Immersed tunnel leakage voiceprint recognition system based on deep learning

Abstract

The invention discloses an immersed tunnel leakage voiceprint recognition system based on deep learning. The system comprises an acoustic signal acquisition module, a signal preprocessing module, a voiceprint feature extraction module, a deep learning recognition module and a leakage positioning and early warning module. According to the invention, the leakage water flow sound and the structural damage elastic wave signal of the immersed tunnel are comprehensively considered, the multi-dimensional voiceprint feature extraction and deep learning model are fused, and the adaptive signal noise reduction and sound source positioning technology is combined, so that the high-precision monitoring and real-time early warning of the leakage of the immersed tunnel are realized; the accuracy and real-time performance of immersed tunnel leakage nondestructive monitoring are improved, and powerful support is provided for follow-up risk prediction and early repair.

Description

Technical Field

[0001] The present invention relates to the technical field of immersed tube tunnel engineering safety monitoring, and in particular to an immersed tube tunnel leakage soundprint recognition system based on deep learning. Background Art

[0002] As critical infrastructure for cross-water transportation, immersed tube tunnels are constantly exposed to the multiple forces of water pressure, erosion, and geological changes. Leakage has become a core risk threatening their structural safety and service life. Traditional leak detection methods rely primarily on manual inspections and single-sensor testing. Manual inspections identify leaks through visual inspection or contact detection equipment (such as hygrometers and infrared thermal imagers). However, due to the airtightness of the tunnel and the inspection cycle, these inspections suffer from significant drawbacks such as low efficiency, poor real-time performance, and a high rate of missed detections. While single-sensor monitoring technologies, such as fiber grating sensors, can achieve continuous monitoring, they are often limited to collecting data at localized points and are susceptible to environmental interference, leading to frequent false alarms.

[0003] In recent years, deep learning technology has demonstrated its potential in acoustic signal processing. For example, deep learning models used in speech recognition (such as CNN and LSTM) are now capable of efficiently extracting complex voiceprint features. The industry is increasingly demanding high-precision, real-time, and intelligent leakage monitoring, especially in scenarios such as large cross-sea tunnels and urban underground utility corridors. There is an urgent need to overcome existing technological bottlenecks and build intelligent monitoring systems that adapt to complex environments and possess self-learning and adaptive capabilities. The development of a soundprint recognition system for immersed tunnel leakage aims to address the industry pain points of traditional leakage monitoring technologies, which suffer from low efficiency, high cost, and poor real-time performance. By integrating high-sensitivity acoustic sensors with deep learning algorithms, it can capture the soundprint characteristics of underwater pipe joints in real time, enabling leak type identification and location, and providing timely warnings. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a deep learning-based immersed tube tunnel leakage soundprint recognition system. This system integrates multi-dimensional soundprint feature extraction and deep learning models, combines adaptive signal noise reduction and sound source localization technology, and realizes high-precision monitoring and real-time early warning of immersed tube tunnel leakage damage.

[0005] The present invention is achieved through the following technical solutions: A deep learning-based immersed tunnel leakage soundprint recognition system, which includes an acoustic signal acquisition module, a signal preprocessing module, a soundprint feature extraction module, a deep learning recognition module, and a leakage location and early warning module; Acoustic signal acquisition module, used to collect acoustic signal data of water leakage and structural vibration in key parts of immersed tube tunnels; The signal preprocessing module is used to perform adaptive noise reduction and frequency domain filtering on the collected acoustic signal data, and then segment the signal to construct a valid voiceprint sample with a timestamp; The voiceprint feature extraction module is used to perform time-frequency analysis, nonlinear analysis and deep learning feature extraction on the obtained voiceprint samples, and then perform feature fusion and dimensionality reduction to generate a multi-dimensional voiceprint feature vector; The deep learning recognition module is used to generate classification results for the voiceprint features extracted by the voiceprint feature extraction module, including an assessment of the leakage damage category and the degree of leakage damage; The leakage location and early warning module is used to realize three-dimensional location of the leakage point and output corresponding early warning information according to the identified leakage damage degree and the preset graded early warning threshold.

