Unmanned aerial vehicle electric power inspection safety supervision and emergency response system

Abstract

The invention relates to the technical field of unmanned aerial vehicle electric power inspection and safety monitoring, and discloses an unmanned aerial vehicle electric power inspection safety supervision and emergency response system which comprises a multi-mode space-time fusion module, a layered anti-interference control module, a federal collaborative decision module, a cross-domain dynamic coupling module and a closed loop verification module. The method comprises the following steps: constructing an environmental perception map through multi-source data fusion, realizing stable flight under electromagnetic interference by adopting layered anti-interference control, dynamically generating an emergency instruction in combination with federal reinforcement learning, optimizing control parameters by utilizing cross-domain coupling, and driving system iteration upgrading through closed-loop feedback. And the inspection safety and the emergency response efficiency in a complex environment are comprehensively improved. According to the method, electromagnetic interference is suppressed through layered anti-interference control, defect positioning precision is improved through multi-modal fusion, emergency response priority is optimized through federal reinforcement learning, system robustness is enhanced through cross-domain coupling, closed-loop iteration driving self-optimization is achieved, and cooperation of unmanned aerial vehicle electric power inspection safety supervision and emergency response is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) power inspection and safety monitoring, and in particular to a UAV power inspection safety supervision and emergency response system. Background Art

[0002] In the field of power system operation and maintenance, drone inspections have become an important means of safety monitoring for facilities such as high-voltage transmission lines and substations. However, existing drone inspection systems still have significant deficiencies in adaptability to complex electromagnetic environments, the accuracy of multi-source sensor data fusion, and the efficiency of emergency response to sudden defects. Traditional systems often use single-modality sensors for environmental perception, making it difficult to simultaneously capture multi-dimensional defect characteristics such as equipment temperature anomalies, partial discharge, and structural deformation. Furthermore, they lack deep correlation with the grid topology, resulting in large deviations in anomaly location and a high risk of missed detections or misjudgments.

[0003] In addition, the problem of interference from strong electromagnetic fields around high-voltage transmission lines on drone navigation and control systems has not been effectively resolved for a long time. Conventional PID control or static anti-interference strategies can easily cause posture instability or even loss of control in dynamic electromagnetic environments, seriously threatening the safety of inspection operations. At the emergency response level, existing methods mostly rely on preset rules or centralized decision-making, making it difficult to dynamically adjust priorities based on real-time defect distribution, drone endurance status, and multi-machine collaboration requirements, resulting in delayed handling of critical defects or irrational resource allocation. At the same time, the various modules of the system often operate independently, lacking cross-domain parameter coupling and closed-loop optimization mechanisms, making it difficult to adapt to the dynamic needs of environmental mutations and equipment status evolution during inspection tasks.

[0004] Therefore, the present invention proposes a UAV power inspection safety supervision and emergency response system to address the shortcomings of the existing technology. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a drone power inspection safety supervision and emergency response system, which solves the problems of insufficient precise control capability of the drone power inspection system in complex electromagnetic interference environments, difficulty in efficient integration and defect location of multi-source heterogeneous data, delayed emergency response in dynamic scenarios, and lack of multi-module collaborative optimization.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a UAV power inspection safety supervision and emergency response system, the system includes the following modules:

[0007] A multimodal spatiotemporal fusion module is used to receive real-time monitoring data from infrared sensors, ultraviolet sensors, and LiDAR, and process the real-time monitoring data through spatiotemporal fusion based on grid topology constraints to generate an environmental perception map containing device connection relationships;

[0008] A layered anti-interference control module is used to generate electromagnetic interference-resistant flight control instructions based on the environmental perception map through three-level coordinated control: low-level robust tracking, mid-level disturbance compensation, and high-level communication fault tolerance;

[0009] a federated collaborative decision-making module, configured to generate a set of emergency action instructions with dynamically adjusted priorities through federated reinforcement learning based on the defect distribution in the environmental perception map and the remaining battery power of the drone;

[0010] The cross-domain dynamic coupling module is used to receive in real time the posture tracking error of the layered anti-disturbance control module, the environmental perception map confidence of the multimodal spatiotemporal fusion module, and the emergency instruction priority score of the federated collaborative decision-making module, and generate a dynamic allocation strategy for control gain, fusion weight, and decision threshold;

[0011] The closed-loop verification module is used to execute the flight control instructions and emergency action instruction set, and synchronously feed back the executed drone posture deviation data and the updated sensor monitoring data to the multimodal spatiotemporal fusion module and the hierarchical anti-disturbance control module to drive the iterative optimization of the system.

[0012] Preferably, the multimodal spatiotemporal fusion module generates an environmental perception map in the following manner:

[0013] Align the real-time monitoring data obtained by infrared sensors, ultraviolet sensors, and LiDAR in time and space dimensions to construct a fourth-order tensor Where H×W represents the spatial resolution; C represents the number of sensor modalities; T represents the length of the time series;

[0014] The fourth-order tensor is decomposed and optimized with grid topology constraints, and its objective function is:

[0015]

[0016] Among them, G is the core tensor; U (n) is the factor matrix of the nth mode; L is the Laplace matrix generated according to the physical connection relationship of the power grid equipment; γ is the regularization coefficient;

[0017] Furthermore, the cross-domain migration from ultraviolet modality to infrared modality is realized by generating adversarial networks. gen The loss function is defined as:

[0018]

[0019] Discriminator D disc The loss function is:

[0020]

[0021] in, They are ultraviolet and infrared modal data respectively; 1 represents a full 1 vector, and the final output is the fused environmental perception map.

