Water supply network acoustic leakage detection method based on enhanced semi-supervised model

Abstract

The invention discloses a water supply network sound leakage detection method based on an enhanced semi-supervised model, and the method comprises the steps: 1, obtaining a data set, and carrying out the weak enhancement operation and strong enhancement operation of data; 2, a prediction network is constructed based on a semi-supervised network framework, the prediction network comprises a marked learning branch and an unmarked learning branch which share weights, and image samples with labels in the weak enhancement sample set are input into the marked learning branch to output corresponding prediction distribution; and step 3, training the prediction network by using the weak enhancement sample set and the enhancement sample set to obtain a pipe network acoustic leakage monitoring model. According to the method provided by the invention, the identification precision and robustness of the acoustic leakage detection model can be ensured under the condition of limited labeled samples.

Description

Technical Field

[0001] The invention belongs to the field of water supply pipe network leakage detection, and in particular relates to a water supply pipe network acoustic leakage detection method based on an enhanced semi-supervised model. Background Art

[0002] The water supply pipeline system is an important part of urban infrastructure, and ensuring its efficient and reliable operation is crucial to public health and economic development. However, with the continuous advancement of urbanization, the scale and complexity of the water supply pipeline network are increasing, and the problem of pipeline leakage has become more and more serious. Pipeline leakage not only wastes precious water resources, but also may cause serious environmental pollution and economic losses. Therefore, the development of efficient, reliable and field-adaptable leakage detection methods is of great significance to ensure the stable operation of the water supply system.

[0003] In recent years, leak detection technology based on vibroacoustic signals has attracted widespread attention in academia and practical applications due to its high sensitivity and non-invasive characteristics. Vibroacoustic signals can reflect subtle changes in the state of water flow in the pipeline, thereby producing obvious characteristics when a leak occurs. However, the development of fully supervised leak detection models that rely solely on manually labeled vibroacoustic signal data is often very expensive, and it is difficult for a single water company to obtain sufficient labeled data. In addition, it is relatively easy to collect a large amount of unlabeled acoustic signal data on site, but it has never been effectively utilized.

[0004] Academic Literature Novel Leakage Detection and Localization Method Based on LineSpectrum Pair and Cubic Interpolation Search[J].WATER RESOURCES MANAGEMENT, 2020, 34(12):3895-911. A new leak detection and localization method based on line spectrum pair and random forest is proposed, which has high accuracy and strong anti-noise ability, but there may be problems of high computational complexity in complex pipe networks.

[0005] Patent document CN118757697A discloses an intelligent detection system and method for water supply network leakage, wherein the system part includes: an ultrasonic module, a preprocessing module, an extraction module and an identification module; the ultrasonic module is used to transmit ultrasonic waves to the water supply network and collect waveform data of the water supply network; the preprocessing module is used to preprocess the waveform data to obtain processed data; the extraction module is used to extract features from the processed data to obtain sound wave features; the identification module is used to perform cluster analysis on the sound wave features to identify leakage in the water supply network. Summary of the invention

[0006] The purpose of the present invention is to provide a water supply network acoustic leakage detection method based on an enhanced semi-supervised model, which can ensure the recognition accuracy and robustness of the acoustic leakage detection model under the condition of limited labeled samples.

[0007] In order to achieve the purpose of the present invention, the following technical solution is provided: a method for detecting acoustic leakage in a water supply network based on an enhanced semi-supervised model, comprising the following steps:

[0008] Step 1: Obtain the on-site acoustic signal of the water supply network and convert it into a corresponding logarithmic Mel spectrum through a filter, select part of the logarithmic Mel spectrum to label whether there is leakage in the water supply network, so as to obtain labeled image samples and unlabeled image samples;

[0009] Perform a weak enhancement operation on all image samples to obtain a corresponding weak enhancement sample set;

[0010] Perform strong enhancement operations on unlabeled image samples to obtain corresponding enhanced sample sets;

[0011] Step 2: construct a prediction network based on a semi-supervised network framework, wherein the prediction network includes a weight-shared labeled learning branch and an unlabeled learning branch, and input the labeled image samples in the weak enhancement sample set into the labeled learning branch to output the corresponding prediction distribution;

[0012] The unlabeled image samples and the enhanced sample set in the weakly enhanced sample set are respectively input into the unlabeled learning branch to output a first prediction result corresponding to the unlabeled image samples in the weakly enhanced sample set and a second prediction result corresponding to the enhanced sample set, a pseudo-label is assigned to the unlabeled image samples corresponding to the prediction results with a confidence level higher than an adaptive threshold in the first prediction result, and the unlabeled image samples with the pseudo-label are added to the labeled learning branch to update the prediction distribution, and the adaptive threshold is gradually reduced by a fixed value according to the number of updates;

[0013] Step 3: Use the weakly enhanced sample set and the enhanced sample set to train the prediction network to obtain a pipe network acoustic leakage monitoring model.

