Water supply network sound signal label cleaning method based on convolution confidence learning

Abstract

The invention discloses a water supply network sound signal label cleaning method based on convolution confidence learning, and the method comprises the steps: obtaining sound signals in a water supply network, carrying out the label labeling of whether the sound signals are leakage or not in a manual mode, and enabling the sound signal data and labels to form a first data set; constructing an initial model based on convolution confidence learning, and training the initial model by using the first data set to obtain a CCL model used for predicting whether leakage exists or not; constructing a counting matrix and carrying out standardization processing to output a joint distribution matrix of noise tags and real tags; and cleaning the joint distribution matrix according to a label cleaning strategy to obtain a final high-quality second data set. According to the method provided by the invention, when confidence learning is applied to acoustic leakage detection, the method cannot only depend on an unsupervised or pure statistical cleaning strategy, and a supervised learning mechanism needs to be introduced to more reliably identify and correct an error tag.

Description

Technical Field

[0001] This invention belongs to the field of water supply network detection technology, and particularly relates to a method for cleaning acoustic signal tags in water supply networks based on convolutional confidence learning. Background Technology

[0002] Water distribution networks (WDNs), as a core component of urban lifeline engineering, are crucial for ensuring people's livelihoods and promoting economic development through their safe and efficient operation. However, factors such as network aging and changes in external loads lead to frequent pipeline leaks, causing a series of problems including significant water waste, infrastructure damage, and environmental pollution. Therefore, developing efficient and accurate pipeline leak detection technologies has significant economic, social, and environmental implications.

[0003] Acoustic leak detection technology has become a hot topic in current leak detection technology research and application due to its advantages such as non-invasiveness and wide applicability. This technology captures specific acoustic wave signals generated by leaks by deploying sensors on key nodes of pipelines (such as valves and fire hydrants) or pipeline walls, and uses signal processing and machine learning (ML) algorithms to identify leak characteristics and locate leak areas. However, building a high-performance and high-precision ML-based acoustic leak detection model still faces a fundamental and severe challenge: the noise problem of the labels of the training dataset, that is, noise labels and the uncertainty they bring. At present, there are two main ways to collect acoustic signal data on site: (1) collecting noise data uploaded by accelerometers and noise recorders. When an anomaly is detected, it is necessary to excavate and verify the specific pipeline on site to determine the label of the acoustic signal (leak or no leak), so the labeling process is time-consuming and laborious. Moreover, under normal circumstances, anomaly detection and on-site excavation verification are not carried out simultaneously. The third-party unit that performs the excavation may tend to report leaks due to cost, performance evaluation and other reasons, which can easily lead to labeling errors and the generation of noise labels. (2) Another approach is to directly collect acoustic signals (especially leak-free signals) from the noise monitoring platform and use the platform's detection results as labels without additional verification. This labeling method is very convenient, but due to limitations in the monitoring platform's recognition capabilities and interference from the on-site environment (traffic, construction, equipment operation, etc.), incorrect labels may appear, resulting in noise labels. Noise labels can cause the ML model to overfit the noise pattern rather than the actual leakage physical mechanism, leading to a higher false alarm rate when the model is applied in new scenarios.

[0004] Patent document CN118052326A discloses a method for predicting water consumption in a campus water supply network based on neural networks. Utilizing deep learning, this method predicts water consumption using a neural network-based approach. First, it collects pipeline characteristics and flow data under normal water supply conditions using sensors or monitoring points. Second, it cleans and filters the data, establishing a time series method. Then, based on the collected data, it constructs and trains a pipeline node flow prediction model based on a BP neural network. Finally, it obtains predicted values, enabling reasonable assessment and scheduling of water consumption. By comparing the predicted values ​​with the actual values, it further predicts pipeline leakage.