[0022] Preferably, the layered disturbance rejection control module generates flight control instructions in the following manner:

[0023] Based on the UAV posture state vector Including three-dimensional position and three-axis attitude angular velocity, construct the disturbed dynamic equation:

[0024]

[0025] Where f(x) is the nominal dynamic model; Δf(x) is the electromagnetic interference term; B is the control input matrix; u is the control command; d is the external disturbance;

[0026] Design Lyapunov function V(x) = x T Px, where P>0 is obtained by solving the matrix inequality PA T +A T P+Q-PBB T P<0 is confirmed. is the linearization matrix of the nominal system; Q>0 is the design weight matrix;

[0027] The underlying robust control law is:

[0028]

[0029] Where K is the feedback gain matrix; ρ>‖d‖ max is the disturbance compensation coefficient;

[0030] The mid-level disturbance compensation estimates the disturbance term Δf(x) in real time through an adaptive observer:

[0031]

[0032] in, is the interference estimation value; Γ>0 is the adaptive gain matrix; σ>0 is the attenuation factor; when the communication signal strength S(t) is lower than the threshold S th When , it switches to high-level communication fault-tolerant mode, and the control input is updated as follows:

[0033]

[0034] Where e = x des -x is the trajectory tracking error, x des Generated by the mission planning module; J is the UAV inertia matrix; k p With k d is the proportional-derivative gain matrix.

[0035] Preferably, the dynamic update rule of the adaptive gain matrix Γ is:

[0036]

[0037] Guaranteed exponential convergence of the observer under time-varying electromagnetic interference.

[0038] Preferably, the proportional-differential gain matrix satisfies:

[0039] K p =diag(k p1 ,k p2 ,k p3 ,k p4 ,k p5 ,k p6 ),K d =diag(k d1 ,k d2 ,k d3 ,k d4 ,k d5 ,k d6 );

[0040] where each diagonal element k pi ,k di Dynamic adjustment based on the drone mass m and inertia tensor I:

[0041]

[0042] Preferably, the federal collaborative decision-making module generates the emergency action instruction set in the following manner:

[0043] Each drone trains the Q function Q locally i (s,a;θ i ),in:

[0044] Status s = {defect level, weather conditions, remaining power};

[0045] Action a∈{route change, reinforcement request, emergency landing};

[0046] θ i are local model parameters;

[0047] The central server aggregates the global Q function:

[0048]

[0049] Where N is the number of drones participating in federated learning; W is the priority weight matrix; λ is the penalty coefficient; f critic (s) = [defect density (s1), response delay (s)] Tis the critical state feature vector, defect density (s1) is the number of high defect level devices in a unit area, and response delay (s) is the estimated time delay from action triggering to completion;

[0050] Dynamically adjust the action priority using a normalized exponential function:

[0051]

[0052] Among them, τ>0 is the temperature coefficient, which controls the balance between exploration and exploitation; A collection of actions.

[0053] Preferably, the priority weight matrix W is updated according to the following rules:

[0054] Back propagation updates weights based on global policy evaluation error:

[0055]

[0056] Among them, α is the learning rate; r target is the target reward value, which is calculated from the average reward of historical successful response cases.

[0057] Preferably, the cross-domain dynamic coupling module generates a dynamic allocation strategy in the following manner:

[0058] Constructing the coupling matrix Its elements correspond to the posture tracking error of the hierarchical anti-disturbance control module, the environmental perception map reconstruction error of the multimodal spatiotemporal fusion module, and the emergency instruction priority score of the federal collaborative decision-making module. They are specifically defined as:

[0059]

[0060] Where x is the UAV posture state vector; P is the Lyapunov function weight matrix; and are the original and reconstructed fourth-order tensors respectively; is the electromagnetic interference estimation value; G and L are the tensor decomposition core tensor and the grid topology Laplace matrix respectively; W is the priority weight matrix; J is the inertia matrix;

[0061] Perform singular value decomposition on M: M=UΣV T , extract the maximum singular value σ max and its corresponding left singular vector u1, generating the weight vector w = σ max u1; dynamically adjust the control gain matrix K, tensor decomposition regularization coefficient γ, and decision temperature coefficient τ according to w. The update rule is:

[0062] K←K+η1w1,γ←γ+η2w2,τ←τ+η3w3;

[0063] Among them, η1, η2, η3 are preset adjustment steps; w1, w2, w3 are the first three components of the weight vector w.

[0064] Preferably, the closed-loop verification module is optimized in the following ways:

[0065] After executing the flight control instructions and emergency action instructions, the actual posture x is calculated. actual and the desired pose x des Deviation:

[0066] Δx=x des -x actual ;

[0067] Among them, x des Generated by the hierarchical disturbance rejection control module; x actual Acquired through the fusion of inertial navigation unit and GPS;

[0068] The updated sensor monitoring data is injected into the multimodal spatiotemporal fusion module in the form of fourth-order tensor slices. The update rule is:

[0069]

[0070] Among them, M mask Represents the binary mask of the sensor effective area; T calib is the multimodal data calibration matrix; ⊙ represents the Hadamard product;

[0071] The posture deviation Δx fed back to the hierarchical anti-disturbance control module is used to dynamically adjust the Lyapunov function weight matrix P. The update strategy is:

[0072] P←P+μ·diag(ΔxΔx T );

[0073] Where μ>0 is the adaptive learning rate, which is inversely proportional to the current communication delay.

[0074] The present invention also provides a UAV power inspection safety supervision and emergency response method, which includes the following steps:

[0075] S1. Synchronously collect real-time monitoring data from infrared sensors, ultraviolet sensors, and LiDAR, perform spatiotemporal fusion processing based on grid topology constraints, and generate an environmental perception map that includes device connection relationships;

[0076] S2. Generate electromagnetic interference-resistant flight control instructions based on the environmental perception map through three-level coordinated control: low-level robust tracking, mid-level disturbance compensation, and high-level communication fault tolerance;

[0077] S3: Based on the defect distribution in the environmental perception map and the remaining battery power of the drone, a set of emergency action instructions with adjustable priorities is dynamically generated through federated reinforcement learning.

[0078] S4. Obtain the flight control command's posture tracking error, environmental perception map confidence, and emergency command priority score in real time, build a cross-domain coupling matrix, and dynamically allocate control gains, fusion weights, and decision thresholds.

