Pipe leakage detection method based on double-channel multi-head self-attention feature fusion

Abstract

The application discloses a pipeline leakage detection method based on double-channel multi-head self-attention feature fusion. The method comprises the following steps: firstly, using hydrophones arranged at both ends of the pipeline to collect corresponding pipeline acoustic signals; then, based on the two collected pipeline acoustic signals, corresponding feature vectors are extracted and a feature sample is constructed, and a training data set is constructed; then, the pipeline leakage detection model is trained using the training data set, and a trained pipeline leakage detection model is obtained; finally, the pipeline acoustic signals at both ends of the pipeline to be detected are continuously collected and corresponding feature samples are generated, the current feature sample is input into the trained pipeline leakage detection model, and the pipeline is continuously monitored and detected according to the pipeline detection result output by the model. The application solves the problem that the leakage point of the pipeline is difficult to identify in the prior art, and can significantly improve the accuracy and reliability of the pipeline leakage detection in an interference environment.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of pipeline safety monitoring, and particularly relates to a pipeline leakage detection method based on double-channel multi-head self-attention feature fusion. BACKGROUND

[0002] A pipeline is a critical infrastructure for transporting fluid media such as oil, natural gas, and water, and its safe operation is of great importance. Pipeline leakage not only causes huge economic losses and environmental pollution, but also may even lead to serious safety accidents. Therefore, timely and accurate detection of pipeline leakage is a core link to ensure pipeline safety.

[0003] Among numerous pipeline leakage detection technologies, the analysis method based on acoustic / vibration signals has become a mainstream research direction and an ideal choice for realizing real-time monitoring due to its non-invasive, high efficiency, and high accuracy. Early detection methods mainly rely on direct extraction of signal features from the time domain (such as standard deviation, root mean square) or frequency domain (such as Fourier transform). However, pipeline leakage signals are usually non-stationary and nonlinear, and are often disturbed by background noise generated by pumps, valves, and other sound sources, which limits the recognition ability of traditional time domain or frequency domain analysis methods under complex working conditions.

[0004] In order to more accurately capture the dynamic characteristics of leakage signals, researchers have turned to more complex time-frequency analysis methods. These methods can provide information on both time and frequency dimensions of signals, thereby enabling more in-depth feature extraction. For example, Short-time Fourier Transform (STFT), Wavelet Transform, Hilbert-Huang Transform (HHT), and other acoustic signal processing methods.

[0005] Although these pipeline leakage monitoring methods have made significant progress, they still face severe challenges in practical applications. First, in real pipeline network environments, weak leakage signals are easily overwhelmed by strong background noise (such as traffic, pump noise), making feature extraction difficult and reducing detection reliability. Second, many existing models have difficulty in effectively distinguishing between leakage signals and other non-leakage noise sources, resulting in high false positive rates. For cross-correlation-based positioning methods that rely on time delay estimation, when the leakage point is located at the midpoint of the double sensor or the pipeline has no leakage, the peak of the cross-correlation function appears at zero delay, causing "zero delay ambiguity", which makes the method unable to distinguish between these two completely different physical states, leading to serious misjudgment or omission.

[0006] Therefore, there is an urgent need for a pipeline leakage detection method that can break out of the traditional signal processing framework. SUMMARY

[0007] To address the shortcomings in accuracy and robustness of traditional time-frequency detection methods and the "zero-delay ambiguity" problem of cross-correlation methods, this invention provides a pipeline leak detection method based on dual-channel multi-head self-attention feature fusion. Starting from the intrinsic physical mechanism of signals, the proposed method can make accurate and robust judgments on the true state of pipelines even under strong noise and complex and variable field conditions, thus effectively overcoming the limitations of existing technologies.

[0008] The objective of this invention is achieved through the following technical solution:

[0009] I. A Pipeline Leakage Detection Method Based on Dual-Channel Multi-Head Self-Attention Feature Fusion

[0010] 1) Acquire the corresponding acoustic signals of the pipe using hydrophones installed at both ends of the pipe. Then, extract the corresponding feature vectors based on the two acquired acoustic signals of the pipe and construct a feature sample, thereby constructing a training dataset.

