Intelligent leakage detection method based on sound signals of water supply pipeline

Abstract

The invention relates to an intelligent leakage detection method based on a water supply pipeline sound signal, and relates to the technical field of water supply pipe network leakage monitoring. The method comprises the following steps: acquiring a sound signal of a water supply pipeline, and extracting an acoustic feature vector of the sound signal; the sound signals are manually labeled, a sound sample set with labels is obtained, and the labels comprise a first label representing leakage and a second label representing normal; and constructing a leakage probability model, and training the leakage probability model through the sound sample set. And inputting pipeline sound signals collected in real time into the leakage probability model, and judging whether the water supply pipeline leaks or is normal according to the predicted leakage probability output by the leakage probability model. The method has the effect of improving the recognition efficiency and accuracy of water supply pipeline leakage state detection.

Description

Technical Field

[0001] This application relates to the technical field of water supply network leakage monitoring, and in particular to an intelligent leakage detection method based on sound signals from water supply pipelines. Background Technology

[0002] Leaks in urban underground water supply pipelines are a common problem in water supply systems worldwide. This problem not only leads to the loss of a large amount of precious water resources, causing incalculable waste, but also triggers a series of chain reactions, such as secondary disasters like road collapses, seriously threatening the safety of urban infrastructure and the lives and property of residents. At the same time, it also disrupts the stable operation of the water supply system, affecting residents' normal water use and causing numerous inconveniences to urban life.

[0003] Currently, existing methods for detecting leaks in water supply pipelines mainly include: The listening rod method, a relatively traditional detection method, relies heavily on the experience of the inspectors. Inspectors hold a listening rod, attach it to the pipe, and listen to the sound emanating from within to determine if a leak is present. However, this method is highly subjective; different inspectors, due to differences in experience, may interpret the same sound very differently. Furthermore, manual inspection is extremely inefficient, requiring inspectors to check each section individually, consuming significant time and manpower. More importantly, the listening rod method often fails to detect minute leaks or leaks deeply buried underground. In addition, this method is highly susceptible to environmental noise interference; in noisy urban environments, accurately judging the sound of pipes becomes exceptionally difficult.

[0004] Traditional acoustic monitoring methods have achieved a degree of automation in detection. They involve installing hydrophones or vibration sensors on pipes to collect sound signals, then using a threshold method to judge these signals; if the signal exceeds a set threshold, a leak is identified. While this method reduces human intervention and improves detection efficiency, its overly simplistic logic and lack of intelligent analysis capabilities make it prone to misinterpreting environmental noise such as passing vehicles or construction noise as pipe leak signals, leading to a persistently high false alarm rate. Summary of the Invention

[0005] To improve the efficiency and accuracy of identifying leakage conditions in water supply pipelines, this application provides an intelligent leakage detection method based on sound signals from water supply pipelines.

[0006] The intelligent leakage detection method based on sound signals from water supply pipelines provided in this application adopts the following technical solution: A smart leakage detection method based on sound signals from water supply pipelines includes the following steps: Acquire the sound signal from the water supply pipeline and extract the acoustic feature vector of the sound signal; The sound signals are manually labeled to obtain a set of labeled sound samples, the labels including a first label representing loss and a second label representing normal; A leakage probability model is constructed and trained using the sound sample set. The leakage probability model is used to output the predicted leakage probability. The training objective of the leakage probability model is to minimize the cost function, which is constructed based on the loss function. The loss function is used to measure the difference between the model's predicted leakage probability and the manually labeled tags. The predicted leakage probability is then optimized iteratively. The real-time collected pipe sound signal is input into the leakage probability model, and the predicted leakage probability output by the leakage probability model is used to determine whether the water supply pipe is leaking or normal.

[0007] By adopting the above technical solution, the leakage detection is transformed into a probability output, which effectively avoids the drawbacks of the traditional fixed threshold method in complex working conditions, such as high false alarm and false alarm rates. At the same time, the introduction of a cost function based on the expected economic loss allows the optimization objective to be quantified during the model training stage. After on-site deployment, the economic efficiency can be optimized without repeated parameter tuning, enabling unattended real-time monitoring and significantly improving the safety coverage and response time of the pipeline network.

