A method for measuring underwater defects of high dams based on multimodal data fusion

Abstract

The present invention discloses a method for measuring underwater defects of high dams based on multimodal data fusion, including: obtaining water flow velocity data, water pressure data and water quality turbidity data, using an adaptive median filter algorithm for preprocessing, and constructing a hydrodynamic field characteristic model; then generating an initial trajectory based on the hydrodynamic field characteristic model and the detection target area data, and performing trajectory optimization and stability evaluation in combination with dynamic constraints; then adjusting sonar parameters for data collection according to turbidity influence and trajectory parameters; then performing spatiotemporal alignment through extended Kalman filtering, and using an attention mechanism for feature fusion and optimization; finally performing multi-scale consistency verification and precise positioning mapping to obtain the final defect detection result. The present invention solves the problem of defect detection accuracy and reliability in complex underwater environments.

Description

Technical Field

[0001] The invention belongs to the field of underwater engineering detection, and in particular to a high dam underwater defect measurement method based on multi-modal data fusion. Background Art

[0002] As an important infrastructure of water conservancy projects, the safety of dams is directly related to the safety of people in downstream areas. Underwater defect detection is a key link in assessing the safety status of dams, especially for high dams over 200 meters, whose underwater environment has the characteristics of high water pressure, strong water flow, high turbidity, etc., which brings great challenges to the detection work. Accurate detection and positioning of underwater defects are of great significance for timely discovery of safety hazards, guidance of maintenance and reinforcement, and ensuring the safe operation of dams. At the same time, the early detection and precise positioning of underwater defects of high dams are also the basis for realizing health monitoring and predictive maintenance of dams throughout their life cycle.

[0003] At present, underwater defect detection mainly adopts single-mode detection methods such as sonar imaging, pressure sensing and water quality monitoring. Sonar detection detects the geometric features of the dam surface by emitting and receiving sound waves, but the signal quality is seriously attenuated in strong water flow and high turbidity environment; pressure sensing infers the structural state of the dam by measuring the change of water pressure, but it is easily affected by the change of water depth and produces measurement deviation; water quality monitoring can reflect the leakage status of local areas, but it is difficult to accurately locate the defect. Although some studies have tried to combine multiple detection methods, most of them remain at the simple data superposition level and lack systematic multimodal data fusion methods.

[0004] In practical applications, the existing technology has the following specific problems: First, in the preprocessing of water velocity data, the traditional fixed window filtering method cannot adapt to sudden water flow disturbances, resulting in poor outlier processing effects; second, the influence of water depth on pressure measurement is not considered in the process of extracting pressure gradient features, resulting in feature extraction deviation; third, in the spatial interpolation process of turbidity data, a simple linear interpolation method is used, which cannot accurately characterize the nonlinear distribution characteristics of turbidity; fourth, only motion constraints are considered in trajectory planning optimization, ignoring the influence of hydrodynamic disturbances on the posture stability of the detection equipment; fifth, no feature credibility evaluation mechanism is established in the feature fusion process, resulting in unreliable features interfering with the fusion results; finally, in the collaborative processing of multi-source data, the lack of a unified dimension alignment method affects the effect of feature fusion. These technical problems seriously restrict the accuracy and reliability of underwater defect detection. Summary of the invention

[0005] The purpose of the invention is to provide a method for measuring underwater defects of high dams based on multimodal data fusion, in order to solve at least one technical problem existing in the prior art.

[0006] The technical solution is a method for measuring underwater defects of high dams based on multimodal data fusion, which includes the following steps:

[0007] S1, obtain real-time water flow velocity data, water pressure data and water turbidity data, use adaptive median filter algorithm for preprocessing, and build a hydrodynamic field characteristic model;

[0008] S2. Generate an initial trajectory point sequence based on the hydrodynamic field characteristic model and the predetermined detection target area data; perform trajectory optimization and stability evaluation in combination with the initial trajectory point sequence and a preset set of dynamic constraint conditions to obtain optimized trajectory parameters and stability evaluation data;

[0009] S3. Based on the water turbidity data and the optimized trajectory parameters, the sonar transmission power and gain parameters are calculated to obtain the sonar acquisition parameter set; according to the sonar acquisition parameter set and the optimized trajectory parameters, collaborative data collection and processing are performed to obtain collaboratively enhanced multimodal data;

[0010] S4, performing spatiotemporal alignment processing on the multimodal data to obtain registered data; based on the registered data, stability assessment data and pre-stored historical detection data, using the attention mechanism to perform feature fusion and optimization to obtain fused and optimized defect feature data;

[0011] S5. Perform multi-scale consistency verification and precise positioning mapping on the defect feature data to obtain the final defect detection result.

[0012] Beneficial effects: The present invention overcomes the limitations of a single sensor, ensures the reliability of the detection process, and improves detection accuracy; realizes high-precision detection of dam body defects, and provides strong technical support for the safe operation of reservoir dams; not only solves the problems of poor reliability and low accuracy of traditional detection methods in complex environments, but also improves detection efficiency through a systematic data processing process, and has strong engineering practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a flow chart of the present invention.

[0014] Figure 2 This is a flow chart of step S1 of the present invention.

[0015] Figure 3 This is a flow chart of step S2 of the present invention.

[0016] Figure 4 This is a flow chart of step S3 of the present invention.

[0017] Figure 5 This is a flow chart of step S4 of the present invention.

[0018] Figure 6This is a flow chart of step S5 of the present invention. DETAILED DESCRIPTION

[0019] The present application will be described in more detail below in conjunction with specific embodiments. Figure 1 As shown, the present application proposes a method for measuring underwater defects of high dams based on multimodal data fusion, comprising the following steps:

[0020] S1. Acquire real-time water velocity data, water pressure data and water turbidity data containing spatial position and time information; perform outlier filtering on the real-time water velocity data to obtain pre-processed water velocity data; construct a pressure gradient field based on the water pressure data to obtain a pressure gradient feature vector; construct a spatial distribution model based on the water turbidity data to obtain a turbidity influence coefficient matrix; perform comprehensive processing on the pre-processed water velocity data, pressure gradient feature vector and turbidity influence coefficient matrix to obtain a hydrodynamic field feature model;

[0021] S2. Generate an initial trajectory point sequence based on the hydrodynamic field characteristic model and the predetermined detection target area data; obtain the dynamic parameter data of the submersible, calculate the dynamic constraints based on the dynamic parameter data and the hydrodynamic field characteristic model, and obtain a dynamic constraint condition set; perform trajectory optimization based on the initial trajectory point sequence and the dynamic constraint condition set to obtain optimized trajectory parameters; perform stability calculation based on the optimized trajectory parameters and the hydrodynamic field characteristic model to obtain stability evaluation data;

[0022] S3, obtaining the current position data of the submersible, and calculating the sonar transmission power and gain parameters based on the current position data, the turbidity influence coefficient matrix and the optimized trajectory parameters to obtain the sonar acquisition parameter set; performing collaborative data acquisition and processing according to the sonar acquisition parameter set and the optimized trajectory parameters to obtain collaboratively enhanced multimodal data;

[0023] S4, performing spatiotemporal alignment processing on the multimodal data to obtain registered data; performing feature extraction and fusion based on the registered data and pre-stored historical detection data to obtain a fused feature vector; optimizing the fused feature vector and stability assessment data to obtain fused and optimized defect feature data;

[0024] S5. Perform multi-scale consistency verification on the fused and optimized defect feature data to obtain verified defect feature data; perform precise positioning and mapping based on the verified defect feature data and pre-stored spatial reference data to obtain the final defect detection result.

[0025] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0026] S11, obtaining real-time water flow velocity data including spatial position and time information; using an adaptive median filter algorithm to process outliers on the real-time water flow velocity data to obtain pre-processed water flow velocity data;

[0027] S12, obtaining water pressure data including spatial position and time information; based on the water pressure data, calculating the rate of change of pressure in three-dimensional space to obtain a pressure gradient characteristic vector;

[0028] S13, obtaining water quality turbidity data including spatial position and time information; performing spatial interpolation processing on the water quality turbidity data to obtain a continuously distributed turbidity influence coefficient matrix;

[0029] S14, performing weighted fusion processing on the pre-processed water flow velocity data, pressure gradient characteristic vector and turbidity influence coefficient matrix to obtain a hydrodynamic field characteristic model that characterizes the hydrodynamic environment.

[0030] In one embodiment of the present application, real-time water flow velocity data v(x, y, z, t) is obtained, where x, y, z are spatial coordinates and t is time, and an adaptive median filter is used to remove outliers to obtain pre-processed water flow velocity data v'(x, y, z, t). Based on the water pressure data p(x, y, z, t) and the water depth data d(x, y, z), a three-dimensional pressure gradient field is constructed to obtain a pressure gradient characteristic vector ▽p(x, y, z, t). Based on the water quality turbidity data τ(x, y, z, t), a turbidity spatial distribution model is established to obtain a turbidity influence coefficient matrix T(x, y, z). Based on the pre-processed water flow velocity data v'(x, y, z, t), the pressure gradient characteristic vector ▽p(x, y, z, t) and the turbidity influence coefficient matrix T(x, y, z), a comprehensive hydrodynamic field characteristic model is constructed, and the hydrodynamic field characteristic model F(x, y, z, t) is output.

