Intelligent multi-mode inspection method and system based on inspection robot

Abstract

The application discloses an intelligent multi-mode inspection method and system based on an inspection robot, and belongs to the technical field of intelligent inspection robots. The method comprises the following steps: synchronously acquiring point cloud data and multispectral image data through a panoramic laser radar; performing voxel downsampling and normal vector estimation on the point cloud data to generate an environment complexity Fourier descriptor; performing multi-scale Retinex enhancement and frequency domain saliency detection on the multispectral image data to extract a phase shift amount of a multispectral structured light fringe; inputting the environment complexity Fourier descriptor and the phase shift amount into a preset chaotic mapping decision network to adaptively trigger a high-frequency sampling mode or a low-frequency inspection mode, and outputting a composite state vector. The system comprises a memory and a processor, and a sensor array is dynamically adjusted in collection frequency by a gain coefficient in the composite state vector. The application realizes adaptive perception of the environment complexity and multi-mode switching of the inspection robot.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection robot technology, specifically to an intelligent multi-mode inspection method and system based on inspection robots. Background Technology

[0002] Traditional inspection robots, when performing inspection tasks in unstructured or dynamically changing environments, typically employ fixed-frequency data acquisition strategies and preset inspection paths. This fixed approach struggles to adjust sensor sampling and motion planning in real-time to adapt to environmental changes, resulting in insufficient information acquisition in complex scenarios and excessive resource consumption in simpler ones. Existing environmental perception methods based on single LiDAR or visual sensors often rely on statistical histograms, geometric features, or depth features for feature extraction, failing to adequately represent the frequency domain complexity and texture details of the environment. This leads to ambiguous scene discrimination and difficulty in accurately triggering multi-mode switching. Some solutions use simple threshold comparisons or rule bases for mode switching, resulting in delayed responses to environmental changes, and threshold settings rely on human experience, leading to poor adaptability.

[0003] Furthermore, in terms of inspection path planning and robotic arm control, existing methods often decouple environmental perception and motion control, lacking an end-to-end adaptive mechanism from perception to action. For the analysis of structured light and multispectral images, current technologies mostly remain at the level of defect detection or target recognition, failing to fuse the structured light phase shift and the spatially complex features of laser point clouds for decision-making. The application of chaotic mapping and pulse-coupled neural networks in robot decision-making remains rare; most decision networks employ traditional neural networks or finite state machines, making it difficult to generate flexible and noise-resistant decision outputs in highly uncertain environments. Therefore, a method is needed that can deeply fuse point cloud frequency domain features and multispectral phase features, and achieve adaptive switching between high-frequency sampling and low-frequency inspection through a nonlinear chaotic decision-making mechanism, while directly using the decision results for robotic arm trajectory generation and chassis path replanning. Summary of the Invention

[0004] This invention provides an intelligent multi-mode inspection method and system based on an inspection robot, aiming to solve the problems of incomplete environmental feature representation, lagging mode switching, and separation of perception and motion control in existing inspection methods. By performing voxel downsampling and normal vector field extraction on point cloud data and converting it to the spherical harmonic domain and frequency domain, a Fourier descriptor for environmental complexity is constructed to fully express the complex spatial structure information of the environment. Multispectral images are enhanced and saliency detected, and structured light phase shifts are extracted to reflect surface micro-changes. Based on this, Logistic and Henon mappings are used to perform chaotic iteration on the above features, generating chaotic sequences. A decision chaotic seed is formed through XOR operation, driving a pulse-coupled neural network to adaptively determine the high-frequency sampling or low-frequency inspection mode based on the pulse duty cycle, outputting a composite state vector containing mode identifiers and gain coefficients. This decision result further guides the real-time adjustment of the robotic arm's scanning trajectory and the chassis obstacle avoidance path, realizing a closed loop of perception-decision-control.

[0005] To achieve the above objectives, this invention provides the following technical solution: This invention provides an intelligent multi-mode inspection method based on an inspection robot. It simultaneously acquires point cloud data and multispectral image data of the working environment using a panoramic LiDAR mounted on the robot's base. The point cloud data undergoes voxel downsampling and normal vector estimation to generate an environmental complexity Fourier descriptor. The multispectral image data is then subjected to multi-scale Retinex enhancement and frequency domain saliency detection to extract the phase shift of the multispectral structured light fringes. The environmental complexity Fourier descriptor and the phase shift are input into a preset chaotic mapping decision network, adaptively triggering either a high-frequency sampling mode or a low-frequency inspection mode, outputting a composite state vector containing a mode identifier and gain coefficients. This method can autonomously switch inspection modes based on environmental complexity and spectral phase changes. In complex environments, it automatically increases the sampling frequency to obtain fine data, while maintaining low-frequency inspection in simple areas to save computational and energy resources, thus balancing inspection efficiency and accuracy.

[0006] As a technical solution of this invention, the following method is used when generating a Fourier descriptor for environmental complexity by voxel downsampling and normal vector estimation of point cloud data: The point cloud data is divided into a three-dimensional mesh according to a preset voxel side length; the average point coordinates and point density within each voxel are extracted to construct a voxelized point cloud matrix; based on the voxelized point cloud matrix, the eigenvector corresponding to the minimum eigenvalue of the covariance matrix of each point within the voxel is calculated using principal component analysis, and used as a local normal vector to generate a normal vector field; the normal vector field is mapped to a spherical harmonic function basis, and a Fourier transform is performed on the spherical harmonic function basis to obtain a frequency domain coefficient matrix; then, a logarithmic polar coordinate transformation is performed on the frequency domain coefficient matrix to obtain the Fourier descriptor for environmental complexity. This descriptor can accurately characterize the complexity of environmental geometric features, providing a stable feature representation for subsequent decision-making.

[0007] Preferably, when inputting the Fourier descriptor of environmental complexity and the phase shift into the chaotic mapping decision network, the Logistic mapping is used to perform chaotic iteration on the Fourier descriptor of environmental complexity to generate a first chaotic sequence, and the Henon mapping is used to perform chaotic iteration on the phase shift to generate a second chaotic sequence. The first and second chaotic sequences are then XORed to obtain a decision chaotic seed, which is used as the initial membrane potential of a preset pulse-coupled neural network to drive the pulse-coupled neural network to fire a pulse sequence. The high-frequency sampling mode or low-frequency inspection mode is determined based on the duty cycle of the pulse sequence, and the sampling mode is gain-modulated based on the peak phase of the pulse sequence to output a composite state vector. By fusing chaotic features from different sources and leveraging the nonlinear pulse firing characteristics of the pulse-coupled neural network, the mode switching exhibits high sensitivity and noise resistance, enabling timely response to subtle changes in the environment.

[0008] Based on the aforementioned composite state vector, the method of this invention further drives the robotic arm joints of the inspection robot to perform motion trajectory planning and generate a joint angle sequence. Specifically, the mode identifier bit in the composite state vector is decoded to obtain the scanning cone angle and step increment of the high-frequency sampling mode or the low-frequency inspection mode; using the scanning cone angle, step increment, and gain coefficient as input, the Jacobian matrix of the robotic arm links is constructed using the DH parameter method; singular value decomposition is performed on the Jacobian matrix, and the right singular vector corresponding to the minimum singular value is extracted as the obstacle avoidance direction vector; cubic spline interpolation is used to interpolate the step increment under the constraint of the obstacle avoidance direction vector to generate a joint angle sequence containing timestamps. This allows the robotic arm's motion during the inspection process to adapt to the requirements of the sampling mode and actively avoid obstacles, improving operational safety.

