All-terrain self-adaptive hydropower AI bionic robot intelligent inspection method and system

Abstract

The invention discloses an intelligent inspection method and system for an all-terrain self-adaptive hydroelectric AI bionic robot, and relates to the technical field of robots, and the method comprises the steps: constructing a terrain stiffness matrix based on a laser radar, an IMU and a foot end force sensor, and driving a bionic joint group to carry out the intelligent switching of a wheel mode, a six-foot gait mode and an adsorption mode; for equipment defect identification, visible light, infrared and laser point cloud data are fused, and crack / corrosion classification is realized by adopting a contrast learning migration model; a dynamic risk grid map is generated in combination with the equipment fault history, the environment corrosion rate and the task emergency degree, the inspection path is optimized through an MOEA / D-TOPSIS algorithm, and multi-target balance of energy consumption and path length is achieved on the premise that the coverage rate is guaranteed; the positioning system integrates Beidou RTK, UWB anchor points and inertial navigation, and combines a dynamic SLAM enhancement technology, so that centimeter-level positioning precision under strong electromagnetic interference and satellite rejection environments is ensured, and an all-weather and all-terrain intelligent inspection solution is provided for hydroelectric facilities.

Description

Technical Field

[0001] This invention relates to the field of robotics, specifically to an intelligent inspection method and system for an all-terrain adaptive hydropower AI bionic robot. Background Technology

[0002] As a crucial power generation site, the stable operation of equipment in hydropower plants is of paramount importance. Traditional equipment inspections in hydropower plants primarily rely on manual labor. However, the efficiency of manual inspections is increasingly insufficient to meet the high-frequency inspection needs of the large number of devices in hydropower plants. Furthermore, in complex environments, such as densely distributed equipment, confined spaces, and dark, damp conditions, the accuracy and safety of manual inspections are difficult to guarantee. Additionally, hydropower plants possess a wide variety of equipment with complex operating conditions, making it difficult for manual inspections to quickly and accurately detect potential faults and anomalies.

[0003] With the development of artificial intelligence and robotics, although some robots have been applied to equipment inspection, existing technologies still face many challenges in the complex environment of hydropower plants. These challenges include: insufficient equipment recognition accuracy in complex environments, making it difficult to accurately identify subtle equipment anomalies; limited autonomous movement capabilities of robots in complex terrain and access control environments, making it difficult to achieve efficient inspection of the entire area; and inadequate positioning accuracy and multi-source sensing capabilities to meet the needs of refined inspection in hydropower plants. Therefore, there is an urgent need for an all-terrain adaptive hydropower AI bionic robot intelligent inspection solution to address these issues. Summary of the Invention

[0004] To address the aforementioned technical challenges, this paper presents an intelligent inspection method and system for an AI-powered biomimetic robot that is adaptive to all terrains for hydropower. This technical solution resolves the issues of insufficient implicit feature extraction capabilities and difficulties in multi-source data fusion.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] In a first aspect, the present invention provides an intelligent inspection method for an all-terrain adaptive hydropower AI bionic robot, comprising:

[0007] Acquire and generate a terrain stiffness matrix based on data from lidar, IMU, and foot force sensor;

[0008] Simultaneously acquire visible light images, infrared thermal imaging information, and laser point cloud information, and output crack / corrosion classification results based on a contrastive learning transfer model;

[0009] Based on the crack / corrosion classification results, a dynamic risk grid map is constructed by integrating equipment failure history, environmental corrosion rate, and task urgency.

[0010] Based on the dynamic risk grid map, three-source positioning is fused, and dynamic SLAM enhancement is performed to obtain the enhanced map;

[0011] Based on the enhanced map, MOEA / D-TOPSIS path optimization is performed, simultaneously satisfying the preset proportion of path length being less than the perimeter of the area, and controlling coverage and energy consumption within preset ranges to obtain the optimized path.

[0012] Multimodal motion decisions are made based on the optimized path and terrain stiffness matrix.

[0013] Furthermore, data from lidar, IMU, and foot force sensors are acquired to generate a terrain stiffness matrix, including:

[0014] A 16-line rotating lidar is used to scan a preset range around the robot at a preset frequency to generate a three-dimensional point cloud terrain elevation map, and to capture the terrain features of slope, convexity and depression in real time to obtain lidar data.

[0015] It has a built-in six-axis inertial measurement unit, which outputs the robot's attitude angle and angular velocity data, detects attitude fluctuations caused by terrain, and obtains IMU data;

[0016] A three-dimensional force sensor is embedded in each foot to measure the contact force vector between the foot and the ground in real time, calculate the distribution of ground reaction force, and obtain foot force sensor data.

[0017] Furthermore, a terrain stiffness matrix is ​​generated based on data from lidar, IMU, and foot force sensors, including:

[0018] Time synchronization is achieved by aligning the time of data from the lidar, IMU, and foot force sensor via hardware trigger signals.

[0019] Spatial calibration involves establishing a transformation matrix between the lidar coordinate system, the IMU body coordinate system, and the foot local coordinate system to unify the data into the robot's global coordinate system.

[0020] Feature extraction, fusion of point cloud elevation data and foot force feedback, and calculation of the estimated value of the elastic modulus of the terrain surface;

[0021] Stiffness classification and quantification divides the terrain into grid cells, assigns a stiffness value to each cell, and classifies them into rigid surfaces, semi-rigid surfaces, and flexible surfaces, thus constructing a stiffness matrix.

[0022] Water reflection filtering is performed by applying range-intensity joint filtering to lidar point cloud data.

[0023] Metal surface correction: In the metal structure area of ​​the power station, millimeter-wave radar-assisted point cloud correction is used.

[0024] Motion artifact suppression is achieved by compensating for point cloud distortion caused by the robot's own motion using IMU angular velocity data.

[0025] The matrix data is structured and output as a two-dimensional stiffness distribution map, with a confidence score for each grid cell.

[0026] When multiple consecutive grid stiffness values ​​are detected to be less than a preset threshold, an avoidance command is sent to the motion control system; when the grid stiffness gradient changes abruptly and exceeds the preset threshold, it is marked as a terrain discontinuity and preload control is activated.

[0027] Furthermore, visible light images, infrared thermal imaging information, and laser point cloud information are acquired simultaneously, including:

[0028] A global shutter industrial camera is used to capture RGB images of the equipment surface; a ring LED fill light is applied, and an adaptive homomorphic filter is used to obtain a visible light image;

[0029] Equipped with an uncooled microbolometer, it generates a temperature distribution map; to address the interference from metal surface reflection, it employs a dynamic emissivity correction algorithm and compensates for temperature measurement errors based on ambient temperature and humidity to obtain infrared thermal imaging information;

[0030] Using a ToF lidar, a divergence angle beam is emitted to acquire sub-millimeter level three-dimensional point clouds on the device surface; to address water refraction interference, a Snell's law compensation model is applied to correct underwater point cloud distortion, thus obtaining laser point cloud information.

[0031] Furthermore, based on the contrastive learning transfer model, the crack / corrosion classification results are output, including:

[0032] Synchronization pulse signals are generated using FPGA to align the timestamps of the three sensors; a multi-sensor joint coordinate system is established based on a checkerboard calibration board.

[0033] For visible light images, texture features, edge features, and color anomaly regions are extracted; for infrared thermal imaging information, temperature gradient fields and morphological contours of overheated areas are extracted; for laser point clouds, surface curvature changes and normal vector deviations are calculated; a multi-channel fusion tensor is constructed to map visible light texture, infrared temperature gradient, and point cloud curvature to a unified dimensional space, and key features are enhanced through an attention weighting mechanism.

[0034] The contrastive learning transfer model includes a backbone network, which is pre-trained on the ImageNet-21K dataset using Vision Transformer; contrastive learning optimization is performed by introducing the NT-Xent loss function to narrow the distance between features of the same type of defect and push apart features of different types.

[0035] Fine-tuning of the hydropower scenario was performed, and the migration strategy was to freeze the weights of the pre-set proportion layer before ViT and only fine-tune the top Transformer block and classification head; small sample training was performed, based on defect samples of hydropower equipment, and hard sample mining was applied.

[0036] The classification decision logic is implemented, with a fusion tensor as input and a defect probability vector as output. A dual threshold is set for judgment: if the confidence level is greater than a preset proportion, the classification result is directly output. If the confidence level is within the preset proportion range, local reconstruction of the 3D point cloud is activated, and classification is performed after secondary verification based on the Poisson surface reconstruction algorithm.

[0037] Crack parameters were calculated. The length was obtained by fitting a B-spline curve along the normal vector direction of the point cloud, and the depth was obtained by combining the infrared temperature gradient and the height difference of the point cloud.

[0038] Enhanced adaptability for hydropower scenarios: High-reflectivity surface treatment; applying polarized light imaging to the stainless steel casing of generators, combining point cloud intensity information to separate real defects from reflective artifacts; underwater imaging compensation; enhanced penetration through blue-green light bands when inspecting the inner wall of pressure steel pipes; dynamic target filtering; using optical flow motion detection for swimming fish and floating debris, combined with morphological filtering to eliminate non-static targets.

[0039] Furthermore, based on the crack / corrosion classification results, and by integrating equipment failure history, environmental corrosion rate, and task urgency, a dynamic risk grid map is constructed, including:

[0040] Acquire historical equipment failure data, connect to the hydropower plant's SCADA system, extract past equipment failure records, and construct a failure probability density function; environmental corrosion rate, acquired in real time through weather stations and corrosion sensors, is divided into atmospheric corrosion and water corrosion; task urgency, parsed maintenance work order data, is classified by priority.

