Underwater multi-mode identification method and system for hidden microdefects of high dam

Abstract

The invention provides an underwater multi-modal identification method and system for hidden micro defects of a high dam, and the method comprises the following steps: obtaining displacement monitoring data, stress monitoring data and temperature monitoring data, generating a multi-source monitoring data matrix through time synchronization processing, extracting abnormal features through employing a self-adaptive weight method, and carrying out the recognition of the abnormal features. Obtaining a suspicious area list with priority; planning a sonar scanning path according to the suspicious area list, obtaining edge feature data through multi-stage signal enhancement processing, and determining an accurate defect area; acoustic data and optical data are collected, and a fusion feature matrix is obtained through space-time registration and feature fusion; and carrying out feature discrimination on the fusion feature matrix, determining defect types and levels, obtaining a positioning result and a reliability evaluation result through multi-level analysis, and finally generating a defect identification report. According to the invention, the accuracy and reliability of underwater hidden micro defect identification are improved.

Description

Technical Field

[0001] The invention relates to a multi-modal data processing technology, in particular to an underwater multi-modal identification method and system for high dam hidden defects. Background Art

[0002] With the continuous expansion of the scale of water conservancy project construction, the safe operation of dams faces increasingly severe challenges. Underwater micro defects of high dams are one of the important factors threatening dam safety. These defects are often small in size, unclear in features, and develop rapidly. Timely discovery and accurate identification of underwater micro defects are of great significance to ensure the safe operation of dams, extend their service life, and reduce maintenance costs.

[0003] At present, underwater defect detection mainly adopts single-mode detection methods, such as sonar scanning, optical imaging or displacement monitoring. Traditional sonar detection detects underwater structures by emitting sound waves and receiving echo signals, but it is often affected by equipment performance and underwater environment, and often has problems such as limited resolution and severe noise interference; although optical detection can obtain intuitive image information, the imaging quality is seriously reduced in turbid water bodies and the detection distance is limited; although displacement monitoring can reflect the structural deformation in real time, the monitoring points are limited and it is difficult to cover all possible defect locations. These single detection methods have shown obvious limitations in practical applications.

[0004] In actual engineering applications, the current detection technology still has the following specific problems: First, the data collected by different sensors have inconsistent sampling frequencies and uneven data quality, which makes it difficult to fuse multi-source data and poor extraction of abnormal features; second, in complex underwater environments, sonar signals are easily affected by factors such as temperature stratification, suspended matter and bubbles, causing attenuation and distortion of echo signals, affecting the accurate extraction of defect features; third, due to the complex and changeable defect morphology, the existing feature extraction methods are difficult to simultaneously take into account the feature expressions of different types of defects such as cracks, erosion, and leakage, reducing the accuracy of recognition; fourth, in the process of multimodal data processing, the time and space alignment accuracy of different types of data is insufficient, affecting the feature fusion effect; fifth, the lack of a scientific reliability evaluation mechanism makes it difficult to quantitatively evaluate the accuracy and stability of the detection results, affecting the credibility of the detection results. The existence of the above technical problems has seriously restricted the effect and application promotion of underwater micro-defect detection. Summary of the invention

[0005] The purpose of the invention is to provide an underwater multimodal identification method and system for high dam subtle defects, in order to solve at least one technical problem existing in the prior art.

[0006] The technical solution is an underwater multi-modal identification method for high dam subtle defects, which includes the following steps:

[0007] Step S1, obtaining displacement monitoring data, stress monitoring data and temperature monitoring data from the sensor network database in real time, and generating a multi-source monitoring data matrix through time synchronization processing; extracting abnormal features from the multi-source monitoring data matrix using an adaptive weight calculation method to generate an abnormal feature matrix; inputting the abnormal feature matrix into a pre-configured multidimensional analysis model, calculating a suspicious area coordinate set, and generating a credibility score at the same time; based on the credibility score, prioritizing the suspicious area coordinate set, and outputting a suspicious area list;

[0008] Step S2, obtaining a list of suspicious areas and planning a scanning path according to the list of suspicious areas; based on the scanning path, obtaining original sonar data, and obtaining enhanced sonar data through multi-level signal enhancement processing; based on the enhanced sonar data, extracting edge feature data, performing multi-feature analysis, and determining the precise defect area and defect feature description;

[0009] Step S3, based on the precise defect area and defect feature description, collect acoustic data and optical data, and obtain registration data through spatiotemporal alignment; enhance the registration data to obtain enhanced feature data; based on the enhanced feature data, extract and fuse features to obtain multimodal feature data; based on the multimodal feature data, generate a fusion feature matrix and feature credibility index;

[0010] Step S4: Based on the fusion feature matrix and feature credibility index, the defect identification result is obtained through feature identification; based on the defect identification result, a multi-level analysis is performed to determine the defect type and defect level; based on the defect type and defect level, the precise spatial position is calculated to obtain the positioning result; based on the positioning result, a comprehensive evaluation is performed to obtain the reliability evaluation result, and a defect identification report and a warning parameter update list are generated.

[0011] An underwater multi-modal identification system for high dam subtle defects, comprising:

[0012] at least one processor; and,

[0013] a memory communicatively connected to at least one of the processors; wherein,

[0014] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the underwater multi-modal identification method of high dam subtle defects.

[0015] Beneficial effects: The present invention realizes rapid and accurate positioning of suspicious areas, improves the quality and reliability of detection data, and enhances the integrity and accuracy of feature expression; it realizes accurate identification and quantitative evaluation of defects, and solves technical problems such as data heterogeneity, environmental interference, difficulty in feature extraction, and low recognition accuracy in underwater defect detection of dams. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of the present invention.

[0017] Figure 2 It is a flow chart of step S1 of the present invention.

[0018] Figure 3 It is a flow chart of step S2 of the present invention.

[0019] Figure 4 It is a flow chart of step S3 of the present invention.

[0020] Figure 5 It is a flow chart of step S4 of the present invention. DETAILED DESCRIPTION

[0021] like Figure 1 As shown, the present application proposes an underwater multi-modal identification method for high dam hidden defects, comprising the following steps:

[0022] Step S1, obtaining displacement monitoring data, stress monitoring data and temperature monitoring data from the sensor network database in real time, and generating a multi-source monitoring data matrix through time synchronization processing; extracting abnormal features from the multi-source monitoring data matrix using an adaptive weight calculation method to generate an abnormal feature matrix; inputting the abnormal feature matrix into a pre-configured multidimensional analysis model, calculating a suspicious area coordinate set, and generating a credibility score at the same time; based on the credibility score, prioritizing the suspicious area coordinate set, and outputting a suspicious area list;

[0023] Step S2, obtaining a list of suspicious areas and planning a scanning path according to the list of suspicious areas; based on the scanning path, obtaining original sonar data, and obtaining enhanced sonar data through multi-level signal enhancement processing; based on the enhanced sonar data, extracting edge feature data, performing multi-feature analysis, and determining the precise defect area and defect feature description;

[0024] Step S3, based on the precise defect area and defect feature description, collect acoustic data and optical data, and obtain registration data through spatiotemporal alignment; enhance the registration data to obtain enhanced feature data; based on the enhanced feature data, extract and fuse features to obtain multimodal feature data; based on the multimodal feature data, generate a fusion feature matrix and feature credibility index;

[0025] Step S4: Based on the fusion feature matrix and feature credibility index, the defect identification result is obtained through feature identification; based on the defect identification result, a multi-level analysis is performed to determine the defect type and defect level; based on the defect type and defect level, the precise spatial position is calculated to obtain the positioning result; based on the positioning result, a comprehensive evaluation is performed to obtain the reliability evaluation result, and a defect identification report and a warning parameter update list are generated.

[0026] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0027] Step S11, reading displacement monitoring data corresponding to a three-dimensional spatial position from a displacement sensor acquisition node, reading stress monitoring data at the same position from a stress sensor acquisition node, and reading temperature monitoring data at the same position from a temperature sensor acquisition node; performing sampling frequency detection on the displacement monitoring data, stress monitoring data, and temperature monitoring data, respectively, to generate a sampling feature vector; constructing linear interpolation parameters based on the sampling feature vector, resampling and time-aligning the displacement monitoring data, stress monitoring data, and temperature monitoring data, to generate time-aligned data; organizing the time-aligned data according to spatial position and sensor type, and outputting a multi-source monitoring data matrix;

[0028] The time alignment process is as follows: T(i, j) = Tb(i) +α [D(i, j) / Dmax] [1 + β Q(i, j)]; where T(i, j) is the alignment time point; Tb(i) is the reference time series; D(i, j) is the data delay; Dmax is the maximum allowable delay; Q(i, j) is the data quality factor; α is the delay compensation coefficient (0.51.0); β is the quality adjustment factor (0.20.4). S(i) =Σ[wk |T(i, k)-T(i, k-1)|] exp(-γ V(i)); where S(i) is the synchronization score; wk is the time weight; V(i) is the data variability; γ is the stability coefficient (0.3~0.6).

[0029] The process of organizing the spatial position includes: spatial coordinate unified transformation algorithm: P'(x, y, z) = RP(x, y, z) + t +Δ(x, y, z); where P'(x, y, z) is the transformed coordinate; P(x, y, z) is the original coordinate; R is the rotation matrix; t is the translation vector; Δ(x, y, z) is the nonlinear correction term. Δ(x, y, z) = Σ[wi B(x, y, z)] [1 + α E(x, y, z)]; where B(x, y, z) is the basis function value; E(x, y, z) is the environmental impact factor; wi is the basis function weight; α is the environmental correction coefficient (0.1~0.3).