[0006] In the above technical solution, for collecting leakage water sound, a broadband hydrophone array is used; for collecting structural vibration sound, a combination of a three-axis accelerometer and a piezoelectric acoustic emission sensor is used.

[0007] In the above technical solution, the signal preprocessing module adopts a multi-stage signal processing architecture, including: Adaptive noise reduction processing: The LMS algorithm is used to construct a noise reference channel to coherently cancel out steady-state interference from water flow noise and traffic vibration. Frequency domain filtering: Design a Butterworth bandpass filter to eliminate low-frequency mechanical vibration and high-frequency electromagnetic interference; Intelligent signal segmentation: Based on the short-time energy and zero-crossing rate double threshold detection method, valid voiceprint fragments are extracted and converted into a standardized format to construct a valid voiceprint sample with a timestamp.

[0008] In the above technical solution, the voiceprint feature extraction module adopts a hybrid feature extraction system, including: Time-frequency analysis layer: Morlet wavelet transform is used to extract 6-dimensional wavelet energy ratio features, 128-dimensional MFCC coefficient features are generated through Mel filter bank, and the zero-crossing rate and short-time energy envelope are calculated to generate 2-dimensional dynamic envelope features; Nonlinear analysis layer: Apply recursive quantitative analysis (RQA) to obtain the three-dimensional RQA features of the certainty coefficient, laminar flow and entropy value of the voiceprint sample; Deep learning layer: Build a deep learning model to extract 256-dimensional deep voiceprint features of voiceprint samples.

[0009] In the above technical solution, the deep learning recognition module adopts a multimodal neural network architecture, including: Spatial feature extraction unit: Configured with a 4-layer dilated convolutional network, it expands the receptive field through dilated convolution and effectively captures the local correlation of voiceprint features; Time series modeling unit: Build a bidirectional LSTM network and use the attention mechanism to enhance the transient feature expression of voiceprint features.

[0010] In the above technical solution, the leakage location and early warning module includes: Sound source positioning layer: used to achieve three-dimensional spatial positioning of leakage points; Hierarchical early warning mechanism: Set three-level response thresholds to simultaneously trigger sound and light alarms, SCADA system linkage, and maintenance personnel APP push.

[0011] In the above technical solution, the sound source localization layer adopts an improved TDOA algorithm: based on the generalized cross-correlation-phase transformation optimization delay estimation, the delay difference By sensor Determine the peak value of the received signal cross-correlation function; locate the peak value by non-uniform grid search, combined with the sensor array coordinates and , construct an overdetermined system of equations: ; in is the speed of sound in water, and the least squares method is used to solve the three-dimensional coordinates of the leakage point .

[0012] The advantages and beneficial effects of the present invention are: This invention integrates time-frequency analysis, nonlinear feature extraction, and deep learning technologies to construct a multidimensional voiceprint feature engineering system. Combined with adaptive noise reduction and frequency-domain filtering, it significantly improves the comprehensiveness and accuracy of leak monitoring. An improved TDOA algorithm and beamforming technology are used to achieve three-dimensional positioning of leak points. Risk levels are dynamically divided using three-level warning thresholds, which simultaneously trigger audible and visual alarms, SCADA system linkage, and app push notifications for maintenance personnel, forming a closed-loop "monitoring-warning-disposal" management system. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is the architecture diagram of the immersed tube tunnel leakage soundprint recognition system based on deep learning. DETAILED DESCRIPTION

[0014] In order to enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention are further described below with reference to specific embodiments.

[0015] The present invention designs a deep learning-based immersed tube tunnel leakage soundprint recognition system, which includes an acoustic signal acquisition module, a signal preprocessing module, a soundprint feature extraction module, a deep learning recognition module, and a leakage positioning and early warning module.

