Nuclear emergency robot PSO-depth neural network fusion planning method

Abstract

The application discloses a nuclear emergency robot PSO-depth neural network fusion planning method, a radiation field, a geometric field and a signal field are sampled through an anti-radiation sensing module arranged in the nuclear emergency robot, and a 4D dose map is constructed based on the sampling data; adaptive node insertion is carried out through a low-dose trajectory planning module arranged in the nuclear emergency robot, and mixed optimization is carried out through a PSO-depth neural network, so that an optimal trajectory of the nuclear emergency robot is solved; a nuclear material carrying module arranged in the nuclear emergency robot performs a grabbing-carrying-placing operation according to the obtained optimal trajectory; monitoring data is fed back to a remote terminal through online dose monitoring, and the method constructs a'sensing-calculation-carrying' integrated three-level architecture, so that a sensing-planning-execution closed loop is completed locally on the robot, and the remote terminal is only used for supervision, thereby reducing communication dependence.

Description

Technical Field

[0001] This invention relates to the field of nuclear accident rescue and handling technology, and in particular to a nuclear emergency robot PSO-deep neural network fusion planning method. Background Technology

[0002] Currently, nuclear accident rescue robots (NERR) have undergone three generations of development: the first generation only had internal maintenance functions and no radiation protection; the second generation added lead shielding, but was large and had limited functionality; the third generation achieved miniaturization and intelligence, but still generally uses the "shortest path" or "shortest time" criteria (RRT, A, spline) for trajectory planning, without considering "cumulative dose" as an optimization variable, resulting in the robot receiving a dose of >1.2 Sv in a single mission in an environment >100 Gy / h. Furthermore, traditional CMOS cameras experience gamma fog saturation at 10 Gy / h, with a recognition success rate of <65%.

[0003] Therefore, the existing technical solutions have the following problems: ① Dose-trajectory disconnect: Existing plans only focus on geometry or time, without considering radiation costs, resulting in excessive dose load on the robot; ② Computational complexity mismatch: High-order splines or RRT* require more than 3,000 iterations in 6-dimensional joint space to converge, while the kernel emergency task requires replanning to be completed in less than 200 ms; ③ Lack of radiation perception: Gamma fog saturation causes visual failure, and GM tubes can only provide a point dose and cannot reconstruct the three-dimensional distribution.

[0004] In view of this, the present invention is hereby proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a nuclear emergency robot PSO-deep neural network fusion planning method to solve the aforementioned technical problems in the prior art. The method of this invention constructs an integrated three-level architecture of "sensing-computing-moving," enabling the perception-planning-execution closed loop to be completed locally on the robot, with remote terminals only used for supervision, thus reducing communication dependence.

[0006] The objective of this invention is achieved through the following technical solution: A nuclear emergency response robot PSO-deep neural network fusion planning method, the method comprising: Step 1: Sample the radiation field, geometric field, and signal field using the radiation-resistant sensing module installed in the nuclear emergency robot, and construct a 4D dose map based on the sampled data; Step 2: Adaptive node insertion is performed through the low-dose trajectory planning module set in the nuclear emergency robot, and hybrid optimization is performed through PSO-deep neural network to obtain the optimal trajectory of the nuclear emergency robot; Step 3: The nuclear material handling module in the nuclear emergency robot performs the grasping-handling-placement operation according to the optimal trajectory obtained in Step 2; Step 4: Transmit the monitoring data back to the remote terminal via online dose monitoring.

[0007] Compared with existing technologies, the method provided by this invention constructs an integrated three-level architecture of "sensing-computing-moving", which enables the perception-planning-execution closed loop to be completed locally on the robot, while the remote terminal is only used for supervision, reducing communication dependence and meeting the needs of multi-task collaboration. Attached Figure Description

[0008] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A schematic diagram of the nuclear emergency robot PSO-deep neural network fusion planning method provided in an embodiment of the present invention. Detailed Implementation

[0010] 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 a part of the embodiments of the present invention, and not all of them, and do not constitute a limitation on the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0011] First, the following explanations are provided for the terms that may be used in this article: The term "and / or" means that either or both can be achieved simultaneously. For example, X and / or Y means that it includes both "X" or "Y" as well as the three cases of "X and Y".

[0012] The terms "comprising," "including," "containing," "having," or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.) should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.

[0013] The term "composed of" excludes any technical features not expressly listed. When used in a claim, it closes the claim to exclude all technical features other than those expressly listed, except for associated conventional impurities. If the term appears only in a clause of a claim, it limits the claim to the elements expressly listed in that clause; elements recited in other clauses are not excluded from the overall claim.

