Anti-interference adaptive control system and method of tower inspection robot in dynamic environment

Abstract

The application relates to the technical field of robot control, and discloses a dynamic environment tower inspection robot anti-interference adaptive control system and method. In view of the problems that the interference is difficult to decouple under high-altitude strong wind and guide rail deformation, and the instability caused by the gravity center deviation, the application contains modules such as dynamics solving, state estimation and adaptive control. The wind load and displacement fluctuation are taken into an extended state vector to perform Kalman filtering to realize physical level decoupling of the interference; a nonlinear optimization is solved to generate an augmented reference trajectory balanced in the coding gravity center and the normal force, and a sliding mode control torque is output based on the Lyapunov theory to eliminate instruction conflicts. The application realizes multi-target collaborative robust control, and guarantees the safe and intelligent inspection of high-voltage towers.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and in particular to an anti-interference adaptive control system and method for a tower inspection robot in dynamic environments. Background Technology

[0002] Safety inspection of high-voltage transmission line towers is a crucial link in ensuring the stable operation of the power grid. Automated inspections using climbing robots along the tower's fall arrestor rails have become an industry trend. However, the high-altitude natural environment of the towers is extremely harsh. During operation, the robot is inevitably subjected to the combined effects of multiple dynamic environmental factors, including strong winds, tower foundation vibrations, and flexible deformation of the rails. Under these dynamic conditions, existing robot control systems struggle to effectively decouple external environmental interference from the mixed noise collected by sensors. This results in a significant decrease in the system's perception accuracy of its own true motion state and external forces, consequently affecting the accuracy of subsequent control. Meanwhile, when the robot encounters extreme conditions such as strong crosswinds or large-angle tilts, the center of gravity of the whole machine often deviates significantly from the center line of the guide rail, resulting in an extremely uneven distribution of contact force between each drive wheel and the guide rail. This can easily lead to serious safety accidents such as wheel slippage or even the whole machine overturning and derailing. Traditional control strategies usually use multiple independent sub-controllers to handle trajectory tracking, anti-slip control and center of gravity compensation tasks respectively. There is a lack of underlying dynamic coordination mechanism between the sub-controllers, which can easily cause control command conflicts under complex working conditions, making it impossible for the system to maintain a robust and stable operating state. Summary of the Invention

[0003] The purpose of this invention is to provide an anti-interference adaptive control system and method for a tower inspection robot in a dynamic environment, so as to solve the technical problems of existing robots in high-altitude strong winds and flexible deformation of guide rails, such as difficulty in decoupling interference, instability caused by center of gravity shift, and conflict of multi-target control commands.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] An anti-interference adaptive control system for a tower inspection robot in a dynamic environment includes a perception layer, a decision layer, an execution layer, and an energy management layer. The perception layer is equipped with multimodal sensors, and the execution layer includes a drive actuator for driving the robot to move along the tower guide rail. The control system includes a multiphysics coupled dynamics solution module, a multimodal perception and state estimation module, an anti-interference adaptive control module, and an adaptive dual-mode drive execution module. The multiphysics coupled dynamics solution module presets the structural parameters of the robot body and the guide rail and constructs dynamic equations. The multimodal perception and state estimation module uses extended Kalman filtering combined with multimodal sensor data to output the robot's generalized coordinate estimate and the decoupled environmental interference estimate. The anti-interference adaptive control module calculates the center of gravity offset based on the generalized coordinate estimate, generates an augmented reference trajectory that positions the center of gravity on the guide rail centerline and uniformly distributes the normal force of the drive wheels, constructs a Lyapunov candidate function based on the augmented reference trajectory, and calculates the total control torque including equivalent control torque and robust switching torque. The adaptive dual-mode drive execution module decomposes the total control torque and controls the action of the drive actuator.

[0006] As a preferred technical solution of the present invention, the multiphysics coupling dynamics solution module establishes a five-degree-of-freedom multibody dynamics system using the Lagrange equation. The five generalized coordinates are displacement along the guide rail direction, displacement perpendicular to the guide rail direction, roll angle around the guide rail axis, swing arm joint angle, and joint arm angle. The potential energy term of the system includes gravitational potential energy, elastic deformation potential energy of the guide rail based on equivalent stiffness, and magnetic attraction potential energy based on nonlinear decay of the air gap magnetic circuit equation.

[0007] As a preferred technical solution of the present invention, the state vector of the extended Kalman filter in the multimodal perception and state estimation module is a twelve-dimensional vector containing generalized coordinates, generalized velocity, wind load disturbance estimate and guide rail displacement fluctuation estimate. The discretized dynamic equation is used as the state transition equation, and the Jacobian matrix is ​​used to calculate the error covariance in the prediction update cycle to output the pure pose and velocity estimates as well as the feedforward compensation of the separated wind load and guide rail displacement fluctuation.

[0008] As a preferred technical solution of the present invention, the anti-interference adaptive control module applies corrections to the swing arm joint angle and joint arm angle in the basic reference trajectory by solving a nonlinear optimization problem to obtain the augmented reference trajectory. The objective function of the nonlinear optimization problem includes the square term of the center of gravity offset and the square term of the deviation of the actual normal force of each drive wheel from the target uniform normal force.

[0009] As a preferred technical solution of the present invention, the Lyapunov candidate function consists of the kinetic energy metric term of the augmented tracking error on the sliding surface and the disturbance upper bound estimation error term. The equivalent control torque in the total control torque includes the feedforward compensation of the modeled dynamics term and the environmental disturbance estimate. The robust switching torque includes the adaptive estimate of the disturbance upper bound and the continuous approximation function for eliminating high-frequency flutter.

[0010] As a preferred technical solution of the present invention, the adaptive dual-mode drive execution module includes a magnetic drive unit suitable for steel guide rails and a ratchet drive unit suitable for non-ferromagnetic guide rails. The control system switches the drive mode according to the guide rail material identified by the inductive sensor. In the steel-aluminum mixed transition section, a gradual switching strategy is adopted in which the magnetic attraction force and the ratchet clamping force increase and decrease synchronously and linearly.

[0011] As a preferred technical solution of the present invention, the anti-interference adaptive control module further includes a model predictive control trajectory optimization submodule. This submodule runs on the outer loop of the adaptive sliding mode controller and solves the optimization problem constrained by joint limits, torque saturation and obstacle safety distance through a sequential quadratic programming algorithm, thereby generating a smooth reference trajectory for multiple future prediction steps online.

[0012] As a preferred technical solution of the present invention, the system also includes an energy management and scheduling module, which adopts a hybrid power supply architecture of lithium battery and supercapacitor. Based on the robot's current load rate, the state of charge of the lithium battery and the remaining energy ratio of the supercapacitor, the supercapacitor discharge power ratio is output through a fuzzy inference engine, and the supercapacitor is controlled to release a large current pulse when there is a momentary high power demand.

[0013] An anti-interference adaptive control method for a tower inspection robot under dynamic environment, applied to the above-mentioned control system, includes the following steps:

[0014] The system initializes and loads the dynamic model parameters, and determines the driving mode based on the guide rail material;

[0015] Continuously perform multimodal environmental perception, and collect body motion state data, contact force data, and environmental meteorological data;

[0016] The extended Kalman filter is executed with a preset control cycle. The state is predicted based on the dynamic equation and updated through multi-source observations. The decoupled pure system pose estimate and environmental disturbance estimate are output.

[0017] The center of gravity offset is calculated based on the system pose estimation value. The nonlinear static optimization problem is solved online to generate an augmented reference trajectory that achieves center of gravity repositioning and normal force equilibrium. The total control torque of the adaptive sliding mode is calculated based on Lyapunov stability theory.

[0018] The total control torque is decomposed into speed commands for each drive wheel and joint servo angle commands, which drive the robot to perform anti-interference adaptive actions.

[0019] As a preferred technical solution of the present invention, the method also performs multi-level safety protection steps. In each control cycle, the derivative value of the Lyapunov function is calculated. If the derivative value is greater than zero and continues for more than a preset time, the upper limit of the disturbance is increased and deceleration protection is triggered. When the center of gravity offset exceeds the absolute safety threshold or the normal force of any drive wheel drops below the critical value, a hardware interrupt is triggered to make the robot perform rail-hugging braking and cut off the power supply of the drive motor.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] This invention constructs an extended state vector incorporating wind load disturbances and guide rail displacement fluctuations, and introduces an extended Kalman filter. Utilizing a multiphysics coupled dynamics model constraint, it achieves precise decoupling of external interference signals and sensor measurement noise at the physical mechanism level, significantly improving the robot's state estimation accuracy and anti-interference identification capability under strong wind and vibration environments. Furthermore, by integrating the implicit encoding of center of gravity offset compensation and drive wheel normal force balancing targets into the online generation process of the augmented reference trajectory, this invention transforms traditional multi-objective conflict control into a single Lyapunov trajectory tracking problem. This allows the adaptive sliding mode control law to mathematically guarantee the coordinated convergence of trajectory tracking, anti-slip, and anti-tipping subtasks, completely eliminating the potential for control command conflicts under extreme conditions and significantly improving the global robustness and stability of the robot system and the safety of high-altitude operations. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall architecture of the control system of the present invention;

[0023] Figure 2 This is a schematic diagram of the adaptive control method of the present invention.

