Deep learning pipeline leak localization method based on threshold triggered hit event data

Abstract

The present application relates to the technical field of pipeline leakage monitoring and positioning, and particularly relates to a deep learning pipeline leakage positioning method based on threshold triggered Hit event data. The method comprises the following steps: obtaining multiple groups of Hit event data and performing time window decomposition; extracting two-channel statistical features and cross-channel correlation features of each time window to form a multi-dimensional feature vector, and a plurality of multi-dimensional feature vectors of continuous time windows form a sequence feature matrix as an original training sample; performing disturbance mapping on the cross-channel correlation features of the sequence feature matrix to generate an enhanced sample with a continuous position label; training a deep learning regression model by using the original training sample and the enhanced sample; inputting a to-be-tested sequence feature matrix into the trained model to output leakage position estimation values and leakage confidence of each time window, and adaptively fusing the outputs of adjacent time windows to obtain a final leakage position. The present application can realize stable and reliable pipeline leakage positioning under the condition of only relying on threshold triggered Hit event data.

Description

Technical Field

[0001] This invention relates to the field of duct and pressure pipeline leakage monitoring and location technology, specifically to a deep learning pipeline leakage location method based on threshold-triggered Hit event data. Background Technology

[0002] Existing pipeline leak detection systems often employ a threshold-triggered acquisition strategy in engineering sites, recording only Hit event parameters (such as timestamp, channel number, energy, amplitude, and RMS) that exceed a threshold, without saving continuous waveforms. While this improves transmission efficiency, it makes it difficult to apply methods that rely on complete waveforms for cross-correlation, time difference estimation, and fine-grained spectral analysis.

[0003] For data containing only Hit event tables, existing positioning methods mainly rely on the simple energy ratio method and the time difference method based on event pairing. The simple energy ratio method is easy to implement, but when there are low energy windows, noise disturbance windows, or large differences in triggering between the two channels, the fluctuation of the original logarithmic energy ratio intensifies, leading to unstable position estimation. The time difference method based on event pairing is susceptible to asynchronous triggering, event sparsity, and spurious events, often resulting in mismatches, omissions, or inability to pair.

[0004] Deep learning methods have been used to locate pipeline leaks, but most methods still take continuous waveforms as input, which makes it difficult to directly adapt to scenarios with only a sequence of hit events.

[0005] Therefore, how to achieve stable and reliable pipeline leak location by relying solely on threshold-triggered Hit event data is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a deep learning pipeline leak localization method based on threshold-triggered Hit event data. It can achieve stable and reliable pipeline leak localization by relying solely on threshold-triggered Hit event data, thereby improving localization accuracy and robustness.

[0007] To solve the above-mentioned technical problems, the technical solution of the present invention is: a deep learning pipeline leak localization method based on threshold-triggered Hit event data, comprising:

[0008] Multiple sets of Hit event data are acquired. Each set of data includes two-channel signals collected by sensors upstream and downstream of the pipeline leak location. Each set of data is decomposed into time windows with a fixed window length. For each time window, two-channel statistical features and cross-channel correlation features are extracted to form a multi-dimensional feature vector. The multi-dimensional feature vectors of several consecutive time windows constitute a sequence feature matrix, which serves as the original training sample. Among them, the cross-channel correlation features include the two-channel energy ratio after stability correction. The stability correction is used to suppress abnormal fluctuations in the energy ratio under low-energy windows and noise windows.

[0009] Based on the physical mapping relationship between pipe length and the energy ratio of the two channels, the cross-channel correlation features of the sequence feature matrix in the original training samples are perturbed and mapped to generate enhanced samples with continuous position labels.

[0010] A deep learning regression model is trained using the original training samples and augmented samples to obtain the trained model;

[0011] The Hit event data to be tested is constructed into a feature matrix of the sequence to be tested, which is input into the trained model. The model outputs the estimated leakage location and leakage confidence for each time window, and the outputs of adjacent time windows are smoothly fused by adaptive confidence index to obtain the final leakage location.

[0012] Furthermore, Hit event data includes event timestamp, channel number, energy, amplitude, and RMS parameter.

[0013] Furthermore, statistical characteristics include event count N, energy, and E. sum Energy mean E mean Maximum energy E max Energy standard deviation E std Maximum amplitude A max Amplitude mean A mean Amplitude standard deviation A std RMS mean mean Maximum RMS value max Mean of event timestamps μ γ σ, the standard deviation of event timestamps γ Mean event interval μ Δγ σ, the standard deviation of the event interval Δγ skewness of time distribution γ Kurt's time distribution γ .