[0009] When decoding the mode identifier, the binary code of the mode identifier can be extracted and Gray code decoded to obtain the mode index value. A nonlinear mapping table between the mode index value and the scan cone angle and step increment is established, and the corresponding parameters are looked up in the table. When the mode index value is outside the boundary of the mapping table, radial basis function interpolation is used to extrapolate the scan cone angle and step increment to generate extended parameters. This method ensures smooth transition of motion parameters when the mode space changes continuously, avoiding mechanical shock.

[0010] When generating joint angle sequences using cubic spline interpolation, a system of equations for the cubic spline interpolation function is constructed, with the step increment as the time variable and the initial and target angles of the robotic arm joints as node values. An obstacle avoidance direction vector is introduced into the boundary conditions of this system, and its components are used as constraints on the first derivative of the spline curve. Solving the system of equations with the obstacle avoidance direction vector in mind yields the angle values ​​for each interpolation node, which are then arranged by timestamp to generate the joint angle sequence. The resulting trajectory maintains continuous acceleration and smooth motion while satisfying obstacle avoidance constraints.

[0011] As another technical solution of this invention, a global path replanning is performed on the tracked chassis of the inspection robot driven by a composite state vector to generate an obstacle avoidance path point set. The specific process is as follows: the gain coefficient in the composite state vector is thresholded to obtain an obstacle density index; the obstacle density index, the current pose of the inspection robot, and a preset grid map are input into a fast expanding random tree algorithm, using the obstacle density index as an adjustment factor for the expansion step size, so that the expansion step size of the fast expanding random tree algorithm is equal to the base step size multiplied by the reciprocal of the obstacle density index, generating exploration tree nodes; redundant nodes are pruned from the exploration tree nodes, and the node sequence of the shortest path from the starting node to the target node is extracted as the obstacle avoidance path point set. In this way, the step size of the path planning can be adaptively adjusted according to the density of environmental obstacles, using a small step size for fine-grained search in dense obstacle areas and a large step size for rapid expansion in open areas, balancing planning quality and speed.

[0012] When thresholding the gain coefficient, the gain coefficient is compared with a first preset threshold and a second preset threshold, where the first preset threshold is greater than the second preset threshold. If the gain coefficient is greater than the first preset threshold, the output is a high density index; if the gain coefficient is less than or equal to the first preset threshold and greater than or equal to the second preset threshold, the output is a medium density index; if the gain coefficient is less than the second preset threshold, the output is a low density index. These high, medium, and low density indices are used as obstacle density indices. This hierarchical approach enables path planning to have a hierarchical response capability to obstacle distribution.

[0013] When pruning redundant nodes in the exploration tree, the tree is traversed backwards from the target node, and the parent node of each node is extracted to construct a parent node linked list. Dijkstra's shortest path search is performed on the parent node linked list, calculating the Euclidean distance between adjacent nodes and summing them to obtain the total path length. The node sequence corresponding to the path branch with the smallest total path length is selected. Then, the Douglas-Puk algorithm is applied to this node sequence to simplify it, removing intermediate nodes whose distance from the straight-line segment is less than a preset threshold, thus obtaining the obstacle avoidance path point set. The resulting path guarantees global optimality and reduces unnecessary intermediate turning points through simplification, facilitating smooth operation of the tracked chassis.

[0014] This invention also provides an intelligent multi-mode inspection system based on an inspection robot, including a memory and a processor. The memory contains a program for an intelligent multi-mode inspection method based on the inspection robot, which, when executed by the processor, implements the aforementioned method steps. The system also includes a three-axis gimbal, an integrated sensor array, and a tracked chassis. The three-axis gimbal is fixed to the end effector of the inspection robot's robotic arm. The integrated sensor array includes a panoramic LiDAR, a visible light camera, and an infrared thermal imager, and the acquisition frequency of the integrated sensor array is dynamically adjusted by the gain coefficient in the composite state vector. Through multi-sensor integration and adaptive frequency adjustment, the system can acquire multi-dimensional, high-quality perception data in different inspection modes, enhancing its adaptability to complex environments.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0016] By generating a Fourier descriptor for environmental complexity through voxel downsampling and normal vector estimation, the spatial distribution of point clouds is transformed into frequency domain invariants. Specifically, after voxelizing the point cloud, the normal vector within each voxel is calculated to construct a normal vector field. This field is then mapped to a spherical harmonic basis and subjected to a Fourier transform to obtain the frequency domain coefficient matrix. Finally, a logarithmic polar coordinate transformation is used to obtain the Fourier descriptor for environmental complexity. This approach transforms the spatial rotation and scale changes of the environmental structure into a frequency domain translation, giving the descriptor rotation invariance and scale robustness. Compared to traditional point cloud histograms or geometric moment features, it significantly improves sensitivity and discriminative power to changes in environmental complexity, producing significant differences between cluttered and regular scenes, and providing stable and highly discriminative feature inputs for subsequent decision-making.

[0017] The Fourier descriptor of environmental complexity and the phase shift of multispectral structured light are subjected to chaotic iteration through Logistic and Henon mappings, respectively, to obtain two chaotic sequences. A decision-making chaotic seed is generated by bitwise XOR and used as the initial membrane potential of the pulse-coupled neural network to drive pulse firing. Based on the duty cycle of the pulse sequence, a high-frequency sampling mode or a low-frequency inspection mode is adaptively triggered, and the gain is modulated by the pulse peak phase to output a composite state vector. This chaotic mapping decision network introduces a decision-making process that fuses two heterogeneous features into a nonlinear dynamic orbit. Its initial value sensitivity and pseudo-randomness enable identifiable decision outputs under similar but different environmental conditions, avoiding the step oscillations and lags of traditional threshold decision-making. The synchronous firing characteristics of the pulse-coupled neurons make mode switching smoother, and the continuous modulation of the gain coefficient allows for stepless adjustment of robot sensing sampling and motion speed, thus achieving true adaptive inspection. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a flowchart of an intelligent multi-mode inspection method based on inspection robots;

[0020] Figure 2 This is a flowchart of the method for generating Fourier descriptors for environment complexity;

[0021] Figure 3 It is a flowchart of the inspection mode decision-making process based on chaotic mapping and pulse coupling neural network;

[0022] Figure 4 This is a flowchart of the process for generating the joint angle sequence of a robotic arm based on a composite state vector.

[0023] Figure 5 This is a schematic diagram of an intelligent multi-mode inspection system based on an inspection robot. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] See Figure 1 This invention provides an intelligent multi-mode inspection method and system based on an inspection robot. The method uses a panoramic LiDAR mounted on the robot's base to simultaneously acquire point cloud data and multispectral image data of the working environment. The point cloud data undergoes voxel downsampling and normal vector estimation to generate an environmental complexity Fourier descriptor. The multispectral image data is then subjected to multi-scale Retinex enhancement and frequency domain saliency detection to extract the phase shift of the multispectral structured light fringes. The environmental complexity Fourier descriptor and the phase shift are input into a pre-defined chaotic mapping decision network, which adaptively triggers either a high-frequency sampling mode or a low-frequency inspection mode, outputting a composite state vector.

