Unmanned aerial vehicle-based concrete structure safety quantitative nondestructive intelligent monitoring method and system

Abstract

This invention relates to the field of non-destructive intelligent monitoring technology for concrete structures, and discloses a method and system for non-destructive intelligent monitoring of concrete structure safety quantification based on unmanned aerial vehicles (UAVs). The method controls a spherical UAV equipped with a visual sensor to perform a full-area cruise scan of the concrete structure surface, and uses image processing algorithms to intelligently identify corrosion areas. Based on the identification results, a monitoring point sequence is planned, and a dynamic flight path is generated. At each monitoring point, the UAV adjusts its attitude so that at least two spring sensors simultaneously and effectively contact the concrete surface, forming a closed electrical loop. Multi-frequency electrical excitation signals are applied through the electrical loop, and impedance spectra are measured to obtain corrosion quantification parameters, which are then transmitted to the control center in real time. This achieves fully automated monitoring of concrete structure corrosion, improving detection efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of non-destructive intelligent monitoring technology for concrete structures, specifically to a method and system for quantitative non-destructive intelligent monitoring of concrete structure safety based on unmanned aerial vehicles (UAVs). Background Technology

[0002] Concrete structures, as a core component of modern infrastructure, are widely used in bridges, high-rise buildings, tunnels, and other engineering projects. However, long-term exposure to harsh environments such as moisture and salt can easily cause the steel reinforcement inside the concrete to corrode, leading to a decrease in the structural load-bearing capacity. Therefore, regular safety monitoring of concrete structures is crucial. Currently, traditional testing methods mainly rely on manual labor, requiring inspectors to use scaffolding, suspended platforms, or ropes to approach the structural surface and visually observe or measure the corrosion using instruments.

[0003] This method has significant shortcomings: for some concrete structures, such as tall bridge piers or large-span roofs, manual inspection is difficult to reach, creating numerous blind spots; high-altitude operations are risky, inefficient, time-consuming, and labor-intensive, and the results are influenced by subjective experience, lacking quantitative standards. Although non-destructive testing technologies such as ultrasonic or radar have been applied, manual operation of the equipment is still required, preventing large-scale automation. In recent years, drone technology has provided a new approach to structural monitoring, but existing drones are mostly limited to image acquisition and cannot deeply quantify corrosion parameters. Therefore, developing an automated and precise non-destructive testing method has become an urgent need for the industry. Summary of the Invention

[0004] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a method and system for quantitative and non-destructive intelligent monitoring of concrete structure safety based on unmanned aerial vehicles (UAVs), thereby solving the problems of low efficiency and high risk in the monitoring of high-altitude or large-span concrete structures in the prior art.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0006] In a first aspect, the present invention provides a method for quantitative and non-destructive intelligent monitoring of the safety of concrete structures based on unmanned aerial vehicles (UAVs), comprising:

[0007] After the spherical drone takes off, it uses its onboard visual sensors to perform a full-area cruise scan of the concrete structure surface, acquires surface image data, and identifies potential corrosion areas through image processing algorithms.

[0008] Based on the identification results, plan the sequence of monitoring points;

[0009] A dynamic flight path is generated based on the distribution of monitoring points. The spherical UAV is controlled to fly along the path and adjust its attitude when it arrives at the monitoring point so that no less than two spring sensors can simultaneously and effectively contact the concrete surface to form a closed electrical circuit.

[0010] By applying multi-frequency electrical excitation signals through an electrical circuit, the impedance spectrum and potential response are measured to obtain quantification parameters of corrosion.

[0011] The rust quantification parameters are transmitted in real time to the control center via a wireless module for analysis.

[0012] Preferably, in one possible implementation of the first aspect, identifying potential rust areas includes:

[0013] Multi-view surface image data acquired by a spherical UAV visual sensor is used to perform image sequence registration to compensate for view distortion caused by changes in UAV flight attitude.

[0014] The image sequence registration uses a feature point matching algorithm to extract SIFT feature descriptors and perform RANSAC optimization to eliminate mismatches.

[0015] The registered images are fused at multiple scales to generate a high-resolution panoramic image. The fusion weights are adaptively calculated based on the local contrast and signal-to-noise ratio of the images.

[0016] A deep learning segmentation model is used to classify panoramic images at the pixel level. The deep learning segmentation model is a variant of U-Net based on the attention mechanism. Its encoder part introduces an SE module to enhance feature representation, and the decoder part uses dilated convolution to expand the receptive field and identify the boundaries of the rusted area.

[0017] Preferably, in one possible implementation of the first aspect, the RANSAC optimization process used in the image sequence registration is used to calculate the fundamental matrix and estimate the interior point set, and its objective function is:

[0018]

[0019] in Based on the matrix, For the estimated fundamental matrix, and It is the first A pair of matching feature points, For transpose operation, It is an indicator function. Distance threshold;

[0020] The weight calculation for the multi-scale fusion adopts Laplace pyramid decomposition, and the weight formula for each layer is as follows:

[0021]

[0022] in It is the first Layer in position The fusion weight, It is the image gradient. For A local neighborhood window centered on the center, Therefore Points within a local neighborhood window centered on the center. It is a smoothing constant.

