localization model; the leak detection and localization model includes: a data processing module, a two-layer lightweight convolutional neural network, a two-layer gated recurrent unit network, a fully connected layer, and a Softmax layer connected in sequence; The data processing module includes: a one-dimensional convolutional layer, an activation function module, and a pooling layer connected in sequence; Each layer of the lightweight convolutional neural network includes: two Fire modules connected in sequence; each Fire module includes: a 1×1 convolutional kernel and a dimension expansion module connected in sequence, the dimension expansion module includes: a 1×1 convolutional branch and a 3×3 convolutional branch, the 1×1 convolutional branch includes multiple 1×1 convolutional kernels, and the 3×3 convolutional branch includes multiple 3×3 convolutional kernels; Each gated loop unit network includes two gated loop units connected in sequence, and each gated loop unit includes a parallel update gate and a reset gate.

2. The method for detecting and locating leaks in a pressurized pipeline as described in claim 1, characterized in that, Training the leak detection and localization model specifically includes: Acquire high-frequency pressure signal data and corresponding leak detection and location results under single-point leakage conditions in pressurized pipelines; The high-frequency pressure signal data of all pressurized pipelines under single-point leakage conditions are standardized to obtain a high-frequency pressure signal dataset; The high-frequency pressure signal dataset is divided into training set, validation set and test set according to a set ratio; Input the training set into the leak detection and localization model to obtain the leak detection and localization prediction results; The cross-entropy loss function is determined based on the leak detection and location prediction results and the actual leak detection and location results. The hyperparameters of the leak detection and location model are updated through multiple iterations using the backpropagation algorithm. After each iteration, the model performance is evaluated using a validation set, and the hyperparameters of the leak detection and location model are adjusted. The iteration stops when the maximum number of iterations or the validation accuracy convergence condition is reached, resulting in the pre-trained leak detection and location model. The leak detection and localization model after initial training was independently validated using a test set, resulting in the final trained leak detection and localization model.

3. The method for detecting and locating leaks in a pressurized pipeline as described in claim 1, characterized in that, The convolution operation on the high-frequency pressure signal sequence is used to jointly model and fuse the spatiotemporal features of the high-frequency pressure signals from multiple sensors to obtain local features, specifically including: A one-dimensional convolution operation is performed on the high-frequency pressure signal sequence to perform weighted operations on the local time windows of all high-frequency pressure signals, thereby obtaining a multi-channel feature map. A nonlinear mapping is performed on the multi-channel feature map, and a one-dimensional average pooling operation is used to reduce the dimensionality and compress the feature map after the nonlinear mapping to obtain local features.

4. The method for detecting and locating leaks in a pressurized pipeline as described in claim 1, characterized in that, The method of using a convolutional neural network to compress local features to determine the time delay and amplitude difference of the leakage signal during propagation, thereby obtaining leakage features, specifically includes: The leakage features are obtained by performing multiple feature compressions on the local features; For each feature compression, a one-dimensional convolution operation is performed on the input feature to compress its dimension, resulting in the first intermediate feature. Then, one-dimensional and three-dimensional convolution operations are performed on the first intermediate feature to expand its dimension at different scales, resulting in the second and third intermediate features. Finally, the second and third intermediate features are concatenated and fused along the channel dimension to obtain the output feature. Specifically, local features are used as input features for the first feature compression, the output features of the previous feature compression are used as input features for the next feature compression, and the output features of the last feature compression are used as leakage features.

5. The method for detecting and locating leaks in a pressurized pipeline as described in claim 1, characterized in that, The method employs a gated recurrent unit network to dynamically weight leakage features in order to capture the temporal dependencies of leakage features and obtain fused features, specifically including: The leakage characteristics are dynamically weighted multiple times to obtain the fused characteristics; For each dynamic weighting, the input features of the current dynamic weighting are weighted and fused with the hidden state of the previous dynamic weighting to obtain the temporal representation; the temporal representation is weighted and fused with the output features of the previous dynamic weighting to obtain the output features. Specifically, the leakage feature is used as the input feature for the first dynamic weighting, the output feature of the previous dynamic weighting and the initial input feature of the next dynamic weighting are used together as the input feature for the next dynamic weighting, and the output feature of the last dynamic weighting is used as the fusion feature.

6. The method for detecting and locating leaks in a pressurized pipeline as described in claim 1, characterized in that, The process of locating leaks based on fusion features to obtain leak detection and location results for the pressurized pipeline under test specifically includes: A linear transformation is performed on the fused features to obtain multiple candidate leakage locations and their corresponding prediction scores; The predicted scores are normalized to obtain the probability distribution of each candidate leak location, and the candidate leak location corresponding to the highest probability distribution is taken as the leak detection and localization result.

7. The method for detecting and locating leaks in a pressurized pipeline as described in claim 1, characterized in that, This also includes normalizing the high-frequency pressure signal after acquiring it during the operation of the pressurized pipeline, based on the following formula: ； in, x i For the first i One set of raw data, y i For the first i A standardized set of data, The average of all the original data. σ ( x ) represents the standard deviation of all the original data.

Images