[0026] Example 1:

[0027] In specific implementation, please refer to Figure 2 Point cloud data acquired by panoramic LiDAR is a collection of a large number of discrete points in three-dimensional space, each point containing spatial coordinates. And possible intensity information. First, voxel downsampling and normal vector estimation are performed on the point cloud data to generate a Fourier descriptor for environment complexity.

[0028] The point cloud data is divided into 3D meshes according to a preset voxel side length. The average point coordinates and point density within each voxel are extracted to construct a voxelized point cloud matrix. The preset voxel side length can be a fixed value set based on prior knowledge of the inspection environment; for example, the voxel side length is set to 0.05 meters in an indoor substation environment and 0.1 meters in an outdoor pipe gallery environment. The space containing the point cloud data is divided into several cubic units along the three coordinate axes according to the preset voxel side length, with each cubic unit being a voxel. For a given voxel, the 3D coordinates of all points falling within that voxel are counted, and the arithmetic mean of these point coordinates is calculated to obtain the average point coordinates of that voxel. Simultaneously, the number of points within the voxel is counted, and the point density is obtained by dividing the number of points by the voxel volume. The average point coordinates and point density of all voxels are organized into a three-dimensional matrix according to the voxel index. For each voxel, the matrix elements are four-dimensional vectors. This three-dimensional matrix is ​​the voxelized point cloud matrix. During the construction process, if a voxel does not contain any points, its average point coordinates are marked as empty, and the point density is set to zero.

[0029] Based on the voxelized point cloud matrix, principal component analysis is used to calculate the eigenvector corresponding to the minimum eigenvalue of the covariance matrix of each voxel, which serves as the local normal vector to generate a normal vector field. For each non-empty voxel in the voxelized point cloud matrix, the original point cloud coordinates within that voxel are used as input to construct the covariance matrix. Specifically, suppose a voxel contains... There are points, each with coordinates of . Calculate the centroid coordinates And construct the covariance matrix Covariance matrix for Real symmetric matrix. Solve for the covariance matrix using the Jacobi iteration method or QR decomposition method. From the eigenvalues ​​and corresponding eigenvectors, we obtain three eigenvalues. and the corresponding feature vector , , Select the smallest eigenvalue. corresponding feature vector As the local normal vector of the point cloud surface within this voxel, Normalize to a unit vector. Traverse all non-empty voxels and assign a local normal vector to each voxel. All local normal vectors of voxels constitute a three-dimensional normal vector field. The normal vectors at each position in the normal vector field are stored in the form of a three-dimensional vector.

[0030] The normal vector field is mapped to a spherical harmonic function basis. A Fourier transform is performed on the spherical harmonic function basis to obtain the frequency domain coefficient matrix. A logarithmic polar coordinate transformation is then performed on the frequency domain coefficient matrix to obtain the Fourier descriptor of the environmental complexity. Specifically, the mapping process involves representing the unit normal vector at each voxel position in the normal vector field as the azimuth angle in spherical coordinates. and polar angle , The range of values ​​is , The range of values ​​is Define spherical harmonic functions. ,in It is a non-negative integer order. are integers and satisfy The expression for the spherical harmonic function is:

[0031]

[0032] in, To relate Legendre polynomials, The imaginary unit, the order The upper limit of the value is set according to the precision of the environment complexity description, and in a specific implementation The upper limit of its value is set to 5. For the normal vector direction of each voxel in the normal vector field... Calculate its projection coefficients on the basis of each order of spherical harmonic function to obtain a series of spherical harmonic coefficients. The spherical harmonic coefficients of all voxels are organized into a multidimensional array called the spherical harmonic coefficient field. Discrete Fourier transforms are performed on the spherical harmonic coefficient field along each of the three spatial dimensions to transform the spatially transformed spherical harmonic coefficients to the frequency domain, resulting in a frequency domain coefficient matrix. The elements of the frequency domain coefficient matrix are complex numbers, with indices corresponding to the spatial frequency components and the spherical harmonic order. Next, a logarithmic polar transformation is performed on the frequency domain coefficient matrix to convert the frequency domain coordinates to Cartesian coordinates. Mapping to logarithmic polar coordinates ,in , and The direction angle in the frequency space is denoted as . The transformed coefficients are rearranged according to the logarithmic radius and direction to form a Fourier descriptor for environmental complexity. This descriptor encodes the geometric complexity and structural directionality of the environment and is rotation- and scale-invariant.

[0033] Example 2:

[0034] In specific implementation, please refer to Figure 3The Fourier descriptor of environmental complexity and the phase offset are input into a pre-defined chaotic mapping decision network. This chaotic mapping decision network is composed of a Logistic mapping module, a Henon mapping module, an XOR operation module, and a pulse-coupled neural network module cascaded in sequence. The Fourier descriptor of environmental complexity is a multi-dimensional complex array, and the phase offset is an angle value in radians, with a value range of [-π, π].

[0035] The first chaotic sequence is generated by performing chaotic iteration on the Fourier descriptor of environmental complexity using a Logistic mapping. First, the Fourier descriptor of environmental complexity is compressed by calculating the root mean square (RMS) value of the complex modulus of all elements, denoted as the RMS descriptor value. The RMS descriptor value is then mapped to the open interval (0,1) using a preset maximum-minimum normalization method. The maximum and minimum parameters used in the mapping process are pre-calibrated according to the inspection environment type. In a specific implementation, for an indoor electrical inspection environment, the maximum parameter is set to 8.5, and the minimum parameter is set to 0.3. The normalized values ​​are used as the initial state values ​​for the Logistic mapping. The iterative equation for the Logistic mapping is:

[0036]

[0037] in, Indicates the first The state value generated in the next iteration. Indicates the first The state value generated in the next iteration. The branch parameter for the Logistic mapping is set to 3.99. This value is based on the fact that when the branch parameter is 3.99, the dynamics of the Logistic mapping enter a completely chaotic state, and the output sequence exhibits extreme sensitivity to initial values ​​and a non-periodic ergodic behavior. Using the initial value, iterate 256 times to obtain the result. The sequence of real numbers formed is denoted as the first chaotic sequence.

[0038] A second chaotic sequence is generated by performing chaotic iteration on the phase offset using Henon mapping. The phase offset is extracted and denoted as... .Will The two initial state components are mapped to the Henon mapping: Let the initial... ,initial The iterative relationship of the Henon mapping is: , .in, and The first The two state components of the next iteration. and The first The two state components of the next iteration. and These are the control parameters for Henon mapping. Among them... Set to 1.4, parameter The value is set to 0.3 because this parameter combination represents the classic standard Henon map, which can generate stable odd attractors and ensure the chaotic properties of the sequence. , Using the initial value, iterate 256 times according to the Henon mapping iteration relationship described above, and take the value generated in each iteration. Components, obtained from The resulting sequence of real numbers is denoted as the second chaotic sequence.