[0016] The acoustic signal acquisition module is used to collect acoustic signal data of leakage water flow sound and structural vibration sound at key parts of the immersed tube tunnel; it includes: Leakage water sound collection: using a broadband hydrophone array; Structural vibration and acoustic acquisition: Using a combination of a triaxial accelerometer and a piezoelectric acoustic emission sensor, it is possible to capture elastic wave signals (i.e., structural vibration and acoustic wave signals) generated by structural damage such as concrete cracking.

[0017] The signal preprocessing module is used to perform adaptive noise reduction and frequency domain filtering on the collected raw acoustic signal data. The processed signal data retains the useful information of the acoustic signal. Then, the acoustic signal data after noise reduction and filtering is divided into short-time segments, and effective voiceprint segments are extracted through a double threshold detection method of short-time energy and zero-crossing rate. These segments are stored as voiceprint samples, laying the foundation for subsequent voiceprint feature extraction and modeling. The signal preprocessing module adopts a multi-stage signal processing architecture, including: Adaptive noise reduction processing: The LMS algorithm is used to construct a noise reference channel to coherently cancel steady-state interference such as water flow noise and traffic vibration; Frequency domain filtering: Design a Butterworth bandpass filter to eliminate low-frequency mechanical vibration and high-frequency electromagnetic interference; Intelligent signal segmentation: Based on the short-time energy and zero-crossing rate double threshold detection method, valid voiceprint fragments are extracted and converted into a standardized format to construct a valid voiceprint sample with a timestamp.

[0018] The voiceprint feature extraction module is used to perform time-frequency analysis, nonlinear analysis, and deep learning feature extraction on the obtained voiceprint samples, and then perform feature fusion and dimensionality reduction to generate multi-dimensional voiceprint feature vectors. These voiceprint feature vectors are standardized and used as input data for training and inference, and input to the deep learning recognition module. The voiceprint feature extraction module adopts a hybrid feature extraction system, including: Time-frequency analysis layer: Morlet wavelet transform is used to extract 6-dimensional wavelet energy ratio features, 128-dimensional MFCC coefficient features are generated through Mel filter bank, and the zero-crossing rate and short-time energy envelope are calculated to generate 2-dimensional dynamic envelope features; Nonlinear analysis layer: Recursive quantitative analysis (RQA) is applied to obtain the three-dimensional RQA features of the certainty coefficient, laminar flow and entropy value of the voiceprint sample; Deep learning layer: Build a 5-layer 1D-CNN deep learning model to extract 256-dimensional deep voiceprint features of voiceprint samples; Feature fusion and dimensionality reduction: Time-frequency features (6-dimensional wavelet energy ratio features + 128-dimensional MFCC features + 2-dimensional dynamic envelope features), nonlinear features (3-dimensional RQA features), and deep learning features (256-dimensional deep voiceprint features) are input into the fully connected layer for fusion, and then the dimension is reduced to a 128-dimensional mixed feature vector through principal component analysis (PCA).

[0019] The deep learning recognition module is used to generate classification results for the voiceprint features extracted by the voiceprint feature extraction module, including an assessment of the leakage damage category and the degree of leakage damage. These assessment results are passed to the leakage location and warning module. The deep learning recognition module adopts a multimodal neural network architecture and includes: Spatial feature extraction unit: Configured with a 4-layer dilated convolutional network, it expands the receptive field through dilated convolution and effectively captures the local correlation of voiceprint features; Time series modeling unit: Build a bidirectional LSTM network and use the attention mechanism to enhance the transient feature expression of voiceprint features.

[0020] The leakage location and warning module is used to achieve three-dimensional positioning of the leakage point and output corresponding warning information based on the identified leakage damage level and the preset graded warning threshold. The leakage location and warning module includes: Sound source localization layer: uses an improved TDOA algorithm combined with beamforming technology to achieve three-dimensional spatial positioning; Gradual warning mechanism: Set three-level response thresholds (leakage rate <1L / s is yellow warning, leakage rate 1-5L / s is orange warning, leakage rate >5L / s is red warning), and simultaneously trigger sound and light alarms, SCADA system linkage, and maintenance personnel APP push.