[0005] Patent document CN117033353A discloses a method for cleaning monitoring data of water supply networks. The method includes: assessing the data quality of water supply network monitoring data, identifying failed monitoring nodes and providing maintenance prompts; performing correlation analysis on the monitoring data of all nodes, extracting relevant nodes for each node and constructing a corresponding variable candidate library; establishing a data cleaning model based on BP neural network to realize abnormal data detection and repair; and establishing a data cleaning model based on LSTM network to realize data auxiliary cleaning. Summary of the Invention

[0006] The purpose of this invention is to provide a method for cleaning acoustic signal tags in water supply networks based on convolutional confidence learning. When applying confidence learning to acoustic leak detection, this method cannot rely solely on unsupervised or purely statistical cleaning strategies; it also needs to introduce a supervised learning mechanism to more reliably identify and correct erroneous tags.

[0007] To achieve the objectives of this invention, the following technical solution is provided, comprising the following steps: Acquire acoustic signals from the water supply network and manually label whether the acoustic signals indicate leakage. Combine the acoustic signal data and labels to form the first dataset. An initial model based on Convolutional Confident Learning (CCL) is constructed, and the initial model is trained using the first dataset to obtain a CCL model for predicting whether there is a leak. The CCL model calculates the predicted probability of the input acoustic signal through cross-validation, uses the average predicted probability of all acoustic signals as the corresponding confidence threshold, and classifies all acoustic signals using the confidence threshold to obtain the predicted label corresponding to the acoustic signal. Using the noise labels (artificially labeled labels) of the input acoustic signal as the vertical axis and the corresponding predicted labels (considered to be real labels) as the horizontal axis, a counting matrix is ​​constructed and standardized to output the joint distribution matrix of noise labels and real labels. The joint distribution matrix is ​​cleaned according to the label cleaning strategy to remove problematic labels of acoustic signals in the first dataset, so as to obtain the final high-quality second dataset.

[0008] This invention intelligently identifies and cleans up erroneous labels by adaptively estimating the joint distribution of noise labels and real labels. It aims to reduce the uncertainty of noise labels in the modeling of acoustic leakage detection in water supply networks and develop a higher-performance acoustic leakage detection model. This not only proves that label quality has a key impact on model performance, but also provides a reliable and scalable data cleanup solution for acoustic leakage detection.

[0009] Specifically, the acoustic signal includes noise signal and pipeline parameters of the water supply network.

[0010] Specifically, the expression for the cross-validation calculation is as follows: ; ; ; Among them, the acoustic signal dataset The total number of samples in the sample is The total number of sample categories is defined as That is, leakage or no leakage. It is a dataset The Middle One sample; It is a sample Category The model predicts the probability at that time, and Confidence threshold For category The average predicted probability of all samples in the sample; for the sample The real label is The highest predicted probability among all categories .

[0011] Specifically, the expression for the counting matrix is ​​as follows: ; in, This indicates manually marked noise labels. This represents the predicted label, where m represents the total number of categories. Indicates that the noise label is a category. And the predicted label is a category. The number of samples at that time.

[0012] Specifically, the expression for the joint distribution matrix is ​​as follows: ; in, Describe the joint distribution matrix. Represents a counting matrix. This indicates manually marked noise labels. Indicates the predicted label, This is expressed as the total number of samples, which is equal to... .

[0013] Specifically, the CCL model is constructed using a three-layer convolutional neural network, which includes three convolutional layers and pooling layers.

[0014] Specifically, a Dropout layer is added after the convolutional layer to prevent the model from overfitting.

[0015] Specifically, the signature cleaning strategy includes: Strategy 1: Select the index corresponding to the sample with the highest probability. Sound signals that do not match the noise label are discarded; Strategy 2: Eliminate acoustic signals from the off-diagonal sections of the counting matrix; Strategy 3: For the off-diagonal cells of the counting matrix, according to the maximum interval Sort and select Individual sound signals are eliminated; In It is a sample Category The model predicts the probability at that time. It is a sample Category The model predicts the probability at that time. It is the joint distribution matrix The manually labeled noise is categorized. And the predicted label is a category. When a value less than 1 is compared with the total number of samples. Multiplication results in a single numerical value; Strategy 4: For each category The samples are sorted according to the lowest predicted probability, and then selected. Individual sound signals are eliminated. It is the joint distribution matrix The manually labeled noise is categorized. And the predicted label is a category. For an element less than 1, the sum of all elements in its category is equal to the total number of samples. Multiplication results in a single numerical value.