[0008] Preferably, the expression for the leakage probability model is: ; Where x is the acoustic feature vector, and θ is the weight vector parameter of the leakage probability model. It is an activation function, and , This indicates the predicted probability of leakage.

[0009] By adopting the above technical solution and using the logistic regression model, on the one hand, the model is simple and has low computational complexity, making it very suitable for deployment on resource-constrained edge devices or real-time systems that require rapid response; on the other hand, the output is an intuitive predicted leakage probability value, which not only provides a continuous and precise measurement basis for subsequent threshold judgment, but also makes the model interpretable, that is, the weight size can reflect the importance of different acoustic signals to leakage judgment.

[0010] Preferably, it also includes setting a decision threshold v, if If it is determined to be a leak, the leak probability model outputs the first label; if If the result is normal, the leakage probability model outputs a second label.

[0011] By adopting the above technical solution, when facing different application scenarios, there is no need to retrain the model; adjustments can be made simply by adjusting the decision threshold, which greatly improves the practicality and adaptability of the method.

[0012] Preferably, the expression for the loss function is: ; in, For the i-th sample, This represents the acoustic feature vector in the i-th sample. for Corresponding to the manually labeled tags, It is the predicted leakage probability of the i-th sample by the leakage probability model.

[0013] By adopting the above technical solutions, the leakage probability model can be forced to quickly correct serious errors during training, thereby accelerating convergence and improving the calibration capability of the leakage probability model, so that its output probability value more realistically reflects the possibility of leakage.

[0014] Preferably, the expression for the cost function is: ; Where m is the sample size. This represents the acoustic feature vector in the i-th sample. for The labels that correspond to the manual annotations.

[0015] By adopting the above technical solution, the leakage probability model performs consistently well not only on a few samples but also on the entire data distribution, effectively avoiding overfitting of the leakage probability model to the training data. This is the key to ensuring high detection accuracy even on unknown real pipe sound signals.

[0016] Preferably, the leakage probability model is trained using a gradient descent algorithm to optimize the weight vector parameter θ, and the update formula for the weight vector parameter θ is: ; Where, θ j These are the weight parameters in the leakage probability model. It is the set learning rate.

[0017] By adopting the above technical solution, the learning rate can be adjusted. It allows for fine-tuning while running, eliminating the need to recall equipment and solving the pain point of traditional offline models' inability to adapt to environmental changes.

[0018] Preferably, the leakage probability model is implemented through multiple iterations, including the following steps: Initialize the weight vector parameter θ; The total cost of the leakage prediction model calculated by the cost function on the training set of the sound sample set based on the current weight vector parameters is calculated. Determine whether the total cost meets the convergence condition. If it does, terminate the iteration and output the current weight vector parameters. If it does not meet the condition, update the weight vector parameters, and then calculate the total cost of the leakage probability model on the entire training set based on the recalculated updated weight vector parameters, and proceed to the next iteration.

[0019] By adopting the above technical solution, it is ensured that the model can start from a random state and autonomously evolve to the optimal state through a data-driven approach.

[0020] Preferably, the convergence condition is that the absolute value of the difference between the total costs calculated in two adjacent iterations is less than a preset threshold.

[0021] By adopting the above technical solution, the subjective and arbitrary nature of human intervention during the training process is eliminated by using an objective convergence criterion based on the cost change threshold.

[0022] Preferably, the system also includes an intelligent detection step, which includes setting multi-level alarm thresholds and triggering different levels of early warning or alarm operations based on the threshold range into which the predicted leakage probability falls.

[0023] By adopting the above technical solutions, different levels of operation and maintenance response can be triggered based on the threshold range into which the predicted leakage probability falls, thereby optimizing the allocation of human and material resources and improving emergency management efficiency.

[0024] Preferably, the multi-level alarm thresholds include at least a warning threshold, an alarm threshold, and an emergency alarm threshold; when the predicted leakage probability falls within the warning threshold range, the system performs a log recording or background notification operation; when the predicted leakage probability falls within the alarm threshold range, the system triggers and pushes alarm information; when the predicted leakage probability falls within the emergency alarm threshold range, the system triggers the highest level alarm and initiates linked emergency operations.

[0025] By adopting the above technical solution, setting multi-level alarm thresholds including early warning, alarm, and emergency alarm, and triggering different levels of response operations based on the threshold range where the predicted leakage probability is located, refined hierarchical management of pipeline leakage risk is achieved.