[0079] S5. Execute the flight control instructions and emergency action instruction set, collect updated posture data and sensor monitoring results and feed them back to steps S1 and S2, and drive the system for iterative optimization.

[0080] The present invention provides a UAV power inspection safety supervision and emergency response system. It has the following beneficial effects:

[0081] 1. The present invention effectively suppresses the impact of strong electromagnetic interference generated by high-voltage transmission lines on the posture of drones through the three-level collaborative mechanism of layered anti-interference control modules, solving the problem that traditional drones are susceptible to electromagnetic interference and cause loss of control during power inspections, ensuring the continuity of inspection operations and flight safety. It is particularly suitable for high-electromagnetic risk scenarios such as substations and high-voltage corridors.

[0082] 2. The present invention uses a multimodal spatiotemporal fusion module based on grid topology constraints to deeply couple infrared, ultraviolet, and LiDAR data with device connection relationships, breaking through the blind spots of single sensor detection and significantly improving the positioning accuracy of equipment defects (such as insulator rupture and wire overheating). It provides a high-confidence environmental perception map for safety supervision and reduces the risks of missed detection and false detection.

[0083] 3. The federal collaborative decision-making module of the present invention uses a distributed reinforcement learning framework, combines real-time defect distribution and drone endurance status, and dynamically generates a priority-adjustable emergency instruction set to achieve multi-drone collaborative task allocation and resource scheduling, ensuring the rapid processing of high-priority defects and shortening the response cycle from anomaly discovery to emergency intervention.

[0084] 4. The cross-domain dynamic coupling module of the present invention analyzes multi-dimensional indicators such as control error, perception confidence and decision priority in real time, dynamically allocates control gain, data fusion weight and decision threshold, solves the performance imbalance problem caused by the isolation of traditional system modules, enables the system to maintain stable operation under complex environmental changes, and adapt to the dynamic needs of different inspection tasks.

[0085] 5. The closed-loop verification module of the present invention achieves self-improvement of system performance by executing feedback data to drive continuous optimization of multimodal fusion, anti-disturbance control and collaborative decision-making, avoids the performance degradation problem of traditional open-loop systems caused by environmental time changes, and ensures the long-term reliability of drone power inspection safety supervision and emergency response capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 This is a system architecture diagram of the present invention;

[0087] Figure 2 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0088] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0089] See also Figure 1 The embodiment of the present invention provides a UAV power inspection safety supervision and emergency response system, which includes the following modules:

[0090] A multimodal spatiotemporal fusion module is used to receive real-time monitoring data from infrared sensors, ultraviolet sensors, and LiDAR, and process the real-time monitoring data through spatiotemporal fusion based on grid topology constraints to generate an environmental perception map containing device connection relationships;

[0091] In this embodiment, the multimodal spatiotemporal fusion module is used to perform spatiotemporal alignment and physical constraint fusion on heterogeneous sensor data collected by drones to generate an environmental perception map that accurately reflects the connection relationship between power grid equipment. The specific implementation method is as follows:

[0092] The infrared sensor acquires the temperature distribution data of the power equipment, the ultraviolet sensor captures the corona discharge signal, and the LiDAR generates 3D point cloud data. The multi-source data is aligned in time and space:

[0093] Spatial alignment: Based on GPS positioning and point cloud registration technology, the monitoring areas of different sensors are mapped to a unified geographic coordinate system and divided into H×W grid spatial resolutions;

[0094] Time alignment: Based on the LiDAR sampling frequency, infrared and ultraviolet data are interpolated to form synchronized time series data with a time window length of T;

[0095] Construct the processed data into a fourth-order tensor Among them, C=3 represents three sensor modes, each element Represents the monitoring value of the cth sensor at the grid (h, w) at the tth time.

[0096] In order to integrate the physical connection relationship of the power grid, the Laplace matrix based on the connection state of the device node is introduced Its construction rules are:

[0097] If the grid (h i ,w i ) and (h j ,w j ) There is a direct physical connection (such as wires, insulators) to the corresponding power equipment, then L (i,j) =-1;

[0098] Diagonal element L (i,i) =∑ j≠i |L (i,j) |, the rest of the elements are 0;

[0099] Tensor Perform Tucker decomposition with topological constraints, and optimize the objective function as follows:

[0100]

[0101] Among them, the core tensor Characterize the fused low-dimensional features; is the spatial factor matrix; is the modal factor matrix; is the time series factor matrix; γ is the regularization coefficient, which is used to balance the reconstruction error and the topological constraint strength.

[0102] In view of the complementarity of ultraviolet and infrared data, a generative adversarial network (GAN) is used to achieve cross-modal transfer:

[0103] Generator G gen UV data As input, output pseudo infrared data

[0104] Discriminator D disc Distinguishing real infrared data and generate data

[0105] The generator loss function is defined as:

[0106]

[0107] The discriminator loss function is:

[0108]

[0109] in, They are ultraviolet and infrared modal data respectively; 1 represents a full 1 vector, and the final output is the fused environmental perception map.

[0110] Combine the decomposed core tensor G with the migration data The fusion process extracts the device state feature vector of each grid, aggregates the information of adjacent nodes through the graph convolutional network (GCN), and generates an environmental perception map containing the following elements:

[0111] Device connection topology: identifies physical connection relationships based on the L matrix;

[0112] Abnormal area marking: integrating temperature, discharge, and geometric deformation data to locate defects such as insulator rupture and conductor overheating;

[0113] Time series evolution trend: Displays device status changes over time to support fault prediction.