[0039] Perform a bitwise XOR operation between the first chaotic sequence and the second chaotic sequence to obtain the decision chaotic seed. The specific operation process is as follows: traverse the sequence indices; for the... The elements in the first chaotic sequence In the second chaotic sequence Convert each 16-bit binary number separately. The conversion method is to multiply the real number by 65535, round down, and take the lower 16 bits. Perform a bitwise XOR operation on the two 16-bit binary numbers to obtain a new 16-bit binary number. Then convert this binary number to an unsigned integer, which will be used as the first... The XOR result is 256 times. This process is repeated 256 times to obtain a sequence of unsigned integers of length 256. This sequence is the decision chaos seed.

[0040] A decision chaos seed is used as the initial membrane potential of a pre-defined pulse-coupled neural network (PWMN), driving the PWMN to fire a pulse sequence. The PWMN employs a single-neuron structure, comprising five parts: a feed input channel, connection input channels, a modulation domain, membrane potential, and a dynamic threshold. The pre-defined feed input of the PWMN is set to a constant value of 1.0, and the connection strength coefficient is... Set to 0.15, Used to adjust the degree of influence of the connection input on the modulation domain; membrane potential decay time constant. Set to 0.08, Controlling the exponential decay rate of the membrane potential without pulse input; dynamic threshold decay time constant. Set to 0.1, Controlling the exponential decay rate of the dynamic threshold when there is no pulse input; threshold amplification factor Set to 10.0. Determines the magnitude of the rise in the dynamic threshold after each pulse delivery; membrane potential amplification factor. Set to 0.5. The magnitude of the membrane potential rise after each pulse is determined. These parameters are set based on the balance between the network's pulse firing rate and the sensitivity to the input stimulus. The first element of the decision chaos seed is used as the initial value of the membrane potential. dynamic threshold initial value Set to 0, initial value of pulse output Set to 0. At each discrete time step Calculate sequentially: Connect input ,in For the first The input for the next step is determined by the pulse output of the previous step. For the first The pulse output value of the step; modulation domain ,in, The modulation threshold at step t is modulated by both the feed input and the connection input; membrane potential ,in For the first The membrane potential value of the step is obtained by superimposing the membrane potential decay and pulse contribution of the previous step; dynamic threshold ,in For the first The dynamic threshold of the step is obtained by superimposing the threshold decay of the previous step with the pulse contribution; comparison and ,like but ,otherwise ,in For the first The pulse output value of the step. Simultaneously, after calculating the pulse output, the membrane potential is additively perturbed using the next element in the decision chaos seed sequence, i.e. ,in For the decision chaotic seed sequence Each element is normalized to the value in the interval [0, 0.2]. For the first The initial membrane potential value is used for each iteration. This process is repeated for 256 time steps to obtain the pulse sequence. .

[0041] The high-frequency sampling mode or low-frequency inspection mode is determined based on the duty cycle of the pulse sequence. The duty cycle is defined as the proportion of time steps with a value of 1 out of the total 256 time steps in the pulse sequence. A duty cycle determination threshold is set. The duty cycle is 0.35. If the duty cycle of the pulse sequence is greater than or equal to 0.35, it is determined to be a high-frequency sampling mode, and the mode identifier bit is set to binary "1"; if the duty cycle of the pulse sequence is less than 0.35, it is determined to be a low-frequency inspection mode, and the mode identifier bit is set to binary "0".

[0042] Gain modulation is applied to high-frequency sampling mode or low-frequency inspection mode based on the peak phase of the pulse sequence, outputting a composite state vector. The instantaneous phase of the pulse sequence is calculated using Hilbert transform to extract the analytic signal, and the instantaneous phase value at each time step is calculated to obtain the instantaneous phase sequence. The probability density distribution of the instantaneous phase values ​​in the interval [0, 2π) is statistically analyzed, and the phase value corresponding to the maximum probability density value is found; this phase value is the peak phase. The gain coefficient is calculated using a pre-defined linear mapping relationship: Gain Coefficient This allows the gain coefficient to range from [0.5, 1.0]. The mode identifier and the gain coefficient are combined to form a composite state vector, which is stored as a structure, where the mode identifier is a single byte and the gain coefficient is a double-precision floating-point number.

[0043] Example 3:

[0044] In specific implementation, please refer to Figure 4 The inspection robot's robotic arm performs motion trajectory planning based on a composite state vector-driven system, generating a sequence of joint angles. The robotic arm is a multi-degree-of-freedom serial joint structure, with each joint driven by an independent servo motor, and the joint angles are fed back in real time via encoders.

[0045] The pattern identifier bits in the composite state vector are decoded to obtain the scan cone angle and step increment of the high-frequency sampling mode or the low-frequency inspection mode. Specifically, the binary code of the pattern identifier bits is extracted. The pattern identifier bits are single-byte values ​​with a binary code length of 8 bits. Gray code decoding is performed on the binary code to obtain the pattern index value. The Gray code decoding process is as follows: Let the highest bit of the binary code be the highest bit of the Gray code, denoted as... The highest bit of the pattern index value Directly equal to For the remaining bits, calculate them sequentially from the most significant bit to the least significant bit. ,in This represents the XOR operation. The first binary encoding Bit, The first value of the pattern index Bit, The first value of the pattern index Bit binary value, Decrease from 6 to 0. After decoding, an 8-bit binary number is obtained. Convert it to a decimal integer, which is the pattern index value. The pattern index value ranges from 0 to 255.

[0046] A nonlinear mapping table is established between pattern index values, scan cone angle, and step increment. The nonlinear mapping table is a one-dimensional lookup table structure containing 256 entries. Each entry corresponds to a pattern index value, storing a scan cone angle value and a step increment value. The scan cone angle is defined as the angle between the field of view axis of the end effector sensor and the vertical axis of the end effector base, in degrees, with a storage range of 15 to 75 degrees. The step increment is defined as the angular step size along the scanning direction when the end effector performs a single step, in degrees, with a storage range of 0.1 to 2.0 degrees. The nonlinear mapping table is constructed as follows: In the offline phase, multiple typical scenarios are first arranged in the inspection environment. The optimal combination of scan cone angle and step increment is manually set for each scenario, collecting a total of 256 data points. These 256 data points are sorted from low to high according to the optimal scan cone angle and assigned pattern index values ​​from 0 to 255, thus forming a complete nonlinear mapping table.

[0047] The scan cone angle and step increment corresponding to the pattern index value are retrieved from the nonlinear mapping table. The decoded pattern index value is used as the lookup address to directly read the scan cone angle and step increment stored at that address in the nonlinear mapping table.

[0048] If the pattern index value is outside the boundary of the nonlinear mapping table, radial basis function interpolation is used to extrapolate the scan cone angle and step increment, generating expanded scan cone angle and step increment. Within the boundary of the nonlinear mapping table refers to pattern index values ​​falling within the range of 0 to 255; outside the boundary refers to pattern index values ​​less than 0 or greater than 255. When the pattern index value is less than 0, radial basis function interpolation is used for extrapolation. The radial basis function chosen is the Gaussian kernel function, which has the following form: ,in These are the radial basis function values. For shape parameters, The Euclidean distance between the input variable and the center point of the radial basis function is used. The five entries with the smallest mode index values ​​in the nonlinear mapping table are selected as extrapolation reference points. The mode index values ​​of these five reference points and their corresponding scan cone angles are used as input-output pairs to construct a radial basis function interpolation model. After solving for the interpolation coefficients, mode index values ​​less than 0 are substituted into the interpolation model to extrapolate the expanded scan cone angle. The step increment is extrapolated in the same way. When the mode index value is greater than 255, the five entries with the largest mode index values ​​in the nonlinear mapping table are selected as extrapolation reference points, and the same extrapolation process is performed to obtain the expanded scan cone angle and step increment.

