A two-stage leakage monitoring method and device for water distribution network based on acoustic signals

Abstract

The present invention discloses a two-stage leakage monitoring method for a water distribution network based on acoustic signals, comprising the following steps: obtaining on-site acoustic signals of leakage events and normal operation, and constructing corresponding log-mel spectrum graphs; labeling the log-mel spectrum graphs to form a data set; constructing a prediction network comprising two stages; training the prediction network using the data set to obtain a leakage event prediction model; inputting the on-site acoustic signals of the pipelines in the area to be analyzed into the leakage event prediction model to obtain prediction results for leakage events in each pipeline in the area. The present invention also provides a two-stage leakage monitoring device for a water distribution network. The method provided by the present invention can make the optimal decision for pipeline leakage management from a large amount of alarm information to eliminate current leakage events.

Description

Technical Field

[0001] The present invention belongs to the technical field of urban pipe network management, and in particular relates to a two-stage leakage monitoring method and device for a water distribution pipe network based on acoustic signals. Background Art

[0002] As a critical urban infrastructure, water distribution networks (WDNs) play an irreplaceable role in efficiently delivering clean, safe drinking water from water plants to consumers. However, internal and external factors, such as pipeline material defects and aging, lead to frequent leaks and bursts, resulting in severe economic losses and secondary contamination of drinking water, posing a significant challenge to the sustainable management of water resources. Therefore, there is a need to develop more comprehensive, effective, and advanced leak monitoring frameworks and systems, combined with online acoustic signal acquisition equipment, to identify and manage potential leaks and promote the sustainable development of WDNs.

[0003] Patent document CN116658837A discloses a water supply network leakage location system, comprising a terminal device, a cloud platform, and a mobile terminal. The terminal device includes a sensing unit and a circuit board. The sensing unit uses a vibration sensor to convert pipeline leakage noise from vibration signals into electrical signals. The circuit board includes a main control chip that switches between two different operating modes, including online monitoring mode and offline field detection mode, according to the program in the main control chip. In the online monitoring mode, the terminal device and the cloud platform operate for daily monitoring of the monitoring area. Vibration sensors are arranged at various points in the monitoring area to regularly collect pipeline noise signals and transmit them to the cloud platform. The cloud platform uses a positioning algorithm to accurately locate the leakage point and displays the daily monitoring results on the cloud platform. In the offline detection mode, the terminal device, the mobile terminal, and the cloud platform operate for real-time on-site leakage detection and location. The sensing terminal is arranged at the point to be detected. The mobile terminal issues a detection task to the terminal device. The terminal device completes the pipeline noise data collection and transmits the data to the cloud platform. The cloud platform uses a positioning algorithm to accurately locate the leakage point, and the cloud platform returns the calculation results to the mobile terminal for display.

[0004] Patent document CN113483276A discloses an uncertainty assessment method and system for leak location, comprising: receiving input pipeline parameters, obtaining sound velocity data and its confidence interval corresponding to the current pipeline based on the pipeline parameters; collecting noise signals generated by the leak location on the current pipeline multiple times, and calculating the leakage noise propagation time difference and its confidence interval based on the noise signal; calling the sound velocity data and the leakage noise propagation time difference to calculate the leak location of the current pipeline, and inferring the confidence interval of the pipeline leak location based on the confidence interval of the sound velocity data and the confidence interval of the time difference for output. Summary of the Invention

[0005] The purpose of the present invention is to provide a two-stage leakage monitoring method and device for a water distribution network based on acoustic signals, which can make the optimal decision for pipeline leakage management from a large amount of alarm information on site to eliminate current leakage events.

[0006] In order to achieve the first object of the present invention, the following technical solution is provided: a two-stage leakage monitoring method for a water distribution network based on acoustic signals, comprising the following steps:

[0007] Acquire historical data, which includes on-site acoustic signals of leakage events and normal operation, and construct corresponding logarithmic Mel-frequency spectrograms frame by frame along the time axis based on the on-site acoustic signals;

[0008] The logarithmic Mel-spectrogram is labeled with three leakage level assessments: mild leakage, moderate leakage, and severe leakage, and the on-site acoustic signal, the corresponding logarithmic Mel-spectrogram, and the labels are combined into a data set;

[0009] A prediction network is constructed based on the coupled residual convolution and Transformer network framework. The prediction network consists of two stages, where the first stage includes a data preprocessing module, a local feature extraction module, a global feature extraction module, and a feature fusion module.