[0023] Preferably, in one possible implementation of the first aspect, the algorithmic structure of the deep learning segmentation model includes:

[0024] The encoder consists of multiple residual blocks. The output feature map of each residual block is recalibrated through a channel attention mechanism, the formula of which is:

[0025]

[0026] in It is the first Output feature map after channel calibration It is the first Input feature maps of each channel, yes Scalar features after global average pooling. and Here are the weight matrices for the two fully connected layers. It is the ReLU activation function. It is the sigmoid function. This represents element-wise multiplication;

[0027] The decoder part uses transposed convolutions to upsample the deep feature maps, and then uses the upsampled feature maps... The same resolution feature map passed from the corresponding layer of the encoder through skip connections The features are concatenated along the channel dimension to obtain the fused feature map. :

[0028]

[0029] in, This is the function for concatenation operations;

[0030] For the fused feature map Sequential depthwise separable convolution processing is performed, specifically including:

[0031] right Each input channel is processed independently using depthwise convolution;

[0032] Pointwise convolution is used to linearly combine the outputs of the depthwise convolution along the channel dimension to generate the output feature map of the current layer of the decoder. Its operation is represented as:

[0033]

[0034] in It is the decoder output. Represents depthwise convolution. This represents pointwise convolution.

[0035] Preferably, in one possible implementation of the first aspect, the planned monitoring point sequence includes:

[0036] Based on the identified potential corrosion areas, a corrosion risk index is calculated for each area. This corrosion risk index integrates the area area, corrosion confidence level, historical monitoring data, and environmental factors, using a weighted summation model.

[0037]

[0038] in It is a region The risk index, It is a region normalized area, It is a region The confidence level of corrosion It is a region Historical data scores, It is a region Environmental factors, , , , These are the weighting coefficients, determined through principal component analysis;

[0039] Using risk indices as priorities, a genetic algorithm is used to solve for the monitoring point sequence. The objective function is a weighted combination of maximizing total risk coverage and minimizing path length, expressed as:

[0040]

[0041] in It is a binary variable representation area Whether it is overwritten, It is a point Time Euclidean distance, These are path decision variables. It is a punishment factor;

[0042] The optimal monitoring point sequence is generated iteratively.

[0043] Preferably, in one possible implementation of the first aspect, generating the dynamic flight path includes:

[0044] Based on the monitoring point sequence, a three-dimensional spatial path map is constructed, where nodes are the coordinates of the monitoring points, and edge weights take into account flight energy consumption and obstacle risks.

[0045] The initial shortest path is calculated using Dijkstra's algorithm, and then dynamically adjusted using model prediction control. The prediction model is as follows:

[0046]

[0047] in It is a state vector. , , This indicates the position coordinates of the drone in three-dimensional space. , , Indicates the velocity components in the corresponding directions. The derivative of the state vector. A function describing the dynamic characteristics of an unmanned aerial vehicle (UAV). It controls the input vector;

[0048] The optimization problem is expressed as:

[0049]

[0050] in It is a reference trajectory. and It is a positive definite weight matrix. and It is the weighted Euclidean norm. It predicts the time domain. Represents a time variable.

[0051] Preferably, in one possible implementation of the first aspect, the attitude adjustment utilizes airborne IMU and visual odometry data, represents the current attitude using quaternions, and determines the target attitude by the surface normal vector of the monitoring point.

[0052] The adjustment process uses a sliding mode control law:

[0053]

[0054] in It is the control torque. It is the gain matrix. For symbolic functions, It is a sliding surface, defined as a linear combination of attitude error quaternions:

[0055]

[0056] in It is a three-dimensional vector representing the difference between the current angular velocity and the target angular velocity. It is a positive definite diagonal matrix. The three-dimensional attitude error vector is obtained by converting the error quaternions calculated from the current attitude quaternion and the target attitude quaternion.

[0057] Synchronous and effective contact with the concrete surface is achieved by controlling the extension and retraction of the spring sensor, with the amount of extension and retraction adjusted based on the laser rangefinder readings.

[0058]

[0059] in It refers to the change in the length of the telescoping. This is the current distance. This is a reference contact distance. and These are the proportional and derivative control coefficients, respectively.

[0060] Preferably, in one possible implementation of the first aspect, determining whether the posture meets the contact requirement includes:

[0061] Calculate the contact force vector between the tip of each spring sensor and the concrete surface;

[0062] Contact force estimation is based on spring deformation and Hooke's law:

[0063]

[0064] in It is the first The contact force vector of a spring sensor. It is the spring stiffness coefficient. It is deformation displacement. For the first A unit direction vector pointing axially from the spring to the concrete surface;

[0065] The determination is made by calculating the cosine of the angle between the resultant force vector and the surface normal vector:

[0066]

[0067] in It is the deviation angle. It is the surface normal vector. It is a threshold; if it is not met, the attitude is readjusted.

[0068] Preferably, in one possible implementation of the first aspect, obtaining the rust quantification parameters includes:

[0069] Impedance spectrum data were measured by applying multi-frequency sinusoidal electrical excitation signals through an electrical circuit, with a frequency range from 0.1 Hz to 10 kHz.