[0030] Step S12, dividing the multi-source monitoring data matrix into a core data part representing the main change trend and a peripheral data part representing the fluctuation change, calculating the adaptive weight coefficients of the core data part and the peripheral data part; multiplying the core data part and the peripheral data part by the corresponding adaptive weight coefficients respectively and adding them to obtain weighted monitoring data; based on the weighted monitoring data, calculating the distance between each monitoring point and its adjacent monitoring point; converting the distance into a correlation weight value, the correlation weight value decreases as the distance between the monitoring points increases; multiplying the correlation weight value by the weighted monitoring data to obtain an abnormal feature matrix;

[0031] The multi-source monitoring data matrix is ​​divided into a core data part representing the main trend of change and a peripheral data part representing the fluctuation change. The following algorithm is used: Core data and peripheral data decomposition algorithm: U(x, y) =Σ[wi Di(x, y)] [1+α F(x, y)]; where U(x, y) is the decomposition eigenvalue; Di(x, y) is the i-th layer decomposition result; wi is the level weight; F(x, y) is the frequency feature; α is the feature adjustment coefficient (0.2~0.4). Mc = U(x, y) [1 + β |gradU(x, y)|]; Mp = M is Mc; where Mc is the core data; Mp is the peripheral data; M is the original data; gradU(x, y) is the feature gradient; β is the gradient weight (0.1~0.3), and grad is the gradient operator.

[0032] Step S13: for each monitoring point in the abnormal feature matrix, calculate its local correlation value with surrounding monitoring points and its global correlation value with all monitoring points; construct an enhanced feature matrix based on a weighted combination of the local correlation value and the global correlation value;

[0033] Step S14: Based on the enhanced feature matrix, a dynamic threshold of each monitoring point is calculated, and the threshold changes as the distance from the reference point increases; the monitoring points exceeding the corresponding dynamic threshold are recorded to form a suspicious area coordinate set;

[0034] Step S15: for each area in the suspicious area coordinate set, respectively calculate its displacement anomaly degree, stress anomaly degree and temperature anomaly degree; combine the displacement anomaly degree, stress anomaly degree and temperature anomaly degree through adaptive weights to obtain a credibility score; based on the credibility score, sort the suspicious area coordinate set to generate a prioritized suspicious area list.

[0035] The process of combining these three abnormal degrees through adaptive weights to obtain the credibility score is as follows: Credibility multi-index fusion calculation: C(r) =λ1 P1(r) +λ2 P2(r) +λ3 P3(r) [1 +α S(r)]; where C(r) is the credibility score; P1, P2, P3 are displacement, stress, and temperature indicators; λ1, λ2, λ3 are dynamic weights; S(r) is spatial correlation; α is the correlation coefficient (0.2~0.4). λi = exp(-|Pi-μi| / σi) [1 + β Ti]; where μi is the historical mean; σi is the standard deviation; Ti is the time stability; β is the stability weight (0.1~0.3).

[0036] This embodiment obtains three types of monitoring data, namely displacement, stress and temperature, in real time, and combines time synchronization processing and adaptive weight calculation methods to solve the technical problems of strong heterogeneity of dam monitoring data, inconsistent sampling frequency, and uneven data quality. Specifically, by using linear interpolation for time synchronization, data with different sampling frequencies can be aligned to a unified time series, avoiding misjudgment caused by data time series misalignment; by calculating the adaptive weight coefficients of the core data part and the peripheral data part, the influence of the main trend and fluctuation changes is effectively balanced, and the accuracy of anomaly detection is improved; by calculating the distance correlation weights between monitoring points, the influence of spatial position relationships on abnormal features is fully considered, so that the abnormal feature matrix finally generated not only reflects the changing laws of the time dimension, but also contains the correlation characteristics of the spatial dimension. This embodiment improves the accuracy of suspicious area identification, reduces the false alarm rate, and can give a reliable credibility score, providing effective priority guidance for subsequent precise detection.

[0037] In one embodiment of the present application, an abnormal feature enhancement function can also be used to calculate and construct an enhanced feature value, and form an enhanced feature matrix: F(x, y) = Fb(x, y) [1 +α L(x, y)] [1 +β G(x, y)]; wherein F(x, y) is an enhanced feature; Fb(x, y) is a basic feature; L(x, y) is a local abnormality, that is, a local correlation value; G(x, y) is a global abnormality, a global correlation value; α, β are weight coefficients (0~0.5). L(x, y)=Σ[wi |f(x, y)-f(xi, yi)| / d(i)]; wherein f(x, y) is a feature value; d(i) is a spatial distance; wi is a distance weight; xi, yi are neighborhood point coordinates.

[0038] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0039] Step S21, based on the priority and spatial coordinates of each suspicious area in the suspicious area list, calculate the horizontal scanning angle and the vertical scanning angle to generate a sonar scanning path; control the sonar device to collect the echo signal of the target area according to the sonar scanning path to obtain the original sonar data including the distance value and the signal strength;

[0040] Step S22, based on the original sonar data, calculating the signal mean and variance of the local area; based on the mean, determining the signal gain coefficient; based on the variance, determining the noise suppression coefficient; multiplying the signal gain coefficient with the original sonar data to obtain gain sonar data; multiplying the noise suppression coefficient with the gain sonar data to obtain enhanced sonar data;

[0041] Step S23, based on the enhanced sonar data, calculate the gradient values ​​in the horizontal direction and the vertical direction; based on the gradient values, calculate the edge strength and direction angle, and determine the adaptive threshold in combination with the statistical characteristics of the local area; record the edge points above the adaptive threshold as edge feature data;

[0042] The precise edge feature extraction is as follows: G(x, y) = sqrt(Gx 2 + Gy 2 ) [1 +α D(x,y)]; where G(x,y) is the edge strength; Gx, Gy are the horizontal and vertical gradients; D(x,y) is the direction consistency; α is the direction weight (0.2~0.4). E(x,y) = G(x,y) exp(-β |θ(x,y)-θm|); where E(x,y) is the edge feature; θ(x,y) is the local direction; θm is the main direction; β is the direction weight (0.3~0.6).

[0043] Step S24, based on the edge feature data, calculating the intensity feature, shape feature and texture feature of each edge point; matching the intensity feature, shape feature and texture feature with a preset defect feature template to obtain a feature matching probability; marking an area where the feature matching probability is higher than a preset threshold as a precise defect area;

[0044] The specific process of feature matching is as follows: Calculate the feature matching probability: P(f, t) = exp(-||ft|| 2 / σ1 2 ) [1 + α S(f, t)]; where P(f, t) is the matching probability; f is the feature to be matched; t is the template feature; S(f, t) is the structural similarity; α is the similarity weight (0.2~0.4); σ1 is the scale parameter. M(i) =Σ[wj P(fi, tj)] [1 +β C(i)]; where M(i) is the comprehensive matching degree; wj is the feature weight; C(i) is the context consistency; β is the context weight (0.1~0.3).

[0045] Step S25: Based on the precise defect area, the geometric dimensions, edge clarity, echo intensity and spatial continuity of the area are calculated respectively; the geometric dimensions, edge clarity, echo intensity and spatial continuity are combined to form a defect feature description, and the reliability index of the feature is calculated at the same time to generate a defect detection report including the defect location, characteristics and reliability.

[0046] This implementation performs sonar scanning path planning and data collection and processing based on a list of suspicious areas, solving the technical problems of complex underwater environments, susceptibility to interference of sonar signals, and low detection efficiency. By introducing environmental parameters such as water depth, temperature, and salinity to calculate the sound velocity profile, and optimizing the transmission power and receiving gain in combination with the target distance and water turbidity, the adaptive adjustment of sonar parameters is achieved, and the signal-to-noise ratio of the echo signal is improved; the improved ant colony algorithm is used to optimize the detection sequence, and the scanning density is dynamically adjusted according to the regional priority, which not only ensures the detection accuracy of key areas, but also improves the overall detection efficiency; in the signal processing link, the influence of platform shaking and sound velocity changes is eliminated through attitude compensation and sound path correction, and the gain adjustment is performed using overlapping sliding windows and piecewise linear mapping, which effectively improves the quality and detail performance of the sonar image. This embodiment enables clear and reliable sonar data to be obtained even in complex underwater environments, providing high-quality basic data for defect identification.

[0047] According to one aspect of the present application, step S21 is further:

[0048] Step S211, read the spatial coordinate information of the suspicious area list, and based on the spatial coordinate information, use an adaptive grid decomposition algorithm to divide the detection area into predetermined sub-areas to generate grid division data; read the current water flow data and obstacle data from the pre-stored environmental database, calculate the detection difficulty coefficient of each sub-area in combination with the grid division data, and output the difficulty coefficient matrix; perform weighted combination of the priority information in the suspicious area list and the difficulty coefficient matrix to generate a regional priority matrix; based on the regional priority matrix, use an improved ant colony algorithm for optimization to output an initial scanning path;

[0049] The adaptive grid decomposition algorithm is as follows: C(i, j) = Cb [1 +α D(i, j)] [1+β O(i, j)][1 + γ P(i, j)]; where: C(i, j) is the grid size; Cb is the basic grid size; D(i, j) is the depth change rate; O(i, j) is the occlusion; P(i, j) is the priority coefficient; α, β, γ are adjustment factors (0~1). Q(i, j) = exp(-λ1 V(i, j)) exp(-λ2 R(i, j)); where Q(i, j) is the grid quality score; V(i, j) is the visibility index; R(i, j) is the resolution requirement; λ1, λ2 are weight coefficients.