[0026] In summary, this application includes at least one of the following beneficial technical effects: 1. By combining acoustic signals with a logistic regression probability model, the reliance on human listening experience is reduced, and the standardization and efficiency of detection are improved. 2. By designing loss and cost functions to optimize and train the model, the leakage probability model can learn the subtle features of leakage sounds more accurately, effectively distinguish leakage from normal states, and reduce false alarms and false negatives. 3. Through online detection processes and multi-level alarm threshold settings, leakage can be quickly identified and corresponding warnings can be issued in the early stages, saving time for timely repairs and reducing water waste. Attached Figure Description

[0027] Figure 1 This is an overall flowchart of an embodiment of this application; Figure 2 This is a flowchart of step S2 in the embodiments of this application; Figure 3 This is a flowchart of step S5 in the embodiments of this application; Detailed Implementation

[0028] The following combination Figures 1-3 This application will be described in further detail.

[0029] This application discloses an intelligent leakage detection method based on sound signals from water supply pipelines.

[0030] Reference Figure 1 A smart leakage detection method based on sound signals from water supply pipelines includes the following steps: S1: Acquire sound signals from the water supply pipeline; S2: The collected sound signals are manually labeled, and a sound sample set of the water supply pipeline is constructed; S3: Design a leakage probability model based on the sound sample set.

[0031] S4: Design the loss function.

[0032] S5: Construct a cost function based on the loss function and train a leakage probability model.

[0033] S6: Input the real-time audio signal into the leakage probability model for online intelligent detection and alarm.

[0034] Specifically, step S1 acquires the sound signal from the water supply pipe, including the following: By deploying an acoustic sensor network at key nodes in the pipeline, high-quality raw sound and vibration signals that accurately reflect the pipeline's operating status can be continuously and stably acquired. Specifically, high-sensitivity acoustic sensors are deployed at key monitoring points in the water supply pipeline (such as valves, connections, and areas prone to leakage). The acoustic sensors can be piezoelectric accelerometers or hydrophones, and are firmly coupled to the pipe wall or inserted into the water to ensure effective capture of specific frequency sound vibrations propagating in the pipe wall or water caused by leakage.

[0035] The data acquisition process continues at a preset sampling rate and duration. The sensor's built-in preprocessing circuitry performs initial analog filtering to suppress high-frequency noise. Subsequently, the sound signals acquired by the acoustic sensor are uploaded to a cloud server in real-time or periodically via wireless IoT. The data packets uploaded to the cloud server include key metadata such as the acoustic sensor's device ID, geographic coordinates, acquisition timestamp, and pipeline operating conditions (e.g., pressure, flow rate), thus forming the source data stream for all subsequent intelligent analysis and processing.

[0036] Reference Figure 2 Step S2 involves manually labeling the collected sound signals and constructing a sound sample set for the water supply pipeline, including the following steps: S21: Manually label the sound signals and assign missing or normal labels to different sound signals.

[0037] Professionals conduct the core manual annotation work through a dedicated interactive annotation platform. This platform synchronously displays the time-domain waveform, spectrogram, and Mel-spectrum for each audio segment, possibly supplemented by operating data such as pipeline pressure and flow rate for that period. Annotation experts (usually experienced leak detection engineers) listen through headphones and combine this with the visual spectrogram to make a comprehensive judgment, assigning labels according to a set of clear criteria: when a continuous and regular high-frequency "hissing" sound or a "clicking" sound synchronized with water flow pulsation, which are typical characteristics of leakage, is present, it is labeled with the first label 1; if it is a steady sound of flowing water, pump background noise, or a sound without abnormal characteristics, it is labeled with the second label 0.

[0038] S22: Verify the accuracy of missing or normal labels.

[0039] To ensure the objectivity and consistency of the labels, a multi-person independent annotation-arbitration verification mechanism is typically adopted. This means that the same audio segment is processed independently by multiple annotators, and in case of disagreement, a senior expert makes the final decision, or the audio is cross-validated with actual repair records. To address the inherent imbalance problem of missing samples being far fewer than normal samples, data augmentation techniques are used to expand the minority class samples. For example, slight time stretching is applied to the missing audio, controllable background noise is added, or frequency band fine-tuning is performed to simulate sound variations under different operating conditions, thereby enhancing the robustness of the missing probability model. Finally, all annotated samples, their corresponding multi-dimensional feature vectors (feature extraction can be performed at this step or before training the missing probability model), manual labels, and metadata are systematically organized and randomly divided into training, validation, and test sets in a ratio such as 70:15:15, forming a sound sample set that can directly drive the training of the missing probability model.