[0021] Furthermore, the acoustic signal acquisition module is used to build a multi-dimensional acoustic monitoring network for immersed tube tunnels, deploying a high-sensitivity acoustic sensor array at pipe joints, settlement joints, and weak areas prone to leakage. Specifically, it includes: Leakage water flow sound acquisition: using a wide-band hydrophone array, and through titanium alloy waterproof packaging and pressure-resistant structure design. The original sound pressure signal output by the hydrophone is p(t), where t represents time, combined with the sound wave propagation equation (ρ is water density, c is the speed of sound, and v(t) is the particle vibration velocity), and the velocity and flow characteristics of the leakage water flow are inverted.

[0022] Structural vibration sound acquisition: A triaxial accelerometer and a piezoelectric acoustic emission sensor are combined to capture elastic wave signals generated by structural damage such as concrete cracking and steel corrosion through time domain synchronous triggering technology. The propagation speed of elastic waves in concrete is , from the elastic modulus of the material E and density Decision, that is , combined with the waveform arrival time difference , the damage source can be located.

[0023] The layout of the sensor array is based on the 3D model of the tunnel segment, using a non-uniform grid deployment strategy. d Satisfies the spatial sampling theorem ( is the minimum monitoring wavelength), ensuring the full-band acoustic signal acquisition without aliasing, and achieving signal phase consistency control through multi-channel synchronous transmission technology.

[0024] Furthermore, the signal preprocessing module adopts a multi-stage signal processing architecture, specifically including: Adaptive noise reduction processing: The noise reference channel is constructed through the least mean square (LMS) algorithm to coherently cancel steady-state interference such as water flow noise and traffic vibration. Noise reference signal With the main signal Enter the adaptive filter, and the weight update formula is: ; in, is the filter weight vector, is the convergence factor, is the error signal. Through iterative optimization, the noise reduction signal is output .

[0025] Frequency domain filtering: Design an 8th-order Butterworth bandpass filter with the following transfer function: ; The passband range is 100Hz-15kHz, the stopband attenuation is ≥40dB / dec, and the passband ripple is ≤1dB. A bilinear transformation converts the analog filter into a digital filter, eliminating low-frequency mechanical vibration (<100Hz) and high-frequency electromagnetic interference (>15kHz), while retaining the characteristic frequency band of the leakage soundprint.

[0026] Intelligent signal segmentation: Extract effective voiceprint segments based on short-time energy and zero-crossing rate double threshold detection method. and zero-crossing rate The calculation formula is: ; Among them, the frame length N =512, dynamically set energy threshold , and the zero-crossing rate threshold ;when and The valid voiceprint segment is determined when the segment lasts 100ms-2s, and the segment is converted into 16bit / 48kHz standardized format to build a voiceprint sample library with timestamp.

[0027] Furthermore, the voiceprint feature extraction module builds a hybrid feature extraction system, which specifically includes: (1) Time-frequency analysis layer: Continuous wavelet transform (CWT) is used to extract multi-scale voiceprint features. The wavelet basis function is: ; in As the center frequency, set 6 scale sub-bands (corresponding to the frequency band 50Hz-10kHz) and calculate the energy ratio of each sub-band , characterize the energy distribution of different frequency bands, and thus obtain the 6-dimensional wavelet energy ratio feature.

[0028] Mel-frequency cepstral coefficient (MFCC): The voiceprint spectrum is nonlinearly compressed using a 40-channel Mel filter bank (frequency range 100Hz-15kHz) to extract 128-dimensional MFCC coefficient features. The calculation formula is: ; in is the Mel filter bank energy output, .

[0029] Zero-crossing rate and short-time energy envelope: Calculate the zero-crossing rate of a 20ms frame (960 points @ 48kHz) and short-term energy , generating 2D dynamic envelope features.

