An intelligent monitoring and identification method for submarine tunnel damage based on acoustic and optical collaborative perception

Abstract

The present invention relates to the field of civil engineering structural health monitoring, and specifically to a method for intelligent monitoring and identification of submarine tunnel damage based on acoustic-optical collaborative sensing. The method sequentially comprises deploying a three-dimensional acoustic-optical sensing network, active acoustic wave detection, reconstruction of a local sound velocity distribution cloud map, identification of abnormal sound velocity areas, focused sampling of optical fiber responses, and establishment of an intelligent damage diagnosis model. The present invention simultaneously deploys a three-dimensional acoustic array and a distributed optical fiber grid within the submarine tunnel structure, achieving functional complementarity between the two, thereby achieving seamless collaborative sensing for coarse crack location and fine strain location. By employing active acoustic detection, triggered optical fiber high-frequency sampling, and a temperature-strain decoupling algorithm, the method effectively eliminates interference from seawater noise and thermal expansion, accurately distinguishes between "dry cracks" and "water-seepage cracks," and performs real-time monitoring of the submarine tunnel structure, enabling full-process identification of micro-damage, expansion, and failure, and establishing an active early warning mechanism.

Description

Technical Field

[0001] The present invention relates to the field of civil engineering structure health monitoring, and in particular to an intelligent monitoring and identification method for submarine tunnel damage based on acoustic and optical collaborative perception. Background Art

[0002] With the widespread construction of undersea tunnels for transportation, energy, and communications, their long-term service safety faces severe challenges. Due to the high-pressure, highly corrosive, and highly turbulent marine environment, undersea tunnel structures are highly susceptible to various types of damage, such as concrete cracking, leakage at segment joints, and steel corrosion and voids. Failure to promptly identify these early micro-damages can lead to severe structural failure and even catastrophic accidents. Therefore, developing highly sensitive, robust, and robust undersea tunnel structural health monitoring and identification technologies with real-time feedback is crucial.

[0003] Existing monitoring and identification methods mainly include acoustic nondestructive testing technology and distributed fiber optic sensing technology, but both have the following technical bottlenecks in the complex submarine environment:

[0004] The sensing method is single and the detection capability is limited: acoustic wave monitoring has high accuracy in the rough positioning of cracks, but is easily affected by seawater noise, water flow disturbances, etc., and it is difficult to accurately identify tiny cracks or deeply buried voids; distributed fiber optic monitoring can achieve large-scale strain and temperature monitoring, but is insensitive to deep voids or leakages inside tunnel structures, and there is a risk of missed detection.

[0005] Environmental factors cause serious interference and signal decoupling is difficult: seawater leakage not only changes the local sound velocity distribution within the tunnel structure, but also causes sudden changes in the temperature field, resulting in coupling distortion of acoustic and optical fiber signal characteristics. Traditional analysis methods find it difficult to distinguish between critical safety conditions such as "dry cracks" and "water seepage cracks."

[0006] The damage evolution process cannot be tracked and the intelligent diagnostic capability is weak: Existing damage monitoring systems are mostly passive or periodic sampling monitoring modes. It is difficult to capture the entire process of damage from initiation to expansion in real time, and often miss the key window for early warning. There is a lack of active warning mechanisms based on early signs of damage, making it difficult to identify the entire process of micro-damage-extension-failure. In addition, there is a lack of cross-modal data fusion and intelligent classification algorithms. When diagnosing multiple types of damage, it can only rely on empirical thresholds or simple rules. The classification accuracy and anti-interference ability are limited, and the accuracy of damage type identification is not high. Summary of the Invention

[0007] The purpose of the present invention is to address the shortcomings of the existing technology and provide an intelligent monitoring and identification method for submarine tunnel damage based on acoustic and optical collaborative perception. It integrates acoustic active detection and optical fiber high-frequency sampling functions, has multi-physical field decoupling capabilities and intelligent damage monitoring and identification methods, breaks through the key bottlenecks of the existing technology in submarine tunnel damage monitoring and identification, realizes accurate identification and risk warning of complex, multi-source, and multi-scale damage states, and solves the problems existing in the above-mentioned existing technology.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] An intelligent monitoring and identification method for submarine tunnel damage based on acoustic and optical collaborative perception includes the following steps:

[0010] Step 1: Deploy a 3D acoustic-optical sensing network: A sensing layer is installed on the tunnel segment. The sensing layer, a distributed optical fiber strain analyzer for optical fiber signal acquisition and encoding, an edge computing unit, and a shore platform for presenting data analysis results are electrically connected. The sensing layer includes an acoustic sensing layer and an optical fiber sensing layer. The acoustic sensing layer includes an acoustic transducer array, and the optical fiber sensing layer includes distributed optical fibers.