[0070] The impedance is fitted using an equivalent circuit model, which is an improved Randle circuit that includes diffused impedance elements. Its impedance expression is as follows:

[0071]

[0072] in It is an electrolyte resistor. It is a double-layer capacitor. It is a charge transfer resistor. It is the Warburg diffusion impedance. It is the diffusion coefficient. It is angular frequency. The imaginary unit;

[0073] Then, the polarization resistance is calculated according to the Stern-Geary equation. Corrosion current density The derivation is as follows:

[0074]

[0075] Where B is the Stern-Geary constant;

[0076] Corrosion depth prediction uses Faraday's law:

[0077]

[0078] in It is the depth of corrosion. It is the molar mass of the steel reinforcement. It refers to the time it takes for the steel bars to corrode. It is the number of electrons transferred. It is Faraday's constant. It refers to the density of the reinforcing steel.

[0079] Predicting the corrosion section loss rate based on corrosion depth:

[0080]

[0081] in, Indicates the corrosion cross-sectional loss rate. This represents the initial cross-sectional area of ​​the reinforcing steel bars in a concrete structure. Let be the initial radius of the reinforcing bar. = This represents the remaining effective cross-sectional area and remaining effective radius of the reinforcing steel in a concrete structure after corrosion. .

[0082] Secondly, the present invention provides a UAV-based quantitative and non-destructive intelligent monitoring system for the safety of concrete structures, the system being used to implement the UAV-based quantitative and non-destructive intelligent monitoring method for the safety of concrete structures as described in the first aspect, comprising:

[0083] The visual sensor module uses a multi-view camera array mounted on a spherical drone to collect image data of the concrete structure surface;

[0084] The image processing module integrates a GPU accelerator to execute image registration, fusion, and deep learning segmentation algorithms to identify rusted areas.

[0085] The planning module, which includes a risk calculation unit and a path optimizer, is used to generate monitoring point sequences and dynamic flight paths;

[0086] The control module, which includes an attitude controller and a spring sensor driver, is used to adjust the drone's attitude and ensure synchronized contact.

[0087] The electrical measurement module, equipped with a multi-frequency signal generator and an impedance analyzer, is used to apply electrical excitation and measure the impedance spectrum;

[0088] The wireless transmission module uses the 5G communication protocol to send corrosion quantification parameters to the control center in real time.

[0089] The beneficial effects of this invention are: it enables fully automated monitoring of concrete structure corrosion through a spherical drone, effectively overcoming the limitations of manual inspection.

[0090] The drone is equipped with a visual sensor to perform full-area scanning, and uses image registration and deep learning algorithms to intelligently identify rusted areas, avoiding the risks of high-altitude operations and improving detection efficiency.

[0091] The monitoring point sequence is optimized based on the risk index, and dynamic path planning ensures coverage of key areas and reduces redundant flights.

[0092] Adjust the drone's attitude to ensure that at least two spring sensors make simultaneous and effective contact with the concrete surface, forming a stable electrical circuit. By measuring the impedance spectrum through multi-frequency electrical excitation, quantitative parameters such as corrosion current density, corrosion depth, and corrosion cross-sectional loss rate can be obtained non-destructively, improving the accuracy and comparability of the data.

[0093] Data is sent to the control center in real time, enabling remote analysis and early warning, and shortening the monitoring cycle.

[0094] The entire system requires no manual intervention, reducing safety risks, and non-destructive testing protects structural integrity. This invention provides reliable technical support for the health management of concrete structures and extends the service life of infrastructure. Attached Figure Description

[0095] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0096] Figure 1 This application provides a flowchart of a method for quantitative and non-destructive intelligent monitoring of the safety of concrete structures based on unmanned aerial vehicles (UAVs).

[0097] Figure 2 This application provides a structural diagram of a non-destructive intelligent monitoring system for the safety quantification of concrete structures based on unmanned aerial vehicles (UAVs).

[0098] Reference numerals: 1-Vision sensor module, 2-Image processing module, 3-Planning module, 4-Control module, 5-Electrical measurement module, 6-Wireless transmission module. Detailed Implementation

[0099] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0100] Example 1: As Figure 1 As shown, this invention provides a non-destructive intelligent monitoring method for the safety quantification of concrete structures based on unmanned aerial vehicles (UAVs), comprising:

[0101] After the spherical drone takes off, it uses its onboard visual sensors to perform a full-area cruise scan of the concrete structure surface, acquires surface image data, and identifies potential corrosion areas through image processing algorithms.

[0102] In this embodiment, the visual sensor employs a multi-view camera array, capable of capturing high-resolution images from different perspectives, covering the entire surface area of ​​the concrete structure. During flight, the UAV uses an onboard inertial measurement unit and a global positioning system for real-time positioning. Since changes in the UAV's flight attitude can cause image distortion, image sequence registration is required for the acquired multi-view surface image data to compensate for these distortion effects.

[0103] Image sequence registration employs a feature point matching algorithm. Specifically, it extracts scale-invariant feature transform descriptors (SIFT) to detect and describe key points in the image. Subsequently, a RANSAC optimization process is used to eliminate mismatched points and improve registration accuracy. The RANSAC optimization process calculates the fundamental matrix and estimates the interior point set; its objective function is expressed as: ,in Based on the matrix, For the estimated fundamental matrix, and It is the first A pair of matching feature points, For transpose operation, It is an indicator function. The distance threshold is set to 0.05 pixels to balance accuracy and computational efficiency. Through iterative random sampling and consistency checks, the RANSAC algorithm can remove outlier matches, ensuring robust estimation of the fundamental matrix.