[0049] Using the scanning cone angle, step increment, and gain coefficient as inputs, the Jacobian matrix of the robotic arm links is constructed using the Denavit-Hartenberg parameter method. The number of joints in the inspection robot arm is denoted as... Each joint Corresponding to a set of DH parameters: Link length Linkage torsion angle Linkage offset and joint angle In practice, The value is set to 6, meaning the robotic arm has 6 rotary joints. Based on the physical dimensions of the robotic arm links, the value of each link is calibrated. , , The homogeneous transformation matrix between the coordinate systems of each link is determined. The homogeneous transformation matrices from the robot arm base to the end effector are multiplied together to obtain the pose matrix of the end effector relative to the base coordinate system. The position vector of the end effector is then adjusted for each joint angle. Taking the partial derivatives, we obtain the position Jacobian matrix. Each element; the attitude angle of the end effector for each joint angle. Taking the partial derivatives, we obtain the attitude Jacobian matrix. Each element of the position Jacobian matrix. and posture Jacobian matrix By piecing them together, a complete structure can be formed. Viacobi matrix The scanning cone angle determines the target orientation of the end effector, the step increment determines the joint space span between adjacent path points, and the gain coefficient is introduced as a damping factor in the pseudo-inverse operation of the Jacobian matrix.

[0050] Singular value decomposition is performed on the Jacobian matrix, and the right singular vector corresponding to the minimum singular value is extracted as the obstacle avoidance direction vector. The singular value decomposition expression is: ,in for An orthogonal matrix whose column vectors are left singular vectors; for A diagonal matrix, with diagonal elements as singular values Arranged from largest to smallest; for An orthogonal matrix whose column vectors are right singular vectors. Select... Minimum singular value on the diagonal , correspond In the matrix Right singular vectors with the same column index are denoted as... , This is the obstacle avoidance direction vector. The obstacle avoidance direction vector indicates the direction of motion in joint space that has the least impact on the end effector position. When the end effector of the robotic arm moves in this direction, the disturbance to the overall configuration is minimized, which is beneficial for avoiding obstacles.

[0051] Cubic spline interpolation is used to interpolate the step increment under obstacle avoidance direction vector constraints, generating a joint angle sequence containing timestamps. Specifically, a system of equations for the cubic spline interpolation function is constructed, with the step increment as the time variable and the initial and target angles of the robotic arm joints as node values. The initial angle is read from the encoders of each joint of the robotic arm, and the target angle is obtained through inverse kinematics based on the scanning cone angle and the current pose. Let the time variable interval corresponding to the step increment be... ,in , The total motion time is obtained by dividing the step increment by the preset end-effector linear velocity. The motion is uniformly inserted within the interval. Each internal node, total number of nodes The value is 50. Let the first... The time independent variable at each node is The joint angle value is , For each joint, construct a cubic spline interpolation function. , In each sub-interval The above is a cubic polynomial. The system of equations for the cubic spline interpolation function includes: the function value at the endpoints of each subinterval is equal to the nodal value, totaling... There are 1 equation; the first derivative is continuous at the interior nodes, totaling 1. There are 10 equations; the second derivative is continuous at the internal nodes, totaling 100 equations. There are equations. The total number of the above equations is . There are 1, and the total number of unknown coefficients for cubic splines is 2. To make the system of equations solvable, two boundary conditions are added. An obstacle avoidance direction vector is introduced into the boundary conditions. The components are used as constraints on the first derivative of the spline curve. Obstacle avoidance direction vector. for dimensional vector, at the starting node At this point, the first derivative of each joint angle is set and The components are proportional, and the proportionality coefficient is... Set as the desired obstacle avoidance speed at the end point. It is determined by the product of the gain coefficient and the step increment.

[0052] Solve the system of equations for the cubic spline interpolation function after introducing the obstacle avoidance direction vector to obtain the angle value of each interpolation node. Solve the linear system constructed from the above equations using either the chasing method or Gaussian elimination to obtain all coefficients of the cubic spline polynomial for each joint. Substitute the time independent variable of each interpolation node into the corresponding cubic spline polynomial to calculate the joint angle value of that node. Repeat the above process for each joint, obtaining a set of equations for each interpolation node. A joint angle vector containing the joint angle values. All interpolation nodes are arranged in timestamp order, with each timestamp corresponding to a joint angle vector, forming a joint angle sequence containing timestamps. Timestamps start from 0 and increase at equal intervals, with the step size equal to the total motion time divided by the number of nodes. The generated sequence of joint angles, including timestamps, is sent by the robotic arm controller to the servo drivers of each joint in time-stamp order, driving the robotic arm to complete the motion trajectory.

[0053] Example 4:

[0054] In practical implementation, the tracked chassis of the inspection robot, driven by a composite state vector, performs global path replanning to generate a set of obstacle avoidance path points. The tracked chassis uses a differential drive method, achieving forward, backward, and turning movements by adjusting the speed of the left and right tracks.

[0055] The gain coefficient in the composite state vector is thresholded to obtain the obstacle density index. The gain coefficient is a double-precision floating-point number with a value range of [0.5, 1.0], and its value is generated by linearly mapping the peak phase output of the pulse-coupled neural network. The gain coefficient is compared with a first preset threshold and a second preset threshold, where the first preset threshold is greater than the second preset threshold. The first preset threshold is set to 0.85, and the second preset threshold is set to 0.65. The basis for setting the first and second preset thresholds is that the gain coefficient reflects the complexity of the current environment and the potential distribution density of obstacles. When the gain coefficient is greater than 0.85, it indicates that the obstacles around the inspection robot are densely distributed and the passage is narrow. When the gain coefficient is between 0.65 and 0.85, it indicates that there are scattered obstacles in the environment but there are passable areas. When the gain coefficient is less than 0.65, it indicates that the environment is open and there are few obstacles.

[0056] If the gain coefficient is greater than the first preset threshold of 0.85, the output is a high density index, assigned a value of 3.0. If the gain coefficient is less than or equal to the first preset threshold of 0.85 and greater than or equal to the second preset threshold of 0.65, the output is a medium density index, assigned a value of 2.0. If the gain coefficient is less than the second preset threshold of 0.65, the output is a low density index, assigned a value of 1.0. The index with assigned values ​​is used as the obstacle density index, which represents the degree of obstacle clustering in the environment.