[0030] (2) Nonlinear analysis layer: Recursive Quantitative Analysis (RQA): Based on the phase space reconstruction of the voiceprint signal (embedding dimension ,Delay ), calculate the 3D RQA features, including: Coefficient of certainty: ,in is the diagonal length distribution, ; Laminar flow: , characterizes the signal stationarity; Entropy: , reflecting the signal complexity.

[0031] (3) Deep learning layer: 1D-CNN network architecture: Build a 5-layer convolutional network with the following parameters: Layer 1: Convolution kernel size , number of channels , step length , activation function ReLU; Layers 2-4: Reduce kernel size layer by layer , the number of channels doubled , step length ; Layer 5: Global maximum pooling, outputting a 256-dimensional deep voiceprint feature vector.

[0032] Finally, feature fusion and dimensionality reduction are performed: the time-frequency features (6-dimensional wavelet energy ratio features + 128-dimensional MFCC features + 2-dimensional dynamic envelope features), nonlinear features (3-dimensional RQA features) and deep learning features (256-dimensional deep voiceprint features) are input into the fully connected layer for fusion, and then the dimensionality is reduced to a 128-dimensional mixed feature vector through principal component analysis (PCA).

[0033] Furthermore, the deep learning recognition module adopts a multimodal neural network architecture, specifically including: (1) Spatial feature extraction unit Configure a 4-layer hole convolutional network and use the expansion rate The receptive field is expanded layer by layer to effectively capture the local correlation of the voiceprint spectrum. The dilated convolution operation of the first layer is defined as: ; in, is the input feature map, for Convolution kernel , is the bias term. Through dilated convolution, the receptive field of the fourth layer is expanded to [ 1 ], significantly improving the ability to capture local details of the voiceprint spectrum (such as short-term harmonics and transient pulses). Each layer is followed by batch normalization and a ReLU activation function, with the number of output channels increasing to 32, 64, 128, and 256, respectively.

[0034] (2) Timing Modeling Unit A bidirectional long short-term memory (Bi-LSTM) network is built, and the attention mechanism is used to enhance the transient feature expression of the leakage signal. The network structure is as follows: Bi-LSTM layer: 128 hidden units per layer, with input being the temporal feature sequence output by the spatial feature extraction unit , the forward and backward LSTM state update formulas are: ; in is the forward hidden state, is the backward hidden state, the final temporal feature .

[0035] Attention Mechanism: Through Learnable Weights Calculating attention scores : ; The weighted time series features are , focusing on transient leakage signals (such as sudden water flow sound and structural crack sound).

[0036] Furthermore, the leakage location and early warning module specifically includes: (1) Sound source localization layer: Improved TDOA algorithm: Optimized delay estimation based on generalized cross-correlation-phase transform (GCC-PHAT), delay difference By sensor The peak value of the cross-correlation function of the received signal is determined by: ; in is the frequency domain signal of sensor i, is the delay parameter. The peak is located by non-uniform grid search, combined with the sensor array coordinates , construct an overdetermined system of equations: ; in is the speed of sound in water, and the least squares method is used to solve the three-dimensional coordinates of the leakage point .

[0037] Beamforming-enhanced positioning: Based on delay-sum beamforming (DSBF), multi-channel signals are spatially filtered to suppress multipath interference and improve azimuth resolution.

[0038] (2) Graded early warning mechanism: Three-level threshold setting: Yellow warning (leakage ): Trigger local sound and light alarm, SCADA system records data; Orange warning ( ): Start the emergency ventilation system, link the BIM model to highlight the leakage point, and push the work order to the maintenance personnel APP; red warning ( ): Broadcast evacuation instructions throughout the tunnel, close adjacent pipe sections, start the drainage pump group, and the response delay is <3 seconds.

[0039] The above is an exemplary description of the present invention. It should be noted that, without departing from the core of the present invention, any simple deformation, modification or other equivalent replacement that can be made by other skilled in the art without expending creative labor falls within the scope of protection of the present invention.