[0114] A layered anti-interference control module is used to generate electromagnetic interference-resistant flight control instructions based on the environmental perception map through three-level coordinated control: low-level robust tracking, mid-level disturbance compensation, and high-level communication fault tolerance;

[0115] In this embodiment, the layered anti-disturbance control module generates high-precision flight control instructions in a complex environment with strong electromagnetic interference and unstable communication through a three-level coordinated control mechanism of low-level robust tracking, middle-level disturbance compensation, and high-level communication fault tolerance:

[0116] Define the drone posture state vector as Where x, y, and z are three-dimensional position coordinates; φ (roll angle), θ (pitch angle), and ψ (yaw angle) are the three-axis Euler angles. Considering electromagnetic interference and external disturbances, the nonlinear disturbed dynamic equation is established:

[0117]

[0118] in, is the nominal dynamic model, which consists of the UAV mass m, inertia tensor And the aerodynamic coefficient is determined in the following form:

[0119]

[0120] is the electromagnetic interference term, the additional torque and force generated by the alternating magnetic field induced current of the high-voltage transmission line, modeled as Δf(x) = E·x, where is the time-varying interference coefficient matrix;

[0121] To control the input matrix, the thrusts u1, u2, u3, and u4 of the four rotors are mapped to six-degree-of-freedom motion. The specific structure is:

[0122]

[0123] Where l is the length of the swing arm; κ is the thrust-torque conversion coefficient;

[0124] is the control instruction, corresponding to the rotation speed of the four rotors;

[0125] is the external disturbance, including wind disturbance, sensor noise, etc., satisfying ‖d‖≤d max .

[0126] The underlying robust tracking control law is designed based on Lyapunov stability theory to design a globally asymptotically stable robust control law:

[0127] Construct a Lyapunov function:

[0128] V(x)=x T Px, P>0;

[0129] in, is a symmetric positive definite matrix, which is determined by solving the linear matrix inequality (LMI):

[0130] PA+A T P+Q-PBB T P < 0;

[0131] in, is the Jacobian matrix of the nominal system at the equilibrium point x = 0;

[0132] To design a weight matrix for adjusting the state convergence rate, a diagonal matrix Q = diag(q1,q2,…,q6) is usually taken.

[0133] Derive the robust control law by minimizing the time derivative of the Lyapunov function The underlying control law is obtained:

[0134]

[0135] in, is the feedback gain matrix, which is designed by pole placement method to satisfy K= in, is the pseudo-inverse of B; ρ>||d|| max is the disturbance compensation coefficient, which is used to offset the upper limit of external disturbance d max impact.

[0136] Adaptive compensation of electromagnetic interference in the middle layer. An adaptive observer is designed to estimate the electromagnetic interference term Δf(x) in real time, avoiding reliance on an accurate interference model:

[0137] Observer dynamic equation:

[0138]

[0139] in, is the interference estimate; is the adaptive gain matrix, and the dynamic update rule is: β>0; where β is the attenuation factor, which controls the memory decay rate of the gain matrix; σ>0 is the attenuation factor, which adjusts the parameter of the estimation error convergence speed.

[0140] By constructing a composite Lyapunov function (in, is the estimation error), it can be proved that under the update rule, the estimation error exponentially converges to zero.

[0141] In high-level communication fault-tolerant control mode, when the wireless communication signal strength S(t) is lower than the threshold S th Time (S th The communication module switches to inertial navigation mode based on the channel bit error rate and delay calculation:

[0142] Control input update:

[0143]

[0144] Where e = x des -x is the trajectory tracking error, x des Generated by the mission planning module; is the inertial matrix of the UAV, defined as J=diag(m,m,m,I xx ,I yy ,I zz );k p =diag(k p1 ,k p2 ,k p3 ,k p4 ,k p5 ,k p6 ) and k d =diag(k d1 ,k d2 ,k d3 ,k d4 ,k d5 ,k d6 ) is the proportional-derivative gain matrix, whose diagonal elements are dynamically adjusted according to the mass and inertia parameters:

[0145]

[0146] This adjustment strategy ensures that the control gains automatically match the dynamic characteristics of the drone under different load conditions (such as carrying detection equipment or battery power consumption).

[0147] a federated collaborative decision-making module, configured to generate a set of emergency action instructions with dynamically adjusted priorities through federated reinforcement learning based on the defect distribution in the environmental perception map and the remaining battery power of the drone;

[0148] In this embodiment, the federated collaborative decision-making module implements multi-UAV collaborative decision-making through a distributed reinforcement learning framework. It combines the defect distribution and real-time status parameters of the environmental perception map to dynamically generate a set of emergency action instructions with adjustable priorities. The following is a complete description of the federated learning architecture, strategy optimization, parameter definition, and implementation logic:

[0149] Federated reinforcement learning architecture, local Q function training: Each drone acts as an independent intelligent agent and maintains a local Q function Q i (s,a;θ i ), its state space s, action space a and network parameters are defined as follows:

[0150] State space s = {s1,s2,s3}:

[0151] s1 (defect level): Calculated based on the device heating temperature T, corona discharge intensity U, and geometric deformation error δ in the environmental perception map. The formula is:

[0152]

[0153] Among them, α T ,α U ,α δ is the normalized weight coefficient, T min 、T max is the safe temperature range of the equipment; U th is the corona discharge threshold.

[0154] s2 (weather conditions): wind speed v w , humidity h, and rainfall intensity r constitute the vector s2=[v w ,h,r] T ;

[0155] s3 (remaining power): expressed as a percentage,

[0156] Action space a∈{a1,a2,a3}:

[0157] a1 (route change): Generate a new path point sequence p new =PathPlan(s1,s3);

[0158] a2 (reinforcement request): Send a request signal to the designated drone assembly

[0159] a3 (Emergency Landing): Trigger the landing protocol and release the emergency identification signal.

[0160] Network structure: A two-layer fully connected neural network is used to implement the Q function, with an input layer dimension of d in = dim(s), hidden layer dimension d hidden =64, output layer dimension d out =3, the activation function is ReLU.

