A Leakage Diagnosis Technology for Industrial Groundwater Pipeline Networks Based on Intelligent Learning Models

Abstract

This invention relates to the field of intelligent diagnostic technology, and discloses a leakage diagnosis technology for industrial underground water pipe networks based on an intelligent learning model. The technology includes: cleaning the raw sound data of the underground water pipe network to obtain standard sound data; performing multi-scale feature analysis on the standard sound data to obtain leakage-labeled and non-leakage-labeled data of the underground water pipe network; constructing an intelligent diagnostic model for the underground water pipe network; inputting the standard sound data into the intelligent diagnostic model to obtain leakage status diagnosis results for the underground water pipe network; performing leakage verification on the leakage status diagnosis results to obtain leakage location information for the underground water pipe network; and integrating the leakage status diagnosis results and leakage location information into a leakage diagnosis report for the underground water pipe network. This invention can improve the efficiency of an industrial underground water pipe network leakage diagnosis technology based on an intelligent learning model.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology, and in particular to a technology for diagnosing leakage in industrial underground water pipe networks based on an intelligent learning model. Background Technology

[0002] Existing technologies have significant shortcomings in the raw data processing stage of industrial underground water pipeline network leakage diagnosis. They lack multi-dimensional purification solutions adapted to the acoustic characteristics of the pipeline network, making it difficult to effectively filter out abnormal information such as environmental interference and equipment noise. Furthermore, the noise reduction and compensation methods are limited, resulting in redundant components and biases remaining in the preprocessed data. This makes it impossible to provide accurate and standardized basic data support for subsequent feature analysis, directly affecting the effective extraction of leakage-related features.

[0003] Traditional leak diagnosis technologies lack the ability to comprehensively adapt to the temporal and multi-scale features of sound data, and fail to establish an effective feature weight allocation mechanism. This makes it difficult to focus on key leak features and mine deep data correlations, resulting in insufficient accuracy and sensitivity in leak status identification, and a tendency for missed or false diagnoses. Furthermore, the leak location process does not fully integrate the spatial topology of the pipeline network and historical leak patterns, relying solely on a single diagnostic result, leading to low location accuracy and an inability to quickly pinpoint the specific location of the leak. This fails to meet the practical needs of efficient leak detection and timely repair in industrial pipeline networks. Summary of the Invention

[0004] This invention provides a leakage diagnosis technology for industrial underground water pipe networks based on an intelligent learning model, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a leakage diagnosis technology for industrial underground water pipe networks based on an intelligent learning model, comprising:

[0006] S101. Perform data cleaning on the raw sound data of the underground water pipe network to obtain the standard sound data of the underground water pipe network;

[0007] S102. Perform multi-scale feature analysis on the standard sound data to obtain leakage label data and non-leakage label data of the underground water pipe network;

[0008] S103. Based on the original sound data, the leaked label data, and the non-leaked label data, an intelligent diagnostic model for the underground water pipe network is constructed using a bidirectional convolutional long short-term memory network combined with an attention mechanism.

[0009] S104. Input the standard sound data into the intelligent diagnostic model to obtain the leakage status diagnosis result of the underground water pipe network;

[0010] S105. Based on the spatial topology of the underground water pipe network and historical leakage logs, perform leakage verification on the leakage status diagnosis results to obtain leakage location information of the underground water pipe network.

[0011] S106. Integrate the leakage status diagnosis results and leakage location information into a leakage diagnosis report for the underground water pipe network.

[0012] In a preferred embodiment, the step of data cleaning the raw sound data of the underground water pipe network to obtain the standard sound data of the underground water pipe network includes:

[0013] Based on the acoustic characteristic benchmark parameters of the underground water pipe network, multi-dimensional anomaly screening is performed on the original sound data to obtain preliminary purified data of the original sound data.

[0014] Wavelet decomposition and noise reduction are performed on the preliminary purified data to obtain redundant data from the preliminary purified data.

[0015] Interpolation compensation is performed on the non-redundant data to obtain the standardized data of the non-redundant data;

[0016] The standard data is checked for continuity according to the timestamp order to obtain the standard sound data of the underground water pipe network.

[0017] In a preferred embodiment, the step of performing multi-scale feature analysis on the standard sound data to obtain leakage label data and non-leakage label data of the underground water pipe network includes:

[0018] The temporal statistical features in the standard sound data are used as the temporal feature set of the standard sound data.

[0019] The standard sound data is subjected to frequency domain energy distribution statistics to obtain the frequency domain feature set of the standard sound data;

[0020] The time-domain feature set and the frequency-domain feature set are fused in multiple dimensions to construct a multi-dimensional fused feature vector of the standard sound data.

[0021] Based on a pre-set expert knowledge rule base, pattern matching and classification are performed on the multi-dimensional fused feature vectors to obtain leakage label data and non-leakage label data of the underground water pipe network.

[0022] In a preferred embodiment, the step of constructing an intelligent diagnostic model for the underground water pipe network based on the original sound data, the leakage tag data, and the non-leakage tag data, using a bidirectional convolutional long short-term memory network combined with an attention mechanism, includes:

[0023] The original sound data, the leaked label data, and the non-leaking label data are sample-paired and integrated to obtain the label dataset of the underground water pipe network;

[0024] Perform balanced sampling on the labeled dataset to obtain the validation training dataset of the labeled dataset;

[0025] Based on the model architecture of convolutional neural networks, the hierarchical parameters and connection methods of the bidirectional convolutional long short-term memory network are initialized and configured to obtain the initial network framework of the underground water pipe network.

[0026] An adaptive attention mechanism is embedded in the initial network framework to construct a feature weight allocation benchmark, so as to obtain a preliminary diagnostic model of the groundwater pipe network.

[0027] Based on the validation training dataset, the preliminary diagnostic model is trained with strategy configuration to obtain the intelligent diagnostic model for the underground water pipe network.

[0028] In a preferred embodiment, the convolutional neural network-based model architecture initializes the hierarchical parameters and connection methods of the bidirectional convolutional long short-term memory network to obtain the initial network framework of the underground water pipe network, including:

[0029] Based on the model architecture of the convolutional neural network, the receptive field size and stride of the convolutional kernel of the model architecture are defined, and the number of feature maps output by the convolutional layers in the model architecture is determined to obtain the convolutional structure of the underground water pipe network.

[0030] The bidirectional long short-term memory module is initialized and configured to obtain the temporal learning structure of the underground water pipe network;

[0031] The output feature map sequence of the convolutional structure is used as the input of the temporal learning structure for connection configuration, thereby establishing the forward propagation path of data from feature extraction to temporal modeling;

[0032] Based on the forward propagation path, the convolutional structure and the temporal learning structure are integrated and connected to obtain the initial network framework of the underground water pipe network.

[0033] In a preferred embodiment, embedding an adaptive attention mechanism into the initial network framework to construct a feature weight allocation benchmark to obtain a preliminary diagnostic model of the groundwater pipeline network includes:

[0034] Attention parameter matrices are added at predetermined positions in the initial network framework to enable training and learning from the data.

[0035] The feature sequence output by the temporal learning structure is input into the attention parameter matrix;

[0036] The attention weights of the input feature sequence are obtained by calculating the correlation between the feature vectors at different time steps in the input feature sequence using the attention parameter matrix.

[0037] Based on the attention weights, the feature vectors at each time step in the input feature sequence are weighted and fused to obtain the context feature vector of the input feature sequence.