[0041] Spatial discretization is performed, dividing the inspection area into grid cells, assigning unique coordinates to each cell, and dynamically calculating the risk value. The risk value of each grid cell is calculated by a formula. Crack / corrosion results are superimposed, and if an identified defect exists within the grid, the risk value is multiplied according to the defect level.

[0042] Heatmaps are generated, and risk values ​​are mapped using the HSV color space. The tiered response strategy includes high-risk grids, which immediately trigger the robot to re-inspect them, increasing the data feedback frequency; medium-risk grids, which are included in the nodes that must be passed through in the optimized path; and low-risk grids, which are covered as needed and allowed to be skipped.

[0043] Furthermore, by fusing three-source localization and performing dynamic SLAM enhancement, an enhanced map is obtained, including:

[0044] To obtain BeiDou RTK positioning, a dual-frequency receiver is used to differentially correct ionospheric errors through ground reference stations and output centimeter-level absolute coordinates. To address the multipath effect in the canyon of the hydropower plant, a multipath suppression algorithm is implemented, including reducing the weight by a corresponding proportion when the signal incident angle is greater than a preset threshold angle, carrier phase cycle slip detection, and switching to BeiDou-3 B2b signal when an anomaly occurs.

[0045] UWB anchor point positioning is obtained, a self-calibrating anchor point network is deployed in the inspection area, TDOA technology is adopted, and the robot carries a tag to communicate at a preset frequency; an anchor point drift self-correction mechanism is introduced, a point cloud map is built based on laser SLAM, the anchor point position offset is calculated, and when the offset exceeds the preset distance, the UWB anchor point position is updated online.

[0046] The system acquires inertial navigation (INS) data from a nine-axis MEMS IMU and calculates attitude angles using the fourth-order Runge-Kutta method; it also incorporates a built-in temperature compensation chip to address temperature drift.

[0047] In ideal environments, strong magnetic interference, and satellite denial, the weights of BeiDou RTK, UWB, and INS are adaptively allocated; based on the state vector, the three-source data are fused using the observation equation; when the variance of any source data mutation exceeds a preset multiple of the historical variance mean, the fusion is paused and SLAM pose prediction is enabled.

[0048] Furthermore, based on the enhanced map, MOEA / D-TOPSIS path optimization is performed, simultaneously satisfying a preset proportion where the path length is less than the area perimeter, and controlling coverage and energy consumption within preset ranges, resulting in optimized paths, including:

[0049] Optimization objectives include minimizing path length, maximizing coverage, and minimizing energy consumption; constraint embedding includes mandatory coverage of high-risk areas, time window limits, and physical accessibility.

[0050] MOEA / D-TOPSIS Hybrid Optimization Process:

[0051] Population initialization generates several initial paths, each path being a grid coordinate sequence, with the starting point fixed at the hydropower plant control center and the ending point at the charging pile;

[0052] MOEA / D decomposition and collaboration decomposes a multi-objective problem into several sub-problems; neighborhood definition selects the sub-problems with the closest Euclidean distance as the collaborative neighborhood; sub-problem solving uses a differential evolution operator to generate new paths;

[0053] TOPSIS optimal solution decision-making calculates the approximation distance to the ideal solution for each path and selects the path with the smallest approximation distance as the final output;

[0054] When a new crack is detected and the risk value is greater than a preset threshold, an enhanced detection zone is generated around the risk grid; the current path is interrupted, and the shortest detour path is calculated using the A* algorithm; the TOPSIS weights are updated.

[0055] Furthermore, based on the optimized path and terrain stiffness matrix, multimodal motion decisions are performed, including:

[0056] The wheel movement mode is activated when the slope is less than the preset angle range.

[0057] When the slope is within the preset angle range, it switches to six-legged gait mode;

[0058] When the slope is greater than the preset angle range, the biomimetic adsorption-gait coordination mode is triggered.

[0059] Furthermore, the wheeled mobility mode includes: activating the McCannum wheel drive to achieve a zero turning radius by utilizing the omnidirectional wheel characteristics;

[0060] The six-legged gait mode includes: switching to hydraulically driven six-legged gait, adopting a triangular gait sequence, with the three diagonal legs lifting simultaneously, and the stride being dynamically adjusted to a preset ratio of the foot length;

[0061] The biomimetic adsorption-gait coordinated control mode includes:

[0062] During the adsorption phase, the foot tip is pre-pressed to contact the surface, activating the deformation and locking of the bristle array. The adsorption force is controlled through a force feedback closed loop, while the negative pressure cavity is released to enhance the adhesion.

[0063] During the stepping phase, a support triangle is constructed, and three non-adjacent legs are selected to form a support surface, while the other three legs perform the stepping motion.

[0064] Dynamic gait planning: the stepping foot is raised to a preset height along the normal direction, and the step distance is adjusted according to the slope using a formula.

[0065] Furthermore, based on the optimized path and terrain stiffness matrix, performing multimodal motion decisions also includes:

[0066] When the single-foot slip rate exceeds a preset ratio, the current gait is immediately frozen, and foot position remapping is triggered through IMU attitude feedback to reconstruct the support triangle within a preset time.

[0067] Furthermore, based on the optimized path and terrain stiffness matrix, performing multimodal motion decisions also includes:

[0068] When the slope sensor continuously samples a value that exceeds the threshold, a mode switch is triggered; wheel type switches to foot type, the wheel set is retracted and the wheel axle is locked; foot type switches to adsorption type, the bristle array at the foot end is pre-extended, and the adsorption module is powered on to preheat.

[0069] In the transition zone, a hybrid mode is activated, where the front wheels keep driving while the rear four legs execute the gait.

[0070] A hydrophobic coating is sprayed onto the foot of the slippery area in the turbine housing.

[0071] To address magnetic interference on metal surfaces, a non-magnetic titanium alloy foot structure is adopted in the strong magnetic field environment of the generator compartment.

[0072] In turbulent environment stability control, when the flow velocity in the turbine corridor reaches the threshold velocity, the vortex suppression algorithm is activated, and the thruster vector nozzle deflects.

[0073] The IMU provides real-time feedback on the attitude angle, and the PID controller outputs joint torque compensation commands.

[0074] Secondly, this invention provides an all-terrain adaptive hydropower AI bionic robot intelligent inspection system, comprising:

[0075] Multimodal bionic mobile platform: integrates three-state drive modules of wheel, leg and propeller propulsion, adopts bionic joint group collaborative control algorithm, and is optimized based on biological central pattern generator CPG model; equipped with active deformable track structure, which can automatically adjust the ground area according to the terrain slope;

[0076] Multi-source sensing fusion module: The cross-media environment sensing unit is equipped with lidar, millimeter-wave radar, and multispectral industrial camera;

[0077] Electromagnetic interference suppression layer: The sensor is encased in a Mu metal shielding cover;

[0078] Edge-AI Decision Center: Deploys a lightweight Transformer model to support the identification of subtle device anomalies; integrates a dynamic SLAM enhancement engine to execute the intelligent inspection method described in the first aspect.

[0079] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0080] This invention significantly improves the identification accuracy of hydropower plant equipment inspection in complex environments by constructing artificial intelligence recognition algorithms and multi-source perception technology. It can promptly detect subtle abnormalities in equipment and improve the safety and reliability of equipment operation. The autonomous intelligent path dynamic planning method for quadruped robots in complex environments enables the robot to autonomously plan the optimal path in complex terrain and environment, achieve efficient inspection of the entire area, improve inspection efficiency, and reduce the cost of manual inspection.

[0081] The access control and obstacle crossing schemes and precise positioning methods of the bionic robot ensure the robot's autonomous movement and positioning accuracy in the complex environment of the hydropower plant, enabling the robot to successfully complete inspection tasks and expanding the inspection range.

[0082] The integrated solution of multi-source sensing methods enables comprehensive perception of the environment and equipment information of hydropower plants, which can promptly detect safety hazards such as oil and gas leaks, and improve the timeliness and efficiency of fault handling through remote alarms and visualization.

[0083] The all-terrain adaptive hydropower AI bionic robot intelligent inspection device and system realizes intelligent, automated and remote management of hydropower plant equipment inspection, providing strong support for the intelligent upgrading of hydropower plants, and has broad application prospects and significant economic benefits. Attached Figure Description

[0084] Figure 1 A flowchart for an intelligent inspection method for an AI-inspired bionic robot for all-terrain adaptive hydropower projects.

[0085] Figure 2 This is an internal framework diagram of an all-terrain adaptive hydropower AI bionic robot intelligent inspection system. Detailed Implementation

[0086] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0087] Example 1:

[0088] This embodiment provides an intelligent inspection method for all-terrain adaptive hydropower AI bionic robot, including:

[0089] S1. Generate a terrain stiffness matrix using data from lidar, IMU, and foot force sensor.

[0090] S3. Simultaneously acquire visible light, infrared thermal imaging, and laser point cloud data; output crack / corrosion classification results based on a contrastive learning transfer model.

[0091] S4. Based on the crack / corrosion classification results, integrate equipment failure history, environmental corrosion rate, and task urgency, and allocate them proportionally to construct a dynamic risk grid map;

[0092] S6. Based on a dynamic risk grid map, integrate three-source positioning, BeiDou RTK, UWB anchor points, and inertial navigation; perform dynamic SLAM enhancement to obtain an enhanced map.