[0031] This embodiment establishes a complete hydrodynamic field characteristic model through multimodal fusion processing of real-time water velocity data, water pressure data and water quality turbidity data. Specifically, the use of an adaptive median filter algorithm to process water velocity anomalies can effectively deal with sudden water flow disturbances; the spatial distribution characteristics of the pressure field are accurately captured by constructing the water pressure gradient feature vector; the spatial interpolation method is used to process water quality turbidity data to achieve a continuous expression of the turbidity effect. This embodiment not only takes into account the impact of various environmental factors on underwater detection, but also improves the accuracy and reliability of the hydrodynamic field characteristic model through the complementarity of data. In the complex underwater environment of a high dam, the problem that a single sensor data is susceptible to interference can be effectively overcome, providing stable and reliable environmental characteristic information for subsequent trajectory planning and data acquisition.

[0032] According to one aspect of the present application, step S11 is further:

[0033] S111, acquiring real-time water flow velocity data, segmenting the data sequence of the real-time water flow velocity data according to a preset time window and a preset spatial position to obtain segmented data; based on the segmented data, calculating the time interval and spatial distribution characteristics of each segment of data to obtain a calculation result; according to the calculation result, determining an initial window size and a sliding step length; based on the initial window size and the sliding step length, using a sliding window method to block the segmented data to obtain a data window sequence;

[0034] S112. Based on the data window sequence, the mean, standard deviation and interquartile range of the data in each window are calculated to obtain statistical characteristics; based on the statistical characteristics, an adaptive threshold function is constructed to calculate the adjustment coefficient of the window size; based on the adjustment coefficient, the window size is dynamically adjusted to obtain an optimized window sequence;

[0035] S113, based on the optimized window sequence, calculating the local median of the data in each optimized window; constructing a data distribution model based on the local median; calculating the deviation between the data point and the data distribution model to obtain a deviation value; comparing the deviation value with a preset dynamic threshold, identifying abnormal data points, and obtaining an abnormal value marker sequence;

[0036] S114. Based on the outlier marker sequence, extract the outlier data points and the corresponding real-time water flow velocity data; based on the outlier data points and the corresponding real-time water flow velocity data, calculate the replacement value through the data distribution model; use the exponential weighted average method to smooth the replacement value to obtain the smoothed replacement value; based on the smoothed replacement value, iteratively execute the outlier replacement process until the preset convergence condition is met to obtain the pre-processed water flow velocity data.

[0037] In one embodiment of the present application, a window size adaptive function W(t) = W_base + γ·σ(t)·(1 - exp(-μ·v(t))) is constructed; a data deviation score S(i) = |V(i) - M(t)| / (σ(t) + ε); an outlier determination threshold T(t) = α + β·σ(t)·log(1 + |dv / dt|); a filter output V_f(i) = ∑(w_k·V_k) / (∑w_k); wherein W_base is the base window size; σ(t) is the local standard deviation; v(t) is the local velocity mean; V(i) is the current velocity value; M(t) is the local median; w_k is the distance weight; V_k is the kth velocity value, γ, μ, α, β, ε are control parameters; and dv / dt is the velocity change rate.

[0038] This embodiment achieves high-quality preprocessing of water velocity data. Specifically, a segmented processing strategy based on time window and spatial position is adopted to effectively capture the spatiotemporal distribution characteristics of the data; the window size is dynamically adjusted through an adaptive threshold function to improve the adaptability of the filter; the data distribution model based on the local median is used to identify outliers to ensure the accuracy of the filter; the exponential weighted average method is used for smoothing to improve the continuity of the data. In the underwater environment of the high dam, the water velocity often undergoes mutations and fluctuations. This embodiment can accurately identify and process these abnormal fluctuations through adaptive window size adjustment; the law of change of water velocity is accurately grasped through the construction of a local data distribution model; and the reliability of outlier processing is ensured by an iterative optimization strategy. This embodiment not only effectively removes the influence of noise and outliers, but also well maintains the effective information of the water velocity data, providing high-quality basic data for the subsequent construction of the hydrodynamic field characteristic model. Especially in the strong water flow area, this embodiment shows excellent robustness and can accurately capture the characteristics of the rapid change of water velocity, which is of great significance to ensure the safety and reliability of the detection process.

[0039] According to one aspect of the present application, step S12 is further:

[0040] S121, obtaining water pressure data and water depth data including spatial position and time information; constructing a depth correction coefficient matrix according to the water depth data; performing a product operation on the water pressure data and the depth correction coefficient matrix to obtain corrected water pressure data;

[0041] S122, based on the corrected water pressure data, calculate the pressure change rate in the X-axis direction to obtain the X-axis pressure gradient component; based on the corrected water pressure data, calculate the pressure change rate in the Y-axis direction to obtain the Y-axis pressure gradient component; based on the corrected water pressure data, calculate the pressure change rate in the Z-axis direction to obtain the Z-axis pressure gradient component; vector combination of the X-axis pressure gradient component, the Y-axis pressure gradient component and the Z-axis pressure gradient component to obtain an initial pressure gradient vector;

[0042] S123. Based on the water depth data, interpolation calculation is performed in the pre-stored depth compensation model parameters to obtain a depth compensation coefficient; the initial pressure gradient vector is multiplied by the depth compensation coefficient to obtain a compensated pressure gradient vector; and the compensated pressure gradient vector is normalized to obtain a pressure gradient characteristic vector.

[0043] In one embodiment of the present application, the pressure gradient depth compensation calculation method is: depth correction pressure P_c(x, y, z, t) = P(x, y, z, t)·α(z)·β(t); where α(z) = 1 + k1·z + k2·z 2is the depth correction coefficient; β(t) = exp(-λ·|dz / dt|) is the dynamic compensation factor; the pressure gradient vector G(x, y, z) = [ΨP_c / Ψx, ΨP_c / Ψy, ΨP_c / Ψz]; the normalized pressure gradient feature F_p = G / ||G||; P(x, y, z, t) is the original pressure measurement; x, y, z are the spatial coordinates; t is the time; k1, k2 are the depth correction parameters; λ is the dynamic response coefficient; dz / dt is the vertical motion velocity, and Ψ is the partial derivative.

[0044] This embodiment realizes depth adaptive correction of water pressure data by integrating water depth data into the calculation process of pressure gradient feature vector. First, the original water pressure data is calibrated using the depth correction coefficient matrix to effectively eliminate the pressure measurement deviation caused by water depth changes; then, the spatial distribution characteristics of the pressure field are accurately captured through the decomposition and calculation of the three-dimensional pressure gradient components; finally, the depth compensation model parameters are introduced for secondary correction to further improve the accuracy of the pressure gradient characteristics. This embodiment not only solves the problem of ignoring the influence of water depth in traditional pressure gradient calculations, but also provides standardized feature expressions through normalization processing, providing high-quality pressure field feature data for subsequent multimodal data fusion. In the underwater environment of high dams, the accuracy and reliability of pressure field feature extraction are improved.

[0045] According to one aspect of the present application, step S13 is further:

[0046] S131. Acquire water turbidity data, determine a three-dimensional grid division scheme based on the spatial range of the monitoring area; calculate the matching degree between the grid density and the sampling point distribution; dynamically adjust the grid division parameters; generate a regular grid point coordinate sequence, and obtain grid sampling point data.

[0047] S132, obtaining grid sampling point data and original water quality turbidity data, calculating the Euclidean distance from each grid point to all turbidity sampling points; constructing a Gaussian weight function based on the distance value; calculating the influence weight of each sampling point on the grid point; normalizing the weight to obtain a spatial weight matrix.

[0048] S133, obtaining grid sampling point data, water quality turbidity data and spatial weight matrix, and weighting the turbidity values ​​of the sampling points using spatial weights; calculating the initial interpolation results of each grid point; constructing a local error evaluation function; optimizing the interpolation parameters based on the error function to obtain initial turbidity distribution data.

[0049] S134, obtaining initial turbidity distribution data, constructing a radial basis function kernel matrix; calculating a smoothing coefficient; performing iterative smoothing processing; performing range constraints on the smoothing results; normalizing the processed data to the interval [0, 1], and obtaining a turbidity influence coefficient matrix.

[0050] In one embodiment of the present application, a spatial weight function W(r) = exp(-r 2 / 2h 2 )·(1 + Δ·cos(πr / R)); interpolated turbidity value T(x, y, z) = ∑(W_i·T_i·K(x, y, z)) / ∑W_i; kernel function K(x, y, z) = (1 - ||d|| / R) 2 ·(1 + η·||▽T||); smoothness constraint E(T) = ∑W_i·(T - T_i) 2 +λ·∑||▽ 2 T|| 2 ; where r is the spatial distance; h is the bandwidth parameter; R is the influence radius; T_i is the turbidity value of the sampling point; d is the spatial position difference vector; ▽T is the turbidity gradient; η, Δ, λ are control parameters, and W_i is the weight associated with the i-th sampling point.