[0011] Step 2: Active acoustic wave detection: The edge computing unit triggers the transducers in the acoustic transducer array to periodically send acoustic coded pulses in sequence, and calculates the transit time of each pair of transducers;

[0012] Step 3: Reconstruct the local sound speed distribution cloud map;

[0013] Step 4: Identify the abnormal sound velocity area;

[0014] Step 5: Fiber Response Focused Sampling: Increase the fiber sampling rate in the area of ​​abnormal acoustic velocity, collect the total strain and temperature sequence actually measured by the optical fiber, input the temperature-strain coupling data into the temperature compensation module, correct the original strain signal affected by temperature, and obtain the true structural strain response. The acoustic wave velocity, optical fiber strain, and temperature data are then integrated to determine whether water seepage cracks or dry cracks are present.

[0015] Step 6. Establish an intelligent damage diagnosis model: The acoustic time-frequency spectrum data and optical fiber timing characteristics are spliced ​​and fused through a dual-channel deep learning network. After passing through a fully connected layer and softmax classification, the regression branch simultaneously outputs the crack width quantification value and leakage risk level.

[0016] Furthermore, in step 1, there are two optical fiber sensing layers, both of which are located outside the tunnel segment, wherein the inner optical fibers are spaced apart from each other along the circumference of the tunnel segment, and the outer optical fibers are spaced apart from each other along the axial direction of the tunnel segment;

[0017] The acoustic transducer array includes a plurality of groups of transducers spaced apart from each other in the axial direction of the tunnel segment, each group of transducers has a plurality of transducers, and the transducers are spaced apart from each other in the circumferential direction of the tunnel segment.

[0018] Furthermore, in the step 2, two transducers are selected from the acoustic transducer array to form a pair of transducers ( ), measure each pair of transducers according to the following formula ( ) of the transit time :

[0019]

[0020] in, For each pair of transducers ( )’s transit time;

[0021] Representative transducer To transducer The physical straight-line distance between

[0022] It is the average propagation velocity of sound waves in the tunnel segments and concrete / steel composite structure under healthy conditions;

[0023] is the actual speed of sound due to damage on a certain propagation path;

[0024] It is reflected as the increase or decrease of time delay relative to the reference speed of sound.

[0025] Furthermore, in step 3, the tunnel section to be monitored is discretized into M voxel units in three-dimensional space. , written as:

[0026]

[0027] Set path The geometric length within the kth voxel unit is:

[0028]

[0029] The relationship between the lengths of all paths and voxel units can form a sparse matrix A:

[0030]

[0031] For the → For the transducer, the total travel time measured is for:

[0032]

[0033] in, is the speed of sound of the k-th voxel unit;

[0034] At uniform reference speed of sound The calculated theoretical travel time is :

[0035]

[0036] Time difference for:

[0037]

[0038] Define the acoustic impedance of the kth voxel unit for:

[0039]

[0040] Then the travel time difference set of all paths can be written as a linear system:

[0041]

[0042] Where, P is the number of transducer pairs;

[0043] s is the vector to be solved;

[0044] is a known observation vector;

[0045] Tikhonov regularization is introduced to solve the following optimization problem:

[0046]

[0047] in, is the regularization parameter;

[0048] R is the smoothing matrix;

[0049] The analytical solution to this problem is:

[0050]

[0051] After solving s, the sound speed of each voxel unit can be restored :

[0052]

[0053] Will By mapping in a three-dimensional grid, we can obtain a local sound speed distribution cloud map.

[0054] Furthermore, in step 4, for longitudinal waves:

[0055] In the reconstructed sound velocity cloud map, the reference sound velocity For reference, calculate the relative deviation field:

[0056]

[0057] in, is the sound velocity change of the k-th voxel unit;

[0058] Take the change of sound speed The absolute value of The voxel unit is a candidate abnormality;

[0059] For Lamb waves:

[0060] By measuring the dispersion curve shift , analyze the gap between layers:

[0061]

[0062] in, is the angular frequency of the emitted sound pulse;

[0063] is the speed of sound in seawater;

[0064] is the incident angle of the incident wave at the interface between the tunnel segment and the seawater;

[0065] is the propagation velocity of the plate wave in the tunnel segment;

[0066] when When ≥0.1rad / m, it is determined that there is a void.