[0161] Local training process, experience replay buffer pool: store historical transfer samples (s t ,a t ,r t ,s t+1 ), with a capacity of N buffer ;

[0162] Target network parameter update: Every C target Step 1: Set the online network parameters θ i Copy to target network

[0163] Loss function: Minimize the temporal difference error:

[0164]

[0165] in, is the experience replay buffer pool; γ∈[0,1) is the discount factor, which weighs current and future rewards; The Q value output by the target network.

[0166] Global policy aggregation and priority optimization, global Q function generation: the central server performs the aggregation period T agg Collect the local Q functions of each drone and generate a global Q function:

[0167]

[0168] Where N is the number of drones participating in federated learning; f critic (s) = [defect density (s1), response delay (s)] T is the critical state feature vector, defect density (s1) is the number of high defect level devices in a unit area, and response delay (s) is the estimated time delay from action triggering to completion; is the priority weight matrix, encoding the trade-off between defect density and response delay; λ>0 is the regularization penalty coefficient, which suppresses overfitting of the strategy.

[0169] Dynamic priority calculation, based on the global Q function, uses a normalized exponential function with a temperature coefficient to calculate the action priority:

[0170]

[0171] Among them, τ>0 is the temperature coefficient, which controls the balance between exploration and exploitation; A collection of actions.

[0172] The weight matrix is adaptively updated, and the update rule is: the priority weight matrix W is optimized by back propagation of the policy evaluation error:

[0173]

[0174] Among them, α∈(0,1) is the learning rate, which controls the weight update step size; r target is the target reward value, calculated from the historical successful case library:

[0175]

[0176] M is the number of historical successful cases; is the reward value for the kth successful response, and the calculation formula is:

[0177]

[0178] Among them, β1, β2, and β3 are reward weight coefficients.

[0179] Historical case library management, case storage: record the status of successful responses k 、Action a k , reward r k and environmental parameters;

[0180] Retrieval mechanism: Retrieve the most similar case based on the Euclidean distance of the current state s. The formula is:

[0181] Implementation process:

[0182] Federation aggregation period: preferably, the central server agg = Global Q function aggregation is performed once every 10 minutes;

[0183] Temperature coefficient adjustment: Initially set τ = 1.0, and linearly decay to τ with each training round. min =0.1;

[0184] Reward weight coefficient: Dynamically adjusted according to the task type. For example, in a defect inspection task, β1 = 0.6, β2 = 0.2, and β3 = 0.2 are set.

[0185] Network training hyperparameters: learning rate η = 0.001, discount factor γ = 0.99, target network update cycle C target =100.

[0186] The cross-domain dynamic coupling module is used to receive in real time the posture tracking error of the layered anti-disturbance control module, the environmental perception map confidence of the multimodal spatiotemporal fusion module, and the emergency instruction priority score of the federated collaborative decision-making module, and generate a dynamic allocation strategy for control gain, fusion weight, and decision threshold;

[0187] In this embodiment, the cross-domain dynamic coupling module is used to integrate multi-domain information such as control error, perception confidence, and decision priority in real time to generate a dynamic allocation strategy to optimize overall system performance. The core of this module is to construct a cross-domain coupling matrix and implement adaptive parameter adjustment based on matrix feature analysis. The specific implementation method is as follows:

[0188] Matrix dimension and element mapping, building coupling matrix Its elements are mapped to the core performance indicators of the three modules: hierarchical anti-disturbance control, multimodal spatiotemporal fusion, and federal collaborative decision-making:

[0189]

[0190] The first row represents the interaction between the control and perception modules:

[0191] M 11 =x T Px: pose tracking error, is the UAV posture state vector, P>0 is the Lyapunov function weight matrix;

[0192] Environmental perception map reconstruction error, is the original fourth-order tensor, To reconstruct the tensor, ‖·‖ F is the Frobenius norm;

[0193] M 13 =0: Reserved bit for subsequent expansion.

[0194] The second row represents the correlation between the perception and decision modules:

[0195] The L2 norm of the electromagnetic interference estimate, Output of the adaptive observer of the hierarchical disturbance rejection control module;

[0196] M 22 =tr(G T LG): Topological consistency measure of tensor decomposition, G is the core tensor, and L is the grid topology Laplace matrix;

[0197] M 23 =max a Priority(a): The highest priority action score, output by the federated collaborative decision module.

[0198] The third row represents the coupling effect between the decision-making and control modules:

[0199] M 31 =0: reserved bit;

[0200] M 32 =tr(W T W): trace of the priority weight matrix, reflecting the stability of the decision-making strategy;

[0201] M 33 =rank(J): UAV inertia matrix The rank of , characterizes the controllability of the dynamic model.

[0202] Singular value decomposition and weight vector generation:

[0203] Matrix decomposition process: Perform singular value decomposition (SVD) on the coupling matrix M:

[0204] M=UΣV T ;

[0205] in, Left singular vector matrix; Diagonal matrix, the elements are the singular values σ1≥σ2≥σ3≥0 in descending order; Right singular vector matrix.

[0206] Feature extraction rules:

[0207] Extract the largest singular value σ max =σ1, which characterizes the main energy direction of cross-domain coupling;

[0208] Extract the corresponding left singular vector (the first column of U), which reflects the contribution weight of each module to the main energy direction;

[0209] Generate dynamic weight vector w = σ max u1, where components w1, w2, and w3 are associated with the parameter adjustment requirements of the control, perception, and decision modules, respectively.

[0210] Dynamic parameter adjustment strategy:

[0211] Control gain adjustment, update the feedback gain matrix K of the hierarchical anti-disturbance control module based on the weight component w1:

[0212] K←K+η1w1·I;

[0213] Where η1>0: preset control gain adjustment step size; The identity matrix ensures that the gain adjustment is compatible with the original control structure.

[0214] Regularization coefficient optimization, based on the weight component w2, adjusts the tensor decomposition regularization coefficient γ of the multimodal spatiotemporal fusion module:

[0215] γ←γ+η2w2;

[0216] Among them, η2>0: the regularization coefficient adjusts the step size to dynamically balance the reconstruction error and the topological constraint strength.