[0024] Illustrations: 100. Multiphysics Coupled Dynamics Solution Module; 101. Robot Kinematics Parameter Library Submodule; 102. Guide Rail Physical Property Parameter Library Submodule; 103. Lagrange Dynamics Equation Construction Submodule; 104. Critical Instability Frequency Identification Submodule; 200. Multimodal Perception and State Estimation Module; 201. Sensor Array Submodule; 202. Multimodal Data Spatiotemporal Alignment and Feature Fusion Submodule; 203. Extended Kalman Filter State Estimation Submodule; 204. Center of Gravity Offset Calculation Submodule; 300. Anti-interference Adaptive Control Module; 301. Reference Trajectory Generation Submodule; 302. Tracking Error and Sliding Mode Surface Construction Submodule; 303. Lyapunov function construction and adaptive sliding mode control law calculation submodule; 304. Model predictive control trajectory optimization submodule; 305. Control command decomposition and allocation submodule; 400. Adaptive dual-mode drive execution module; 401. Magnetic drive unit submodule; 402. Ratchet drive unit submodule; 403. Differential drive wheel assembly submodule; 404. Swing arm and joint servo submodule; 500. Energy management and scheduling module; 501. Hybrid power supply module; 502. Fuzzy logic-based energy scheduling submodule; 600. Host computer communication and monitoring module; 601. Wireless communication submodule; 602. Remote monitoring and emergency intervention submodule. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the following embodiments are only for explaining the invention and are not intended to limit the scope of protection of the invention.

[0026] like Figure 1As shown, this invention provides an anti-interference adaptive control system for a tower inspection robot in dynamic environments. The system adopts a four-level closed-loop control architecture, consisting of a perception layer, a decision-making layer, an execution layer, and an energy management layer. Each layer interacts bidirectionally via a 1Mbps CAN bus. The core control computation runs on an embedded main controller based on the ARM Cortex-A72 architecture. This controller has a clock speed of 1.5GHz, 4GB of DDR4 memory, a fixed control cycle of 5ms, and a control frequency of 200Hz, thus meeting the stringent real-time response requirements of the robot in high-altitude, strong wind environments. The system also includes an edge AI computing module equipped with an embedded GPU. This module, based on the NVIDIA Jetson Orin Nano platform, has 40 TOPS of INT8 inference computing power and 1024 CUDA cores, dedicated to performing deep learning inference tasks for visual and infrared images. The main controller and the edge AI computing module achieve high-speed data interaction through a USB 3.0 interface. The main controller focuses on 200Hz high-frequency control loop calculation and sensor data fusion, while the edge AI computing module focuses on deep neural network inference calculation. The two complement each other in terms of function and computing power.

[0027] The entire system is functionally divided into a multiphysics coupled dynamics solution module 100, a multimodal sensing and state estimation module 200, an anti-interference adaptive control module 300, an adaptive dual-mode drive execution module 400, an energy management and scheduling module 500, and a host computer communication and monitoring module 600. Logically, each module sequentially undertakes the functions of mathematical description of the controlled object, acquisition of environmental information and system state estimation, control decision-making and command calculation, execution of physical actions, energy supply guarantee, and remote monitoring and safety management, forming a complete technical chain from modeling to sensing, from control to execution, and from energy to monitoring. The implementation details of each module are explained below.

[0028] The multiphysics coupled dynamics solution module 100 constitutes the mathematical foundation layer of the entire control system. All parameters and calculation programs of this module are pre-loaded in the flash memory of the main controller. Parameter loading is completed during the system power-on initialization phase, and during the operation phase, it continuously provides real-time dynamics calculation services to the multimodal perception and state estimation module 200 and the anti-interference adaptive control module 300.

[0029] The multiphysics coupled dynamics solution module 100 includes a robot kinematics parameter library submodule 101. This submodule stores all structural parameters of the robot body. Regarding the lengths of each link, the main frame length... Take 280mm, swing arm length Take 150mm, articulated arm length Take 120mm. Regarding the mass of each connecting rod, the main body mass... The weight is 2.8 kg, and the swing arm mass is... The weight of the articulated arm is 0.6 kg. The weight of the drive wheel assembly is 0.4 kg. The weight of the sensor and circuit board is 0.7 kg. The weight is 0.5kg, and the total weight of the whole machine is... It weighs 5.0 kg. The moment of inertia of each link is accurately calculated based on the CAD model and stored in numerical form. The constraint of the limit angle of each joint is: the rotation range of the swing arm. to Range of motion of the articulated arm to Among the drive wheel parameters, wheel diameter 60mm, wheel width The value is 25mm; the coefficient of friction corresponding to steel guide rails. A value of 0.35 corresponds to the coefficient of friction when using aluminum guide rails. Take 0.28.

[0030] The multiphysics coupled dynamics solution module 100 also includes a guide rail physical property parameter library submodule 102, which stores the physical parameters of various types of guide rails involved in actual engineering. The steel guide rails are made of Q235 steel with a modulus of elasticity of [missing information]. The value is 200 GPa, and the corresponding moment of inertia is stored. Linear density and structural damping coefficient The aluminum guide rail is made of 6061-T6 aluminum alloy, with a modulus of elasticity of [missing information]. The corresponding moment of inertia is 69 GPa. Linear density and structural damping coefficient It is also stored as a parameter. The geometric parameters of the guide rail are uniformly defined as: cross-sectional width. It is 40mm high. It is 30mm thick. The spacing between installation supports is 3mm. The guide rail has a surface roughness range of Ra1.6 to Ra6.3 and a radius of curvature range of R300mm to R500mm. These parameters provide an indispensable physical basis for the subsequent establishment of dynamic equations and the calculation of guide rail flexible deformation.

[0031] The core component of the multiphysics coupled dynamics solution module 100 is the Lagrange dynamics equation construction submodule 103. This submodule abstracts the robot-guide rail system into a five-degree-of-freedom multibody dynamics system for modeling. In this embodiment, the five generalized coordinates selected are: displacement along the guide rail direction... Displacement perpendicular to the guide rail direction Roll angle about the guide rail axis Arm joint angle and joint arm angle The generalized coordinate vector of the system is denoted as .

[0032] kinetic energy of the system It includes the translational kinetic energy and rotational kinetic energy of each link, and its expression is:

[0033]

[0034] in For the first The position vector of the center of mass of each link in the inertial coordinate system is calculated recursively by the forward kinematics method based on the DH parameters; For the first The angular velocity vectors of the links; For the first The inertia tensor matrix of each link.

[0035] The potential energy of the system It consists of three components:

[0036]

[0037] The first component is gravitational potential energy:

[0038]

[0039] In the formula For the first The height coordinates of the center of mass of the connecting rod. The second component is the elastic deformation energy of the guide rail.

[0040]

[0041] in The equivalent stiffness of the guide rail is calculated using the Euler-Bernoulli beam theory from the parameters in submodule 102 of the guide rail physical property parameter library:

[0042]

[0043] The first component represents the flexible deformation of the guide rail at the robot's contact point. The introduction of this potential energy term allows the model to capture the influence of the elastic displacement of the guide rail caused by robot loads and external excitations on the system's dynamic behavior, which is key to its difference from the traditional rigid guide rail assumption. The third component is the magnetic attraction potential energy. The derivation of the magnetic attraction potential energy is based on the equivalent magnetic circuit model: the actual magnetic flux density generated by the permanent magnet in the air gap. It is not equal to the remanence of a permanent magnet. It is not determined by the air gap magnetic reluctance, but by the combined influence of the magnetic reluctance of the other parts of the magnetic circuit. According to Kirchhoff's law for magnetic circuits, the air gap magnetic flux density can be expressed as:

[0044]

[0045] In the formula The remanence of the neodymium iron boron permanent magnet is 1.2T. The thickness of the permanent magnet along the magnetization direction. This represents the relative restoring permeability of the permanent magnet (typically around 1.05). Let be the relative permeability of the guide rail steel material. This expression clearly shows the air gap magnetic flux density. With air gap distance The magnetic flux density increases rather than decreases linearly; this decrease is entirely determined by the physical laws of the magnetic circuit. Based on the air gap magnetic flux density, the magnetic attraction force on the pole face is given by Maxwell's formula for attraction:

[0046]

[0047] In the formula The area of ​​the magnetic poles. Let be the permeability of free space. The magnetic attraction potential energy is defined as the integral of the magnetic attraction force along the air gap direction:

[0048]

[0049] The upper limit of integration is taken as the zero point of potential energy at infinity. The above modeling accurately describes the nonlinear effect of air gap change on magnetic flux density through magnetic circuit equations, rather than simplifying it to an undefined decay function. This provides a controlled object model with a clear physical basis for the closed-loop air gap control of magnetic attraction force in the subsequent adaptive dual-mode drive execution module.