[0010] Phase II includes a leakage level assessment module;

[0011] The data preprocessing module is used to convert the input live sound signal into a corresponding logarithmic Mel spectrum;

[0012] The local feature extraction module is used to extract the spectrum feature sequence of the logarithmic Mel spectrum graph;

[0013] The global feature extraction module and the local feature extraction module output the long-range dependency between the spectral feature sequences to output the corresponding global feature vector;

[0014] The fusion feature module is used to perform a fusion operation on the input spectrum feature sequence and the global feature vector to obtain a fusion feature;

[0015] The leakage level assessment module makes predictions based on the input fusion features to output assessment results, wherein the assessment results include a mild leakage assessment result, a moderate leakage assessment result, and a severe leakage assessment result;

[0016] The first stage of the prediction network is trained using the dataset, and the corresponding leakage amount is introduced into the dataset as a label to train the second stage of the prediction network, so as to obtain a leakage event prediction model for leakage event assessment;

[0017] The on-site acoustic signals of the pipelines in the area to be analyzed are input into the leakage event prediction model to obtain the prediction results of leakage events of each pipeline in the area.

[0018] Drawing on analytical logic from the field of meteorology, this paper proposes a two-stage leakage monitoring framework based on acoustic signals. Unlike conventional leak detection models, the proposed monitoring framework not only determines whether a pipeline is leaking and the probability of leakage, but also provides an assessment of the leakage level. That is, by integrating information such as leakage probability, leakage presence, and leakage level, managers can make optimal decisions on pipeline leakage management based on a large amount of on-site alarm information.

[0019] Specifically, the generation process of the logarithmic Mel-spectrogram is as follows:

[0020] The live sound signal is converted into power spectrum by combining fast Fourier transform and square function;

[0021] A dot product operation is performed on the power spectrum on a plurality of Mel-level triangular filter banks to obtain a logarithmic Mel-level spectrum map of each frame.

[0022] Specifically, the data preprocessing module includes a Filter Bank filter.

[0023] Specifically, the local feature extraction module uses a CNN with a residual block to perform multi-level local feature extraction on the logarithmic spectrum. The residual block includes two 3x3 convolutional layers with the same number of output channels, and the output end of each convolutional layer is connected to a batch normalization layer and a ReLU activation function.

[0024] Specifically, the global feature extraction module includes an encoder, stacked self-attention and fully connected layers, and the encoder performs layer normalization on the input spectral feature sequence to obtain frequency domain feature data, and performs weighted calculation on the frequency domain feature data based on the stacked self-attention mechanism and the fully connected layer to obtain a global feature vector.

[0025] Specifically, the fusion feature is the leakage probability of the pipeline, which is obtained by performing nonlinear fitting on the input local spectrum feature sequence and global feature vector through the sigmoid function of the fully connected layer.

[0026] Specifically, the label of the leakage amount is expressed as follows:

[0027]

[0028] The unit of leakage is m 3 / d.

[0029] For the second purpose of the present invention, the following technical solution is provided: a two-stage leakage monitoring device for a water distribution network, which is implemented by the above-mentioned two-stage leakage monitoring method for a water distribution network based on acoustic signals, and its specific steps are as follows: a large amount of on-site alarm information is input into a leakage event prediction model to output a leakage level assessment for each pipeline, providing a reference for arranging the optimal decision-making plan, wherein the optimal decision-making plan includes determining when the pipeline should be repaired, the urgency of the repair, and the priority of allocating resources.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] It has high accuracy and stability for acoustic leak detection in water distribution networks, can more effectively capture the local and global frequency domain characteristics of pipeline acoustic signals, and provide more accurate leakage probability estimation and leakage level assessment. At the same time, the monitoring framework couples residual convolution with Transformer, making the two complement each other's advantages, and realizes rapid classification of on-site pipeline acoustic signals and leakage information mining. By integrating information such as leakage probability, leakage and leakage level, managers can make the optimal decision for pipeline leakage management among a large amount of on-site alarm information. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Schematic diagram of two-stage leak monitoring based on meteorological knowledge provided for this example

[0033] Figure 2 A flow chart of a two-stage leakage monitoring method for a water distribution network based on acoustic signals provided in this embodiment;

[0034] Figure 3 Flowchart of generating logarithmic Mel-spectrogram for the Filter Bank provided in this embodiment;

[0035] Figure 4 A schematic diagram of the residual convolution structure provided in this embodiment;

[0036] Figure 5 A comparison chart of the evaluation results of each model provided in this embodiment;

[0037] Figure 6 A schematic diagram of leakage probability distribution provided in this embodiment.