[0093] S5. Based on the enhanced map, perform MOEA / D-TOPSIS path optimization, simultaneously satisfying the preset proportion of path length being less than the perimeter of the area, and controlling coverage and energy consumption within the preset range to obtain the optimized path;

[0094] S2. Based on the optimized path and terrain stiffness matrix, perform multimodal motion decisions. When the slope is less than a preset angle, activate wheeled movement; when the slope is within a preset angle range, switch to hexapod gait; when the slope is greater than a preset angle, trigger biomimetic adsorption-gait coordination.

[0095] Preferably, S1 specifically includes:

[0096] The LiDAR scanning uses a 16-line rotating LiDAR to scan a preset range around the robot at a preset frequency, generating a three-dimensional point cloud terrain elevation map and capturing the terrain features of slope, convexity, and depression in real time.

[0097] IMU attitude perception, with a built-in six-axis inertial measurement unit, outputs robot body attitude angle and angular velocity data, and detects attitude fluctuations caused by terrain;

[0098] Foot force feedback: Each foot is embedded with a three-dimensional force sensor to measure the contact force vector between the foot and the ground in real time and calculate the distribution of ground reaction force.

[0099] Time synchronization: Align the time of the three types of sensors through hardware trigger signals; Spatial calibration: Establish the transformation matrix of the lidar coordinate system, IMU body coordinate system, and foot local coordinate system to unify the data to the robot's global coordinate system; Feature extraction: Fuse point cloud elevation data with foot force feedback to calculate the estimated value of the elastic modulus of the terrain surface.

[0100] Stiffness classification and quantification divides the terrain into grid cells, assigns a stiffness value to each cell, and classifies them into rigid surfaces, semi-rigid surfaces, and flexible surfaces, thus constructing a stiffness matrix.

[0101] Water reflection filtering: range-intensity joint filtering is applied to the lidar point cloud; metal surface correction: millimeter-wave radar is used to assist in point cloud correction in the metal structure area of ​​the power station; motion artifact suppression: point cloud distortion caused by the robot's own motion is compensated by IMU angular velocity data.

[0102] The matrix data is structured and output as a two-dimensional stiffness distribution map, with a confidence score for each grid cell.

[0103] When multiple consecutive grid stiffness values ​​are detected to be less than a preset threshold, an avoidance command is sent to the motion control system; when the grid stiffness gradient changes abruptly and exceeds the preset threshold, it is marked as a terrain discontinuity and preload control is activated.

[0104] Preferably, S2 specifically includes:

[0105] Visible light imaging employs a global shutter industrial camera to capture RGB images of the equipment surface; ring LED supplementary lighting is used, along with adaptive homomorphic filtering.

[0106] Infrared thermal imaging, equipped with an uncooled microbolometer, generates a temperature distribution map; to address interference from metal surface reflection, a dynamic emissivity correction algorithm is used, and temperature measurement errors are compensated based on ambient temperature and humidity;

[0107] Laser point cloud scanning uses a ToF lidar to emit a divergence angle beam and acquire sub-millimeter-level three-dimensional point clouds on the surface of the device; to address water refraction interference, a Snell's law compensation model is applied to correct underwater point cloud distortion.

[0108] A spatiotemporal synchronization mechanism is implemented by generating synchronization pulse signals through FPGA to align the timestamps of the three sensors; a multi-sensor joint coordinate system is established based on a checkerboard calibration board.

[0109] Visible light images are used to extract texture features, edge features, and color anomaly areas; infrared images are used to extract temperature gradient fields and morphological contours of overheated areas; laser point clouds are used to calculate surface curvature changes and normal vector deviations; a multi-channel fusion tensor is constructed to map visible light texture, infrared temperature gradient, and point cloud curvature to a unified dimensional space, and key features are enhanced through an attention weighting mechanism.

[0110] Preferably, S2 specifically includes:

[0111] The pre-trained architecture includes a backbone network, which is pre-trained on the ImageNet-21K dataset using Vision Transformer; contrastive learning optimization is performed by introducing the NT-Xent loss function to narrow the distance between features of the same type of defect and push apart features of different types.

[0112] For fine-tuning of the hydropower scenario, the migration strategy is to freeze the weights of the pre-set proportion layer before ViT and only fine-tune the top Transformer block and classification head; small sample training is conducted based on defect samples of hydropower equipment and hard sample mining is applied.

[0113] The classification decision logic takes a fusion tensor as input and outputs a defect probability vector. It sets two thresholds for judgment: if the confidence level is greater than a preset proportion, it directly outputs the classification result. If the confidence level is within the preset proportion range, it activates local reconstruction of the 3D point cloud and classifies the data after secondary verification based on the Poisson surface reconstruction algorithm.

[0114] Crack parameters were calculated: the length was obtained by fitting a B-spline curve along the point cloud normal vector direction; the depth was obtained by combining the infrared temperature gradient and the point cloud height difference; and corrosion levels were classified according to the crack parameters.

[0115] The hydropower scenario adaptability enhancement is carried out in three steps: strong reflective surface treatment, applying polarized light imaging to the stainless steel casing of the generator, and combining point cloud intensity information to separate real defects from reflective artifacts; underwater imaging compensation, using blue-green light band penetration enhancement when inspecting the inner wall of pressure steel pipes; and dynamic target filtering, using optical flow motion detection for swimming fish and floating debris, combined with morphological filtering to eliminate non-static targets.

[0116] Preferably, S3 specifically includes:

[0117] Historical equipment failure data is integrated into the hydropower plant's SCADA system to extract past equipment failure records and construct a failure probability density function; environmental corrosion rate is acquired in real time through weather stations and corrosion sensors, and is divided into atmospheric corrosion and water corrosion; task urgency is analyzed from maintenance work order data and classified by priority.

[0118] Spatial discretization is used to divide the inspection area into grid cells, assigning unique coordinates to each cell. Dynamic risk values ​​are calculated, with the risk value of each grid cell derived from a formula. Crack / corrosion results are superimposed, and if defects identified in step S2 exist within the grid, the risk value is multiplied according to the defect level.

[0119] Heatmaps are generated, and risk values ​​are mapped using the HSV color space. The tiered response strategy includes high-risk grids, which immediately trigger the robot to re-inspect, increasing the data feedback frequency; medium-risk grids, which are included in the nodes that must be passed through in the optimized path; and low-risk grids, which are covered as needed and can be skipped.

[0120] Preferably, S4 specifically includes:

[0121] Beidou RTK positioning, equipped with a dual-frequency receiver, uses ground reference station differential correction of ionospheric error to output centimeter-level absolute coordinates; in response to the multipath effect in the canyon of hydropower plant, a multipath suppression algorithm is used, including reducing the weight by a corresponding proportion when the signal incident angle is greater than a preset threshold angle, carrier phase cycle slip detection, and switching to Beidou third generation B2b signal when abnormal.

[0122] UWB anchor point positioning deploys a self-calibrating anchor point network in the inspection area, adopts TDOA technology, and the robot carries tags to communicate at a preset frequency; an anchor point drift self-correction mechanism is introduced, a point cloud map is built based on laser SLAM, the anchor point position offset is calculated, and when the offset exceeds the preset distance, the UWB anchor point position is updated online.

[0123] Inertial navigation INS, nine-axis MEMS IMU, attitude angles are calculated using the fourth-order Runge-Kutta method; built-in temperature compensation chip is used to address temperature drift;

[0124] In ideal environments, strong magnetic interference, and satellite denial, the weights of BeiDou RTK, UWB, and INS are adaptively allocated; based on the state vector, the three-source data are fused using the observation equation; when the variance of any source data mutation exceeds a preset multiple of the historical variance mean, the fusion is paused and SLAM pose prediction is enabled.

[0125] Preferably, S5 specifically includes:

[0126] Optimization objectives include minimizing path length, maximizing coverage, and minimizing energy consumption; constraint embedding includes mandatory coverage of high-risk areas, time window limits, and physical accessibility.

[0127] MOEA / D-TOPSIS Hybrid Optimization Process:

[0128] Population initialization generates several initial paths, each path being a grid coordinate sequence, with the starting point fixed at the hydropower plant control center and the ending point at the charging pile;

[0129] MOEA / D decomposition and collaboration decomposes a multi-objective problem into several sub-problems; neighborhood definition selects the sub-problems with the closest Euclidean distance as the collaborative neighborhood; sub-problem solving uses a differential evolution operator to generate new paths;

[0130] TOPSIS optimal solution decision-making calculates the approximation distance to the ideal solution for each path and selects the path with the smallest approximation distance as the final output;

[0131] When S2 detects a new crack and the risk value is greater than the preset threshold, it generates an enhanced detection zone around the risk grid; interrupts the current path and calculates the shortest detour path using the A* algorithm; and updates the TOPSIS weights.

[0132] Preferably, S6 specifically includes:

[0133] The preset angle threshold setting of the slope classification decision mechanism, based on big data analysis of hydropower plant terrain, divides the slope into three decision intervals:

[0134] Flat terrain mode activates McCannum wheel drive, utilizing the omnidirectional wheel characteristics to achieve zero turning radius; medium slope mode switches to hydraulically driven six-legged gait, adopting a triangular gait sequence, with the three diagonal legs lifting simultaneously, and the stride dynamically adjusted to a preset ratio of foot length; extreme slope mode triggers biomimetic adsorption-gait coordinated control, with the foot end generating van der Waals force adsorption based on micro-nano structure biomimetic bristles.