[0014] The technical solution provided by this invention will be described in detail below. Contents not described in detail in the embodiments of this invention are prior art known to those skilled in the art. Where specific conditions are not specified in the embodiments of this invention, they shall be performed according to conventional conditions in the art or conditions recommended by the manufacturer. Reagents or instruments used in the embodiments of this invention whose manufacturers are not specified are all conventional products that can be purchased commercially.

[0015] like Figure 1 The diagram shown is a schematic flowchart of the nuclear emergency robot PSO-deep neural network fusion planning method provided in an embodiment of the present invention. The method includes: Step 1: Sample the radiation field, geometric field, and signal field using the radiation-resistant sensing module installed in the nuclear emergency robot, and construct a 4D dose map based on the sampled data; In this step, by inputting the acquired four-channel conditional image and spatiotemporal coordinates, the transient neural graphics processor outputs three branches: adult density, dose rate, and uncertainty, based on a sparse 4D voxel mesh. Specifically: First, based on the structure of nuclear facilities and the location of emergency base stations, a communication coverage prediction map is generated offline using a ray tracing algorithm, and the boundaries of blind spots are marked. The measured RSSI value of the current location is output in real time through Quectel RM500Q-GL industrial-grade 5G module. The measured value is compared with the predicted value and then the local field strength distribution is corrected by Kalman filtering. When the measured value deviates from the prediction by more than 10 dB or enters an unmodeled area, a local map recalculation is performed to ensure that the communication field strength map and the dose rate map remain spatiotemporally aligned. The multimodal fusion unit employs a "radio-visual fusion neural field" to reconstruct a four-dimensional dose-geometry joint field in scenarios with γ-fog visibility <0.1 m, using the γ-camera pixel-dose rate map Iγ and the neutron tube count density map In to construct a radiation condition image. Geometric occupancy dual-channel map Igeo was obtained by voxelizing and projecting the Pradar point cloud from a millimeter-wave radar. x Igeo y The four elements are stitched together to form a conditional image. ;in, {H×W} represents the set of real numbers; H represents the image height, i.e. the number of pixel rows in the vertical direction, and W represents the image width, i.e. the number of pixel columns in the horizontal direction. The spatial coordinate x is encoded using multi-resolution hashing as follows: ; Among them, e space Multi-resolution hash coding represents spatial coordinates, mapping spatial coordinate x to a high-dimensional differentiable feature vector; H(·) represents the multi-resolution hash coding function; x is the spatial coordinate; L, T, F, N min , N max These are encoded hyperparameters used to control the hierarchy, capacity, feature dimensions, and resolution range of multi-resolution grids. The output spatial dimension is represented as L layers × F-dimensional features = L·F-dimensional real vector; The time coordinate t, sinusoidally encoded, is as follows: ; Among them, e time γ(t) represents the time-encoded feature vector, which serves as one of the inputs to the neural network; γ(t) represents the encoding function, which takes time t as input and outputs a high-dimensional vector; t represents the current time, which is the time taken from the start of the task. Represents a real number vector space with dimension 2K (K in terms of sine and cosine). Conditional Image I cond Mapped by the convolutional encoder as: ; Among them, e image Represents the image encoding feature vector, which is the four-channel conditional image I. cond Compressed into a compact high-level semantic representation; D represents img A 3D real vector space, with typical values ​​D img = 64; e space e time e image The three are concatenated to form the joint input z = C(e space , e time , e image After joint features are extracted by a shared multilayer perceptron h = Fshared(z), the system is decoupled into three branches: Volume density branch: σ geo (x, t) = σ sigmoid (W geo ^T h + b geoThe output space occupancy probability is used for collision detection; Where, σ geo The volume density branch output represents the geometric occupancy probability at a spatial point (x, y, z), used for collision detection; σ sigmoid () denotes the Sigmoid activation function, used to map any real number to the interval (0,1); W geo ^T h +b geo denoted by linear transformation, used to map shared features h to scalar logits; T represents the capacity of each layer's hash table, i.e., the maximum number of learnable feature vectors that layer can store; h represents the output of the shared multilayer perceptron, which is the joint feature representation after fusing spatial, temporal, and image information; b geo The bias term is an additive constant in the linear transformation that allows the activation function input to be shifted. Dose rate branch: D(x, t) = σ softplus (W rad ^T h + b rad The real-time dose rate is output for path dose integration. Where D(x, t) is the dose rate branch output, representing the instantaneous radiation dose rate of the spatial point (x,y,z) at time t, used for path dose integration; σ softplus () represents the Softplus activation function, which maps any real number to (0, +∞) and ensures that the output is strictly positive; Uncertainty branch: ε(x, t) = exp(W unc T h + b unc The output estimated variance is used as a calculation by the Kriging solver; Where ε(x, t) is the uncertainty branch output, representing the variance of the dose rate estimate, used for the prior variance of the Kriging solver. ; exp(·) represents the exponential activation function, used to map any real number to (0, +∞), ensuring that the variance is strictly positive; W unc T h + b unc This represents a linear transformation used to map shared features h to scalar logits; The resulting 4D dose map contains the following data layers: (1) Core dose layer: including gamma dose rate, neutron dose rate, total equivalent dose, spatial resolution 10 cm³, temporal resolution 1 s; (2) Energy Spectrum Information Layer: Includes 1024 channels of γ energy spectrum for nuclide identification, and thermal neutron / fast neutron ratio for neutron energy spectrum hardening correction; (3) Metadata layer: including GPS timestamps, measurement uncertainties, and radiation source ID tags; (4) Dynamic feature layer: including dose rate time derivative, used for leak detection; spatial gradient, used for hotspot location; All data are uniformly encoded using Instant-NGP sparse 4D voxel meshes, with a 30-second sliding window to maintain temporal consistency. When output to the trajectory planning module, the data is downsampled to a 32×32 raster format.