[0057] The obstacle density index, the current pose of the inspection robot, and a preset grid map are input into a fast expanding random tree algorithm. The obstacle density index is used as an adjustment factor for the expansion step size to generate exploration tree nodes. The preset grid map is a two-dimensional occupancy grid map, where each grid cell stores a probability occupancy value. Grid cells with an occupancy value greater than 0.5 are considered obstacles. The current pose of the inspection robot is provided by the synchronous localization and mapping module, including the x-coordinate, y-coordinate, and orientation angle in the grid map coordinate system. The current pose is used as the starting node of the fast expanding random tree algorithm, and the target node is the endpoint coordinate specified by the task. The expansion step size of the fast expanding random tree algorithm is equal to the base step size multiplied by the reciprocal of the obstacle density index. The base step size is set according to the physical resolution of the grid map. When the grid resolution is 0.05 meters / grid, the base step size is set to 10 grid lengths, i.e., 0.5 meters. The expansion step size is calculated as: Expansion step size = Base step size × (1 / Obstacle density index value). When the obstacle density index is 3.0, the expansion step size is 0.5 × (1 / 3.0) ≈ 0.167 meters; when the obstacle density index is 2.0, the expansion step size is 0.25 meters; and when the obstacle density index is 1.0, the expansion step size is 0.5 meters. In each iteration of the fast expanding random tree algorithm, a sampling point is randomly generated in a uniform distribution within the free space of the grid map. All nodes in the current exploration tree are traversed, and the node with the closest Euclidean distance to the sampling point is identified as the nearest node. The expansion proceeds from the nearest node towards the sampling point, with the expansion distance being the currently set expansion step size, resulting in candidate new nodes. It is checked whether the straight path between the nearest node and the candidate new node intersects with the obstacle area in the grid map. If they do not intersect, the candidate new node is added to the exploration tree, and the nearest node is recorded as the parent node of the candidate new node; if they intersect, the expansion is discarded. Repeat the iteration until there is a node in the exploration tree that is less than the distance to the target node by the expansion step, or until the number of iterations reaches the preset limit, which is set to 5000 times.

[0058] Redundant nodes are pruned in the exploration tree to extract the sequence of nodes along the shortest path from the starting node to the target node, which serves as the obstacle avoidance path set. Starting from the target node, the exploration tree is traversed in reverse order, extracting the parent node of each node. Starting from the target node, the current node is recorded, its parent node is obtained, and added to the parent node linked list. This process is repeated until the current node becomes the starting node. The parent node linked list stores all nodes along a connected path from the target node to the starting node, in reverse order.

[0059] Perform Dijkstra's shortest path search on the parent node's linked list, calculate the Euclidean distance between adjacent nodes in the parent node's linked list, and sum the Euclidean distances to obtain the total path length. If there are multiple connectable nodes near the target node in the exploration tree, multiple candidate paths from the target node to the starting node are obtained. Calculate the total path length of each candidate path, and select the node sequence corresponding to the path branch with the smallest total path length as the node sequence of the shortest path.

[0060] The Douglas-Pock algorithm is applied to simplify the node sequence, removing intermediate nodes whose distance to the straight line segment is less than a preset threshold. The preset simplification threshold is set to 0.3 meters, based on the minimum turning radius and path smoothness requirements of the inspection robot's tracked chassis. The Douglas-Pock algorithm is executed as follows: connect the first and last nodes of the node sequence to form a straight line segment; calculate the vertical distance from all intermediate nodes in the node sequence to this straight line segment; find the maximum vertical distance and its corresponding intermediate node. If the maximum vertical distance is greater than the simplification threshold of 0.3 meters, retain the intermediate node and divide the node sequence into two segments using this intermediate node as the dividing point; recursively perform the above process on each segment. If the maximum vertical distance is less than or equal to the simplification threshold of 0.3 meters, remove all intermediate nodes between the first and last nodes, retaining only the first and last nodes as the simplified path points for that segment. After recursion, the node sequence after removing redundant nodes is obtained; this node sequence is the obstacle avoidance path point set. The obstacle avoidance path point set contains an ordered list of path point coordinates from the starting node to the target node. Each coordinate contains the horizontal and vertical coordinates in the grid map. The tracked chassis controller of the inspection robot performs servo tracking with the path points as local targets in sequence.

[0061] Example 5:

[0062] In specific implementation, please refer to Figure 5The intelligent multi-mode inspection system based on the inspection robot includes a memory and a processor. The memory contains a program for the intelligent multi-mode inspection method based on the inspection robot. When the program is executed by the processor, it implements the following steps: synchronously acquiring point cloud data and multispectral image data of the working environment through a panoramic LiDAR mounted on the inspection robot base; performing voxel downsampling and normal vector estimation on the point cloud data to generate an environmental complexity Fourier descriptor; performing multi-scale Retinex enhancement and frequency domain saliency detection on the multispectral image data to extract the phase shift of the multispectral structured light stripes; inputting the environmental complexity Fourier descriptor and the phase shift into a preset chaotic mapping decision network to adaptively trigger a high-frequency sampling mode or a low-frequency inspection mode, and outputting a composite state vector.

[0063] The memory uses a non-volatile storage medium, specifically a solid-state drive or an embedded multimedia card, with a storage capacity of no less than 64GB. It stores the executable binary file of the intelligent multi-mode inspection method program based on the inspection robot, preset voxel side length configuration parameters, a pre-calculated coefficient table of spherical harmonic function basis, preset values ​​of Logistic mapping branch parameter R, preset values ​​of Henon mapping parameters a and b, connection strength coefficient β of the pulse-coupled neural network, membrane potential decay time constant α_E, dynamic threshold decay time constant α_T, preset values ​​of threshold amplification coefficient V_T and membrane potential amplification coefficient V_E, preset values ​​of duty cycle determination threshold τ, nonlinear mapping table, preset values ​​of the first and second preset thresholds, preset grid map, and basic step size parameters of the fast expanding random tree algorithm.

[0064] The processor employs an embedded central processing unit, specifically a multi-core processor based on ARM or x86 architecture, with a clock speed of no less than 1.5GHz. The processor establishes data communication connections with the memory, panoramic LiDAR, three-axis gimbal drive module, robotic arm joint servo driver, and tracked chassis motor driver via a system bus. After power-on, the processor loads the intelligent multi-mode inspection method program based on the inspection robot from the memory into memory and executes it.

[0065] During operation, the panoramic LiDAR is mounted on the front crossbeam of the inspection robot's base, 0.8 meters above the ground. The panoramic LiDAR comprises a laser emission module, a multispectral imaging module, and a synchronization control module. The laser emission module emits near-infrared pulsed laser light with a wavelength of 905 nanometers. The multispectral imaging module contains three independent spectral channels with center wavelengths of 450 nanometers, 550 nanometers, and 680 nanometers, corresponding to the blue, green, and red spectral bands, respectively. The synchronization control module triggers laser emission and multispectral image acquisition with a time synchronization accuracy of 10 microseconds, ensuring that the timestamps of each frame of point cloud data and each frame of multispectral image data are consistent. The point cloud data is processed at 30 frames per second, with each frame containing approximately 300,000 three-dimensional spatial points. Each point stores spatial coordinates (x, y, z) and a reflection intensity value. The multispectral image data is processed at 30 frames per second, with each frame having a resolution of 1920 pixels × 1080 pixels. Each pixel contains 16-bit grayscale values ​​across three spectral channels. Point cloud data and multispectral image data acquired by the panoramic LiDAR are transmitted in real time to a circular data buffer in the processor's memory via a gigabit Ethernet interface.