[0050] The system's Lagrangian is defined as:

[0051]

[0052] Considering guide rail damping and joint friction, the Rayleigh dissipation function is defined as follows:

[0053]

[0054] In the formula This is the equivalent viscous damping coefficient of the guide rail, which physically represents the ability of the guide rail to dissipate energy during bending deformation due to internal material damping and friction of the supporting structure. This coefficient is selected from the parameter library of the guide rail physical property parameter library submodule 102 based on the guide rail material: for steel guide rails... Aluminum guide rails The values ​​were obtained through free vibration attenuation experiments on guide rail samples. Typical values ​​for steel guide rails are... Typical values ​​for aluminum guide rails . ( The coefficient of viscous friction between the swing arm joint and the articulated arm joint is the speed-related damping torque generated during joint rotation due to bearing friction, seal resistance, and internal losses in the reducer. In this embodiment... , The values ​​were obtained by measuring and calibrating the driving torque when the joint was rotating at a low speed under no load.

[0055] Generalized force vector acting on the system It is composed of four superimposed components:

[0056]

[0057] in This is the driving torque vector generated by each drive motor. The wind load is a generalized force, calculated from the measured values ​​of the wind speed sensor using aerodynamic formulas.

[0058]

[0059] air density in the formula Take the standard value, drag coefficient Take 1.2 as the robot's frontal area. for . This refers to the inertial force caused by the vibration of the tower foundation. The excitation of the tower foundation, expressed in terms of displacement, is as follows:

[0060]

[0061] In the formula The lateral vibration displacement of the tower body at the robot mounting base, amplitude The range is 5 to 20 mm, frequency The frequency range is 0.1 to 2 Hz, covering the typical galloping frequency band of the tower under wind load. The inertial force generated by the tower vibration on the robot's links is obtained by taking the second time derivative of the foundation displacement, i.e., the foundation acceleration. The inertial forces acting on each link are This is thus mapped to the generalized coordinate space to form a generalized force. . The force is the wheel-rail contact force, including the normal force. and tangential friction Two components describe the mechanical coupling relationship between the drive wheel and the guide rail.

[0062] According to the second kind of Lagrange equation:

[0063]

[0064] After differentiating and rearranging with respect to each generalized coordinate, we obtain the standard form of the system's dynamic equations:

[0065]

[0066] In the formula for Symmetric positive definite inertia matrix, The matrix of Coriolis force and centrifugal force. The vector of the gravity term. Let the damping and friction force vector be... For the driving torque input vector, This represents the combined external disturbance vector, incorporating three physical field effects: wind load, tower vibration, and guide rail displacement fluctuation. The dynamic equation comprehensively describes, from an energy perspective, how the flexible deformation of the guide rail affects the force distribution on the drive wheel and the overall center of gravity position through the contact force transmission chain, providing an accurate mathematical description of the controlled object for subsequent state estimation and controller design. It is worth noting that the time derivative of matrix M... Coriolis force matrix They satisfy the antisymmetric property, that is, for any vector have This property will be utilized in subsequent Lyapunov stability analysis.

[0067] The multiphysics coupled dynamics solution module 100 also includes a critical instability frequency identification submodule 104, which obtains the system's natural frequency information by linearizing the aforementioned dynamic equations near the equilibrium point. Let Taking a first-order Taylor expansion of the dynamic equations, we obtain the linearized vibration equations:

[0068]

[0069] Solve its characteristic equation The natural frequencies of the system at each order can be obtained. In this embodiment, the calculated first-order natural frequency is approximately 5.2 Hz. When the external excitation frequency approaches this value, the system is at risk of resonant instability. Therefore, the critical instability frequency is set to 5 Hz as an upper limit constraint for the bandwidth design of the control system.

[0070] The multimodal sensing and state estimation module 200 is responsible for environmental information acquisition, multi-sensor data fusion, and optimal estimation of the system state. The output of this module is high-precision system state information after noise filtering and interference decoupling, which serves as the feedback reference for the anti-interference adaptive control module 300 to implement closed-loop control.

[0071] The multimodal perception and state estimation module 200 includes a sensor array submodule 201, which carries various types of sensors to cover the measurement needs of different physical quantities. A six-axis inertial measurement unit is mounted at the robot's center of mass, with a sampling rate of 200Hz and an angular velocity range of... Accelerate measurement range Built-in digital motion processor, outputting three-axis acceleration in real time. and triaxial angular velocity The binocular industrial vision camera is mounted on the top of the robot in a forward-facing position, with a resolution of [missing information]. With a frame rate of 30fps, employing a global shutter to eliminate motion blur, and an IP67 protection rating, it is used for guide rail geometry feature recognition and obstacle detection. The infrared thermal imaging module is coaxially mounted with the binocular camera, achieving a resolution of [missing information]. The noise equivalent temperature difference is no greater than 50mK, and the frame rate is 25fps, used for detecting abnormal temperature areas and visual compensation in low-light environments. Four ultrasonic ranging sensor probes are distributed at the four corners of the robot body, operating at a frequency of 40kHz, with a range of 20mm to 4000mm, and a measurement accuracy of [missing information]. A 100Hz sampling rate (mm) is used to accurately measure the distance between the robot and the guide rail and tower. A three-dimensional force / torque sensor is installed at the connection flange between the drive wheel assembly and the main frame, with a three-axis force range of 0 to 500N and a three-axis torque range of 0 to 20Nm, and a sampling rate of 100Hz, to measure the distribution of wheel-rail contact force in real time. A miniature wind speed and direction sensor is installed at the highest point on top of the robot, with a range of 0 to 60m / s and an accuracy of [missing information]. With a sampling rate of 10Hz and a speed of m / s, the system acquires real-time information on wind speed and direction in the current environment.

[0072] The multimodal data spatiotemporal alignment and feature fusion submodule 202 processes multi-source heterogeneous sensor data from the sensor array submodule 201. In the time dimension, all sensors achieve unified time synchronization via GPS second pulses or hardware timer interrupts from the main controller, with a time synchronization accuracy better than 0.1ms. For sensors with different sampling rates, linear interpolation is used to unify all data to the system's 5ms control cycle time reference, thereby eliminating data misalignment caused by inconsistent sampling times. In the spatial dimension, the robot's body coordinate system is used as a unified reference system. This coordinate system has its origin at the center of mass, with the X-axis along the guide rail direction, the Y-axis perpendicular to the guide rail surface, and the Z-axis determined according to the right-hand rule. The mounting extrinsic parameter matrices (including rotation matrices) of each sensor relative to the body coordinate system are also specified. Translation vector The parameters are determined and stored in the parameter library by hand-eye calibration at the factory. The original measurement values ​​in the sensor coordinate system are transformed to the body coordinate system by a homogeneous transformation matrix.