[0067] Furthermore, in step 5, the total strain actually measured by the optical fiber is , the actual measured temperature of the optical fiber is T_fiber, and the interference of thermal expansion is removed by the following formula:

[0068]

[0069] in, is the real structural strain after removing the temperature effect;

[0070] is the total strain measured on the optical fiber;

[0071] is the thermal expansion coefficient of concrete;

[0072] The difference between the actual measured temperature of the optical fiber T_fiber and the ambient temperature of seawater T_sea;

[0073] When |T_fiber-T_sea|≤0.5℃ and the local sound speed are satisfied in the region When the two conditions of ≥1400m / s are met, it is judged as a water seepage crack, otherwise it is a dry crack.

[0074] Furthermore, in step 6, the intelligent damage diagnosis model includes an acoustic feature extraction channel and a fiber feature extraction channel, wherein the acoustic branch is based on a 3D convolutional neural network, and the output vector is recorded as:

[0075]

[0076] The fiber branching is based on the long short-term memory network, and the output vector is recorded as:

[0077]

[0078] The fusion feature vector f is obtained by splicing the two modal branch features along the dimension merged as follows:

[0079]

[0080] After passing through the fully connected layer, the score vector for each damage type is output:

[0081]

[0082] Among them, z is the rating vector;

[0083] W cls is the weight matrix of the fully connected layer;

[0084] f merged is the fusion feature vector;

[0085] b cls is the bias vector of the classification branch;

[0086] R 6 Indicates that the number of damage types output in the classification branch is 6;

[0087] The probability distribution of 6 types of damage is obtained by the Softmax function :

[0088]

[0089] in, Expressed as categories;

[0090] The damage types are transverse cracks, longitudinal cracks, voids, corrosion, spalling, and composite damage;

[0091] The regression branch is parallel to the Softmax classification branch, and the other output branch of the network is used to predict the structural quantization parameters in the form of:

[0092]

[0093] in, is the model's predicted output for the input sample;

[0094] W reg is the weight matrix of the fully connected layer;

[0095] f merged is the fusion feature vector;

[0096] b reg is the bias vector of the classification branch;

[0097] R 3 It means that the number of continuous variables predicted in the regression branch is 3: crack width, void area, and seepage rate.

[0098] Furthermore, the acoustic branch input size is (t×f×sensor_num)=(50 frames×128 frequency points×6 sensors), and it undergoes 32 5×5×5 convolutions, 2×2×1 maximum pooling, and 64 3×3×3 convolutions to extract spatial-spectral features.

[0099] The fiber branch input (time_steps=100, fiber_points=100), two layers of LSTM units, followed by TimeDistributed(Dense(64)), extracts the time series strain / temperature change pattern.

[0100] Compared with the prior art, the present invention has the following beneficial effects:

[0101] 1. By simultaneously deploying a three-dimensional acoustic array and a distributed fiber optic grid within the submarine tunnel structure, the two complement each other's functions, thereby achieving seamless collaborative perception of coarse crack positioning and fine strain positioning.

[0102] 2. The use of active acoustic detection, triggered optical fiber high-frequency sampling and temperature-strain decoupling algorithm effectively eliminates the interference of seawater noise and thermal expansion, accurately distinguishes "dry cracks" from "water seepage cracks", and conducts real-time monitoring of the submarine tunnel structure, effectively capturing the entire process of damage from initiation to expansion, realizing the full process identification of micro-damage-expansion-failure, and establishing an active early warning mechanism.

[0103] 3. By building a deep learning diagnostic model based on 3D-CNN and LSTM, the acoustic and optical spatiotemporal spectral features are integrated and analyzed, which greatly improves the recognition accuracy and real-time warning capabilities of multiple types of damage.

[0104] 4. The intelligent tunnel damage identification system integrates an acoustic sensing layer, a fiber optic sensing layer, a distributed fiber optic analyzer, and an edge computing unit to form an engineered system that can be deployed on-site. It not only meets the hardware requirements of high voltage resistance and corrosion resistance, but also takes into account algorithm upgrades and convenient maintenance, thereby comprehensively improving the sensitivity, anti-interference and intelligence level of submarine tunnel health monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] Figure 1 A schematic diagram of the overall structure of the tunnel segment provided by the present invention;

[0106] Figure 2 An axial front view of the tunnel segment provided by the present invention;

[0107] Figure 3 A top view of the tunnel segment provided by the present invention;

[0108] Figure 4 This is a schematic diagram of the structure of the intelligent damage monitoring and identification system provided by the present invention;

[0109] Figure 5 A structural diagram of the deep learning model provided by the present invention;

[0110] Figure 6 The present invention provides a flow chart of an intelligent identification method for submarine tunnel damage based on acoustic-optical collaborative perception and multi-physical field coupling.