[0014] The present invention makes full use of unlabeled data and acoustic data enhancement, and is based on a semi-supervised network framework to reduce the dependence on labeled data, thereby ensuring the accuracy and reliability of acoustic leakage detection under limited labeled data conditions.

[0015] Specifically, the weak enhancement operation includes one or more of brightness adjustment, horizontal flipping or contrast adjustment on the image sample.

[0016] Specifically, the strong enhancement operation uses SpecAugment to adjust the image samples from both time domain and frequency domain perspectives.

[0017] Specifically, the adjustment includes one or more of time distortion based on the time domain, time masking based on the time domain, or frequency masking based on the frequency domain.

[0018] Specifically, the prediction network adopts a three-layer residual convolutional neural network, and the residual block expression in the three-layer residual convolutional neural network is as follows:

[0019]

[0020] Y n =X n +F(X n ,W n )

[0021]

[0022] Among them, X n and X n+1 are the input and output of the residual block, F is the residual function, f is the ReLU activation function, and X N represents the features learned in the residual block, W i is the weight coefficient vector of the residual block.

[0023] Specifically, the calculation formula of the pseudo label is as follows:

[0024]

[0025] Among them, I is the indicator function; q = p(y|x) is the probability vector predicted by the model; H is the cross entropy function; pse = argmax(q) is the pseudo label.

[0026] Specifically, during the training process, a consistency regularization method is used for the unlabeled learning branch to ensure that the first prediction result and the second prediction result are consistent in form, and the consistency regularization method is only executed when the confidence of the prediction result is greater than the adaptive threshold.

[0027] Specifically, the comprehensive loss function is used to jointly guide the training of the labeled learning branch and the unlabeled learning branch.

[0028] Specifically, the expression of the comprehensive loss function is as follows:

[0029] L=L S +λ u ×L u

[0030]

[0031] Among them, L S is the loss function of the label learning branch; L u is the loss function of the unlabeled learning branch; λ u is the weight coefficient, m is the amount of data in a batch, y i is the true label, f(x i ) is the predicted value, k is the number of categories; pse ij is the pseudo label after weak enhancement; is the probability vector after strong enhancement.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] By performing weak and strong data enhancement operations on datasets containing a small number of labels and training models based on a semi-supervised model framework, ESSM can significantly improve the accuracy and reliability of acoustic leak detection when there is limited labeled data, and better support the optimized operation and reliability management of water supply networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A flow chart of a water supply network acoustic leakage detection method based on an enhanced semi-supervised model provided in this embodiment;

[0035] Figure 2 The weakly enhanced logarithmic Mel spectrum provided in this embodiment;

[0036] Figure 3 The logarithmic Mel spectrum diagram before and after strong enhancement provided in this embodiment;

[0037] Figure 4 A schematic diagram of the structure of the semi-supervised network framework provided for this embodiment;

[0038] Figure 5 A schematic diagram of a residual convolutional neural network provided for this embodiment;

[0039] Figure 6 The enhanced spectrograms of Cutout data of different sizes provided in this embodiment;

[0040] Figure 7 The impact of different data enhancement methods provided in this embodiment on the semi-supervised network framework. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme 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 here 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 claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in 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.

[0042] This embodiment proposes a method for detecting acoustic leakage in water supply networks based on an enhanced semi-supervised model, while taking into account reducing the reliance on labeled data, making full use of unlabeled data, and acoustic data enhancement. By effectively utilizing unlabeled acoustic data, ESSM can significantly improve the accuracy and reliability of acoustic leakage detection when there is limited labeled data, and can even exceed the fully supervised model, which can better support the optimized operation and reliability management of water supply networks.

[0043] like Figure 1 As shown, the water supply network acoustic leakage detection method based on the enhanced semi-supervised model provided in this embodiment has the following specific steps:

[0044] Step 1: Obtain the on-site acoustic signal of the water supply network and convert it into a corresponding logarithmic Mel spectrum through a filter, select part of the logarithmic Mel spectrum to label whether there is leakage in the water supply network, so as to obtain labeled image samples and unlabeled image samples;

[0045] Perform a weak enhancement operation on all image samples to obtain a corresponding weak enhancement sample set;

[0046] A strong enhancement operation is performed on the unlabeled image samples to obtain a corresponding enhanced sample set.

