Method for rapidly identifying and positioning air leakage point of vacuum film in vacuum preloading project based on artificial intelligence assistance

Abstract

The invention discloses a method for rapidly identifying and positioning a vacuum film air leakage point in vacuum preloading engineering based on artificial intelligence assistance. The method comprises the steps that S1, system initialization and self-inspection are carried out; s2, the unmanned aerial vehicle adopts a rasterized route to cover the whole vacuum preloading site, and cruises the whole area through a back-and-forth parallel sweeping method; s3, sound signals are collected in real time through an audio collection module, and the collected sound signals are uploaded to a signal processing unit; s4, preprocessing sound data, extracting an MFSC coefficient from a preprocessed sound signal, inputting the MFSC coefficient into the well-trained convolution-cycle-attention mechanism hybrid neural network model, and distinguishing gas leakage sound and noise; s5, when the signal processing unit detects suspected air leakage sound, triggering the unmanned aerial vehicle to hover and circle, further determining the position of an air leakage point by utilizing array orientation, and then performing film repairing; meanwhile, the ground monitoring station receives the position and air leakage alarm information of the unmanned aerial vehicle in real time and visually displays the information on an interface.

Description

Technical Field

[0001] This invention belongs to the field of geotechnical engineering technology, specifically relating to a method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading projects based on artificial intelligence assistance. Background Technology

[0002] Vacuum preloading is a foundation treatment technology suitable for soft soil foundations with high water content. Due to its mature technology and reliable results, it is widely used in the eastern coastal areas of my country. During vacuum preloading, drainage boards are vertically inserted into the soft soil to be treated and connected to a vacuum pump via vacuum pipelines on the ground surface. After the drainage boards and pipelines are laid, a vacuum membrane is placed over the ground surface, completely isolating the construction system from the atmosphere. During construction, the vacuum generated by the vacuum pump is transmitted through pipelines to the membrane and further through the drainage boards to the deeper soil layers. This promotes pore water seepage to the drainage boards and discharge to the surface through pipelines, allowing the soil to drain and consolidate, ultimately increasing the strength of the foundation soil.

[0003] To ensure efficient drainage and consolidation of the soil during vacuum preloading, a high vacuum level must be maintained in the membrane construction system. Therefore, the integrity and sealing of the impermeable vacuum membrane are crucial for achieving this high vacuum. However, after vacuuming, the vacuum membrane shrinks and adheres tightly to the soil surface. Soft soil foundations experience significant settlement and deformation during treatment, and coupled with on-site foot traffic and mechanical impacts, vacuum membrane damage and leakage have become common problems during construction. Typically, monitoring for vacuum membrane leakage involves manual inspection, listening for leaks on the ground surface, and then repairing the membrane. However, for large-scale soft soil foundation treatment projects, manually checking for vacuum membrane leaks is extremely difficult and time-consuming, significantly hindering construction efficiency and increasing labor and time costs.

[0004] To address the aforementioned issues, there is an urgent need to design a method for identifying air leakage in the vacuum membrane during vacuum preloading, so as to achieve efficient leak repair at the vacuum preloading construction site. Summary of the Invention

[0005] To address the aforementioned technical problems in existing technologies, this invention provides a method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading projects using artificial intelligence-assisted technology. This invention utilizes a drone equipped with an audio sensor to collect on-site sound signals, converts them from the time domain to the frequency domain, and extracts the energy spectrum. Taking advantage of the specific frequency range of vacuum membrane leak sounds, and leveraging the feature learning capabilities of a hybrid neural network with a convolutional-recurrent-attention mechanism, the invention effectively distinguishes between membrane leak sounds and non-leak sounds. The leak point is then located based on the direction of the sound source. After location, manual membrane repair at the specific point is performed, achieving efficient detection of vacuum membrane leaks at the vacuum preloading construction site.

[0006] The technical solution adopted in this invention is: A method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading engineering using artificial intelligence-assisted methods, characterized by the following steps: S1. System initialization and self-test; S2. The UAV uses a gridded flight path to cover the entire vacuum pre-compression site and cruises the entire area using a back-and-forth parallel sweep method. S3. Acquire sound signals in real time through the audio acquisition module and upload the acquired sound signals to the signal processing unit; S4. Preprocess the sound data, extract the MFSC coefficients from the preprocessed sound signal, and input them into a well-trained convolutional-recurrent-attention hybrid neural network model to distinguish between air leakage sound and noise. S5. When the signal processing unit detects a suspected air leak sound, it triggers the drone to hover and circle, and uses array orientation to further determine the location of the air leak before performing membrane repair. At the same time, the ground monitoring station receives the drone's location and air leak alarm information in real time and displays them intuitively on the interface, realizing visualization and human-machine interaction.