[0066] The processor retrieves a frame of point cloud data from the circular data buffer, performs voxel downsampling and normal vector estimation, and generates a Fourier descriptor of environmental complexity. The specific execution method is consistent with the steps described in Example 1.

[0067] The processor retrieves a frame of multispectral image data aligned with the timestamp of the point cloud data from the circular data buffer and performs multi-scale Retinex enhancement. Multi-scale Retinex enhancement employs a Gaussian wrap function at three scales, with standard deviations set to 15 pixels, 80 pixels, and 250 pixels, respectively. The three spectral channels are processed separately, and the illumination component estimate at each scale is calculated. Then, the reflection component estimates at the three scales are weighted and averaged with weighting coefficients of 0.33, 0.33, and 0.34, respectively, to obtain the enhanced multispectral image. Frequency domain saliency detection is performed on the enhanced multispectral image. The average image of the three spectral channels is calculated, and a two-dimensional discrete Fourier transform is performed on the average image to obtain the amplitude spectrum and phase spectrum. The logarithm of the amplitude spectrum is taken to obtain the logarithmic amplitude spectrum. The difference between the local mean of the logarithmic amplitude spectrum and the logarithmic amplitude spectrum is calculated to obtain the spectral residual. A two-dimensional inverse discrete Fourier transform is performed on the spectral residual and the phase spectrum, followed by Gaussian smoothing to obtain the saliency map. The multispectral structured light stripe region is located in the saliency map. The multispectral structured light stripes are projected by the structured light projector mounted on the inspection robot. The stripe pattern is a sinusoidal phase-shifted stripe. The intensity value of each pixel in the stripe region under the three spectral channels is extracted in the saliency map. The phase value of each pixel is calculated using the three-step phase-shifting method. The phase value of all pixels in the stripe region is taken as the arithmetic mean to obtain the phase shift of the multispectral structured light stripe. The phase shift is expressed in radians and the value range is [-π, π].

[0068] The processor inputs the Fourier descriptor of environmental complexity and the phase offset of multispectral structured light fringes into a preset chaotic mapping decision network, adaptively triggering either a high-frequency sampling mode or a low-frequency inspection mode, and outputting a composite state vector. The system also includes a three-axis gimbal, an integrated sensor array, and a tracked chassis. The three-axis gimbal is fixed to the end effector of the inspection robot's arm and includes a yaw axis motor, a roll axis motor, and a pitch axis motor. Each motor is equipped with an absolute photoelectric encoder with an angle detection accuracy of 0.01 degrees. The three-axis gimbal controller communicates with the processor via a CAN bus, receiving target attitude commands and executing closed-loop servo control.

[0069] The integrated sensor array includes a panoramic LiDAR, a visible light camera, and an infrared thermal imager. The visible light camera uses a global shutter CMOS sensor with a resolution of 2592 pixels × 1944 pixels and an adjustable frame rate ranging from 1 frame / second to 60 frames / second. The infrared thermal imager uses an uncooled vanadium oxide focal plane detector with a resolution of 640 pixels × 512 pixels, a thermal sensitivity of 0.05 degrees Celsius, a temperature measurement range of -20 degrees Celsius to 150 degrees Celsius, and an adjustable frame rate ranging from 1 frame / second to 30 frames / second. The panoramic LiDAR, visible light camera, and infrared thermal imager are connected to a field-programmable gate array (FPGA) chip via a unified synchronous trigger interface. The FPGA chip receives the acquisition frequency command from the processor and generates an independent trigger pulse for each sensor.

[0070] The acquisition frequency of the integrated sensor array is dynamically adjusted by the gain coefficient in the composite state vector. The processor extracts the gain coefficient from the composite state vector. The gain coefficient is a double-precision floating-point number with a value range of [0.5, 1.0]. This gain coefficient is mapped to the acquisition frequency using the following relationship: Acquisition Frequency = Minimum Acquisition Frequency + Gain Coefficient × (Maximum Acquisition Frequency - Minimum Acquisition Frequency). In a specific implementation, the minimum acquisition frequency of the panoramic LiDAR is set to 5 frames / second, and the maximum acquisition frequency is set to 30 frames / second; the minimum acquisition frequency of the visible light camera is set to 1 frame / second, and the maximum acquisition frequency is set to 60 frames / second; the minimum acquisition frequency of the infrared thermal imager is set to 1 frame / second, and the maximum acquisition frequency is set to 30 frames / second. When the gain coefficient is 0.5, the panoramic LiDAR acquisition frequency is 17.5 frames / second; when the gain coefficient is 1.0, the panoramic LiDAR acquisition frequency is 30 frames / second. Each time the processor receives a new composite state vector, it updates the acquisition frequency command and writes the updated command into the register of the field-programmable gate array (FPGA) chip via a serial peripheral interface. The FPGA chip then generates a trigger pulse at the new frequency in the next synchronization cycle, enabling dynamic adjustment of the integrated sensor array's acquisition frequency. The tracked chassis is powered by independently driven brushless DC motors on both sides and is equipped with rubber tracks to adapt to various ground conditions. The chassis controller receives speed and steering commands from the processor via an RS485 bus.

[0071] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An intelligent multi-mode inspection method based on an inspection robot, characterized in that, The method includes: The panoramic LiDAR mounted on the base of the inspection robot synchronously acquires point cloud data and multispectral image data of the working environment. The point cloud data is subjected to voxel downsampling and normal vector estimation to generate a Fourier descriptor for environment complexity, specifically: The point cloud data is divided into a three-dimensional mesh according to a preset voxel side length. The average point coordinates and point density within each voxel are extracted to construct a voxelized point cloud matrix. Based on the voxelized point cloud matrix, the eigenvector corresponding to the minimum eigenvalue of the covariance matrix of each point within the voxel is calculated using principal component analysis, and used as a local normal vector to generate a normal vector field. The normal vector field is mapped to a spherical harmonic function basis, and a Fourier transform is performed on the spherical harmonic function basis to obtain a frequency domain coefficient matrix. The frequency domain coefficient matrix is ​​then transformed into logarithmic polar coordinates to obtain the Fourier descriptor of the environmental complexity. Multiscale Retinex enhancement and frequency domain saliency detection are performed on the multispectral image data to extract the phase shift of the multispectral structured light stripes; The Fourier descriptor of environmental complexity and the phase offset are input into a preset chaotic mapping decision network, which adaptively triggers either a high-frequency sampling mode or a low-frequency inspection mode, and outputs a composite state vector, specifically: The environment complexity Fourier descriptor is chaotically iterated using a Logistic mapping to generate a first chaotic sequence; the phase offset is chaotically iterated using a Henon mapping to generate a second chaotic sequence; the first chaotic sequence and the second chaotic sequence are bitwise XORed to obtain a decision chaotic seed; the decision chaotic seed is used as the initial membrane potential of a preset pulse-coupled neural network to drive the pulse-coupled neural network to emit a pulse sequence; the high-frequency sampling mode or low-frequency inspection mode is determined according to the duty cycle of the pulse sequence; and the high-frequency sampling mode or low-frequency inspection mode is gain-modulated based on the peak phase of the pulse sequence to output the composite state vector, which includes a mode identifier bit and a gain coefficient.

