A deep learning-based immersed tube tunnel leakage acoustic print recognition system

Abstract

The application discloses a kind of based on deep learning's immersed tube tunnel leakage acoustic print recognition system, the system includes acoustic signal acquisition module, signal pre-processing module, acoustic print feature extraction module, deep learning identification module and leakage positioning and early warning module;The present application considers the sound of leakage water flow and structure damage elastic wave signal of immersed tube tunnel by comprehensively, fuses multidimensional acoustic print feature extraction and deep learning model, combines adaptive signal noise reduction and acoustic source positioning technology, realizes high-precision monitoring and real-time early warning to immersed tube tunnel leakage, not only improves the accuracy and real-time of immersed tube tunnel leakage nondestructive testing, also provides strong support for subsequent risk prediction and advance repair.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of safety monitoring of immersed tunnel engineering, and particularly relates to an immersed tunnel leakage voiceprint recognition system based on deep learning. BACKGROUND

[0002] As a key infrastructure for water-crossing traffic, immersed tunnels have long been subjected to the combined effects of water pressure, water flow scouring and geological changes, and leakage has become a core risk threatening the structural safety and service life of immersed tunnels. Traditional leakage detection methods mainly rely on manual inspection and single sensor detection: manual inspection identifies leakage through visual inspection or contact-type detection equipment (such as a hygrometer, an infrared thermal imager), but is limited by the airtightness of the tunnel and the inspection cycle, and has significant defects such as low efficiency, poor real-time performance, and high missed detection rate; while single-sensor monitoring technologies such as fiber grating sensors can achieve continuous monitoring, but are mostly limited to local data collection and are prone to false positives due to environmental interference.

[0003] In recent years, deep learning technology has shown potential in the field of acoustic signal processing, for example, deep learning models (such as CNN, LSTM) in speech recognition can efficiently extract complex voiceprint features. The industry's demand for high-precision, real-time, and intelligent leakage monitoring is increasingly urgent, especially in large cross-sea tunnels and urban underground comprehensive pipe galleries, and it is urgent to break through the existing technical bottlenecks and build an intelligent monitoring system that can adapt to complex environments and has self-learning and adaptive capabilities. The development of an immersed tunnel leakage voiceprint recognition system aims to address the industry's pain points of low efficiency, high cost, and poor real-time performance of traditional leakage monitoring technology. By integrating high-sensitivity acoustic sensing and deep learning algorithms, the system can capture real-time leakage voiceprint features of underwater pipe joint seams, enabling leakage type identification, positioning, and timely warning. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art and provide an immersed tunnel leakage voiceprint recognition system based on deep learning. The system integrates multi-dimensional voiceprint feature extraction and deep learning models, combines adaptive signal noise reduction and sound source positioning technology, and realizes high-precision monitoring and real-time warning of immersed tunnel leakage damage.

[0005] The present application is implemented by the following technical solutions:

[0006] An immersed tunnel leakage voiceprint recognition system based on deep learning, the system comprising an acoustic signal acquisition module, a signal preprocessing module, a voiceprint feature extraction module, a deep learning recognition module, and a leakage positioning and warning module;

[0007] The acoustic signal acquisition module is used to acquire acoustic signal data of leakage water flow sound and structural vibration sound at key positions of the immersed tunnel;

[0008] a signal preprocessing module, configured to perform adaptive noise reduction and frequency domain filtering on the collected acoustic signal data, segment the signal, and construct effective voiceprint samples with timestamps;

[0009] The voiceprint feature extraction module is configured to perform time-frequency analysis, nonlinear analysis, and deep learning feature extraction on the obtained voiceprint samples, and then perform feature fusion and dimension reduction to generate multi-dimensional voiceprint feature vectors.

[0010] The deep learning recognition module is configured to generate a classification result of the voiceprint features extracted by the voiceprint feature extraction module, including an evaluation of the leakage damage category and the leakage damage degree.