[0040] Step S3, design the leakage probability model, including the following: The expression for the probability model of water supply pipe leakage is:

[0041] in, Let x represent the predicted leakage probability, and be the multidimensional acoustic feature vector of the sound signal, that is, the multidimensional acoustic feature vector extracted from the sound signal of the water supply pipeline, which can be expressed as: ,in, Leakage acoustic characteristics can include time-domain statistics, frequency-domain characteristics, and time-frequency analysis features of the sound signal. θ represents the weight vector parameters of the leakage probability model, which can be expressed as a vector form. The weight vector parameter θ is a core parameter that the leakage probability model needs to learn from the data. It is a bias term, corresponding to a virtual feature with a value of 1. , The calculation essentially involves linearly combining all feature weights to obtain a decision score z that can distinguish between leakage and normal states.

[0042] Let be the activation function, denoted as The activation function is used to smoothly compress and map the decision score z to the (0,1) interval. The activation function curve is S-shaped, and when z is a large positive number (meaning the feature strongly indicates leakage), It approaches 1 infinitely; when z is a very large negative number (a strong characteristic indicating normality). It approaches 0 infinitely; when z=0, It equals 0.5. The output of the leakage probability model represents the predicted leakage probability of the pipeline being in a leaky state given a multi-dimensional acoustic feature vector x. At this point, It is a continuous probability value.

[0043] To make a clear binary judgment of leakage or normality, a decision threshold v is also set, and in this embodiment, the decision threshold v is set to 0.5. If If it is determined to be a leak, the leak probability model outputs the first label 1; if If the result is negative, it is considered normal, and the leakage probability model outputs the second label 0. That is:

[0044] In practical engineering deployments, the threshold can be adjusted according to specific scenarios to achieve different business objectives: lowering the decision threshold (e.g., setting it to 0.3) makes the leakage probability model more sensitive and improves the leakage detection rate, but may be accompanied by more false alarms; raising the decision threshold (e.g., setting it to 0.7) makes the leakage probability model more conservative and can effectively reduce false alarms, but may miss some minor or early-stage leakage. This flexible threshold adjustment mechanism allows the same trained leakage probability model to adapt to application scenarios with different stringency requirements, from water supply mains to community pipe networks.

[0045] Step S4: Design the loss function, including the following: The loss function used to measure the difference between the predicted leakage probability and the true label during the training process of the leakage probability model is set as follows:

[0046] in, For the i-th sample, This represents the acoustic feature vector in the i-th sample. for Corresponding to manually labeled loss tags (1 represents loss, 0 represents normal). This represents the predicted leakage probability of the i-th sample by the leakage probability model. The value of the loss function essentially measures the difference between the probability distribution predicted by the leakage probability model and the true deterministic distribution. The closer the predicted leakage probability is to the true label, the smaller this difference (i.e., the loss).

[0047] When the true label is leakage (y=1), the loss function simplifies to: At this point, the leakage probability predicted by the leakage probability model... The closer to 1, The smaller the value (approaching 0), the more accurate the prediction and the smaller the loss. Conversely, if the leakage probability model incorrectly gives a very low probability, for example... If -log(0.1) is a very large positive value, it means that the prediction is inaccurate.

[0048] When the true label is missing (y=0), the loss function simplifies to: At this point, the leakage probability predicted by the leakage probability model... The closer to 0, The larger the value (approaching 1), the more accurate the prediction and the smaller the loss. Conversely, if the leakage probability model incorrectly gives a very high probability, for example... If -log(0.9) is a very small positive value, it means that the prediction is inaccurate.

[0049] refer to Figure 3Step S5 involves constructing a cost function based on the loss function and training a leakage probability model, including the following steps: S51: Define the cost function.

[0050] The total cost of the entire training set is set to the average loss of all samples, and is defined as the cost function. And the cost function Designed as follows:

[0051] Where m is the total number of labeled samples in the training set, i.e.