[0011] 2) Train the pipeline leak detection model using the training dataset to obtain a trained pipeline leak detection model;

[0012] 3) Continuously collect the acoustic signals of the pipeline at both ends of the pipeline to be tested and generate corresponding feature samples. Input the current feature samples into the trained pipeline leakage detection model and perform continuous monitoring and detection of the pipeline based on the pipeline detection results output by the model.

[0013] In step 1), the corresponding feature vectors are extracted based on the acquired acoustic signals from the two pipes, including:

[0014] The acoustic signals of the pipeline collected by the hydrophones at each end of the pipeline are preprocessed and then converted into log-Mel spectrograms. The log-Mel spectrograms are then input into the audio neural network, which outputs the corresponding feature vectors.

[0015] The pipeline leak detection model includes a connected Transformer encoder module and a classification layer; each feature sample is combined with a learnable classification token to construct a feature sequence, and a learnable position embedding is added to the feature sequence. Each feature sequence and its corresponding position embedding are used as input to the Transformer encoder module, and the Transformer encoder module outputs a classification token, which is used as input to the classification layer.

[0016] The continuous monitoring and detection of pipelines based on the pipeline detection results output by the model specifically includes:

[0017] If the model outputs a pipeline detection result indicating no leakage, then the model continues to collect the pipeline acoustic signals at both ends of the pipeline to be detected and generates the corresponding pipeline detection result; if the model outputs a pipeline detection result indicating leakage, then the model performs cross-correlation calculations based on the two pipeline acoustic signals of the pipeline to be detected to obtain the actual leakage point of the pipeline.

[0018] The step of obtaining the actual leak point of the pipeline by performing cross-correlation calculations based on the acoustic signals of the two pipelines to be tested includes:

[0019] First, bandpass filtering and fast Fourier transform are performed on the two acoustic signals of the pipeline to be tested to obtain the corresponding spectral functions. Then, the cross-power spectral density of the two spectral functions is calculated. Next, the cross-power spectral density is subjected to inverse fast Fourier transform to obtain the time-domain cross-correlation function. Then, the time delay corresponding to the peak of the time-domain cross-correlation function is found, and the location of the leak point is calculated by combining the sound speed and the sensor distance.

[0020] II. A Pipeline Leak Detection System Based on Dual-Channel Multi-Head Self-Attention Feature Fusion

[0021] The acoustic signal acquisition unit is used to acquire the acoustic signals of the pipe at both ends.

[0022] The feature sample construction unit is used to extract the corresponding feature vectors based on the acquired acoustic signals from the two pipes and construct a feature sample.

[0023] The pipeline leak detection unit is used to store the pipeline leak detection model and train the model using a training dataset composed of feature samples.

[0024] The pipeline location unit is used to monitor pipelines and locate leak points based on the pipeline detection results output by the pipeline leak detection model.

[0025] III. A computer device

[0026] The device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the pipeline leak detection method based on dual-channel multi-head self-attention feature fusion.

[0027] IV. A computer-readable storage medium

[0028] The medium stores a computer program, which, when executed by a processor, implements the steps of the pipeline leak detection method based on dual-channel multi-head self-attention feature fusion.

[0029] V. A computer program product

[0030] The product includes a computer program / instruction that, when executed by a processor, implements the steps of the pipeline leak detection method based on dual-channel multi-head self-attention feature fusion.

[0031] The beneficial effects of this invention are:

[0032] 1. To address the "zero-delay ambiguity" problem of cross-correlation classification methods, this invention utilizes the feature extraction capabilities of neural networks to classify the original problem. Compared to traditional methods that optimize within the existing framework, this invention can more accurately distinguish between leakage sound and background noise, significantly reducing the false alarm rate and false negative rate of the system.

[0033] 2. This invention no longer analyzes signals from a single sensor in isolation, but instead introduces a Transformer encoder architecture. By constructing an input feature sequence containing classification tokens and dual-channel features, and utilizing a multi-head self-attention mechanism, the model can explicitly model the intrinsic physical correlation and propagation characteristics between the pipe acoustic signals collected by the two hydrophones. This deep interaction based on the attention mechanism achieves more efficient information complementarity and nonlinear fusion, thereby making more robust and accurate joint judgments.