[0038] Figure 7 The following are the recognition results of the untrained and pretrained RCT-Net provided in this embodiment at different leakage levels. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0040] like Figure 1 As shown, when monitoring pipeline leaks, online acoustic signals can be combined with data-driven models to estimate the probability of a pipeline leak. Once a leak is confirmed, the severity of the leak is associated with the leak level, such as minor leak, moderate leak, or severe leak, providing more complete leak information (including leak probability, leak status, and leak level). Assessing the leak level helps assess the potential impact of water leaks on the environment and facilities in WDNs and take appropriate measures to optimize the priority of pipeline leak management. Cross-analyzing water pipe leak monitoring issues and meteorological disciplines can provide a diversified approach to leak monitoring research, providing new insights into solving existing problems in leak monitoring field applications (such as incomplete leak information and unclear pipeline leak management priorities).

[0041] like Figure 2 FIG. 1 is a flow chart of a two-stage leakage monitoring method for a water distribution network based on acoustic signals provided in this embodiment:

[0042] Historical data including leakage events and live acoustic signals of normal operation are acquired, and corresponding logarithmic Mel-spectrograms are constructed frame by frame along the time axis based on the live acoustic signals.

[0043] The log-Mel spectrum graph is labeled with three leakage levels: mild leakage, moderate leakage, and severe leakage. The on-site acoustic signal, the corresponding log-Mel spectrum graph, and the labels form a dataset.

[0044] A prediction network is constructed based on the coupled residual convolution and Transformer network framework. The prediction network consists of two stages, where the first stage includes a data preprocessing module, a local feature extraction module, a global feature extraction module, and a feature fusion module.

[0045] The second phase includes the leakage level assessment module.

[0046] The data preprocessing module is used to convert the input live sound signal into a corresponding logarithmic Mel spectrum.

[0047] Local feature extraction module, used to extract the spectral feature sequence of the logarithmic Mel spectrum graph;

[0048] The global feature extraction module and the long-range dependency between the spectral feature sequences output by the local feature extraction module are used to output the corresponding global feature vector.

[0049] The fusion feature module is used to fuse the input spectrum feature sequence and the global feature vector to obtain the fusion feature.

[0050] The leakage level assessment module makes predictions based on the input fusion features to output assessment results, which include a mild leakage assessment result, a moderate leakage assessment result, and a severe leakage assessment result.

[0051] The first stage of the prediction network is trained using the dataset, and the corresponding leakage amount is introduced into the dataset as a label to train the second stage of the prediction network, so as to obtain a leakage event prediction model for leakage event assessment.

[0052] The on-site acoustic signals of the pipelines in the area to be analyzed are input into the leakage event prediction model to obtain the prediction results of leakage events of each pipeline in the area.

[0053] Furthermore, in this embodiment, the specific execution process is as follows:

[0054] Since in acoustic leak detection, the on-site acoustic signal is usually formed into a spectrogram after STFT and combined with a data-driven model to carry out leakage identification. However, STFT will produce an energy leakage effect in the frequency domain, and requires a large number of Fourier transforms, with high computational complexity. The spectrogram formed has a large number of redundant features, which will greatly affect the speed and accuracy of the leakage detection model. The Filter Bank (filter bank) adopted in this embodiment is a signal processing technology consisting of a group of filters and samplers (down or up samplers). Each filter is responsible for extracting information within a specific frequency range from the input leakage signal, and can perform frequency decomposition of the signal to varying degrees. This design enables Filter Bank to capture the global features and local details of the signal at the same time, which helps to more fully understand the structure of the signal and remove noise within a specific frequency range. On the basis of STFT, Filter Bank can reduce the frequency domain leakage effect by appropriately designing filters and improve the accuracy of the Mel spectrum graph. In addition, by applying different filters at different frequencies and directions to remove the correlation in the leakage signal, it helps to reduce the redundant information in the leakage sound signal, thereby achieving feature dimensionality reduction, which is more suitable for carrying out leakage detection in combination with a data-driven model and improving the computational efficiency and accuracy of the leakage detection model.