[0038] The context feature vectors are then fed into the subsequent processing layer of the initial network framework to complete the embedding of the adaptive attention mechanism, thereby obtaining a preliminary diagnostic model of the underground water pipe network.

[0039] In a preferred embodiment, the attention weight is calculated using the following formula:

[0040] ;

[0041] In the formula, For the first The feature vector at the i-th time step is related to the i-th time step. The attention weights of the feature vectors at each time step For the first The time step and the first Relevance score at each time step The total number of time steps for the input feature sequence. For the first The feature vector at the i-th time step is related to the i-th time step. The correlation score of the feature vectors at each time step It is a natural exponential function. This is a temporary index variable used for summation.

[0042] In a preferred embodiment, the step of inputting the standard sound data into the intelligent diagnostic model to obtain the leakage status diagnostic result of the underground water pipe network includes:

[0043] Temporal features are extracted from the standard sound data to construct a temporal feature sequence that matches the input dimension of the intelligent diagnostic model;

[0044] The time-series feature sequence is input into the intelligent diagnostic model, and the forward propagation reasoning process of the intelligent diagnostic model is executed.

[0045] Obtain the original diagnostic confidence vector generated by the output layer of the intelligent diagnostic model;

[0046] Based on a preset diagnostic threshold, the original diagnostic confidence vector is used to determine its state, thereby obtaining the leakage status diagnosis result of the underground water pipe network.

[0047] In a preferred embodiment, the step of verifying the leakage status diagnosis results based on the spatial topology of the underground water pipe network and historical leakage logs to obtain leakage location information of the underground water pipe network includes:

[0048] Based on the geographical coordinates and connection relationships of the monitoring points in the underground water pipeline network, a spatial topology map of the underground water pipeline network is constructed.

[0049] Extract the locations of leakage events recorded in the historical time periods adjacent to the current diagnosis time period from the historical leakage log of the underground water pipe network as the historical leakage location set;

[0050] The locations of monitoring points marked as having leaks in the leakage status diagnosis results are mapped onto the spatial topology map to obtain the candidate leakage point set of the underground water pipe network;

[0051] Spatial topology verification is performed on the monitoring points of the candidate leakage point set;

[0052] Based on the results of the spatial topology verification, the selected monitoring point location information is integrated into the leakage location information of the underground water pipe network.

[0053] In a preferred embodiment, integrating the leakage status diagnosis results with the leakage location information into a leakage diagnosis report for the underground water pipe network includes:

[0054] According to the preset report template, the leakage probability and status category in the leakage status diagnosis result, as well as the suspected leakage pipe segment and confidence level in the leakage location information, are structured and mapped and filled.

[0055] Based on the mapping and filling results, the correlation between the leakage status diagnosis results and the leakage location information is established and integrated to obtain the integrated data record of the underground water pipe network;

[0056] The integrity and logical consistency of the fields in the integrated data records are verified to obtain a leakage diagnosis report for the underground water pipe network.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] 1. This invention employs a multi-stage, progressive data cleaning process. Based on the acoustic characteristic benchmark parameters of underground water pipe networks, it conducts multi-dimensional anomaly screening. Combined with wavelet decomposition noise reduction, interpolation compensation, and continuity verification, it effectively improves the purification level of the original sound data, forming standardized and non-redundant standard sound data, providing high-quality data support for subsequent diagnostic analysis. Simultaneously, by constructing feature vectors through multi-dimensional fusion of time-domain and frequency-domain features, and relying on an expert knowledge rule base, it achieves accurate classification of labeled data, enhancing the comprehensiveness and accuracy of leakage-related feature identification.

[0059] 2. The intelligent diagnostic model constructed in this invention, combining a bidirectional convolutional long short-term memory network with an adaptive attention mechanism, can deeply mine the temporal correlation features and key information of sound data. Through dynamic feature weight allocation, it enhances the model's ability to capture leakage features, significantly optimizing the accuracy and efficiency of leakage status diagnosis. Furthermore, by leveraging the spatial topology of the underground water pipe network and historical leakage logs for leakage verification, it achieves precise location of the leak. Then, by structurally integrating the diagnostic results and location information, it generates a standardized leakage diagnosis report, providing comprehensive and reliable technical support for the rapid handling and efficient repair of pipe network leaks. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating an industrial underground water pipeline leakage diagnosis technology based on an intelligent learning model, as provided in an embodiment of the present invention.

[0061] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0062] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0063] This application provides a technology for diagnosing leakage in industrial underground water pipe networks based on an intelligent learning model. The executing entity of this technology includes, but is not limited to, at least one electronic device that can be configured to execute the technology provided in this application, such as a server or a terminal. In other words, the technology can be executed by software or hardware installed on a terminal device or a server-side device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cluster of cloud servers. The server can be an independent server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0064] Reference Figure 1 The diagram shown is a flowchart illustrating a leakage diagnosis technology for industrial groundwater pipe networks based on an intelligent learning model, according to an embodiment of the present invention. In this embodiment, the leakage diagnosis technology for industrial groundwater pipe networks based on an intelligent learning model includes:

[0065] S101. Perform data cleaning on the raw sound data of the underground water pipe network to obtain the standard sound data of the underground water pipe network;

[0066] In this embodiment of the invention, the step of cleaning the raw sound data of the underground water pipe network to obtain the standard sound data of the underground water pipe network includes:

[0067] Based on the acoustic characteristic benchmark parameters of the underground water pipe network, multi-dimensional anomaly screening is performed on the original sound data to obtain preliminary purified data of the original sound data.

[0068] Wavelet decomposition and noise reduction are performed on the preliminary purified data to obtain redundant data from the preliminary purified data.

[0069] Interpolation compensation is performed on the non-redundant data to obtain the standardized data of the non-redundant data;

[0070] The standard data is checked for continuity according to the timestamp order to obtain the standard sound data of the underground water pipe network.

[0071] Based on the acoustic characteristic benchmark parameters of the underground water pipe network, anomaly screening is performed on the original sound data in three dimensions: frequency, amplitude, and duration. The acoustic characteristic benchmark parameters include preset normal pipe network sound frequency range, normal pipe network sound amplitude range, and normal pipe network sound duration range. During screening, the frequency value, amplitude value, and duration value of each data point in the original sound data are extracted one by one. The frequency value is compared with the preset normal pipe network sound frequency range, and data points with frequency values ​​exceeding the range are eliminated. The amplitude value is compared with the preset normal pipe network sound amplitude range, and data points with amplitude values ​​exceeding the range are eliminated. The duration value is compared with the preset normal pipe network sound duration range, and data points with duration values ​​exceeding the range are eliminated. The remaining sound data after the above three-dimensional screening is the preliminary purified data of the original sound data.

[0072] The initial purified data is input into the wavelet decomposition process, which decomposes the initial purified data into wavelet coefficients of multiple different scales. A fixed wavelet coefficient threshold is set, and the relationship between the absolute value of each wavelet coefficient and the threshold is judged one by one. Wavelet coefficients with an absolute value lower than the threshold are set to zero, while wavelet coefficients with an absolute value higher than the threshold are retained. Then, all the processed wavelet coefficients are subjected to inverse wavelet transform and resynthesized into a continuous audio data segment. This resynthesized audio data is the non-redundant data of the initial purified data.