[0047] More specifically, the acoustic signal data set in this embodiment comes from the water supply networks of multiple southwestern cities in China, including a total of 17,602 audio signals, including 8,801 leakage signals and 8,801 non-leakage signals, and each signal duration is 1 second. Since each audio data is manually labeled, there will be some inevitable errors. Each acoustic signal sample is converted into a logarithmic Mel spectrum by a filter bank to facilitate the basic model to identify its frequency domain characteristics. The parameters are shown in Table 1. In order to meet the needs of semi-supervised model training, 90% (including labeled and unlabeled) of the audio signals are selected for training the semi-supervised model, and 10% (labeled) of the data is used as a test set to verify the performance of the trained semi-supervised model.

[0048] Table 1

[0049]

[0050] Given that labeled acoustic signals in water supply networks are often scarce and expensive, while unlabeled acoustic signal data is relatively abundant but underutilized, data augmentation helps bridge the gap between the two. Weak augmentation usually perturbs the input spectrogram slightly without destroying its basic time-frequency characteristics.

[0051] like Figure 2 As shown, it is a weak enhancement operation demonstration diagram provided by this embodiment, wherein Figure 2 a in the figure is a schematic diagram of brightness adjustment. Figure 2 The b in the figure is a schematic diagram of horizontal flipping. Figure 2 Figure c is a schematic diagram of contrast adjustment.

[0052] Moreover, the above three data enhancement methods will not bring too much interference to the time-frequency characteristics of the logarithmic Mel-spectrogram.

[0053] like Figure 3 As shown, it is a strong enhancement operation demonstration diagram provided by this embodiment, wherein Figure 3 a in is the original logarithmic Mel spectrum. Figure 3 b is a schematic diagram of time distortion. Figure 3 c in the figure is a schematic diagram of time masking. Figure 3 The d in the figure is a schematic diagram of frequency masking. Figure 3 Figure e is a schematic diagram of the superposition of the first three methods.

[0054] like Figure 4As shown, the semi-supervised network framework provided by this embodiment is used in this embodiment to construct a prediction network. The prediction network adopts a three-layer residual convolutional neural network, which includes a labeled learning branch and an unlabeled learning branch. This dual-path strategy ensures the effective use of the rich information contained in the field labeled and unlabeled acoustic signals. First, a large number of field acoustic signals are collected, and some signals are manually labeled by experts after data processing. Second, in the labeled learning branch, we apply weak enhancement technology to generate weakly enhanced samples from the labeled acoustic signal data set. Then it is input into the basic model to generate a predicted distribution. In this process, a key step involves calculating and optimizing the supervised loss between the model prediction and the true label to ensure its effectiveness for the labeled data. Third, the unlabeled learning branch is used to process unlabeled acoustic signal samples. Since there is no real label, each spectrogram will be weakly and strongly enhanced at the same time. It is then input into the basic model to generate two different sets of predictions for calculating the consistency regularization loss (unsupervised loss). Fourth, in the unlabeled branch, in order to improve the quality and reliability of pseudo-label generation, an adaptive threshold strategy for dynamically assigning pseudo-labels is introduced, which mainly judges the prediction results after weak enhancement, and can more accurately utilize a large amount of unlabeled acoustic signal data. Fifth, the two training branches are jointly guided by the total loss function to reduce the overfitting of the prediction network to a small amount of labeled data, thereby improving the generalization ability and overall performance of the final model.

[0055] like Figure 5 As shown, it is the residual convolutional neural network provided by this embodiment. Each residual block in the residual convolutional neural network can be expressed as follows:

[0056] Y n =X n +F(X n ,W n ) (1)

[0057] X n+1 =f(Y n ) (2)

[0058] Among them, X n and X n+1 are the input and output of the residual block, F is the residual function, and f is the ReLU activation function. From (1) and (2), we can derive equation (3) to obtain the features learned in the residual block.

[0059]

[0060] In the labeled learning branch, the labeled logarithmic Mel-frequency spectrogram is also weakly enhanced before being input into the basic model for training, mainly to reduce overfitting in the supervised learning process. In addition, for the leakage detection problem, the loss function of the supervised part is obtained by calculating the cross entropy loss during the training process, which can be expressed as equation (4).