[0111] Wherein, the accompanying drawings are marked as follows:

[0112] 1. Tunnel segment; 2. Transducer; 3. Optical fiber. DETAILED DESCRIPTION

[0113] In order to help those skilled in the art better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0114] It should be noted that when an element is referred to as being “fixed on” or “disposed on” another element, it may be directly on the other element or indirectly on the other element. When an element is referred to as being “connected to” another element, it may be directly connected to the other element or indirectly connected to the other element.

[0115] It should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on this application.

[0116] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or to implicitly indicate the quantity of the technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features.

[0117] For easier understanding, see Figures 1 to 6 This embodiment provides an intelligent identification method for submarine tunnel damage based on acoustic-optical collaborative sensing and multi-physical field coupling, comprising the following steps:

[0118] Step 1: Deploy a 3D acoustic and optical sensing network

[0119] The intelligent damage monitoring and identification system consists of a sensing layer, a distributed fiber optic strain analyzer (BOTDA), an edge computing unit, and a shore platform, all of which are electrically connected. The sensing layer comprises an acoustic sensing layer and a fiber optic sensing layer located within the tunnel structure. The distributed fiber optic strain analyzer, connected to a distributed fiber optic sensor network located within tunnel segment 1, serves as the optical fiber signal acquisition and decoding device. The edge computing unit is the core component for achieving "on-site intelligent identification" and "multi-source data coupling processing," deploying an intelligent damage diagnosis model. The shore platform presents data analysis results.

[0120] Specifically, the optical fiber sensing layer includes two layers of distributed optical fibers 3 (BOTDA / BOTDR) fixedly embedded on the outside of the tunnel segment 1 (i.e., between the tunnel segment 1 and the concrete protective layer), wherein the inner optical fibers 3 are arranged at intervals of 20 cm along the circumference of the tunnel segment 1 and extend in the axial direction of the tunnel segment 1, and the outer optical fibers 3 are arranged at intervals of 20 cm along the axial direction of the tunnel segment 1 and wrap the inner optical fibers 3 in the circumferential direction of the tunnel segment 1. The two layers of distributed optical fibers 3 form a grid arrangement of optical fibers 3. Preferably, the diameter of the optical fiber 3 is 0.9 mm, and the outer layer is a polyethylene coating. Furthermore, the temperature resistance range of the optical fiber 3 is -40°C to 85°C, and the maximum strain measurement range is ±5,000 , with a spatial resolution of ≤10cm. The acoustic sensing layer comprises a piezoelectric acoustic transducer array fixedly mounted on the inner side of tunnel segment 1. Six transducers 2 are arranged in groups of 5m apart in the axial direction of tunnel segment 1. Each group of transducers 2 consists of six transducers, spaced 60° apart circumferentially around the tunnel segment 1. The frequency response of transducers 2 ranges from 10kHz to 100kHz, stimulating composite signals of longitudinal, shear, and Lamb waves. The acoustic transducer array is connected to the acoustic transceiver control unit via a coaxial submarine connector and is centrally powered and controlled by the edge computing unit.

[0121] Step 2: Active acoustic detection

[0122] The multi-channel acoustic transmitter is triggered by the edge computing unit, so that the transducers 2 in the acoustic transducer array send Chirp coded pulses (10kHz→50kHz longitudinal wave, 50kHz→100kHz Lamb wave) in sequence. Specifically, each transducer 2 is both a transmitter and a receiver. Two transducers 2 are selected from the acoustic transducer array to form a pair of transducers ( ), where the transducer As the transmitter, the edge computing unit controls the emission of coded sound waves, and the transducer As the receiving end, the receiver is the transducer The signal returned by the transmitted sound wave after passing through the tunnel structure.

[0123] The receiving end measures each pair of transducers according to the following formula ( ) of the transit time :

[0124]

[0125] in, For each pair of transducers ( )’s transit time;

[0126] Representative transducer To transducer The physical straight-line distance between

[0127] It is the benchmark sound velocity measured in advance in a healthy state, that is, when the structure is intact. It corresponds to the average propagation velocity of sound waves in the tunnel segment concrete / steel composite structure when there are no cracks or voids. It can be obtained through theory, experiments or initial measurements.

[0128] The actual sound velocity caused by damage on a propagation path. When cracks or voids exist, the wave velocity will shift due to the reduced medium stiffness. This can be obtained by measuring the sound wave propagation time and combining it with the structural geometry to identify areas of abnormal sound velocity (i.e., potential damage areas) in the structure.