[0016] Specifically, it also includes strategy five, which eliminates sound signals that simultaneously satisfy strategies three and four.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: Supervised convolutional confidence learning (CCL) is introduced into the modeling process for acoustic leakage detection in water supply networks, aiming to effectively identify and process noise labels and significantly reduce their impact on the uncertainty of model training. From the perspective of data quality, we will build a more reliable and robust data preprocessing and model training paradigm for acoustic leak detection in water supply networks, and promote the application of label cleaning technology from theoretical methods to multi-scenario applications in on-site pipeline networks. Attached Figure Description

[0018] Figure 1 This is a flowchart of the water supply network acoustic signal tag cleaning method based on convolutional confidence learning provided in this embodiment; Figure 2 This is a flowchart of the convolutional confidence learning provided in this embodiment; Figure 3 The Mel spectrum diagrams are for some problem samples in multiple pipeline networks provided in this embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] like Figure 1 The image shows the water supply network acoustic signal tag cleaning method based on convolutional confidence learning provided in this embodiment. Its specific steps include data processing, tag cleaning, and case analysis.

[0021] Data processing section: Manually labeled acoustic signal samples were collected from the pipeline network. These labels included those determined through excavation verification and those directly measured using the water company's monitoring platform. First, the acoustic signal dataset (each signal converted to a 65×65 Mel-ray spectrogram) was divided into a training set (to be cleaned) and a test set (for testing the cleaning effect) at an 8:2 ratio. Since the test set was used to test the label cleaning effect, it had to be manually inspected by experts and noisy labels removed to ensure absolute purity. The training set to be cleaned was then subjected to a five-fold cross-partitioning and further divided into training and validation sets at an 8:2 ratio for CCL training and validation.

[0022] Label cleaning process: like Figure 2 As shown, the process of convolutional confidence learning provided in this embodiment includes: (1) counting: estimating the joint distribution of noise labels and real labels; (2) cleaning: arranging the acoustic signal samples according to the confidence level, finding and filtering out problematic samples; (3) retraining: after filtering out problematic samples, retraining the model with relatively clean samples and verifying it.

[0023] Counting: In this embodiment, CCL does not directly improve the capabilities of the model architecture itself, but rather provides a better learning environment for the model by purifying the training data, thereby unleashing its inherent potential.

[0024] More specifically, in estimating the joint distribution of noise labels and true labels, sample labels in the dataset that have undergone initial labeling (verification labels after data mining or obtained from a monitoring platform) but may contain errors are defined as noise labels. .

[0025] Additionally, define the actual tag as That is, the true label is predicted and estimated by cross-validation of the CCL model.

[0026] If the noise label does not match the true label, it is defined as a problem label. (Acoustic signal dataset) The total number of samples in the sample is The total number of sample categories is defined as That is, leakage or no leakage.

[0027] To estimate the joint distribution of noise and real labels, we first perform cross-validation on the acoustic dataset based on the CCL model to calculate the predicted probability of the samples. .

[0028] The expression for cross-validation calculation is as follows: ; ; ; Among them, the acoustic signal dataset The total number of samples in the sample is The total number of sample categories is defined as That is, leakage or no leakage. It is a dataset The Middle One sample; It is a sample Category The model predicts the probability at that time, and Confidence threshold For category The average predicted probability of all samples in the sample; for the sample The real label is The highest predicted probability among all categories .