[0104] The registered image sequences are then fused at multiple scales to generate a high-resolution panoramic image. The multi-scale fusion employs the Laplacian pyramid decomposition method, dividing each image into pyramid layers of different scales. Fusion weights are then adaptively calculated based on local contrast and signal-to-noise ratio. The specific formula for weight calculation is as follows: ,in It is the first Layer in position The fusion weight, It is the image gradient. For A local neighborhood window centered on the center, Therefore Points within a local neighborhood window centered on the center. This is a smoothing constant, set to 0.001 to prevent division by zero errors. This weighting formula prioritizes preserving detailed regions of the image, thereby improving the clarity and information content of the panoramic image. During the fusion process, each pyramid layer is linearly superimposed according to its weight, ultimately reconstructing a seamless panoramic image.

[0105] After generating the panoramic image, a deep learning segmentation model is used to perform pixel-level classification of the image to identify potential corrosion areas. This model is based on a variant of U-Net with an attention mechanism. Its encoder part consists of multiple residual blocks. After each residual block outputs a feature map, a channel attention mechanism is introduced for recalibration to enhance the feature representation capability. The formula for the channel attention mechanism is expressed as: ,in It is the first Output feature map after channel calibration It is the first Input feature maps of each channel, yes Scalar features after global average pooling. and Here are the weight matrices for the two fully connected layers. It is the ReLU activation function. It is the sigmoid function. This represents element-wise multiplication. By adaptively adjusting channel weights, features related to corrosion are highlighted while irrelevant information is suppressed. The encoder portion progressively downsamples the feature map to expand the receptive field and capture global contextual information.

[0106] The decoder uses transposed convolutions to upsample deep feature maps, gradually restoring spatial resolution. (Upsampled feature map) The same resolution feature map passed from the corresponding layer of the encoder through skip connections The features are concatenated along the channel dimension to obtain the fused feature map. The operation is represented as: ,in For the concatenation operation function, skip connections preserve detailed information from the encoder, avoiding information loss during upsampling. Subsequently, the fused feature maps are... Sequential depthwise separable convolution is performed to reduce computational complexity and improve model efficiency. Depthwise separable convolution involves two steps: first, ... Each input channel is processed independently using depthwise convolution to extract spatial features; then pointwise convolution is used to linearly combine the outputs of the depthwise convolution along the channel dimension to generate the output feature map of the current layer of the decoder. Its operation is represented as: ,in It is the decoder output. Represents depthwise convolution. This represents pointwise convolution. The decoder also introduces dilated convolution to expand the receptive field, capturing multi-scale information without increasing the number of parameters. Finally, the model outputs the class probability for each pixel, and thresholding is used to segment the boundaries of the rusted region.

[0107] The entire image processing algorithm runs in real time on an onboard GPU accelerator. During the training phase, the deep learning model uses a labeled dataset of concrete structure images, optimizing parameters through backpropagation and gradient descent. The loss function employs a combination of cross-entropy loss and Dice loss to improve segmentation accuracy. During the inference phase, the model performs forward propagation on the panoramic image, outputting a binary mask to identify the location and shape of potential corrosion areas.

[0108] Based on the identification results, a sequence of monitoring points is planned.

[0109] In this embodiment, the planning process first calculates the corrosion risk index for each potential corrosion area identified by the image processing module, quantifying its monitoring priority. The corrosion risk index is comprehensively evaluated using a weighted summation model, the model expression of which is: ,in, It is a region The risk index, It is a region The normalized area is obtained by dividing the actual pixel area by the total image area. It is a region The corrosion confidence level is determined by the probability value output by the deep learning segmentation model, reflecting the degree of confidence that the area belongs to corrosion. It is a region The historical data score is assigned based on the corrosion evolution trend of the area in previous monitoring records. If historical data is unavailable, the default value is used. It is a region Environmental factors, comprehensively considering on-site environmental parameters such as humidity, temperature, and chloride ion concentration, are acquired in real time through sensor data. Weighting coefficients. , , , Principal component analysis was used to determine that this method reduces the dimensionality of a large amount of historical monitoring data and extracts the main influencing factors to optimize weight allocation.

[0110] Prioritizing risk indices, a genetic algorithm is employed to solve for the monitoring point sequence. The goal is to minimize the drone flight path length while covering high-risk areas, thereby improving monitoring efficiency. The genetic algorithm abstracts the monitoring point sequence planning into a combinatorial optimization problem, with the objective function being a weighted combination of maximizing total risk coverage and minimizing path length, expressed as: ,in, It is a binary decision variable, when the region The value is 1 if the device is covered by at least one monitoring point, and 0 otherwise. Point Time The Euclidean distance is calculated based on the three-dimensional coordinates; It is a path decision variable; if the drone starts from point... Flight destination The value is then 1; It is a penalty factor used to balance the weight between risk coverage and path length. Its value is adjusted through pre-experimentation to avoid overfitting.

[0111] During algorithm initialization, multiple candidate sequences are randomly generated as a population, with each sequence representing a possible monitoring point access order. During iteration, selection, crossover, and mutation operations simulate natural evolution. Selection retains superior individuals based on fitness values ​​(i.e., objective function values), crossover swaps some sequences to explore new solutions, and mutation randomly adjusts the point order to maintain diversity. Fitness is recalculated after each iteration until convergence to an approximate optimal solution, generating the monitoring point sequence.