[0016] Step 2: Adaptive node insertion is performed through the low-dose trajectory planning module set in the nuclear emergency robot, and hybrid optimization is performed through PSO-deep neural network to obtain the optimal trajectory of the nuclear emergency robot; In this step, adaptive node insertion is performed using the low-dose trajectory planning module set up in the nuclear emergency robot. The specific process is as follows: Setting up the nuclear emergency robot status Dosage field ; Where x is the state vector; q is the generalized position; q̇ is the generalized velocity; and the dimension is 12. For dose field mapping, D is the dose field function symbol, representing the mapping relationship from spatial location to dose rate. Let the domain and range of the function be defined. Let be the three-dimensional Euclidean space, which is the domain of the function, representing all possible spatial locations. The set of non-negative real numbers, i.e., the range of the function, indicates that the dose rate can only be positive or zero. The symbol represents a mapping relationship between the independent variable on the left and the function value on the right. The trajectory is parameterized into N B-splines, represented as: ; in, The basis function is a cubic B-spline, which ensures the second-order continuity of the trajectory (no abrupt changes in velocity and acceleration) and meets the requirement of "smooth start-stop" in nuclear emergency response. is the coordinate vector of the control points; u is the normalized parameter of the B-spline curve; p(u) is the path point; The entire trajectory does not pass through any control points, naturally avoiding joint singularities, making it suitable for operation in narrow passages.

[0017] Stack the 16 points in xyz order into a vector, and use it as a unified optimization variable. The joint optimization objective function of the three objectives is expressed as: ; The first term in the formula is a penalty for long paths, used to reduce joint wear and energy consumption; The second term integrates the real-time dose rate D[p(u)] along the path to directly quantify the cumulative dose received by the nuclear emergency robot; The third term, Comm(p(u)), is a predicted 5G signal strength value, used to prevent robots from losing contact when entering communication dead zones. Its expression is: ; Here, RSSI(p(u)) is the received signal strength indicator, indicating the 5G signal strength at path point p(u); -80 is the communication blind zone determination threshold, in dBm. When RSSI < -80 dBm, it is determined to be a weak communication zone, triggering a penalty; when RSSI ≥ -80 dBm, the penalty is 0; max(0, ·) is the truncation function, used to penalize weak signal areas, while the penalty for strong signal areas is 0, to avoid negative penalties causing the optimizer to be biased towards extremes; In the fourth item, T is the trajectory execution time; α, β, γ, and δ are dimensionless normalization coefficients. The coefficient β is calculated from the task dose limit, and the coefficient δ is dynamically adjusted according to the remaining battery capacity. The criteria for adaptive node insertion are: When any of the following conditions are met Insert a new node under certain conditions; Where ΔD is the dose rate difference between two adjacent path segments; when it is >0.1 mSv / h, it indicates a steep dose gradient, requiring denser control points to precisely avoid radiation source hotspots; κ is the curvature. When RSSI < -80 dBm, it indicates that the robot arm is about to enter a weak communication zone, and encrypting nodes in advance can reserve more degrees of freedom for subsequent network outage replanning. Whenever the adaptive criterion triggers node insertion, the B-spline order remains unchanged. Only the newly inserted node u_new is written into the node vector, and the current number of effective control points n′=n+1 is recalculated. The calculated effective control points Pi are stacked in xyz order to obtain a new optimization vector. The new optimization vector P is the core optimization variable in trajectory planning, 3n′ is the dimension of the optimization vector P, and n′ is the number of effective control points after the adaptive node insertion; The old particle swarm is then discarded, and particles are resampled with a new optimization vector P as the mean and a variance of 0.5σ, and the iteration continues.