[0073] Based on the completed spatiotemporal alignment, the multimodal data spatiotemporal alignment and feature fusion submodule 202 performs deep feature fusion on visual and infrared image data. This fusion computation is executed on the edge AI computing module, utilizing the parallel computing capabilities of its embedded GPU to complete real-time inference of the deep neural network. The fusion is implemented using a multimodal fusion network based on the Transformer architecture. The network receives binocular visual images. and infrared thermal imaging As input, feature maps are extracted by their respective convolutional encoders. The encoder backbone network adopts a ResNet-18 structure and loads pre-trained weights. The output feature map dimension is [dimensionality missing]. and To accommodate the computational constraints of embedded platforms, the feature maps of both modalities are spatially downsampled to [a specific value] before entering the Transformer. The resolution is then flattened into a sequence for concatenation, and the length of the concatenated sequence is [length missing]. The sequence is fed into a Transformer encoder, which includes a 4-layer encoder, an 8-head attention mechanism, and a 512-dimensional hidden layer. The Transformer's self-attention mechanism automatically learns the correlation weights between cross-modal features, enabling the network to automatically increase the contribution weight of infrared features in low-light scenes, while focusing on utilizing detailed information from visual texture features under normal lighting conditions. The entire fusion network is deployed on the edge AI computing module after TensorRT quantization optimization, with an INT8 quantization inference latency of less than 30ms, meeting the real-time processing requirements at 30fps. The output of the Transformer encoder is processed by the decoder and divided into two paths: one path connects to the detection head to detect surface defects on the guide rail, including typical defect types such as rust, cracks, and loose bolts, with a detection accuracy of up to 0.1mm and an average accuracy of no less than 98%; the other path outputs geometric feature parameters of the guide rail, such as the radius of curvature, the position and height of welding bosses, and the position of branch nodes. The inference results are transmitted back to the main controller via a USB 3.0 interface for use by subsequent control modules. To enhance the model's generalization ability under all weather conditions, generative adversarial networks were used during the training phase to augment the data of samples in adverse scenarios such as low light, rain, and fog.

[0074] The core computing unit of the multimodal sensing and state estimation module 200 is the extended Kalman filter state estimation submodule 203. This submodule runs on the main controller, and its design concept is to incorporate wind load disturbance and guide rail displacement fluctuation as extended state variables into the state vector for joint estimation. This achieves accurate identification and decoupling of environmental disturbances while completing the optimal fusion of multi-sensor data.

[0075] The extended state vector of the system defined by the extended Kalman filter state estimation submodule 203 is a twelve-dimensional vector:

[0076] The first five components are generalized coordinates, the middle five components are the corresponding generalized velocities, and the last two components are estimates of the wind-borne disturbance forces. Estimated value of guide rail displacement fluctuation Incorporating disturbances into the state vector is a key technique for achieving disturbance-noise decoupling: disturbances are no longer considered as unobservable external inputs, but rather participate in the filtering estimation process as part of the system state, enabling EKF to simultaneously output the clean system motion state and the separated disturbance estimate after each recursive update.

[0077] The state transition equations of EKF are based on the dynamic equations of submodule 103 of the Lagrange dynamic equations. The continuous-time dynamic equations are then processed using the fourth-order Runge-Kutta method. Discretizing the step size ms, we get:

[0078]

[0079] Nonlinear function From the dynamic equation The discretized form is given. For the state transition of the perturbation, a random walk model is used to describe it:

[0080]

[0081]

[0082] This choice reflects the prior assumption that the disturbance changes slowly between adjacent sampling times. (Process noise vector) Assume it follows a zero-mean Gaussian distribution Its covariance matrix The setting can be adjusted online according to system operating conditions—increase the setting when the wind speed sensor indicates high wind speed. The component increases when the vision module detects that the guide rail has entered a curved section. The component is used to reflect the difference in the severity of disturbance changes under different operating conditions.

[0083] The EKF observation equation maps the state vector to the space observable by the sensor. Observation vector Composed of multi-sensor measurements after spatiotemporal alignment:

[0084]

[0085] in:

[0086] These correspond to six-dimensional data (three-axis acceleration and angular velocity) output from the inertial measurement unit, four-dimensional data (four-channel ultrasonic ranging values), and two-dimensional data (two sets of normal contact forces of the drive wheels). Nonlinear observation function. The state vector is mapped to the observation space based on the physical installation location and kinematic relationships of each sensor. Observation noise vector. Similarly, it is assumed that the distribution follows a zero-mean Gaussian distribution. Its covariance matrix The noise characteristics are set according to the factory calibration of each sensor.

[0087] The EKF recursive algorithm executes a complete prediction-update cycle within each 5ms control cycle. In the prediction step, the state is predicted in one step using the dynamic model and the current control input to obtain the predicted state. ; Calculate the nonlinear state transition function Jacobian matrix at the current estimate The matrix is The square matrix was calculated using the numerical difference method; utilizing Process noise covariance Update the prediction error covariance matrix:

[0088]

[0089] In the update step, the nonlinear observation function is calculated. Jacobian matrix at the predicted state ; Calculate the new information covariance matrix:

[0090] Calculate the Kalman gain:

[0091]

[0092] The predicted state is corrected using innovation and Kalman gain:

[0093]

[0094] The error covariance matrix is ​​updated using the Joseph stable form to ensure numerical stability.

[0095]

[0096] The output of EKF is divided into two sets of information: one set is the pure generalized coordinate estimate. and generalized velocity estimates One set consists of the first ten components of the extended state vector, which are used by the controller as a feedback reference; the other set consists of wind load disturbance estimates. and guide rail displacement fluctuation estimate These are the last two components of the extended state vector, which are used by the controller as a basis for feedforward compensation. Through this design, the problem of partial overlap between environmental interference signals and sensor noise in the frequency domain is fundamentally solved. EKF does not rely on simple frequency domain filtering to separate the two, but uses the physical laws contained in the system dynamics model to distinguish between disturbance response and measurement noise, thus achieving precise decoupling in a physical sense.

[0097] The multimodal sensing and state estimation module 200 also includes a center of gravity offset calculation submodule 204. This submodule calculates the position of each link's center of mass in the body coordinate system using forward kinematics based on the pure generalized coordinate estimates output by the EKF, and then calculates the position vector of the entire machine's center of gravity according to the mass-weighted average formula.

[0098]

[0099] Center of gravity offset Defined as the Euclidean distance between the center of gravity and the centerline of the guide rail in a plane perpendicular to the direction of the guide rail:

[0100]

[0101] in For the first The position of the center of mass of each link in the body coordinate system The coordinates of the guide rail centerline in the body coordinate system are determined by the ultrasonic ranging value. This represents the Euclidean distance in the YZ plane. This serves as the input to the augmented reference trajectory optimization stage in the anti-interference adaptive control module 300. The system... A three-tiered progressive safety threshold system was established: Level 1 warning threshold. mm, secondary critical threshold mm and Level 3 absolute safety threshold mm. The first-level early warning threshold triggers enhanced compensation and remote alarm push; the second-level critical threshold triggers system deceleration; and the third-level absolute safety threshold triggers hardware rail-holding braking. The severity of these three thresholds increases sequentially, corresponding one-to-one with the graded safety protection mechanisms in subsequent steps S4 and S8.

[0102] The anti-interference adaptive control module 300 is the core of the control decision-making of the entire system. It undertakes the key function of calculating the control commands of each actuator in real time based on the state estimation results and ensuring that the robot maintains robust and stable operation under dynamic disturbances.

[0103] The anti-interference adaptive control module 300 includes a reference trajectory generation submodule 301, which generates the desired five-dimensional generalized coordinate trajectory based on the inspection task planning and guide rail geometry information. and its first and second time derivatives and In the linear guide section, the desired trajectory is set to uniform motion, with a climbing speed of 25 m / min, or 0.417 m / s. Lateral displacement and roll angle All are set to zero. and The default clamping angle is used. In curved and obstacle-crossing sections, the reference trajectory is generated online by the model prediction control trajectory optimization submodule 304 and then downloaded to the reference trajectory generation submodule 301.

[0104] The tracking error and sliding surface construction submodule 302 defines the error signal and sliding surface required by the controller. The tracking error and sliding surface construction submodule 302 first generates the basic reference trajectory output by submodule 301 based on the reference trajectory. The augmented reference trajectory is generated by combining the center of gravity offset information output by the center of gravity offset calculation submodule 204. Specifically, based on the basic reference trajectory, the arm joint angle is adjusted. and joint arm angle A real-time correction is applied so that, on the corrected reference trajectory, the overall center of gravity of the machine, calculated by forward kinematics, lies precisely on the center line of the guide rail, and the normal forces of each drive wheel are uniformly distributed. The correction amount is obtained by solving the following nonlinear optimization problem in each control cycle:

[0105]

[0106] The constraint is the joint limit. , .in This is the center of gravity offset calculated using the positive kinematics and mass-weighted average formula at a given joint angle. For the corresponding number Normal force of each drive wheel To ensure a uniform distribution of the target normal force, the number of drive wheels in this embodiment... Weighting coefficient Take 1000, Take 500. The objective function... and All calculations are performed using forward kinematic equations and static equilibrium equations that incorporate trigonometric functions; therefore, the objective function is about... and The nonlinear function. However, since this optimization problem contains only two decision variables and the constraint space is a finite closed interval, the actual solution uses the Gauss-Newton method combined with the analytical Jacobian matrix, and uses the chain rule of forward kinematics to analytically calculate the objective function. and The gradient and approximate Hessian matrix are used to avoid the additional overhead of numerical differencing. The optimal solution from the previous control cycle is used as the initial value for hot start, ensuring continuous joint angle changes with minimal variation per cycle (joint angle changes do not exceed a certain value within 5ms). Under the condition that the Gauss-Newton method typically converges to the condition within 1 to 2 iterations. The required accuracy means that a single iteration involves only two forward kinematics calculations and one... Solving the system of linear equations takes less than [time missing] in total computation time. The augmented reference trajectory is denoted as... Its first and second time derivatives and Calculated using the finite difference method.