[0073] First, sort out the timestamp sequence without redundant data and mark the positions where data is missing in the timestamp sequence. For each missing data position, extract the value of the previous and next valid data points. Using the timestamp interval between the two valid data points as the calculation basis, use linear interpolation to calculate the corresponding padding value for the missing position. Accurately fill the corresponding missing position with the calculated padding value. After filling the values ​​of all missing positions, the resulting audio data is the standardized data without redundant data.

[0074] The standardized data is sorted in chronological order according to timestamps. Then, adjacent data points are selected sequentially, and the numerical difference between the next and previous data points is calculated. This calculated difference is compared to a preset continuity difference threshold. If the difference does not exceed the threshold, the next data point is retained; otherwise, it is discarded. This process of calculating, comparing, and discarding differences between adjacent data points continues until the numerical differences between all adjacent data points do not exceed the preset continuity difference threshold. After all verification operations are completed, the remaining audio data is the standard audio data for the underground water pipe network.

[0075] S102. Perform multi-scale feature analysis on the standard sound data to obtain leakage label data and non-leakage label data of the underground water pipe network;

[0076] In this embodiment of the invention, the step of performing multi-scale feature analysis on the standard sound data to obtain leakage label data and non-leakage label data of the underground water pipe network includes:

[0077] The temporal statistical features in the standard sound data are used as the temporal feature set of the standard sound data.

[0078] The standard sound data is subjected to frequency domain energy distribution statistics to obtain the frequency domain feature set of the standard sound data;

[0079] The time-domain feature set and the frequency-domain feature set are fused in multiple dimensions to construct a multi-dimensional fused feature vector of the standard sound data.

[0080] Based on a pre-set expert knowledge rule base, pattern matching and classification are performed on the multi-dimensional fused feature vectors to obtain leakage label data and non-leakage label data of the underground water pipe network.

[0081] The temporal statistical features of the standard audio data are extracted, including mean, variance, peak value, trough value, waveform factor, and peak factor. To extract the mean, the values ​​of all data points in the standard audio data are summed, and the sum is divided by the total number of data points. To extract the variance, the difference between each data point and the mean is calculated, and each difference is squared. All squared results are summed, and the sum is divided by the total number of data points to obtain the variance. To extract the peak value, the values ​​of all data points in the standard audio data are iterated through, and the data with the largest value is selected. When extracting peak values, the value of the data point is the peak value. When extracting valley values, the values ​​of all data points in the standard sound data are traversed, and the data point with the smallest value is selected as the valley value. When extracting waveform factors, the mean of the absolute values ​​of all data point values ​​is calculated first, and then the mean is divided by the square root of the variance to obtain the waveform factor. When extracting peak factors, the peak value is divided by the mean of the absolute values ​​of all data point values ​​to obtain the peak factor. The set formed by integrating the mean, variance, peak value, valley value, waveform factor, and peak factor extracted above is the time-domain feature set of the standard sound data.

[0082] The standard sound data is converted into frequency component data in the frequency domain by performing frequency domain transformation. The transformation process involves decomposing the time-domain signal corresponding to the standard sound data into a combination of sinusoidal signals of different frequencies. Each sinusoidal signal corresponds to a frequency value and an amplitude value. Then, fixed frequency intervals are divided, with the interval between frequency intervals set to 10Hz. Starting from 0Hz, continuous frequency intervals are divided sequentially. The amplitude values ​​of all sinusoidal signals contained in each frequency interval are squared. Then, the results of all squared operations in each frequency interval are accumulated to obtain the energy value corresponding to each frequency interval. At the same time, the start frequency and end frequency of each frequency interval are recorded. The start frequency, end frequency and corresponding energy value of all frequency intervals are used as frequency domain features. The set of all frequency domain features is the frequency domain feature set of the standard sound data.

[0083] First, all feature values ​​in the time-domain feature set are normalized. The normalization method is to select the maximum and minimum values ​​of each feature in the time-domain feature set as the normalization boundary of the feature, subtract the minimum value of the feature from the specific value of the feature, and then divide the difference by the difference between the maximum and minimum values ​​of the feature to obtain the normalized time-domain feature value. The same normalization method is used to normalize all feature values ​​in the frequency-domain feature set to obtain the normalized frequency-domain feature values. Then, in a fixed order of time-domain features first and then frequency-domain features, all the normalized time-domain feature values ​​and normalized frequency-domain feature values ​​are arranged in sequence to form a continuous one-dimensional data sequence. This one-dimensional data sequence is the multi-dimensional fusion feature vector of the standard sound data.

[0084] The pre-defined expert knowledge rule base includes leakage pattern rules and non-leakage pattern rules. Leakage pattern rules are fixed value ranges corresponding to the normalized feature values ​​in the multi-dimensional fused feature vector, while non-leakage pattern rules are another set of fixed value ranges corresponding to the normalized feature values ​​in the multi-dimensional fused feature vector. The value ranges of leakage pattern rules and non-leakage pattern rules do not overlap. Each normalized feature value in the multi-dimensional fused feature vector is sequentially compared with a leakage pattern rule in the expert knowledge rule base to determine whether each normalized feature value falls within the value range corresponding to a leakage pattern rule. If all normalized feature values... If all normalized feature values ​​meet the value requirements of the leakage pattern rule, then the standard audio data corresponding to the multi-dimensional fusion feature vector is marked as leakage label data. If at least one normalized feature value does not meet the value requirements of the leakage pattern rule, then each normalized feature value in the multi-dimensional fusion feature vector is compared with the non-leakage pattern rule in the expert knowledge rule base in turn to determine whether each normalized feature value is within the value range corresponding to the non-leakage pattern rule. If all normalized feature values ​​meet the value requirements of the non-leakage pattern rule, then the standard audio data corresponding to the multi-dimensional fusion feature vector is marked as non-leakage label data.

[0085] S103. Based on the original sound data, the leaked label data, and the non-leaked label data, an intelligent diagnostic model for the underground water pipe network is constructed using a bidirectional convolutional long short-term memory network combined with an attention mechanism.

[0086] In this embodiment of the invention, the step of constructing an intelligent diagnostic model for the underground water pipe network based on the original sound data, the leaked tag data, and the non-leaking tag data, using a bidirectional convolutional long short-term memory network combined with an attention mechanism, includes:

[0087] The original sound data, the leaked label data, and the non-leaking label data are sample-paired and integrated to obtain the label dataset of the underground water pipe network;

[0088] Perform balanced sampling on the labeled dataset to obtain the validation training dataset of the labeled dataset;

[0089] Based on the model architecture of convolutional neural networks, the hierarchical parameters and connection methods of the bidirectional convolutional long short-term memory network are initialized and configured to obtain the initial network framework of the underground water pipe network.

[0090] An adaptive attention mechanism is embedded in the initial network framework to construct a feature weight allocation benchmark, so as to obtain a preliminary diagnostic model of the groundwater pipe network.

[0091] Based on the validation training dataset, the preliminary diagnostic model is trained with strategy configuration to obtain the intelligent diagnostic model for the underground water pipe network.

[0092] The convolutional neural network-based model architecture initializes the hierarchical parameters and connection methods of the bidirectional convolutional long short-term memory network to obtain the initial network framework of the underground water pipeline network, including:

[0093] Based on the model architecture of the convolutional neural network, the receptive field size and stride of the convolutional kernel of the model architecture are defined, and the number of feature maps output by the convolutional layers in the model architecture is determined to obtain the convolutional structure of the underground water pipe network.