[0018] This process ensures that the PSO dimension and the number of B-spline nodes are consistent in real time, preventing array out-of-bounds errors.

[0019] The above-mentioned hybrid optimization using PSO and deep neural networks yields the optimal trajectory of the nuclear emergency robot, specifically: A one-dimensional covariance supernetwork is used. First, two layers of one-dimensional convolution are used to extract the local spatial features of the 4D dose map to obtain a 64-dimensional latent vector z. Then, the coordinates of 16 B-spline control points P and the low-rank covariance factor L are output by three linear supernetwork branches respectively. The total number of covariance parameters is 320. An analytical kriging solver is used for millisecond-level uncertainty correction without iteration. The input interface is a dual-channel image: channel 0 is the dose rate map. Channel 1 is a communication field strength diagram. Sparse sampling ≤ 64 points, node features include ΔD, Δ²D, and ΔRSSI; ΔD represents the first gradient of the dose rate, used to describe the rate and direction of change of the dose field in space; Δ²D represents the second gradient of the dose rate, describing the curvature and local extremum characteristics of the dose field, used to identify radiation source hotspots and dose field inflection points; ΔRSSI represents the spatial rate of change of the received signal strength, describing the attenuation or enhancement trend of the communication field in the path direction, used to predict the boundary of the communication blind zone; The dual-channel image is first unfolded row by row into a one-dimensional signal of length 32, forming a [B, 2, 32] tensor; then it is gradually downsampled to 1 dimension through two layers of one-dimensional convolution to obtain a 64-dimensional latent vector z. The three linear hypernetwork branches include: ① P branch: Linear(64→48), indicating a fully connected layer (linear transformation layer), with an input dimension of 64 and an output dimension of 48, and no activation function; it directly outputs the coordinates of 16 control points; This layer is the P branch of the one-dimensional covariance supernetwork, responsible for mapping the 64-dimensional latent vector z to the coordinate vector of the 48-dimensional B-spline control points; ② Branch: Linear(64→256), outputs and reshapes into a low-rank factor of [64, 4]. ; ③ Branch: Linear(64→64), diagonalized after smoothing with Softplus rectified linear units to ensure positive definiteness; Final low-rank covariance factor matrix L: This low-rank covariance factor matrix L is used to construct the covariance matrix Σ in the Kriging solver. in, As a diagonal factor, by The branch output is obtained by diagonalizing it after passing through a smooth rectified linear unit Softplus; The matrix has dimensions of 64 rows and 8 columns, where 64 corresponds to ≤64 sparse sampling points, and 8 is the rank of the low-rank decomposition; [·, ·] represents the matrix concatenation operation. Finally, the Kriging solver was analyzed: Directly provide the closed-form solution ; in, The output of the network is the Cholesky decomposition form of the covariance matrix, which is constructed as a symmetric positive definite covariance matrix through a low-rank factor matrix L; L is a low-rank factor matrix. Let L be the transpose of L; Let Σ be the cross-covariance vector between the point to be estimated and the known points; since Σ is positive definite and has a fixed size, the inverse can be completed analytically in one step; substituting this into the posterior mean expression... ; Among them, P ref The reference control point is the B-spline control point output from the coarse-scale optimization stage of PSO (Particle Swarm Optimization), serving as the reference true value for Kriging correction; λ is the weight vector of the Kriging solver, and each component of λ... The correction weights of the i-th sampling point to the estimated point are quantified. A large absolute value indicates a strong influence of the sampling point (closer distance, larger covariance). This indicates that the impact is negligible (due to large distance or small covariance). in, And the corresponding 2σ confidence interval, without the need for iteration or numerical optimization; The estimated value of μ after Kriging correction is still denoted as μ for use by the trajectory generator. The entire network outputs "path point + uncertainty" in one forward pass, without any other hidden layers or additional operations. In the specific implementation, during the inference phase, the one-dimensional covariance hypernetwork is quantized on Jetson AGX Orin using TensorRT-INT8; if a small number of observations need to be updated online subsequently, the Woodbury identity can be used for fast recalculation. ; When the analytical Kriging solver finds that the maximum 2σ root of the submatrix at each 3×3 position of the network output Σ is greater than 5cm, it immediately reverts to the PSO global research phase to ensure that the nuclear emergency robot always operates in the "high confidence" corridor.