[0051] This embodiment achieves high-precision turbidity impact modeling. First, through the dynamic grid division strategy, a spatial reference frame matching the sampling point distribution is established; the spatial weight matrix constructed based on the Gaussian weight function accurately expresses the spatial correlation; the error evaluation function is used to optimize the interpolation parameters to improve the interpolation accuracy; the radial basis function kernel matrix is ​​used for smooth processing to enhance the continuity of the data. In the underwater environment of the high dam, the spatial distribution of water turbidity often shows obvious non-uniformity and mutation characteristics. This embodiment accurately captures this spatial variation feature through adaptive grid division; the spatial transfer relationship of turbidity is accurately expressed by the construction of the spatial weight matrix; the accuracy and smoothness of the data are effectively balanced through parameter optimization and smooth processing. This embodiment not only achieves high-quality reconstruction of turbidity data, but also provides standardized impact assessment indicators through normalization processing, which provides a reliable basis for the adaptive adjustment of sonar parameters. Especially in areas with complex and changeable water quality conditions, this embodiment can accurately depict the spatial distribution law of turbidity, providing important support for the optimization of detection strategies.

[0052] In another embodiment of the present application, step S134 is further as follows:

[0053] S1341, obtaining initial turbidity distribution data, constructing a radial basis function kernel matrix, and obtaining a spatial interpolation operator; calculating the gradient of the interpolation error, optimizing the smoothing coefficient based on the gradient descent method, and obtaining an optimized smoothing parameter; performing iterative smoothing processing, and performing range constraints on the smoothing result to obtain smoothed turbidity data;

[0054] S1342. Obtain pre-stored numerical precision control parameters, determine the quantization level according to the numerical precision control parameters, perform piecewise linear quantization on the smoothed turbidity data to obtain quantized turbidity data; normalize the quantized turbidity data to the interval [0, 1], use piecewise linear interpolation to maintain the local characteristics of the data, and obtain a turbidity influence coefficient matrix.

[0055] This embodiment achieves high-fidelity conversion of turbidity data. Based on smooth processing, the accuracy of data conversion is ensured by piecewise linear quantization and local feature preservation interpolation. This embodiment effectively solves the problem of accuracy loss in data transmission, and improves the reliability of data conversion by optimizing the design of quantization levels and maintaining local features. In particular, when processing continuously changing turbidity data, the local features of the data can be accurately maintained, providing a high-quality control basis for sonar parameter adjustment.

[0056] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0057] S21, based on the hydrodynamic field characteristic model and the predetermined detection target area data, a path sequence is generated by using a heuristic search algorithm to obtain an initial trajectory point sequence;

[0058] S22, obtaining dynamic parameter data of the submersible, and calculating control constraints using dynamic equations based on the dynamic parameter data and a hydrodynamic field characteristic model to obtain a set of dynamic constraint conditions;

[0059] S23, based on the initial trajectory point sequence and the set of dynamic constraint conditions, using the model predictive control method to perform optimization calculation to obtain optimized trajectory parameters;

[0060] S24. Based on the optimized trajectory parameters and the hydrodynamic field characteristic model, the trajectory stability index is calculated to obtain stability evaluation data.

[0061] In one embodiment of the present application, based on the hydrodynamic field characteristic model F(x, y, z, t) and the detection target area R, the A* algorithm is used to generate the initial path to obtain the initial trajectory point sequence P0(t). Based on the submersible dynamic parameters D and the hydrodynamic field characteristic model F(x, y, z, t), the dynamic feasibility constraints are calculated and the dynamic constraint condition set C(t) is output. Based on the initial trajectory point sequence P0(t) and the dynamic constraint condition set C(t), the trajectory is optimized using model predictive control (MPC) to obtain the optimized trajectory P(t)*. Based on the optimized trajectory P(t)* and the hydrodynamic field characteristic model F(x, y, z, t), the trajectory stability index is calculated to obtain the stability evaluation data S(t).

[0062] This embodiment adopts a trajectory optimization method based on heuristic search and model predictive control to achieve stable trajectory planning in complex hydrodynamic environments. Trajectory generation is guided by the hydrodynamic field characteristic model to ensure the reasonable distribution of trajectory points; optimization calculations are performed in combination with the dynamic constraints of the submersible to ensure the executability of the trajectory; and dynamic optimization of the trajectory is achieved through the stability evaluation mechanism. This embodiment not only takes into account the integrity of the detection task, but also ensures the actual feasibility of the trajectory through dynamic constraints and stability evaluation. Especially in the underwater strong water flow environment of a high dam, it can effectively balance the requirements of detection coverage and motion stability, thereby improving detection efficiency and reliability.

[0063] According to one aspect of the present application, step S23 is further:

[0064] S231. Obtain the initial trajectory point sequence and establish the motion state equation of the submersible; calculate the state transfer matrix; construct the system noise model; design the prediction time domain length; generate the prediction sequence and obtain the state prediction model.

[0065] S232, obtain the state prediction model, construct the trajectory tracking error term; design the control amount penalty term; establish the control increment penalty term; determine the weight coefficients of each item; combine to form a complete optimization objective function to obtain the optimization objective function.

[0066] S233. Obtain the set of dynamic constraints and the optimization objective function, extract the control force constraints; process the motion state constraints; construct the collision avoidance constraints; convert into the standard optimization problem form to obtain the constrained optimization model.

[0067] S234, obtain the constrained optimization model and initialize the solution parameters; perform the interior point method iterative solution; check the convergence conditions; update the Lagrange multiplier; output the optimal control sequence to obtain the optimized control sequence.

[0068] S235, obtain the optimized control sequence, extract the first control quantity; update the system state; perform rolling prediction in the time domain; repeat the optimization solution process; integrate the optimization results, and obtain the optimized trajectory parameters.

[0069] In one embodiment of the present application, a state prediction model X(k+1) = A(k)·X(k) + B(k)·U(k) + F(k)·D(k) is constructed; the objective function J = ∑(||X(k) - X_r(k)|| Q + ||U(k)|| R + ||ΔU(k)|| S ); motion constraint g(X,U)≤0; optimization iteration ΔU = -(H + ρI) -1·(▽J + μ·▽g); where X(k) is the state vector; U(k) is the control input; D(k) is the disturbance term; A(k), B(k), F(k) are system matrices; X_r(k) is the reference trajectory; Q, R, S are weight matrices; H is the Hessian matrix; I is the identity matrix; ρ, μ are optimization parameters, ΔU(k) ​​is the control input increment, ▽J is the gradient of the objective function, and ▽g is the gradient of the motion constraint.

[0070] This embodiment can accurately predict the motion trajectory of the submersible by obtaining the initial trajectory point sequence and establishing the motion state equation of the submersible, thereby achieving accurate trajectory tracking. By constructing the trajectory tracking error term, the control amount penalty term and the control increment penalty term, and determining the weight coefficients of each item, a complete optimization objective function can be formed to achieve optimization control. By obtaining the set of dynamic constraints and the optimization objective function, extracting the control force constraints, processing the motion state constraints, constructing the collision avoidance constraints, and converting them into the form of a standard optimization problem, an optimization model that meets various constraints can be obtained. By initializing the solution parameters, executing the interior point method iterative solution, checking the convergence conditions, and updating the Lagrange multiplier, the optimization problem can be efficiently solved to obtain the optimal control sequence. By obtaining the optimized control sequence, extracting the first control amount, updating the system state, rolling the prediction time domain, and repeating the optimization solution process, the control strategy can be dynamically adjusted to ensure the stable operation of the submersible in a complex environment. By integrating the optimization results, the optimized trajectory parameters can be obtained, thereby achieving efficient, stable and safe operation of the submersible. This embodiment can improve the trajectory tracking accuracy and operating efficiency of the submersible through precise prediction, optimized control and dynamic adjustment, while ensuring its safety in complex environments.

[0071] According to one aspect of the present application, step S24 is further:

[0072] S241. Obtain optimized trajectory parameters and hydrodynamic field characteristic model, extract water flow velocity gradient at trajectory point; calculate hydrodynamic disturbance force and moment; establish disturbance transfer model; analyze disturbance components in each degree of freedom, and obtain disturbance characteristic sequence.

[0073] S242. Obtain disturbance characteristic sequence and establish the attitude dynamics equation of the submarine; calculate the fluctuation response of each attitude angle; analyze the statistical distribution characteristics of the attitude angle; evaluate the attitude stability margin and obtain the attitude stability index.

[0074] S243. Obtain the attitude stability index and calculate the effective detection range of the detection equipment; analyze the impact of attitude fluctuations on the detection range; evaluate the overlap of adjacent detection areas; calculate the coverage integrity of the detection area and obtain the coverage stability index.

[0075] S244. Obtain attitude stability index and coverage stability index, establish a multi-index evaluation system; calculate index weight coefficients; perform weighted fusion operations; generate a comprehensive score to obtain stability evaluation data.