[0112] A dynamic flight path is generated based on the distribution of monitoring points. The spherical UAV is controlled to fly along the path and adjust its attitude when it arrives at the monitoring point so that no less than two spring sensors can simultaneously and effectively contact the concrete surface to form a closed electrical circuit.

[0113] In this embodiment, the process of generating a dynamic flight path first constructs a three-dimensional spatial path map based on the monitoring point sequence output by the planning module. The nodes of the path map represent the three-dimensional coordinates of each monitoring point in the global coordinate system. Edge weights comprehensively consider flight energy consumption and obstacle risk factors. The energy consumption model is calculated based on the UAV's dynamic characteristics, and obstacle risk is assessed by acquiring point cloud data and evaluating collision probabilities through real-time scanning of the environment by an onboard LiDAR. The Dijkstra algorithm is used to calculate the initial shortest path. This algorithm iteratively expands from the starting point, gradually solving for the minimum cost path to each monitoring point, ensuring optimal path length.

[0114] However, due to the presence of dynamic obstacles and airflow disturbances in the actual flight environment, model predictive control is required for dynamic adjustments. The predictive model of model predictive control describes the dynamic characteristics of the UAV, and the state vector is defined as... ,in , , This indicates the position coordinates of the drone in three-dimensional space. , , Represents the velocity component in the corresponding direction. Derivative of the state vector. ,in A nonlinear function describing the six-degree-of-freedom dynamic characteristics of an unmanned aerial vehicle (UAV). It is the control input vector, including throttle thrust and torque in three axes.

[0115] The optimization problem is expressed as follows: ,in It is a reference trajectory, generated by interpolation of the initial path. and It is a positive definite weight matrix, used to balance trajectory tracking accuracy and control energy consumption. and It is the weighted Euclidean norm. It predicts the time domain. This represents the time variable. The optimization problem is solved within each control cycle, using numerical methods such as sequential quadratic programming to generate control commands and achieve real-time smooth path adjustment.

[0116] The weighted Euclidean norm introduces a weight matrix to adjust the importance of different dimensions based on the standard Euclidean norm. The calculation process is as follows: ,in For transpose operation, the same applies. .

[0117] When controlling a spherical UAV to fly along a dynamic flight path, the onboard flight controller drives the propeller speed and adjusts the UAV's attitude based on the control output of the model predictive control output. During flight, positioning is achieved through fusion of the Global Positioning System (GPS) and the Inertial Measurement Unit (INS). Simultaneously, the Visual Odometry (VAU) estimates relative motion using camera sequence images to compensate for the accumulated errors of the INS. When the UAV reaches the vicinity of a preset monitoring point, it enters the attitude adjustment phase. Attitude adjustment utilizes real-time data provided by the onboard INS and VAU, representing the current attitude using quaternions to avoid the singularity problem of Euler angle representation.

[0118] The target pose is determined by the surface normal vector of the concrete at the monitoring point. This surface normal vector is calculated using a stereo vision algorithm based on images acquired by a multi-camera array, specifically by extracting surface feature points and fitting a local plane. The adjustment process utilizes a sliding mode control law, whose control torque... ,in It is the gain matrix, designed using Lyapunov stability theory to ensure convergence, and the sign function. This is a piecewise function that switches between the two sides of the sliding surface. Sliding surface Defined as a linear combination of attitude errors, i.e. ,in It is a three-dimensional vector representing the difference between the current angular velocity and the target angular velocity. It is a positive definite diagonal matrix used to adjust error weights. The three-dimensional attitude error vector is obtained by transforming the error quaternions calculated from the current attitude quaternions and the target attitude quaternions using a quaternion logarithmic mapping method. The sliding mode control law drives the system state to slide along the sliding surface, exhibiting robustness to external disturbances.

[0119] Synchronous contact with the concrete surface is achieved by controlling the extension and retraction of spring sensors. These spring sensors are evenly distributed across the surface of the spherical UAV, each containing a spring mechanism and an electrode head. The extension and retraction amount is adjusted based on laser rangefinder readings, which are pointed at the concrete surface to measure the distance between the UAV and the surface in real time. The control algorithm employs a proportional-derivative controller, and the extension and retraction length varies... ,in This is the current distance. It is a reference contact distance, set according to the sensor's mechanical design. and These are the proportional and derivative control coefficients, respectively. The Ziegler-Nichols tuning method is used to optimize and ensure a fast, overshoot-free response. This represents a time variable. During the adjustment process, the drone gradually approaches the surface, and the spring sensors extend in sequence, with their tips contacting the surface to form stable support.

[0120] Determining whether the posture meets the contact requirements involves calculating the contact force vector between the tip of each spring sensor and the concrete surface. The contact force estimation is based on spring deformation and Hooke's law. The contact force vector of each spring sensor ,in It is the spring stiffness coefficient, determined by the material properties. It is deformation displacement, measured by a built-in displacement sensor. For the first The unit direction vector of each spring pointing axially towards the concrete surface is obtained from the UAV attitude matrix transformation. Subsequently, the resultant force vector of all contact forces is calculated, and then compared with the surface normal vector. The cosine of the included angle is used for judgment. ,in It is the deviation angle. It is the surface normal vector. This is the threshold value; setting it to 0.95 corresponds to approximately 18 degrees of deviation. If... Greater than This indicates that the resultant force direction is aligned with the normal vector, and the contact stability meets the requirements; if not, the attitude adjustment process is re-triggered, and the attitude is fine-tuned through the sliding mode control law until the conditions are met.