[0052] Minimize the cost function using an optimization algorithm. This allows for iterative updates of the weight vector parameters θ, and by calculating the average value, it can be ensured that the update direction is based on the overall data trend.

[0053] The optimization algorithm formula is designed, and the optimization algorithm formula simultaneously updates all weight vector parameters θ of the leakage probability model. The specific expression is as follows:

[0054] Where, θ j is the weight parameter in the leakage probability model, which is iteratively updated during the training phase using the gradient descent algorithm until it converges to the optimal value. The learning rate is set manually to control the magnitude of each update step. If If it's too large, it will lead to excessively large update increments, eventually preventing convergence; if... If the value is too small, the update increment will be too small, and more iterations will be needed to converge.

[0055] For cost function After differentiation, we can obtain the specific partial derivatives:

[0056] Therefore, the formula for updating the weight vector parameter θ is:

[0057] This is used to obtain the weight vector parameters θ of the leakage probability model required by the leakage probability model, and to obtain the trained leakage probability model.

[0058] S52: Initialize the weight vector parameters.

[0059] The weight vector parameters are typically initialized as a random vector that is close to zero but not entirely zero (e.g., sampled from a normal distribution with a mean of 0 and a variance of 0.01).

[0060] S53: Calculate the total cost of the leakage probability model in the current iteration on the entire training set.

[0061] Using the weight vector parameters θᵗ of the current iteration round (denoted as round t), perform a one-time calculation on all m samples in the training set. Calculate the predicted value for each sample i. Then, based on the loss function and cost function, the total cost J(θ)ᵗ of the current leakage probability model on the entire training set is calculated.

[0062] S54: Convergence Judgment and Termination Decision.

[0063] After each cost J(θᵗ) is calculated, the algorithm must determine whether the optimization is complete, and convergence is determined by setting a threshold for cost changes. The absolute difference |J(θ) between the current round's cost and the previous round's cost is calculated. t - J(θ) t-1 If the calculated absolute difference is less than a preset threshold ε, then convergence is considered complete. The threshold ε is a very small positive number, such as e^(-ε / ε). -6 Once convergence is complete, the iteration loop is terminated immediately, and the current weight vector parameter θ is set. t Save as the optimal leakage probability model weight vector parameters θ*. If convergence is not complete, jump to S55.

[0064] S55: Update all weight vector parameters. Update all weight vector parameters using the optimization algorithm formula. After updating all weight vector parameters, obtain a set of weight vector parameters θ. t+1 , This is used to initiate the next iteration and jump to S53, using the weight vector parameter θ. t+1 A new cycle begins.

[0065] S6: Input the real-time audio signal into the leakage probability model for online intelligent detection and alarm.

[0066] Acoustic sensors deployed on the pipeline continuously collect sound signals and automatically extract multidimensional acoustic feature vectors. Subsequently, the multidimensional acoustic feature vectors are input into a trained leakage probability model, which outputs a real-time leakage prediction probability between 0 and 1.

[0067] Based on the leakage probability prediction output by the leakage probability model, and according to the prediction results output in step S3, the system performs online intelligent detection and alarm. To achieve refined response, multiple alarm thresholds are set. Specifically, a warning threshold (e.g., 0.5-0.7), an alarm threshold (e.g., 0.7-0.9), and an emergency alarm threshold (e.g., >0.9) are set. When the leakage probability is predicted... When the value falls within the warning threshold, the system may only log in the background or provide a notification in the management interface, without triggering an active alarm. This is used to monitor suspected early or minor leaks. When predicting the probability of leakage... When the value falls within the alarm threshold range, the system generates a standard alarm event and pushes it to the user's mobile phone or computer. This is based on the predicted leakage probability. When the alarm level is between the emergency alarm thresholds, the system immediately triggers the highest level alarm and may simultaneously initiate related operations, such as shutting down area valves or notifying the emergency response team.