[0020] Step 3: The nuclear material handling module in the nuclear emergency robot performs the grasping-handling-placement operation according to the optimal trajectory obtained in Step 2; In this step, before performing the grabbing operation, the nuclear emergency robot first undergoes relevant verifications, including: Trajectory verification: The optimal path point p(u) is fed into the inverse kinematics solver to verify whether the joint angle q(u) = IK(p(u)) of each path point is within the mechanical limit [q^{min}, q^{max}]. Dynamic verification: Calculate the joint moments at each path point Verify whether the maximum torque of the servo motor is exceeded. ; Dose prediction: Estimated cumulative dose is calculated by integrating along the trajectory. Confirm the estimated cumulative dose Sv is the single-task dose limit; Then, a communication handshake is performed, and a task start signal is sent to the remote terminal through the RM500Q module to confirm that RSSI>-80 dBm and establish a two-way communication link. During the grasping phase of the nuclear emergency robot, visual localization is performed first. In a gamma fog environment, millimeter-wave radar fills in the visual failure area and outputs centimeter-level point clouds. Then, a 4D dose-geometry joint field is reconstructed through multimodal fusion neural field to identify the target position of the nuclear material. End-effector alignment is performed, and the expected pose T_{grasp} of the end effector is calculated based on the target position. Finally, the corresponding joint angle q_{grasp} = IK(T_{grasp}) is solved through inverse kinematics to generate the grasping trajectory p_{grasp}(u). During the grasping phase of the nuclear emergency robot, the end effector (lead-boron polyethylene shielded gripper) of the nuclear emergency robot closes, and the gripping force is adaptively adjusted according to the weight of the nuclear material; the gripping stability signal is fed back by the grasping confirmation sensor, triggering the handling phase; During the handling phase, the joints of the nuclear emergency robot's robotic arm move along a B-spline trajectory q(u), with u advancing at a constant speed from 0 to 1; the PID controller adjusts the joint torque in real time to track errors.

[0021] Step 4: Transmit the monitoring data back to the remote terminal via online dose monitoring.

[0022] In step 4, the method further includes a communication interruption fallback mechanism, specifically: A radiation control emergency skill map is set up locally; the nodes of the radiation control emergency skill map are basic skills, including obstacle avoidance, grabbing, and placement operations; the edges are the conditions for skill transfer. When the received signal strength indicator RSSI < -95dBm and lasts for 3 seconds, a local map search is triggered. The cumulative dose increment ΔE within 30 seconds is estimated along the backtracking path to ensure safe withdrawal to the shielded area without remote command within 30 seconds. The dose field stored in the nuclear emergency robot is considered a "high dose unknown area" after 90 seconds from the last successful cloud update. The dose weight of all edges in the radiation control emergency skill map is forcibly set to +∞. At this time, the nuclear emergency robot immediately stops and waits for communication to be restored. If it is not restored after 120 seconds, an emergency power outage shutdown is triggered.

[0023] It is worth noting that the contents not described in detail in the embodiments of the present invention belong to the prior art known to those skilled in the art.

[0024] In summary, the method described in this embodiment of the invention can be extended to scenarios such as nuclear fuel reprocessing, spent fuel transportation, and high-level radioactive waste leakage, meeting the needs of multi-task collaboration. Since the number of parameters in the one-dimensional covariance hypernetwork is relatively small, this method is also applicable to MCU-level edge computing nodes and can be extended to limited hardware scenarios such as onboard dose planning for spent fuel transportation vehicles in the future.

[0025] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A nuclear emergency robot PSO-deep neural network fusion planning method, characterized in that, The method includes: Step 1: Sample the radiation field, geometric field, and signal field using the radiation-resistant sensing module installed in the nuclear emergency robot, and construct a 4D dose map based on the sampled data; Step 2: Adaptive node insertion is performed through the low-dose trajectory planning module set in the nuclear emergency robot, and hybrid optimization is performed through PSO-deep neural network to solve for the optimal trajectory of the nuclear emergency robot; Step 3: The nuclear material handling module in the nuclear emergency robot performs the grasping-handling-placement operation according to the optimal trajectory obtained in Step 2; Step 4: Transmit the monitoring data back to the remote terminal via online dose monitoring.