[0061]

[0062] Among them, L s is the supervised loss, m is the amount of data in a batch, y i is the true label, f(x i ) is the predicted value of the basic model.

[0063] In the unlabeled learning branch, a small amount of labeled acoustic signals on site can provide some effective information for the basic model during the training process. This information enables the model to generate a preliminary cognitive pattern of the acoustic signal. When the trained model is used to predict unlabeled data in the unsupervised branch, the prediction results with high confidence can be used as pseudo labels, thereby converting the unlabeled data into part of the labeled data. This can add additional information during the training process and improve the performance and generalization ability of the model. The pseudo label calculation formula is shown in Equation (5).

[0064]

[0065] Among them, I is the indicator function; q = p(y|x) is the probability vector predicted by the model; H is the cross entropy function; pse = argmax(q) is the pseudo label.

[0066] Traditional semi-supervised models set fixed thresholds when generating pseudo labels, which cannot be adjusted according to changes in training data and may perform poorly in different data sets or at different training stages. In this work, the threshold is dynamically adjusted based on the progress of training. First, a higher initial threshold is set to ensure the accuracy of pseudo labels in the early stage of basic model training. Then, as the basic model is improved, the threshold is gradually lowered, and more unlabeled acoustic signal data can be used for leakage detection.

[0067] This embodiment adopts a consistency regularization strategy to ensure that the basic model can maintain consistent prediction results after introducing different disturbances (weak enhancement and strong enhancement) to the unlabeled acoustic signal. Through consistency regularization, a large amount of unlabeled acoustic signal data can be effectively utilized to improve the generalization ability and stability of the semi-supervised acoustic leakage detection model. The calculation formula (6) of the consistency regularization loss in the unsupervised learning branch is as follows:

[0068]

[0069] Among them, L uis the consistency regularization loss of the unsupervised branch (i.e., unsupervised loss); m is the total number of samples; k is the number of categories; pse ij is the pseudo label after weak enhancement; is the probability vector predicted by the strongly enhanced base model.

[0070] It is worth noting that not every sample will be calculated for unsupervised loss after weak and strong enhancement. Only when the confidence of the sample predicted by the basic model after weak enhancement is greater than the set adaptive threshold, its consistency regularization loss (i.e., unsupervised loss) will be calculated.

[0071] By jointly training the labeled learning branch and the unlabeled branch, the potential information in the limited labeled acoustic signals and the unlabeled acoustic signals can be more effectively utilized. From equations (4) and (6), the overall optimization objective function of ESSM is shown in equation (7).

[0072] L=L S +λ u ×L u (7)

[0073] Among them, L S is the loss function of the label learning branch; L u is the loss function of the unlabeled learning branch; λ u is the weight coefficient.

[0074] In order to prove the superiority of the proposed pipe network acoustic leakage monitoring model (ESSM model) trained in this embodiment, the commonly used semi-supervised model and fully supervised model are compared under the same annotation rate. As can be seen from Table 2, the fully supervised model shows extremely high accuracy and specificity when using 100% annotated data, reaching 98.30% and 98.61% respectively. This shows that with the support of sufficient annotated data, fully supervised learning can maximize model performance. However, when the proportion of annotated data is reduced to 20% or less, the performance of the fully supervised model decreases significantly. In particular, when there is only 5% annotated data, the accuracy and drop to 88.80% and 87.83% respectively. In contrast, semi-supervised models such as ESSM and Fixmatch perform better when there is less annotated data, and the model deviation in multiple cross-validations is small, which proves the application potential and advantages of semi-supervised models in acoustic leakage detection compared with fully supervised models. Especially ESSM, its performance is close to or even exceeds the performance of the fully supervised model under high annotated data conditions.

[0075] ESSM shows excellent performance at 5%, 10% and 20% labeled data ratios, outperforming other semi-supervised models. Especially at 20% label rate, its accuracy, precision, F1-score, sensitivity and specificity all reach the highest, which are 98.45%, 98.46%, 98.45%, 98.32% and 98.59% respectively. In addition, when the label rate is 20%, the average deviation of each evaluation index of ESSM is the smallest, which further illustrates its strong stability in leakage detection.