[0007] Furthermore, in step S1, system initialization and self-test specifically include: power supply check of hardware devices, communication line connection test, sensor function verification, and software system loading and verification; after debugging, a free flight test is conducted to test the coverage and false alarm rate of the UAV cruise monitoring.

[0008] Furthermore, in step S4, the preprocessing of the audio data specifically includes filtering, denoising, and normalization to ensure data quality.

[0009] Furthermore, in step S2, the UAV flies at an altitude of 0.5-1.0 meters and a speed of 1-2 m / s.

[0010] Furthermore, in step S3, the audio acquisition module is composed of high-precision microphones arranged in an array, which simultaneously receive sound signals and acquire sound signals at a high sampling rate; the high-precision microphones upload the acquired sound data to the signal processing unit, preprocess the audio data and extract features, and input it into the trained convolution-recurrent-attention mechanism hybrid neural network model to determine whether there is air leakage and the corresponding confidence level.

[0011] Furthermore, in step S4, the specific process for preprocessing the acquired raw audio signal is as follows: the data is divided into frames of appropriate length with overlap between frames to maintain time resolution; a pre-designed bandpass filter is applied to each frame of audio to enhance wind noise resistance; normalization is then performed to balance the volume level and prevent model deviation due to distance or gain differences; a short-time Fourier transform (STFT) is then performed on the audio signal to convert each frame of audio signal into a two-dimensional time-frequency graph; and finally, Mel frequency spectrum coefficients are extracted as features to distinguish the air leakage sound from other sounds.

[0012] Furthermore, in step S4, a hybrid neural network model with convolutional-recurrent-attention mechanisms is selected as the algorithm for air leakage sound diagnosis. The specific algorithm includes: (1) Model building: Input layer: Inputs preprocessed feature data; Convolutional layer: After inputting feature data, features are extracted through a set of one-dimensional convolutional kernels; Pooling layer: Reduces the dimensionality of the feature sequence output by the convolutional layer, thus reducing computational cost; Bidirectional LSTM layer: Further models the feature sequences extracted by the convolutional layer to capture the long-distance dependencies of the audio signal in the time dimension; Attention mechanism layer: Weights the importance of each time step, focusing on moments highly correlated with leakage; Fully connected classification layer: transforms the attention-weighted feature vectors into a probability distribution; Output layer: Outputs the final prediction result; (2) Loss function: The cross-entropy loss function is used to measure the difference between the class probability distribution output by the model and the true label; Optimization algorithm: The Adam optimizer is used to update model parameters, accelerating the convergence speed; Batch training: The preprocessed audio signal is divided into multiple batches for training, and each batch contains a certain number of samples; (3) Model evaluation: The model performance is evaluated using accuracy, average loss and confidence metrics; Model optimization: Data augmentation is used to diversify training samples and improve the model's generalization ability; the model structure, parameters or training strategies are adjusted based on the evaluation results to improve diagnostic accuracy.

[0013] Furthermore, in step S2, the microphone array is installed on the drone, and the microphone array is used to determine the direction of the sound source. Based on the spatial time delay sound source localization model, the generalized cross-correlation function is used for time delay estimation, and then a beamforming method is introduced to improve the ability to locate the air leak point.

[0014] Furthermore, in step S5, the ground monitoring station consists of PC-based real-time monitoring software, which implements a cross-platform interface based on Web technology; The cross-platform interface includes map display, drawing of the drone's current location and flight trajectory, marking of leak points and the time and coordinates of suspected leak points, drone status information and environmental parameters, and a list of alarm logs, recording the time, coordinates, and intensity of each leak event for on-site personnel to archive and analyze.

[0015] A system for implementing the above-described method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading engineering based on artificial intelligence is characterized by comprising: The drone platform is equipped with an audio acquisition module and a signal processing unit. It uses a gridded flight path with parallel navigation to cover the entire vacuum pre-compression area. In the whole architecture, the drone performs the main calculations and realizes real-time local decision-making, while the ground end performs visualization and human-machine interaction.