[0068] The implementation principle of the intelligent leakage detection method based on sound signals of water supply pipelines in this application is as follows: Physical changes in the pipeline state are converted into analyzable signals using acoustic sensors. Next, a data foundation is established using manual annotation to establish the correspondence between sound features and leakage states. Then, the weight vector parameters of the leakage probability model are optimized by minimizing a cost function constructed based on a loss function, and the mapping relationship between sound features and leakage states is automatically learned and solidified from the data. Finally, the leakage probability model is deployed in a real-world scenario to achieve automatic analysis, probability prediction, and tiered alarm for real-time sound signals. This embodiment encodes traditional leak detection experience into a reusable algorithm model, thereby achieving automation, intelligence, and precision in the detection process.

[0069] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for intelligent leakage detection based on sound signals from water supply pipelines, characterized in that: Includes the following steps: Acquire the sound signal from the water supply pipeline and extract the acoustic feature vector of the sound signal; The sound signals are manually labeled to obtain a set of labeled sound samples, the labels including a first label representing loss and a second label representing normal; A leakage probability model is constructed and trained using the sound sample set. The leakage probability model is used to output the predicted leakage probability. The training objective of the leakage probability model is to minimize the cost function, which is constructed based on the loss function. The loss function is used to measure the difference between the model's predicted leakage probability and the manually labeled tags. The predicted leakage probability is then optimized iteratively. The real-time collected pipe sound signal is input into the leakage probability model, and the predicted leakage probability output by the leakage probability model is used to determine whether the water supply pipe is leaking or normal.

2. The intelligent leakage detection method based on sound signals from water supply pipelines according to claim 1, characterized in that: The expression for the leakage probability model is: ； Where x is the acoustic feature vector, and θ is the weight vector parameter of the leakage probability model. It is an activation function, and , This indicates the predicted probability of leakage.

3. The intelligent leakage detection method based on sound signals from water supply pipelines according to claim 2, characterized in that: It also includes setting a decision threshold v, if If it is determined to be a leak, the leak probability model outputs the first label; if If the result is normal, the leakage probability model outputs a second label.

4. The intelligent leakage detection method based on sound signals from water supply pipelines according to claim 1, characterized in that: The expression for the loss function is: ; in, For the i-th sample, This represents the acoustic feature vector in the i-th sample. for Corresponding to the manually labeled tags, It is the predicted leakage probability of the i-th sample by the leakage probability model.

5. The intelligent leakage detection method based on sound signals from water supply pipelines according to claim 4, characterized in that: The expression for the cost function is: ; Where m is the sample size. This represents the acoustic feature vector in the i-th sample. for The labels that correspond to the manual annotations.

6. The intelligent leakage detection method based on sound signals from water supply pipelines according to claim 5, characterized in that: The leakage probability model is trained using a gradient descent algorithm to optimize the weight vector parameter θ. The update formula for the weight vector parameter θ is as follows: ； Where, θ j These are the weight parameters in the leakage probability model. It is the set learning rate.

7. The intelligent leakage detection method based on sound signals from water supply pipelines according to claim 6, characterized in that: The leakage probability model is implemented through multiple iterations, including the following steps: Initialize the weight vector parameter θ; The total cost of the leakage prediction model calculated by the cost function on the training set of the sound sample set based on the current weight vector parameters is calculated. Determine whether the total cost meets the convergence condition. If it does, terminate the iteration and output the current weight vector parameters. If it does not meet the condition, update the weight vector parameters, and then calculate the total cost of the leakage probability model on the entire training set based on the recalculated updated weight vector parameters, and proceed to the next iteration.

8. The intelligent leakage detection method based on sound signals of water supply pipelines according to claim 7, characterized in that: The convergence condition is that the absolute value of the difference between the total costs calculated in two adjacent iterations is less than a preset threshold.

9. The intelligent leakage detection method based on sound signals from water supply pipelines according to claim 1, characterized in that: It also includes an intelligent detection step, which includes setting multi-level alarm thresholds and triggering different levels of early warning or alarm operations based on the threshold range into which the predicted leakage probability falls.

10. The intelligent leakage detection method based on sound signals from water supply pipelines according to claim 9, characterized in that: The multi-level alarm thresholds include at least a warning threshold, an alarm threshold, and an emergency alarm threshold; when the predicted leakage probability falls within the warning threshold range, the system performs a log recording or background notification operation; when the predicted leakage probability falls within the alarm threshold range, the system triggers and pushes alarm information; when the predicted leakage probability falls within the emergency alarm threshold range, the system triggers the highest-level alarm and initiates linked emergency operations.

Images