[0076] Fixmatch also performs well at 10% and 20% labeled data ratios, especially in terms of precision and F1-score. This is due to its strong and weak data enhancement and consistency regularization, which effectively utilizes unlabeled data to improve model performance. Mean Teacher's performance is close to Fixmatch, especially at 20% label rate. SimCLR performs slightly worse at low labeled data ratios, especially at 5% labeled data, with its accuracy and sensitivity of only 92.16% and 92.01%. This may be because SimCLR relies on contrastive learning, and the learning efficiency of feature representation is limited when there is less labeled data.

[0077] Table 2 Comparison of recognition results of each model

[0078]

[0079] In order to prove the superiority of SpecAugment, a data augmentation method in the ESSM model, a comparative analysis was conducted between the Cutout method and the weak augmentation method (i.e., replacing the strong augmentation of ESSM with weak augmentation).

[0080] like Figure 6 As shown in Figure 2, the logarithmic Mel spectrum is enhanced by Cutout of different sizes. Figure 6 a in is Cutout-10, Figure 6 b in is Cutout-15, Figure 6 c in the figure is Cutout-20. As the size of Cutout increases, the masked area on the logarithmic Mel-spectrogram becomes larger and larger, and the time-frequency features lost in the spectrogram after data enhancement also increase.

[0081] like Figure 7As described above, the impact of different data augmentation methods on ESSM identification of pipeline leakage under three label rates shows that in the case of 5% labeled data, SpecAugment's accuracy (94.24%), precision (94.26%), F1-score (94.12%), sensitivity (94.80%), and specificity (93.68%) are all higher than Cutout and weak enhancement. This shows that when there is less labeled data, SpecAugment is most effective and can better improve the various indicators of the model. SpecAugment enhances the diversity of data through techniques such as time domain masking, frequency domain masking, and time distortion, which not only enables the model to learn features at different times and frequencies, but also enhances the model's robustness to timing and frequency changes, and more comprehensively improves the model's generalization ability.

[0082] In contrast, the masking technique in the Cutout method mainly randomly masks certain areas of the input image or feature map, and the mask size is set to (10, 15, 20). Although Cutout increases data diversity by masking part of the image area, due to the fixed mask size, it cannot flexibly mask different areas in the time domain and frequency domain like SpecAugment, nor can it achieve time warping.

[0083] When the amount of labeled data increases to 10%, ESSM still performs best, especially in accuracy and sensitivity, leading Cutout and weak enhancement by about 1 percentage point respectively. This shows that SpecAugment not only has a significant advantage when there is less data, but also maintains its leading position even when the amount of labeled data increases. When the amount of labeled data reaches 20%, the performance of ESSM is further improved. In contrast, although Cutout and weak enhancement also improve at this annotation ratio, they still lag behind SpecAugment in terms of F1-score, sensitivity, and specificity.

[0084] In water supply networks, the cost of collecting unlabeled data is relatively small compared to labeled acoustic signal data. Therefore, it is meaningful to explore the impact of the amount of unlabeled data on improving the accuracy of ESSM leak detection while keeping the amount of labeled data unchanged (10%).

[0085] Table 3

[0086]

[0087] As shown in Table 3, keeping the proportion of labeled data unchanged and gradually increasing the amount of unlabeled data, the performance of ESSM is continuously improved. As the amount of unlabeled data increases from (0-80%), the recognition accuracy of ESSM is improved by about 5%, which shows that unlabeled data has a significant contribution to the accuracy improvement of the semi-supervised model. In addition, when the unlabeled data is 40%, the accuracy of ESSM has increased by about 4% compared with the fully supervised model. However, when the unlabeled data is further increased from 40% to 80%, the various indicators of ESSM have only increased by about 1%. The results show that the improvement of the accuracy of ESSM with the number of unlabeled data is not unlimited. It may be that with the continuous increase in the number of unlabeled data, ESSM's grasp of the distribution and time-frequency characteristics of acoustic signal data has gradually become sufficient, and the marginal contribution of the newly added unlabeled data to the model has gradually decreased.

[0088] In summary, acoustic leak detection based on deep learning plays an irreplaceable role in the optimized operation and reliability management of water supply networks. However, the limited scale of labeled acoustic signals on site restricts the recognition accuracy and robustness of the acoustic leak detection model, and a large number of unlabeled acoustic signals are not effectively utilized. The present invention only requires a small amount of labeled data to carry out acoustic leak detection in water supply networks, and has high generalization, accuracy and robustness. At the same time, it can greatly reduce the high acoustic signal labeling costs of water companies and improve the utilization efficiency of on-site data, which is of great significance for enhancing the stable and efficient operation of water supply networks and reducing resource consumption. Even when the same number of labeled acoustic signals are collected on site, the semi-supervised model modeling using the present invention can obtain more robust and accurate recognition results compared to the fully supervised model.