[0107] Based on the augmented reference trajectory, define the augmented tracking error vector:

[0108]

[0109]

[0110] The augmented sliding mode surface vector is defined as:

[0111]

[0112] in It is a positive definite diagonal matrix, each The rate of error convergence on the sliding mode is determined. In this embodiment, the three parameters correspond to the displacement degrees of freedom. This corresponds to a convergence time constant of approximately 50ms to ensure a fast response in displacement tracking; the two parameters corresponding to the angular degrees of freedom... .

[0113] The Lyapunov function construction and adaptive sliding mode control law calculation submodule 303 is the core algorithm unit of the anti-disturbance adaptive control module 300, and its design process follows the strict Lyapunov stability theory.

[0114] Lyapunov function construction and adaptive sliding mode control law calculation submodule 303 constructs the following Lyapunov candidate function. :

[0115]

[0116] First item To broaden the measurement of the "kinetic energy" of tracking error on the sliding surface, The inertia matrix is ​​positive definite, ensuring this term is positive. The second term... Here is the adaptive gain estimation error term, where The upper bound estimate of the disturbance With the true upper realm difference, The learning rate for the adaptive law is set to 50. This function is for... or All states of the function are strictly positive, satisfying the positive definiteness condition of the Lyapunov candidate function.

[0117] By designing an augmented reference trajectory, the objectives of center of gravity offset compensation and drive wheel anti-slip control are implicitly encoded into the tracking error. Chinese: When When sufficiently small, the system's generalized coordinates are close to ,and The fourth and fifth components are precisely the optimal joint angles that allow the center of gravity to be located on the guide rail centerline and the normal force to be evenly distributed. This design concept of "transforming multi-objective optimization into a single tracking control problem through augmented reference trajectory" is one of the core innovations that distinguishes this invention from existing technologies. In traditional control schemes, center of gravity offset compensation and drive wheel anti-slip control are often handled by their respective independent sub-controllers. There is a lack of mathematical coordination mechanism between their optimization objectives and control commands, and under extreme conditions, control command conflicts may even lead to system divergence. This invention encodes the center of gravity offset and normal force balance objectives into the augmented reference trajectory, transforming the multi-objective control problem into a single trajectory tracking problem. This ensures that the control law derived based on Lyapunov theory mathematically guarantees the cooperative convergence of all sub-objectives, fundamentally eliminating the possibility of control command conflicts. At the same time, this design avoids the relative order mismatch problem caused by directly introducing position-related penalty terms into the Lyapunov function, thus ensuring the mathematical rigor of the subsequent control law derivation.

[0118] right Find the time derivative:

[0119]

[0120] Substituting the dynamic equations and the definition of the augmented sliding surface into... ,Will Expand to control input Related forms. Utilizing The antisymmetric property, namely Simplifying, we get:

[0121]

[0122] based on The unfolding form, design of the total control torque It is composed of the superposition of two components:

[0123]

[0124] Equivalent control torque The modeled dynamics of the compensation system and the disturbance feedforward estimated by EKF:

[0125]

[0126] In the formula The disturbance estimation vector output by the extended Kalman filter state estimation submodule 203 is used as a feedforward compensation term in the control law calculation. This means that the controller does not have to rely entirely on the feedback mechanism to suppress disturbances, which significantly improves the speed and accuracy of compensation for identified disturbances.

[0127] Robust switching torque Used to suppress uncertainties such as perturbation estimation errors and unmodeled dynamics:

[0128]

[0129] in This is an adaptive estimate of the upper bound of the perturbation. This represents the boundary layer thickness parameter. Here, a continuous approximation function is used instead of the sign function in traditional sliding mode control. This avoids the high-frequency chatter problem caused by hard switching. When Much larger At that time, this term is approximately It maintained strong robustness; when As it approaches zero, the term naturally approaches zero, thus avoiding discontinuity at the zero point.

[0130] Substituting the two control torques mentioned above into... The expression yields:

[0131]

[0132] In order to make The terms containing adaptive gain error are also eliminated or transformed into non-positive terms, and the adaptive law is designed as follows:

[0133]

[0134] in for - The correction term coefficient is set to 0.1 to prevent drift in the parameter estimates. The intuitive meaning of this adaptive law is: when external disturbances increase, leading to a decrease in the sliding mode surface vector norm... When the value increases, the upper bound estimate of the disturbance is... This is increased accordingly to enhance robust compensation capability; when the disturbance decreases, When approaching zero, pass - The correction gradually decays to avoid over-control and energy waste. During operation... Set lower bound protection To prevent it from becoming negative.

[0135] when It can be proven that:

[0136]

[0137] The above inequality shows It is monotonically non-increasing, which can be proven from this. and All signals are bounded. It should be noted that boundary layer parameters are introduced into the robust switching torque. To eliminate high-frequency flutter, an adaptive law is introduced. - Correction terms prevent parameter drift. These two engineering treatments ensure that the system state does not strictly converge to the equilibrium point, but rather converges to a bounded neighborhood of the equilibrium point, meaning the system satisfies the eventually uniformly bounded (UUB) property. Specifically, the augmented sliding mode surface vector... The final bounded range is determined by and Joint decision: Ultimately constrained by Within the same order of magnitude neighborhood. By the definition of an augmented sliding surface. ,when When bounded, tracking error Ultimately, they are also bound by... Within a bounded neighborhood of the same magnitude, where .

[0138] In this embodiment, The corresponding steady-state tracking error band width is in Within the mm range, the steady-state residual of the center of gravity offset is within All within mm meet the requirements of engineering applications. and The value of represents a design trade-off between system robustness (flutter suppression, parameter drift protection) and tracking accuracy: reducing It can reduce the steady-state error band but will exacerbate the high-frequency chatter of the actuator and increase its frequency. This can more effectively prevent parameter drift but will increase the upper bound of the steady-state error. The values ​​used in this embodiment... and Through extensive simulations and experiments, the method was optimized to eliminate flutter while keeping the steady-state error within an acceptable range for engineering applications.

[0139] Lyapunov function construction and adaptive sliding mode control law calculation submodule 303, in actual operation, affects the total control torque. Each component of the function is saturated and limited to ensure it does not exceed the physical limits of the corresponding actuator. At the end of each control cycle, the current Lyapunov function value is calculated. and its derivative It is then output to the host computer communication module for remote stability monitoring. If If the duration exceeds 10 control cycles (50ms), the system will trigger a safety alarm and... The value is increased to 1.5 times the current value to enhance robustness, and the event is pushed to the ground operator's terminal via the host computer communication module. This triggering condition is consistent with the first-level criterion in the graded safety protection mechanism in step S8.

[0140] The anti-interference adaptive control module 300 also includes a model predictive control trajectory optimization submodule 304, which operates at a frequency of 20Hz in the outer loop of the adaptive sliding mode controller and executes once every 50ms, for future... The movement state of each step is predicted using a rolling prediction method. In this embodiment, the number of steps is predicted. If we set it to 3, the prediction time domain is 150ms.

[0141] The optimization problem solved by MPC is defined as follows:

[0142]

[0143] Constraints include joint limit constraints:

[0144]

[0145] Torque saturation constraint:

[0146]

[0147] Dynamic constraints:

[0148]

[0149] Obstacle safety distance constraints:

[0150]

[0151] The state tracking weight matrix Lateral displacement and roll angle degrees of freedom are given high weights to highlight the safety-first design philosophy; control energy consumption weight matrix Energy consumption weighting ; This represents the predicted energy consumption. To predict the minimum distance from each point on the trajectory to the detected obstacles; mm is a safety margin.