[0050] The improved ant colony algorithm (S211) is as follows: τij(t+1) = (1-ρ) τij(t) + Δτij(t) [1 + α Q(i, j)]; where τij(t) is the pheromone concentration; ρ is the volatility coefficient (0.10.3); Δτij(t) is the pheromone increment; Q(i, j) is the path quality; α is the quality weight (0.20.4). P(i, j) = [τij(t)] α [ηij(t)] β [1 +γ V(i,j)] / Σ[τik(t)] α [ηik(t)] β ; Where P(i, j) is the transition probability; ηij(t) is the heuristic information; V(i, j) is the visibility adjustment term; α, β are the weights of pheromone and heuristic information; γ is the adjustment coefficient (0.1~0.3). E(i, j) = w1 L(i, j) + w2D(i, j) + w3 C(i, j); Where E(i, j) is the path evaluation function; L(i, j) is the path length; D(i, j) is the detection difficulty; C(i, j) is the coverage efficiency; w1, w2, w3 are the evaluation weights. K(t) = Kb [1+exp(-λ t)] [1 + μ S(t)]; Where K(t) is the convergence control factor; Kb is the basic convergence coefficient; S(t) is the search stagnation; λ is the time scale parameter; μ is the stagnation weight (0.1~0.3).

[0051] Step S212: Generate sound velocity profile data based on the initial scanning path and pre-stored water depth data; calculate the attenuation coefficient of the sonar beam at different depths based on the sound velocity profile data, and generate an attenuation coefficient table; calculate the optimal transmission power and receiving gain from the target area distance data based on the attenuation coefficient table; perform parameter optimization based on the optimal transmission power and receiving gain in combination with the pre-stored turbidity data, and generate a parameter configuration table;

[0052] Step S213: Based on the parameter configuration table and the regional priority matrix, the optimal pitch angle and horizontal rotation angle of each detection position are calculated to generate a multi-angle scanning sequence; the sonar rotation mechanism is controlled to move according to the multi-angle scanning sequence to collect echo signal data in real time;

[0053] Step S214, acquiring attitude data from the inertial navigation system, and based on the attitude data, performing attitude compensation on the echo signal data to generate compensated signal data; extracting sound path correction parameters from the sound velocity profile data, performing sound path correction on the compensated signal data to generate corrected signal data; performing time-series alignment and spatial splicing on all corrected signal data to output original sonar data.

[0054] This embodiment solves the technical problems of strong blindness and low efficiency of underwater detection through multi-level sonar scanning path planning and parameter optimization. The detection area is divided by an adaptive grid decomposition algorithm, and the detection difficulty coefficient is calculated in combination with water flow and obstacle data, so that the path planning is more in line with the actual working conditions; by analyzing the sound velocity profile and attenuation characteristics, an adaptive adjustment mechanism for sonar parameters is established, and the transmission power and receiving gain are dynamically optimized according to the target distance and water turbidity, thereby improving the quality of the sonar signal; based on the improved ant colony algorithm, the detection sequence is optimized, and the scanning density is adjusted through the regional priority matrix to achieve a reasonable allocation of detection resources; through attitude compensation and sound range correction, the influence of platform movement and sound velocity changes is eliminated, ensuring the accuracy of sonar data. This embodiment not only improves the detection efficiency, but also ensures the data quality, laying a solid foundation for subsequent defect identification.

[0055] In another embodiment of the present application, in step S211, the following method can also be used to optimize the scanning path and obtain the initial scanning path: R(θ, φ) = Rb [1 +α D(θ, φ)] [1+β O(θ, φ)]; where R(θ, φ) is the scanning radius; Rb is the basic scanning radius; D(θ, φ) is the depth change rate; O(θ, φ) is the occlusion degree; α, β are adjustment coefficients (0~0.5). E(p) = w1 L(p) + w2 T(p) + w3 C(p); where E(p) is the path evaluation value; L(p) is the path length; T(p) is the time consumption; C(p) is the coverage degree; w1, w2, w3 are weight coefficients.

[0056] According to one aspect of the present application, step S22 is further:

[0057] Step S221, dividing the original sonar data into overlapping local analysis windows, calculating the signal mean, variance and peak distribution in each local analysis window, and generating a statistical feature matrix; based on the statistical feature matrix, using an adaptive window algorithm to optimize the window size, and outputting optimized statistical data;

[0058] The adaptive window algorithm is as follows: W(x, y) = Wb [1 +α E(x, y)] [1+β N(x, y)]; where W(x, y) is the window size; Wb is the basic window size; E(x, y) is the edge intensity factor; N(x, y) is the noise level; α is the edge adjustment coefficient (0.30.7); β is the noise suppression coefficient (0.20.5). H(x, y) = exp(-|G(x, y)| / σ1) exp(-|L(x, y)| / σ2); where H(x, y) is the window uniformity; G(x, y) is the grayscale gradient; L(x, y) is the local variance; σ1, σ2 are scale parameters.

[0059] Step S222: based on the optimized statistical data, the signal dynamic range of each window is calculated to generate a dynamic range vector; based on the dynamic range vector, a piecewise linear mapping function is constructed to generate a gain mapping table; the gain control parameter of each signal segment is extracted from the gain mapping table, the original sonar data is segmentedly gain adjusted according to the gain control parameter, and the gain adjustment data is output;

[0060] Step S223, performing time-frequency analysis on the gain adjustment data to extract the frequency characteristics and energy distribution of the noise; based on the frequency characteristics and energy distribution, constructing a multi-layer noise model to identify system noise, environmental noise and interference noise, and generating a noise feature vector and a noise suppression coefficient;

[0061] Step S224: Based on the noise feature vector and the noise suppression coefficient, an adaptive filter group is constructed to process the signal in frequency bands to obtain processed data; based on the processed data, a wavelet threshold method is used to eliminate noise, and the signal is reconstructed to obtain enhanced sonar data.

[0062] The specific wavelet threshold method is: T(j, k) =σj [log(N) / N]α [1 +β |W(j, k)| / σj]; where T(j, k) is the adaptive threshold of the kth wavelet coefficient of the jth layer; σj is the standard deviation of the noise in the jth layer; N is the signal length; W(j, k) is the wavelet coefficient; α is the scale adjustment factor (0.30.7); β is the coefficient adjustment factor (0.10.5).

[0063] This embodiment solves the technical problems of poor quality and unclear detail features of underwater sonar signals through multi-level signal enhancement and adaptive parameter optimization. Overlapping sliding windows are used for local analysis, and the optimal window size is automatically determined by calculating statistical features, avoiding blind spots and redundancy caused by fixed windows; a piecewise linear mapping method based on the dynamic range of the signal is designed, and the gain control parameters are adaptively calculated to enhance the detail performance of weak signal areas and avoid saturation distortion in strong signal areas; time-frequency analysis is introduced to identify different types of noise features, and wavelet threshold method is used for hierarchical noise suppression, which retains effective signals while reducing background noise; multi-layer filter groups are used to realize signal frequency band processing, and adaptive weighted reconstruction is performed according to the feature importance of different frequency bands, thereby improving the signal-to-noise ratio and resolution. This embodiment not only improves the quality and clarity of sonar images, but also maintains the authenticity and integrity of defect features, providing a high-quality data foundation for subsequent feature extraction.

[0064] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0065] Step S31: Based on the precise defect area and defect feature description, control the acoustic sensor to collect acoustic raw data of the target area at different angles, and control the underwater camera to collect optical raw data at the corresponding position, and record the collection time and spatial coordinates of each data; based on the collection time and spatial coordinates, align the acoustic raw data and the optical raw data to generate registration data;

[0066] The process of collecting optical and acoustic raw data is as follows: according to the spatial coordinates of the precise defect area, the water quality parameter data and water flow velocity data of the current water area are read from the database, and the detection space is divided into working sub-areas using the adaptive grid method. The water quality parameter data and water flow velocity data are input into the obstacle avoidance evaluation model, the difficulty coefficient of each sub-area is calculated, and the regional difficulty matrix is ​​generated. According to the regional difficulty matrix, the dynamic obstacle avoidance path planning algorithm is used to generate the initial path of the equipment, and the initial path of the equipment is input into the energy consumption optimization model to generate the equipment operation path and equipment control parameters. According to the equipment operation path, the acoustic sensor is controlled to collect the acoustic raw data of the target area at different angles, and the underwater camera is controlled to collect optical raw data at the corresponding position according to the equipment control parameters.

[0067] The process of aligning the acoustic raw data and the optical raw data based on the acquisition time and spatial coordinates is as follows: constructing a spatiotemporal registration matrix: T(p, q) = Tb [1 +α D(p, q)] [1+β V(p, q)]; where T(p, q) is the spatiotemporal transformation matrix; Tb is the basic transformation matrix; D(p, q) is the spatial displacement; V(p, q) is the velocity influence; α, β are adjustment factors (0~0.5). R(p, q) = exp(-|gradT(p, q)| / σ1) exp(-|H(p, q)| / σ2); where R(p, q) is the registration reliability; gradT is the transformation gradient; H(p, q) is the registration entropy; σ1, σ2 are control parameters.