[0016] The audio acquisition module is used to acquire sound signals in real time and transmit the acquired sound signals to the signal processing unit; The signal processing unit is used to preprocess the acquired sound signals, including filtering, denoising, and normalization, to ensure data quality; it extracts MFSC coefficients from the preprocessed sound data and inputs them into a well-trained convolutional-recurrent-attention hybrid neural network model (CNN-LSTM-Attention) to distinguish between air leaks and noise. The communication and ground monitoring center receives real-time information on the drone's location and air leak alarms, and displays it intuitively on the interface.

[0017] The technical concept of this invention is to adopt a two-layer mechanism of "detection + positioning". After the detection model detects the sound of air leakage, the microphone array positioning algorithm is immediately activated to provide the spatial coordinates of the air leakage point. The positioning result is fed back to the ground monitoring station to guide the UAV to conduct detailed inspection and mark alarms. The acoustic imaging method of array beamforming is used to overlay the sound source location on the map, reducing the inspection time of the inspection personnel and further improving the efficiency and speed of the repair work.

[0018] Compared with the prior art, the beneficial effects of the present invention are reflected in: 1. In this invention, a hybrid neural network model is used to automatically learn and identify air leak sounds from complex background noise. Compared with the traditional method of listening with the human ear, the hybrid neural network model with convolutional-recurrent-attention mechanism (CNN-LSTM-Attention) has powerful feature extraction capabilities and environmental adaptability. It works closely with the UAV flight control to greatly improve the ability to perceive air leaks. 2. In this invention, a dual-layer mechanism of "detection + positioning" is adopted. After the detection model detects the sound of air leakage, the microphone array positioning algorithm is immediately activated to provide the spatial coordinates of the air leakage point. The positioning result is fed back to the ground monitoring station to guide the UAV for detailed inspection and marking alarms. The acoustic imaging method of array beamforming is used to overlay the sound source location on the map, reducing the inspection time of the inspection personnel and further improving the efficiency and speed of the repair work. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the implementation of the method of this invention; Figure 2 This is a flowchart of the leakage identification and location process of the present invention. Detailed Implementation

[0020] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0022] The present invention will now be described in detail with reference to the accompanying drawings and exemplary embodiments.

[0023] refer to Figure 1 and Figure 2 The present invention discloses a method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading engineering based on artificial intelligence, the method comprising the following steps: S1. System initialization and self-test; S2. The UAV uses a gridded flight path to cover the entire vacuum pre-compression site and cruises the entire area using a back-and-forth parallel sweep method. S3. Acquire sound signals in real time through the audio acquisition module and upload the acquired sound signals to the signal processing unit; S4. Preprocess the sound data, extract the MFSC coefficients from the preprocessed sound signal, and input them into a well-trained convolutional-recurrent-attention hybrid neural network model to distinguish between air leakage sound and noise. S5. When the signal processing unit detects a suspected air leak sound, it triggers the drone to hover and circle, and uses array orientation to further determine the location of the air leak before performing membrane repair. At the same time, the ground monitoring station receives the drone's location and air leak alarm information in real time and displays them intuitively on the interface, realizing visualization and human-machine interaction.

[0024] In one embodiment, in step S1, system initialization and self-test specifically include: power supply check of hardware devices, communication line connection test, sensor function verification, and software system loading and verification; after debugging, a free flight test is performed to test the coverage and false alarm rate of the UAV cruise monitoring.

[0025] In one embodiment, step S4 specifically includes preprocessing the audio data, including filtering, denoising, and normalization, to ensure data quality.

[0026] In one embodiment, in step S2, the drone flies at an altitude of 0.5-1.0 meters and at a speed of 1-2 m / s.

[0027] In one embodiment, in step S3, the audio acquisition module is composed of high-precision microphones arranged in an array, which simultaneously receive sound signals and acquire sound signals at a high sampling rate; the high-precision microphones upload the acquired sound data to the signal processing unit, preprocess the audio data and extract features, and input it into the trained convolution-recurrent-attention mechanism hybrid neural network model to determine whether there is air leakage and the corresponding confidence level.