[0135] The biomimetic adsorption-gait coordinated control includes: adsorption stage, where the foot pre-presses into the contact surface, activates the bristle array deformation locking, and controls the adsorption force through a force feedback closed loop, while releasing the negative pressure cavity to enhance adhesion; stepping stage, a support triangle is constructed, and three non-adjacent feet are selected to form a support surface, while the other three feet perform the lifting action; dynamic gait planning, where the lifting foot raises a preset height along the normal direction, and the stepping distance is adjusted according to the slope using a formula;

[0136] The fault tolerance mechanism is switched. When the single-foot slip rate is greater than the preset ratio, the current gait is immediately frozen, and the foot position remapping is triggered through IMU attitude feedback to reconstruct the support triangle within a preset time.

[0137] Preferably, S6 specifically includes:

[0138] The multimodal seamless switching logic includes: transition condition detection, triggering mode switching when the continuous sampling value of the slope sensor crosses the threshold; wheeled to footed switching, retracting the wheel set and locking the axle; footed to adsorption switching, pre-extending the bristle array at the foot end, and powering on the adsorption module for preheating;

[0139] Energy consumption optimization strategy: In the transition zone, activate hybrid mode, i.e., front wheel keep drive and rear quadruped executes gait;

[0140] Enhanced adaptability to complex scenarios includes: anti-slip on wet surfaces, with a hydrophobic coating sprayed on the foot end in the wet and slippery area of ​​the turbine housing; magnetic interference response on metal surfaces, employing a non-magnetic titanium alloy foot end structure in the strong magnetic field environment of the generator nacelle; turbulent environment stabilization control, activating the vortex suppression algorithm and deflecting the thruster vector nozzle when the flow velocity in the turbine corridor reaches the threshold velocity; and real-time feedback of attitude angle by the IMU, with the PID controller outputting joint torque compensation commands.

[0141] Example 2:

[0142] Reference Figure 1 As shown, this embodiment provides an intelligent inspection method for all-terrain adaptive hydropower AI bionic robots, including:

[0143] S1. Generate a terrain stiffness matrix using data from lidar, IMU, and foot force sensor.

[0144] S2. Based on the terrain stiffness matrix, perform multimodal motion decisions. When the slope is less than a preset angle, activate wheeled movement; when the slope is within a preset angle range, switch to hexapod gait; when the slope is greater than a preset angle, trigger biomimetic adsorption-gait coordination.

[0145] S3. Simultaneously acquire visible light, infrared thermal imaging, and laser point cloud data; output crack / corrosion classification results based on a contrastive learning transfer model.

[0146] S4. Based on the crack / corrosion classification results, integrate equipment failure history, environmental corrosion rate, and task urgency, and allocate them proportionally to construct a dynamic risk grid map;

[0147] S5. Based on the dynamic risk grid map, perform MOEA / D-TOPSIS path optimization, and simultaneously meet the preset proportion of path length being less than the perimeter of the area, and control coverage and energy consumption within the preset range.

[0148] S6 integrates three-source positioning: BeiDou RTK, UWB anchor point, and inertial navigation; and performs dynamic SLAM enhancement.

[0149] It should be noted that the perception layer (S1, S3, S6) integrates the terrain stiffness matrix (S1), multimodal equipment defect data (S3), and three-source localization and SLAM (S6) to construct a multidimensional environmental cognition model; after spatiotemporal registration, sensor data such as lidar and infrared thermal imaging form a unified digital twin of the hydropower scene with a resolution of millimeters, covering mechanical, thermal, and geometric multi-physics field features.

[0150] Decision layers (S2, S4, S5): Based on perception data, through a dynamic hierarchical response mechanism (slope threshold triggering motion mode switching) and risk-driven path optimization (MOEA / D-TOPSIS algorithm), the transition from local motion control to global task planning is achieved; for example, the biomimetic adsorption decision of S2 directly affects the climbing energy consumption distribution in the S5 path, while the risk map of S4 dynamically adjusts the coverage priority of S5.

[0151] Execution and feedback: The robot's motion state (such as foot slip rate) and defect identification results (such as crack confidence) are transmitted back to the perception layer in real time, triggering parameter iterative optimization (such as increasing the refresh rate of the terrain stiffness matrix from 100Hz to 200Hz), forming a self-evolution capability.

[0152] Improved energy efficiency: The S5's MOEA / D optimization algorithm reduces inspection energy consumption to ≤4kJ / m, and combined with the S2's hybrid power mode (wheel + foot coordination), the range is increased by 40%. The underwater micro hydroelectric generator (50mm impeller diameter) outputs 40W of power at a flow velocity ≥1.5m / s, supporting uninterrupted operation.

[0153] Maintenance cost control: S3's small sample transfer model can achieve commercial-grade recognition accuracy (F1-score≥0.93) with only 500 labeled images, reducing data labeling costs by 90%; S4's risk prediction reduces the frequency of preventive maintenance by 50% and unplanned downtime losses by 80%.

[0154] Long-term adaptability: The dynamic SLAM map (S6) is updated in conjunction with the equipment health index (HI), and it automatically iterates with the transformation of power plant equipment (such as synchronizing the coordinates of newly installed pressure steel pipes to the grid map), avoiding redundant modeling.

[0155] S1 specifically includes:

[0156] The LiDAR scanning uses a 16-line rotating LiDAR to scan a preset range around the robot at a preset frequency, generating a three-dimensional point cloud terrain elevation map and capturing the terrain features of slope, convexity, and depression in real time.

[0157] IMU attitude perception, with a built-in six-axis inertial measurement unit, outputs robot body attitude angle and angular velocity data, and detects attitude fluctuations caused by terrain;

[0158] Foot force feedback: Each foot is embedded with a three-dimensional force sensor to measure the contact force vector between the foot and the ground in real time and calculate the distribution of ground reaction force.

[0159] Time synchronization: Align the time of the three types of sensors through hardware trigger signals; Spatial calibration: Establish the transformation matrix of the lidar coordinate system, IMU body coordinate system, and foot local coordinate system to unify the data to the robot's global coordinate system; Feature extraction: Fuse point cloud elevation data with foot force feedback to calculate the estimated value of the elastic modulus of the terrain surface.

[0160] Stiffness classification and quantification divides the terrain into grid cells, assigns a stiffness value to each cell, and classifies them into rigid surfaces, semi-rigid surfaces, and flexible surfaces, thus constructing a stiffness matrix.

[0161] Water reflection filtering is applied to lidar point clouds using a combined range-intensity filter.

[0162] Metal surface correction: In the metal structure area of ​​the power station, millimeter-wave radar-assisted point cloud correction is used.

[0163] Motion artifact suppression is achieved by compensating for point cloud distortion caused by the robot's own motion using IMU angular velocity data.

[0164] The matrix data is structured and output as a two-dimensional stiffness distribution map, with a confidence score for each grid cell.

[0165] When multiple consecutive grid stiffness values ​​are detected to be less than a preset threshold, an avoidance command is sent to the motion control system; when the grid stiffness gradient changes abruptly and exceeds the preset threshold, it is marked as a terrain discontinuity and preload control is activated.

[0166] It should be noted that the selection criteria for the LiDAR are as follows: using a 16-line rotating LiDAR (instead of a 32-line or 64-line LiDAR) is a balance between accuracy and power consumption. Within a 10m scanning range, the 0.1° angular resolution can identify surface undulations of ≥1cm (such as concrete joints and rust pits), while the power consumption is controlled at 45W (64-line LiDAR > 120W), which is suitable for the robot's endurance requirements.

[0167] Necessity of IMU attitude compensation: The common vibration environment in hydropower plants causes high-frequency shaking of the robot body. The six-axis IMU outputs the attitude angle at a frequency of 100Hz and solves it in real time through quaternion differential equations, suppressing point cloud distortion error by up to 80%.

[0168] A Mechanical Perspective from Foot Force Feedback: A 3D Force Sensor (500N Range) Not Only Measures Contact Force, But Also Infers Ground Deformation Through Hertzian Contact Theory.

[0169]

[0170] In the formula, d represents the depth of ground deformation; This is the normal component of the contact force, i.e., the pressure perpendicular to the contact surface; The radius of curvature of the foot tip is the curvature of the foot when it contacts the ground. This is the equivalent elastic modulus, a material property used to describe the elastic characteristics of a material.

[0171] Hardware synchronization: The PPS pulse signal is generated by the FPGA and transmitted to each sensor through the LVDS differential circuit to ensure that the timestamp alignment deviation is <1ms;

[0172] Coordinate system transformation, establishing a transformation chain from lidar to IMU to the foot, essentially involves solving for rigid body transformations on the Lie group SE(3):

[0173]

[0174] In the formula, This is the global transformation matrix, representing the transformation chain from the lidar coordinate system to the IMU coordinate system and then to the foot coordinate system; Let be the transformation matrix from lidar to IMU, representing the transformation of the lidar coordinate system relative to the IMU coordinate system; This is the transformation matrix from the IMU to the foot, representing the transformation of the IMU coordinate system relative to the foot coordinate system; Let T be the transformation matrix of the foot coordinate system, representing the transformation of the foot coordinate system; where T is a 4×4 homogeneous transformation matrix, which is calibrated offline by the hand-eye calibration method (AX=XB equation) and updated online by the extended Kalman filter (EKF).