[0034] 3. This invention utilizes PANNs models pre-trained on massive amounts of general audio data for feature extraction. Consequently, only a neural network module containing a Transformer encoder module and a classification layer (such as MLP) needs to be trained for a specific task, which significantly reduces the technical threshold and time cost of model development. Attached Figure Description

[0035] Figure 1 This is a schematic diagram illustrating an application scenario of the pipeline leak detection system provided in an embodiment of the present invention.

[0036] Figure 2 A flowchart of a pipeline leak detection system provided in an embodiment of the present invention.

[0037] Figure 3 This is a schematic diagram of the network architecture of the pipeline leakage detection model provided in an embodiment of the present invention.

[0038] Figure 4 This is a flowchart of the cross-correlation localization method.

[0039] Figure 5 The curves are the result of cross-correlation calculations on the acoustic signals of the pipeline collected by two hydrophones. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] like Figure 2 and Figure 3 As shown, the pipeline leakage detection method based on dual-channel multi-head self-attention feature fusion proposed in this invention includes the following steps:

[0042] 1) such as Figure 1 As shown, a hydrophone is deployed on a section of water pipe. The hydrophone is used to acquire the corresponding acoustic signals of the pipe in real time. Then, based on the two acquired acoustic signals, feature vectors are extracted to construct a feature sample, which is then used to build a training dataset. Each training sample in the training dataset consists of a feature sample and a corresponding label. The label specifically indicates whether the pipe is leaking or not.

[0043] Alternatively, the hydrophone is designed to be installed in an invasive manner on a pipe, with the probe in contact with the water to listen for sound signals in the water.

[0044] In one feasible implementation, feature vectors are extracted based on the acquired acoustic signals from the two pipes, including:

[0045] The acoustic signals from the pipes, collected by hydrophones at each end of the pipe, are preprocessed and then converted into log-Mel spectrograms. These log-Mel spectrograms are then input into pre-trained audio neural networks (PANNs). The PANNs output high-dimensional feature vectors representing the intrinsic acoustic characteristics of the pipe's acoustic signals. The preprocessing includes signal denoising, signal normalization, signal resampling, and slicing. Signal resampling aims to adjust the signal's sampling rate to match the sampling rate required by the PANN input; both sampling rates must be the same. Slicing involves dividing or padding the signal to form fixed-length audio segments.

[0046] The audio neural network architecture is CNN14, denoted as PANNs-CNN14. This architecture is a 14-layer convolutional neural network specifically designed for audio pattern recognition. It is pre-trained using a large and diverse audio dataset. The network parameters of PANNs-CNN14 are shown in Table 1. It consists of six convolutional blocks and pooling layers stacked alternately. Each convolutional block contains two 3×3 convolutional layers, supplemented by batch normalization (BN) and ReLU activation functions. After a series of convolution and pooling operations, a global pooling layer aggregates the feature maps. This is followed by an embedding layer containing 2048 neurons, the output of which is the 2048-dimensional high-dimensional feature vector extracted in this invention. The network weights of PANNs-CNN14 are frozen and used only as a fixed feature extractor.

[0047] Table 1 Network Parameter Table of Audio Neural Network

[0048]

[0049] 2) Train the pipeline leak detection model using the training dataset to obtain a trained pipeline leak detection model;

[0050] In one feasible implementation, the pipeline leak detection model includes a connected Transformer encoder module and a classification layer. Each feature sample is combined with a learnable classification token ([CLS] Token) to construct a feature sequence of length 3, and a positional encoding is added to this feature sequence. Each feature sequence and its corresponding positional encoding are used as input to the Transformer encoder module, which outputs a classification token (2048 dimensions) as input to the classification layer. This invention achieves deep interaction and nonlinear fusion of dual-channel features through the self-attention mechanism of the Transformer encoder module.

[0051] Specifically:

[0052] The core of the Transformer encoder module lies in the multi-head self-attention mechanism, which allows the model to focus on information at different locations in the input sequence in order to capture long-distance dependencies.