[0068] Step S32, reading the acoustic part of the registration data, calculating the high-frequency component and the low-frequency component; based on the high-frequency component and the low-frequency component, enhancing them respectively, and performing weighted combination to obtain enhanced acoustic data; reading the optical part of the registration data, enhancing the contrast and texture features respectively, and performing weighted combination to obtain enhanced optical data; calculating the complementary enhancement coefficient based on the preset quality evaluation index, combining the enhanced acoustic data and the enhanced optical data to obtain enhanced feature data;

[0069] Step S33: based on the enhanced feature data, extracting spectrum features, time domain features and energy features from the enhanced acoustic data to generate acoustic feature data; extracting edge features, texture features and shape features from the enhanced optical data to generate optical feature data; calculating the similarity between the acoustic feature data and the optical feature data to generate a feature mutual verification matrix;

[0070] Step S34: based on the acoustic feature data and the optical feature data, generate a quality assessment result, calculate feature weights, and obtain weighted features; fuse the weighted features to obtain multimodal feature data; based on the multimodal feature data, screen according to preset feature importance to generate a fusion feature matrix;

[0071] The process of calculating feature weights and fusing them includes: calculating feature fusion weights: W(i, j) = Wb(i, j)[1 +α Q(i, j)] [1+β D(i, j)]; where W(i, j) is the fusion weight; Wb(i, j) is the basic weight; Q(i, j) is the feature quality; D(i, j) is the feature redundancy; α, β are adjustment coefficients (0~0.5). Q(i, j) = exp(-|F(i, j)-μ| / σ1)exp(-|gradF(i, j)| / σ2); where F(i, j) is the feature value; μ is the mean; gradF is the feature gradient; σ1, σ2 are scale parameters.

[0072] Step S35, calculate the geometric size, physical characteristics, dynamic response and spatial distribution of each feature in the fused feature matrix to generate a feature description vector; based on the feature description vector, calculate the credibility of the acoustic feature data, optical feature data and feature mutual verification matrix respectively, and obtain the feature credibility index by weighted combination.

[0073] This embodiment solves the technical problems of limited information and poor anti-interference ability of a single detection method through multimodal data collection and fusion processing. Path planning is performed based on water quality parameters and water flow velocity, and a dynamic obstacle avoidance algorithm is used to generate the optimal operation path, thereby improving the efficiency and safety of data collection. In the data collection stage, the acoustic sensor and low-light camera are collaboratively controlled, and the collection parameters are dynamically adjusted according to environmental conditions to ensure the quality of acoustic and optical data. In terms of data processing, a spatiotemporal registration algorithm is used to achieve precise alignment of heterogeneous data, and the reliability weight is calculated through a quality assessment model to ensure the accuracy of data fusion. In the feature extraction stage, the feature performance of acoustic and optical data is enhanced respectively, and feature fusion is performed through an adaptive weight method, giving full play to the complementary advantages of the two modal data. This embodiment not only improves the reliability of detection, but also enhances the environmental adaptability of the system.

[0074] According to one aspect of the present application, step S31 is further:

[0075] Step S311, read the spatial coordinates of the precise defect area, and divide the detection space into working sub-areas using an adaptive grid method; calculate the passage difficulty coefficient based on the working sub-areas, the pre-stored water flow velocity and the obstacle distribution; based on the passage difficulty coefficient, use a dynamic obstacle avoidance path planning algorithm to generate an initial path for the submersible; based on the initial path of the submersible, use a pre-configured energy consumption model for optimization to obtain an operation path for the submersible;

[0076] The adaptive grid method is as follows: S(i, j) = Sb [1 +α D(i, j)] [1 +β V(i, j)]; where S(i, j) is the grid size; Sb is the basic grid size; D(i, j) is the depth change rate; V(i, j) is the water velocity influencing factor; α, β are adjustment coefficients (0~1). Q(i, j) = exp(-γ C(i, j)) exp(-η O(i, j)); where Q(i, j) is the grid quality score; C(i, j) is the complexity index; O(i, j) is the obstacle density; γ, η are weight coefficients.

[0077] Step S312: Based on the operation path of the submersible and the defect feature description, the optimal working distance of the acoustic sensor and the low-light camera is calculated, and an equipment layout plan is generated; based on the equipment layout plan, a multi-device collaborative control algorithm is used to calculate the working sequence of each sensor and output collaborative acquisition instructions;

[0078] Step S313, read the real-time water quality parameters and ambient light intensity, dynamically adjust the exposure time, gain value and color enhancement parameters of the low-light camera, and generate a camera parameter table; based on the camera parameter table, obtain the reflection characteristics of the target area, calculate the optimal fill light angle and intensity, and output the fill light control instruction;

[0079] Step S314: based on the collaborative acquisition instruction and the fill light control instruction, receiving the original acoustic data and original image data collected by each sensor, extracting the timestamp and spatial position mark of the acoustic data, extracting the acquisition time and camera posture information of the image data; based on the timestamp, spatial position mark, acquisition time and camera posture information, using the time-space registration algorithm to generate a time-space correspondence table;

[0080] The specific spatiotemporal registration algorithm is: E(R, t) = ws Es(R, t) + wt Et(R, t) + wc Ec(R, t); where Es(R, t) is the spatial registration error term; Et(R, t) is the temporal synchronization error term; Ec(R, t) is the feature consistency error term; ws, wt, wc are weight coefficients and ws+wt+wc=1. Es(R, t) = Σ||ps(i)-(R pt(i)+t)|| 2 exp(-d(i) / ds); where ps(i), pt(i) are the spatial coordinates of the corresponding feature points; R is the rotation matrix; t is the translation vector; d(i) is the distance between feature points; and ds is the spatial scale parameter.

[0081] Step S315: Based on the time-space correspondence table, the signal-to-noise ratio and clarity of the acoustic data are calculated to obtain the acoustic quality index; based on the time-space correspondence table, the clarity and contrast of the image data are calculated to obtain the image quality index; based on the acoustic quality index and the image quality index, a multi-index fusion evaluation method is used to generate the registration data and data quality report.

[0082] In another embodiment of the present application, the acoustic raw data and optical raw data collected by each sensor are received, and the acoustic wave emission time, reception time and device posture data are read from the acoustic sensor to generate an acoustic space-time tag. The image acquisition time, camera posture and GPS location information are read from the underwater camera to generate an optical space-time tag. The acoustic space-time tag and the optical space-time tag are input into the space-time registration model, and the spatial mapping relationship and time correspondence relationship of the two data are calculated to generate a registration mapping matrix. The acoustic raw data and the optical raw data are aligned in time and space according to the registration mapping matrix, and the space-time alignment data is output.

[0083] For the acoustic part of the spatiotemporal alignment data, the peak ratio, signal-to-noise ratio and spectrum clarity of the signal are calculated to generate an acoustic quality vector. For the optical part of the spatiotemporal alignment data, the clarity, contrast and illumination uniformity of the image are calculated to generate an optical quality vector. The acoustic quality vector and the optical quality vector are input into the quality assessment model, the reliability weight of each data is calculated, and a data weight matrix is ​​generated. The spatiotemporal alignment data is weightedly fused according to the data weight matrix, and the registration data and the registration quality report are output.

[0084] This embodiment solves the technical problems of difficult underwater multimodal data collection and unstable quality through precise equipment layout and collaborative control. An adaptive grid division method based on water quality parameters and obstacle distribution is proposed, and the path of the submersible is optimized by calculating the passage difficulty coefficient, thereby improving operation efficiency and safety. The optimal working distance calculation model for acoustic sensors and low-light cameras is designed, and the precise layout of the acquisition equipment is realized in combination with the multi-device collaborative control algorithm, ensuring the spatial coverage and time synchronization of data acquisition. The exposure parameters and fill light strategy of the camera are dynamically adjusted according to the real-time water quality parameters and ambient light intensity, effectively improving the image quality. By analyzing the signal-to-noise ratio of acoustic data and the clarity of image data, a real-time evaluation mechanism for data quality is established to ensure the availability of the collected data. This embodiment not only ensures the integrity and synchronization of the data, but also improves the reliability and efficiency of the acquisition process.

[0085] According to one aspect of the present application, step S32 is further:

[0086] Step S321, performing multi-scale wavelet decomposition on the acoustic signal in the registration data, extracting characteristic components of different frequency bands, calculating the energy distribution of each characteristic component, and generating a frequency domain characteristic matrix; based on the frequency domain characteristic matrix, analyzing the frequency band correlation, and obtaining frequency domain correlation data; wherein the characteristic components of different frequency bands include high frequency components, medium frequency components, and low frequency components;

[0087] Step S322: based on the frequency domain feature matrix and the frequency domain correlation data, the high frequency components are directionally enhanced, the intermediate frequency components are selectively retained, the low frequency components are noise suppressed, and the processed components of each frequency band are integrated to generate enhanced acoustic data;

[0088] Step S323, performing geometric correction and illumination balance on the optical image in the registration data, constructing an adaptive histogram model, compensating for uneven illumination effects, and generating corrected image data;

[0089] Step S324, applying a multi-scale contrast enhancement algorithm to the corrected image data, adaptively adjusting enhancement parameters in combination with statistical features of the local area, optimizing detail and texture performance, and outputting enhanced optical data;

[0090] The multi-scale contrast enhancement algorithm is as follows: E(x, y) = I(x, y) +λ Σ[wk Dk(x, y)]; where E(x, y) is the enhanced image; I(x, y) is the original image; Dk(x, y) is the detail image of the kth scale; wk is the scale weight coefficient; λ is the enhancement intensity factor (0.5~2.0). Dk(x, y) = Gk(x, y)-Gk+1(x, y); where: Gk(x, y) is the kth layer Gaussian pyramid image; wk = exp(-|Dk(x, y)| / σk); σk is the kth layer scale parameter.

[0091] Or, E(x, y) = I(x, y) [1 + α C(x, y)] [1 + β T(x, y)]; where E(x, y) is the enhancement result; I(x, y) is the original image; C(x, y) is the contrast term; T(x, y) is the texture enhancement term; α, β are enhancement coefficients (0.3~0.6). T(x, y) = Σ[wi Li(x, y)] exp(-γ N(x, y)); where Li(x, y) is the Laplacian feature of the i-th layer; N(x, y) is the noise evaluation; wi is the layer weight; γ is the noise suppression coefficient (0.2~0.4).