[0076] In one embodiment of the present application, a disturbance transfer function H(ω) = (K_p·s + K_i) / (I*·s 2 + D·s + K_p·s + K_i); attitude stability index S_a = ∑(θ_rms·w_f) / (1 + λ·|dθ / dt|); coverage stability index S_c = exp(-κ·var(θ))·(A_overlap / A_total); comprehensive score E = ω1·S_a +ω2·S_c + ω3·(1 - max|θ| / θ_max); where ω is frequency; K_p, K_i are control parameters; I* is the moment of inertia; D is the damping coefficient; θ_rms is the attitude root mean square value; w_f is the frequency weight; θ is the attitude angle; A_overlap is the overlapping area; κ, λ, ω1, ω2, ω3 are weight coefficients, s is a complex variable in the complex frequency domain, A_total is the total area, and θ_max is the maximum attitude angle.

[0077] This embodiment can accurately analyze the components of disturbances in each degree of freedom and obtain a detailed disturbance feature sequence by extracting the water velocity gradient at the trajectory point, calculating the hydrodynamic disturbance force and torque, and establishing a disturbance transfer model. By establishing the attitude dynamics equation of the submersible, calculating the fluctuation response of each attitude angle, analyzing the statistical distribution characteristics of the attitude angle, and evaluating the attitude stability margin, an accurate attitude stability index can be obtained. By calculating the effective detection range of the detection equipment, analyzing the impact of attitude fluctuations on the detection range, evaluating the overlap of adjacent detection areas, and calculating the coverage integrity of the detection area, a detailed coverage stability index can be obtained. By establishing a multi-index evaluation system, calculating the index weight coefficient, performing a weighted fusion operation, and generating a comprehensive score, comprehensive stability evaluation data can be obtained. By obtaining the optimized trajectory parameters and the hydrodynamic field characteristic model, the control strategy of the submersible can be optimized to improve its operating efficiency and stability in a complex environment. By rolling the prediction time domain and repeating the optimization solution process, the control strategy can be dynamically adjusted to ensure that the submersible remains stable in a changing environment. This embodiment can improve the operating efficiency and stability of the submersible through precise disturbance analysis, attitude stability assessment and coverage stability assessment, combined with optimized control strategies and dynamic adjustment capabilities.

[0078] In another embodiment of the present application, step S241 is further:

[0079] S2411, obtaining optimized trajectory parameters and a hydrodynamic field characteristic model, extracting a three-dimensional coordinate sequence of trajectory points according to the optimized trajectory parameters, substituting the three-dimensional coordinate sequence into the hydrodynamic field characteristic model, and obtaining water flow velocity data at the trajectory points; calculating the water flow velocity difference between adjacent trajectory points to obtain a water flow velocity gradient sequence;

[0080] S2412, calculating the hydrodynamic disturbance force at each trajectory point according to the water flow velocity gradient sequence to obtain a disturbance force sequence; obtaining the pre-stored submersible shape parameters, calculating the hydrodynamic force arm according to the submersible shape parameters, performing a cross multiplication operation on the disturbance force sequence and the hydrodynamic force arm to obtain a disturbance moment sequence; combining the disturbance force sequence and the disturbance moment sequence to obtain an initial disturbance feature;

[0081] S2413, performing wavelet transform on the initial disturbance characteristics, extracting disturbance components of different scales, and obtaining multi-scale disturbance data; grouping the multi-scale disturbance data according to frequency, calculating the energy distribution of each frequency group, and obtaining a disturbance feature sequence.

[0082] This embodiment constructs a complete hydrodynamic disturbance feature evaluation system by introducing disturbance torque analysis. The disturbance force is calculated based on the water velocity gradient, the disturbance torque is analyzed in combination with the submersible shape parameters, and the disturbance components of different frequencies are extracted through multi-scale wavelet transform, and finally a comprehensive disturbance feature sequence is obtained. This dual disturbance analysis method including force and torque accurately depicts the influence mechanism of hydrodynamic force on the attitude of the submersible, especially when dealing with disturbance problems under strong water flow conditions. The energy distribution characteristics of the disturbance are revealed through frequency domain analysis, which provides a reliable basis for attitude stability evaluation. This embodiment not only improves the integrity of disturbance feature extraction, but also enhances the accuracy of stability evaluation through energy distribution analysis, which is of great significance to improving the reliability of the detection process.

[0083] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0084] S31, obtaining the current position data of the submersible, and calculating the transmission power and gain parameters of the sonar based on the current position data and the turbidity influence coefficient matrix to obtain a sonar acquisition parameter set;

[0085] S32. Based on the sonar acquisition parameter set and the optimized trajectory parameters, sonar data is collected and enhanced to obtain collaboratively enhanced multimodal data.

[0086] In one embodiment of the present application, based on the turbidity influence coefficient matrix T(x, y, z) and the current position p(t), the sonar transmission power and receiving gain are optimized to obtain the sonar acquisition parameter set Θs(t). Based on the sonar acquisition parameter set Θs(t) and the optimized trajectory P(t)*, the acquisition strategy is adjusted through sonar feedback to output the collaboratively enhanced multimodal data M(x, y, z, t).

[0087] This embodiment realizes adaptive adjustment of sonar parameters and collaborative enhancement of data through the coordinated control of the turbidity influence coefficient matrix and the optimized trajectory parameters. The sonar parameters are dynamically adjusted based on the turbidity characteristics of the current position, effectively overcoming the impact of turbid water on sonar detection; the trajectory parameters are used to guide the data collection process to ensure the integrity and continuity of data collection. This embodiment shows strong adaptability in the complex underwater environment of the high dam, and can automatically adjust the detection strategy according to changes in environmental conditions, thereby improving data quality and detection effects. Especially in areas with poor water quality conditions, the signal-to-noise ratio can be effectively improved through parameter optimization and data enhancement to ensure the availability of detection data.

[0088] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0089] S41, based on the multimodal data, using extended Kalman filtering to perform spatiotemporal alignment processing to obtain registered data;

[0090] S42, obtaining the registered data and the pre-stored historical detection data, and using the attention mechanism to perform feature fusion to obtain a fused feature vector;

[0091] S43. Based on the fused feature vector and the stability assessment data, multi-level fusion optimization processing is performed to obtain defect feature data after fusion optimization.

[0092] In one embodiment of the present application, based on the collaboratively enhanced multimodal data M(x, y, z, t), the extended Kalman filter is used for spatiotemporal alignment to obtain the registered data M'(x, y, z, t). Based on the registered data M'(x, y, z, t) and the historical detection feature library H, a feature fusion network with an attention mechanism is used to obtain a fused feature vector F(t). Based on the fused feature vector F(t) and the stability assessment data S(t), through a confidence-weighted multi-level fusion, the fused and optimized defect feature data D(x, y, z, t) is output.

[0093] This embodiment adopts a multimodal data fusion method of extended Kalman filtering and attention mechanism to achieve high-quality defect feature extraction. The time series differences of multi-source data are eliminated through spatiotemporal alignment processing; the attention mechanism is used to highlight important features and suppress irrelevant information; and multi-level fusion optimization is performed in combination with stability assessment data to improve the reliability of features. This embodiment not only overcomes the problem that single modal data is easily disturbed by the environment, but also improves the accuracy of defect detection through the complementarity of features. In the complex underwater environment of high dams, multi-source heterogeneous data can be effectively processed, and stable and reliable defect features can be extracted, providing strong support for the final defect identification.

[0094] According to one aspect of the present application, step S42 is further:

[0095] S421, constructing a multi-scale feature extractor based on the registered data; using the multi-scale feature extractor, calculating a local feature descriptor; extracting global context information of the registered data, combining the local feature descriptor and the global context information, and obtaining an original feature set;

[0096] S422, based on the original feature set, construct an attention query matrix, calculate the correlation score between features; based on the correlation score, generate an attention mask; based on the original feature set and the attention mask, perform an attention weighting operation to obtain attention weight data;

[0097] S423, obtaining historical detection data, and constructing a feature matching measurement function based on the historical detection data and the attention weight data; based on the feature matching measurement function, calculating the similarity between feature pairs to obtain a similarity score; based on the similarity score, using a nearest neighbor matching algorithm to obtain a nearest neighbor matching result; based on the nearest neighbor matching result, screening reliable matching pairs to obtain a feature matching result;

[0098] S424, based on the feature matching result, analyzing the discriminative performance of the feature to obtain a performance analysis result; based on the performance analysis result, enhancing the key feature component to obtain an enhanced key feature component; based on the performance analysis result, identifying and suppressing redundant features to obtain a suppressed feature set; based on the enhanced key feature component and the suppressed feature set, optimizing the feature expression to obtain enhanced feature data;

[0099] S425. Calculate the feature principal component based on the enhanced feature data; select the optimal feature subspace according to the variance contribution rate of the feature principal component; project the enhanced feature data onto the optimal feature subspace, perform dimensionality reduction transformation, and obtain the feature data after dimensionality reduction; normalize the feature data after dimensionality reduction to obtain a fused feature vector.