[0121] When at least two spring sensors simultaneously contact the concrete surface, the electrode head becomes conductive with the surface, forming a closed electrical circuit. The closed circuit state is verified by a circuit continuity detection module to ensure reliable electrical connection.

[0122] By applying multi-frequency electrical excitation signals through an electrical circuit, the impedance spectrum and potential response are measured to obtain quantification parameters of corrosion.

[0123] In this embodiment, the electrical measurement module is equipped with a multi-frequency signal generator and an impedance analyzer to achieve electrical excitation and impedance measurement. Specifically, after the spherical UAV stably contacts the concrete surface at the monitoring point, a closed electrical loop is formed, and the electrical measurement module begins to operate. The multi-frequency signal generator generates multi-frequency sinusoidal electrical excitation signals with a frequency range from 0.1Hz to 10kHz, covering the low-frequency to high-frequency bands, capturing various electrochemical responses of the corrosion process. The excitation signal is applied to the concrete surface through the electrode head of the spring sensor, forming an electrical connection with the internal reinforcing steel. Simultaneously, the impedance analyzer synchronously measures the impedance spectrum data and potential response in the loop, with the sampling frequency adaptively adjusted according to the signal frequency to ensure data accuracy.

[0124] The measured impedance spectral data were then transmitted to the onboard processing unit for fitting analysis. An equivalent circuit model was used to fit the impedance data; this model employed a modified Randle circuit, incorporating diffuse impedance elements, to describe the electrochemical behavior of the steel reinforcement corrosion interface. The impedance expression is as follows: ,in . Indicates electrolyte resistance. Indicates double-layer capacitance. Represents charge transfer resistance. It is the Warburg diffusion impedance. It is the diffusion coefficient. It is angular frequency. The unit is the imaginary unit. The fitting process uses a nonlinear least squares iterative optimization method to minimize the residual between the measured value and the model prediction, thereby extracting the circuit parameters.

[0125] Based on the fitted circuit parameters, corrosion quantification parameters are calculated. First, according to the Stern-Geary equation, the polarization resistance is... Then, the corrosion current density for Where B is the Stern-Geary constant, with a value of 26 mV. The corrosion current density directly reflects the instantaneous corrosion degree of the reinforcing steel. To further assess the long-term corrosion impact, Faraday's law is used to predict corrosion depth, with the formula: .in, It is the depth of corrosion. It is the molar mass of the steel reinforcement. It is the time of steel bar corrosion, estimated based on historical monitoring data or environmental exposure time. It is the number of electrons transferred. It is Faraday's constant. It refers to the density of the reinforcing steel.

[0126] Based on corrosion depth Predicting the loss rate of the corroded section This quantifies the degree of cross-sectional loss in reinforcing steel. The cross-sectional loss rate represents the percentage reduction in the effective cross-sectional area of ​​the reinforcing steel due to corrosion, and is calculated using the following formula: , This represents the initial cross-sectional area of ​​the reinforcing steel bars in a concrete structure. Let be the initial radius of the reinforcing bar. = This represents the remaining effective cross-sectional area and remaining effective radius of the reinforcing steel in a concrete structure after corrosion. .

[0127] These parameters together quantify the severity and progression of corrosion.

[0128] The entire measurement process is non-destructive testing, and the intensity of the electrical excitation signal is controlled within a safe range to avoid causing additional damage to the structure.

[0129] The rust quantification parameters are transmitted in real time to the control center via a wireless module for analysis.

[0130] In this embodiment, after acquiring the corrosion quantification parameters, the wireless module initiates the data transmission process. Employing the 5G communication protocol, which features high bandwidth and low latency, ensures reliable real-time data transmission. The corrosion quantification parameters include key indicators such as polarization resistance, corrosion current density, corrosion depth, and corrosion cross-sectional loss rate. These parameters are encapsulated into standard data packets, with timestamps and monitoring point location information added. The data packets are encrypted and have a checksum appended to prevent tampering or loss during transmission.

[0131] The control center receives data streams from the drones. First, the data packets are decrypted and verified to confirm data integrity and source authenticity. Then, a parsing algorithm extracts quantifiable corrosion parameters and stores them in a structured database. The database uses a time-series data model, linking historical monitoring records and environmental parameters, supporting querying and backtracking. The parsing process includes parameter validity checks, such as range verification and outlier filtering, to ensure data quality. For each monitoring point, the system calculates parameter change trends in real time and compares them with preset thresholds to generate preliminary assessment results.

[0132] The analyzed data is further used to generate quantitative monitoring reports. These reports include a summary of corrosion status, risk level classification, and trend prediction charts. The control center's visual interface dynamically displays the monitoring results, supporting multi-dimensional data drill-down, such as analyzing corrosion evolution by time series or spatial distribution. When parameters exceed safety thresholds, the system automatically triggers an early warning mechanism, notifying management personnel via SMS or email. Simultaneously, the data is integrated with engineering models to assess the structure's remaining lifespan and develop maintenance strategies. All analysis results and reports are archived in a knowledge base, providing data support for structural health management.