[0029] Then, calculate the counting matrix. The calculation formula is as follows: The manually labeled noise tags of the input sound signal are used as the vertical axis, and the corresponding predicted tags are used as the horizontal axis. ; in, This indicates manually marked noise labels. This represents the predicted label, where m represents the total number of categories. Indicates that the noise label is a category. And the predicted label is a category. The number of samples at that time.

[0030] Through standardized counting matrix The joint distribution matrix of noise labels and unknown true labels is obtained. Its expression is as follows: ; in, Describe the joint distribution matrix. Represents a counting matrix. This indicates artificially labeled noise. Indicates the predicted label, This is expressed as the total number of samples, which is equal to... .

[0031] In this embodiment, the CCL classifier employs a three-layer convolutional neural network. Its structure mainly includes three convolutional layers and pooling layers, with kernel sizes of 16, 32, and 64 respectively. Dropout layers are added after each convolutional layer to prevent overfitting, with parameters of 0.2, 0.2, and 0.5 respectively. The optimizer used is Adam, with a learning rate of 0.0001, a batch size of 64, and 500 training epochs.

[0032] Cleaning: Identify and filter out samples with problematic labels.

[0033] After obtaining the joint distribution of noise and true labels in the training set, appropriate label cleaning strategies can be adopted to reduce label uncertainty during model training. Different cleaning strategies correspond to different label cleaning thresholds and will exhibit different cleaning effects on the same dataset. Commonly used cleaning strategies can be categorized into the following five types: Strategy 1: Filtering The sample, i.e., the selection The index corresponding to the highest probability Samples that do not match the noise label.

[0034] Strategy 2: Filtering the counting matrix Samples in the off-diagonal section.

[0035] Strategy 3: For counting matrices Off-diagonal units, according to the maximum interval Sort, select Each sample is filtered.

[0036] Strategy 4: For each category The samples are sorted according to the lowest predicted probability, and then selected. Individual sound signals are eliminated.

[0037] Strategy 5: It requires meeting the screening conditions of both Strategy 3 and Strategy 4. This strategy usually filters out fewer but more confident noisy labels.

[0038] Retraining: After filtering out problem labels, the CCL model is retrained using the cleaned training set, validated using a clean test set, and compared with the performance of the uncleaned model.

[0039] To better illustrate the beneficial effects of the method provided in this embodiment, the recognition performance of the model is measured using the commonly used metrics in the field of acoustic leakage detection: accuracy, specificity, and sensitivity, as shown in Tables 1 and 2. Table 1 shows the confusion matrix, and Table 2 shows the performance evaluation metrics of the CCL model.

[0040] More specifically, this embodiment also collected 11,499 1-second acoustic signal data points from four real pipeline networks in China. The sampling frequency of the noise recorder was 8192 Hz, and the sensitivity was no less than 1200 pC / g. Each signal data sample had a manually labeled noise tag, as detailed in Table 3. Taking the real pipeline network A in Table 3 as an example, we will compare the changes in model performance after tag cleaning to illustrate the effectiveness of CCL in leak detection and tag uncertainty reduction research. The cleaning effect of CCL on pipeline network A is very significant. Before tag cleaning, the accuracy, sensitivity, and specificity of CCL on the independent test set were 92.99%, 91.05%, and 94.20%, respectively.