[0089] In addition, the terms "upper", "lower", "inner", "outer", "front", "rear" 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.

[0090] 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. All equivalent changes or modifications made according to the structure, characteristics and principles described in the patent application scope of the present invention should be included in the patent application scope of the present invention.

[0091] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A water supply network acoustic leakage detection method based on enhanced semi-supervised model, characterized in that: The following steps are involved: Step 1: Obtain the on-site acoustic signal of the water supply network and convert it into a corresponding logarithmic Mel spectrum through a filter, select part of the logarithmic Mel spectrum to label whether there is leakage in the water supply network, so as to obtain labeled image samples and unlabeled image samples; Perform a weak enhancement operation on all image samples to obtain a corresponding weak enhancement sample set; Perform strong enhancement operations on unlabeled image samples to obtain corresponding enhanced sample sets; Step 2: construct a prediction network based on a semi-supervised network framework, wherein the prediction network includes a weight-shared labeled learning branch and an unlabeled learning branch, and input the labeled image samples in the weak enhancement sample set into the labeled learning branch to output the corresponding prediction distribution; The unlabeled image samples and the enhanced sample set in the weakly enhanced sample set are respectively input into the unlabeled learning branch to output a first prediction result corresponding to the unlabeled image samples in the weakly enhanced sample set and a second prediction result corresponding to the enhanced sample set, a pseudo-label is assigned to the unlabeled image samples corresponding to the prediction results with a confidence level higher than an adaptive threshold in the first prediction result, and the unlabeled image samples with the pseudo-label are added to the labeled learning branch to update the prediction distribution, and the adaptive threshold is gradually reduced by a fixed value according to the number of updates; Step 3: Use the weakly enhanced sample set and the enhanced sample set to train the prediction network to obtain a pipe network acoustic leakage monitoring model.

2. The water supply network acoustic leakage detection method based on enhanced semi-supervised model according to claim 1 is characterized in that: The weak enhancement operation includes one or more of brightness adjustment, horizontal flipping or contrast adjustment on the image sample.

3. The water supply network acoustic leakage detection method based on enhanced semi-supervised model according to claim 1 is characterized in that: The strong enhancement operation uses SpecAugment to adjust the image samples from both time domain and frequency domain perspectives.

4. The water supply network acoustic leakage detection method based on enhanced semi-supervised model according to claim 3 is characterized in that: The adjustment includes one or more of time warping based on the time domain, time masking based on the time domain, or frequency masking based on the frequency domain.

5. The water supply network acoustic leakage detection method based on enhanced semi-supervised model according to claim 1 is characterized in that: The prediction network adopts a three-layer residual convolutional neural network, and the residual block expression in the three-layer residual convolutional neural network is as follows: Y n =X n +F(X n ,W n ) Among them, X n and X n+1 are the input and output of the residual block, F is the residual function, f is the ReLU activation function, and X N represents the features learned in the residual block, W i is the weight coefficient vector of the residual block.

6. The water supply network acoustic leakage detection method based on enhanced semi-supervised model according to claim 1 is characterized in that: The calculation formula of the pseudo label is as follows: Among them, I is the indicator function; q = p(y|x) is the probability vector predicted by the model; H is the cross entropy function; pse = argmax(q) is the pseudo label.

7. The water supply network acoustic leakage detection method based on enhanced semi-supervised model according to claim 1 is characterized in that: During the training process, a consistency regularization method is used for the unlabeled learning branch to ensure that the first prediction result and the second prediction result are consistent in form. The consistency regularization method is only executed when the confidence of the prediction result is greater than the adaptive threshold.

8. The water supply network acoustic leakage detection method based on enhanced semi-supervised model according to claim 1 is characterized in that: A comprehensive loss function is used to jointly guide the training of the labeled learning branch and the unlabeled learning branch.

9. The water supply network acoustic leakage detection method based on enhanced semi-supervised model according to claim 8 is characterized in that: The expression of the comprehensive loss function is as follows: L=L S +λ u ×L u Among them, L S is the loss function of the label learning branch; L u is the loss function of the unlabeled learning branch; λ u is the weight coefficient, m is the amount of data in a batch, y i is the true label, f(x i ) is the predicted value, k is the number of categories; pse ij is the pseudo label after weak enhancement; is the probability vector after strong enhancement.

Images