2. The nuclear emergency robot PSO-deep neural network fusion planning method according to claim 1, characterized in that, In step 1, by inputting the acquired four-channel conditional image and spatiotemporal coordinates, the transient neural graphics processor outputs three branches: adult density, dose rate, and uncertainty, based on a sparse 4D voxel mesh. Specifically: First, based on the structure of nuclear facilities and the location of emergency base stations, a communication coverage prediction map is generated offline using a ray tracing algorithm, and the boundaries of blind spots are marked. The measured RSSI value of the current location is output in real time through Quectel RM500Q-GL industrial-grade 5G module. The measured value is compared with the predicted value and then the local field strength distribution is corrected by Kalman filtering. When the measured value deviates from the prediction by more than 10 dB or enters an unmodeled area, a local map recalculation is performed to ensure that the communication field strength map and the dose rate map remain spatiotemporally aligned. The multimodal fusion unit employs a "radiovisual fusion neural field" to reconstruct a four-dimensional dose-geometry joint field in scenarios with γ-fog visibility < 0.1 m. The radiation condition image is constructed using the γ-camera pixel-dose rate map Iγ and the neutron tube count density map In. Geometric occupancy dual-channel map Igeo was obtained by voxelizing and projecting the Pradar point cloud from a millimeter-wave radar. x Igeo y The four elements are stitched together to form a conditional image. ;in, {H×W} represents the set of real numbers; H represents the image height, i.e. the number of pixel rows in the vertical direction, and W represents the image width, i.e. the number of pixel columns in the horizontal direction. The spatial coordinate x is encoded using multi-resolution hashing as follows: ； Among them, e space Multi-resolution hash coding represents spatial coordinates, mapping spatial coordinate x to a high-dimensional differentiable feature vector; H(·) represents the multi-resolution hash coding function; x is the spatial coordinate; L, T, F, N min , N max These are encoded hyperparameters used to control the hierarchy, capacity, feature dimensions, and resolution range of multi-resolution grids. The output space dimension is represented as L layers × F-dimensional features = L·F-dimensional real vector; The time coordinate t, sinusoidally encoded, is as follows: ； Among them, e time γ(t) represents the time-encoded feature vector, which serves as one of the inputs to the neural network; γ(t) represents the encoding function, which takes time t as input and outputs a high-dimensional vector; t represents the current time, which is the time taken from the start of the task. Represents a real number vector space with a dimension of 2K; Conditional Image I cond Mapped by the convolutional encoder as: ； Among them, e image Represents the image encoding feature vector, which is the four-channel conditional image I. cond Compressed into a compact high-level semantic representation; D represents img A 3D real vector space, with typical values ​​D img = 64; e space e time e image The three are concatenated to form the joint input z = C(e space , e time , e image After joint features are extracted by a shared multilayer perceptron h = Fshared(z), the system is decoupled into three branches: Volume density branch: σ geo (x, t) = σ sigmoid (W geo ^T h + b geo The output space occupancy probability is used for collision detection; Where, σ geo The volume density branch output represents the geometric occupancy probability at a spatial point (x, y, z), used for collision detection; σ sigmoid () denotes the Sigmoid activation function, used to map any real number to the interval (0,1); W geo ^T h + b geo denoted by linear transformation, used to map shared features h to scalar logits; T represents the capacity of each layer's hash table, i.e., the maximum number of learnable feature vectors that layer can store; h represents the output of the shared multilayer perceptron, which is the joint feature representation after fusing spatial, temporal, and image information; b geo The bias term is an additive constant in the linear transformation that allows the activation function input to be shifted. Dose rate branch: D(x, t) = σ softplus (W rad ^T h + b rad The real-time dose rate is output for path dose integration. Where D(x, t) is the dose rate branch output, representing the instantaneous radiation dose rate of the spatial point (x,y,z) at time t, used for path dose integration; σ softplus () represents the Softplus activation function, which maps any real number to (0, +∞) and ensures that the output is strictly positive; Uncertainty branch: ε(x, t) = exp(W unc T h + b unc The output estimated variance is used as a calculation by the Kriging solver; Where ε(x, t) is the uncertainty branch output, representing the variance of the dose rate estimate, used for the prior variance of the Kriging solver. ; exp(·) represents the exponential activation function, used to map any real number to (0, +∞), ensuring that the variance is strictly positive; W unc T h + b unc This represents a linear transformation used to map shared features h to scalar logits; The resulting 4D dose map contains the following data layers: (1) Core dose layer: including gamma dose rate, neutron dose rate, total equivalent dose, spatial resolution 10 cm³, temporal resolution 1 s; (2) Energy Spectrum Information Layer: Includes 1024 γ energy spectrum channels for nuclide identification, and thermal neutron / fast neutron ratio for neutron energy spectrum hardening correction; (3) Metadata layer: including GPS timestamps, measurement uncertainties, and radiation source ID tags; (4) Dynamic feature layer: including dose rate time derivative, used for leak detection; spatial gradient, used for hotspot location; All data are uniformly encoded using Instant-NGP sparse 4D voxel meshes, with a 30-second sliding window to maintain temporal consistency. When output to the trajectory planning module, the data is downsampled to a 32×32 raster format.