[0217] Decision temperature coefficient adaptation, adjusts the temperature coefficient τ of the federated collaborative decision module based on the weight component w3:

[0218] τ←τ+η3w3;

[0219] Where η3>0: The temperature coefficient adjusts the step size to control the balance rate between exploration and exploitation of actions.

[0220] Implementation process and parameter configuration:

[0221] Coupling matrix update frequency: preferably, every T update =Updated once per second, synchronized with the system master control cycle;

[0222] SVD calculation optimization: Jacobi iteration method is used to achieve real-time decomposition, avoiding the sensitivity of traditional QR algorithm to matrix conditions;

[0223] Adaptive step size mechanism: The adjustment step sizes η1, η2, and η3 can decay over the system running time, for example:

[0224]

[0225] in, is the initial step size and ξ is the decay coefficient, which is used to converge to the steady state gradually.

[0226] A closed-loop verification module is used to execute the flight control instructions and emergency action instruction set, and synchronously feed back the executed UAV posture deviation data and updated sensor monitoring data to the multimodal spatiotemporal fusion module and the hierarchical anti-disturbance control module to drive the iterative optimization of the system;

[0227] In this embodiment, the closed-loop verification module is used to execute the flight control instructions and emergency action instruction sets, and drive system dynamic optimization through real-time feedback of posture deviations and sensor data, forming a closed-loop iterative improvement mechanism. Its specific implementation is as follows:

[0228] Pose deviation calculation and data fusion:

[0229] Deviation calculation logic, after executing flight control instructions, obtains the actual position of the UAV through multi-source positioning data fusion:

[0230] Desired pose: generated by the hierarchical anti-disturbance control module, denoted as xdes =[x d ,y d ,z d ,φ d ,θ d ,ψ d ] T ;

[0231] Actual pose: calculated by integrating the inertial navigation unit (IMU) with the global positioning system (GPS), denoted as x actual =[x a ,y a ,z a ,φ a ,θ a ,ψ a ] T ;

[0232] Pose deviation: Calculate the Euclidean deviation vector between the actual pose and the expected pose: Δx = x des -x actual =[Δx,Δy,Δz,Δφ,Δθ,Δψ] T ;

[0233] Among them, x des Generated by the hierarchical disturbance rejection control module; x actual The positioning is obtained by integrating the inertial navigation unit with GPS.

[0234] The positioning data fusion method uses the Kalman filter to fuse the IMU angular velocity, accelerometer data and GPS position information. The state equation and observation equation are:

[0235]

[0236] z=Hx actual +v;

[0237] in, is the state transfer matrix, which is determined by the UAV kinematic model; To control the input matrix, map the rotor thrust to the attitude change; is the process noise, the covariance matrix Q is calibrated by the IMU accuracy; z is the observation vector (GPS position and IMU attitude angle); is the observation matrix, extracting the measurable state; For the observed voice, the covariance matrix R is determined by the sensor error characteristics.

[0238] Sensor data update rules:

[0239] Multimodal tensor dynamic injection, injecting updated infrared, ultraviolet, and LiDAR monitoring data into the fourth-order tensor according to the spatiotemporal alignment rules The update formula is:

[0240]

[0241] Among them, M mask ∈{0,1} H×W is a binary mask matrix that identifies the effective coverage area of the sensor (1 indicates valid, 0 indicates invalid); ⊙ is the Hadamard product (element-wise multiplication), which is used to mask invalid data in the non-detection area; The multimodal data calibration matrix is used to convert the pose deviation Δx into the spatial compensation of each sensor mode, which is determined by the pre-calibrated sensor-motion coupling relationship.

[0242] Feedback-driven parameter optimization, the posture deviation Δx is fed back to the hierarchical anti-disturbance control module, and the Lyapunov function weight matrix P is dynamically updated:

[0243] P←P+μ·diag(ΔxΔx T );

[0244] Where diag(·) is a diagonal matrix constructed by extracting the diagonal elements of the matrix; μ>0 is the adaptive learning rate, which is inversely proportional to the current communication delay τ and is calculated as follows:

[0245]

[0246] Where μ0 is the baseline learning rate; κ is the attenuation coefficient, which controls the impact of communication delay on the learning rate; and τ is the real-time communication delay, which is periodically measured and reported by the communication module.

[0247] The physical meaning of the update strategy:

[0248] When the communication delay τ is low, increase μ to speed up the weight matrix adjustment and improve the control response speed;

[0249] When τ is high, reduce μ to avoid high-frequency oscillation and ensure system stability.

[0250] See also Figure 2 The present invention also provides a method for safety supervision and emergency response of UAV power inspection, which includes the following steps:

[0251] S1. Synchronously collect real-time monitoring data from infrared sensors, ultraviolet sensors, and LiDAR, perform spatiotemporal fusion processing based on grid topology constraints, and generate an environmental perception map that includes device connection relationships;

[0252] The system synchronously collects real-time monitoring data from infrared sensors, ultraviolet sensors, and LiDAR. Infrared data is used to capture abnormal temperature distribution of power equipment, ultraviolet data detects corona discharge intensity, and LiDAR generates high-precision three-dimensional point clouds to characterize equipment geometric deformation. Based on the topological constraints of the power grid (such as the physical connection relationship of equipment and the layout of electrical nodes), multi-source data is aligned and integrated through a spatiotemporal fusion algorithm to eliminate the spatiotemporal inconsistencies of sensor observations. The fused data generates an environmental perception map that includes equipment location, connection status, abnormal markers (such as broken insulators and overheated conductor areas), and dynamic risk level assessments, providing a global environmental cognition basis for subsequent control and decision-making.