[0129] It is reflected as the increase or decrease of time delay relative to the reference speed of sound.

[0130] Step 3: Reconstruct the local sound speed distribution cloud map

[0131] Discretize the tunnel section to be monitored (or several continuous sections along the tunnel axis) into M voxel units in three-dimensional space. , written as:

[0132]

[0133] Each pair of acoustic transducers → All possible propagation paths are considered as a ray and intersect with these voxel cells.

[0134] Set path The geometric length within the kth voxel unit is:

[0135]

[0136] The relationship between the lengths of all paths and voxel units forms a sparse matrix A:

[0137]

[0138] For the → For the transducer, the measured total travel time (Time-of-Flight) for:

[0139]

[0140] in, is the speed of sound of the k-th voxel unit.

[0141] At uniform reference speed of sound Calculated theoretical travel time for:

[0142]

[0143] Time difference for:

[0144]

[0145] Define the acoustic impedance of the kth voxel for:

[0146]

[0147] Then the travel time difference set of all paths can be written as a linear system:

[0148]

[0149] Where, P is the number of transducer pairs;

[0150] s is the vector to be solved;

[0151] is the known observation vector.

[0152] Considering data noise and inversion stability, Tikhonov regularization is introduced to solve the following optimization problem:

[0153]

[0154] in, is the regularization parameter, which can be determined by cross-validation or L-curve regularization method;

[0155] R is a smoothing matrix used to suppress high-frequency oscillations of the solution.

[0156] The analytical solution to this problem is:

[0157]

[0158] After solving s, the sound speed of each voxel unit can be restored :

[0159]

[0160] Will By mapping in a three-dimensional grid, we can obtain a local sound speed distribution cloud map.

[0161] Step 4: Identify abnormal sound velocity areas

[0162] 1) For longitudinal waves

[0163] In the reconstructed sound velocity cloud map, the reference sound velocity For reference, calculate the relative deviation field:

[0164]

[0165] in, is the sound velocity change of the k-th voxel unit.

[0166] Take the change of sound speed The absolute value of The voxel unit is a candidate abnormality.

[0167] 2) For Lamb waves

[0168] By measuring the dispersion curve shift , analyze the gap between layers:

[0169]

[0170] in, is the angular frequency of the emitted sound pulse ( , f is the frequency);

[0171] is the speed of sound in seawater (about 1500 m / s), used to calculate critical leakage conditions;

[0172] is the incident angle of the incident wave at the interface between the tube segment and the seawater. When the critical angle is exceeded, energy leakage into the water occurs;

[0173] It is the propagation speed of plate waves in tunnel segments. Unlike longitudinal waves or shear waves, plate waves are particularly sensitive to interlayer voids, so their leakage characteristics can reflect void defects.

[0174] when When ≥0.1rad / m, it is determined that there is a void.

[0175] Step 5: Fiber Response Focusing Sampling

[0176] When a change in sound velocity is detected or Abnormal (the trigger threshold is the absolute value of the sound speed change ≥±2% or dispersion curve offset ≥0.1rad / m), the edge computing unit immediately increases the BOTDA sampling rate of the corresponding grid node from the normal 1Hz to 100Hz, continuously samples for 10s, and obtains the total strain actually measured by the optical fiber with high precision. And the actual measured temperature T_fiber sequence of the optical fiber.

[0177] The temperature-strain coupling data is input into the temperature compensation module to correct the original strain signal affected by temperature, eliminate the temperature effect, and obtain the true structural strain response.

[0178] Specifically, the thermal expansion interference is removed by the following formula:

[0179]

[0180] in, is the real structural strain after removing the temperature effect;

[0181] The total strain measured for the optical fiber includes structural strain and thermal expansion and contraction caused by temperature changes;

[0182] is the thermal expansion coefficient of concrete (approximately 7×10^-6 / °C), which is used to convert the fiber temperature response into equivalent strain;

[0183] It is the difference between the actual measured temperature of the optical fiber T_fiber and the ambient temperature of seawater T_sea. Subtracting the two can obtain the false strain introduced by the thermal effect, thereby eliminating interference.

[0184] When |T_fiber-T_sea|≤0.5℃ (i.e. | |≤0.5℃) and local sound velocity When the two conditions of ≥1400m / s are met, it is judged as a water seepage crack, otherwise it is a dry crack.

[0185] Step 6: Establish an intelligent damage diagnosis model

[0186] The system integrates acoustic time-frequency spectrum data and optical fiber timing characteristics through a dual-channel deep learning network:

[0187] The intelligent damage diagnosis model includes two functional channels: an acoustic feature extraction channel (based on a 3D convolutional neural network, or 3D-CNN) and a fiber feature extraction channel (based on a long short-term memory network, or LSTM).