[0092] Step S325, calculating the complementary features of the enhanced acoustic data and the enhanced optical data, and constructing feature fusion weights; based on the feature fusion weights, using an adaptive weighting method to combine the enhanced acoustic data and the enhanced optical data to generate enhanced feature data.

[0093] The specific process of generating enhanced feature data is as follows: M(i, j) =λ1 F1(i, j) +λ2 F2(i, j) +θ I(F1, F2); where M(i, j) is the fusion feature; F1 and F2 are two features; λ1 and λ2 are feature weights; I(F1, F2) is the mutual information term; θ is the mutual information weight (0.1~0.3). I(F1, F2) =Σ[P(x, y) log(P(x, y) / P(x)P(y))] [1 + α C(x, y)]; where P(x, y) is the joint probability; P(x) and P(y) are the edge probabilities; C(x, y) is the local correlation; α is the correlation coefficient (0.2~0.4).

[0094] This implementation solves the technical problems of underwater heterogeneous data features that are not prominent and have poor complementarity through independent enhancement and collaborative fusion of acoustic and optical data. Multi-scale wavelet decomposition is used to extract the frequency domain features of acoustic signals, and differentiated enhancement of different frequency bands is achieved by analyzing the frequency band correlation, thereby increasing the amount of information in acoustic data; an adaptive geometric correction and illumination equalization algorithm is designed to effectively compensate for the impact of uneven illumination and improve the contrast and detail performance of the image; a multi-scale contrast enhancement algorithm is introduced to adaptively adjust the enhancement parameters according to local area features, thereby optimizing the texture performance of the image; adaptive feature fusion weights are designed by analyzing the complementary features of acoustic and optical data, thereby maintaining their respective advantageous features and achieving effective information complementarity. This embodiment improves the expressiveness and distinguishability of features, and provides high-quality data support for subsequent feature fusion.

[0095] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0096] Step S41: Based on the feature credibility index, the fused feature matrix is ​​compared and calculated with the pre-stored structural similarity data, time series correlation data and spatial position correlation data to obtain a feature discrimination matrix; based on the feature weights in step S34, the feature discrimination matrix is ​​weighted and calculated to output the defect discrimination result;

[0097] Step S42: based on the defect identification result, respectively calculating the matching degree with the preset crack feature library, erosion feature library, leakage feature library and deformation feature library to obtain feature matching data; based on the feature matching data and the defect identification result, determining the defect type and classifying the defect level;

[0098] Step S43: Based on the defect type and defect level, combined with the pre-stored sonar positioning data, optical positioning data and monitoring positioning data, calculate the spatial positioning data according to their respective reliability weights; based on the spatial positioning data, perform positioning error analysis to generate error evaluation data; based on the error evaluation data, determine the final positioning result and positioning confidence interval;

[0099] Step S44, calculating the recognition accuracy index of the defect discrimination result, the spatial precision index of the positioning result, the characteristic time stability index and the spatial consistency index, and generating evaluation index data; based on the evaluation index data, obtaining a comprehensive reliability score through weighted combination to form a reliability evaluation result;

[0100] Step S45: Based on the reliability assessment results, defect type, defect level and pre-stored historical evolution trend data, calculate the warning index value and generate warning level data; based on the warning level data, update the key monitoring point list, monitoring parameter set and warning threshold to form a warning parameter update list; based on the warning parameter update list, generate a defect identification report.

[0101] This embodiment solves the technical problems of low underwater defect recognition accuracy, poor positioning accuracy, and difficult reliability evaluation through multi-level feature analysis and comprehensive evaluation. In the feature discrimination link, the fused feature matrix is ​​compared with the pre-stored structural similarity, time series correlation, and spatial position association data in multiple dimensions, and the robustness of defect discrimination is improved through adaptive adjustment of feature weights; in the defect classification process, multi-feature library matching and multi-level fuzzy evaluation methods are adopted, which can not only accurately identify the defect type, but also give a reasonable grade classification; in terms of spatial positioning, the positioning accuracy is improved through weighted fusion and error compensation of multi-source positioning data, and a reliable positioning confidence interval can be given; in the reliability evaluation link, a complete evaluation system is established by analyzing multi-dimensional indicators such as recognition accuracy, spatial accuracy, temporal stability, and spatial consistency, providing a quantitative basis for the credibility of the detection results; at the same time, the early warning parameters are dynamically updated based on the evaluation results, realizing the adaptive optimization of the detection system. This embodiment realizes the accurate identification and reliable positioning of underwater defects, and establishes a scientific reliability evaluation mechanism.

[0102] According to one aspect of the present application, step S42 is further:

[0103] Step S421, read edge feature data from the defect discrimination result, extract the crack area using the adaptive threshold segmentation algorithm based on regional growth, and generate crack contour data. Input the crack contour data into the improved Hough transform model, calculate the main direction and branch direction of the crack, and generate crack direction data. Apply the sub-pixel edge positioning algorithm to the crack contour data, calculate the precise coordinates of each contour point, and generate contour accuracy data. Extract the depth feature from the fusion feature matrix output by S3, calculate the crack width and length in combination with the contour accuracy data, and output the crack size data.

[0104] The improved Hough transform is specifically: A(ρ, θ) = Σ w(i) Δ(ρ-xi cos(θ)-yi sin(θ)); where w(i) = exp(-|gradI(i)| / σ1) exp(-|κ(i)| / σ2); gradI(i) is the image gradient intensity; κ(i) is the local curvature; σ1, σ2 are scale parameters; xi, yi are the coordinates of the edge points.

[0105] Step S422: Apply the region growing algorithm to the fused feature matrix to divide the damaged area, extract the boundary curve of each damaged area, use the improved area calculation model to obtain the damaged area data, and calculate the damaged depth data based on the image grayscale gradient analysis.

[0106] The specific region growing algorithm is: T(x, y) = μ(x, y) + α σ(x, y) [1 + β g(x, y)]; where T(x, y) is the local adaptive threshold; μ(x, y) is the local region mean; σ(x, y) is the local region standard deviation; g(x, y) is the local gradient strength; α is the mean adjustment coefficient (1.02.0); β is the gradient influence factor (0.20.8). If |I(x, y)-I(s, t)|< T(x, y) and C(x, y) > Tc, then R(x, y) = 1; otherwise R(x, y) = 0; where R(x, y) is the region growing result; I(x, y) is the target pixel value; I(s, t) is the seed point pixel value; C(x, y) is the connectivity constraint; Tc is the connectivity threshold.

[0107] The improved area calculation model is as follows: A(R) = Σ[w(i, j) a(i, j)] [1 + α K(i, j)]; where A(R) is the area of ​​the region; w(i, j) is the pixel weight; a(i, j) is the pixel area; K(i, j) is the curvature correction term; α is the curvature influence coefficient (0.1~0.4). w(i, j) = exp(-|gradI(i, j)| / σ1) exp(-|θ(i, j)| / σ2); where gradI(i, j) is the gradient strength; θ(i, j) is the direction difference; σ1, σ2 are adjustment parameters.

[0108] Step S423: Combine the crack width data, crack length data and damage depth data to establish a three-dimensional defect reconstruction model, calculate the defect volume to obtain defect volume data, and analyze the defect extension trend to generate a defect development prediction.

[0109] The process of establishing a defect 3D reconstruction model can be as follows: Perform 3D defect reconstruction calculation: V(x, y, z) = Vb(x, y, z) [1 + α H(x, y, z)] [1 + β S(x, y, z)]; where V(x, y, z) is the reconstructed voxel value; Vb(x, y, z) is the basic voxel value; H(x, y, z) is the depth feature; S(x, y, z) is the surface feature; α, β are feature weights (0.2~0.4). R(x, y, z) = exp(-|gradV(x, y, z)| / σ1) exp(-|K(x, y, z)| / σ2); where R(x, y, z) is the reconstruction reliability; grad V is the voxel gradient; K(x, y, z) is the local curvature; σ1, σ2 are control parameters.

[0110] Step S424: match the defect volume data and damage area data with the pre-stored feature template library, use a multi-feature fusion classification method to identify the defect type, and output the defect type and type credibility.

[0111] Step S425: Based on the defect type, defect volume data and pre-stored rating standards, the geometric feature damage index and the structural impact index are calculated, and the defect grade and grade credibility are determined using a multi-level fuzzy evaluation method.

[0112] This embodiment solves the technical problems of low accuracy in underwater defect type identification and unreasonable grading through multi-feature library matching and multi-level fuzzy evaluation. Adaptive threshold segmentation algorithm is used to extract crack contours, and the improved Hough transform is combined to achieve accurate calculation of crack direction, improving the accuracy of feature extraction; sub-pixel edge positioning technology and depth information fusion are used to achieve high-precision measurement of crack width and length, providing a reliable quantitative basis for defect evaluation; regional growth algorithm is used to accurately divide the damaged area, combined with improved area calculation model and grayscale gradient analysis, to achieve accurate measurement of damaged area and depth; based on defect three-dimensional reconstruction technology, not only the defect volume is calculated, but also the extension trend of the defect is analyzed, providing a scientific basis for predicting defect development; defect identification is performed through multi-feature fusion classification method, and defect level is determined in combination with multi-level fuzzy evaluation, which not only ensures the accuracy of classification, but also achieves the rationality of grading. This embodiment improves the accuracy and credibility of defect identification.