[0094] The bidirectional long short-term memory module is initialized and configured to obtain the temporal learning structure of the underground water pipe network;

[0095] The output feature map sequence of the convolutional structure is used as the input of the temporal learning structure for connection configuration, thereby establishing the forward propagation path of data from feature extraction to temporal modeling;

[0096] Based on the forward propagation path, the convolutional structure and the temporal learning structure are integrated and connected to obtain the initial network framework of the underground water pipe network.

[0097] The step of embedding an adaptive attention mechanism into the initial network framework to construct a feature weight allocation benchmark to obtain a preliminary diagnostic model of the groundwater pipeline network includes:

[0098] Attention parameter matrices are added at predetermined positions in the initial network framework to enable training and learning from the data.

[0099] The feature sequence output by the temporal learning structure is input into the attention parameter matrix;

[0100] The attention weights of the input feature sequence are obtained by calculating the correlation between the feature vectors at different time steps in the input feature sequence using the attention parameter matrix.

[0101] Based on the attention weights, the feature vectors at each time step in the input feature sequence are weighted and fused to obtain the context feature vector of the input feature sequence.

[0102] The context feature vectors are then fed into the subsequent processing layer of the initial network framework to complete the embedding of the adaptive attention mechanism, thereby obtaining a preliminary diagnostic model of the underground water pipe network.

[0103] The formula for calculating the attention weight is as follows:

[0104] ;

[0105] In the formula, For the first The feature vector at the i-th time step is related to the i-th time step. The attention weights of the feature vectors at each time step For the first The time step and the first Relevance score at each time step The total number of time steps for the input feature sequence. For the first The feature vector at the i-th time step is related to the i-th time step. The correlation score of the feature vectors at each time step It is a natural exponential function. This is a temporary index variable used for summation.

[0106] No. The feature vector at the i-th time step is related to the i-th time step. The attention weights of the feature vectors at each time step are derived through corresponding calculations, and their calculation depends on the first time step. The time step and the first The correlation score at the nth time step, the total number of time steps in the input feature sequence, and the result of the natural exponential function calculation are used. The calculation process involves first applying the natural exponential function to the nth time step. The time step and the first The correlation scores at each time step are transformed, and then the transformed results are compared with all the time steps. The sum of the correlation scores of each time step and the corresponding time step after transformation by the natural exponential function is compared to the final ratio, which is the attention weight. The core function of this parameter is to quantify the weight distribution relationship between feature vectors of different time steps, and its source process is directly related to the correlation analysis of feature sequences.

[0107] No. The time step and the first The source of the correlation score at the time step is the 1st time step. The time step and the first The feature vectors at the nth time step are obtained by correlation analysis. Specifically, the feature vectors at the nth time step are obtained by performing correlation analysis on the feature sequence output by the time-series learning structure. The feature vector at the nth time step and the nth time step Each time step feature vector is input into the attention parameter matrix. The difference in the values ​​of corresponding elements of two feature vectors is measured by the preset initial parameters in the matrix. The smaller the difference, the higher the calculated correlation score. The source of this parameter is the temporal feature correlation of the underground water pipe network sound data, that is, the sound features of adjacent or related time steps should have a higher degree of fit, and thus correspond to a higher correlation score.

[0108] The total number of time steps in the input feature sequence comes from the statistical results of the time steps in the input feature sequence. The input feature sequence is output by the temporal learning structure, and its number of time steps is consistent with the number of sampling timestamps of the original sound data of the underground water pipe network. The specific statistical process is to traverse the set of feature vectors arranged in the order of time steps in the input feature sequence and count the total number of feature vectors. This number is the total number of time steps in the input feature sequence. The source is based on the acquisition scheme of the original sound data. During acquisition, sound data is recorded at fixed time intervals, and each time interval corresponds to a sampling timestamp, which in turn corresponds to a time step in the feature sequence, ensuring that the total number of time steps in the input feature sequence can completely cover all the temporal information of the feature sequence.

[0109] No. The source of the correlation score between the feature vector of the 1st time step and the feature vector of the corresponding time step is related to the 2nd time step. The time step and the first The source logic of the correlation scores at each time step is consistent; they are all results of correlation analysis between feature vectors. Specifically, the process involves... The feature vector at the nth time step and the corresponding feature vector at the time step in the input feature sequence are input into the attention parameter matrix. Preset initial parameters within the matrix are used to measure the difference in values ​​of corresponding elements between the two vectors, thereby obtaining a relevance score. The value of the corresponding time step ranges from 1 to the total number of time steps in the input feature sequence. This value is derived to complete the attention parameter matrix for the nth time step. A comprehensive calculation of the correlation between the feature vector at each time step and the feature vectors at all time steps ensures that subsequent summation operations can cover all time steps of the feature sequence.

[0110] The natural exponential function is derived from the quantification requirement of feature correlation scores. The reason for choosing this function is that it can convert correlation scores in any real number range into non-negative values, while preserving the degree of difference between different scores, avoiding the interference of negative scores on subsequent weight calculations, ensuring that the weight values ​​are all positive and can accurately reflect the strength of correlation between feature vectors. The process of using this function is to take each correlation score as input and output the corresponding exponential result.

[0111] The temporary index variable is used to implement the first... The summation calculation of the correlation scores between the feature vector at the nth time step and the feature vectors at all time steps is performed, with values ​​increasing sequentially from 1 to the total number of time steps in the input feature sequence. Each value corresponds to a time step in the input feature sequence. This is based on the principle that the summation operation needs to traverse all time steps to ensure the comprehensiveness of the correlation analysis and avoid missing any feature vector at any time step. The correlation information of the feature vectors at each time step.

[0112] The significance of this formula is to normalize the correlation scores of feature vectors at different time steps in the input feature sequence, thereby obtaining the attention weight of each time step feature vector to other time step feature vectors. Its relevance to the content is reflected in the attention weight calculated by this formula, which can quantify the importance of feature vectors at different time steps. This allows the subsequent weighted fusion process to focus on feature information that is more relevant and more valuable for diagnosing leaks in underground water pipe networks. It provides a core quantitative basis for constructing a feature weight allocation benchmark, ensuring that the adaptive attention mechanism can accurately act on the feature sequence output by the temporal learning structure, improving the effectiveness of subsequent context feature vectors, and thus supporting the accurate judgment of the leakage status of underground water pipe networks by the preliminary diagnostic model.

[0113] The raw audio data is assigned a unique sample identifier in chronological order of collection, with each sample identifier corresponding to a continuous segment of raw audio data. Simultaneously, leaked and non-leaked tagged data are assigned the same sample identifier as their corresponding raw audio data, ensuring that each tagged data is accurately associated with its source raw audio data. Then, the sample identifiers are matched one by one, combining raw audio data with the corresponding leaked or non-leaked tagged data to form "raw audio data - tagged data" sample pairs. The set formed by integrating all matched sample pairs constitutes the tagged dataset for the underground water pipe network.

[0114] First, count the number of leaked and non-leaked label sample pairs in the label dataset. Define the category with fewer samples as the baseline category and record its sample count. From the category with more samples, randomly select sample pairs equal to the baseline category's sample count, ensuring each sample pair has the same probability of being selected and that no pairs are selected repeatedly. Merge the selected sample pairs with all sample pairs from the baseline category; the merged set of sample pairs constitutes the validation training dataset for the label dataset.