[0188] Among them, the acoustic branch is based on 3D-CNN, with an input size of (t×f×sensor_num) = (50 frames×128 frequency points×6 sensors). It undergoes 32 5×5×5 convolutions, 2×2×1 maximum pooling, and 64 3×3×3 convolutions to extract spatial-spectral features.

[0189] The first convolutional layer: 32 convolution kernels of size 5×5×5, extracting joint patterns across time, frequency, and space;

[0190] Maximum pooling layer: 2×2×1, retaining spatial structure features;

[0191] The second convolutional layer: 64 convolution kernels of size 3×3×3 are used to extract local acoustic changes;

[0192] The output vector is denoted as

[0193]

[0194] Among them, the dimension is N1.

[0195] Specifically, the 5×5×5 convolution kernel has a large receptive field in the three dimensions of time-frequency-space, and can extract deep linkage features across frames and frequency bands;

[0196] The 3×3×3 convolution kernel is used to capture local mutations at a finer scale;

[0197] The first layer uses a smaller number of 32 channels to save computation, and the second layer uses 64 channels to enrich feature expression.

[0198] The fiber branching is based on LSTM, with input (time_steps=100, fiber_points=100), two layers of LSTM units, followed by TimeDistributed(Dense(64)) to extract the temporal strain / temperature change pattern.

[0199] Two layers of LSTM units, each layer contains 128 hidden state units;

[0200] The TimeDistributed(Dense(64)) operation is used at the end to compress the output of each time step;

[0201] The input represents the strain / temperature changes at 100 monitoring points collected by the optical fiber in 100 time steps;

[0202] The output vector is denoted as:

[0203]

[0204] Among them, the dimension is N2.

[0205] After the above acoustic branch outputs and optical fiber branch outputs are spliced ​​in the fusion layer (Concatenate), they pass through the fully connected layer and softmax classification (six types of damage: transverse cracks, longitudinal cracks, voids, rust, spalling, and combined damage). The regression branch simultaneously outputs the quantitative value of the crack width and the leakage risk level (level 1-5).

[0206] Specifically, the acoustic branch output vector and the optical fiber branch output vector are spliced ​​at the fusion layer as follows:

[0207] When the acoustic branch and the optical fiber branch output eigenvectors respectively:

[0208]

[0209] The fused feature vector obtained by concatenating the two modal branch features along the dimension is as follows:

[0210]

[0211] After splicing, a fully connected layer (FC) is input to achieve deep coupling of cross-modal features.

[0212] After passing through the fully connected layer, the score vector for each damage type is output:

[0213]

[0214] Among them, z is the rating vector;

[0215] W cls is the weight matrix of the fully connected layer;

[0216] f merged is the fusion feature vector;

[0217] b cls is the bias vector of the classification branch;

[0218] R 6 Indicates that the number of damage type categories output in the classification branch is 6.

[0219] The probability distribution of six types of damage (transverse cracks, longitudinal cracks, voids, corrosion, spalling, and composite damage) is obtained through the Softmax function. :

[0220]

[0221] Among them, exp represents the indicator function;

[0222] Indicates the categories.

[0223] The damage type often affects the assessment of tunnel structures and subsequent maintenance strategies. Different damage types have different impacts on tunnels. Therefore, the probability distribution of the six damage types not only helps the system make the final damage type classification, but also provides more detailed data support for subsequent quantitative assessments and risk warnings.

[0224] The regression branch is parallel to the Softmax classification branch, and the other output branch of the network is used to predict the structural quantization parameters in the form of:

[0225]

[0226] in, is the model's predicted output for the input sample;

[0227] W reg is the weight matrix of the fully connected layer;

[0228] f merged is the fusion feature vector;

[0229] b reg is the bias vector of the classification branch;

[0230] R 3 Indicates that the number of continuous variables predicted in the regression branch is 3: crack width, void area, and seepage rate. That is, the output includes: crack width (mm), void area (cm²), and seepage rate (mL / min).