[0175] Stiffness-graded hydroelectric compatibility:

[0176] Rigid surfaces (80-100): corresponding to concrete dams of hydropower stations (elastic modulus 25GPa) and steel plate linings;

[0177] Semi-rigid surface (50-80): covering compacted backfill soil (8GPa) and epoxy resin coating;

[0178] Flexible surface (0-50): specifically refers to the silt (0.5GPa) in the flood discharge tunnel and the vegetation-covered area.

[0179] Anti-interference mechanism: Wavelength selection for water reflection filtering: 905nm laser (not 1550nm) is used because of its low absorption rate in water (attenuation coefficient 0.2m⁻¹). Combined with the intensity-distance joint threshold: point clouds with intensity <30 and distance abrupt change >10cm are judged as water surface reflection noise; near-infrared band is additionally used to assist in identification in the flood discharge channel area (water absorption peak at 1450nm).

[0180] Metal surface correction: Millimeter-wave radar (77GHz) penetrates the paint layer to detect the substrate metal and performs ICP registration with the laser point cloud. When the offset is greater than 2mm, correction is triggered to solve the misjudgment caused by paint peeling in the inspection of pressure steel pipes.

[0181] Decision triggering mechanism:

[0182] Avoidance instruction: When the stiffness of three consecutive grids is <30 (equivalent elastic modulus <3GPa), calculate the soil shear strength based on the Coulomb-Mohr criterion. If the soil shear strength is <0.3·F foot If the maximum thrust at the foot is detected, an avoidance command is sent and the area is marked as an instability risk zone.

[0183] Bionic optimization of preload control: A preload strategy inspired by arachnids triggered by terrain discontinuities (stiffness gradient > 60 / cm).

[0184] The micro-spiky structure at the foot end (imitating cricket bristles) is embedded in the micropores of the ground; the hydraulic system outputs high-frequency pulsating pressure (20Hz, amplitude ±10%), which enhances the surface biting force by up to 40%.

[0185] S2 specifically includes:

[0186] The preset angle threshold setting of the slope classification decision mechanism, based on big data analysis of hydropower plant terrain, divides the slope into three decision intervals:

[0187] Flat terrain mode activates McCannum wheel drive, utilizing the omnidirectional wheel characteristics to achieve zero turning radius; medium slope mode switches to hydraulically driven six-legged gait, adopting a triangular gait sequence, with the three diagonal legs lifting simultaneously, and the stride dynamically adjusted to a preset ratio of foot length; extreme slope mode triggers biomimetic adsorption-gait coordinated control, with the foot end generating van der Waals force adsorption based on micro-nano structure biomimetic bristles.

[0188] The biomimetic adsorption-gait coordinated control includes: adsorption stage, where the foot pre-presses into the contact surface, activates the bristle array deformation locking, and controls the adsorption force through a force feedback closed loop, while releasing the negative pressure cavity to enhance adhesion; stepping stage, a support triangle is constructed, and three non-adjacent feet are selected to form a support surface, while the other three feet perform the lifting action; dynamic gait planning, where the lifting foot raises a preset height along the normal direction, and the stepping distance is adjusted according to the slope using a formula;

[0189] Switch the fault tolerance mechanism. When the single-foot slip rate is greater than the preset ratio, immediately freeze the current gait and trigger foot position remapping through IMU attitude feedback to reconstruct the support triangle within a preset time.

[0190] The multimodal seamless switching logic includes: transition condition detection, triggering mode switching when the continuous sampling value of the slope sensor crosses the threshold; wheeled to footed switching, retracting the wheel set and locking the axle; footed to adsorption switching, pre-extending the bristle array at the foot end, and powering on the adsorption module for preheating;

[0191] Energy consumption optimization strategy: In the transition zone, activate hybrid mode, i.e., front wheel keep drive and rear quadruped executes gait;

[0192] Enhanced adaptability to complex scenarios includes: anti-slip on wet surfaces, with a hydrophobic coating sprayed on the foot end in the wet and slippery area of ​​the turbine housing; magnetic interference response on metal surfaces, employing a non-magnetic titanium alloy foot end structure in the strong magnetic field environment of the generator nacelle; turbulent environment stabilization control, activating the vortex suppression algorithm and deflecting the thruster vector nozzle when the flow velocity in the turbine corridor reaches the threshold velocity; and real-time feedback of attitude angle by the IMU, with the PID controller outputting joint torque compensation commands.

[0193] It should be noted that 30° is the slope limit and 60° is the climbing threshold;

[0194] Motion modes:

[0195] The wheeled omnidirectional mobility features McCann wheels with four-wheel independent vector control. Each wheel is driven by a brushless motor (peak torque 35 N·m). The electromechanical principle that enables zero-radius turning is the diagonal wheel differential in opposite directions (left front wheel + right rear wheel counterclockwise, right front wheel + left rear wheel clockwise).

[0196] The six-legged gait energy-saving design, with the Tripod Gait, selects three diagonal legs for stepping. This combination ensures that the robot's center of mass projection always falls within the supporting triangle, reducing hydraulic joint power consumption by 30%.

[0197] Bionic Adsorption – Gait Synergy:

[0198] The micro-nano structure is biomimetic. The bristle array adopts the multi-level fractal structure of a gecko's foot (the first-level bristles are 200 μm in diameter and the second-level nanofibers are 20 nm in diameter). When the density is 14,000 bristles / cm², it generates an adsorption force of ≥150N. Its essence is the cumulative effect of intermolecular van der Waals forces.

[0199] The negative pressure enhancement mechanism releases a -20kPa negative pressure chamber at the center of the foot on a wet metal surface, and utilizes the water film sealing property to increase adhesion by 80%, solving the problem of pure van der Waals force failure in a liquid film environment.

[0200] Stability control of dynamic gait:

[0201] The supporting triangular region construction rule, selecting non-adjacent triplets (such as left front + right middle + left rear) maximizes the supporting area, covering ≥80% of the body projection. Its geometric constraints are:

[0202]

[0203] In the formula, The area is the support area, i.e., the area of ​​the selected tripod support region; m is the mass of the object; g is the gravitational acceleration, usually taken as 9.8 m / s². The allowable surface compressive strength of concrete is given here as 0.3 MPa (megapascals). The tilt angle is the degree to which an object is tilted relative to the horizontal plane.

[0204] The normal step height of 5-8cm is designed based on the typical obstacle height in hydropower plants (such as cable troughs and weld protrusions), and obstacle avoidance is achieved in real time through a Z-axis laser TOF sensor.

[0205] Slip tolerance:

[0206] Slip ratio detection, when the foot tangential velocity , and the speed of the body ,satisfy At that time, slippage is determined;

[0207] The key to the remapping acceleration mechanism, which completes position replanning within 0.3 seconds, is the use of the RRT*-Connect algorithm, combined with IMU attitude angular velocity. Predicting body pose:

[0208]

[0209] In the formula, This refers to the change in position. The velocity of the main body, that is, the velocity of the robot or object itself; The attitude angular velocity measured by the IMU (Inertial Measurement Unit) is used to describe the rotational speed of the object; r is the vector of the foot relative to the center of mass, that is, the position vector from the robot's center of mass to the foot. It is a time variable;

[0210] Wheel to Foot Switching Sequence:

[0211] 0-0.2 seconds: The electromagnetic lock releases the wheel assembly, and the hydraulic rod retracts at a speed of 50mm / s;

[0212] 0.2 to 0.5 seconds: The knee joint hydraulic cylinder is pressurized to 20 MPa, and the foot structure unfolds.

[0213] Preheating of the adsorption module: The bristle array is pre-stretched and energized to 100V high voltage, which causes the piezoelectric ceramic to undergo micro-deformation, increasing the contact area by 30%.

[0214] Anti-slip chemical-mechanical fusion:

[0215] The hydrophobic coating formulation, with a fluorosilicone nano-solution (C8F17CH2CH2Si(OCH3)3) having a contact angle >150°, forms a water-resistant film on the turbine housing steel plate (friction coefficient 0.25→0.62), with an injection volume of 0.1 ml / stool·time.

[0216] The material used to combat magnetic interference is Ti-6Al-4V titanium alloy, which has a permeability of <1.001 at the foot end and an eddy current loss of <3W in a 200μT magnetic field, thus avoiding the overheating and runaway of traditional steel joints.

[0217] S3 specifically includes:

[0218] The pre-trained architecture includes a backbone network, which is pre-trained on the ImageNet-21K dataset using Vision Transformer; contrastive learning optimization is performed by introducing the NT-Xent loss function to narrow the distance between features of the same type of defect and push apart features of different types.

[0219] For fine-tuning of the hydropower scenario, the migration strategy is to freeze the weights of the pre-set proportion layer before ViT and only fine-tune the top Transformer block and classification head; small sample training is conducted based on defect samples of hydropower equipment and hard sample mining is applied.

[0220] The classification decision logic takes a fusion tensor as input and outputs a defect probability vector. It sets two thresholds for judgment: if the confidence level is greater than a preset proportion, it directly outputs the classification result. If the confidence level is within the preset proportion range, it activates local reconstruction of the 3D point cloud and classifies the data after secondary verification based on the Poisson surface reconstruction algorithm.

[0221] Crack parameters were calculated: the length was obtained by fitting a B-spline curve along the point cloud normal vector direction; the depth was obtained by combining the infrared temperature gradient and the point cloud height difference; and corrosion levels were classified for further assessment.