[0028] In one embodiment, the specific process for preprocessing the acquired raw audio signal in step S4 is as follows: the data is divided into frames of appropriate length with overlap between frames to maintain time resolution; a pre-designed bandpass filter is applied to each frame of audio to enhance wind noise resistance; normalization is then performed to balance the volume level and prevent model deviation due to distance or gain differences; a short-time Fourier transform (STFT) is then performed on the audio signal to convert each frame of audio signal into a two-dimensional time-frequency graph; and the Mel frequency spectrum coefficients are further extracted as features to distinguish the air leakage sound from other sounds.

[0029] In one embodiment, in step S4, a hybrid neural network model with convolutional-recurrent-attention mechanisms is selected as the algorithm for air leakage sound diagnosis. The specific algorithm includes: (1) Model building: Input layer: Inputs preprocessed feature data; Convolutional layer: After inputting feature data, features are extracted through a set of one-dimensional convolutional kernels; Pooling layer: Reduces the dimensionality of the feature sequence output by the convolutional layer, thus reducing computational cost; Bidirectional LSTM layer: Further models the feature sequences extracted by the convolutional layer to capture the long-distance dependencies of the audio signal in the time dimension; Attention mechanism layer: Weights the importance of each time step, focusing on moments highly correlated with leakage; Fully connected classification layer: transforms the attention-weighted feature vectors into a probability distribution; Output layer: Outputs the final prediction result; (2) Loss function: The cross-entropy loss function is used to measure the difference between the class probability distribution output by the model and the true label; Optimization algorithm: The Adam optimizer is used to update model parameters, accelerating the convergence speed; Batch training: The preprocessed audio signal is divided into multiple batches for training, and each batch contains a certain number of samples; (3) Model evaluation: The model performance is evaluated using accuracy, average loss and confidence metrics; Model optimization: Data augmentation is used to diversify training samples and improve the model's generalization ability; the model structure, parameters or training strategies are adjusted based on the evaluation results to improve diagnostic accuracy.

[0030] Specifically, a hybrid neural network model using convolutional-recurrent-attention mechanisms (CNN-LSTM-Attention) is selected as the algorithm for air leak diagnosis. The specific methods include: (1) Model building: Input layer: Input data size is (B, F, T) The three-dimensional feature tensor; in B Indicates batch size, F This represents the dimension of the MFSC features extracted in each frame. F =251, T Indicates the number of time frames. T =40; The first convolutional layer uses 64 two-dimensional convolutional kernels of size 3×3 with a stride of 1 and padding of 1. The feature map size remains unchanged, and the output data size is 64×251×40, which generates 64 feature channels. The size of each channel is consistent with the original time-frequency dimension. The ReLU activation function is used for non-linear mapping. First pooling layer: Use max pooling with a window size of 2×2 and a stride of 2 to downsample the convolution output and obtain a feature map with a size of 64×125×20. The second convolutional layer uses 128 convolutional kernels of size 3×3, with a stride of 1 and padding of 1. The output feature map size is 128×125×20. The layer uses the ReLU activation function for non-linear mapping. The second pooling layer uses max pooling with a window size of 2×2 and a stride of 2 to downsample the convolution output, resulting in a feature map with a size of 128×62×10. Bidirectional LSTM layer (BiLSTM): Its input dimension is 128, the hidden layer has a unidirectional dimension of 128, and the output after bidirectional concatenation is 256 features per time step. The final output tensor size of the LSTM is... B ×62×256; Attention Mechanism Layer: This layer applies an attention weighting mechanism to the LSTM output. The output features at each time step are linearly mapped to obtain attention scores, which are then normalized using the Softmax function to obtain attention weights. The weighted feature vectors are summed over the time dimension to obtain a result of size . B A context vector of 256.

[0031] Fully connected classification layer: The context vector output by the attention module is input into the Linear layer to complete the final classification task.

[0032] (2) Loss function: The cross-entropy loss function is used to measure the difference between the class probability distribution output by the model and the true label; Optimization algorithm: The Adam optimizer is used to update model parameters, accelerating the convergence speed; Batch training: Divide the data into multiple batches for training, with each batch containing a certain number of samples.

[0033] The data representation of the cross-entropy loss function is as follows: (1) in, For the model to the true category Predicted probability value Data representation of the Adam optimizer: (2) (3) (4) (5) (6) in: Indicates the learning rate; , These are the exponential decay coefficients estimated by the first and second moments, respectively; To prevent stable terms with a denominator of zero, it is often set to 10. -8 .