[0100] In one embodiment of the present application, attention weight α(i, j) = softmax(Q·K T / sqrt(d))·V; feature similarity S(i,j) = exp(-||F_i - F_j|| 2 / σ 2 )·(1 + ρ·cos(F_i, F_j)); enhanced feature F_e = F + β·(W_a·F)·(1 + γ·||▽F||); dimensionality reduction projection Y = W·(F_e - μ) / (σ + ε); where Q, K, V are query, key, and value matrices; d is the feature dimension; F_i, F_j are feature vectors; W_a is the attention weight matrix; ▽F is the feature gradient; μ, σ are the mean and standard deviation; β, γ, ρ, ε are control parameters, T is transposed.

[0101] This embodiment achieves efficient feature expression and fusion through the fusion strategy of multi-scale feature extraction and attention mechanism. First, a multi-scale feature extractor is used to perform hierarchical analysis on the data, effectively capturing defect features of different scales; the correlation between features is calculated through the attention query matrix, highlighting the importance of key information; feature alignment is performed based on the feature matching metric function to ensure the comparability of features from different sources; feature enhancement and dimensionality reduction transformation are used to optimize the expression form of features. In the complex underwater environment of high dams, defect features often show multi-scale and multi-modal characteristics. This embodiment accurately captures defect information at different scales through multi-scale feature extraction; the attention mechanism is used to effectively highlight the discriminative feature components; and the feature matching and enhancement are used to improve the distinguishing ability of features. Especially when processing multi-source heterogeneous data such as sonar, pressure, and turbidity, this embodiment accurately grasps the contribution of different modal data through adaptive adjustment of attention weights; through feature principal component analysis and normalization processing, a compact and discriminative feature expression is constructed, providing a reliable feature basis for subsequent defect identification. This embodiment not only improves the expressiveness of features, but also ensures the stability and reliability of features through multi-level optimization processing, and exhibits excellent feature extraction performance in complex environments.

[0102] In another embodiment of the present application, step S421 is further:

[0103] S4211, obtaining the registered data, dividing the registered data into data blocks of a fixed size, and obtaining a data block sequence; obtaining a pre-stored feature extraction model library, matching the data blocks in the feature extraction model library according to the dimensional features of the data blocks, and obtaining a set of candidate feature extractors;

[0104] S4212, calculating the information gain of each candidate feature extractor on the current data block to obtain a feature extractor scoring sequence; selecting N feature extractors with the highest scores according to the feature extractor scoring sequence to obtain a preferred feature extractor set;

[0105] S4213. Input the data block sequence into the preferred feature extractor set to obtain a local feature sequence; use global average pooling to aggregate the local feature sequence to obtain global features; perform channel attention operation on the local feature sequence and the global feature to obtain the original feature set.

[0106] This embodiment realizes the adaptive optimization selection of feature extractors by introducing a feature extraction model library and an information gain evaluation mechanism. First, model matching is performed based on the dimensional features of the data block, then the optimal feature extractor combination is selected through information gain calculation, and finally the channel attention mechanism is used to fuse local and global features. This embodiment effectively solves the problem of poor adaptability of traditional fixed parameter feature extractors, and improves the efficiency and quality of feature extraction by constructing a set of optimal feature extractors. Especially when processing multimodal data, this embodiment can adaptively select the optimal feature extraction strategy according to data features, providing a high-quality original feature set for subsequent feature fusion.

[0107] In another embodiment of the present application, step S426 is also included, obtaining a fused feature vector, calculating the dimensional distribution of the feature vector, and obtaining dimensional statistical features; obtaining a pre-stored dimensional alignment template, selecting the most suitable alignment scheme according to the dimensional statistical features, and obtaining dimensional transformation parameters; constructing a projection matrix according to the dimensional transformation parameters, projecting the fused feature vector to a unified feature space, and obtaining initial alignment features; calculating the covariance matrix of the initial alignment features, performing eigenvalue decomposition, and obtaining a principal component projection matrix; multiplying the initial alignment features by the principal component projection matrix to obtain a reduced-dimensional feature vector; performing zero-mean and variance normalization processing on the reduced-dimensional feature vector to obtain a standardized feature vector.

[0108] This embodiment achieves a unified expression of feature vectors by introducing a dimensional alignment template and a principal component projection strategy. An alignment scheme is selected based on dimensional statistical features, dimensionality reduction is achieved through projection transformation and principal component analysis, and finally standardized processing is performed to obtain a unified feature expression. This embodiment effectively solves the problem of inconsistent feature dimensions from different sources, and retains the key information of the features through dimensionality reduction processing of principal component projection. In the process of multi-source feature fusion, this embodiment can ensure the consistency of feature expression and provide a standardized data basis for subsequent feature optimization.

[0109] According to one aspect of the present application, step S43 is further:

[0110] S431, obtaining the fused feature vector and stability assessment data, constructing a feature credibility assessment model, and calculating the stability influencing factor; generating a feature weight coefficient according to the stability influencing factor; performing credibility scoring based on the feature weight coefficient and the fused feature vector to obtain credibility weight data;

[0111] S432, sorting the fused feature vectors according to the credibility weight data to obtain a sorted feature list; dividing different feature groups based on the sorted feature list; reconstructing the feature relationship between the feature groups to obtain reconstructed feature relationship data; optimizing the feature combination based on the reconstructed feature relationship data to obtain reconstructed feature data;

[0112] S433, constructing a multi-level network structure based on the reorganized feature data, defining the inter-layer connection relationship of the multi-level network structure; constructing a feature transfer function based on the inter-layer connection relationship; establishing a feedback mechanism based on the feature transfer function and the multi-level network structure, and obtaining a multi-level network model;

[0113] S434, based on the multi-layer network model, extracting the feature expression of each layer, calculating the inter-layer information gain; based on the inter-layer information gain and the feature expression of each layer, fusing the complementary information to obtain the fused feature expression; optimizing the fused feature expression to obtain the inter-layer fusion result;

[0114] S435. Based on the inter-layer fusion result, perform feature optimization to obtain the optimized feature set; based on the optimized feature set, eliminate redundant information to obtain the de-redundant feature set; based on the de-redundant feature set, enhance the discriminant features to obtain the enhanced feature set; based on the enhanced feature set, perform normalization processing to obtain the defect feature data after fusion and optimization.

[0115] In one embodiment of the present application, a feature credibility assessment model C(i) = exp(-λ·E_s)·(1- exp(-μ·SNR)) is constructed; inter-layer information gain G(l) = KL(P_l||P_{l-1}) + η·MI(F_l, F_{l-1}); optimization target L = ∑(||F_l - F_{l-1}|| 2 + α·TV(F_l) + β·R(F_l))·w_l; where E_s is the stability score; SNR is the signal-to-noise ratio; P_l is the feature distribution of the lth layer; KL is the KL divergence; MI is the mutual information; TV is the total variation; R is the regularization term; F_l is the feature of the lth layer; w_l is the layer weight; λ, μ, η, α, β are control parameters.

[0116] This embodiment adopts the optimization method of multi-level feature fusion and credibility assessment to achieve high-reliability defect feature optimization. Specifically, the reliability of the feature is accurately evaluated through the feature credibility assessment model; the feature is reorganized based on the credibility weight to strengthen the contribution of reliable features; the multi-level network structure is used for feature transfer to achieve multi-level optimization of the feature; and the transmission effect of the feature is optimized through inter-layer information gain analysis. In the process of underwater detection of high dams, due to the influence of environmental factors, the reliability of different features often varies. This embodiment accurately identifies the feature components with high reliability through the credibility assessment mechanism; the multi-level network structure is used to achieve layer-by-layer optimization and purification of features; and the introduction of the feedback mechanism improves the stability of feature optimization. Especially when processing multi-source heterogeneous data, this embodiment effectively integrates complementary information from different sources through feature reorganization and multi-layer transfer; and constructs a feature expression with strong discriminative ability through inter-layer fusion and optimization. This embodiment not only improves the reliability of defect features, but also ensures the discriminativeness and stability of features through multi-level optimization processing, which is of great significance to improving the accuracy of defect detection. In complex underwater environments, this embodiment can effectively cope with the impact of data noise and interference, and through credibility evaluation and multi-layer optimization, provides high-quality defect feature expression, laying a solid foundation for the final defect identification.

[0117] like Figure 6 As shown, according to one aspect of the present application, step S5 is further:

[0118] S51, based on the defect feature data after fusion optimization, perform multi-scale feature consistency verification to obtain verified defect feature data;

[0119] S52: Based on the verified defect feature data and the pre-stored spatial reference data, perform spatial mapping and error compensation processing to obtain a final defect detection result.

[0120] In one embodiment of the present application, based on the defect feature data D (x, y, z, t) after fusion optimization, a multi-scale feature consistency check is performed to obtain the verified defect feature D' (x, y, z, t). Based on the verified defect feature D' (x, y, z, t) and the spatial reference data G, an accurate spatial mapping relationship is constructed to obtain the final defect detection result R (x, y, z, t).

[0121] This embodiment achieves high-precision defect detection result output through a combination strategy of multi-scale consistency verification and precise positioning mapping. The multi-scale feature verification mechanism is adopted to effectively reduce the false detection rate; precise positioning is performed in combination with spatial reference data to improve positioning accuracy; and the reliability of the detection results is further improved through error compensation processing. This embodiment shows excellent performance in the complex underwater environment of high dams, and can accurately identify and locate various defects, providing a reliable basis for dam safety assessment. In particular, for the detection of tiny defects, multi-scale analysis and precise positioning are used to improve detection sensitivity and accuracy.