[0011] The leakage positioning and warning module is configured to realize three-dimensional positioning of the leakage point, and output corresponding warning information according to the identified leakage damage degree and the preset grading warning threshold.

[0012] In the above technical solution, for the collection of the leakage water flow sound, a wideband hydrophone array is used; for the collection of the structural vibration sound, a three-axis accelerometer and a piezoelectric acoustic emission sensor are combined.

[0013] In the above technical solution, the signal preprocessing module adopts a multi-stage signal processing architecture, including:

[0014] Adaptive noise reduction processing: a noise reference channel is constructed through an LMS algorithm to perform coherent cancellation on water flow noise and traffic vibration steady-state interference;

[0015] Frequency domain filtering processing: a Butterworth band-pass filter is designed to eliminate low-frequency mechanical vibration and high-frequency electromagnetic interference;

[0016] Intelligent signal segmentation: based on the double-threshold detection method of short-time energy and zero-crossing rate, effective voiceprint segments are extracted, and standardized format conversion is performed to construct effective voiceprint samples with timestamps.

[0017] In the above technical solution, the voiceprint feature extraction module adopts a hybrid feature extraction system, including:

[0018] Time-frequency analysis layer: Morlet wavelet transform is used to extract 6-dimensional wavelet energy ratio features, a 128-dimensional MFCC coefficient feature is generated through a Mel filter bank, and 2-dimensional dynamic envelope features are calculated based on zero-crossing rate and short-time energy envelope;

[0019] Nonlinear analysis layer: recursive quantitative analysis RQA is applied to obtain 3-dimensional RQA features of the voiceprint samples, including the determinacy coefficient, laminarity, and entropy value;

[0020] Deep learning layer: a deep learning model is built to extract 256-dimensional deep voiceprint features of the voiceprint samples.

[0021] In the above technical solution, the deep learning recognition module adopts a multi-modal neural network architecture, comprising:

[0022] The spatial feature extraction unit is configured with a 4-layer cavity convolution network, which expands the receptive field through dilated convolution and effectively captures the local correlation of the voiceprint feature.

[0023] The time series modeling unit is built with a bidirectional LSTM network, and the attention mechanism is used to strengthen the transient feature expression of the voiceprint feature.

[0024] In the above technical solution, the leakage positioning and early warning module comprises:

[0025] The sound source positioning layer is used to realize three-dimensional spatial positioning of the leakage point.

[0026] The hierarchical early warning mechanism sets three response thresholds to trigger sound and light alarms, SCADA system linkage and maintenance personnel APP push simultaneously.

[0027] In the above technical solution, the sound source positioning layer adopts an improved TDOA algorithm: based on generalized cross-correlation-phase transform optimized time delay estimation, time delay difference The sensor pair The peak value of the cross-correlation function of the received signal is determined; the peak value is located by non-uniform grid search, and the sensor array coordinates And , an over-determined equation set is constructed:

[0028] ;

[0029] Where is the sound speed in water, and the least square method is used to solve the three-dimensional coordinates of the leakage point.

[0030] The advantages and beneficial effects of the present application are:

[0031] The present application fuses time-frequency analysis, nonlinear feature extraction and deep learning technology to construct a multi-dimensional voiceprint feature engineering system, combines adaptive noise reduction and frequency domain filtering processing, and significantly improves the comprehensiveness and accuracy of leakage monitoring. The improved TDOA algorithm and beamforming technology are used to realize three-dimensional positioning of the leakage point. The risk level is dynamically divided by three early warning thresholds, and sound and light alarms, SCADA system linkage and maintenance personnel APP push are triggered simultaneously to form a "monitoring-early warning-disposal" closed-loop management. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The figure is the architecture diagram of the deep learning-based immersed tunnel leakage voiceprint recognition system. DETAILED DESCRIPTION

[0033] In order for the person skilled in the art to better understand the technical scheme of the present application, the technical scheme of the present application will be further described below in combination with specific embodiments.