[0152] This optimization problem is solved online using a sequential quadratic programming algorithm. The computation time for each iteration is controlled within 2ms, and 5 to 10 iterations can be completed within a 50ms update cycle to ensure convergence. The output of MPC is the optimized future... Step reference trajectory The trajectory is passed to the reference trajectory generation submodule 301 and replaces the original simple reference trajectory, providing a smoother and more energy-efficient tracking target for the inner-loop adaptive sliding mode controller. Through a dual-loop collaborative architecture where the outer-loop MPC is responsible for globally optimal trajectory planning and the inner-loop SMC is responsible for locally robust tracking, the system maintains robust stability under extreme conditions while improving trajectory smoothness and energy utilization efficiency. For typical obstacles such as welding bosses up to 15mm high, the MPC pre-plans a smooth obstacle-crossing trajectory, starting to rise 30mm from the obstacle and resuming normal travel 30mm after crossing it, achieving an obstacle-crossing success rate of over 99%.

[0153] The control command decomposition and allocation submodule 305 outputs the five-dimensional generalized torque vector from the Lyapunov function construction and adaptive sliding mode control law calculation submodule 303. This is broken down into specific control commands for each actuator. The driving torque along the guide rail direction... The speed commands were converted into the left and right drive wheel speeds. and The base rotation speed is determined by the desired climbing speed, according to The differential component is calculated to be determined by the lateral deviation compensation. (Swing arm joint torque) and joint arm torque These are converted into angle adjustment commands for the corresponding servo motors. Regarding drive mode selection, the control command decomposition and allocation submodule 305 automatically switches based on the guide rail material identification result: when the inductive sensor determines the guide rail is made of steel, the magnetic drive module dominates, bearing no less than 70% of the normal force, and the control objective is to maintain the air gap distance. Constant to achieve The system maintains stable adsorption force. When the rail is identified as aluminum, the ratchet drive module takes the lead, aiming to maintain a pawl engagement depth of no less than 1.5mm to provide a clamping force of no less than 200N. In the steel-aluminum mixed transition section, the vision module anticipates material changes 200mm in advance, and the system adopts a 50ms gradual switching strategy. The linear decrease in magnetic attraction force and the linear increase in ratchet clamping force are synchronized to avoid sudden changes in normal force.

[0154] The adaptive dual-mode drive execution module 400 is responsible for converting the digital control commands of the anti-interference adaptive control module 300 into physical execution actions.

[0155] The magnetic drive unit submodule 401 is suitable for use with steel guide rails. This unit uses a neodymium iron boron (N52) permanent magnet array to... Arranged in a specific pattern, the total magnetic pole area is The permanent magnet array adopts a Halbach arrangement, meaning that the magnetization directions of adjacent permanent magnets rotate alternately. This arrangement increases the magnetic flux density on the working surface by approximately 20% compared to traditional uniform magnetization arrangements, with a measured magnetic flux density of 0.6T. The static adsorption force on one side under a 1mm air gap can reach 350N. Precise adjustment of the air gap is achieved via a miniature electric actuator with a stroke of 0 to 5mm and a positioning accuracy of [missing information]. mm, response time less than 50ms. The closed-loop control of the magnetic attraction force adopts a PID strategy, using the actual attraction force measured by the force sensor as the feedback quantity, with 300N as the target value. Internally, the controller uses the Lagrange dynamics equation to construct the magnetic circuit model established in submodule 103 to calculate the target air gap value corresponding to the target attraction force. Then, the air gap PID loop drives the electric actuator to perform adjustment, with a control bandwidth of 50Hz and a steady-state accuracy of [missing information]. Right now N.

[0156] The ratchet drive unit submodule 402 is suitable for use with guide rails made of non-ferromagnetic materials such as aluminum. The core component of this unit is a flexible pawl mechanism. The pawl is made of 17-4PH stainless steel with a hardness of HRC42 to ensure wear resistance. The pawl tooth pitch is 2mm and the tooth depth is 2mm. Spring preload causes the pawl teeth to engage with the anti-slip rack on the side of the guide rail, achieving forward drive and reverse self-locking functions. The clamping force is adjusted via a cam-spring combination structure. A servo motor drives the cam to rotate, changing the spring compression, thus allowing the clamping force to be continuously adjustable within the range of 150N to 300N with high adjustment precision. N.

[0157] The differential drive wheel assembly submodule 403 comprises two independent drive wheels, left and right. The wheel body is made of polyurethane-coated rubber with a Shore A hardness of 70 and a diameter of 60mm, balancing frictional adhesion with the guide rail surface and wear resistance. Each drive wheel is independently driven by a brushless DC motor with a rated power of 30W, a maximum speed of 500rpm, and a built-in encoder resolution of 2000 lines / revolution. Straight-line and siding correction functions are achieved by controlling the speed difference between the left and right wheels: during straight-line movement, both wheels rotate at the same speed. Speed ​​difference during correction ,in Lateral displacement estimated by EKF Its reference value The difference, i.e., the lateral displacement tracking error, is the proportional coefficient. The speed control is dynamically determined by the anti-interference adaptive control module 300 based on the current operating conditions. The speed control accuracy reaches [a certain level]. The RPM is ensured by the motor's built-in encoder in conjunction with a PID speed loop.

[0158] The swing arm and joint servo submodule 404 comprises two parts: a swing arm servo motor and a joint arm servo motor. The swing arm servo motor has a power of 15W, a reduction ratio of 1:100, and is equipped with a 14-bit absolute encoder at the output, achieving a resolution of [resolution missing]. The articulated arm's servo motor has a power of 10W, a reduction ratio of 1:80, and is equipped with an encoder of the same specification. Both servo motors achieve... The position control accuracy is guaranteed by a multi-turn absolute encoder combined with a PID position loop, with a control bandwidth of 100Hz. The angle adjustment of the swing arm and articulated arm is the main physical execution means of center of gravity offset compensation. When the anti-interference adaptive control module 300 calculates the need to adjust the center of gravity position through augmented reference trajectory optimization, the adjustment command is ultimately reflected in the angle increment of these two servo motors.

[0159] The Energy Management and Scheduling Module 500 provides a reliable and efficient power supply for the entire system and dynamically schedules energy allocation strategies according to operating conditions.

[0160] The energy management and dispatch module 500 includes a hybrid power supply module 501, which employs a hybrid energy storage architecture that complements lithium batteries and supercapacitors. The lithium battery pack uses 18650 cells configured in a 3-series, 2-parallel configuration, with a nominal voltage of 11.1V, a capacity of 5200mAh, an energy density of approximately 57.7Wh, and a weight of approximately 300g. It serves as the primary energy source, providing continuous and stable power output with a rated discharge power of 50W. The supercapacitor pack uses a modular series configuration with a 12V / 100F specification. Although its energy density is only about 2Wh, its power density exceeds 5kW / kg, and it weighs approximately 200g. It serves as an auxiliary energy source, providing instantaneous high-power pulses for obstacle crossing and wind load resistance, with a peak discharge power of up to 200W and a duration of 2 to 5 seconds. The lithium battery and the supercapacitor are connected by a bidirectional DC-DC converter with an efficiency of not less than 95% and a switching frequency of 200kHz, enabling bidirectional energy flow. Under low load conditions, the lithium battery charges the supercapacitor in trickle mode via DC-DC, while under high power demand conditions, the supercapacitor releases a large current pulse to the load via DC-DC.

[0161] The fuzzy logic-based energy scheduling submodule 502 uses a fuzzy inference engine to achieve intelligent energy allocation. This engine receives three input variables: the robot's current load rate. (Range from 0% to 100%, language variables are divided into three levels: "low", "medium", and "high"), calculated by dividing the cumulative real-time power of each motor by the total rated power of the system; lithium battery state of charge. (Range 20% to 100%, linguistic variables are divided into three levels: "low", "medium", and "high"), estimated online by the coulomb counting method combined with the open-circuit voltage method of the battery management system; supercapacitor remaining energy ratio (Range from 0% to 100%, language variables are divided into three levels: "empty," "half," and "full"). The output variable is the percentage of supercapacitor discharge power. (Range 0% to 100%). A total of 27 fuzzy rules cover all combinations of input variables, and defuzzification uses the centroid method. Typical rules include: when... For "high" and "Low" and When it is "full", Take the "large" value, meaning the supercapacitor discharges at high power to protect the lithium battery from over-discharge damage; when "Low" and For "high" and When it is "empty", Taking "zero", the lithium battery independently supplies power and simultaneously charges the supercapacitor; when For "high" and For "middle" and When it is "half", Taking the "middle" approach, both systems discharge together. Under the premise that the overall weight of the machine is controlled within 5kg, the total usable energy of the hybrid power supply scheme is approximately 60Wh, which, calculated at an average power of 15W, can support at least 4 hours of continuous inspection operations.