[0053] In this embodiment, the input feature sequence is first mapped to a query matrix (Q), a key matrix (K), and a value matrix (V) using three linear transformation matrices. To prevent the inner product from becoming too large and causing the Softmax function to enter the saturation region, this invention employs a scaled dot product attention mechanism. The calculation formula is as follows:

[0054]

[0055] in, Let be the dimension of the key vector. Scaling factor This indicates the transpose operation. This represents the Softmax function.

[0056] The multi-head attention mechanism projects the query matrix Q, key matrix K, and value matrix V onto the feature dimension h times (i.e., h heads), performs the attention calculations mentioned above on each head separately, and finally concatenates the outputs of all heads before performing linear projection, thereby extracting a rich feature representation containing dual-channel signal interaction information.

[0057] The classification layer (MLP) receives the classification token ([CLS] Token) output by the Transformer encoder module. After dimensionality reduction by the fully connected layer and processing by the sigmoid activation function within the classification layer, it outputs a probability value between 0 and 1. This probability value represents the likelihood of a leak in the pipe.

[0058] During the model training phase, this invention employs the Binary Cross-Entropy Loss (BCE Loss) function as the objective function to measure the difference between the model's predicted output and the true label. The BCE Loss function effectively handles binary classification problems, and its mathematical expression is as follows:

[0059]

[0060] in, This represents the average loss value of the current batch; N represents the batch size during training. This represents the true label of the i-th sample when there is a leak in the pipe. =1, no leakage =0; This represents the probability that the model predicts the i-th sample as "leaking".

[0061] This loss function penalizes samples that are predicted incorrectly (i.e., when...). but Close to 0, or but When the value approaches 1, the loss value increases significantly, driving the model parameters to be continuously updated during backpropagation, making the predicted probability distribution approximate the true distribution.

[0062] In terms of specific model training parameter settings, this invention employs the Adam (Adaptive Moment Estimation) optimizer for gradient descent optimization. The Adam optimizer combines the advantages of momentum and RMSProp algorithms, enabling adaptive adjustment of the learning rate for each parameter. The specific training hyperparameter settings are as follows:

[0063] Initial learning rate: set to It also incorporates a cosine annealing strategy for dynamic adjustments to more precisely search for the optimal solution in the later stages of training.

[0064] Batch Size: Set to 64 to balance memory usage and gradient descent stability.

[0065] Training epochs: Set to 100 epochs.

[0066] Regularization strategy: To prevent overfitting, a Dropout layer with a dropout rate of 0.1 is introduced into the Transformer encoder module, and the weight decay coefficient is set in the optimizer. .

[0067] The training process specifically includes: randomly dividing the constructed training dataset into a training set and a validation set in an 8:2 ratio. In each training round, the training set is input into the model to calculate the loss function and update the weights; then, the validation set is used to evaluate the model's accuracy, precision, recall, and F1 score. If the loss value on the validation set does not decrease within 10 consecutive epochs, an early stopping mechanism is triggered to preserve the model parameters with optimal generalization ability.

[0068] The pipeline leakage detection model of this invention has an end-to-end architecture. It uses a unified loss function (binary cross-entropy) and backpropagation algorithm to jointly optimize and train the Transformer encoder module and classification layer until the model converges.

[0069] 3) Continuously collect the acoustic signals of the pipeline at both ends of the pipeline to be tested and generate corresponding feature samples. Input the current feature samples into the trained pipeline leakage detection model and perform continuous monitoring and detection of the pipeline based on the pipeline detection results output by the model.

[0070] In one feasible implementation, continuous monitoring and inspection of the pipeline are performed based on the pipeline inspection results output by the model, specifically including:

[0071] The model outputs either a leaking or non-leaking pipe detection result. If the model outputs a non-leaking pipe detection result, the acoustic signals from both ends of the pipe to be detected are collected and the corresponding pipe detection result is generated. If the model outputs a leaking pipe detection result, the actual leak point of the pipe is obtained by cross-correlation calculation based on the two acoustic signals of the pipe to be detected.