[0122] In another embodiment of the present application, step S1 may also be:

[0123] S1a, obtain real-time water flow velocity data containing spatial position and time information, calculate the median of the local data window according to the historical data of each sampling point, compare the median with the preset threshold, and dynamically adjust the filter window size; calculate a new local median based on the adjusted filter window, and mark the data whose deviation from the original data and the median exceeds the threshold as an outlier; replace the marked outlier with the local median, repeat the above process until the outlier processing is completed, and obtain the pre-processed water flow velocity data.

[0124] S1b. Acquire water pressure data and water depth data containing spatial position and time information, divide the acquisition space into several grid units, and calculate the pressure difference of each grid unit in the three directions of x, y, and z; calculate the pressure change rate in each direction according to the pressure difference of adjacent grid units and the grid spacing; combine the pressure change rates in the three directions of x, y, and z into a vector, normalize the vector, and obtain the pressure gradient characteristic vector.

[0125] S1c, obtain water quality turbidity data containing spatial position and time information, determine the grid point distribution of interpolation calculation; calculate the distance from each grid point to be interpolated to all known turbidity sampling points; calculate the initial turbidity value of each grid point based on the inverse weighted distance; use radial basis function to smooth the initial turbidity value; normalize the smoothed turbidity value to the interval [0, 1] to obtain a continuously distributed turbidity influence coefficient matrix.

[0126] S1d. Obtain the preprocessed water velocity data, pressure gradient characteristic vector and turbidity influence coefficient matrix, and construct weight coefficients of the three characteristic data; multiply the preprocessed water velocity data by the corresponding weight coefficient to obtain the velocity characteristic component; multiply the pressure gradient characteristic vector by the corresponding weight coefficient to obtain the pressure characteristic component; multiply the turbidity influence coefficient matrix by the corresponding weight coefficient to obtain the turbidity characteristic component; superimpose and fuse the three characteristic components, and perform normalization processing to obtain a hydrodynamic field characteristic model that characterizes the hydrodynamic environment.

[0127] This embodiment pre-processes the water velocity data by dynamically adjusting the filter window size, which can effectively adapt to the data characteristics under different flow field conditions. The median value in the local data window is used as a reference, and compared with the preset threshold to identify the outlier, which can accurately capture the sudden water flow changes. Through iterative replacement processing, the gradual convergence of the outlier is ensured, and the excessive smoothing that may be caused by a single processing is avoided. It not only maintains the effective information of the original data, but also effectively removes the measurement noise and interference, and improves the data quality and reliability of subsequent analysis. The acquisition space is divided into grid units and the three-dimensional pressure difference is calculated, so as to achieve an accurate characterization of the water pressure distribution. By calculating the pressure change rate between adjacent grid units, the pressure gradient characteristics of the local area can be accurately reflected. The pressure change rate vector is normalized to eliminate the difference in pressure scales in different directions, so that the pressure gradient feature vector has good directional consistency and numerical stability, and provides a reliable description of the pressure field characteristics for subsequent multimodal data fusion. The interpolation method based on inverse distance weighting is adopted to achieve continuous processing of discrete turbidity data. The radial basis function is used for smoothing, which effectively eliminates local oscillations in the interpolation process and ensures the physical rationality of the turbidity distribution. By normalizing the processed turbidity values ​​to the interval [0, 1], a unified metric is established, so that the turbidity influence coefficient matrix can be directly used in the subsequent feature fusion process, improving the robustness of data processing. By constructing the weight coefficients of the three characteristic data, the adaptive fusion of water velocity, pressure gradient and turbidity characteristics is achieved. The product operation of each characteristic component and the corresponding weight coefficient ensures the balance of contribution of different physical quantities in the fusion process. The final hydrodynamic field characteristic model not only includes the motion characteristics of the flow field, but also reflects the pressure distribution and medium characteristics, providing a complete feature description framework for the accurate characterization of the underwater environment.

[0128] In another embodiment of the present application, step S2 may also be:

[0129] S2a, obtain the hydrodynamic field characteristic model and the predetermined detection target area data, and divide the detection area into multiple sub-area grids; calculate the heuristic cost function value of each grid point, which is composed of the actual cost from the point to the starting point and the estimated cost from the point to the target point; add the starting point to the open list, and traverse the adjacent points of the grid point with the smallest cost in the open list; calculate the cost function values ​​of the adjacent points and update the open list and the closed list; repeat the above process until the target point is found, backtrack to obtain the path point set, convert the path point set into time series data, and obtain the initial trajectory point sequence.

[0130] S2b. Obtain the propeller configuration, center of gravity position, moment of inertia and other dynamic parameter data and hydrodynamic field characteristic model of the submersible, and calculate the maximum available thrust and torque at each spatial position based on the submersible dynamic equation; consider the disturbance force and disturbance torque of the hydrodynamic field on the submersible, and calculate the actually available control force and control torque; according to the constraint range of the control force and control torque, combined with the motion acceleration limit, construct a complete set of dynamic constraint equations to obtain a set of dynamic constraint conditions.

[0131] S2c, obtain the initial trajectory point sequence and the set of dynamic constraints, construct the state equation in the prediction time domain; set the weighted matrix of trajectory tracking error, control quantity and control increment; construct the target optimization function based on the state equation and the weighted matrix; solve the optimization problem under the dynamic constraints to obtain the optimal control sequence in the prediction time domain; extract the first control quantity from the optimal control sequence and execute it, and loop the rolling optimization process to finally obtain the optimized trajectory parameters.

[0132] S2d. Obtain the optimized trajectory parameters and hydrodynamic field characteristic model, calculate the hydrodynamic disturbance intensity at the trajectory point; calculate the attitude fluctuation range of the submersible based on the disturbance intensity; calculate the coverage range change of the detection sensor according to the attitude fluctuation; comprehensively consider the attitude stability index and the detection coverage stability index to obtain the stability evaluation data.

[0133] This embodiment can find the optimal detection path in a complex hydrodynamic environment through a path planning method based on a heuristic cost function. The evaluation method that combines the actual cost with the estimated cost effectively balances the efficiency and accuracy of the path search. By converting the time sequence of the path point set, the initial trajectory point sequence obtained not only meets the spatial continuity requirements, but also takes into account the time constraints, providing a reasonable initial solution for the subsequent trajectory optimization. By considering the dynamic parameters and hydrodynamic field characteristics of the submersible, an accurate estimation of the actual available control force and control torque is achieved. The complete dynamic constraint equation group constructed in combination with the motion acceleration limit accurately describes the motion characteristics of the submersible in a complex water environment. This set of constraints based on the physical model ensures the executability of the subsequent trajectory optimization results and improves the reliability of motion control. The model predictive control method is used for trajectory optimization to achieve forward-looking control of the submersible motion. By setting a reasonable weighting matrix, the trajectory tracking accuracy and control cost are balanced. Based on the rolling optimization strategy, the real-time nature of the control sequence is guaranteed and the robustness of the optimization results is improved. The optimized trajectory parameters finally obtained not only meet the dynamic constraints, but also have good control performance. The trajectory stability is evaluated by combining the hydrodynamic field characteristic model, which realizes a comprehensive analysis of the detection quality. By calculating the attitude fluctuation range and the change in the detection sensor coverage range, the impact of external disturbances on the detection effect is accurately reflected. Taking into account the attitude stability and detection coverage stability, the evaluation data obtained provides a reliable basis for the optimization of the detection scheme.

[0134] In another embodiment of the present application, step S3 may also be:

[0135] S3a. Obtain the turbidity influence coefficient matrix and the current position data of the submersible, and extract the local turbidity value from the turbidity influence coefficient matrix according to the current position; calculate the attenuation coefficient of the sound wave in the current environment based on the local turbidity value; adjust the sonar transmission power in combination with the attenuation coefficient; calculate the optimal receiving gain value according to the transmission power and the expected echo intensity; quantify the calculated transmission power and receiving gain, and establish a parameter lookup table; combine the transmission power, receiving gain and the corresponding spatial position information to obtain the sonar acquisition parameter set.

[0136] S3b, obtain the sonar acquisition parameter set and the optimized trajectory parameters, extract the corresponding transmission power and receiving gain parameters from the sonar acquisition parameter set according to the current trajectory point position; use the extracted parameters to collect sonar data to obtain original sonar echo data; perform time gain compensation and noise suppression processing on the original sonar echo data to obtain enhanced sonar data; dynamically adjust the scanning angle and scanning range of subsequent acquisition points according to the target echo characteristics in the sonar data; associate and bind the processed sonar data with the acquisition parameters, location information and timestamp to obtain collaboratively enhanced multimodal data.