[0162] The host computer communication and monitoring module 600 includes a wireless communication submodule 601 and a remote monitoring and emergency intervention submodule 602.

[0163] The wireless communication submodule 601 uses a 4G / 5G cellular network module as the main communication link, with an uplink bandwidth of no less than 10Mbps. It is also equipped with a LoRa backup link to provide low-power, long-range communication capabilities with a coverage radius of 3km. The system transmits data in real time, including: robot pose, velocity, and acceleration information updated at a frequency of 10Hz; and Lyapunov function values. and Remote stability monitoring is updated at a frequency of 5Hz; drive wheel contact force distribution and center of gravity offset are updated at a frequency of 10Hz; detection images and defect identification results are transmitted back on demand and compressed using H.265 encoding; battery The status of the supercapacitor is updated at a frequency of 1 Hz.

[0164] The remote monitoring and emergency intervention submodule 602 enables ground operators to view real-time trend graphs and video streams of various robot status parameters through a monitoring interface. The system's remote alarm mechanism corresponds to the three-level progressive safety threshold system defined in the center of gravity offset calculation submodule 204: when... Exceeding the Level 1 warning threshold When the center of gravity shifts by mm, the system automatically pushes a level one warning to the operator's terminal, prompting them to pay attention to the trend of center of gravity shift; when Persistent abnormality or Exceeding the second critical threshold When the speed reaches a certain level (mm), the system will send a level-two alarm and automatically activate deceleration protection measures. Operators can remotely send emergency braking commands, and the robot will immediately initiate the rail-holding braking procedure upon receiving the command.

[0165] Based on the above system architecture, this invention also provides an anti-interference adaptive control method for a tower inspection robot in a dynamic environment, such as... Figure 2 As shown, the method executes the following steps sequentially according to the transmission path of the signal and data stream.

[0166] In step S1, the system completes initialization and guide rail recognition. After the robot is installed on the iron tower's fall arrestor rail, it is powered on and started. The main controller loads all parameters from the multiphysics coupled dynamics solution module 100 from flash memory, and the edge AI computing module simultaneously loads and warms up the deep learning model. The sensor array submodule 201 begins collecting data. The inductive sensor identifies whether the guide rail is made of steel or aluminum, and the binocular vision camera acquires the geometric features in front of the guide rail. The control command decomposition and allocation submodule 305 automatically selects either the magnetic drive mode or the ratchet drive mode based on the guide rail material. The EKF state estimation submodule 203 initializes the state vector to the initial sensor sampling values. Error covariance matrix Initialized as a large diagonal matrix to represent the high uncertainty of the initial estimate. Adaptive gain. Initialize to a conservative value, taking 50% of the maximum possible disturbance amplitude of the system. (Lyapunov function initial value) Once the data is calculated and recorded, the system enters normal operating mode.

[0167] In step S2, the system continuously performs multimodal environmental perception and data acquisition. Within each 5ms control cycle, the sensor array submodule 201 synchronously acquires triaxial acceleration and angular velocity data from the inertial measurement unit, four-channel ultrasonic ranging values, normal contact force and tangential driving force data from the drive wheel assembly, and wind speed and direction data. Binocular visual images and infrared thermal images are acquired at a lower frequency. The multimodal data spatiotemporal alignment and feature fusion submodule 202 performs timestamp alignment and coordinate system-one transformation on all sensor data. The visual and infrared images are deeply fused by the edge AI computing module using Transformer network inference, outputting the guide rail surface defect detection results and the geometric feature parameters of the front guide rail.

[0168] In step S3, the extended Kalman filter state estimation is decoupled from environmental disturbances. The extended Kalman filter state estimation submodule 203 executes a prediction-update recursive algorithm on the main controller with a 5ms period. In the prediction step, the Lagrange dynamics equations are used based on the state estimate from the previous time step. and current control input A state-one-step prediction is performed, which implicitly incorporates the effects of physical factors such as guide rail flexibility deformation and tower vibration. The update step utilizes spatiotemporally aligned multi-sensor observations. Through Kalman gain The optimal weighted correction prediction result. Due to the disturbance... and Incorporating these as extended state variables into the estimation, the EKF update simultaneously outputs clean estimates of robot pose and velocity. and And the estimated wind load disturbance after separation and guide rail displacement fluctuation estimate The center of gravity offset calculation submodule 204 calculates the current center of gravity offset based on the pure pose estimate. Thus, the coupling effect between environmental interference, sensor noise, and control response has been quantitatively decoupled.

[0169] In step S4, the adaptive sliding mode control law is calculated in real time. The tracking error and sliding surface construction submodule 302 first calculates the pure pose estimate output by the EKF. Based on the information of the center of gravity offset, an augmented reference trajectory is generated by solving a nonlinear static optimization problem. This trajectory, based on pose tracking, simultaneously encodes the optimal joint angles that ensure center of gravity alignment and uniform distribution of normal force. Subsequently, the Lyapunov function construction and adaptive sliding mode control law calculation submodule 303 executes core control calculations with a 5ms cycle. Based on the augmented reference trajectory... Calculate augmented tracking error and augmented sliding surface vector Then calculate the equivalent control torque in sequence. and robust switching torque The total control torque is obtained by superimposing the two values ​​and then applying saturation limiting. The equivalent control torque is based on the augmented reference trajectory. The calculation implicitly includes the control objectives of center of gravity offset compensation and driving wheel normal force balance. Adaptive law Synchronously update the upper bound estimate of the perturbation Center of gravity offset compensation and drive wheel anti-slip control are achieved through augmented reference trajectory. The online updates continue to function: within each control cycle, the static optimization problem is resolved for the optimal joint angle based on the current EKF estimate. Small offsets naturally require small corrections, while large offsets automatically increase the correction, eliminating the need for explicit threshold switching logic. When the distance exceeds the Level 1 warning threshold by 2.5mm, the system will augment the weighting coefficients in the reference trajectory optimization problem. and Increased to 1.5 times the original value to enhance compensation and make the correction of the optimal joint angle more aggressive. Calculated at the end of the control cycle. and The monitoring system status is continuously within the normal range of eventual bounded convergence.

[0170] In step S5, control command decomposition and dual-mode drive are performed. The control command decomposition and allocation submodule 305 decomposes the five-dimensional generalized torque vector into specific commands for each actuator. The driving torque along the guide rail direction... The speed commands are converted to the left and right drive wheels, with a base speed of approximately 133 rpm. The differential component is determined by the lateral deviation compensation, and the speed accuracy is... rpm. The torque of the swing arm and articulated arm is converted into angle adjustment commands, realizing... Precision position control. Based on the target normal force distribution determined by the augmented reference trajectory optimization problem and the current drive mode, the normal contact force distribution is adjusted to control the target... A smooth transition of the driving mode is achieved through a 50ms gradual switching strategy in the steel-aluminum hybrid transition section.

[0171] In step S6, the model predictive control trajectory optimization submodule 304 performs trajectory rolling optimization at 50ms intervals. Based on the dynamic model, it predicts the motion state for the next 3 steps (150ms), and combines this with obstacle information detected by the vision module. Under constraints such as joint limits, torque saturation, and safe distance, it solves for the optimal reference trajectory through sequential quadratic programming. This optimized trajectory is then transmitted to the reference trajectory generation submodule 301, providing a better tracking target for the inner-loop adaptive sliding mode controller.

[0172] In step S7, the energy management and scheduling module 500 continuously monitors and schedules the system energy at 100ms intervals. During the normal, uniform climbing phase, the load rate is approximately 30%, with the lithium battery independently supplying power at approximately 15W and charging the supercapacitor in trickle mode. When encountering strong winds or obstacles, the load rate surges, and the fuzzy inference engine quickly determines the discharge power ratio of the supercapacitor. With a response time of less than 10ms, the supercapacitor releases a high-current pulse with a peak value of up to 15A through a bidirectional DC-DC converter to power the lithium battery in tandem, meeting the instantaneous high-power demand without causing over-discharge damage to the lithium battery.