[0231] Table 1 Leakage risk level classification

[0232] Risk Level Risk Description Crack width (mm) Leakage flow (L / h·m) Temperature difference (℃) Speed ​​of sound (m / s) Judgment conditions & recommended measures Level 1 Very low risk ＜0.2 ＜0.1 ≤0.2 ＜1400 The crack width is very small, the temperature difference between the optical fiber and the seawater is ≤0.2℃; there is no obvious abnormality in the sound velocity; Recommendation: only routine inspection Level 2 Low risk 0.2-0.5 0.1-0.5 0.2-0.5 1400-1450 Crack width 0.2–0.5 mm; water leakage flow ≤ 0.5 L / h·m; sound velocity slightly increased; Recommendation: Increase the frequency of testing and pay attention to the changing trend Level 3 Medium risk 0.5-1.0 0.5-2.0 0.5-1.0 1450-1500 Crack width 0.5–1.0 mm; water leakage rate 0.5–2.0 L / h·m; significant increase in sound velocity; Recommendation: On-site diving review and initiation of local reinforcement Level 4 Higher risk 1.0-2.0 2.0-5.0 1.0-2.0 1500-1550 Crack width 1.0–2.0 mm; water leakage rate 2.0–5.0 L / h·m; abnormally high sound velocity; Recommendation: Emergency reinforcement and temporary sealing Level 5 Very high risk (critical) ＞2.0 ＞5.0 ＞2.0 ＞1550 Crack width > 2.0mm; water leakage flow > 5.0L / h·m; rapid increase in sound velocity; Recommendation: Evacuate personnel immediately and initiate comprehensive rescue and repair efforts

[0233] In summary, the model first generates a probability distribution of six types of damage for the abnormal area in tunnel segment 1. Subsequently, the regression branch predicts the corresponding physical quantitative parameters in this area. Finally, the leakage risk classification table combines the a priori criticality of the damage type with the actual severity of the quantitative parameters to produce a comprehensive risk rating of 1-5.

[0234] The fully connected layer and Softmax classification are the standard "classification head" designs in deep classification networks: the fully connected layer integrates all previously extracted features, combines scattered local features into a global feature vector, and completes the "feature → category score" mapping. Softmax converts the feature scores output by the fully connected layer into category probability distributions, mapping the scores into trainable and interpretable probability distributions. The combination of the two makes multi-category damage identification both efficient and easy to optimize and deploy.

[0235] Although the present invention has been described using the above preferred embodiments, they are not intended to limit the scope of protection of the present invention. Any person skilled in the art who makes various changes and modifications to the above embodiments without departing from the spirit and scope of the present invention still fall within the scope of protection of the present invention.

Claims

1. An intelligent monitoring and identification method for submarine tunnel damage based on acoustic and optical collaborative perception, characterized in that: The steps include: Step 1: Deploy a 3D acoustic-optical sensing network: A sensing layer is installed on the tunnel segment. The sensing layer, a distributed optical fiber strain analyzer for optical fiber signal acquisition and encoding, an edge computing unit, and a shore platform for presenting data analysis results are electrically connected. The sensing layer includes an acoustic sensing layer and an optical fiber sensing layer. The acoustic sensing layer includes an acoustic transducer array, and the optical fiber sensing layer includes distributed optical fibers. Step 2: Active acoustic wave detection: The edge computing unit triggers the transducers in the acoustic transducer array to periodically send acoustic coded pulses in sequence, and calculates the transit time of each pair of transducers; Step 3: Reconstruct the local sound speed distribution cloud map; Step 4: Identify the abnormal sound velocity area; Step 5: Fiber Response Focused Sampling: Increase the fiber sampling rate in the area of ​​abnormal acoustic velocity, collect the total strain and temperature sequence actually measured by the optical fiber, input the temperature-strain coupling data into the temperature compensation module, correct the original strain signal affected by temperature, and obtain the true structural strain response. The acoustic wave velocity, optical fiber strain, and temperature data are then integrated to determine whether water seepage cracks or dry cracks are present. Step 6: Establish an intelligent damage diagnosis model: The acoustic time-frequency spectrum data and fiber time series features are spliced ​​and fused through a dual-channel deep learning network. After passing through a fully connected layer and softmax classification, the regression branch simultaneously outputs the crack width quantification value and leakage risk level. In step 6, the intelligent damage diagnosis model includes an acoustic feature extraction channel and a fiber feature extraction channel, wherein the acoustic branch is based on a 3D convolutional neural network, and the output vector is recorded as: ； The fiber branching is based on the long short-term memory network, and the output vector is recorded as: ； The fusion feature vector obtained by splicing the two modal branch features along the dimension as follows: ； After passing through the fully connected layer, the score vector for each damage type is output: ； Among them, z is the rating vector; is the weight matrix of the fully connected layer; is the fusion feature vector; is the bias vector of the classification branch; Indicates that the number of damage types output in the classification branch is 6; The probability distribution of 6 types of damage is obtained by the Softmax function ; ； Among them, i represents the i-th category; The damage types are transverse cracks, longitudinal cracks, voids, corrosion, spalling, and composite damage; The regression branch is parallel to the Softmax classification branch, and the other output branch of the network is used to predict the structural quantization parameters in the form of: ； in, is the model's predicted output for the input sample; is the weight matrix of the fully connected layer; is the fusion feature vector; is the bias vector of the regression branch; It means that the number of continuous variables predicted in the regression branch is 3: crack width, void area, and seepage rate.