[0034] (3) Model evaluation: The model performance is evaluated using accuracy, average loss and confidence metrics; Model optimization: Data augmentation is used to diversify training samples and improve the model's generalization ability; the model structure, parameters or training strategies are adjusted based on the evaluation results to improve diagnostic accuracy.

[0035] In one embodiment, in step S2, the microphone array is mounted on the drone. The microphone array is used to determine the direction of the sound source. Based on a spatial time delay sound source localization model, a generalized cross-correlation function is used for time delay estimation. Then, a beamforming method is introduced to improve the ability to locate the leak point. Figure 2 As shown, when the convolution-recurrent-attention hybrid neural network model identifies the pre-processed audio, it immediately triggers the drone to hover and intervenes in the TDOA model when it detects a leak sound, so as to accurately locate the leak point.

[0036] In one embodiment, in step S5, the ground monitoring station consists of PC-based real-time monitoring software, which implements a cross-platform interface based on Web technology. The cross-platform interface includes map display, drawing of the drone's current location and flight trajectory, marking of leak points and the time and coordinates of suspected leak points, drone status information and environmental parameters, and a list of alarm logs, recording the time, coordinates, and intensity of each leak event for on-site personnel to archive and analyze.

[0037] A system for implementing the present invention, which is a method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading engineering based on artificial intelligence assistance, includes: The drone platform is equipped with an audio acquisition module and a signal processing unit. It uses a gridded flight path with parallel navigation to cover the entire vacuum pre-compression area. In the whole architecture, the drone performs the main calculations and realizes real-time local decision-making, while the ground end performs visualization and human-machine interaction.

[0038] The audio acquisition module is used to acquire sound signals in real time and transmit the acquired sound signals to the signal processing unit; The signal processing unit is used to preprocess the acquired sound signals, including filtering, denoising, and normalization, to ensure data quality; it extracts MFSC coefficients from the preprocessed sound data and inputs them into a well-trained convolutional-recurrent-attention hybrid neural network model (CNN-LSTM-Attention) to distinguish between air leaks and noise. The communication and ground monitoring center receives real-time information on the drone's location and air leak alarms, and displays it intuitively on the interface.

[0039] In this invention, a hybrid neural network model is used to automatically learn and identify air leak sounds from complex background noise. Compared with the traditional method of listening with the human ear, the hybrid neural network model with convolution-recurrent-attention mechanism (CNN-LSTM-Attention) has powerful feature extraction capabilities and environmental adaptability. It works closely with the UAV flight control to greatly improve the ability to perceive air leak points.

[0040] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading engineering based on artificial intelligence assistance, characterized in that, The method includes the following steps: S1. System initialization and self-test; S2. The UAV uses a gridded flight path to cover the entire vacuum pre-compression site and cruises the entire area using a back-and-forth parallel sweep method. S3. Acquire sound signals in real time through the audio acquisition module and upload the acquired sound signals to the signal processing unit; S4. Preprocess the sound data, extract the MFSC coefficients from the preprocessed sound signal, and input them into a well-trained convolutional-recurrent-attention hybrid neural network model to distinguish between air leakage sound and noise. S5. When the signal processing unit detects a suspected air leak sound, it triggers the drone to hover and circle, and uses array orientation to further determine the location of the air leak before performing membrane repair. At the same time, the ground monitoring station receives the drone's location and air leak alarm information in real time and displays them intuitively on the interface, realizing visualization and human-machine interaction.

2. The method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading engineering based on artificial intelligence as described in claim 1, characterized in that, In step S1, system initialization and self-test specifically include: power supply check of hardware devices, communication line connection test, sensor function verification, and software system loading and verification; after debugging, a free flight test is conducted to test the coverage and false alarm rate of the UAV cruise monitoring.

3. The method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading engineering based on artificial intelligence as described in claim 1, characterized in that, In step S4, the preprocessing of the audio data specifically includes filtering, noise reduction, and normalization to ensure data quality.

4. The method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading engineering based on artificial intelligence as described in claim 1, characterized in that, In step S2, the UAV flies at an altitude of 0.5-1.0 meters and a speed of 1-2 m / s.