[0115] Based on the classic architecture of convolutional neural networks, the hierarchical composition of a bidirectional convolutional long short-term memory (BSL-S) network is determined, including an input layer, convolutional layers, pooling layers, a BSL-S, and fully connected layers. The connection order of each layer is clarified: the output of the input layer connects to the input of the convolutional layer, the output of the convolutional layer connects to the input of the pooling layer, the output of the pooling layer connects to the input of the BSL-S, and the output of the BSL-S connects to the input of the fully connected layer. The core components of each layer are determined: the input layer receives the feature dimensions corresponding to the raw sound data; the convolutional layer uses a fixed-size kernel to extract local features; the pooling layer uses a fixed-size pooling window to compress feature dimensions; the BSL-S uses a fixed number of memory units to capture temporal features; and the fully connected layer outputs nodes corresponding to the label categories. Through the above-described hierarchical composition, connection order, and core component settings, the initial configuration of the BSL-S layer parameters and connection methods is completed, resulting in the initial network framework of the underground water pipe network.

[0116] An adaptive attention mechanism module is embedded between the bidirectional long short-term memory layer and the fully connected layer in the initial network framework. The module is constructed as follows: it first receives all temporal feature vectors output from the bidirectional long short-term memory layer, calculates the correlation between each temporal feature vector and all other temporal feature vectors, and calculates the correlation based on the similarity between vectors; higher similarity indicates a higher correlation. Based on the calculated correlation, a corresponding weight value is assigned to each temporal feature vector, with higher weight values ​​assigned to those with higher correlation and lower weight values ​​assigned to those with lower correlation. The sum of the weight values ​​of all temporal feature vectors is 1. The assigned weight values ​​are then superimposed on the corresponding temporal feature vectors to form a fused feature vector after weight adjustment. This weight allocation method is the feature weight allocation benchmark. After completing the embedding of the adaptive attention mechanism and the construction of the feature weight allocation benchmark, a preliminary diagnostic model of the underground water pipe network is obtained.

[0117] The validation training dataset is divided into training and validation subsets according to a preset ratio of 70% and 30%, respectively. The partitioning is randomized, ensuring the distribution of the two types of labeled samples within the subsets matches the original validation training dataset. A training strategy is configured with a fixed number of training iterations. In each iteration, a fixed number of samples from the training subset are input into the preliminary diagnostic model. The model outputs predicted labels based on the input samples. The predicted labels are compared with the corresponding ground truth labels, and the number of samples in each batch whose predicted labels match the ground truth labels is counted. This count is used to judge the model's prediction performance. If the prediction performance does not meet the preset requirements, the parameters of each level of the model are adjusted to increase the percentage of correctly predicted samples in subsequent iterations. After each iteration, the model is validated using the validation subset, and the percentage of correctly predicted samples in the validation subset is also counted. If this percentage remains unchanged for three consecutive iterations during validation, training is terminated early. Through the cyclical process of training subsets, adjusting parameters, and validating subsets, the strategy configuration training of the preliminary diagnostic model is completed. When the number of training iterations is reached or the training is terminated early, the training is stopped, and the intelligent diagnostic model of the underground water pipe network is obtained.

[0118] Based on a convolutional neural network (CNN) model architecture and considering the temporal continuity of underground water pipe network sound data, the receptive field size of the convolutional kernel is defined to cover three consecutive sound data points. This receptive field size is determined by the number of sampling points corresponding to the shortest effective fluctuation period of the underground water pipe network sound signal, ensuring complete capture of local sound signal fluctuation features. The convolutional kernel's stride is defined as one sound data point, meaning that after each movement, the convolutional kernel covers only one new sound data point, ensuring no local sound feature information is missed. The number of feature maps output by the convolutional layer is determined to be 16, with each feature map corresponding to a set of independent convolutional kernel parameters used to extract local features of different types of sound signals. Through the definitions of the convolutional kernel's receptive field size, stride, and the determination of the number of convolutional layer output feature maps, a convolutional structure for the underground water pipe network is formed.

[0119] When initializing the bidirectional long short-term memory module, it is first configured to contain memory units in both forward and reverse directions, with 32 memory units in each direction. These memory units store the temporal correlation information of the sound signals. The forward memory units process the input feature information in the order of the sound data's timestamps, while the reverse memory units process the input feature information in reverse order of the sound data's timestamps. All memory units are initialized to zero to ensure that there is no prior temporal information interference in the initial stage. Simultaneously, a forgetting threshold of 0.5 is set for each memory unit to determine whether to retain historical temporal information. That is, if the weight of historical information in a memory unit is lower than 0.5, it is forgotten; if it is higher than or equal to 0.5, it is retained. This configuration yields the temporal learning structure of the underground water pipe network.

[0120] The feature map sequence output by the convolutional structure is normalized to ensure that the temporal dimension of each feature map matches the input temporal dimension of the temporal learning structure. Each time node in each feature map sequence corresponds to a feature vector. The normalized feature map sequence is then directly connected to the input port of the temporal learning structure, ensuring that the feature vectors of each time node output by the convolutional structure are input one-to-one into the forward and backward memory units of the temporal learning structure. The data propagation direction is defined as flowing from the output of the convolutional structure to the input of the temporal learning structure, without changing the temporal order of the feature data during this propagation. Through the above connection configuration, the forward propagation path of data from feature extraction to temporal modeling is established.

[0121] Based on the established forward propagation path, the output of the convolutional structure is fixedly connected to the input of the temporal learning structure, ensuring that data can only flow along the forward propagation path. Simultaneously, the data interaction rules between the convolutional and temporal learning structures are defined: the feature map sequence output by the convolutional structure must be completely input into the temporal learning structure, and the temporal learning structure only receives the feature data output by the convolutional structure, not other external data. The connected structure is checked and confirmed to have no data transmission breakpoints or redundant connections, ensuring the integrity of the entire structure and the smoothness of data transmission. Through the above integration and connection of the convolutional and temporal learning structures, the initial network framework of the underground water pipe network is obtained.

[0122] In the initial network framework, a predetermined position is set between the temporal learning structure and the subsequent processing layer. This position is determined because the feature sequence output by the temporal learning structure needs to be weighted before being fed into the subsequent processing layer, ensuring that the attention mechanism can accurately apply to the temporal features. An attention parameter matrix is ​​added at this position, with initial values ​​set to uniformly distributed small values ​​ranging from 0.01 to 0.05. This range is set to avoid gradient vanishing or exploding during training due to excessively large or small initial weights. By adding this matrix, the network can continuously update the parameters within the matrix during training, thereby achieving training and learning of the correlations between features in the input data.

[0123] The feature sequence output by the temporal learning structure is a set of feature vectors arranged sequentially by time step. Each time step corresponds to a 64-dimensional feature vector, and the length of the feature sequence is consistent with the number of sampling timestamps in the underground water pipe network sound data. This feature sequence is then passed to the input port of the attention parameter matrix one by one in chronological order. During the input process, it is ensured that the dimension of the feature vector at each time step is consistent with the input dimension of the attention parameter matrix, i.e., both are 64-dimensional, to avoid feature information loss or transmission errors due to dimension mismatch, thus completing the input of the feature sequence to the attention parameter matrix.