[0014] Furthermore, cross-channel correlation characteristics also include the difference in the number of events. Energy difference E. Amplitude difference A. RMS difference RMS and timestamp mean difference .

[0015] Furthermore, the formula for calculating the energy ratio is:

[0016] ;

[0017] In the formula, This represents the k-th time window after decomposition with a fixed window length; This represents the sum of Hit event energy for sensor channel j within the k-th time window; This represents the energy stabilization term for channel j; A small bias term is introduced to improve feature stability.

[0018] Furthermore, based on the physical mapping relationship between pipe length and the energy ratio of the two channels, the cross-channel correlation features of the sequence feature matrix in the original training samples are perturbed and mapped to generate enhanced samples with continuous position labels; specifically including:

[0019] Based on the physical mapping relationship between pipe length and the energy ratio of the two channels, a virtual energy ratio is generated:

[0020] ;

[0021] In the formula, Indicates the virtual energy ratio. Indicates the virtual leak location; Represents the mapping coefficients; Indicates disturbance noise;

[0022] The perturbation formulas for the cross-channel correlation features other than the energy ratio are as follows:

[0023] ;

[0024] In the formula, This is a cross-channel correlation feature matrix. To and A random perturbation matrix with consistent dimensions, where each element is independently sampled from a small random distribution with zero mean.

[0025] Furthermore, the deep learning regression model comprises a one-dimensional convolutional neural network, a bidirectional gated recurrent unit, an additive attention mechanism, and a fully connected layer connected in sequence.

[0026] Furthermore, the training loss function for the deep learning regression model is:

[0027] ;

[0028] Indicates position loss. , This represents the absolute error between the estimated leak location and the actual location. This represents the squared error between the estimated and actual location of the leak.

[0029] Indicates confidence loss. 'c' represents the leakage confidence label. This represents the predicted leakage confidence value output by the model.

[0030] Furthermore, the outputs of adjacent time windows are smoothly fused using an adaptive confidence index to obtain the final leakage location; the specific formula is as follows:

[0031] ;

[0032] ;

[0033] In the formula, This represents the adaptive smoothing coefficient of the current window; Confidence weighting coefficient; β represents the base smoothing coefficient; This represents the predicted leakage confidence level for the current window. The output indicates the location of the leak after fusion; This indicates the predicted location of the leak in the current window. This indicates the previous historical output position that has already been merged.

[0034] By adopting the above technical solution, this invention does not rely on continuous original waveforms. It can locate pipeline leaks using only the Hit event table acquired by threshold triggering, reducing the burden of data transmission and storage and adapting to engineering site conditions. By correcting the stability of the cross-channel energy ratio, it effectively suppresses feature fluctuations under low energy windows and noise windows, providing a more robust input for the subsequent regression model. In continuous inference within the time window, the leakage confidence output by the deep learning model is directly introduced into the dynamic calculation of the smoothing coefficient, achieving adaptive fusion of fast tracking at high confidence and suppression of abnormal jumps at low confidence, thus improving the temporal consistency and anti-disturbance capability of the location results. At the same time, the method as a whole consists of lightweight feature extraction, physical mapping enhancement, deep learning regression, and adaptive post-processing. It has a simple structure, is easy to deploy and implement on existing acquisition systems, and has good engineering practicality. Attached Figure Description

[0035] Figure 1 Photos of a pipeline acoustic emission leak detection experiment.

[0036] Figure 2 Schematic diagram of the leak source;

[0037] Figure 3 This is a flowchart of the deep learning pipeline leak localization method based on threshold-triggered Hit event data of the present invention;

[0038] Figure 4 This is a comparison diagram of the actual and predicted leakage locations of the present invention.

[0039] Figure 5 This is a comparison chart of the positioning accuracy of the example and the comparative example;

[0040] Figure 6 This is a comparison chart of the cumulative distribution of absolute positioning errors between this embodiment and the comparative example.