3. The nuclear emergency robot PSO-deep neural network fusion planning method according to claim 1, characterized in that, In step 2, adaptive node insertion is performed using the low-dose trajectory planning module set up in the nuclear emergency robot. The specific process is as follows: Setting up the nuclear emergency robot status Dosage field ; Where x is the state vector; q is the generalized position; q̇ is the generalized velocity; and the dimension is 12. For dose field mapping, D is the dose field function symbol, representing the mapping relationship from spatial location to dose rate. Let the domain and range of the function be defined. Let be the three-dimensional Euclidean space, which is the domain of the function, representing all possible spatial locations. The set of non-negative real numbers, i.e., the range of the function, indicates that the dose rate can only be positive or zero. The symbol represents a mapping relationship between the independent variable on the left and the function value on the right. The trajectory is parameterized into N B-splines, represented as: ； in, The basis function is a cubic B-spline, which ensures the second-order continuity of the trajectory and meets the requirements of "smooth start-stop" in nuclear emergency response. is the coordinate vector of the control points; u is the normalized parameter of the B-spline curve; p(u) is the path point; Stack the 16 points in xyz order into a vector, and use it as a unified optimization variable. The joint optimization objective function of the three objectives is expressed as: ； The first term in the formula is a penalty for long paths, used to reduce joint wear and energy consumption; The second term integrates the real-time dose rate D[p(u)] along the path to directly quantify the cumulative dose received by the nuclear emergency robot; The third term, Comm(p(u)), is a predicted 5G signal strength value, used to prevent robots from losing contact when entering communication dead zones. Its expression is: ； Here, RSSI(p(u)) is the received signal strength indicator, indicating the 5G signal strength at path point p(u); -80 is the communication blind zone determination threshold, in dBm. When RSSI < -80 dBm, it is determined to be a weak communication zone, triggering a penalty; when RSSI ≥ -80 dBm, the penalty is 0; max(0, ·) is the truncation function, used to penalize weak signal areas, while the penalty for strong signal areas is 0, to avoid negative penalties causing the optimizer to be biased towards extremes; In the fourth item, T is the trajectory execution time; α, β, γ, and δ are dimensionless normalization coefficients. The coefficient β is calculated from the task dose limit, and the coefficient δ is dynamically adjusted according to the remaining battery capacity. The criteria for adaptive node insertion are: When any of the following conditions are met Insert a new node under certain conditions; Where ΔD is the dose rate difference between two adjacent path segments; when it is >0.1 mSv / h, it indicates a steep dose gradient, requiring denser control points to precisely avoid radiation source hotspots; κ is the curvature. When RSSI < -80 dBm, it indicates that the robot arm is about to enter a weak communication zone, and encrypting nodes in advance can reserve more degrees of freedom for subsequent network outage replanning. Whenever the adaptive criterion triggers node insertion, the B-spline order remains unchanged. Only the newly inserted node u_new is written into the node vector, and the current number of effective control points n′=n+1 is recalculated. The calculated effective control points Pi are stacked in xyz order to obtain a new optimization vector. The new optimization vector P is the core optimization variable in trajectory planning, 3n′ is the dimension of the optimization vector P, and n′ is the number of effective control points after the adaptive node insertion; The old particle swarm is then discarded, and particles are resampled with a new optimization vector P as the mean and a variance of 0.5σ, and the iteration continues.