[0055] like Figure 3 As shown in the figure, when processing a leakage sound signal, framing and windowing are first performed. A fast Fourier transform (FFT) is then applied to each frame to generate an amplitude spectrum, which is then converted to a power spectrum using a square function. The power spectrum is then dot-producted across multiple Mel-level triangular filter banks to determine the filter's Mel-frequency response, resulting in a logarithmic Mel-spectrogram for each frame. The conversion relationship between Mel-frequency and frequency (Hz) and the size of the logarithmic Mel-spectrogram are expressed as follows:

[0056]

[0057] H = N-mels;

[0058] Among them, f mel is the Mel frequency; f is the leakage signal frequency; W is the width of the logarithmic Mel spectrum; Signal_len is the number of seconds of the leakage signal; Sr is the sampling rate; hop_len represents; H is the height of the logarithmic Mel spectrum; N_mels is the number of Mel filters.

[0059] The prediction network constructed in this embodiment includes two stages. The first stage includes a data preprocessing module, a local feature extraction module, a global feature extraction module and a feature fusion module, and the second stage includes a leakage level assessment module.

[0060] The data preprocessing module is used to convert the input live sound signal into the corresponding logarithmic Mel spectrum.

[0061] Deep convolutional neural networks (DCNNs) have long held a prominent position in leak detection due to their powerful feature extraction capabilities. However, as the network layers deepen, the information acquired decreases layer by layer, making it prone to vanishing or exploding gradients, leading to poor leak detection performance. Therefore, the local feature extraction module in this embodiment employs a CNN with a residual block to perform multi-level local feature extraction on the log-Mel spectrogram of the pipeline acoustic signal to enrich local details and output a sequence of local spectral features in the log-Mel spectrogram.

[0062] The structure of its local feature extraction module is as follows Figure 4 As shown, the residual block first has two 3×3 convolutional layers with the same number of output channels. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function to extract low-level and mid-level features of the spectrum. Simultaneously, the two convolution operations in the residual block are skipped, and the original spectrum features are input directly before the final ReLU activation function after passing through a 1×1 convolutional layer. It is worth noting that the 1×1 convolutional layer does not make any changes to the spatial dimension, but mainly changes the channel dimension of the spectral features. This allows information to propagate directly across multiple layers, preserving key frequency domain features and solving the problems of vanishing gradients and network degradation.

[0063] The global feature extraction module in this embodiment is built based on the Transformer framework and only includes an encoder, stacked self-attention, and fully connected layers. It uses the Transformer encoding layer to capture the global features of the pipeline acoustic signal and the long-range dependencies between different features, thereby enhancing the global perception capability of the leakage monitoring frame.

[0064] In the global feature extraction module, layer normalization is first performed, and then the frequency domain feature data X is weighted through the multi-head self-attention mechanism to obtain a new feature vector F, which can be expressed as:

[0065]

[0066] Q=W Q ×X,K=W K ×X,V=W V ×X

[0067] Among them, W represents the weight matrix corresponding to the input vector.

[0068] After solving the eigenvector F, the eigenvector is fed into the feedforward neural network (FFnn) and processed by the activation function ReLU and the linear activation function to obtain the global eigenvector, which is expressed as follows:

[0069] FFnn(W)=max(0,FW1+b1)W2+b2;

[0070] The feature fusion module provided in this embodiment performs nonlinear fitting through the sigmoid function of the fully connected layer to determine the leakage probability of the pipeline, and finally sets a threshold to determine whether the pipeline is leaking or not.

[0071] The leakage level assessment module predicts and assesses the leakage volume of the pipeline by judging the results of pipeline leakage, providing key information for subsequent analysis of the degree of damage to the pipeline and the consequences of the impact.

[0072] The first step is to convert the previous leakage prediction problem into a leakage level classification, including leakage interval classification (mild, moderate and severe leakage), which is divided as follows:

[0073]

[0074] The unit of leakage is m 3 / d.

[0075] The leakage samples with leakage amount labels are input into the second stage for training, and the weight parameters of the network model are fine-tuned to accelerate the convergence of the network model and efficiently extract spectral features. The input composition of the model (leakage signal) and the output dimension of the prediction result (three leakage levels) are changed.