[0137] This embodiment realizes the adaptation of the sonar system to complex water environments by dynamically adjusting the sonar parameters based on the local turbidity value. The transmit power and receive gain are optimized by calculating the sound wave attenuation coefficient, ensuring the effective acquisition of the sonar signal. The established parameter lookup table simplifies the real-time parameter adjustment process and improves the working efficiency and data acquisition quality of the sonar system. The coordinated optimization of sonar parameters and acquisition trajectories is achieved, and the quality of sonar data is improved. The sonar echo signal is effectively enhanced through time gain compensation and noise suppression processing. The scanning parameters are dynamically adjusted based on the target echo characteristics to improve the pertinence of data acquisition. The correlation binding of multimodal data provides complete information support for subsequent data analysis.

[0138] In another embodiment of the present application, step S4 may also be:

[0139] S4a, obtain collaboratively enhanced multimodal data, extract timestamp and position information from the data; establish an extended Kalman filter state equation, where the state vector contains position, attitude and velocity information; calculate state prediction values ​​based on the timestamp sequence; calculate observation prediction values ​​based on the position information; calculate the Kalman gain and update the state estimate; interpolate and correct the multimodal data based on the updated state estimate; establish a spatiotemporal mapping relationship for the synchronized data to obtain the registered data.

[0140] S4b, obtain the registered data and the pre-stored historical detection data, extract the local feature descriptor from the registered data; calculate the attention weight matrix, which represents the correlation between different features; perform weighted combination of local features according to the attention weight matrix; perform similarity matching between the weighted features and the features in the historical detection data; filter and enhance key features based on the matching results; perform dimensionality reduction and normalization processing on the enhanced features to obtain a fused feature vector.

[0141] S4c, obtain fused feature vectors and stability assessment data, calculate feature credibility weights based on the stability assessment data; reorganize feature vectors according to credibility weights; construct a multi-level fusion network, which includes a feature layer, an attribute layer, and a decision layer; group and cluster the reorganized features at the feature layer; extract statistical properties of feature groups at the attribute layer; integrate information from each layer at the decision layer to obtain fused and optimized defect feature data.

[0142] This embodiment uses an extended Kalman filter to achieve spatiotemporal synchronization of multimodal data, effectively solving the problem of timing differences in data from different sensors. The accuracy of position estimation is improved through the recursive process of state prediction and observation update. Interpolation correction based on the spatiotemporal mapping relationship ensures the consistency and reliability of the data after registration. Adaptive weighted fusion of features is achieved through the attention mechanism, improving the effectiveness of feature extraction. The feature screening method based on similarity matching ensures the accurate identification of key features. Feature dimensionality reduction and normalization processing not only reduce data redundancy, but also maintain the distinctiveness of features, providing high-quality feature representation for subsequent defect detection. The feature recombination method based on stability assessment data improves the reliability of feature representation. The hierarchical processing strategy of the multi-level fusion network realizes step-by-step abstraction from bottom-level features to high-level semantics. Feature grouping, clustering and statistical attribute extraction effectively capture the essential characteristics of defect features and improve the accuracy of detection results.

[0143] In another embodiment of the present application, step S5 may also be:

[0144] S5a. Obtain defect feature data after fusion optimization, decompose the feature data at different scales, and obtain feature subsets at multiple scale levels; calculate local statistical indicators of features at each scale level, including mean, variance, and skewness; construct a feature consistency evaluation function based on statistical indicators; calculate feature correlation indicators between adjacent scale levels; determine the feature consistency threshold based on the correlation indicator; mark features below the threshold as noise and eliminate them; rescale the remaining features to obtain verified defect feature data.

[0145] S5b, obtain the verified defect feature data and the pre-stored spatial reference data, extract the coordinate control point information from the spatial reference data; establish the coordinate transformation matrix from the feature data to the actual space; calculate the position deviation of the feature points in different sensor coordinate systems; optimize the coordinate transformation parameters based on the least squares method; use the optimized transformation parameters to spatially reproject the feature points; calculate the reprojection error and perform error compensation; associate the compensated feature point positions with the corresponding feature attributes; spatially cluster the association results to obtain the final defect detection results.

[0146] This embodiment realizes multi-level analysis of features through a multi-scale decomposition method, thereby improving the comprehensiveness of feature verification. The noise features are effectively identified based on the consistency evaluation of statistical indicators. The feature correlation analysis between adjacent scales ensures the accuracy of feature reconstruction. Through threshold screening and scale reconstruction, the verified features obtained have high reliability. Based on the coordinate transformation of spatial reference data, accurate mapping of feature data to actual space is achieved. The least squares optimization and error compensation mechanism improve the accuracy of spatial positioning. The correlation analysis of feature point positions and attributes realizes the accurate positioning and feature description of defects. Spatial clustering processing ensures the consistency and reliability of the detection results.

[0147] The present invention adopts an adaptive median filtering algorithm to address the limitations of fixed window filtering in water velocity data preprocessing. In the specific implementation, the data is first segmented by time window and spatial position to calculate the spatiotemporal distribution characteristics of the data; then an adaptive threshold function is constructed based on statistical characteristics to dynamically adjust the window size; then an outlier is identified through a data distribution model of the local median; and finally an exponential weighted average method is used for smoothing. This multi-level adaptive processing method effectively solves the problem of data preprocessing under sudden water flow disturbances. In view of the problem of ignoring the influence of water depth in pressure gradient feature extraction, a water depth compensation mechanism is introduced. In the specific implementation, a depth correction coefficient matrix is ​​first constructed to perform a preliminary correction on the water pressure data; then the pressure gradient components in three directions are calculated respectively; then a secondary correction is performed through the depth compensation model parameters; and finally a normalization process is performed to obtain a standardized pressure gradient feature vector. This dual correction mechanism effectively eliminates the pressure measurement deviation caused by water depth changes. In view of the accuracy problem of turbidity spatial interpolation, a multi-level spatial interpolation method is constructed. In the specific implementation, the spatial reference frame is first established through dynamic meshing; then the spatial weight matrix is ​​constructed based on the Gaussian weight function; then the interpolation parameters are optimized through the error evaluation function; and finally the radial basis function is used for smoothing. This refined interpolation method accurately captures the nonlinear distribution characteristics of turbidity. In view of the problem of attitude stability in trajectory planning, a complete stability evaluation system is constructed. In the specific implementation, the trajectory is first optimized through the model predictive control method; then the hydrodynamic disturbance characteristics are analyzed to evaluate the attitude stability and detection coverage integrity; finally, a comprehensive stability evaluation result is obtained through multi-index evaluation. This trajectory planning method considering the influence of hydrodynamics improves the stability of the detection process. In view of the reliability problem in feature fusion, a feature credibility evaluation and multi-layer optimization mechanism are established. In the specific implementation, the important features are first highlighted through the attention mechanism; then the feature credibility evaluation model is established, and the features are reorganized based on the stability evaluation data; finally, the features are optimized through a multi-layer network structure. This credibility-based feature fusion method effectively improves the reliability of the fusion results. In view of the problem of inconsistent dimensions of multi-source data, a dimension alignment processing step is added. In the specific implementation, the optimal alignment scheme is first selected based on the dimensional statistical features; then the dimension is transformed through the projection matrix; the dimension reduction optimization is performed through principal component analysis; and finally the standardization is performed. This systematic dimensional alignment method provides a unified data foundation for feature fusion.

[0148] The present invention constructs a complete high dam underwater defect detection system through the organic combination of multimodal data fusion, adaptive parameter optimization, collaborative enhancement processing, multi-level feature fusion and multi-scale verification positioning. In the data acquisition stage, multimodal sensor fusion is used to overcome the limitations of a single sensor; in the trajectory planning stage, the reliability of the detection process is ensured by dynamic constraints and stability evaluation; in the data processing stage, multi-level fusion and verification mechanisms are used to improve the detection accuracy. The present invention achieves high-precision detection of dam body defects under complex environmental conditions such as strong water flow and high turbidity underwater in high dams, providing strong technical support for the safe operation of reservoir dams. It not only solves the problems of poor reliability and low precision of traditional detection methods in complex environments, but also improves detection efficiency through a systematic data processing process, and has strong engineering practical value.

[0149] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A high dam underwater defect measurement method based on multimodal data fusion, characterized in that: The steps include: S1, obtain real-time water flow velocity data, water pressure data and water turbidity data, use adaptive median filter algorithm for preprocessing, and build a hydrodynamic field characteristic model; S2, generating an initial trajectory point sequence based on the hydrodynamic field characteristic model and predetermined detection target area data; Combine the initial trajectory point sequence and the preset dynamic constraint condition set to perform trajectory optimization and stability evaluation to obtain optimized trajectory parameters and stability evaluation data; S3. Based on the water turbidity data and the optimized trajectory parameters, the sonar transmission power and gain parameters are calculated to obtain the sonar acquisition parameter set; according to the sonar acquisition parameter set and the optimized trajectory parameters, collaborative data collection and processing are performed to obtain collaboratively enhanced multimodal data; S4, performing spatiotemporal alignment processing on the multimodal data to obtain the registered data; Based on the registered data, stability evaluation data and pre-stored historical detection data, the attention mechanism is used to perform feature fusion and optimization to obtain the fused and optimized defect feature data; S5. Perform multi-scale consistency verification and precise positioning mapping on the defect feature data to obtain the final defect detection result; Step S1 is further as follows: S11, obtaining real-time water flow velocity data including spatial position and time information; using an adaptive median filter algorithm to process outliers on the real-time water flow velocity data to obtain pre-processed water flow velocity data; S12, obtaining water pressure data including spatial position and time information; based on the water pressure data, calculating the rate of change of pressure in three-dimensional space to obtain a pressure gradient characteristic vector; S13, obtaining water quality turbidity data including spatial position and time information; performing spatial interpolation processing on the water quality turbidity data to obtain a continuously distributed turbidity influence coefficient matrix; S14, performing weighted fusion processing on the pre-processed water flow velocity data, pressure gradient characteristic vector and turbidity influence coefficient matrix to obtain a hydrodynamic field characteristic model that characterizes the hydrodynamic environment.