[0041] As shown in Table 4, after applying different cleaning strategies to perform label cleaning, CCL achieved different levels of excellent performance on the same independent test set. Strategy 2 achieved accuracy, sensitivity, and specificity of 98.25%, 97.83%, and 98.49%, respectively, representing performance improvements of 5.26%, 6.78%, and 4.29% compared to before cleaning. Furthermore, the standard deviations of the evaluation metrics for Strategy 2 were 0.51, 1.06, and 0.67, respectively, significantly lower than the label differences before cleaning, indicating that the label cleaning process also enhanced the stability and robustness of CCL. Strategy 1 treats all samples where the "model predicted label" and "noise label" are inconsistent as problem samples. As the baseline strategy, Strategy 1 significantly outperforms the situation before cleaning, but it is the worst performing strategy among all strategies. Strategy 3 tends to delete samples that are estimated to have the highest noise probability in the entire dataset. Strategy 4 tends to delete samples whose predicted probability within each class is lower than the class adaptive confidence threshold. All evaluation metrics for Strategy 4 are better than those for Strategy 3, suggesting that "the difference in built-in confidence across classes" may be a key factor in the acoustic leakage detection task. Considering the frequency differences between no-miss and leaked samples, Strategy 4, by taking into account inter-class differences and setting different confidence thresholds for no-miss and leaked samples, may outperform Strategy 3. Strategy 3's global perspective makes it less refined in handling inter-class differences, resulting in slightly lower performance. Furthermore, the accuracy of the model after cleaning with Strategy 5 is not very high, with accuracy, sensitivity, and specificity reaching 97.25%, 96.53%, and 97.69%, respectively. However, its standard deviations across multiple validations are the smallest among the five strategies, at 0.49, 0.95, and 0.60, respectively. This is because Strategy 5 combines the results of Strategy 3 and Strategy 4, removing samples only if they are simultaneously identified as problematic by both strategies. However, its metrics are still lower than Strategy 4, indicating that an overly conservative cleaning strategy may lead to insufficient cleaning. Furthermore, experiments were conducted on three independent real-world water supply networks (B, C, and D) to verify the generality and scalability of the proposed Convolutional Belief Learning (CCL) label cleaning method. All experiments followed the same procedure: the CCL method was used to clean the training data for each network, and then the performance of the model before and after cleaning was evaluated on independent clean test sets.

[0042] As shown in Table 5, the acoustic leakage detection performance was significantly and consistently improved after applying the CCL cleaning method in real pipe networks with different locations and acoustic environmental characteristics. This fully demonstrates that CCL is not dependent on the data distribution of a specific pipe network. Specifically, the model accuracy of the three pipe networks significantly improved after CCL cleaning. The improvements were: pipe network B (+3.81%), pipe network C (+3.01%), and pipe network D (+1.39%). This indicates that the CCL method effectively removed harmful noise samples from the training set, enabling the model to learn a more accurate mapping relationship between acoustic features and labels, thus making more correct overall judgments. The sensitivity index was significantly improved in all pipe networks, especially in pipe network B, where the sensitivity increased from 95.81% to 99.89%, and in pipe network D, it reached 99.29%. Sensitivity measures the model's ability to detect real leaks; therefore, the improvement in the sensitivity index directly means a significant reduction in the false negative rate. like Figure 3 The image shows the spectrum of some problematic samples selected after label cleaning in each pipeline dataset.

[0043] like Figure 3 As shown in (a), the ID 980 sample exhibits a clear continuous frequency and energy distribution around 500 Hz, thus clearly indicating a leakage sample. This is consistent with... Figure 3 The ID 1750 sample shown in (b) is... Figure 3 The judgment of the 1763 samples shown in (c) is the same.

[0044] Figure 3 As shown in (a), the ID 84 sample has an energy distribution in the low-frequency (around 20Hz) part, but the intensity is weak, and the CCL model is quite confident in predicting it as a sample with no omissions.

[0045] Figure 3 As shown in (b), the sample spectrograms of ID 54 all exhibit significant irregular and discontinuous energy distributions, most likely due to high-frequency noise interference. Therefore, the CCL model labels them as samples with no missing data. Figure 3As shown in (c), the spectrum of ID 227 sample also shows a discontinuous energy distribution, and therefore it is also judged as a sample without omissions containing environmental noise.

[0046] Furthermore, the terms "upper," "lower," "inner," "outer," "front," and "rear" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0047] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of the present invention. All equivalent changes or modifications made to the structure, features and principles described in the claims of the present invention should be included in the scope of the claims of the present invention.