[0034] The present application designs a deep learning-based immersed tunnel leakage acoustic print recognition system, which comprises an acoustic signal acquisition module, a signal preprocessing module, an acoustic print feature extraction module, a deep learning recognition module, and a leakage positioning and early warning module.

[0035] The acoustic signal acquisition module is used to acquire acoustic signal data of leakage water flow sound and structural vibration sound at key positions of the immersed tunnel; and comprises:

[0036] Leakage water flow sound acquisition: a wideband hydrophone array is used;

[0037] Structural vibration sound acquisition: a combination of a three-axis accelerometer and a piezoelectric acoustic emission sensor is used, which can capture elastic wave signals (i.e. structural vibration sound wave signals) generated by structural damage such as concrete cracking.

[0038] The signal preprocessing module is used to perform adaptive noise reduction and frequency domain filtering processing on the acquired original acoustic signal data, and the processed signal data retains the useful information of the acoustic signal; then, the acoustic signal data after noise reduction and filtering processing is segmented into short-time segments, and valid acoustic print segments are extracted through a short-time energy and zero-crossing rate double-threshold detection method, which are stored as acoustic print samples, laying a foundation for subsequent acoustic print feature extraction and modeling. The signal preprocessing module adopts a multi-level signal processing architecture, which comprises:

[0039] Adaptive noise reduction processing: a noise reference channel is constructed through an LMS algorithm to coherently cancel water flow noise, traffic vibration and other steady-state interference;

[0040] Frequency domain filtering processing: a Butterworth band-pass filter is designed to eliminate low-frequency mechanical vibration and high-frequency electromagnetic interference;

[0041] Intelligent signal segmentation: based on the short-time energy and zero-crossing rate double-threshold detection method, valid acoustic print segments are extracted, and standardized format conversion is performed to construct valid acoustic print samples with timestamps.

[0042] The acoustic print feature extraction module is used to perform time-frequency analysis, nonlinear analysis and deep learning feature extraction on the obtained acoustic print samples, and then perform feature fusion and dimension reduction to generate multi-dimensional acoustic print feature vectors. After standardization, these acoustic print feature vectors are used as input data for training and reasoning, and are input to the deep learning recognition module. The acoustic print feature extraction module adopts a hybrid feature extraction system, which comprises:

[0043] Time-frequency analysis layer: 6-dimensional wavelet energy ratio features are extracted using Morlet wavelet transform, 128-dimensional MFCC coefficient features are generated through a Mel filter bank, and 2-dimensional dynamic envelope features are calculated by zero-crossing rate and short-time energy envelope;

[0044] Nonlinear analysis layer: 3-dimensional RQA features of determinism, laminarity, and entropy are obtained by applying recurrence quantitative analysis (RQA) to the voiceprint samples;

[0045] Deep learning layer: a 5-layer 1D-CNN deep learning model is built to extract 256-dimensional deep voiceprint features from the voiceprint samples;

[0046] Feature fusion and dimension 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 a fully connected layer for fusion, and then reduced to 128-dimensional mixed feature vectors through principal component analysis (PCA).

[0047] The deep learning recognition module is used to generate classification results of the voiceprint features extracted by the voiceprint feature extraction module, including evaluation of the leakage damage category and the leakage damage degree. These evaluation results are passed to the leakage positioning and warning module. The deep learning recognition module adopts a multi-modal neural network architecture, which includes:

[0048] Spatial feature extraction unit: a 4-layer dilated convolution network is configured to expand the receptive field through dilated convolution, effectively capturing the local correlation of voiceprint features;

[0049] Time series modeling unit: a bidirectional LSTM network is built to strengthen the transient feature expression of voiceprint features using an attention mechanism.