[0124] After receiving the feature sequence, the attention parameter matrix compares the feature vector at each time step with the feature vectors at all other time steps in the feature sequence. The degree of fit between two feature vectors is measured using preset initial parameters within the matrix. The degree of fit is determined by the difference in the values ​​of corresponding elements in the two feature vectors; the smaller the difference, the higher the degree of fit. The degree of fit of each feature vector with all other feature vectors is summarized to obtain the comprehensive fit value for that feature vector. Then, the comprehensive fit values ​​of all feature vectors are normalized by dividing each comprehensive fit value by the sum of all comprehensive fit values, so that the sum of the comprehensive fit values ​​of all feature vectors after processing is 1. The normalized comprehensive fit value for each feature vector is the attention weight of that time step feature vector. The attention weights of all time steps together constitute the attention weight of the input feature sequence.

[0125] The dynamic attention weights are continuously updated during model training to ensure real-time adaptation to changes in the feature sequence. Based on these dynamic attention weights, the feature vectors at each time step in the input feature sequence are weighted and fused. The fusion process involves multiplying the value of each element in the feature vector at each time step by the attention weight corresponding to that time step to obtain the weighted value of each element. Then, the weighted values ​​of the corresponding elements in the feature vectors of all time steps are summed to obtain a new vector with a dimension of 64. This new vector is the context feature vector of the input feature sequence. This vector integrates the important information of features from different time steps and highlights the contribution of high-weight features.

[0126] The output of the context feature vector is fixedly connected to the input of the subsequent processing layer of the initial network framework. During the connection process, it is ensured that the dimension of the context feature vector is consistent with the input dimension of the subsequent processing layer. After the connection is completed, the compatibility of the attention mechanism module with the original structure of the initial network framework is checked to ensure that data can flow smoothly from the temporal learning structure through the attention mechanism module to the subsequent processing layer without data transmission interruptions or redundant connections. At this point, the adaptive attention mechanism has been fully embedded into the initial network framework, and the integrated network framework is the preliminary diagnostic model for the underground water pipe network.

[0127] S104. Input the standard sound data into the intelligent diagnostic model to obtain the leakage status diagnosis result of the underground water pipe network;

[0128] In this embodiment of the invention, the step of inputting the standard sound data into the intelligent diagnostic model to obtain the leakage status diagnostic result of the underground water pipe network includes:

[0129] Temporal features are extracted from the standard sound data to construct a temporal feature sequence that matches the input dimension of the intelligent diagnostic model;

[0130] The time-series feature sequence is input into the intelligent diagnostic model, and the forward propagation reasoning process of the intelligent diagnostic model is executed.

[0131] Obtain the original diagnostic confidence vector generated by the output layer of the intelligent diagnostic model;

[0132] Based on a preset diagnostic threshold, the original diagnostic confidence vector is used to determine its state, thereby obtaining the leakage status diagnosis result of the underground water pipe network.

[0133] When extracting temporal features from standard audio data, the data is first divided at fixed time intervals, with the interval set to 0.1 seconds, consistent with the time interval used during the training of the intelligent diagnostic model. Each time interval corresponds to a time step, and each time step contains 10 consecutive standard audio data points. For each time step, three features are extracted: mean, peak, and trough. To extract the mean, the values ​​of the 10 data points are summed and then divided by 10. To extract the peak, the data point with the largest value among the 10 data points is selected. To extract the trough, the data point with the smallest value among the 10 data points is selected. The input dimension of the intelligent diagnostic model is preset to 64 dimensions. If the feature vector formed by the three features extracted in a single time step has less than 64 dimensions, zero values ​​are added to the end of the feature vector to bring it to 64 dimensions. If the number of time steps obtained after the division is less than the required input sequence length of the model, zero vectors with a dimension of 64 are added to the end of the sequence. Finally, a temporal feature sequence is formed, arranged in chronological order of time steps, with dimensions that perfectly match the input dimension of the intelligent diagnostic model.

[0134] The constructed temporal feature sequence is input into the input layer of the intelligent diagnostic model in chronological order, ensuring that the dimension of the temporal feature sequence matches the receiving dimension of the input layer. During the forward propagation inference process, the temporal feature sequence is first fed into the model's convolutional structure. Local temporal features are extracted from the sequence using pre-defined convolutional kernels, generating a feature map sequence. This feature map sequence is then fed into a temporal learning structure, which captures the temporal correlation information between preceding and following sequences through forward and backward memory units. The feature sequence output from the temporal learning structure is fed into an adaptive attention mechanism module. Attention weights for each time step feature vector are calculated using a pre-trained attention parameter matrix, and the feature sequences are weighted and fused to obtain a context feature vector. Finally, the context feature vector is fed into a fully connected layer, where the features are integrated using connection weights, completing the entire forward propagation inference process.

[0135] The output layer of the intelligent diagnostic model contains two output nodes, corresponding to the leakage and non-leakage states of the underground water pipe network, respectively. The original diagnostic confidence vector is a two-dimensional vector composed of the output values ​​of these two output nodes. The first element of the vector corresponds to the confidence level of the leakage state, and the second element corresponds to the confidence level of the non-leakage state. The confidence level ranges from 0 to 1. By reading the real-time output values ​​of the two output nodes of the output layer and combining the two values ​​in the order of "leakage confidence level - non-leakage confidence level", the original diagnostic confidence vector of the underground water pipe network is obtained.

[0136] The preset diagnostic threshold is set to 0.5. This threshold is based on the optimal classification threshold of the validation set during the training of the intelligent diagnostic model. That is, when the confidence level reaches 0.5, the model's accuracy in judging the state is the highest. When determining the state of the original diagnostic confidence vector, the confidence value corresponding to the leakage state is first extracted from the vector and compared with the diagnostic threshold of 0.5. If the confidence value of the leakage state is greater than or equal to 0.5, the diagnosis result of the groundwater pipe network leakage state is determined to be leakage. If the confidence value of the leakage state is less than 0.5, the confidence value of the corresponding non-leakage state is extracted from the vector, and its value will necessarily be greater than 0.5. In this case, the diagnosis result of the groundwater pipe network leakage state is determined to be non-leakage.

[0137] S105. Based on the spatial topology of the underground water pipe network and historical leakage logs, perform leakage verification on the leakage status diagnosis results to obtain leakage location information of the underground water pipe network.

[0138] In this embodiment of the invention, the step of verifying the leakage status diagnosis results based on the spatial topology of the underground water pipe network and historical leakage logs to obtain leakage location information of the underground water pipe network includes:

[0139] Based on the geographical coordinates and connection relationships of the monitoring points in the underground water pipeline network, a spatial topology map of the underground water pipeline network is constructed.

[0140] Extract the locations of leakage events recorded in the historical time periods adjacent to the current diagnosis time period from the historical leakage log of the underground water pipe network as the historical leakage location set;

[0141] The locations of monitoring points marked as having leaks in the leakage status diagnosis results are mapped onto the spatial topology map to obtain the candidate leakage point set of the underground water pipe network;

[0142] Spatial topology verification is performed on the monitoring points of the candidate leakage point set;

[0143] Based on the results of the spatial topology verification, the selected monitoring point location information is integrated into the leakage location information of the underground water pipe network.