[0253] S2. Generate electromagnetic interference-resistant flight control instructions based on the environmental perception map through three-level coordinated control: low-level robust tracking, mid-level disturbance compensation, and high-level communication fault tolerance;

[0254] Based on the defect information and real-time posture requirements of the environmental perception map, a hierarchical control strategy is used to generate anti-electromagnetic interference instructions:

[0255] Low-level robust tracking: The basic control law is designed based on Lyapunov stability theory to ensure that the drone accurately tracks the preset trajectory without interference;

[0256] Mid-level disturbance compensation: uses an adaptive observer to estimate electromagnetic interference and external disturbances in real time and generate compensation instructions to offset the interference effects;

[0257] High-level communication fault tolerance: When wireless signals are attenuated or interrupted, the system switches to inertial navigation mode and generates fault-tolerant control instructions based on historical trajectories and dynamic model predictions.

[0258] The three-level control strategies work together to ensure the UAV's stable flight and mission execution in a strong electromagnetic environment.

[0259] S3: Based on the defect distribution in the environmental perception map and the remaining battery power of the drone, a set of emergency action instructions with adjustable priorities is dynamically generated through federated reinforcement learning.

[0260] A federated reinforcement learning framework dynamically generates emergency command sets based on the defect density in the environmental perception map, the remaining UAV battery level, and real-time weather conditions. Each UAV, acting as an independent agent, trains a decision-making model based on local historical mission data. A central server aggregates global strategies and calculates action priorities. Priority scoring comprehensively considers defect handling urgency, endurance, and collaborative efficiency, dynamically adjusting the order of command sets (e.g., prioritizing high-risk defects, requesting reinforcements, or urgently returning home), achieving distributed decision-making and optimal global resource allocation.

[0261] S4. Obtain the flight control command's posture tracking error, environmental perception map confidence, and emergency command priority score in real time, build a cross-domain coupling matrix, and dynamically allocate control gains, fusion weights, and decision thresholds.

[0262] The system acquires three cross-domain metrics in real time: the pose tracking error of layered anti-disturbance control, the reconstruction confidence of the environmental perception map, and the priority score of emergency commands. A coupling matrix is constructed to characterize the interactions between modules. Matrix feature analysis extracts the main energy direction, dynamically allocating control gains (to enhance trajectory tracking robustness), fusion weights (to increase the decision weight of high-confidence data), and decision thresholds (to optimize action triggering conditions). This enables coordinated adaptive adjustment of multi-module parameters, ensuring balanced overall system performance in complex scenarios.

[0263] S5, executing the flight control instructions and emergency action instruction set, collecting updated posture data and sensor monitoring results and feeding them back to steps S1 and S2, and driving the system for iterative optimization;

[0264] After executing flight control commands and emergency actions, the drone collects updated actual posture data and environmental monitoring results through multi-source sensors (such as IMU, GPS, and LiDAR). This posture deviation data is fed back to the layered anti-disturbance control module for online optimization of control algorithm parameters. New sensor data is then fed into the multimodal fusion module to refresh the real-time state of the environmental perception map. Through continuous closed-loop verification and data-driven optimization, the system gradually improves control accuracy, perception reliability, and decision-making efficiency, forming an iterative enhancement mechanism that adapts to strong electromagnetic environments and dynamic mission requirements.

[0265] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A UAV power inspection safety supervision and emergency response system, characterized by: The system includes the following modules: A multimodal spatiotemporal fusion module is used to receive real-time monitoring data from infrared sensors, ultraviolet sensors, and LiDAR, and process the real-time monitoring data through spatiotemporal fusion based on grid topology constraints to generate an environmental perception map containing device connection relationships; A layered anti-interference control module is used to generate electromagnetic interference-resistant flight control instructions based on the environmental perception map through three-level coordinated control: low-level robust tracking, mid-level disturbance compensation, and high-level communication fault tolerance; a federated collaborative decision-making module, configured to generate a set of emergency action instructions with dynamically adjusted priorities through federated reinforcement learning based on the defect distribution in the environmental perception map and the remaining battery power of the drone; The cross-domain dynamic coupling module is used to receive in real time the posture tracking error of the layered anti-disturbance control module, the environmental perception map confidence of the multimodal spatiotemporal fusion module, and the emergency instruction priority score of the federated collaborative decision-making module, and generate a dynamic allocation strategy for control gain, fusion weight, and decision threshold; The closed-loop verification module is used to execute the flight control instructions and emergency action instruction set, and synchronously feed back the executed drone posture deviation data and the updated sensor monitoring data to the multimodal spatiotemporal fusion module and the hierarchical anti-disturbance control module to drive the iterative optimization of the system.

2. The UAV power inspection safety supervision and emergency response system according to claim 1 is characterized in that: The multimodal spatiotemporal fusion module generates an environmental perception map in the following way: Align the real-time monitoring data obtained by infrared sensors, ultraviolet sensors, and LiDAR in time and space dimensions to construct a fourth-order tensor Where H×W represents the spatial resolution; C represents the number of sensor modalities; T represents the length of the time series; The fourth-order tensor is decomposed and optimized with grid topology constraints, and its objective function is: Among them, G is the core tensor; U (n) is the factor matrix of the nth mode; L is the Laplace matrix generated according to the physical connection relationship of the power grid equipment; γ is the regularization coefficient; Furthermore, the cross-domain migration from ultraviolet modality to infrared modality is realized by generating adversarial networks. gen The loss function is defined as: Discriminator D disc The loss function is: in, They are ultraviolet and infrared modal data respectively; 1 represents a full 1 vector, and the final output is the fused environmental perception map.