[0076] In the field of leak detection, accuracy, precision, F1-score, specificity, sensitivity, and AUC are commonly used to measure the recognition performance of the model. For specific definitions, see the confusion matrix in Table 1 and the evaluation indicators of the monitoring framework in Table 2.

[0077] Table 1

[0078] Predicted to be a leaked sample Predicted as non-missing sample Leaked Sample i True Positive (TP) False negative (FN) Non-missing samples False Positive (FP) True Negative (TN)

[0079] Table 2

[0080]

[0081] This embodiment further provides a two-stage leakage monitoring device for a water distribution network, which is implemented by the two-stage leakage monitoring method for a water distribution network based on acoustic signals as provided in the above embodiment. The specific process is as follows:

[0082] First, a fixed-size sliding window (1s) is set to obtain on-site acoustic signal data; (2) After denoising and dimensionality reduction by Filter Bank, the data is converted into a two-dimensional Mel-spectrogram and input into the leak detection stage. The local and global features extracted by modules 1, 2, and 3 are fitted to calculate the leakage probability, and then the decision of whether a leak has occurred is made based on the threshold set by the decision maker. If it is determined that no leak has occurred, the first step is repeated; (3) If it is determined that a leak has occurred, the signal sample is transmitted to the leakage assessment stage to predict the specific leakage level (mild, moderate, and severe leakage). (4) After the assessment is completed, comprehensive leakage information (leakage probability, leakage status, and different leakage levels) is provided to the decision makers of the water company to quickly discover leakage events and assist in making on-site leakage management priority decisions, guiding the maintenance and replacement of infrastructure assets.

[0083] In order to better illustrate the technical effects of the method provided in this embodiment, the following specific actual cases are provided:

[0084] Acoustic signal data from pipelines across various regions in several southern cities was collected. The water distribution network was equipped with AX3D piezoelectric accelerometers, with a sensitivity of at least 1200 pC / g. These sensors were evenly distributed throughout the pipes and valves, independent of each other, and transmitted to a signal processing center via the Internet of Things (IoT). Signals were captured at an 8096 Hz sampling rate and a duration of 5 seconds, ensuring detailed representation of the acoustic data. Furthermore, noise logger data collection was primarily scheduled between 2:00 and 4:00 AM to minimize the impact of external ambient noise.

[0085] The dataset includes 828 audio signals, including 309 leakage signals and 519 non-leakage signals. To meet the requirements of data-driven model training, the original audio signals were segmented into equal time intervals (1 second). Ultimately, this work collected a total of 1545 (309 × 5) leakage samples and 2595 (519 × 5) non-leakage signal samples. The collected signal samples come from a variety of scenarios, covering pipe diameters from DN15 to DN500, pipes made of galvanized steel, ductile iron, polyethylene, and various leak locations.

[0086] In order to compare the impact of STFT and Filter Bank feature extraction results on the leakage detection model, this section only uses the SqueezeNet model for 10-fold cross-validation tests of leakage detection. As shown in Table 3:

[0087] Table 3

[0088] Model Accuracy(%) Precision(%) F1-score (%) Specificity(%) Sensitivity(%) Existing technology 94.76 93.95 94.66 93.94 95.63 SqueezeNet 96.36 94.29 95.06 96.68 95.91

[0089] The existing technology uses STFT and SqueezeNet, achieving 94.76% accuracy, 93.95% precision, 94.66% F1-score, 93.94% specificity, and 95.63% sensitivity on the dataset. However, the combination of Filter Bank and SqueezeNet used in this section achieved 96.36% accuracy, 94.29% precision, 95.06% F1-score, 96.68% specificity, and 95.91% sensitivity on the dataset. The results show that each metric in this section outperforms the corresponding metric in the existing technology, with an average increase of approximately 1.1%. In particular, the Accuracy and Specificity metrics are approximately 2% higher than those in the literature. Therefore, compared to STFT, the spectral characteristics of pipe acoustic signals processed by Filter Bank are more pronounced, making it easier for deep learning models to learn detailed frequency domain features and effectively identify leakage signals.

[0090] Regarding the leakage event prediction model, in order to further illustrate the superiority and stability of the RCT-Net model in this work in the acoustic leak detection task, other deep learning models (including ResNet and SqueezeNet of CNN architecture, Transformer, and CNN-LSTM) were compared.