2. The intelligent multi-mode inspection method based on an inspection robot as described in claim 1, characterized in that, The method further includes: Based on the composite state vector, the robotic arm joints of the inspection robot are driven to perform motion trajectory planning, generating a joint angle sequence, specifically: Decode the mode identifier bit in the composite state vector to obtain the scanning cone angle and step increment of the high-frequency sampling mode or the low-frequency inspection mode. Using the scanning cone angle, step increment, and gain coefficient as inputs, the Jacobian matrix of the robotic arm link is constructed using the DH parameter method; The Jacobian matrix is ​​subjected to singular value decomposition, and the right singular vector corresponding to the minimum singular value is extracted as the obstacle avoidance direction vector. Cubic spline interpolation is used to interpolate the step increment under the constraint of the obstacle avoidance direction vector to generate a joint angle sequence containing timestamps.

3. The intelligent multi-mode inspection method based on an inspection robot as described in claim 2, characterized in that, The step of decoding the mode identifier bit in the composite state vector to obtain the scanning cone angle and step increment of the high-frequency sampling mode or low-frequency inspection mode is as follows: Extract the binary code of the pattern identifier bit, perform Gray code decoding on the binary code, and obtain the pattern index value; Establish a nonlinear mapping table between the mode index value and the scan cone angle and step increment, and query the scan cone angle and step increment corresponding to the mode index value in the nonlinear mapping table; If the pattern index value is outside the boundary of the nonlinear mapping table, the scan cone angle and step increment are extrapolated using radial basis function interpolation to generate extended scan cone angle and step increment.

4. The intelligent multi-mode inspection method based on an inspection robot as described in claim 3, characterized in that, The step of using cubic spline interpolation to interpolate the step increment under the obstacle avoidance direction vector constraint to generate a joint angle sequence containing timestamps is as follows: Using the step increment as the time variable and the initial and target angles of the robotic arm joints as node values, a system of equations for a cubic spline interpolation function is constructed. The obstacle avoidance direction vector is introduced into the boundary conditions of the equation system of the cubic spline interpolation function, and each component of the obstacle avoidance direction vector is used as a constraint condition for the first derivative of the spline curve. Solve the system of equations for the cubic spline interpolation function after introducing the obstacle avoidance direction vector to obtain the angle value of each interpolation node. Arrange the angle values ​​of each interpolation node in the order of timestamps to generate the joint angle sequence containing timestamps.

5. The intelligent multi-mode inspection method based on an inspection robot as described in claim 1, characterized in that, The method further includes: Based on the composite state vector, the tracked chassis of the inspection robot is driven to perform global path replanning to generate an obstacle avoidance path point set, specifically: Thresholding the gain coefficients in the composite state vector yields the obstacle density index; The obstacle density index, the current pose of the inspection robot, and the preset grid map are input into the fast expanding random tree algorithm. The obstacle density index is used as the adjustment factor for the expansion step size to generate exploration tree nodes. The expansion step size of the fast expanding random tree algorithm is equal to the base step size multiplied by the reciprocal of the obstacle density index. Redundant nodes are pruned from the exploration tree nodes, and the node sequence of the shortest path from the starting node to the target node is extracted as the obstacle avoidance path point set.

6. The intelligent multi-mode inspection method based on an inspection robot as described in claim 5, characterized in that, The step of thresholding the gain coefficients in the composite state vector to obtain the obstacle density index is as follows: The gain coefficient is compared with a first preset threshold and a second preset threshold, wherein the first preset threshold is greater than the second preset threshold; If the gain coefficient is greater than the first preset threshold, the output is a high density index; If the gain coefficient is less than or equal to the first preset threshold and greater than or equal to the second preset threshold, the output is a medium density index; If the gain coefficient is less than the second preset threshold, the output is a low density index; The high density index, the medium density index, or the low density index are used as the obstacle density index.

7. The intelligent multi-mode inspection method based on an inspection robot as described in claim 5, characterized in that, The step of pruning redundant nodes in the exploration tree and extracting the node sequence of the shortest path from the starting node to the target node as the obstacle avoidance path set is as follows: Starting from the target node, traverse the exploration tree nodes in reverse order, extract the parent node of each node, and construct a parent node linked list; Perform Dijkstra's shortest path search on the parent node linked list, calculate the Euclidean distance between adjacent nodes in the parent node linked list, sum the Euclidean distances to obtain the total path length, and select the node sequence corresponding to the path branch with the smallest total path length; The node sequence is simplified by performing the Douglas-Puk algorithm to remove intermediate nodes in the node sequence that are less than a preset threshold from the straight line segment, thereby obtaining the obstacle avoidance path point set.

8. An intelligent multi-mode inspection system based on an inspection robot, characterized in that, The system includes: The system includes a memory and a processor. The memory contains a program for an intelligent multi-mode inspection method based on an inspection robot. When the processor executes the program, the intelligent multi-mode inspection method based on the inspection robot performs the following steps: The panoramic LiDAR mounted on the base of the inspection robot synchronously acquires point cloud data and multispectral image data of the working environment. The point cloud data is subjected to voxel downsampling and normal vector estimation to generate a Fourier descriptor for environment complexity, specifically: The point cloud data is divided into a three-dimensional mesh according to a preset voxel side length. The average point coordinates and point density within each voxel are extracted to construct a voxelized point cloud matrix. Based on the voxelized point cloud matrix, the eigenvector corresponding to the minimum eigenvalue of the covariance matrix of each point within the voxel is calculated using principal component analysis, and used as a local normal vector to generate a normal vector field. The normal vector field is mapped to a spherical harmonic function basis, and a Fourier transform is performed on the spherical harmonic function basis to obtain a frequency domain coefficient matrix. The frequency domain coefficient matrix is ​​then transformed into logarithmic polar coordinates to obtain the Fourier descriptor of the environmental complexity. Multiscale Retinex enhancement and frequency domain saliency detection are performed on the multispectral image data to extract the phase shift of the multispectral structured light stripes; The Fourier descriptor of environmental complexity and the phase offset are input into a preset chaotic mapping decision network, which adaptively triggers either a high-frequency sampling mode or a low-frequency inspection mode, and outputs a composite state vector, specifically: The environment complexity Fourier descriptor is chaotically iterated using a Logistic mapping to generate a first chaotic sequence; the phase offset is chaotically iterated using a Henon mapping to generate a second chaotic sequence; the first chaotic sequence and the second chaotic sequence are bitwise XORed to obtain a decision chaotic seed; the decision chaotic seed is used as the initial membrane potential of a preset pulse-coupled neural network to drive the pulse-coupled neural network to emit a pulse sequence; the high-frequency sampling mode or low-frequency inspection mode is determined according to the duty cycle of the pulse sequence; and the high-frequency sampling mode or low-frequency inspection mode is gain-modulated based on the peak phase of the pulse sequence to output the composite state vector, which includes a mode identifier bit and a gain coefficient. The system also includes a three-axis gimbal, an integrated sensor array, and a tracked chassis. The three-axis gimbal is fixed to the end of the robotic arm of the inspection robot. The integrated sensor array includes a panoramic lidar, a visible light camera, and an infrared thermal imager. The acquisition frequency of the integrated sensor array is dynamically adjusted by the gain coefficient in the composite state vector.

Images