[0173] In step S8, the system operates a multi-level security protection mechanism throughout the entire process. The first level is software protection: a check is performed every control cycle. ,like If the duration exceeds 10 control cycles (50ms), the system will... Increase climbing speed to 1.5 times the current value and reduce it to 50%; if If the duration exceeds 100 control cycles (500ms), the system determines it to be instability and automatically executes the shutdown procedure. Meanwhile, when... Exceeding the second critical threshold When the climbing speed reaches mm, the system automatically reduces the speed to 50% and sends a secondary alarm. The second level is hardware protection: when Exceeding Level 3 Absolute Safety Threshold If the force of any drive wheel drops below a critical value (below 100N for steel guide rails and below 80N for aluminum guide rails), the hardware interruption circuit immediately triggers the rail-holding brake. The ratchet fully locks to the maximum clamping force of 300N, the magnetic attraction force increases to the maximum of 350N, and all drive motors are de-energized. The third level is communication protection: when any level of protection is triggered, an alarm message is immediately pushed to the ground host computer, and the ground operator can send a remote emergency braking command at any time. If the communication interruption exceeds 60 seconds, the robot automatically stops on the spot and engages the rail-holding brake, waiting for communication to be restored or for manual intervention. The above three-level safety protection mechanism includes a three-level progressive safety threshold system (first-level warning) defined in the center of gravity offset threshold and center of gravity offset calculation submodule 204. mm, secondary critical mm, Level 3 absolute safety (mm) Strictly correspond to the severity level and response measures, with the severity and response measures increasing sequentially.

[0174] The system and method of this invention are applicable to continuous interference superimposed on guide rails in strong winds of force 8 (wind speed 20 m / s). Under test conditions of mm displacement fluctuation, the robot pose error is stably controlled within... Within mm, the center of gravity offset always remains within Within the second-order critical threshold of mm, the fluctuation of single-wheel contact force is suppressed to within 5%, the state estimation convergence time is less than 0.5 seconds, and the swing arm joint angle adjustment accuracy reaches... The differential speed adjustment accuracy of each drive wheel reaches The simultaneous achievement of the above performance indicators is attributed to the organic integration and mutual support of three core technical aspects: using a multi-physics coupled dynamics model as the mathematical foundation, using extended Kalman filtering to achieve precise decoupling of interference and noise, and using an augmented reference trajectory to integrate the center of gravity offset and the anti-slip target of the drive wheels into the tracking control framework, and rigorously proving the eventual uniform boundedness through Lyapunov stability theory.

[0175] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An anti-interference adaptive control system for a tower inspection robot in a dynamic environment, comprising a perception layer, a decision layer, an execution layer, and an energy management layer, wherein the perception layer is equipped with multimodal sensors, and the execution layer includes a drive actuator for driving the robot to move along the tower guide rail, characterized in that: The control system includes a multiphysics coupled dynamics solution module, a multimodal perception and state estimation module, an anti-interference adaptive control module, and an adaptive dual-mode drive execution module. The multiphysics coupled dynamics solution module pre-sets the structural parameters of the robot body and the guide rail and constructs the dynamic equations. The multiphysics coupled dynamics solution module uses the Lagrange equation to establish a five-degree-of-freedom multibody dynamics system. The five generalized coordinates are displacement along the guide rail direction, displacement perpendicular to the guide rail direction, roll angle around the guide rail axis, swing arm joint angle, and joint arm angle. The potential energy term of the system includes gravitational potential energy, elastic deformation potential energy of the guide rail based on equivalent stiffness, and magnetic attraction potential energy based on nonlinear decay of the air gap magnetic circuit equation. The multimodal perception and state estimation module uses extended Kalman filtering combined with multimodal sensor data to output the robot's generalized coordinate estimate and the decoupled environmental disturbance estimate. The state vector of the extended Kalman filter is a twelve-dimensional vector containing generalized coordinates, generalized velocity, wind load disturbance estimate and guide rail displacement fluctuation estimate. The discretized dynamic equation is used as the state transition equation. In the prediction update loop, the Jacobian matrix is ​​used to calculate the error covariance and output the pure pose and velocity estimate as well as the separated wind load and guide rail displacement fluctuation feedforward compensation. The anti-interference adaptive control module calculates the center of gravity offset based on the generalized coordinate estimate. It then applies corrections to the swing arm joint angles and articulated arm angles in the basic reference trajectory by solving a nonlinear optimization problem to obtain the augmented reference trajectory. The objective function of the nonlinear optimization problem includes the square of the center of gravity offset and the square of the deviation of the actual normal force of each drive wheel from the target uniform normal force, ensuring the center of gravity is located on the guide rail centerline and the normal force of the drive wheels is uniformly distributed. Based on the augmented reference trajectory, a Lyapunov candidate function is constructed, and the total control torque, including equivalent control torque and robust switching torque, is calculated. The Lyapunov candidate function consists of the kinetic energy metric term of the augmented tracking error on the sliding surface and the disturbance upper bound estimation error term. The equivalent control torque in the total control torque includes the modeled dynamics term and feedforward compensation of the environmental disturbance estimate. The robust switching torque includes the adaptive estimate of the disturbance upper bound and a continuous approximation function used to eliminate high-frequency flutter. The adaptive dual-mode drive execution module decomposes the total control torque and controls the drive execution mechanism's actions.

2. The anti-interference adaptive control system for the tower inspection robot in a dynamic environment according to claim 1, characterized in that: The adaptive dual-mode drive execution module includes a magnetic drive unit suitable for steel guide rails and a ratchet drive unit suitable for non-ferromagnetic guide rails. The control system switches the drive mode based on the guide rail material identified by the inductive sensor. In the steel-aluminum mixed transition section, a gradual switching strategy is adopted in which the magnetic attraction force and the ratchet clamping force increase and decrease synchronously and linearly.

3. The anti-interference adaptive control system for the tower inspection robot in a dynamic environment according to claim 1, characterized in that: The anti-interference adaptive control module also includes a model predictive control trajectory optimization submodule. This submodule runs on the outer loop of the adaptive sliding mode controller and solves the optimization problem constrained by joint limits, torque saturation, and obstacle safety distance through a sequential quadratic programming algorithm, generating a smooth reference trajectory for multiple future prediction steps online.

4. The anti-interference adaptive control system for the tower inspection robot in a dynamic environment according to claim 1, characterized in that: The system also includes an energy management and scheduling module, which adopts a hybrid power supply architecture of lithium battery and supercapacitor. Based on the robot's current load rate, lithium battery state of charge and supercapacitor remaining energy ratio, the system outputs the supercapacitor discharge power ratio through a fuzzy inference engine, and controls the supercapacitor to release a large current pulse when there is a momentary high power demand.

5. An anti-interference adaptive control method for a tower inspection robot under dynamic conditions, applied to the control system as described in any one of claims 1 to 4, characterized in that, The method includes the following steps: system initialization and loading of dynamic model parameters, and determination of driving mode based on guide rail material; continuous multimodal environmental perception, collection of body motion state data, contact force data and environmental meteorological data; execution of extended Kalman filtering with a preset control cycle, state prediction based on dynamic equations and updating through multi-source observations, outputting decoupled pure system pose estimation and environmental disturbance estimation, wherein the state vector of extended Kalman filtering is a twelve-dimensional vector containing generalized coordinates, generalized velocity, wind load disturbance estimation and guide rail displacement fluctuation estimation, using discretized Lagrange dynamic equations as state transition equations, calculating error covariance using Jacobian matrix in prediction update loop, outputting pure pose and velocity estimation and separated wind load and guide rail displacement fluctuation feedforward compensation; The center of gravity offset is calculated based on the system pose estimation. By solving a nonlinear static optimization problem, corrections are applied to the joint angles of the swing arm and the joint arm in the basic reference trajectory to generate an augmented reference trajectory that achieves center of gravity repositioning and normal force balance. The objective function of the nonlinear optimization problem includes the square term of the center of gravity offset and the square term of the deviation of the actual normal force of each drive wheel from the target uniform normal force. A Lyapunov candidate function is constructed, consisting of the kinetic energy metric term of the augmented tracking error on the sliding surface and the upper bound estimation error term of the disturbance. Based on Lyapunov stability theory, the equivalent control torque including the modeled dynamics term and the feedforward compensation of the environmental disturbance estimate, and the robust switching torque including the upper bound adaptive estimate of the disturbance and the continuous approximation function are calculated. The two are superimposed to form the total adaptive sliding mode control torque. The total control torque is decomposed into the speed commands of each drive wheel and the joint servo angle commands to drive the robot to perform anti-disturbance adaptive actions.

6. The anti-interference adaptive control method for a tower inspection robot in a dynamic environment according to claim 5, characterized in that: The method also performs multi-level safety protection steps. In each control cycle, the derivative value of the Lyapunov function is calculated. If the derivative value is greater than zero and continues for more than a preset time, the upper limit of the disturbance is increased and deceleration protection is triggered. When the center of gravity offset exceeds the absolute safety threshold or the normal force of any drive wheel drops below the critical value, a hardware interrupt is triggered to make the robot perform rail-hugging braking and cut off the power supply to the drive motor.

Images