[0144] First, the latitude and longitude coordinates of all monitoring points in the underground water pipe network were collected, with the coordinates accurate to six decimal places. Simultaneously, the pipe connections between each monitoring point were analyzed, clarifying the pipe diameter, length, and direction of each connection. Based on the geographic coordinates, the location of each monitoring point was marked on a plane coordinate system. Following the analyzed connections, adjacent monitoring points were connected by line segments, and the corresponding pipe information was labeled with the attributes of the line segments. Finally, a spatial topology map of the underground water pipe network was formed, comprehensively reflecting the distribution of monitoring points and the pipe connections.

[0145] The time range for the current diagnostic period is determined. Using this time range as a baseline, an adjacent historical period is defined as the period 30 days prior to the current diagnostic period's start date. This period is chosen because groundwater network leakage events have a certain temporal correlation, and historical leakage data within the adjacent 30 days is valuable for verifying the current diagnosis. From the historical leakage logs of the groundwater network, all leakage events whose recording times fall within the aforementioned adjacent historical periods are selected. The specific geographical location information corresponding to each selected leakage event is extracted, represented by latitude and longitude coordinates. All extracted latitude and longitude coordinates are compiled to form a historical leakage location set for the groundwater network.

[0146] From the leakage status diagnosis results, the identification information of all monitoring points marked as having leaks is extracted, and their corresponding latitude and longitude coordinates are retrieved based on the monitoring point identification information. In the constructed spatial topology map, the location of each corresponding monitoring point is found based on the retrieved latitude and longitude coordinates and specially marked with a solid red dot to distinguish it from the marking style of other non-leaking monitoring points. The set of all specially marked monitoring points constitutes the candidate leakage point set of the groundwater pipe network.

[0147] When performing spatial topology verification on monitoring points of the candidate leak point set, the verification rules are first clarified: the candidate leak point must be located on a valid pipe segment of the underground water pipe network, and the connecting pipe between the candidate leak point and the adjacent monitoring point must be in normal laying condition. For each candidate leak point, its location is checked in the spatial topology map to see if it corresponds to a valid pipe segment. Specifically, the verification method is to confirm whether the latitude and longitude coordinates of the point fall within the range of the marked pipe line segments in the spatial topology map; at the same time, it is checked whether the connecting line segment between the candidate leak point and the adjacent monitoring point exists in the pipe connection relationship of the spatial topology map. If it exists, the connecting pipe is determined to be in normal laying condition. If the candidate leak point meets both the conditions of being located on a valid pipe segment and having a normally laid connecting pipe, the verification passes; if neither condition is met, the verification fails and the candidate leak point is removed.

[0148] Collect the location information of all candidate leak points that pass spatial topology verification, including the latitude and longitude coordinates of each candidate leak point, the pipeline segment number, and the corresponding monitoring point number. Organize this information in a unified format, with each leak point's information recorded as a single record. Each record contains the latitude and longitude coordinates, the pipeline segment number, and the monitoring point number in sequence. The structured information set formed by summarizing and integrating all records constitutes the leak location information of the underground water pipe network.

[0149] S106. Integrate the leakage status diagnosis results and leakage location information into a leakage diagnosis report for the underground water pipe network.

[0150] In this embodiment of the invention, integrating the leakage status diagnosis results and leakage location information into a leakage diagnosis report for the underground water pipe network includes:

[0151] According to the preset report template, the leakage probability and status category in the leakage status diagnosis result, as well as the suspected leakage pipe segment and confidence level in the leakage location information, are structured and mapped and filled.

[0152] Based on the mapping and filling results, the correlation between the leakage status diagnosis results and the leakage location information is established and integrated to obtain the integrated data record of the underground water pipe network;

[0153] The integrity and logical consistency of the fields in the integrated data records are verified to obtain a leakage diagnosis report for the underground water pipe network.

[0154] Based on the pre-set report template, first define the pre-defined structured fields within the template, including diagnosis time, leakage status category, leakage probability, suspected leaking pipe segment number, confidence level, and monitoring point number. Each field has a fixed fill format and data type requirement. The diagnosis time is in the format YYYY-MM-DDHH:MM:SS; the leakage status category is a string type of "leaking" or "non-leaking"; the leakage probability is a numerical type of 0 to 1; the suspected leaking pipe segment number is a string type of "GD-" followed by 6 digits; the confidence level is either a string type of "high confidence" or "medium confidence"; and the monitoring point number is a string type of "JC-" followed by 4 digits. Extract the leakage probability and status category corresponding to each monitoring point from the leakage status diagnosis results. Extract the pipe segment number and confidence level corresponding to each suspected leaking pipe segment from the leakage location information. The confidence level is divided based on the leakage probability: a leakage probability between 0.8 and 1.0 corresponds to "high confidence," and a leakage probability between 0.5 and 0.8 corresponds to "medium confidence." According to the correspondence between the template fields and the extracted data, the leakage probability is filled into the "Leakage Probability" field of the template, the leakage status category is filled into the "Leakage Status Category" field, the suspected leakage pipe section number is filled into the "Suspected Leakage Pipe Section Number" field, and the confidence level is filled into the "Confidence Level" field. At the same time, the current system time is automatically obtained as the diagnosis time and filled into the corresponding field, and the monitoring point number is directly associated and filled into the corresponding field to complete the structured mapping filling.

[0155] Based on the mapped and filled field data, a correlation is established using the monitoring point number as the association key. This is because the monitoring point number uniquely corresponds to the same monitoring point and its associated pipe segment in both the leakage status diagnosis results and leakage location information, ensuring the uniqueness and accuracy of the association. The leakage probability, leakage status category, suspected leaking pipe segment number, confidence level, and diagnosis time corresponding to the same monitoring point number are integrated into a single associated data record. The field data within each record are logically linked through the monitoring point number, meaning one record completely corresponds to the leakage status and location information of one monitoring point. All associated data records corresponding to all monitoring points are summarized to form a set containing multiple complete associated records; this set constitutes the integrated data record of the groundwater pipe network.

[0156] When performing field integrity checks on integrated data records, the required fields are clearly defined as diagnosis time, leakage status category, leakage probability, suspected leaking pipe section number, confidence level, and monitoring point number. Each data record is checked one by one to ensure that it contains valid data for all required fields. The criteria for valid data are that the field data conforms to the preset filling format and data type requirements, and there are no empty values ​​or data with incorrect formatting. If any required field is missing or the data is invalid, the record is marked as incomplete. When performing logical consistency checks, the check rules are set as follows: when the leakage status category is "leakage", there must be a corresponding suspected leaking pipe section number, and the leakage probability and confidence level must match. That is, a leakage probability of 0.8 to 1.0 corresponds to "high confidence", and 0.5 to 0.8 corresponds to "medium confidence". When the leakage status category is "non-leakage", the suspected leaking pipe section number field must be empty, and the confidence level field must also be empty. Each record is checked one by one to see if it conforms to the above rules. If it does not conform, it is marked as an inconsistent record. Incomplete and inconsistent records are removed, and all records that pass the field integrity and logical consistency checks are retained. The retained records are then organized according to the preset report template format, and the resulting document is the underground water pipe network leakage diagnosis report.