2. The method for measuring underwater defects of high dams based on multimodal data fusion according to claim 1 is characterized in that: Step S2 is further as follows: S21, based on the hydrodynamic field characteristic model and the predetermined detection target area data, a path sequence is generated by using a heuristic search algorithm to obtain an initial trajectory point sequence; S22, obtaining dynamic parameter data of the submersible, and calculating control constraints using dynamic equations based on the dynamic parameter data and a hydrodynamic field characteristic model to obtain a set of dynamic constraint conditions; S23, based on the initial trajectory point sequence and the set of dynamic constraint conditions, using the model predictive control method to perform optimization calculation to obtain optimized trajectory parameters; S24. Based on the optimized trajectory parameters and the hydrodynamic field characteristic model, the trajectory stability index is calculated to obtain stability evaluation data.

3. The method for measuring underwater defects of high dams based on multimodal data fusion according to claim 2 is characterized in that: Step S3 is further as follows: S31, obtaining the current position data of the submersible, and calculating the transmission power and gain parameters of the sonar based on the current position data and the turbidity influence coefficient matrix to obtain a sonar acquisition parameter set; S32. Based on the sonar acquisition parameter set and the optimized trajectory parameters, sonar data is collected and enhanced to obtain collaboratively enhanced multimodal data.

4. The method for measuring underwater defects of high dams based on multimodal data fusion according to claim 3 is characterized in that: Step S4 is further as follows: S41, based on the multimodal data, using extended Kalman filtering to perform spatiotemporal alignment processing to obtain registered data; S42, obtaining the registered data and the pre-stored historical detection data, and using the attention mechanism to perform feature fusion to obtain a fused feature vector; S43. Based on the fused feature vector and the stability assessment data, multi-level fusion optimization processing is performed to obtain defect feature data after fusion optimization.

5. The method for measuring underwater defects of high dams based on multimodal data fusion according to claim 4 is characterized in that: Step S5 is further as follows: S51, based on the defect feature data after fusion optimization, perform multi-scale feature consistency verification to obtain verified defect feature data; S52: Based on the verified defect feature data and the pre-stored spatial reference data, perform spatial mapping and error compensation processing to obtain a final defect detection result.

6. The method for measuring underwater defects of high dams based on multimodal data fusion according to claim 5 is characterized in that: Step S11 is further as follows: S111, acquiring real-time water flow velocity data, and segmenting a data sequence of the real-time water flow velocity data according to a preset time window and a preset spatial position to obtain segmented data; Based on the segmented data, the time interval and spatial distribution characteristics of each segment of data are calculated to obtain the calculation results; According to the calculation results, the initial window size and sliding step length are determined; based on the initial window size and sliding step length, the segmented data is divided into blocks using the sliding window method to obtain a data window sequence; S112. Based on the data window sequence, the mean, standard deviation and interquartile range of the data in each window are calculated to obtain statistical characteristics; based on the statistical characteristics, an adaptive threshold function is constructed to calculate the adjustment coefficient of the window size; based on the adjustment coefficient, the window size is dynamically adjusted to obtain an optimized window sequence; S113, based on the optimized window sequence, calculating the local median of the data in each optimized window; constructing a data distribution model based on the local median; calculating the deviation between the data point and the data distribution model to obtain a deviation value; comparing the deviation value with a preset dynamic threshold, identifying abnormal data points, and obtaining an abnormal value marker sequence; S114. Based on the outlier marker sequence, extract the outlier data points and the corresponding real-time water flow velocity data; based on the outlier data points and the corresponding real-time water flow velocity data, calculate the replacement value through the data distribution model; use the exponential weighted average method to smooth the replacement value to obtain the smoothed replacement value; based on the smoothed replacement value, iteratively execute the outlier replacement process until the preset convergence condition is met to obtain the pre-processed water flow velocity data.

7. The method for measuring underwater defects of high dams based on multimodal data fusion according to claim 5 is characterized in that: Step S12 is further as follows: S121, obtaining water pressure data and water depth data including spatial position and time information; constructing a depth correction coefficient matrix according to the water depth data; performing a product operation on the water pressure data and the depth correction coefficient matrix to obtain corrected water pressure data; S122, calculating the pressure change rate in the X-axis direction based on the corrected water pressure data to obtain the X-axis pressure gradient component; Based on the corrected water pressure data, the pressure change rate in the Y-axis direction is calculated to obtain the Y-axis pressure gradient component; based on the corrected water pressure data, the pressure change rate in the Z-axis direction is calculated to obtain the Z-axis pressure gradient component; the X-axis pressure gradient component, the Y-axis pressure gradient component and the Z-axis pressure gradient component are vector-combined to obtain an initial pressure gradient vector; S123. Based on the water depth data, interpolation calculation is performed in the pre-stored depth compensation model parameters to obtain a depth compensation coefficient; the initial pressure gradient vector is multiplied by the depth compensation coefficient to obtain a compensated pressure gradient vector; and the compensated pressure gradient vector is normalized to obtain a pressure gradient characteristic vector.

8. The method for measuring underwater defects of high dams based on multimodal data fusion according to claim 5 is characterized in that: Step S42 is further as follows: S421, constructing a multi-scale feature extractor based on the registered data; using the multi-scale feature extractor, calculating a local feature descriptor; extracting global context information of the registered data, combining the local feature descriptor and the global context information, and obtaining an original feature set; S422, based on the original feature set, construct an attention query matrix and calculate the correlation scores between the features; Generate an attention mask based on the relevance score; Based on the original feature set and the attention mask, perform the attention weighting operation to obtain the attention weight data; S423, obtaining historical detection data, and constructing a feature matching metric function based on the historical detection data and attention weight data; Based on the feature matching metric function, the similarity between feature pairs is calculated to obtain a similarity score; Based on the similarity score, the nearest neighbor matching algorithm is used to obtain the nearest neighbor matching result; Based on the nearest neighbor matching results, reliable matching pairs are screened to obtain feature matching results; S424, based on the feature matching result, analyzing the discrimination performance of the feature to obtain a performance analysis result; According to the performance analysis result, the key characteristic component is enhanced to obtain the enhanced key characteristic component; Based on the performance analysis results, redundant features are identified and suppressed to obtain a suppressed feature set; Based on the enhanced key feature components and the suppressed feature set, the feature expression is optimized to obtain enhanced feature data; S425, calculating feature principal components based on the enhanced feature data; According to the variance contribution rate of the feature principal component, the optimal feature subspace is selected; the enhanced feature data is projected onto the optimal feature subspace, and a dimensionality reduction transformation is performed to obtain the feature data after dimensionality reduction; The feature data after dimension reduction is normalized to obtain the fused feature vector.

9. The method for measuring underwater defects of high dams based on multimodal data fusion according to claim 5 is characterized in that: Step S43 is further as follows: S431, obtaining fused feature vectors and stability assessment data, constructing a feature credibility assessment model, and calculating stability influencing factors; generating feature weight coefficients according to the stability influencing factors; Based on the feature weight coefficient and the fused feature vector, credibility scoring is performed to obtain credibility weight data; S432, sorting the fused feature vectors according to the credibility weight data to obtain a sorted feature list; dividing different feature groups based on the sorted feature list; reconstructing the feature relationship between the feature groups to obtain reconstructed feature relationship data; optimizing the feature combination based on the reconstructed feature relationship data to obtain reconstructed feature data; S433, constructing a multi-level network structure based on the reorganized feature data, defining the inter-layer connection relationship of the multi-level network structure; constructing a feature transfer function based on the inter-layer connection relationship; Based on the characteristic transfer function and multi-level network structure, a feedback mechanism is established to obtain a multi-level network model; S434, based on the multi-layer network model, extracting the feature expression of each layer, calculating the inter-layer information gain; based on the inter-layer information gain and the feature expression of each layer, fusing the complementary information to obtain the fused feature expression; optimizing the fused feature expression to obtain the inter-layer fusion result; S435, based on the inter-layer fusion result, perform feature optimization to obtain an optimized feature set; based on the optimized feature set, eliminate redundant information to obtain a de-redundant feature set; Based on the feature set after removing redundancy, the discriminant features are enhanced to obtain an enhanced feature set; Based on the enhanced feature set, normalization processing is performed to obtain the fused and optimized defect feature data.

Images