Claims

1. A deep learning-based immersed tunnel leakage soundprint recognition system, characterized by: The system includes an acoustic signal acquisition module, a signal preprocessing module, a voiceprint feature extraction module, a deep learning recognition module, and a leakage location and early warning module; Acoustic signal acquisition module, used to collect acoustic signal data of water leakage and structural vibration in key parts of immersed tube tunnels; The signal preprocessing module is used to perform adaptive noise reduction and frequency domain filtering on the collected acoustic signal data, and then segment the signal to construct a valid voiceprint sample with a timestamp; The voiceprint feature extraction module is used to perform time-frequency analysis, nonlinear analysis and deep learning feature extraction on the obtained voiceprint samples, and then perform feature fusion and dimensionality reduction to generate a multi-dimensional voiceprint feature vector; The deep learning recognition module is used to generate classification results for the voiceprint features extracted by the voiceprint feature extraction module, including an assessment of the leakage damage category and the degree of leakage damage; The leakage location and early warning module is used to realize three-dimensional location of the leakage point and output corresponding early warning information according to the identified leakage damage degree and the preset graded early warning threshold.

2. The deep learning-based immersed tunnel leakage soundprint recognition system according to claim 1 is characterized by: For the collection of leakage water sound: a broadband hydrophone array is used; for the collection of structural vibration sound: a combination of a triaxial accelerometer and a piezoelectric acoustic emission sensor is used.

3. The deep learning-based immersed tunnel leakage soundprint recognition system according to claim 1 is characterized by: The signal preprocessing module adopts a multi-level signal processing architecture, including: Adaptive noise reduction processing: The LMS algorithm is used to construct a noise reference channel to coherently cancel out steady-state interference from water flow noise and traffic vibration. Frequency domain filtering: Design a Butterworth bandpass filter to eliminate low-frequency mechanical vibration and high-frequency electromagnetic interference; Intelligent signal segmentation: Based on the short-time energy and zero-crossing rate double threshold detection method, valid voiceprint fragments are extracted and converted into a standardized format to construct a valid voiceprint sample with a timestamp.

4. The deep learning-based immersed tunnel leakage soundprint recognition system according to claim 1 is characterized by: The voiceprint feature extraction module adopts a hybrid feature extraction system, including: Time-frequency analysis layer: Morlet wavelet transform is used to extract 6-dimensional wavelet energy ratio features, 128-dimensional MFCC coefficient features are generated through Mel filter bank, and the zero-crossing rate and short-time energy envelope are calculated to generate 2-dimensional dynamic envelope features; Nonlinear analysis layer: Recursive quantitative analysis (RQA) is applied to obtain the three-dimensional RQA features of the certainty coefficient, laminar flow and entropy value of the voiceprint sample; Deep learning layer: Build a deep learning model to extract 256-dimensional deep voiceprint features of voiceprint samples.

5. The deep learning-based immersed tunnel leakage soundprint recognition system according to claim 1 is characterized by: The deep learning recognition module uses a multimodal neural network architecture, including: Spatial feature extraction unit: Configured with a 4-layer dilated convolutional network, it expands the receptive field through dilated convolution and captures the local correlation of voiceprint features; Time series modeling unit: Build a bidirectional LSTM network and use the attention mechanism to enhance the transient feature expression of voiceprint features.

6. The deep learning-based immersed tunnel leakage soundprint recognition system according to claim 1 is characterized by: Leakage location and early warning module, including: Sound source positioning layer: used to achieve three-dimensional spatial positioning of leakage points; Hierarchical early warning mechanism: Set three-level response thresholds to simultaneously trigger sound and light alarms, SCADA system linkage, and maintenance personnel APP push.

7. The deep learning-based immersed tunnel leakage soundprint recognition system according to claim 6 is characterized by: The sound source localization layer uses an improved TDOA algorithm: based on generalized cross-correlation-phase transformation to optimize delay estimation, delay difference By sensor Determine the peak value of the received signal cross-correlation function; locate the peak value by non-uniform grid search, combined with the sensor array coordinates and , construct an overdetermined system of equations: ； in is the speed of sound in water, and the least squares method is used to solve the three-dimensional coordinates of the leakage point .

Images