[0091] like Figure 5 As shown, RCT-Net achieved the best performance in all classification evaluation metrics, with accuracy, precision, F1-score, specificity, sensitivity, and AUC of 0.9807, 0.9713, 0.9732, 9838, 0.9752, and 0.9970, respectively. The other four deep learning models performed worse than RCT-Net in recognizing mel-spectrogram images. Among the four models other than RCT-Net, the CNN-LSTM model achieved the best overall leakage detection performance, with scores of 0.9705, 0.9547, 0.9581, 0.9751, 0.9620, and 0.9959, respectively. Furthermore, the performance of SqueezeNet and Transformer in various metrics was similar, with all performances lower than the other models. SqueezeNet outperformed Transformer in accuracy, F1-score, and sensitivity, while the opposite was true for precision and specificity.

[0092] Overall, the SqueezeNet and Transformer models, both based on CNN architectures, achieve comparable results across various recognition metrics. This demonstrates that both models can effectively distinguish between leaky and non-leaky pipes, regardless of whether they capture local or global frequency-domain features of the log-mel-spectrogram. Furthermore, ResNet also utilizes a CNN architecture, but its unique residual connection structure preserves more of the original local spectral features of the signal during feature extraction, resulting in superior performance across all recognition metrics. Furthermore, the CNN-LSTM model simultaneously leverages the CNN to capture local features of the mel-spectrogram and the LSTM to extract correlations between different features. Consequently, it achieves more accurate leak detection results compared to models that capture local features (SqueezeNet and ResNet) or global features (Transformer) independently. Compared to the LSTM, the Transformer excels at capturing global features and long-range dependencies. Therefore, RCT-Net, which combines the advantages of both ResNet and Transformer, boasts superior local and global feature capture capabilities compared to the CNN-LSTM model, achieving higher metrics and superior performance in leak detection tasks. In summary, the above results show that the RCT-Net model outperforms other deep learning methods in acoustic leak detection and is able to identify leakage signals with a good balance between accuracy and robustness.

[0093] The leakage event prediction model not only determines whether a pipeline leaks based on the logarithmic Mel spectrum, but also provides a reference for on-site leak detection based on the leakage probability of each signal sample. The leakage probability distribution of all acoustic signal samples in the test set is as follows: Figure 6 shown.

[0094] The RCT-Net model has strong leak detection capabilities, classifying most leak signals (triangles) into the region with probabilities close to [0.95, 1], and most non-leakage signals (circles) into the region [0, 0.05]. A small number of samples (squares) fall between 0.05 and 0.95. Decision-makers can set appropriate thresholds based on their risk tolerance and prior knowledge of the site. This model sets a threshold of 0.5. When the probability value is less than the set threshold, the signal sample is considered non-leakage. When the leak probability exceeds the set threshold, the monitoring framework automatically determines that a pipeline leak has occurred and initiates the second phase of assessment.

[0095] The second stage of the leakage event prediction model focuses on evaluating the leakage level corresponding to the leakage signal. The RCT-Net pre-trained in the first stage is retrained and fine-tuned using the converted log-mel spectrum of the leakage signal, enabling the model to capture the key frequency domain features corresponding to different leakage levels.

[0096] like Figure 7 As shown in the figure, the recognition results of the unpretrained RCT-Net were 0.9732, 0.9758, 0.9737, 0.9724, 0.9844, and 0.9980, respectively. All of these are greater than 0.97, which can meet the needs of on-site leakage level assessment. However, the training process of the unpretrained RCT-Net is longer (3000 rounds) and the convergence process is slower. The recognition results of the pretrained RCT-Net are 0.9837, 0.9862, 0.9851, 0.9906, 0.9845, and 0.9994, respectively. Its training process is shorter (1500 rounds) and its evaluation indicators are higher, averaging approximately 1% higher than those of the unpretrained RCT-Net.

[0097] In summary, the method provided in this embodiment has high accuracy and stability for acoustic leak detection in water distribution networks, can more effectively capture the local and global frequency domain characteristics of pipeline acoustic signals, and provide more accurate leakage probability estimation and leakage level assessment. At the same time, the monitoring framework couples residual convolution with Transformer, so that the advantages of the two complement each other, and realize the rapid classification of on-site pipeline acoustic signals and leakage information mining. When a leak occurs in the water distribution network, it can provide leakage information more efficiently and comprehensively, providing a new perspective for leak management priorities, reducing water losses and guiding water departments towards sustainable development.