[0050] The leakage positioning and warning module is used to realize three-dimensional positioning of the leakage point, and according to the identified leakage damage degree and the preset graded warning threshold, the corresponding warning information is output. The leakage positioning and warning module includes:

[0051] Sound source positioning layer: an improved TDOA algorithm is used in combination with beamforming technology to realize three-dimensional spatial positioning;

[0052] Graded warning mechanism: three response thresholds are set (leakage volume <1 L / s for yellow warning, leakage volume 1-5 L / s for orange warning, and leakage volume >5 L / s for red warning), which synchronously trigger audible and visual alarms, SCADA system linkage, and maintenance personnel APP push.

[0053] Further, the acoustic signal acquisition module is used for constructing a multi-dimensional acoustic monitoring network of the immersed tunnel, deploying a high-sensitivity acoustic sensor array at a pipe joint, a settlement joint and a weak area prone to leakage, and specifically comprising: leakage flow acoustic acquisition: a wideband hydrophone array is adopted, and a titanium alloy waterproof packaging and pressure-resistant structure design are adopted. The original acoustic pressure signal output by the hydrophone is p(t), wherein t represents time, the speed and flow characteristics of the leakage flow are inversed by combining the acoustic wave propagation equation (ρ is the water density, c is the acoustic velocity, and v(t) is the particle vibration velocity).

[0054] Structural vibration acoustic acquisition: a three-axis accelerometer and a piezoelectric acoustic emission sensor are combined, and an elastic wave signal generated by structural damage such as concrete cracking and steel corrosion is captured through time domain synchronous triggering technology. The propagation speed of the elastic wave in the concrete is , which is determined by the material elastic modulus E and the density , that is, , combined with the waveform arrival time difference , the damage source position can be located.

[0055] The layout of the sensor array is based on a three-dimensional model of the tunnel pipe joint, and a non-uniform gridding deployment strategy is adopted, the sensor spacing d satisfies the spatial sampling theorem ( is the minimum monitoring wavelength), ensuring that the full-band acoustic signal is collected without aliasing, and the signal phase consistency control is realized through a multi-channel synchronous transmission technology.

[0056] Further, the signal preprocessing module adopts a multi-level signal processing architecture, specifically comprising:

[0057] Adaptive noise reduction processing: a noise reference channel is constructed through a least mean square (LMS) algorithm to coherently cancel water flow noise, traffic vibration and other steady-state interference. The noise reference signal and the main signal are input into an adaptive filter, and the weight update formula is:

[0058] ;

[0059] wherein, is the filter weight vector, is a convergence factor, is an error signal. Through iterative optimization, the noise reduction signal is output.

[0060] Frequency domain filtering processing: an 8th order Butterworth bandpass filter is designed, and the transfer function thereof is:

[0061] ;

[0062] Wherein, the passband range is 100Hz-15kHz, the stopband attenuation is ≥40dB / dec, and the passband ripple is ≤1dB. Through the bilinear transformation, the analog filter is converted into a digital filter, eliminating low-frequency mechanical vibration (<100Hz) and high-frequency electromagnetic interference (>15kHz), and retaining the leakage voiceprint feature frequency band.

[0063] Signal intelligent segmentation: based on the short-time energy and zero-crossing rate double-threshold detection method, the effective voiceprint segment is extracted. The short-time energy and the zero-crossing rate The calculation formula is:

[0064] ;

[0065] Wherein, the frame length N =512, the dynamic energy threshold , and the zero-crossing rate threshold ; when and , it is determined as an effective voiceprint segment, the segment with a duration of 100ms-2s is extracted, and the 16bit / 48kHz standardization format conversion is carried out, thereby constructing a voiceprint sample library with timestamp.

[0066] Further, the voiceprint feature extraction module constructs a hybrid feature extraction system, which specifically includes:

[0067] (1) Time-frequency analysis layer:

[0068] The continuous wavelet transform (CWT) is used to extract multi-scale voiceprint features, and the wavelet base function is:

[0069] ;

[0070] Wherein is the center frequency, 6 scale subbands (corresponding to the frequency band 50Hz-10kHz) are set, the energy ratio of each subband is calculated, and the energy distribution of different frequency bands is represented, thereby obtaining a 6-dimensional wavelet energy ratio feature.