[0133] Example 2: Figure 2As shown, this invention provides a non-destructive intelligent monitoring system for the safety of concrete structures based on unmanned aerial vehicles (UAVs), comprising:

[0134] The visual sensor module 1 uses a multi-view camera array mounted on a spherical UAV to collect image data of the concrete structure surface;

[0135] Image processing module 2 integrates a GPU accelerator to execute image registration, fusion, and deep learning segmentation algorithms to identify rusted areas;

[0136] Planning module 3, which includes a risk calculation unit and a path optimizer, is used to generate monitoring point sequences and dynamic flight paths;

[0137] Control module 4, which includes an attitude controller and a spring sensor driver, is used to adjust the attitude of the UAV and ensure synchronous contact.

[0138] Electrical measurement module 5, equipped with a multi-frequency signal generator and an impedance analyzer, is used to apply electrical excitation and measure the impedance spectrum;

[0139] The wireless transmission module 6 uses the 5G communication protocol to send the rust quantification parameters to the control center in real time.

[0140] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for quantitative and non-destructive intelligent monitoring of concrete structure safety based on unmanned aerial vehicles (UAVs), characterized in that, include: After the spherical drone takes off, it uses its onboard visual sensors to perform a full-area cruise scan of the concrete structure surface, acquires surface image data, and identifies potential corrosion areas through image processing algorithms. The identification of potential corrosion areas includes: Multi-view surface image data acquired by a spherical UAV visual sensor is used to perform image sequence registration to compensate for view distortion caused by changes in UAV flight attitude. The image sequence registration uses a feature point matching algorithm to extract SIFT feature descriptors and perform RANSAC optimization to eliminate mismatches. The RANSAC optimization process used in the image sequence registration is used to calculate the fundamental matrix and estimate the interior point set. Its objective function is: in Based on the matrix, For the estimated fundamental matrix, and It is the first A pair of matching feature points, For transpose operation, It is an indicator function. Distance threshold; The weight calculation for the multi-scale fusion adopts Laplace pyramid decomposition, and the weight formula for each layer is as follows: in It is the first Layer in position The fusion weight, It is the image gradient. For A local neighborhood window centered on the center, Therefore Points within a local neighborhood window centered on the center. It is a smoothing constant; The registered images are fused at multiple scales to generate a high-resolution panoramic image. The fusion weights are adaptively calculated based on the local contrast and signal-to-noise ratio of the images. A deep learning segmentation model is used to classify panoramic images at the pixel level. The deep learning segmentation model is a variant of U-Net based on the attention mechanism. Its encoder part introduces an SE module to enhance feature representation, and the decoder part uses dilated convolution to expand the receptive field and identify the boundaries of the rusted area. The algorithmic structure of the deep learning segmentation model includes: The encoder consists of multiple residual blocks. The output feature map of each residual block is recalibrated through a channel attention mechanism, the formula of which is: in It is the first Output feature map after channel calibration It is the first Input feature maps of each channel, yes Scalar features after global average pooling. and Here are the weight matrices for the two fully connected layers. It is the ReLU activation function. It is the sigmoid function. This represents element-wise multiplication; The decoder part uses transposed convolutions to upsample the deep feature maps, and then uses the upsampled feature maps... The same resolution feature map passed from the corresponding layer of the encoder through skip connections The features are concatenated along the channel dimension to obtain the fused feature map. : in, This is the function for concatenation operations; For the fused feature map Sequential depthwise separable convolution processing is performed, specifically including: right Each input channel is processed independently using depthwise convolution; Pointwise convolution is used to linearly combine the outputs of the depthwise convolution along the channel dimension to generate the output feature map of the current layer of the decoder. Its operation is represented as: in It is the decoder output. Represents depthwise convolution. This represents pointwise convolution; Based on the identification results, plan the sequence of monitoring points; A dynamic flight path is generated based on the distribution of monitoring points. The spherical UAV is controlled to fly along the path and adjust its attitude when it arrives at the monitoring point so that no less than two spring sensors can simultaneously and effectively contact the concrete surface to form a closed electrical circuit. By applying multi-frequency electrical excitation signals through an electrical circuit, the impedance spectrum and potential response are measured to obtain quantification parameters of corrosion. The rust quantification parameters are transmitted in real time to the control center via a wireless module for analysis.

2. The method for quantitative and non-destructive intelligent monitoring of concrete structure safety based on unmanned aerial vehicles (UAVs) as described in claim 1, characterized in that, The planned monitoring point sequence includes: Based on the identified potential corrosion areas, a corrosion risk index is calculated for each area. This corrosion risk index integrates the area area, corrosion confidence level, historical monitoring data, and environmental factors, using a weighted summation model. in It is a region The risk index, It is a region normalized area, It is a region The confidence level of corrosion It is a region Historical data scores, It is a region Environmental factors, , , , These are the weighting coefficients, determined through principal component analysis; Using risk indices as priorities, a genetic algorithm is used to solve for the monitoring point sequence. The objective function is a weighted combination of maximizing total risk coverage and minimizing path length, expressed as: in It is a binary variable representation area Whether it is covered It is a point Time Euclidean distance, These are path decision variables. It is a punishment factor; The optimal monitoring point sequence is generated iteratively.