[0113] According to one aspect of the present application, step S43 is further:

[0114] Step S431, receive real-time sonar positioning data from the sonar device, receive optical positioning data from the optical system, and read the structural monitoring data at the corresponding time from the monitoring system database. Input these three types of data into the time synchronization model, align them according to their respective timestamps, and generate synchronized positioning data. For each data source in the synchronized positioning data, calculate the reliability weight according to its acquisition conditions and historical stability, and generate a positioning weight matrix. Perform weighted fusion on the synchronized positioning data according to the positioning weight matrix, and output the initial fusion position.

[0115] Step S432: Analyze the systematic error and random error in the synchronous positioning data, establish an error compensation model, use an adaptive Kalman filter algorithm to correct the initial fusion position, and output the compensated position.

[0116] Step S433: convert the compensated position from the equipment coordinate system to the dam structure coordinate system, consider the posture error in the coordinate transformation process, adopt the multi-source data constraint method to improve the conversion accuracy, and generate the structure coordinate data.

[0117] Step S434: Calculate the positioning accuracy evaluation index for the structural coordinate data, establish a three-dimensional error ellipsoid model, analyze the positioning error distribution in each direction, and output the positioning result and the corresponding positioning confidence interval.

[0118] This embodiment solves the technical problems of low accuracy and poor reliability in underwater defect positioning through multi-source positioning data fusion and error compensation. The timestamp alignment method is used to achieve the synchronization of multi-source positioning data, and the accuracy of data fusion is guaranteed by reliability weight calculation; an analysis model of systematic error and random error is designed, and the fusion position is dynamically corrected in combination with an adaptive Kalman filter algorithm to improve positioning accuracy; in the coordinate conversion process, the influence of attitude error is considered, and the accuracy of coordinate transformation is improved through multi-source data constraints; a three-dimensional error ellipsoid model is established to analyze the positioning error distribution, which not only gives the positioning result, but also provides a reliable confidence interval. This embodiment not only improves the accuracy of defect spatial positioning, but also quantifies the uncertainty of the positioning result, providing precise spatial guidance for defect repair.

[0119] According to one aspect of the present application, step S44 is further:

[0120] Step S441: Extract key indicators such as recognition accuracy, feature integrity, and boundary clarity from the defect discrimination results to generate a recognition quality vector. Extract key indicators such as spatial accuracy, repeatability, and stability from the positioning results to generate a positioning quality vector. Normalize the recognition quality vector and the positioning quality vector according to the preset evaluation rules and output a multi-dimensional evaluation index.

[0121] Step S442: read the historical evaluation indicators in the historical database, compare the current multi-dimensional evaluation indicators with the historical evaluation indicators in time series, calculate the short-term volatility indicators using the sliding window method, and generate volatility characteristic data. Apply the trend analysis algorithm to the volatility characteristic data, calculate the change rate and acceleration of each indicator, and generate change characteristic data. Input the volatility characteristic data and the change characteristic data into the time stability evaluation model, and output the time stability data.

[0122] The sliding window method is as follows: F(t) =Σ[w(k) x(tk)] [1 +α D(t)]; where F(t) is the filter output; w(k) is the window weight; x(t) is the input sequence; D(t) is the dynamic adjustment term; α is the adaptive coefficient (0.1~0.3). w(k) = exp(-k 2 / 2σ 2 ) [1 + β E(k)]; k is the time delay; σ is the window width; E(k) is the edge preservation factor; β is the edge weight (0.2~0.5).

[0123] The specific trend analysis algorithm is: P(t) = a exp(-λ t) sin(ωt + φ) + bt + c; where P(t) is the predicted trend value; t is the time variable; a is the amplitude coefficient; λ is the attenuation coefficient; ω is the period factor; φ is the phase angle; b is the linear trend coefficient; c is the bias constant. R(t) = |P(t)-M(t)| / σ(t) [1 +α V(t)]; where R(t) is the reliability index; M(t) is the measurement value; σ(t) is the standard deviation; V(t) is the coefficient of variation; α is the adjustment factor (0.1~0.5).

[0124] Step S443: read the spatial association template from the database, perform spatial mapping on the current multi-dimensional evaluation index according to the spatial association template, and generate spatial distribution data. Calculate the correlation coefficient and consistency index for the index values ​​of adjacent areas in the spatial distribution data to generate regional related data. Input the regional related data into the spatial consistency evaluation model, and output the spatial consistency data in combination with the preset evaluation criteria.

[0125] Step S444: Input the multidimensional evaluation index, time stability data and spatial consistency data into the multi-level fuzzy evaluation model, calculate the fuzzy membership according to the importance of each layer of indicators, and generate a fuzzy evaluation matrix. Apply a comprehensive calculation method to the fuzzy evaluation matrix to obtain the final score of each evaluation index and generate comprehensive score data. Compare the comprehensive score data with the preset reliability level standard and output the reliability evaluation result.

[0126] The comprehensive calculation method for the fuzzy evaluation matrix is ​​as follows: M(i, j) = μ(i, j) [1 + α S(i, j)][1 + β T(i, j)]; where M(i, j) is the fuzzy evaluation value; μ(i, j) is the basic membership; S(i, j) is the spatial correlation; T(i, j) is the temporal correlation; α, β are the correlation weights (0~0.5). R = Σ[wi M(i)] exp(-γ V(i)); where R is the comprehensive evaluation result; wi is the indicator weight; M(i) is the evaluation value of each indicator; V(i) is the evaluation volatility; γ is the stability coefficient (0.2~0.6).

[0127] The specific process of outputting reliability assessment results is as follows: R(i)=Rb(i) [1 +α T(i)] [1 +β S(i)] [1+γ V(i)]; where R(i) is the comprehensive reliability score; Rb(i) is the basic reliability; T(i) is the temporal stability; S(i) is the spatial consistency; V(i) is the degree of fluctuation; α, β, γ are weight coefficients (0~0.5). C(i) =Σ[wj R(i,j)] exp(-λ D(i)); where C(i) is the confidence level; wj is the indicator weight; R(i,j) is the sub-item reliability; D(i) is the data deviation; λ is the deviation impact factor (0.2~0.4).

[0128] This embodiment solves the technical problems of difficult reliability assessment and poor stability of detection results through multi-dimensional indicator evaluation and spatiotemporal characteristic analysis. A multi-dimensional evaluation system including recognition accuracy, feature integrity, boundary clarity, etc. is constructed, and scientific quality evaluation is achieved through indicator normalization processing; a time series analysis method based on a sliding window is designed, and dynamic evaluation of indicators is achieved by calculating short-term volatility and long-term stability; spatial correlation analysis technology is used to evaluate the spatial distribution characteristics of indicators, and the reliability of evaluation is improved by calculating regional consistency; based on a multi-level fuzzy comprehensive evaluation model, the importance of each indicator is considered and a scientific comprehensive evaluation is achieved. This embodiment not only provides a quantification standard for the credibility of detection results, but also establishes a complete quality evaluation system.

[0129] According to one aspect of the present application, step S45 is further:

[0130] Step S451: extract the defect development trend in the historical monitoring data, establish a time series prediction model, analyze the change rules in combination with the reliability assessment results, and generate trend prediction data.

[0131] Step S452: Based on the trend prediction data and the current defect type and defect level, calculate the current value and predicted value of each warning indicator, evaluate the development risk, and output the warning indicator data.

[0132] It also includes dynamic updating of early warning indicators, specifically: P(t) = Pb(t) [1 +α H(t)] [1 +β F(t)]; where P(t) is the early warning indicator value; Pb(t) is the benchmark indicator value; H(t) is the historical trend item; F(t) is the future prediction item; α, β are time weights (0.2~0.4). U(t) = exp(-|dP(t) / dt| / σ1) exp(-|d 2 P(t) / dt 2 | / σ2); where U(t) is the update coefficient; dP(t) / dt is the rate of change; d 2 P(t) / dt 2 is the acceleration; σ1, σ2 are the control parameters.

[0133] Step S453: Based on the early warning indicator data, an adaptive threshold updating algorithm is used to dynamically adjust the thresholds of various early warning indicators and generate an updated threshold list.

[0134] The adaptive threshold update algorithm is as follows: T(k+1) = T(k) + η [μ(k)-T(k)] + β [(σ(k) / σ(k-1))-1]; where T(k+1) is the updated threshold; T(k) is the current threshold; μ(k) is the current mean; σ(k) is the current standard deviation; η is the learning rate (0.10.3); β is the fluctuation adjustment coefficient (0.20.4). W(k) = exp(-|T(k)-T(k-1)| / Δ) exp(-|σ(k)-σ(k-1)| / γ); where W(k) is the update weight; Δ, γ are stability parameters; T(k-1) is the threshold of the previous cycle; σ(k-1) is the standard deviation of the previous cycle.

[0135] Step S454: Based on the updated threshold list, optimize the sampling frequency, measurement accuracy and other parameters of the key monitoring points to generate a monitoring parameter optimization table.

[0136] Step S455: Integrate the early warning indicator data and the early warning threshold, determine the early warning level, update the monitoring and control parameters, and generate an early warning parameter update list and a defect identification report.

[0137] This embodiment solves the technical problems of untimely update of warning parameters and poor optimization effect through trend prediction and parameter adaptive optimization. A time series prediction model based on historical data is constructed, and the change law is analyzed in combination with the reliability assessment results to improve the accuracy of trend prediction; an adaptive threshold update algorithm is designed to dynamically adjust the threshold according to the change of the warning indicator, thereby improving the sensitivity and reliability of the warning; by optimizing the sampling parameters of key monitoring points, the monitoring density of key areas is guaranteed and the system operation efficiency is improved; a dynamic evaluation mechanism for the warning level is established, and adaptive optimization of the system is achieved by updating the monitoring control parameters. This embodiment not only improves the accuracy and real-time performance of the early warning system, but also realizes the continuous optimization of the detection system.