[0071] Mel frequency cepstral coefficient (MFCC): through the 40-channel mel filter bank (frequency range 100Hz-15kHz), the voiceprint spectrum is nonlinearly compressed, and 128-dimensional MFCC coefficient features are extracted, and the calculation formula is:

[0072] ;

[0073] Wherein is the mel filter bank energy output, .

[0074] Zero-crossing rate and short-time energy envelope: calculate the zero-crossing rate of a frame length of 20 ms (960 points @ 48 kHz) and short-time energy to generate a 2-dimensional dynamic envelope feature.

[0075] (2) Nonlinear analysis layer:

[0076] Recurrence quantitative analysis (RQA): based on the phase space reconstruction (embedding dimension , delay ) of the voiceprint signal, calculate the 3-dimensional RQA feature, including:

[0077] Deterministic coefficient: , where is the diagonal line length distribution, ;

[0078] Laminar flow: , representing the signal stationarity;

[0079] Entropy value: , reflecting the signal complexity.

[0080] (3) Deep learning layer:

[0081] 1D-CNN network architecture: build a 5-layer convolutional network, with the following specific parameters:

[0082] Layer 1: convolution kernel size , channel number , step size , activation function ReLU;

[0083] Layers 2-4: reduce kernel size , multiply channel number , step size ;

[0084] Layer 5: global maximum pooling, output a 256-dimensional deep voiceprint feature vector.

[0085] Finally, feature fusion and dimensionality reduction: input 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) into the fully connected layer for fusion, and then reduce them to 128-dimensional hybrid feature vectors through principal component analysis (PCA).

[0086] Further, the deep learning recognition module adopts a multi-modal neural network architecture, specifically including:

[0087] (1) Spatial feature extraction unit

[0088] Configure a 4-layer dilated convolutional network, and adjust the dilation rate. By progressively expanding the receptive field, local correlations in the acoustic signature spectrum are effectively captured. The dilated convolution operation of the l-th layer is defined as follows:

[0089] ;

[0090] in, For the input feature map, for convolution kernel , This is the bias term. Through dilated convolution, the receptive field of the 4th layer is expanded, significantly improving the ability to capture local details of the acoustic spectrum (such as short-time harmonics and transient pulses). Each layer is followed by batch normalization and ReLU activation function, with the number of output channels being 32, 64, 128, and 256 respectively.

[0091] (2) Temporal modeling unit

[0092] A bidirectional long short-term memory (Bi-LSTM) network was constructed, and an attention mechanism was used to enhance the transient feature representation of the leakage signal. The network structure is as follows:

[0093] Bi-LSTM layers: Each layer has 128 hidden units, and the input is the temporal feature sequence output by the spatial feature extraction unit. The forward and backward LSTM state update formulas are as follows:

[0094] ;

[0095] in In the forward-hidden state, For backward hidden states, the final temporal features .

[0096] Attention mechanism: through learnable weights Calculate attention score :

[0097] ;

[0098] The weighted time series characteristics are Pay close attention to transient signals of leakage (such as sudden water flow sounds or sounds generated by structural cracks).

[0099] Furthermore, the leakage location and early warning module specifically includes:

[0100] (1) Sound source localization layer:

[0101] Improved TDOA algorithm: Optimized time delay estimation based on generalized cross-correlation-phase transform (GCC-PHAT), time delay difference... By the sensor pair Receiving signal cross-correlation function peak determination:

[0102] ;

[0103] Where is the frequency domain signal of sensor i, is the time delay parameter. By non-uniform grid search to locate the peak, combined with the sensor array coordinates , construct overdetermined equations:

[0104] ;

[0105] Where is the sound speed in water, and the three-dimensional coordinates of the leakage point are solved by the least square method .