2. The intelligent monitoring and identification method for submarine tunnel damage based on acoustic and optical collaborative perception according to claim 1 is characterized in that: In step 1, there are two optical fiber sensing layers, both of which are located outside the tunnel segment, wherein the inner optical fibers are spaced apart from each other along the circumference of the tunnel segment, and the outer optical fibers are spaced apart from each other along the axial direction of the tunnel segment; The acoustic transducer array includes a plurality of groups of transducers spaced apart from each other in the axial direction of the tunnel segment, each group of transducers has a plurality of transducers, and the transducers are spaced apart from each other in the circumferential direction of the tunnel segment.

3. The intelligent monitoring and identification method for submarine tunnel damage based on acoustic and optical collaborative perception according to claim 1 is characterized in that: In the second step, two transducers are selected from the acoustic transducer array to form a pair of transducers. , measure each pair of transducers according to the following formula Transit time : ； in, For each pair of transducers The transit time; Representative transducer To transducer The physical straight-line distance between It is the average propagation velocity of sound waves in the tunnel segments and concrete / steel composite structure under healthy conditions; is the actual speed of sound due to damage on a certain propagation path; It is reflected as the increase or decrease of time delay relative to the reference speed of sound.

4. The intelligent monitoring and identification method for submarine tunnel damage based on acoustic and optical collaborative perception according to claim 3 is characterized in that: In step 3, the tunnel section to be monitored is discretized into M voxel units in three-dimensional space. , written as: ； Set path The geometric length within the kth voxel unit is: ； The length relationship between all paths and voxel units constitutes a sparse matrix : ； For the For the transducer, the total travel time measured is for: ； in, is the speed of sound of the k-th voxel unit; At uniform reference speed of sound The calculated theoretical travel time is : ； Time difference for: ； Define the acoustic impedance of the kth voxel unit for: ； Then the travel time difference set of all paths can be written as a linear system: ； Where, P is the number of transducer pairs; s is the vector to be solved; is a known observation vector; Tikhonov regularization is introduced to solve the following optimization problem: ； in, is the regularization parameter; is the smoothing matrix; The analytical solution to this problem is: ； After solving s, the sound speed of each voxel unit can be restored : ； Will By mapping in a three-dimensional grid, we can obtain a local sound speed distribution cloud map.

5. The intelligent monitoring and identification method for submarine tunnel damage based on acoustic and optical collaborative perception according to claim 4 is characterized in that: In the fourth step, for longitudinal waves: in the reconstructed sound velocity cloud map, the reference sound velocity For reference, calculate the relative deviation field: ； in, is the sound velocity change of the k-th voxel unit; Take the change of sound speed The absolute value of The voxel unit is a candidate abnormality; For Lamb waves: By measuring the dispersion curve shift , analyze the gap between layers: ； in, is the angular frequency of the emitted sound pulse; is the speed of sound in seawater; is the incident angle of the incident wave at the interface between the tunnel segment and the seawater; is the propagation velocity of the plate wave in the tunnel segment; when When , it is determined that there is a gap.

6. The intelligent monitoring and identification method for submarine tunnel damage based on acoustic and optical collaborative perception according to claim 1 is characterized in that: In step 5, the total strain actually measured by the optical fiber is , the actual temperature measured by the optical fiber is , the thermal expansion interference is removed by the following formula: ; in, is the real structural strain after removing the temperature effect; is the total strain measured on the optical fiber; is the thermal expansion coefficient of concrete; The actual measured temperature of the optical fiber Seawater ambient temperature difference; When the region satisfies and the local speed of sound When both conditions are met, it is judged as a water seepage crack, otherwise it is a dry crack.

7. The intelligent monitoring and identification method for submarine tunnel damage based on acoustic and optical collaborative perception according to claim 1 is characterized in that: In step 6, the acoustic branch input size 50 frames × 128 frequency points × 6 sensors, sequentially undergoing 32 5×5×5 convolutions, 2×2×1 maximum pooling, and 64 3×3×3 convolutions to extract spatial-spectral features; The optical fiber branch input Two layers of LSTM units, followed by TimeDistributed(Dense(64)), extract temporal strain / temperature change patterns.

Images