[0138] According to one aspect of the present application, an underwater multimodal identification system for high dam subtle defects includes:

[0139] at least one processor; and,

[0140] a memory communicatively connected to at least one of the processors; wherein,

[0141] The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the underwater multimodal identification method for high dam subtle defects described in any of the above embodiments.

[0142] The present invention generates sampling feature vectors by detecting the sampling frequencies of displacement, stress and temperature data respectively; then designs linear interpolation parameters according to the sampling feature vectors to achieve data resampling and time alignment; finally, generates a standardized multi-source monitoring data matrix by organizing spatial positions and sensor types. By dividing core data and peripheral data and calculating adaptive weight coefficients, the problem of uneven data quality is effectively solved. This processing mechanism ensures the time synchronization and quality consistency of data from different sources and realizes data synchronization processing.

[0143] The sound velocity profile is calculated by obtaining water depth, temperature and salinity data from the environmental database; then the attenuation coefficient at different depths is calculated based on the sound velocity profile, and the transmission power and receiving gain are optimized in combination with the target distance and water turbidity; finally, the influence of platform motion and sound velocity change is eliminated through attitude compensation and sound path correction. Overlapping sliding windows and piecewise linear mapping are used for gain adjustment, and frequency band processing is achieved through multi-layer filter groups, which effectively improves the quality and reliability of sonar signals and builds a complete environmental adaptation mechanism.

[0144] For crack defects, adaptive threshold segmentation and improved Hough transform are used to accurately extract the trend features; for damaged areas, the area and depth are calculated through regional growing algorithm and gray gradient analysis; by establishing a three-dimensional defect reconstruction model, not only the defect volume is calculated, but also the extension trend is analyzed. Finally, a multi-feature fusion classification method is used to achieve accurate identification of different types of defects and build a multi-level defect feature extraction system.

[0145] By extracting the timestamp and spatial position mark of the acoustic data, the acquisition time and camera posture information of the image data are extracted; then an improved spatiotemporal registration algorithm is used to generate a registration mapping matrix; finally, by calculating the quality indicators of the acoustic and image data, a multi-indicator fusion evaluation method is used to ensure the reliability of the registration results, and an accurate spatiotemporal registration mechanism is constructed.

[0146] By extracting multi-dimensional evaluation indicators from defect identification and location results, analyzing the temporal stability and spatial consistency of the indicators, and finally calculating the comprehensive reliability score through a multi-level fuzzy comprehensive evaluation model, the early warning parameters are dynamically updated based on the reliability evaluation results, achieving continuous optimization of the detection system, ensuring the long-term reliability of the detection results, and establishing a complete reliability evaluation system.

[0147] The present invention solves the key technical problems of data heterogeneity, environmental interference, difficulty in feature extraction, and low recognition accuracy in underwater defect detection of dams by constructing a complete technical chain of "monitoring and early warning-accurate detection-feature extraction-defect identification". Based on the abnormal early warning mechanism of multi-source monitoring data, the rapid and accurate positioning of suspicious areas is achieved through time synchronization processing and adaptive weight calculation; an acoustic and optical collaborative underwater detection method is proposed, and the quality and reliability of detection data are improved through multi-parameter optimization and adaptive control; a multi-modal feature extraction and fusion algorithm is designed, which makes full use of the complementary advantages of acoustic and optical data and enhances the integrity and accuracy of feature expression; a multi-level defect identification and evaluation system is established, and accurate identification and quantitative evaluation of defects are achieved through feature matching, spatial positioning and reliability evaluation. The present invention has strong practicality and adaptability, and can realize the automated and intelligent detection of dam defects in complex underwater environments, providing reliable technical support for dam safety monitoring.

[0148] It should be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

Claims

1. An underwater multimodal identification method for high dam subtle defects, characterized by: The steps include: Step S1, obtaining displacement monitoring data, stress monitoring data and temperature monitoring data from the sensor network database in real time, and generating a multi-source monitoring data matrix through time synchronization processing; extracting abnormal features from the multi-source monitoring data matrix using an adaptive weight calculation method to generate an abnormal feature matrix; inputting the abnormal feature matrix into a pre-configured multidimensional analysis model, calculating a suspicious area coordinate set, and generating a credibility score at the same time; based on the credibility score, prioritizing the suspicious area coordinate set, and outputting a suspicious area list; Step S2, obtaining a list of suspicious areas and planning a scanning path according to the list of suspicious areas; Based on the scanning path, the original sonar data is obtained, and the enhanced sonar data is obtained through multi-level signal enhancement processing; Based on enhanced sonar data, edge feature data is extracted, multi-feature analysis is performed, and precise defect areas and defect feature descriptions are determined; Step S3, based on the accurate defect area and defect feature description, collect acoustic data and optical data, obtain registration data through time and space alignment; enhance the registration data to obtain enhanced feature data; Based on the enhanced feature data, features are extracted and fused to obtain multimodal feature data; Generate fusion feature matrix and feature credibility index based on multimodal feature data; Step S4, based on the fusion feature matrix and the feature credibility index, the defect discrimination result is obtained through feature discrimination; Based on the defect identification results, multi-level analysis is performed to determine the defect type and defect level; Based on the defect type and defect level, the precise spatial position is calculated to obtain the positioning result; Based on the positioning results, a comprehensive evaluation is carried out to obtain the reliability assessment results, generate a defect identification report and a warning parameter update list.

2. The underwater multi-modal identification method for high dam subtle defects according to claim 1 is characterized in that: Step S1 is further as follows: Step S11, reading displacement monitoring data corresponding to a three-dimensional spatial position from a displacement sensor acquisition node, reading stress monitoring data at the same position from a stress sensor acquisition node, and reading temperature monitoring data at the same position from a temperature sensor acquisition node; performing sampling frequency detection on the displacement monitoring data, stress monitoring data, and temperature monitoring data, respectively, to generate a sampling feature vector; constructing linear interpolation parameters based on the sampling feature vector, resampling and time-aligning the displacement monitoring data, stress monitoring data, and temperature monitoring data, to generate time-aligned data; organizing the time-aligned data according to spatial position and sensor type, and outputting a multi-source monitoring data matrix; Step S12, dividing the multi-source monitoring data matrix into a core data part representing the main change trend and a peripheral data part representing the fluctuation change, calculating the adaptive weight coefficients of the core data part and the peripheral data part; multiplying the core data part and the peripheral data part by the corresponding adaptive weight coefficients respectively and adding them to obtain weighted monitoring data; based on the weighted monitoring data, calculating the distance between each monitoring point and its adjacent monitoring point; converting the distance into a correlation weight value; multiplying the correlation weight value by the weighted monitoring data to obtain an abnormal feature matrix; Step S13: for each monitoring point in the abnormal feature matrix, calculate its local correlation value with surrounding monitoring points and its global correlation value with all monitoring points; constructing an enhanced feature matrix based on a weighted combination of local correlation values ​​and global correlation values; Step S14: Calculate the dynamic threshold of each monitoring point based on the enhanced feature matrix; The monitoring points exceeding the corresponding dynamic threshold are recorded to form a suspicious area coordinate set; Step S15: for each area in the suspicious area coordinate set, respectively calculate its displacement anomaly degree, stress anomaly degree and temperature anomaly degree; combine the displacement anomaly degree, stress anomaly degree and temperature anomaly degree through adaptive weights to obtain a credibility score; based on the credibility score, sort the suspicious area coordinate set to generate a prioritized suspicious area list.

3. The underwater multi-modal identification method for high dam subtle defects according to claim 2 is characterized in that: Step S2 is further as follows: Step S21, based on the priority and spatial coordinates of each suspicious area in the suspicious area list, calculate the horizontal scanning angle and the vertical scanning angle to generate a sonar scanning path; control the sonar device to collect the echo signal of the target area according to the sonar scanning path to obtain the original sonar data including the distance value and the signal strength; Step S22, based on the original sonar data, calculating the signal mean and variance of the local area; based on the mean, determining the signal gain coefficient; based on the variance, determining the noise suppression coefficient; multiplying the signal gain coefficient with the original sonar data to obtain gain sonar data; multiplying the noise suppression coefficient with the gain sonar data to obtain enhanced sonar data; Step S23, based on the enhanced sonar data, calculate the gradient values ​​in the horizontal direction and the vertical direction; based on the gradient values, calculate the edge strength and direction angle, and determine the adaptive threshold in combination with the statistical characteristics of the local area; record the edge points above the adaptive threshold as edge feature data; Step S24: based on the edge feature data, calculate the intensity feature, shape feature and texture feature of each edge point; match the intensity feature, shape feature and texture feature with a preset defect feature template to obtain a feature matching probability; Marking the area where the feature matching probability is higher than a preset threshold as a precise defect area; Step S25: Based on the precise defect area, the geometric dimensions, edge clarity, echo intensity and spatial continuity of the area are calculated respectively; the geometric dimensions, edge clarity, echo intensity and spatial continuity are combined to form a defect feature description, and the reliability index of the feature is calculated at the same time to generate a defect detection report including the defect location, characteristics and reliability.