5. The method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading engineering based on artificial intelligence as described in claim 1, characterized in that, In step S3, the audio acquisition module is composed of high-precision microphones arranged in an array, which simultaneously receives sound signals and acquires sound signals at a high sampling rate; the high-precision microphones upload the acquired sound data to the signal processing unit, which preprocesses and extracts features from the audio data, and inputs it into the trained convolution-recurrent-attention mechanism hybrid neural network model to determine whether there is a leaking sound and the corresponding confidence level.

6. The method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading engineering based on artificial intelligence as described in claim 3, characterized in that, In step S4, the specific process for preprocessing the acquired raw audio signal is as follows: the data is divided into frames of appropriate length with overlap between frames to maintain time resolution; a pre-designed bandpass filter is applied to each frame of audio to enhance wind noise resistance; normalization is then performed to balance the volume level and prevent model deviation due to distance or gain differences; a short-time Fourier transform (STFT) is then performed on the audio signal to convert each frame of audio signal into a two-dimensional time-frequency graph; and finally, Mel frequency spectrum coefficients are extracted as features to distinguish the air leakage sound from other sounds.

7. The method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading engineering based on artificial intelligence as described in claim 1, characterized in that, In step S4, a hybrid neural network model with convolutional-recurrent-attention mechanism is selected as the algorithm for air leakage sound diagnosis. The specific algorithm includes: (1) Model building: Input layer: Inputs preprocessed feature data; Convolutional layer: After inputting feature data, features are extracted through a set of one-dimensional convolutional kernels; Pooling layer: Reduces the dimensionality of the feature sequence output by the convolutional layer, thus reducing computational cost; Bidirectional LSTM layer: Further models the feature sequences extracted by the convolutional layer to capture the long-distance dependencies of the audio signal in the time dimension; Attention mechanism layer: Weights the importance of each time step, focusing on moments highly correlated with leakage; Fully connected classification layer: transforms the attention-weighted feature vectors into a probability distribution; Output layer: Outputs the final prediction result; (2) Loss function: The cross-entropy loss function is used to measure the difference between the class probability distribution output by the model and the true label; Optimization algorithm: The Adam optimizer is used to update model parameters, accelerating the convergence speed; Batch training: The preprocessed audio signal is divided into multiple batches for training, and each batch contains a certain number of samples; (3) Model evaluation: The model performance is evaluated using accuracy, average loss and confidence metrics; Model optimization: Data augmentation is used to diversify training samples and improve the model's generalization ability; the model structure, parameters or training strategies are adjusted based on the evaluation results to improve diagnostic accuracy.

8. The method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading engineering based on artificial intelligence as described in claim 5, characterized in that, In step S2, the microphone array is installed on the drone. The microphone array is used to determine the direction of the sound source. Based on the spatial time delay sound source localization model, the generalized cross-correlation function is used for time delay estimation. Then, a beamforming method is introduced to improve the ability to locate the air leak point.

9. The method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading engineering based on artificial intelligence as described in claim 5, characterized in that, In step S5, the ground monitoring station consists of PC-based real-time monitoring software, which implements a cross-platform interface based on Web technology. The cross-platform interface includes map display, drawing of the drone's current location and flight trajectory, marking of leak points and the time and coordinates of suspected leak points, drone status information and environmental parameters, and a list of alarm logs, recording the time, coordinates, and intensity of each leak event for on-site personnel to archive and analyze.

10. A system for implementing the method for rapidly identifying and locating vacuum membrane leaks in vacuum preloading engineering based on artificial intelligence as described in any one of claims 1 to 9, characterized in that, include: The drone platform is used to carry an audio acquisition module and a signal processing unit, and adopts a gridded flight path for parallel navigation to cover the entire vacuum pre-compression area; In the entire architecture, the drone performs the main calculations and enables real-time local decision-making, while the ground terminal performs visualization and human-machine interaction. The audio acquisition module is used to acquire sound signals in real time and transmit the acquired sound signals to the signal processing unit; The signal processing unit is used to preprocess the acquired sound signals, including filtering, noise reduction and normalization, to ensure data quality. MFSC coefficients are extracted from the preprocessed sound data and input into a well-trained convolutional-recurrent-attention hybrid neural network model (CNN-LSTM-Attention) to distinguish between air leakage sound and noise. The communication and ground monitoring center receives real-time information on the drone's location and air leak alarms, and displays it intuitively on the interface.

Images