[0072] Based on the cross-correlation calculation of the acoustic signals from the two pipes under test, the actual leak point of the pipe is obtained, including:

[0073] (A) Bandpass filtering:

[0074] Considering that pipeline leakage signals (especially buried pipelines) typically exhibit narrowband low-frequency characteristics and are easily affected by environmental noise, directly performing cross-correlation calculations may result in indistinct peak values. Therefore, the two acquired raw discrete signals are first analyzed... and (i.e., the acoustic signals from the two pipelines) undergo bandpass filtering. In this embodiment, a fourth-order Butterworth bandpass filter is used, with its bandpass frequency range set to [value missing]. , It has two bandpass frequencies to effectively filter out fluid flow noise (very low frequency) and external environmental interference (high frequency).

[0075] (B) Calculation of cross-correlation function based on FFT:

[0076] To improve computational efficiency and facilitate frequency domain weighting, this invention utilizes Fast Fourier Transform (FFT) to calculate the cross-correlation function. The specific process is as follows: Figure 4 As shown.

[0077] First, perform Fast Fourier Transform on the two filtered signals to obtain two spectral functions. and .

[0078] Next, calculate the two spectral functions. and cross power spectral density The calculation formula is as follows:

[0079]

[0080] in, Represents the spectrum function The complex conjugate of the peaks. To further sharpen the correlation peaks, a weighting function (such as PHAT weighting) can be introduced here for generalized cross-correlation processing.

[0081] Finally, the obtained cross-power spectral density is subjected to inverse fast Fourier transform (IFFT) to obtain the time-domain cross-correlation function. :

[0082]

[0083] Where j represents the imaginary sign; f represents the frequency; and τ represents the time delay.

[0084] The cross-correlation function characterizes the similarity between two signals at different time delays τ.

[0085] (C) Time Delay Estimation:

[0086] Traversing cross-correlation functions Find the moment corresponding to the maximum amplitude, that is, the time difference between the arrival of the signal at the two sensors. :

[0087]

[0088] when At that time, the cross-correlation coefficient reaches its peak, such as Figure 5 As shown. Among them, This indicates the distance from the leak point to the downstream sensor (second hydrophone).

[0089] (D) Leakage location:

[0090] Combining the speed of sound wave propagation in a fluid-filled pipe Given the total distance d between the two hydrophones (i.e., the known total length of the hydrophones laid along the pipe at both ends), calculate the distance from the leak point to the upstream sensor (the first hydrophone). The positioning formula is as follows:

[0091]

[0092] In this embodiment, the speed of sound The calculation can be performed using theoretical formulas based on the pipe material (such as PVC or steel pipe) and fluid parameters. ,in, For the speed of sound in fluid, For fluid bulk modulus, The Young's modulus of the pipe wall. and These are the pipe radius and wall thickness, respectively. Velocity of sound. It can also be obtained through field measurement and calibration at a known location by simulating a leak source.

[0093] Because this invention has eliminated the "no-leakage" state through the pipeline leak detection model, the zero-delay peak value appearing in the cross-correlation calculation can be identified as a "midpoint leak". Finally, the actual leak point of the pipeline detected and located is output, and an alarm is triggered to notify the operation and maintenance personnel to handle the situation.

[0094] This invention proposes a pipeline leak detection system based on dual-channel multi-head self-attention feature fusion, comprising:

[0095] The acoustic signal acquisition unit is used to acquire the acoustic signals of the pipe at both ends.

[0096] The feature sample construction unit is used to extract the corresponding feature vectors based on the acquired acoustic signals from the two pipes and construct a feature sample.

[0097] The pipeline leak detection unit is used to store the pipeline leak detection model and train the model using a training dataset composed of feature samples.

[0098] The pipeline location unit is used to monitor pipelines and locate leak points based on the pipeline detection results output by the pipeline leak detection model.

[0099] This invention proposes a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of a pipeline leak detection method based on dual-channel multi-head self-attention feature fusion.

[0100] This invention proposes a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a pipeline leak detection method based on dual-channel multi-head self-attention feature fusion.