4. The underwater multi-modal identification method of high dam subtle defects according to claim 3 is characterized in that: Step S3 is further as follows: Step S31: Based on the precise defect area and defect feature description, control the acoustic sensor to collect acoustic raw data of the target area at different angles, and control the underwater camera to collect optical raw data at the corresponding position, and record the collection time and spatial coordinates of each data; based on the collection time and spatial coordinates, align the acoustic raw data and the optical raw data to generate registration data; Step S32, reading the acoustic part in the registration data, and calculating the high-frequency component and the low-frequency component; Based on the high-frequency component and the low-frequency component, after enhancing them respectively, the enhanced acoustic data are obtained by weighted combination; The optical part of the registration data is read, and after enhancing the contrast and texture features respectively, enhanced optical data is obtained by weighted combination; Based on the preset quality evaluation index, the complementary enhancement coefficient is calculated, and the enhanced acoustic data and the enhanced optical data are combined to obtain enhanced feature data; Step S33: based on the enhanced feature data, extracting spectrum features, time domain features and energy features from the enhanced acoustic data to generate acoustic feature data; extracting edge features, texture features and shape features from the enhanced optical data to generate optical feature data; calculating the similarity between the acoustic feature data and the optical feature data to generate a feature mutual verification matrix; Step S34: generating a quality assessment result based on the acoustic feature data and the optical feature data, calculating feature weights, and obtaining weighted features; The weighted features are fused to obtain multimodal feature data; Based on multimodal feature data, filter according to preset feature importance and generate a fusion feature matrix; Step S35, calculating the geometric size, physical characteristics, dynamic response and spatial distribution of each feature in the fused feature matrix to generate a feature description vector; Based on the feature description vector, the credibility of the acoustic feature data, optical feature data and feature mutual verification matrix is ​​calculated respectively, and the feature credibility index is obtained by weighted combination.

5. The underwater multi-modal identification method of high dam subtle defects according to claim 4 is characterized in that: Step S4 is further as follows: Step S41: Based on the feature credibility index, the fused feature matrix is ​​compared and calculated with the pre-stored structural similarity data, temporal correlation data and spatial position association data to obtain a feature discriminant matrix; Based on the feature weights in step S34, a feature discrimination matrix is ​​weighted and calculated, and a defect discrimination result is output; Step S42: based on the defect identification result, respectively calculating the matching degree with the preset crack feature library, erosion feature library, leakage feature library and deformation feature library to obtain feature matching data; Determine the defect type and classify the defect level based on feature matching data and defect identification results; Step S43, based on the defect type and defect level, combined with the pre-stored sonar positioning data, optical positioning data and monitoring positioning data, calculate the spatial positioning data; Based on the spatial positioning data, perform positioning error analysis and generate error evaluation data; based on the error evaluation data, determine the final positioning result and positioning confidence interval; Step S44, calculating the recognition accuracy index of the defect discrimination result, the spatial precision index of the positioning result, the characteristic time stability index and the spatial consistency index, and generating evaluation index data; Based on the evaluation index data, a comprehensive reliability score is obtained through weighted combination to form a reliability evaluation result; Step S45, based on the reliability assessment result, defect type, defect level and pre-stored historical evolution trend data, calculate the warning index value and generate warning level data; Based on the warning level data, update the list of key monitoring points, monitoring parameter sets and warning thresholds to form an updated list of warning parameters; Update the list based on the warning parameters and generate a defect identification report.

6. The underwater multi-modal identification method of high dam subtle defects according to claim 5 is characterized in that: Step S21 is further as follows: Step S211, read the spatial coordinate information of the suspicious area list, and based on the spatial coordinate information, use an adaptive grid decomposition algorithm to divide the detection area into predetermined sub-areas to generate grid division data; read the current water flow data and obstacle data from the pre-stored environment database, calculate the detection difficulty coefficient of each sub-area in combination with the grid division data, and output the difficulty coefficient matrix; The priority information in the suspicious area list is weightedly combined with the difficulty coefficient matrix to generate an area priority matrix; Based on the regional priority matrix, an improved ant colony algorithm is used for optimization to output the initial scanning path; Step S212: generating sound velocity profile data based on the initial scanning path and pre-stored water depth data; Based on the sound velocity profile data, the attenuation coefficient of the sonar beam at different depths is calculated and an attenuation coefficient table is generated; Based on the attenuation coefficient table, the optimal transmission power and receiving gain are calculated from the target area distance data; based on the optimal transmission power and receiving gain, the parameters are optimized in combination with the pre-stored turbidity data to generate a parameter configuration table; Step S213: Based on the parameter configuration table and the regional priority matrix, the optimal pitch angle and horizontal rotation angle of each detection position are calculated to generate a multi-angle scanning sequence; the sonar rotation mechanism is controlled to move according to the multi-angle scanning sequence to collect echo signal data in real time; Step S214, acquiring attitude data from the inertial navigation system, and based on the attitude data, performing attitude compensation on the echo signal data to generate compensated signal data; extracting sound path correction parameters from the sound velocity profile data, performing sound path correction on the compensated signal data to generate corrected signal data; performing time-series alignment and spatial splicing on all corrected signal data to output original sonar data.

7. The underwater multi-modal identification method for high dam subtle defects according to claim 5 is characterized in that: Step S22 is further as follows: Step S221, dividing the original sonar data into overlapping local analysis windows, calculating the signal mean, variance and peak distribution in each local analysis window, and generating a statistical feature matrix; based on Statistical feature matrix, using adaptive window algorithm to optimize window size and output optimized statistical data; Step S222: based on the optimized statistical data, the signal dynamic range of each window is calculated to generate a dynamic range vector; based on the dynamic range vector, a piecewise linear mapping function is constructed to generate a gain mapping table; the gain control parameter of each signal segment is extracted from the gain mapping table, the original sonar data is segmentedly gain adjusted according to the gain control parameter, and the gain adjustment data is output; Step S223, performing time-frequency analysis on the gain adjustment data to extract the frequency characteristics and energy distribution of the noise; based on the frequency characteristics and energy distribution, constructing a multi-layer noise model to identify system noise, environmental noise and interference noise, and generating a noise feature vector and a noise suppression coefficient; Step S224: Based on the noise feature vector and the noise suppression coefficient, an adaptive filter group is constructed to process the signal in frequency bands to obtain processed data; based on the processed data, a wavelet threshold method is used to eliminate noise, and the signal is reconstructed to obtain enhanced sonar data.

8. The underwater multi-modal identification method for high dam subtle defects according to claim 5 is characterized in that: Step S31 is further as follows: Step S311, read the spatial coordinates of the precise defect area, and divide the detection space into working sub-areas using an adaptive grid method; calculate the passage difficulty coefficient based on the working sub-areas, pre-stored water flow velocity and obstacle distribution; based on the passage difficulty coefficient, use a dynamic obstacle avoidance path planning algorithm to generate an initial path for the submersible; Based on the initial path of the submersible, the pre-configured energy consumption model is used for optimization to obtain the submersible operation path; Step S312: Based on the operation path of the submersible and the defect feature description, the optimal working distance of the acoustic sensor and the low-light camera is calculated, and an equipment layout plan is generated; based on the equipment layout plan, a multi-device collaborative control algorithm is used to calculate the working sequence of each sensor and output collaborative acquisition instructions; Step S313: read the real-time water quality parameters and ambient light intensity, dynamically adjust the exposure time, gain value and color enhancement parameters of the low-light camera, and generate a camera parameter table; Based on the camera parameter table, obtain the reflection characteristics of the target area, calculate the optimal fill light angle and intensity, and output the fill light control instructions; Step S314: based on the collaborative acquisition instruction and the fill light control instruction, receiving the original acoustic data and original image data collected by each sensor, extracting the timestamp and spatial position mark of the acoustic data, and extracting the acquisition time and camera posture information of the image data; Based on the timestamp, spatial position mark, acquisition time and camera posture information, a spatiotemporal registration algorithm is used to generate a spatiotemporal correspondence table; Step S315: Based on the time-space correspondence table, the signal-to-noise ratio and clarity of the acoustic data are calculated to obtain the acoustic quality index; based on the time-space correspondence table, the clarity and contrast of the image data are calculated to obtain the image quality index; based on the acoustic quality index and the image quality index, a multi-index fusion evaluation method is used to generate the registration data and data quality report.

9. The underwater multi-modal identification method for high dam subtle defects according to claim 5 is characterized in that: Step S32 is further as follows: Step S321, performing multi-scale wavelet decomposition on the acoustic signal in the registration data, extracting characteristic components of different frequency bands, calculating the energy distribution of each characteristic component, and generating a frequency domain characteristic matrix; Based on the frequency domain feature matrix, the frequency band correlation is analyzed to obtain frequency domain correlation data; the characteristic components of different frequency bands include high frequency components, medium frequency components and low frequency components; Step S322: based on the frequency domain feature matrix and the frequency domain correlation data, the high frequency components are directionally enhanced, the intermediate frequency components are selectively retained, the low frequency components are noise suppressed, and the processed components of each frequency band are integrated to generate enhanced acoustic data; Step S323, performing geometric correction and illumination balance on the optical image in the registration data, constructing an adaptive histogram model, compensating for uneven illumination effects, and generating corrected image data; Step S324, applying a multi-scale contrast enhancement algorithm to the corrected image data, adaptively adjusting enhancement parameters in combination with statistical features of the local area, optimizing detail and texture performance, and outputting enhanced optical data; Step S325, calculating the complementary features of the enhanced acoustic data and the enhanced optical data, and constructing feature fusion weights; Based on the feature fusion weights, an adaptive weighted method is used to combine the enhanced acoustic data and the enhanced optical data to generate enhanced feature data.

10. An underwater multi-modal identification system for high dam subtle defects, characterized in that: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the underwater multimodal identification method for high dam subtle defects as described in any one of claims 1 to 9.

Images