3. The method for quantitative and non-destructive intelligent monitoring of concrete structure safety based on unmanned aerial vehicles (UAVs) as described in claim 1, characterized in that, The generated dynamic flight path includes: Based on the monitoring point sequence, a three-dimensional spatial path map is constructed, where nodes are the coordinates of the monitoring points, and edge weights take into account flight energy consumption and obstacle risks. The initial shortest path is calculated using Dijkstra's algorithm, and then dynamically adjusted using model prediction control. The prediction model is as follows: in It is a state vector. , , This indicates the position coordinates of the drone in three-dimensional space. , , Indicates the velocity components in the corresponding directions. The derivative of the state vector. A function describing the dynamic characteristics of an unmanned aerial vehicle (UAV). It controls the input vector; The optimization problem is expressed as: in It is a reference trajectory. and It is a positive definite weight matrix. and It is the weighted Euclidean norm. It predicts the time domain. Represents a time variable.

4. The method for quantitative and non-destructive intelligent monitoring of concrete structure safety based on unmanned aerial vehicles (UAVs) as described in claim 3, characterized in that, The attitude adjustment utilizes airborne IMU and visual odometry data, and the current attitude is represented by quaternions. The target attitude is determined by the surface normal vector of the monitoring point. The adjustment process uses a sliding mode control law: in It is the control torque. It is the gain matrix. For symbolic functions, It is a sliding surface, defined as a linear combination of attitude error quaternions: in It is a three-dimensional vector representing the difference between the current angular velocity and the target angular velocity. It is a positive definite diagonal matrix. The three-dimensional attitude error vector is obtained by converting the error quaternions calculated from the current attitude quaternion and the target attitude quaternion. Synchronous and effective contact with the concrete surface is achieved by controlling the extension and retraction of the spring sensor, with the amount of extension and retraction adjusted based on the laser rangefinder readings. in It refers to the change in the length of the telescoping. This is the current distance. This is a reference contact distance. and These are the proportional and derivative control coefficients, respectively.

5. The method for quantitative and non-destructive intelligent monitoring of concrete structure safety based on unmanned aerial vehicles (UAVs) as described in claim 4, characterized in that, Determining whether the posture meets the contact requirements includes: Calculate the contact force vector between the tip of each spring sensor and the concrete surface; Contact force estimation is based on spring deformation and Hooke's law: in It is the first The contact force vector of a spring sensor. It is the spring stiffness coefficient. It is deformation displacement. For the first A unit direction vector pointing axially from the spring to the concrete surface; The determination is made by calculating the cosine of the angle between the resultant force vector and the surface normal vector: in It is the deviation angle. It is the surface normal vector. It is a threshold; if it is not met, the attitude is readjusted.

6. The method for quantitative and non-destructive intelligent monitoring of concrete structure safety based on unmanned aerial vehicles (UAVs) as described in claim 5, characterized in that, The acquisition of corrosion quantification parameters includes: Impedance spectrum data were measured by applying multi-frequency sinusoidal electrical excitation signals through an electrical circuit, with a frequency range from 0.1 Hz to 10 kHz. The impedance is fitted using an equivalent circuit model, which is an improved Randle circuit that includes diffused impedance elements. Its impedance expression is as follows: in It is an electrolyte resistor. It is a double-layer capacitor. It is a charge transfer resistor. It is the Warburg diffusion impedance. It is the diffusion coefficient. It is angular frequency. The imaginary unit; Then, the polarization resistance is calculated according to the Stern-Geary equation. Corrosion current density The derivation is as follows: Where B is the Stern-Geary constant; Corrosion depth prediction uses Faraday's law: in It is the depth of corrosion. It is the molar mass of the steel reinforcement. It refers to the time it takes for the steel bars to corrode. It is the number of electrons transferred. It is Faraday's constant. It refers to the density of the reinforcing steel. Predicting the corrosion section loss rate based on corrosion depth: in, Indicates the corrosion cross-sectional loss rate. This represents the initial cross-sectional area of ​​the reinforcing steel bars in a concrete structure. Let be the initial radius of the reinforcing bar. = This represents the remaining effective cross-sectional area and remaining effective radius of the reinforcing steel in a concrete structure after corrosion. .

7. A non-destructive intelligent monitoring system for the safety of concrete structures based on unmanned aerial vehicles (UAVs), characterized in that, The system is used to implement the UAV-based non-destructive intelligent monitoring method for the safety of concrete structures as described in any one of claims 1 to 6, including: The visual sensor module uses a multi-view camera array mounted on a spherical drone to collect image data of the concrete structure surface; The image processing module integrates a GPU accelerator to execute image registration, fusion, and deep learning segmentation algorithms to identify rusted areas. The planning module, which includes a risk calculation unit and a path optimizer, is used to generate monitoring point sequences and dynamic flight paths; The control module, which includes an attitude controller and a spring sensor driver, is used to adjust the drone's attitude and ensure synchronized contact. The electrical measurement module, equipped with a multi-frequency signal generator and an impedance analyzer, is used to apply electrical excitation and measure the impedance spectrum; The wireless transmission module uses the 5G communication protocol to send corrosion quantification parameters to the control center in real time.

Images