3. The UAV power inspection safety supervision and emergency response system according to claim 1 is characterized in that: The layered disturbance rejection control module generates flight control instructions in the following manner: Based on the UAV posture state vector Including three-dimensional position and three-axis attitude angular velocity, construct the disturbed dynamic equation: Where f(x) is the nominal dynamic model; Δf(x) is the electromagnetic interference term; B is the control input matrix; u is the control command; d is the external disturbance; Design Lyapunov function V(x) = x T Px, where P>0 is obtained by solving the matrix inequality PA T +A T P+Q-PBB T P<0 confirmed, is the linearization matrix of the nominal system; Q>0 is the design weight matrix; The underlying robust control law is: Where K is the feedback gain matrix; ρ>‖d‖ max is the disturbance compensation coefficient; The mid-level disturbance compensation estimates the disturbance term Δf(x) in real time through an adaptive observer: in, is the interference estimation value; Γ>0 is the adaptive gain matrix; σ>0 is the attenuation factor; when the communication signal strength S(t) is lower than the threshold S th When , it switches to high-level communication fault-tolerant mode, and the control input is updated as follows: Where e = x des -x is the trajectory tracking error, x des Generated by the mission planning module; J is the UAV inertia matrix; k p With k d is the proportional-derivative gain matrix.

4. The UAV power inspection safety supervision and emergency response system according to claim 3 is characterized in that: The dynamic update rule of the adaptive gain matrix Γ is: Guaranteed exponential convergence of the observer under time-varying electromagnetic interference.

5. The UAV power inspection safety supervision and emergency response system according to claim 3 is characterized in that: The proportional-differential gain matrix satisfies: K p =diag(k p1 ,k p2 ,k p3 ,k p4 ,k p5 ,k p6 ),K d =diag(k d1 ,k d2 ,k d3 ,k d4 ,k d5 ,k d6 ); where each diagonal element k pi ,k di Dynamic adjustment based on the drone mass m and inertia tensor I:

6. The UAV power inspection safety supervision and emergency response system according to claim 1 is characterized in that: The federal collaborative decision-making module generates an emergency action instruction set in the following manner: Each drone trains the Q function Q locally i (s,a;θ i ),in: Status s = {defect level, weather conditions, remaining power}; Action a∈{route change, reinforcement request, emergency landing}; θ i are local model parameters; The central server aggregates the global Q function: Where N is the number of drones participating in federated learning; W is the priority weight matrix; λ is the penalty coefficient; f critic (s) = [defect density (s1), response delay (s)] T is the critical state feature vector, defect density (s1) is the number of high defect level devices in a unit area, and response delay (s) is the estimated time delay from action triggering to completion; Dynamically adjust the action priority using a normalized exponential function: Among them, τ>0 is the temperature coefficient, which controls the balance between exploration and exploitation; A collection of actions.

7. The UAV power inspection safety supervision and emergency response system according to claim 6 is characterized in that: The priority weight matrix W is updated by the following rules: Back propagation updates weights based on global policy evaluation error: Among them, α is the learning rate; r target is the target reward value, which is calculated from the average reward of historical successful response cases.

8. The UAV power inspection safety supervision and emergency response system according to claim 1 is characterized in that: The cross-domain dynamic coupling module generates a dynamic allocation strategy in the following way: Constructing the coupling matrix Its elements correspond to the posture tracking error of the hierarchical anti-disturbance control module, the environmental perception map reconstruction error of the multimodal spatiotemporal fusion module, and the emergency instruction priority score of the federal collaborative decision-making module. They are specifically defined as: Where x is the UAV posture state vector; P is the Lyapunov function weight matrix; and are the original and reconstructed fourth-order tensors respectively; is the electromagnetic interference estimation value; G and L are the tensor decomposition core tensor and the grid topology Laplace matrix respectively; W is the priority weight matrix; J is the inertia matrix; Perform singular value decomposition on M: M=UΣV T , extract the maximum singular value σ max and its corresponding left singular vector u1, generating the weight vector w = σ max u1; dynamically adjust the control gain matrix K, tensor decomposition regularization coefficient γ, and decision temperature coefficient τ according to w. The update rule is: K←K+η1w1,γ←γ+η2w2,τ←τ+η3w3; Among them, η1, η2, η3 are preset adjustment steps; w1, w2, w3 are the first three components of the weight vector w.

9. The UAV power inspection safety supervision and emergency response system according to claim 1 is characterized in that: The closed-loop verification module is optimized in the following ways: After executing the flight control instructions and emergency action instructions, the actual posture x is calculated. actual and the desired pose x des Deviation: Δx=x des -x actual ; Among them, x des Generated by the hierarchical disturbance rejection control module; x actual Acquired through the fusion of inertial navigation unit and GPS; The updated sensor monitoring data is injected into the multimodal spatiotemporal fusion module in the form of fourth-order tensor slices. The update rule is: Among them, M mask Represents the binary mask of the sensor effective area; T calib is the multimodal data calibration matrix; ⊙ represents the Hadamard product; The posture deviation Δx fed back to the hierarchical anti-disturbance control module is used to dynamically adjust the Lyapunov function weight matrix P. The update strategy is: P←P+μ·diag(ΔxΔx T ); Where μ>0 is the adaptive learning rate, which is inversely proportional to the current communication delay.

10. A method for safety supervision and emergency response of power inspection by drones, applied to the system according to any one of claims 1 to 9, characterized in that: The method comprises the following steps: S1. Synchronously collect real-time monitoring data from infrared sensors, ultraviolet sensors, and LiDAR, perform spatiotemporal fusion processing based on grid topology constraints, and generate an environmental perception map that includes device connection relationships; S2. Generate electromagnetic interference-resistant flight control instructions based on the environmental perception map through three-level coordinated control: low-level robust tracking, mid-level disturbance compensation, and high-level communication fault tolerance; S3: Based on the defect distribution in the environmental perception map and the remaining battery power of the drone, a set of emergency action instructions with adjustable priorities is dynamically generated through federated reinforcement learning. S4. Obtain the flight control command's posture tracking error, environmental perception map confidence, and emergency command priority score in real time, build a cross-domain coupling matrix, and dynamically allocate control gains, fusion weights, and decision thresholds. S5. Execute the flight control instructions and emergency action instruction set, collect updated posture data and sensor monitoring results and feed them back to steps S1 and S2, and drive the system for iterative optimization.

Images