[0222] The hydropower scenario adaptability enhancement is carried out in three steps: strong reflective surface treatment, applying polarized light imaging to the stainless steel casing of the generator, and combining point cloud intensity information to separate real defects from reflective artifacts; underwater imaging compensation, using blue-green light band penetration enhancement when inspecting the inner wall of pressure steel pipes; and dynamic target filtering, using optical flow motion detection for swimming fish and floating debris, combined with morphological filtering to eliminate non-static targets.

[0223] It should be noted that the ViT-Large backbone network has good hydropower compatibility.

[0224] Input resolution has been optimized to use 384×384 input (non-standard 224×224). Since the average length of cracks in hydroelectric equipment is >15mm (150 pixels are needed to cover them at 0.1mm / px), the high resolution ensures the ability to capture sub-millimeter defect textures.

[0225] Location encoding, based on ImageNet-21K pre-training, adds prior location encoding of metal surfaces—embedding the coordinate probability distribution of high-frequency water and electricity components such as welds and bolts into the Attention layer to improve the sensitivity of local features;

[0226] Decoupling of the defects of the NT-Xent loss function:

[0227] A clustering mechanism for similar defects applies hyperspherical constraints to corrosion defects (pitting corrosion, pitting corrosion, ulcer corrosion), with the spacing between similar eigenvectors being <0.1;

[0228] The heterogeneous feature rejection enhancement is achieved, with the distance between crack and corrosion features >0.8 and the stain artifact pushed to >1.2, thus solving the confusion problem of traditional cross-entropy in morphologically similar defects.

[0229] Physics-driven negative samples are generated based on thermodynamic simulation to create virtual cracks (temperature gradient field + stress distribution), thus expanding the diversity of negative samples.

[0230] Overfitting suppression in migration strategies, and dynamic adjustment of the number of frozen layers:

[0231] Table 1. Overfitting Suppression Table for Transfer Strategies

[0232]

[0233] The gradient redirection mechanism introduces a residual gradient amplifier in the fine-tuning layer, which improves the convergence speed of small sample training by 3 times.

[0234] Enhanced adversarial mining of hard samples: a three-stage mining strategy.

[0235] Initial screening: Samples with a confidence level of 40% to 70% in the first round of training;

[0236] Adversarial generation: Generate deformation crack / corrosion samples (deformation rate ±15%) using DCGAN;

[0237] Thermodynamic verification: Samples generated that do not conform to the characteristics of cracks due to infrared temperature gradients are discarded.

[0238] Semi-supervised self-iteration: For unlabeled samples, the Mean Teacher model is used to generate pseudo-labels, and samples with a confidence level > 85% are included in the training set.

[0239] Dual threshold classification:

[0240] The direct judgment threshold (90%) corresponds to a model prediction variance of <0.01 and a classification error rate of <0.5%; the point cloud reconstruction threshold (60%~90%) covers typical hydropower interference scenarios (oil stain coverage, rust layer obstruction).

[0241] Poisson's five-step surface reconstruction process:

[0242] Extract the point cloud of the target region (sphere with a radius of 5cm); estimate the normal vector (KD-Tree nearest neighbor search + PCA); construct the implicit function field: Solve the Poisson equation to generate an isosurface; extract the crack centerline (skeletalization algorithm); where, This represents the change in the implicit function field, i.e., the gradient change of the implicit function during surface reconstruction. It is the surface normal vector, that is, the direction vector perpendicular to the surface;

[0243] Thermal-mechanical coupling model for deep computation:

[0244]

[0245] In the formula, This refers to the crack depth. The maximum temperature difference at the crack edge is usually obtained through infrared imaging technology; The point cloud height difference is typically obtained using 3D scanning technologies such as LiDAR. It is the coefficient of thermal expansion, which is related to the thermal expansion characteristics of the material; This is the coefficient of thermal expansion of the material, specifically for steel, with a value of 1.2 × 10−5 / ∘C.

[0246] S4 specifically includes:

[0247] Historical equipment failure data is integrated into the hydropower plant's SCADA system to extract past equipment failure records and construct a failure probability density function; environmental corrosion rate is acquired in real time through weather stations and corrosion sensors, and is divided into atmospheric corrosion and water corrosion; task urgency is analyzed from maintenance work order data and classified by priority.

[0248] Spatial discretization is used to divide the inspection area into grid cells, assigning a unique coordinate to each cell. Dynamic risk value is calculated, with the risk value of each grid cell derived from a formula. Crack / corrosion results are superimposed, and if defects identified in step S3 exist within the grid, the risk value is multiplied according to the defect level.

[0249] Heatmaps are generated, and risk values ​​are mapped using the HSV color space. The tiered response strategy includes high-risk grids, which immediately trigger the robot to re-inspect, increasing the data feedback frequency; medium-risk grids, which are included in the nodes that must be passed through in the optimized path; and low-risk grids, which are covered as needed and can be skipped.

[0250] It should be noted that SCADA deep mining involves accessing a 10-year operation and maintenance database of a hydropower plant and fitting the fault probability density function through a Weibull distribution. Spatiotemporal correlation is enhanced by adding a gamma process degradation model to high-frequency fault points such as weld seams of pressure steel pipes, and combining the operating time (in thousands of hours) to predict the remaining lifespan.

[0251] Real-time coupling of environmental corrosion rates:

[0252] Atmospheric corrosion, based on meteorological station data (temperature, humidity, SO2 concentration), the corrosion rate according to ISO 9223 was calculated.

[0253] For water corrosion, pH and Cl⁻ concentrations are obtained through electrochemical sensors, and the corrosion current density is output in real time using the linear polarization resistance method.

[0254] Intelligent hierarchical classification of work orders:

[0255] Level 1 (>0.9): Defects involving generator set shutdown (such as stator grounding faults);

[0256] Level 2 (0.6-0.9): Sub-healthy state affecting efficiency (such as abnormal bearing temperature rise);

[0257] Level 3 (<0.6): Routine preventative maintenance.

[0258] Weighting: Fault history (40%), environmental corrosion (30%), task urgency (30%). The weights are dynamically adjusted according to the equipment type (fault history is the main factor for transformers, and corrosion is the main factor for pressure steel pipes).

[0259] Spatial Discretization Parameter Design

[0260] Grid size optimization: 0.5m×0.5m cells (not 1m²) balance accuracy and computing power, ensuring that the smallest defect (>5cm crack) is covered by a single grid;

[0261] The coordinate mapping rule adopts the UTM projected coordinate system, with the origin locked at the power plant's central control center;

[0262] Dynamic Risk Value Core Algorithm, Basic Risk Formula:

[0263]

[0264] In the formula, The base risk value is the risk level calculated after taking into account various factors. The dynamic weight of equipment failure represents the degree to which equipment failure contributes to the total risk. This refers to the probability of equipment failure, that is, the likelihood of equipment failure. The dynamic weights of environmental factors represent the degree to which environmental factors (such as weather and hydrological conditions) contribute to the total risk. This is an air risk value, which may be related to air quality or atmospheric conditions; This is a hydrological risk value, which may be related to water level, water flow, and other hydrological conditions. The dynamic weight for task urgency represents the degree to which the urgency of the task contributes to the total risk. Task urgency is an indicator of how quickly a task needs to be completed.

[0265] Table 2 Defect Multiplication Mechanism Table:

[0266]

[0267] Heatmap generation, HSV color mapping:

[0268] High risk (R≥0.8): H=0° (red), S=100%, V=100%;

[0269] Medium risk (0.4≤R<0.8): H=30° (orange-yellow gradient);

[0270] Low risk (R<0.4): H=120° (green), V is reduced to saturation according to risk value;

[0271] Human-machine collaborative design, heat map overlay SLAM point cloud model, supports AR glasses perspective viewing (Microsoft HoloLens 3), defect location error <2cm.

[0272] Closed-loop control logic of the hierarchical response strategy:

[0273] Strong intervention for high-risk grids (R≥0.8): Re-inspection path planning, the robot directly drives the target grid with the shortest time trajectory (A* algorithm + dynamic constraints); Sensors operate at full power: laser scanning frequency 100Hz→200Hz, infrared thermal imaging frame rate 30fps→60fps; Data backhaul mechanism, enabling 5G private network QoS guarantee channel, data packet priority is upgraded to DSCP 46 (EF level), latency <50ms;

[0274] Path coupling of medium-risk grids (0.4≤R<0.8): mandatory node constraints, additional hard constraints in MOEA / D-TOPSIS optimization: the path must cover the centroids of all medium-risk grids; detection parameter adaptation: laser scanning frequency downgraded to 50Hz, infrared temperature measurement accuracy maintained at ±0.5℃;

[0275] Intelligent de-loading of low-risk grids (R<0.4): On-demand coverage rule, when the path length is >1.2×perimeter, skipping consecutive low-risk grids (<3); for areas with historically fault-free equipment, the detection frequency is reduced to 30% of the standard value.

[0276] S5 specifically includes:

[0277] Optimization objectives include minimizing path length, maximizing coverage, and minimizing energy consumption; constraint embedding includes mandatory coverage of high-risk areas, time window limits, and physical accessibility.