[0041] Figure 7 This is a comparison chart of the output stability of this embodiment and the comparative embodiment. Detailed Implementation

[0042] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0043] like Figure 3 As shown, a deep learning-based pipeline leak localization method based on threshold-triggered Hit event data includes:

[0044] Multiple sets of Hit event data are acquired. Each set of data includes two-channel signals collected by sensors upstream and downstream of the pipeline leak location. Each set of data is decomposed into time windows with a fixed window length. For each time window, two-channel statistical features and cross-channel correlation features are extracted to form a multi-dimensional feature vector. The multi-dimensional feature vectors of several consecutive time windows constitute a sequence feature matrix, which serves as the original training sample. Among them, the cross-channel correlation features include the two-channel energy ratio after stability correction. The stability correction is used to suppress abnormal fluctuations in the energy ratio under low-energy windows and noise windows.

[0045] Based on the physical mapping relationship between pipe length and the energy ratio of the two channels, the cross-channel correlation features of the sequence feature matrix in the original training samples are perturbed and mapped to generate enhanced samples with continuous position labels.

[0046] A deep learning regression model is trained using the original training samples and augmented samples to obtain the trained model;

[0047] The Hit event data to be tested is constructed into a feature matrix of the sequence to be tested, which is input into the trained model. The model outputs the estimated leakage location and leakage confidence for each time window, and the outputs of adjacent time windows are smoothly fused by adaptive confidence index to obtain the final leakage location.

[0048] In this embodiment, preferably, the Hit event data includes an event timestamp, channel number, energy, amplitude, and RMS parameter.

[0049] In this embodiment, preferably, the statistical characteristics include event count N, energy, and E. sum Energy mean Emean Maximum energy E max Energy standard deviation E std Maximum amplitude A max Amplitude mean A mean Amplitude standard deviation A std RMS mean mean Maximum RMS value max Mean of event timestamps μ γ σ, the standard deviation of event timestamps γ Mean event interval μ Δγ σ, the standard deviation of the event interval Δγ skewness of time distribution γ Kurt's time distribution γ Each channel has 16-dimensional statistical characteristics.

[0050] In this embodiment, preferably, the cross-channel correlation feature also includes the event number difference. Energy difference E. Amplitude difference A. RMS difference RMS and timestamp mean difference The cross-channel correlation feature is a six-dimensional feature.

[0051] Each channel has 16-dimensional statistical features. Cross-channel correlation features are six-dimensional features. A single time window (which can be 10 seconds) has a 38-dimensional feature vector, and a sequence feature matrix can be formed by six consecutive time windows.

[0052] In this embodiment, preferably, the formula for calculating the energy ratio is:

[0053] ;

[0054] In the formula, This represents the k-th time window after decomposition with a fixed window length; This represents the sum of Hit event energy for sensor channel j within the k-th time window; This represents the energy stabilization term for channel j, and its values ​​can be... = 2=10 -6 ; A small bias term, with a value of 0.01, is introduced to improve feature stability.

[0055] By introducing regularization terms ε1 and ε2 into the two channels respectively and superimposing a bias correction term δ, the nonlinear amplification effect of low-energy windows, weak leakage windows and trigger windows near the threshold on energy ratio features is reduced without changing the overall framework of feature construction, thereby improving the stability of cross-channel correlation features.

[0056] In this embodiment, preferably, based on the physical mapping relationship between pipe length and the energy ratio of the two channels, the cross-channel correlation features of the sequence feature matrix in the original training samples are perturbed and mapped to generate enhanced samples with continuous position labels; specifically including:

[0057] Based on the physical mapping relationship between pipe length and the energy ratio of the two channels, a virtual energy ratio is generated:

[0058] ;

[0059] In the formula, Indicates the virtual energy ratio. Indicates the virtual leak location; Represents the mapping coefficient, and its values ​​can be: =0.9; This represents disturbance noise, and its value can be... =0.01;

[0060] The perturbation formulas for the cross-channel correlation features other than the energy ratio are as follows:

[0061] ;

[0062] In the formula, The cross-channel correlation feature matrix is ​​composed of cross-channel correlation features extracted in a 10-second time window. This data comes from the time window feature extraction results of real collected data. For each virtual leak location... Corresponding cross-channel correlation features In essence, it involves adding noise to the original true features, and is only used for data augmentation during the training phase. To and A uniformly dimensional random perturbation matrix, with each element independently sampled from a small random distribution with a zero mean, ranging from [−0.02, 0.02].

[0063] Because laboratory piping can only simulate leaks at fixed locations, the data type input to the neural network is relatively limited. Here, we primarily generate a single feature—the virtual energy ratio—based on the logic that there's an approximately linear mapping relationship between the leak location and the energy ratio between the two channels. Other cross-channel correlated features are only slightly modified. By continuously sampling within the range of [0, L], an enhanced sample covering the entire length of the pipeline is obtained, which makes up for the lack of data with only fixed locations under experimental conditions, and obtains a simulated leak point feature dataset within the entire pipeline to be tested.