[0101] This invention proposes a computer program product. The computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of a pipeline leak detection method based on dual-channel multi-head self-attention feature fusion.

[0102] This invention combines the powerful feature representation capabilities of deep learning with traditional signal processing methods. It first uses neural networks to solve the classification problem of "whether there is a leak" and then uses traditional methods to solve the location problem of "where" after confirming the leak, thus forming a complete, efficient and reliable leak detection solution.

[0103] Finally, it should be noted that the above embodiments and descriptions are only used to illustrate the technical solutions of the present invention and not to limit it. 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 disclosure of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the protection scope of the claims of the present invention.

Claims

1. A pipeline leak detection method based on dual-channel multi-head self-attention feature fusion, characterized in that, Includes the following steps: 1) Acquire the corresponding acoustic signals of the pipe using hydrophones installed at both ends of the pipe. Then, extract the corresponding feature vectors based on the two acquired acoustic signals of the pipe and construct a feature sample, thereby constructing a training dataset. Based on the acquired acoustic signals from the two pipes, the corresponding feature vectors are extracted, including: The acoustic signals of the pipeline collected by the hydrophones at each end of the pipeline are preprocessed and then converted into log-Mel spectrograms. The log-Mel spectrograms are then input into the audio neural network, which outputs the corresponding feature vectors. 2) Train the pipeline leak detection model using the training dataset to obtain a trained pipeline leak detection model; The pipeline leak detection model includes a connected Transformer encoder module and a classification layer; each feature sample is combined with a learnable classification token to construct a feature sequence, and a learnable position embedding is added to the feature sequence. Each feature sequence and its corresponding position embedding are used as input to the Transformer encoder module, and the Transformer encoder module outputs a classification token, which is used as input to the classification layer. 3) Continuously collect the acoustic signals of the pipeline at both ends of the pipeline to be tested and generate corresponding feature samples. Input the current feature samples into the trained pipeline leakage detection model and perform continuous monitoring and detection of the pipeline based on the pipeline detection results output by the model. The continuous monitoring and detection of pipelines based on the pipeline detection results output by the model specifically includes: If the model outputs a pipeline detection result indicating no leakage, then the model continues to collect the pipeline acoustic signals at both ends of the pipeline to be detected and generates the corresponding pipeline detection result; if the model outputs a pipeline detection result indicating leakage, then the model performs cross-correlation calculations based on the two pipeline acoustic signals of the pipeline to be detected to obtain the actual leakage point of the pipeline.

2. The pipeline leakage detection method based on dual-channel multi-head self-attention feature fusion according to claim 1, characterized in that, The step of obtaining the actual leak point of the pipeline by performing cross-correlation calculations based on the acoustic signals of the two pipelines to be tested includes: First, bandpass filtering and fast Fourier transform are performed on the two acoustic signals of the pipeline to be tested to obtain the corresponding spectral functions. Then, the cross-power spectral density of the two spectral functions is calculated. Next, the cross-power spectral density is subjected to inverse fast Fourier transform to obtain the time-domain cross-correlation function. Then, the time delay corresponding to the peak of the time-domain cross-correlation function is found, and the location of the leak point is calculated by combining the sound speed and the sensor distance.

3. A pipeline leak detection system based on dual-channel multi-head self-attention feature fusion for implementing the method of claim 1, characterized in that, include: The acoustic signal acquisition unit is used to acquire the acoustic signals of the pipe at both ends. The feature sample construction unit is used to extract the corresponding feature vectors based on the acquired acoustic signals from the two pipes and construct a feature sample. The pipeline leak detection unit is used to store the pipeline leak detection model and train the model using a training dataset composed of feature samples. The pipeline location unit is used to monitor pipelines and locate leak points based on the pipeline detection results output by the pipeline leak detection model.

4. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the pipeline leakage detection method based on dual-channel multi-head self-attention feature fusion as described in any one of claims 1 to 2.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the pipeline leakage detection method based on dual-channel multi-head self-attention feature fusion as described in any one of claims 1 to 2.

6. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the pipeline leak detection method based on dual-channel multi-head self-attention feature fusion as described in any one of claims 1 to 2.

Images