4. The nuclear emergency robot PSO-deep neural network fusion planning method according to claim 3, characterized in that, In step 2, the optimal trajectory of the nuclear emergency robot is obtained through hybrid optimization using PSO-deep neural networks, specifically: A one-dimensional covariance supernetwork is used. First, two layers of one-dimensional convolution are used to extract the local spatial features of the 4D dose map to obtain a 64-dimensional latent vector z. Then, the coordinates of 16 B-spline control points P and the low-rank covariance factor L are output by three linear supernetwork branches respectively. The total number of covariance parameters is 320. An analytical kriging solver is used for millisecond-level uncertainty correction without iteration. The input interface is a dual-channel image: channel 0 is the dose rate map. Channel 1 is a communication field strength diagram. Sparse sampling ≤ 64 points, node features include ΔD, Δ²D, and ΔRSSI; ΔD represents the first gradient of the dose rate, used to describe the rate and direction of change of the dose field in space; Δ²D represents the second gradient of the dose rate, describing the curvature and local extremum characteristics of the dose field, used to identify radiation source hotspots and dose field inflection points; ΔRSSI represents the spatial rate of change of the received signal strength, describing the attenuation or enhancement trend of the communication field in the path direction, used to predict the boundary of the communication blind zone; The dual-channel image is first unfolded row by row into a one-dimensional signal of length 32, forming a [B, 2, 32] tensor; then it is gradually downsampled to 1 dimension through two layers of one-dimensional convolution to obtain a 64-dimensional latent vector z. The three linear hypernetwork branches include: ① P branch: Linear(64→48), indicating a fully connected layer with an input dimension of 64 and an output dimension of 48, without an activation function; directly outputs the coordinates of 16 control points; This layer is the P branch of the one-dimensional covariance supernetwork, responsible for mapping the 64-dimensional latent vector z to the coordinate vector of the 48-dimensional B-spline control points; ② Branch: Linear(64→256), outputs and reshapes into a low-rank factor of [64, 4]. ; ③ Branch: Linear(64→64), diagonalized after smoothing with Softplus rectified linear units to ensure positive definiteness; Final low-rank covariance factor matrix L: This low-rank covariance factor matrix L is used to construct the covariance matrix Σ in the Kriging solver. in, As a diagonal factor, by The branch output is obtained by diagonalizing it after passing through a smooth rectified linear unit Softplus; The matrix has dimensions of 64 rows and 8 columns, where 64 corresponds to ≤64 sparse sampling points, and 8 is the rank of the low-rank decomposition; [·, ·] represents the matrix concatenation operation. Finally, the Kriging solver was analyzed: Directly provide the closed-form solution ; in, The output of the network is the Cholesky decomposition form of the covariance matrix, which is constructed as a symmetric positive definite covariance matrix through a low-rank factor matrix L; L is a low-rank factor matrix. Let L be the transpose of L; Let Σ be the cross-covariance vector between the point to be estimated and the known points; since Σ is positive definite and has a fixed size, the inverse can be completed analytically in one step; substituting this into the posterior mean expression... ; λ is the weight vector of the Kriging solver, and each component of λ The correction weights of the i-th sampling point to the estimated point are quantified; in, And the corresponding 2σ confidence interval, without the need for iteration or numerical optimization; The Kriging-corrected estimate of μ is still denoted as μ for use by the trajectory generator; When the analytical Kriging solver finds that the maximum 2σ root of the submatrix at each 3×3 position of the network output Σ is greater than 5cm, it immediately reverts to the PSO global research phase to ensure that the nuclear emergency robot always operates in the "high confidence" corridor.

5. The nuclear emergency robot PSO-deep neural network fusion planning method according to claim 3, characterized in that, In step 3, before performing the grabbing operation, the nuclear emergency robot first conducts relevant verifications, including: Trajectory verification: The optimal path point p(u) is fed into the inverse kinematics solver to verify whether the joint angle q(u) = IK(p(u)) of each path point is within the mechanical limit [q^{min}, q^{max}]. Dynamic verification: Calculate the joint moments at each path point Verify whether the maximum torque of the servo motor is exceeded. ; Dose prediction: Estimated cumulative dose is calculated by integrating along the trajectory. Confirm the estimated cumulative dose Sv is the single-task dose limit; Then, a communication handshake is performed, and a task start signal is sent to the remote terminal through the RM500Q module to confirm RSSI > -80 dBm and establish a two-way communication link. During the grasping phase of the nuclear emergency robot, visual localization is performed first. In a gamma fog environment, millimeter-wave radar fills in the visual failure area and outputs centimeter-level point clouds. Then, a 4D dose-geometry joint field is reconstructed through multimodal fusion neural field to identify the target position of the nuclear material. End-effector alignment is performed, and the expected pose T_{grasp} of the end effector is calculated based on the target position. Finally, the corresponding joint angle q_{grasp} = IK(T_{grasp}) is solved through inverse kinematics to generate the grasping trajectory p_{grasp}(u). During the grasping phase of the nuclear emergency robot, the end effector of the nuclear emergency robot closes, and the clamping force is adaptively adjusted according to the weight of the nuclear material; the grasping confirmation sensor feeds back a clamping stability signal, triggering the handling phase; During the handling phase, the joints of the nuclear emergency robot's robotic arm move along a B-spline trajectory q(u), with u moving at a constant speed from 0 to 1; the PID controller adjusts the joint torque in real time to track errors.

6. The nuclear emergency robot PSO-deep neural network fusion planning method according to claim 3, characterized in that, In step 4, the method further includes a communication interruption fallback mechanism, specifically: A radiation control emergency skill map is set up locally; the nodes of the radiation control emergency skill map are basic skills, including obstacle avoidance, grabbing, and placement operations; the edges are the conditions for skill transfer. When the received signal strength indicator RSSI < -95dBm and lasts for 3 seconds, a local map search is triggered. The cumulative dose increment ΔE within 30 seconds is estimated along the backtracking path to ensure safe withdrawal to the shielded area without remote command within 30 seconds. The dose field stored in the nuclear emergency robot is considered a "high dose unknown area" after 90 seconds from the last successful cloud update. The dose weight of all edges in the radiation control emergency skill map is forcibly set to +∞. At this time, the nuclear emergency robot immediately stops and waits for communication to be restored. If it is not restored after 120 seconds, an emergency power outage shutdown is triggered.

Images