[0064] In this embodiment, preferably, the deep learning regression model includes a one-dimensional convolutional neural network, a bidirectional gated recurrent unit, an additive attention mechanism, and a fully connected layer connected in sequence.

[0065] A one-dimensional convolutional neural network is used to extract local time window patterns, a bidirectional gated recurrent unit is used to capture cross-window temporal dependencies, and an additive attention mechanism is used to highlight key time windows that contribute significantly to leak localization. The model finally outputs the leak location estimate and leak confidence simultaneously through a fully connected layer.

[0066] In this embodiment, preferably, a joint optimization method of position loss and confidence loss is adopted during the training phase, and the training loss function of the deep learning regression model is:

[0067] ;

[0068] Indicates position loss. , This represents the absolute error between the estimated leak location and the actual location. This represents the squared error between the estimated and actual location of the leak.

[0069] Indicates confidence loss. 'c' represents the leakage confidence label; c=1 for leaked samples and c=0 for samples without leakage. This represents the predicted leakage confidence value output by the model. The value range is [0,1].

[0070] In this embodiment, preferably, the outputs of adjacent time windows are smoothly fused using an adaptive confidence index to obtain the final leakage location; the specific formula is as follows:

[0071] ;

[0072] ;

[0073] In the formula, This represents the adaptive smoothing coefficient of the current window; The confidence weighting coefficient can be 0.5; β represents the base smoothing coefficient, which can be 0.1. This represents the predicted leakage confidence level for the current window. The output indicates the location of the leak after fusion; This indicates the predicted location of the leak in the current window. This indicates the previous historical output position that has already been merged.

[0074] Specifically, considering the potential for abnormal fluctuations between outputs of adjacent time windows, directly outputting window by window would lead to unstable results. This step involves a weighted fusion of the current window's output and historical outputs to improve the robustness of the output. The specific fusion method involves constructing the smoothing formula above, resulting in:

[0075] when At higher levels, The corresponding increase allows the fused output to track the current prediction result more quickly; when At lower levels, The corresponding reduction means that the fusion output relies more on historical stable results, thereby reducing the impact of occasional abnormal windows on the final localization result.

[0076] The solutions involved in the above embodiments will be described in detail below with reference to specific examples and comparative examples.

[0077] The pipeline acoustic emission leak detection system includes sensors, a preamplifier, an acoustic emission signal acquisition instrument, a pressure leak test pipeline, and cables. Sensors are installed upstream and downstream of the leak point in the test pipeline. The signals detected by the sensors are transmitted to the acoustic emission signal acquisition instrument through the preamplifier.

[0078] The acoustic emission signal acquisition instrument is a PCA-6006 acoustic emission instrument from the United States, with 4 channels, 533MHz, Pentium CPU, filter operating frequency range of 60-500kHz, frequency response: 100KHz-2.1MHz+ / -1.0dB; minimum noise threshold: 18dB; maximum noise amplitude: 100dB; sensor frequency band: 20-100KHz; preamplifier: 40dB.

[0079] The pressure leakage test piping system includes an air compressor, valves, 20# carbon steel metal pipes, and a pressure gauge. The pressure pipe has an outer diameter of 32.85mm, a wall thickness of 3mm, and a yellow paint protective layer on the outer wall. The pressure pipe is a rectangular spiral, 5800mm long and 1500mm wide, equipped with nozzles and valves to facilitate simulating leakage sources. The pipe inlet is connected to the air compressor pump. The pressure gauge on the pipe allows for easy reading of the test pressure conditions. The pipe's variable pressure is 0-1.2MPa. The acoustic emission leakage detection test site is as follows: Figure 1 As shown.

[0080] Example 1: A smart method for duct leakage location based on threshold-triggered event characteristics, such as... Figure 3 As shown, it includes the following steps:

[0081] S1, as Figure 1 and Figure 2As shown, a 5.8m straight pipe section was taken on the air-filled pipeline system, and probes 1 and 2 (AE sensors) were placed at positions of 800mm and 5000mm respectively. Figure 2 The leak source shown is located at 2300 mm. The pipeline pressure is set to 0.3 MPa. The acquisition system uses a threshold-triggered method to record Hit event parameters, and the recorded items include at least the event timestamp, channel number, energy, amplitude, and RMS parameter. Experimental data include triggered leak conditions, continuous leak conditions, and no-leak background conditions.