[0157] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0158] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for diagnosing leakage in industrial groundwater pipe networks based on an intelligent learning model, characterized in that, The method includes: S101. Perform data cleaning on the raw sound data of the underground water pipe network to obtain the standard sound data of the underground water pipe network; S102. Perform multi-scale feature analysis on the standard sound data to obtain leakage label data and non-leakage label data of the underground water pipe network, including: Based on the acoustic characteristic benchmark parameters of the underground water pipe network, multi-dimensional anomaly screening is performed on the original sound data to obtain preliminary purified data of the original sound data. Wavelet decomposition and noise reduction are performed on the preliminary purified data to obtain redundant data from the preliminary purified data. Interpolation compensation is performed on the non-redundant data to obtain the standardized data of the non-redundant data; The standard data is checked for continuity according to the timestamp order to obtain the standard sound data of the underground water pipe network; S103. Based on the original sound data, the leaked label data, and the non-leaking label data, an intelligent diagnostic model for the underground water pipe network is constructed using a bidirectional convolutional long short-term memory network combined with an attention mechanism, including: The original sound data, the leaked label data, and the non-leaking label data are sample-paired and integrated to obtain the label dataset of the underground water pipe network; Perform balanced sampling on the labeled dataset to obtain the validation training dataset of the labeled dataset; Based on the model architecture of convolutional neural networks, the hierarchical parameters and connection methods of the bidirectional convolutional long short-term memory network are initialized and configured to obtain the initial network framework of the underground water pipe network. An adaptive attention mechanism is embedded in the initial network framework to construct a feature weight allocation benchmark, so as to obtain a preliminary diagnostic model of the groundwater pipe network. Based on the validation training dataset, the preliminary diagnostic model is trained with strategy configuration to obtain the intelligent diagnostic model of the underground water pipe network. S104. Input the standard sound data into the intelligent diagnostic model to obtain the leakage status diagnosis result of the underground water pipe network; S105. Based on the spatial topology of the underground water pipe network and historical leakage logs, perform leakage verification on the leakage status diagnosis results to obtain leakage location information of the underground water pipe network. S106. Integrate the leakage status diagnosis results and leakage location information into a leakage diagnosis report for the underground water pipe network.

2. The method for diagnosing leakage in industrial underground water pipe networks based on an intelligent learning model as described in claim 1, characterized in that, The multi-scale feature analysis of the standard sound data yields leakage label data and non-leakage label data for the underground water pipe network, including: The temporal statistical features in the standard sound data are used as the temporal feature set of the standard sound data. The standard sound data is subjected to frequency domain energy distribution statistics to obtain the frequency domain feature set of the standard sound data; The time-domain feature set and the frequency-domain feature set are fused in multiple dimensions to construct a multi-dimensional fused feature vector of the standard sound data. Based on a pre-set expert knowledge rule base, pattern matching and classification are performed on the multi-dimensional fused feature vectors to obtain leakage label data and non-leakage label data of the underground water pipe network.

3. The method for diagnosing leakage in industrial underground water pipe networks based on an intelligent learning model as described in claim 1, characterized in that, The convolutional neural network-based model architecture initializes and configures the hierarchical parameters and connection methods of the bidirectional convolutional long short-term memory network to obtain the initial network framework of the underground water pipe network, including: Based on the model architecture of the convolutional neural network, the receptive field size and stride of the convolutional kernel of the model architecture are defined, and the number of feature maps output by the convolutional layers in the model architecture is determined to obtain the convolutional structure of the underground water pipe network. The bidirectional long short-term memory module is initialized and configured to obtain the temporal learning structure of the underground water pipe network; The output feature map sequence of the convolutional structure is used as the input of the temporal learning structure for connection configuration, thereby establishing the forward propagation path of data from feature extraction to temporal modeling; Based on the forward propagation path, the convolutional structure and the temporal learning structure are integrated and connected to obtain the initial network framework of the underground water pipe network.

4. The method for diagnosing leakage in industrial groundwater pipe networks based on an intelligent learning model as described in claim 3, characterized in that, The step of embedding an adaptive attention mechanism into the initial network framework to construct a feature weight allocation benchmark to obtain a preliminary diagnostic model of the groundwater pipeline network includes: Attention parameter matrices are added at predetermined positions in the initial network framework to enable training and learning from the data. The feature sequence output by the temporal learning structure is input into the attention parameter matrix; The attention weights of the input feature sequence are obtained by calculating the correlation between the feature vectors at different time steps in the input feature sequence using the attention parameter matrix. Based on the attention weights, the feature vectors at each time step in the input feature sequence are weighted and fused to obtain the context feature vector of the input feature sequence. The context feature vectors are then fed into the subsequent processing layer of the initial network framework to complete the embedding of the adaptive attention mechanism, thereby obtaining a preliminary diagnostic model of the underground water pipe network.

5. The method for diagnosing leakage in industrial underground water pipe networks based on an intelligent learning model as described in claim 4, characterized in that, The formula for calculating the attention weight is as follows: ； In the formula, For the first The feature vector at the i-th time step is related to the i-th time step. The attention weights of the feature vectors at each time step For the first The time step and the first Relevance score at each time step The total number of time steps for the input feature sequence. For the first The feature vector at the i-th time step is related to the i-th time step. The correlation score of the feature vectors at each time step It is a natural exponential function. This is a temporary index variable used for summation.

6. The method for diagnosing leakage in industrial underground water pipe networks based on an intelligent learning model as described in claim 1, characterized in that, The step of inputting the standard sound data into the intelligent diagnostic model to obtain the leakage status diagnosis result of the underground water pipe network includes: Temporal features are extracted from the standard sound data to construct a temporal feature sequence that matches the input dimension of the intelligent diagnostic model; The time-series feature sequence is input into the intelligent diagnostic model, and the forward propagation reasoning process of the intelligent diagnostic model is executed. Obtain the original diagnostic confidence vector generated by the output layer of the intelligent diagnostic model; Based on a preset diagnostic threshold, the original diagnostic confidence vector is used to determine its state, thereby obtaining the leakage status diagnosis result of the underground water pipe network.

7. The method for diagnosing leakage in industrial underground water pipe networks based on an intelligent learning model as described in claim 1, characterized in that, Based on the spatial topology of the underground water pipe network and historical leakage logs, the leakage status diagnosis results are verified to obtain leakage location information of the underground water pipe network, including: Based on the geographical coordinates and connection relationships of the monitoring points in the underground water pipeline network, a spatial topology map of the underground water pipeline network is constructed. Extract the locations of leakage events recorded in the historical time periods adjacent to the current diagnosis time period from the historical leakage log of the underground water pipe network as the historical leakage location set; The locations of monitoring points marked as having leaks in the leakage status diagnosis results are mapped onto the spatial topology map to obtain the candidate leakage point set of the underground water pipe network; Spatial topology verification is performed on the monitoring points of the candidate leakage point set; Based on the results of the spatial topology verification, the selected monitoring point location information is integrated into the leakage location information of the underground water pipe network.

8. The method for diagnosing leakage in industrial underground water pipe networks based on an intelligent learning model as described in claim 1, characterized in that, The process of integrating the leakage status diagnosis results with the leakage location information into a leakage diagnosis report for the underground water pipe network includes: According to the preset report template, the leakage probability and status category in the leakage status diagnosis result, as well as the suspected leakage pipe segment and confidence level in the leakage location information, are structured and mapped and filled. Based on the mapping and filling results, the correlation between the leakage status diagnosis results and the leakage location information is established and integrated to obtain the integrated data record of the underground water pipe network; The integrity and logical consistency of the fields in the integrated data records are verified to obtain a leakage diagnosis report for the underground water pipe network.

Images