[0278] MOEA / D-TOPSIS Hybrid Optimization Process:

[0279] Population initialization generates several initial paths, each path being a grid coordinate sequence, with the starting point fixed at the hydropower plant control center and the ending point at the charging pile;

[0280] MOEA / D decomposition and collaboration decomposes a multi-objective problem into several sub-problems; neighborhood definition selects the sub-problems with the closest Euclidean distance as the collaborative neighborhood; sub-problem solving uses a differential evolution operator to generate new paths;

[0281] TOPSIS optimal solution decision-making calculates the approximation distance to the ideal solution for each path and selects the path with the smallest approximation distance as the final output;

[0282] When S3 detects a new crack and the risk value is greater than the preset threshold, it generates an enhanced detection zone around the risk grid; interrupts the current path and calculates the shortest detour path using the A* algorithm; and updates the TOPSIS weights.

[0283] It should be noted that the MOEA / D-TOPSIS hybrid optimization is as follows:

[0284] Scenario-driven strategies for population initialization:

[0285] The starting point and the endpoint are locked. The starting point is fixed at the central control center (coordinates [0,0]) and the endpoint is the nearest charging pile (distance ≤20m).

[0286] Heuristic initial path:

[0287] 50% population: generated based on risk heatmap (high-risk raster density weighted);

[0288] 50% population: Constructing the maximum coverage path using the Voronoi diagram method.

[0289] Electromechanical coupling optimization based on MOEA / D decomposition:

[0290] Subproblem decomposition method: The target space is uniformly divided into N subproblems (N = population size), and the weight vector λ i =(λi1 ,λ i2 ,λ i3 ) satisfies ∑λ ij =1; where λ is the value of 1. i λ represents the weight vector of the i-th subproblem; i1 ,λ i2 ,λ i3 Represents the weight vector λ i The components in ∑λ ij =1 indicates that the weight vector λ i The sum of all components must be equal to 1. This is a fundamental property of the weight vector, ensuring that the sum of the weights of each subproblem is 1.

[0291] Path repair mechanism: If the mutated path violates the constraints, B-spline interpolation is used to smoothly bypass the obstacle.

[0292] TOPSIS Optimal Solution Dynamic Decision-Making:

[0293] Ideal solution distance calculation:

[0294]

[0295] In the formula, The distance to the ideal positive ideal solution of the i-th solution; The distance between the ideal and negative ideal solutions of the i-th solution; The weight of the j-th objective; Let be the normalized value of the i-th solution on the j-th objective; This is the positive ideal value (optimal value) for the j-th objective. The negative ideal value (worst-case value) for the j-th objective.

[0296] Proximity: Select C i →1 is the optimal solution.

[0297] Strengthen the rules for generating detection zones:

[0298] Triggering conditions: S3 identifies cracks of level 2 or higher and the local risk value R ≥ 0.7;

[0299] Region construction: Centered on the defect grid, expand a 3×33×3 neighborhood (a total of 9 grids).

[0300] Detection accuracy upgrade: laser scanning lines increased from 16 to 32, and infrared resolution increased from 640×480 to 1280×1024.

[0301] Interruption recovery mechanism: After a new path is inserted, the original path is executed in segments (preserving the covered area); the charging pile endpoint is dynamically reset to the nearest available node.

[0302] S6 specifically includes:

[0303] Beidou RTK positioning, equipped with a dual-frequency receiver, uses ground reference station differential correction of ionospheric error to output centimeter-level absolute coordinates; in response to the multipath effect in the canyon of hydropower plant, a multipath suppression algorithm is used, including reducing the weight by a corresponding proportion when the signal incident angle is greater than a preset threshold angle, carrier phase cycle slip detection, and switching to Beidou third generation B2b signal when abnormal.

[0304] UWB anchor point positioning deploys a self-calibrating anchor point network in the inspection area, adopts TDOA technology, and the robot carries tags to communicate at a preset frequency; an anchor point drift self-correction mechanism is introduced, a point cloud map is built based on laser SLAM, the anchor point position offset is calculated, and when the offset exceeds the preset distance, the UWB anchor point position is updated online.

[0305] Inertial navigation INS, nine-axis MEMS IMU, attitude angles are calculated using the fourth-order Runge-Kutta method; built-in temperature compensation chip is used to address temperature drift;

[0306] In ideal environments, strong magnetic interference, and satellite denial, the weights of BeiDou RTK, UWB, and INS are adaptively allocated; based on the state vector, the three-source data are fused using the observation equation; when the variance of any source data mutation exceeds a preset multiple of the historical variance mean, the fusion is paused and SLAM pose prediction is enabled.

[0307] It should be noted that BeiDou RTK:

[0308] Ionospheric error: Dual-frequency receiver selection, using a B1I+B2a dual-frequency combination (not a single frequency), and utilizing the dispersion effect to calculate the ionospheric delay.

[0309] Ground reference station layout rules: reference station spacing ≤ 5km (hydropower canyon terrain); fixed station coordinate accuracy 0.1ppm (1mm error for 10km baseline).

[0310] Canyon multipath effect:

[0311] Incident angle threshold control: when the incident angle θ > 60°, the weight is reduced to w = cos2θ (vertical reflection signals are completely eliminated); signals with a signal-to-noise ratio (SNR) < 45dB in the cliff reflection zone are directly discarded.

[0312] Cycle slip detection and third-generation signal switching: When the carrier phase residual is >0.25 cycles, switch to BeiDou third-generation B2b signal (anti-interference gain 20dB); the switching delay is <100ms to ensure positioning continuity.

[0313] Example 3:

[0314] Reference Figure 2As shown, this embodiment provides an all-terrain adaptive hydropower AI bionic robot intelligent inspection system, including:

[0315] Multimodal bionic mobile platform: integrates three-state drive modules of wheel, leg and propeller propulsion, adopts bionic joint group collaborative control algorithm, and is optimized based on biological central pattern generator CPG model; equipped with active deformable track structure, which can automatically adjust the ground area according to the terrain slope;

[0316] Multi-source sensing fusion module: The cross-media environment sensing unit is equipped with lidar, millimeter-wave radar, and multispectral industrial camera;

[0317] Electromagnetic interference suppression layer: The sensor is encased in a Mu metal shielding cover;

[0318] Edge-AI Decision Center: Deploys lightweight Transformer models to support the identification of subtle device anomalies; integrates a dynamic SLAM enhancement engine.

[0319] It should be noted that the edge-AI decision-making center:

[0320] Lightweight Transformer defect detection model, ViT-Base optimized architecture:

[0321] Input resolution 384×384, 12 Transformer Blocks; position encoding, embedding device prior coordinates;

[0322] Dynamic knowledge distillation:

[0323] Teacher model: ViT-Large (ImageNet-21K pre-trained);

[0324] Student model: Only 12M parameters after pruning (inference latency < 8ms / Tesla T4).

[0325] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0326] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0327] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0328] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0329] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An intelligent inspection method for all-terrain adaptive hydropower AI bionic robots, characterized in that, include: Acquire and generate a terrain stiffness matrix based on data from lidar, IMU, and foot force sensor; Simultaneously acquire visible light images, infrared thermal imaging information, and laser point cloud information, and output crack / corrosion classification results based on a contrastive learning transfer model; Based on the crack / corrosion classification results, a dynamic risk grid map is constructed by integrating equipment failure history, environmental corrosion rate, and task urgency. Based on the dynamic risk grid map, three-source positioning is fused, and dynamic SLAM enhancement is performed to obtain the enhanced map; Based on the enhanced map, MOEA / D-TOPSIS path optimization is performed, simultaneously satisfying the preset proportion of path length being less than the perimeter of the area, and controlling coverage and energy consumption within preset ranges to obtain the optimized path. Multimodal motion decisions are made based on the optimized path and terrain stiffness matrix.

2. The all-terrain adaptive hydropower AI bionic robot intelligent inspection method according to claim 1, characterized in that, Data from lidar, IMU, and foot force sensors is acquired to generate a terrain stiffness matrix, including: A 16-line rotating lidar is used to scan a preset range around the robot at a preset frequency to generate a three-dimensional point cloud terrain elevation map, and to capture the terrain features of slope, convexity and depression in real time to obtain lidar data. It has a built-in six-axis inertial measurement unit, which outputs the robot's attitude angle and angular velocity data, detects attitude fluctuations caused by terrain, and obtains IMU data; A three-dimensional force sensor is embedded in each foot to measure the contact force vector between the foot and the ground in real time, calculate the distribution of ground reaction force, and obtain foot force sensor data.

3. The all-terrain adaptive hydropower AI bionic robot intelligent inspection method according to claim 2, characterized in that, A terrain stiffness matrix is ​​generated based on data from lidar, IMU, and foot force sensors, including: Time synchronization is achieved by aligning the time of data from the lidar, IMU, and foot force sensor via hardware trigger signals. Spatial calibration involves establishing a transformation matrix between the lidar coordinate system, the IMU body coordinate system, and the foot local coordinate system to unify the data into the robot's global coordinate system. Feature extraction, fusion of point cloud elevation data and foot force feedback, and calculation of the estimated value of the elastic modulus of the terrain surface; Stiffness classification and quantification divides the terrain into grid cells, assigns a stiffness value to each cell, and classifies them into rigid surfaces, semi-rigid surfaces, and flexible surfaces, thus constructing a stiffness matrix. Water reflection filtering is performed by applying range-intensity joint filtering to lidar point cloud data. Metal surface correction: In the metal structure area of ​​the power station, millimeter-wave radar-assisted point cloud correction is used. Motion artifact suppression is achieved by compensating for point cloud distortion caused by the robot's own motion using IMU angular velocity data. The matrix data is structured and output as a two-dimensional stiffness distribution map, with a confidence score for each grid cell. When multiple consecutive grid stiffness values ​​are detected to be less than a preset threshold, an avoidance command is sent to the motion control system; when the grid stiffness gradient changes abruptly and exceeds the preset threshold, it is marked as a terrain discontinuity and preload control is activated.