[0106] Beamforming enhances positioning: based on delay-and-sum beamforming (DSBF), spatial filtering is performed on multi-channel signals to suppress multipath interference and improve azimuth resolution.

[0107] (2) Hierarchical early warning mechanism:

[0108] Three-level threshold setting: yellow warning (leakage amount ): trigger local sound and light alarm, SCADA system records data; orange warning (leakage amount ): start emergency ventilation system, link BIM model high-light leakage point, push work order to maintenance personnel APP; red warning (leakage amount ): full tunnel broadcast evacuation instruction, close adjacent pipe section traffic, start drainage pump set, response delay <3 seconds.

[0109] The above is an exemplary description of the present application, it should be noted that without departing from the core of the present application, any simple modification, modification or other equivalent replacement which can not cost creative labor of those skilled in the art falls within the protection scope of the present application.

Claims

1. A deep learning-based acoustic signature recognition system for immersed tunnel leakage, characterized in that: 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; The acoustic signal acquisition module is used to collect acoustic signal data of seepage water flow sound and structural vibration sound in key parts of the immersed tunnel; The signal preprocessing module is used to perform adaptive noise reduction and frequency domain filtering on the acquired acoustic signal data, and then segment the signal to construct effective voiceprint samples with timestamps. 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. The deep learning recognition module is used to generate classification results of 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 the three-dimensional location of the leakage point. At the same time, it outputs corresponding early warning information based on the identified degree of leakage damage and the preset graded early warning threshold.

2. The deep learning-based acoustic signature recognition system for immersed tunnel leakage according to claim 1, characterized in that: For sound acquisition of leaking water flow: a wideband hydrophone array is used; for sound acquisition of structural vibration: a combination of a triaxial accelerometer and a piezoelectric acoustic emission sensor is used.

3. The deep learning-based acoustic signature recognition system for immersed tunnel leakage according to claim 1, characterized in that: The signal preprocessing module adopts a multi-level signal processing architecture, including: Adaptive noise reduction processing: A noise reference channel is constructed using the LMS algorithm 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 dual threshold detection method of short-time energy and zero-crossing rate, effective voiceprint segments are extracted and standardized format conversion is performed to construct effective voiceprint samples with timestamps.

4. The deep learning-based acoustic signature recognition system for immersed tunnel leakage according to claim 1, characterized in that: The voiceprint feature extraction module employs 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 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 3D RQA features of the deterministic coefficient, laminar flow, and entropy value of the voiceprint samples; Deep learning layer: Build a deep learning model to extract 256-dimensional deep voiceprint features from voiceprint samples.

5. The deep learning-based acoustic signature recognition system for immersed tunnel leakage according to claim 1, characterized in that: The deep learning recognition module employs a multimodal neural network architecture, including: Spatial feature extraction unit: Configured with a 4-layer dilated convolutional network, which expands the receptive field through dilated convolution to capture the local correlation of voiceprint features; Temporal modeling unit: A bidirectional LSTM network is built, and an attention mechanism is used to enhance the transient feature representation of voiceprint features.

6. The deep learning-based acoustic signature recognition system for immersed tunnel leakage according to claim 1, characterized in that: The leakage location and early warning module includes: Sound source localization layer: used to achieve three-dimensional spatial localization of the leakage point; Tiered early warning mechanism: Set three response thresholds to simultaneously trigger audible and visual alarms, SCADA system linkage, and push notifications to maintenance personnel via APP.

7. The deep learning-based acoustic signature recognition system for immersed tunnel leakage according to claim 6, characterized in that: The sound source localization layer employs an improved TDOA algorithm: based on generalized cross-correlation-phase transformation to optimize time delay estimation, and the time delay difference... By sensor The peak value of the cross-correlation function of the received signal is determined; the peak value is located by searching a non-uniform grid, combined with the sensor array coordinates. and Construct an overdetermined system of equations: ； in To find the speed of sound in water, the three-dimensional coordinates of the leakage point are determined using the least squares method. .

Images