[0098] In addition, the terms "upper", "lower", "inner", "outer", "front", and "back" are used for descriptive purposes only and should not be understood as indicating or implying relative importance. Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the present invention.

[0099] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of implementation of the present invention. Any equivalent changes or modifications made based on the structure, features and principles described in the scope of the patent application of the present invention should be included in the scope of the patent application of the present invention.

[0100] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-described embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A two-stage leakage monitoring method for a water distribution network based on acoustic signals, characterized in that: The following steps are involved: Acquire historical data, which includes on-site acoustic signals of leakage events and normal operation, and construct corresponding logarithmic Mel-frequency spectrograms frame by frame along the time axis based on the on-site acoustic signals; The logarithmic Mel-spectrogram is labeled with three leakage level assessments: mild leakage, moderate leakage, and severe leakage, and the on-site acoustic signal, the corresponding logarithmic Mel-spectrogram, and the labels are combined into a data set; A prediction network is constructed based on the coupled residual convolution and Transformer network framework. The prediction network consists of two stages, where the first stage includes a data preprocessing module, a local feature extraction module, a global feature extraction module, and a feature fusion module. Phase II includes a leakage level assessment module; The data preprocessing module is used to convert the input live sound signal into a corresponding logarithmic Mel spectrum; The local feature extraction module is used to extract the local feature sequence of the spectrum of the logarithmic Mel spectrum graph; The global feature extraction module and the local feature extraction module output the long-range dependency between the spectral feature sequences to output the corresponding global feature vector; The feature fusion module is used to fuse the input spectrum feature sequence and the global feature vector to obtain a fused feature; The leakage level assessment module makes predictions based on the input fusion features to output assessment results, wherein the assessment results include a mild leakage assessment result, a moderate leakage assessment result, and a severe leakage assessment result; The first stage of the prediction network is trained using the dataset, and the corresponding leakage amount is introduced into the dataset as a label to train the second stage of the prediction network, so as to obtain a leakage event prediction model for leakage event assessment; The on-site acoustic signals of the pipelines in the area to be analyzed are input into the leakage event prediction model to obtain the prediction results of leakage events of each pipeline in the area.

2. The two-stage leakage monitoring method for water distribution network based on acoustic signals according to claim 1 is characterized in that: The generation process of the logarithmic Mel-spectrogram is as follows: The live sound signal is converted into power spectrum by combining fast Fourier transform and square function; A dot product operation is performed on the power spectrum on a plurality of Mel-level triangular filter banks to obtain a logarithmic Mel-level spectrum map of each frame.

3. The two-stage leakage monitoring method for water distribution network based on acoustic signals according to claim 1 is characterized in that: The data preprocessing module includes a Filter Bank filter.

4. The two-stage leakage monitoring method for water distribution network based on acoustic signals according to claim 1 is characterized in that: The local feature extraction module uses a CNN with a residual block to perform multi-level local feature extraction on the logarithmic spectrum. The residual block includes two 3x3 convolutional layers with the same number of output channels. The output end of each convolutional layer is connected to a batch normalization layer and a ReLU activation function.

5. The two-stage leakage monitoring method for water distribution network based on acoustic signals according to claim 1 is characterized in that: The global feature extraction module includes an encoder, a stacked self-attention layer and a fully connected layer. The encoder performs layer normalization on the input spectral feature sequence to obtain frequency domain feature data, and performs weighted calculation on the frequency domain feature data based on the stacked self-attention mechanism and the fully connected layer to obtain a global feature vector.

6. The two-stage leakage monitoring method for water distribution network based on acoustic signals according to claim 1 is characterized in that: The fusion feature is the leakage probability of the pipeline, which is obtained by performing nonlinear fitting on the input local spectrum feature sequence and global feature vector through the sigmoid function of the fully connected layer.

7. The two-stage leakage monitoring method for water distribution network based on acoustic signals according to claim 1 is characterized in that: The label is expressed as follows: ; The unit of leakage is .

8. A two-stage leakage monitoring device for a water distribution network, characterized in that: This is achieved by the two-stage leakage monitoring method for a water distribution network based on acoustic signals as described in any one of claims 1 to 7.

Images