[0048] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. 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 foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for water distribution network acoustic signal label cleaning based on convolution belief learning, characterized in that, Includes the following steps: Acquire acoustic signals from the water supply network and manually label whether the acoustic signals indicate leakage. Combine the acoustic signal data and labels to form the first dataset. Construct an initial model based on convolutional confidence learning, and train the initial model using the first dataset to obtain a CCL model for predicting whether it is a leak. The CCL model calculates the predicted probability of the input acoustic signal through cross-validation, uses the average predicted probability of all acoustic signals within the category as the corresponding confidence threshold, and classifies all acoustic signals using the confidence threshold to obtain the predicted label corresponding to the acoustic signal. Using the noise labels of the input acoustic signal as the vertical axis and the corresponding predicted labels as the horizontal axis, a counting matrix is ​​constructed and standardized to output the joint distribution matrix of noise labels and true labels. The joint distribution matrix is ​​cleaned according to the label cleaning strategy to remove problematic labels of acoustic signals in the first dataset, so as to obtain the final high-quality second dataset.

2. The convolutional belief learning based water distribution network acoustic signal label cleaning method of claim 1, wherein, The acoustic signal includes noise signal and pipeline parameters of the water supply network.

3. The method for cleaning acoustic signal tags in water supply networks based on convolutional confidence learning according to claim 1, characterized in that, The expression for cross-validation calculation is as follows: ； ； ； Among them, the acoustic signal dataset The total number of samples in the sample is The total number of sample categories is defined as That is, leakage or no leakage. It is a dataset The Middle One sample; It is a sample Category The model predicts the probability at that time, and Confidence threshold For category The average predicted probability of all samples in the sample; for the sample The real label is The highest predicted probability among all categories .

4. The method for cleaning acoustic signal tags in water supply networks based on convolutional confidence learning according to claim 1, characterized in that, The expression for the counting matrix is ​​as follows: ； in, This indicates manually marked noise labels. This represents the predicted label, where m represents the total number of categories. Indicates that the noise label is a category. And the predicted label is a category. The number of samples at that time.

5. The method for cleaning acoustic signal tags in water supply networks based on convolutional confidence learning according to claim 1, characterized in that, The expression for the joint distribution matrix is ​​as follows: ； in, Describe the joint distribution matrix. Represents a counting matrix. This indicates manually marked noise labels. Indicates the predicted label, This is expressed as the total number of samples, which is equal to... .

6. The method for cleaning acoustic signal tags in water supply networks based on convolutional confidence learning according to claim 1, characterized in that, The CCL model is constructed using a three-layer convolutional neural network, which includes three convolutional layers and pooling layers.

7. The method for cleaning acoustic signal tags in water supply networks based on convolutional confidence learning according to claim 6, characterized in that, Add a Dropout layer after the convolutional layer to prevent the model from overfitting.

8. The method for cleaning acoustic signal tags in water supply networks based on convolutional confidence learning according to claim 1, characterized in that, The signature cleaning strategy includes: Strategy 1: Select the index corresponding to the highest probability of the sample prediction Sound signals that do not match the noise label are discarded; Strategy 2: Eliminate acoustic signals from the off-diagonal sections of the counting matrix; Strategy 3: For the off-diagonal cells of the counting matrix, according to the maximum interval Sort and select Individual sound signals are eliminated; In It is a sample Category The model predicts the probability at that time. It is a sample Category The model predicts the probability at that time. It is the joint distribution matrix The manually labeled noise is categorized. And the predicted label is a category. When a value less than 1 is compared with the total number of samples. Multiplication results in a single numerical value; Strategy 4: For each category The samples are sorted according to the lowest predicted probability, and selected... Individual sound signals are eliminated. It is the joint distribution matrix The manually labeled noise is categorized. And the predicted label is a category. For an element less than 1, the sum of all elements in its category is equal to the total number of samples. Multiplication results in a single numerical value.

9. The method for cleaning acoustic signal tags in water supply networks based on convolutional confidence learning according to claim 8, characterized in that, It also includes strategy five, which eliminates acoustic signals that simultaneously satisfy strategy three and strategy four.

Images