[0082] S2, using a fixed window length of 10 s, decomposes the Hit event data into time windows and constructs a continuous time window sequence with adjacent time windows as sliding steps. For each time window, 16-dimensional statistical features are extracted from channel 1 and channel 2 respectively, and a 6-dimensional cross-channel correlation feature is constructed, thus forming a single-window 38-dimensional feature vector. Further, six consecutive time windows are stacked in chronological order to form a 6×38 sequence feature matrix, and multiple sequence feature matrices constitute the original training samples. For the k-th time window, let the sum of the energies of the two channels be... and The two-channel energy ratio after stability correction ,in and This is a regularization term with a value of 10. -6 , This is the bias correction term, with a value of 0.01;

[0083] S3. Based on the physical mapping relationship between pipeline length and leakage location, virtual leakage locations are sampled in the interval [0,L] to generate virtual energy ratio (target energy ratio). Small multiplicative random perturbations are applied to the other features in the sequence feature matrix except for the energy ratio to obtain enhanced samples with continuous location labels.

[0084] S4. The deep learning regression model is trained using the original training samples and augmented samples to obtain the trained model. The training phase adopts a joint optimization method of position loss and confidence loss.

[0085] S5: Construct the Hit event data to be tested into a feature matrix of the sequence to be tested, input it into the trained model, and output the estimated leakage location and leakage confidence for each time window;

[0086] Step S6: Perform adaptive confidence index smoothing fusion on the outputs of adjacent time windows to obtain the final leakage location.

[0087] The 60 sets of Hit event data obtained in the experiment were used as the original training samples. Based on the physical mapping relationship between pipe length and leak location, 120 sets of enhanced samples with continuous location labels were generated. The model was trained and tested using the original training samples and the enhanced samples. The results show that the leak location accuracy of the model (error not exceeding 0.2 m) reaches 96.2%, and the comparison between the actual leak location and the predicted location is as follows: Figure 4 As shown.

[0088] Comparative Example 1: The main difference between Comparative Example 1 and Example 1 is that the virtual leak location was not generated through the physical mapping relationship between pipe length and leak location. The model positioning accuracy (error not exceeding 0.2m) was 27.2%, which is much lower than that of Example 1. Specifically, as follows... Figure 5 As shown,

[0089] Comparative Example 2: The main difference between Comparative Example 2 and Example 1 is that the cross-channel feature energy ratio was not improved; the energy ratio was calculated using the traditional formula, which is: , and These are the energy levels of the two channels, respectively. Compared to the energy ratio in Example 2, the model positioning accuracy (error not exceeding 0.2m) is 90.9%, which is lower than that in Example 1, specifically as follows: Figure 5 As shown.

[0090] Comparative Example 3: The main difference between Comparative Example 3 and Example 1 is that adaptive confidence exponential smoothing was not performed; instead, fixed coefficient smoothing was used. The model localization accuracy (error not exceeding 0.2m) was 89.9%, which is lower than that of Example 1. Specifically, as follows... Figure 5 As shown.

[0091] Comparison of positioning accuracy between Example 1 and Comparative Examples 1, 1, and 3 Figure 6 As shown, Example 1 has the best positioning accuracy, followed by Comparative Example 2 and Comparative Example 3, while Comparative Example 1 has the worst accuracy.

[0092] Output stability comparison of Example 1 with Comparative Examples 2 and 3 Figure 7 As shown, the standard deviation of the output fluctuation in Example 1 is 0.097m, the standard deviation of the output fluctuation in Comparative Example 2 is 0.124m, and the standard deviation of the output fluctuation in Comparative Example 3 is 0.132m. It can be seen that Example 1 has higher stability and robustness compared with Comparative Example 2 and Comparative Example 3.

[0093] Through comparison and verification of Example 1, Comparative Example 1, Comparative Example 2 and Comparative Example 3, it is proven that the method proposed in this invention has good applicability and can achieve accurate location of duct leaks based on threshold recording Hit events.