4. The all-terrain adaptive hydropower AI bionic robot intelligent inspection method according to claim 1, characterized in that, Simultaneously acquire visible light images, infrared thermal imaging information, and laser point cloud information, including: A global shutter industrial camera is used to capture RGB images of the equipment surface; a ring LED fill light is applied, and an adaptive homomorphic filter is used to obtain a visible light image; Equipped with an uncooled microbolometer, it generates a temperature distribution map; to address the interference from metal surface reflection, it employs a dynamic emissivity correction algorithm and compensates for temperature measurement errors based on ambient temperature and humidity to obtain infrared thermal imaging information; Using a ToF lidar, a divergence angle beam is emitted to acquire sub-millimeter level three-dimensional point clouds on the device surface; to address water refraction interference, a Snell's law compensation model is applied to correct underwater point cloud distortion, thus obtaining laser point cloud information.

5. The all-terrain adaptive hydropower AI bionic robot intelligent inspection method according to claim 4, characterized in that, Based on a contrastive learning transfer model, the output crack / corrosion classification results include: Synchronization pulse signals are generated using FPGA to align the timestamps of the three sensors; a multi-sensor joint coordinate system is established based on a checkerboard calibration board. For visible light images, texture features, edge features, and color anomaly regions are extracted; for infrared thermal imaging information, temperature gradient fields and morphological contours of overheated areas are extracted; for laser point clouds, surface curvature changes and normal vector deviations are calculated; a multi-channel fusion tensor is constructed to map visible light texture, infrared temperature gradient, and point cloud curvature to a unified dimensional space, and key features are enhanced through an attention weighting mechanism. The contrastive learning transfer model includes a backbone network, which is pre-trained on the ImageNet-21K dataset using Vision Transformer; contrastive learning optimization is performed by introducing the NT-Xent loss function to narrow the distance between features of the same type of defect and push apart features of different types. Fine-tuning of the hydropower scenario was performed, and the migration strategy was to freeze the weights of the pre-set proportion layer before ViT and only fine-tune the top Transformer block and classification head; small sample training was performed, based on defect samples of hydropower equipment, and hard sample mining was applied. The classification decision logic is implemented, with a fusion tensor as input and a defect probability vector as output. A dual threshold is set for judgment: if the confidence level is greater than a preset proportion, the classification result is directly output. If the confidence level is within the preset proportion range, local reconstruction of the 3D point cloud is activated, and classification is performed after secondary verification based on the Poisson surface reconstruction algorithm. Crack parameters were calculated. The length was obtained by fitting a B-spline curve along the normal vector direction of the point cloud, and the depth was obtained by combining the infrared temperature gradient and the height difference of the point cloud. Enhanced adaptability for hydropower scenarios: High-reflectivity surface treatment; applying polarized light imaging to the stainless steel casing of generators, combining point cloud intensity information to separate real defects from reflective artifacts; underwater imaging compensation; enhanced penetration through blue-green light bands when inspecting the inner wall of pressure steel pipes; dynamic target filtering; using optical flow motion detection for swimming fish and floating debris, combined with morphological filtering to eliminate non-static targets.

6. The all-terrain adaptive hydropower AI bionic robot intelligent inspection method according to claim 1, characterized in that, Based on crack / corrosion classification results, and integrating equipment failure history, environmental corrosion rate, and task urgency, a dynamic risk grid map is constructed, including: Acquire historical equipment failure data, connect to the hydropower plant's SCADA system, extract past equipment failure records, and construct a failure probability density function; environmental corrosion rate, acquired in real time through weather stations and corrosion sensors, is divided into atmospheric corrosion and water corrosion; task urgency, parsed maintenance work order data, is classified by priority. Spatial discretization is performed, dividing the inspection area into grid cells, assigning unique coordinates to each cell, and dynamically calculating the risk value. The risk value of each grid cell is calculated by a formula. Crack / corrosion results are superimposed, and if an identified defect exists within the grid, the risk value is multiplied according to the defect level. Heatmaps are generated, and risk values ​​are mapped using the HSV color space. The tiered response strategy includes high-risk grids, which immediately trigger the robot to re-inspect them, increasing the data feedback frequency; medium-risk grids, which are included in the nodes that must be passed through in the optimized path; and low-risk grids, which are covered as needed and allowed to be skipped.

7. The all-terrain adaptive hydropower AI bionic robot intelligent inspection method according to claim 1, characterized in that, By fusing three-source localization and performing dynamic SLAM enhancement, an enhanced map is obtained, including: To obtain BeiDou RTK positioning, a dual-frequency receiver is used to differentially correct ionospheric errors through ground reference stations and output centimeter-level absolute coordinates. To address the multipath effect in the canyon of the hydropower plant, a multipath suppression algorithm is implemented, including reducing the weight by a corresponding proportion when the signal incident angle is greater than a preset threshold angle, carrier phase cycle slip detection, and switching to BeiDou-3 B2b signal when an anomaly occurs. UWB anchor point positioning is obtained, a self-calibrating anchor point network is deployed in the inspection area, TDOA technology is adopted, and the robot carries a tag to communicate at a preset frequency; an anchor point drift self-correction mechanism is introduced, a point cloud map is built based on laser SLAM, the anchor point position offset is calculated, and the UWB anchor point position is updated online when the offset exceeds a preset distance. The system acquires inertial navigation (INS) data from a nine-axis MEMS IMU and calculates attitude angles using the fourth-order Runge-Kutta method; it also incorporates a built-in temperature compensation chip to address temperature drift. In ideal environments, strong magnetic interference, and satellite denial, the weights of BeiDou RTK, UWB, and INS are adaptively allocated; based on the state vector, the three-source data are fused using the observation equation; when the variance of any source data mutation exceeds a preset multiple of the historical variance mean, the fusion is paused and SLAM pose prediction is enabled.

8. The all-terrain adaptive hydropower AI bionic robot intelligent inspection method according to claim 1, characterized in that, Based on the enhanced map, MOEA / D-TOPSIS path optimization is performed, simultaneously satisfying a preset proportion where the path length is less than the area perimeter, and controlling coverage and energy consumption within preset ranges to obtain optimized paths, including: Optimization objectives include minimizing path length, maximizing coverage, and minimizing energy consumption; constraint embedding includes mandatory coverage of high-risk areas, time window limits, and physical accessibility. MOEA / D-TOPSIS Hybrid Optimization Process: Population initialization generates several initial paths, each path being a grid coordinate sequence, with the starting point fixed at the hydropower plant control center and the ending point at the charging pile; MOEA / D decomposition and collaboration decomposes a multi-objective problem into several sub-problems; neighborhood definition selects the sub-problems with the closest Euclidean distance as the collaborative neighborhood; sub-problem solving uses a differential evolution operator to generate new paths; TOPSIS optimal solution decision-making calculates the approximation distance to the ideal solution for each path and selects the path with the smallest approximation distance as the final output; When a new crack is detected and the risk value is greater than a preset threshold, an enhanced detection zone is generated around the risk grid; the current path is interrupted, and the shortest detour path is calculated using the A* algorithm; the TOPSIS weights are updated.

9. The all-terrain adaptive hydropower AI bionic robot intelligent inspection method according to claim 1, characterized in that, Based on the optimized path and terrain stiffness matrix, multimodal motion decisions are performed, including: The wheel movement mode is activated when the slope is less than the preset angle range. When the slope is within the preset angle range, it switches to six-legged gait mode; When the slope is greater than the preset angle range, the biomimetic adsorption-gait coordination mode is triggered. Wheeled movement modes include: activating the McCannum wheel drive to achieve a zero turning radius by utilizing the omnidirectional wheel characteristics; The six-legged gait mode includes: switching to hydraulically driven six-legged gait, adopting a triangular gait sequence, with the three diagonal legs lifting simultaneously, and the stride being dynamically adjusted to a preset ratio of the foot length; The biomimetic adsorption-gait coordinated control mode includes: During the adsorption phase, the foot tip is pre-pressed to contact the surface, activating the deformation and locking of the bristle array. The adsorption force is controlled through a force feedback closed loop, while the negative pressure cavity is released to enhance the adhesion. During the stepping phase, a support triangle is constructed, and three non-adjacent legs are selected to form a support surface, while the other three legs perform the stepping motion. Dynamic gait planning: the stepping foot is raised to a preset height along the normal direction, and the step distance is adjusted according to the slope using a formula.

10. An all-terrain adaptive hydropower AI bionic robot intelligent inspection system, characterized in that, include: Multimodal bionic mobile platform: integrates three-state drive modules of wheel, leg and propeller propulsion, adopts bionic joint group collaborative control algorithm, and is optimized based on biological central pattern generator CPG model; equipped with active deformable track structure, which can automatically adjust the ground area according to the terrain slope; Multi-source sensing fusion module: The cross-media environment sensing unit is equipped with lidar, millimeter-wave radar, and multispectral industrial camera; Electromagnetic interference suppression layer: The sensor is encased in a Mu metal shielding cover; Edge-AI Decision Center: Deploys a lightweight Transformer model to support the identification of subtle device anomalies; integrates a dynamic SLAM enhancement engine to perform the methods described in any one of claims 1-9.

Images