[0094] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A deep learning-based pipeline leak localization method based on threshold-triggered Hit event data, characterized in that, include: Multiple sets of Hit event data are acquired. Each set of data includes two-channel signals collected by sensors upstream and downstream of the pipeline leak location. Each set of data is decomposed into time windows with a fixed window length. For each time window, two-channel statistical features and cross-channel correlation features are extracted to form a multi-dimensional feature vector. The multi-dimensional feature vectors of several consecutive time windows constitute a sequence feature matrix, which serves as the original training sample. Among them, the cross-channel correlation features include the two-channel energy ratio after stability correction. The stability correction is used to suppress abnormal fluctuations in the energy ratio under low-energy windows and noise windows. Based on the physical mapping relationship between pipe length and the energy ratio of the two channels, the cross-channel correlation features of the sequence feature matrix in the original training samples are perturbed and mapped to generate enhanced samples with continuous position labels. A deep learning regression model is trained using the original training samples and augmented samples to obtain the trained model; The Hit event data to be tested is constructed into a feature matrix of the sequence to be tested, which is input into the trained model. The model outputs the estimated leakage location and leakage confidence for each time window, and the outputs of adjacent time windows are smoothly fused by adaptive confidence index to obtain the final leakage location.

2. The deep learning pipeline leak localization method based on threshold-triggered Hit event data according to claim 1, characterized in that, Hit event data includes event timestamp, channel number, energy, amplitude, and RMS parameter.

3. The deep learning pipeline leak localization method based on threshold-triggered Hit event data according to claim 2, characterized in that, Statistical characteristics include event count N, energy, and E. sum Energy mean E mean Maximum energy E max Energy standard deviation E std Maximum amplitude A max Amplitude mean A mean Amplitude standard deviation A std RMS mean mean Maximum RMS value max Mean of event timestamps μ γ σ, the standard deviation of event timestamps γ Mean event interval μ Δγ σ, the standard deviation of the event interval Δγ skewness of time distribution γ Kurt's time distribution γ .

4. The deep learning pipeline leak localization method based on threshold-triggered Hit event data according to claim 3, characterized in that, Cross-channel correlation features also include event number difference Energy difference E. Amplitude difference A. RMS difference RMS and timestamp mean difference .

5. The deep learning pipeline leak localization method based on threshold-triggered Hit event data according to claim 1, characterized in that, The formula for calculating the energy ratio is: ； In the formula, This represents the k-th time window after decomposition with a fixed window length; This represents the sum of Hit event energy for sensor channel j within the k-th time window; This represents the energy stabilization term of channel j; A small bias term is introduced to improve feature stability.

6. The deep learning pipeline leak localization method based on threshold-triggered Hit event data according to claim 1, characterized in that, Based on the physical mapping relationship between pipe length and the energy ratio of the two channels, the cross-channel correlation features of the sequence feature matrix in the original training samples are perturbed and mapped to generate enhanced samples with continuous position labels; specifically including: Based on the physical mapping relationship between pipe length and the energy ratio of the two channels, a virtual energy ratio is generated: ； In the formula, Indicates the virtual energy ratio. Indicates the virtual leak location; Represents the mapping coefficients; Indicates disturbance noise; The perturbation formulas for the cross-channel correlation features other than the energy ratio are as follows: ； In the formula, This is a cross-channel correlation feature matrix. To and A random perturbation matrix with consistent dimensions, where each element is independently sampled from a small random distribution with zero mean.

7. The deep learning pipeline leak localization method based on threshold-triggered Hit event data according to claim 1, characterized in that, The deep learning regression model consists of a one-dimensional convolutional neural network, a bidirectional gated recurrent unit, an additive attention mechanism, and a fully connected layer connected in sequence.

8. The deep learning pipeline leak localization method based on threshold-triggered Hit event data according to claim 7, characterized in that, The training loss function for a deep learning regression model is: ； Indicates position loss. , This represents the absolute error between the estimated leak location and the actual location. This represents the squared error between the estimated and actual location of the leak. Indicates confidence loss. 'c' represents the leakage confidence label. This represents the predicted leakage confidence value output by the model.

9. The deep learning pipeline leak localization method based on threshold-triggered Hit event data according to claim 1, characterized in that, The outputs of adjacent time windows are smoothly fused using an adaptive confidence index to obtain the final leak location; the specific formula is as follows: ； ； In the formula, This represents the adaptive smoothing coefficient of the current window; Confidence weighting coefficient; β represents the base smoothing coefficient; This represents the predicted leakage confidence level for the current window. The output indicates the location of the leak after fusion. This indicates the predicted location of the leak in the current window. This indicates the previous historical output position that has already been merged.

Images