Multi-source sensor fused adaptive navigation system

Abstract

The invention relates to the technical field of autonomous navigation and robot environment perception, and discloses a multi-source sensor fused adaptive navigation system, which comprises a multi-source sensor space-time synchronization module, a quantum particle filtering positioning estimation module, a space-time element learning controller module, a cross-modal quantum fusion module and an adaptive navigation control module. Multi-source data space-time alignment is realized through Lie group SE (3) calibration and dynamic time warping; the positioning robustness of particle filtering is improved based on quantum state coding and annealing optimization; dynamically distributing a fusion weight and injecting a physical constraint by utilizing a meta-learning network; feature level fusion of laser radar, vision and inertial data is realized by means of a quantum entanglement mechanism; and constructing closed-loop adaptive navigation by combining model predictive control and quantum purity trigger feedback. According to the method, the navigation reliability problem caused by misalignment of multi-modal sensor data fusion, divergence of state estimation and insufficient cross-modal relevance in a dynamic environment is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous navigation and robot environment perception, and in particular to an adaptive navigation system with multi-source sensor fusion. Background Art

[0002] With the widespread adoption of autonomous navigation systems in complex and dynamic environments, multi-source sensor fusion technology has become a key approach to improving environmental perception and positioning accuracy. Traditional approaches often rely on single-modality sensors or simple weighted fusion strategies. When dealing with scenarios such as sudden changes in illumination, occlusion interference, and rapid motion, they suffer from systemic flaws such as inconsistent spatiotemporal references, divergent state estimates, and weak cross-modal correlation.

[0003] Especially in the collaborative process of multimodal data such as lidar, vision and inertial measurement units, existing technologies are difficult to effectively overcome the feature space mismatch problem caused by sensor heterogeneity, and lack an online adaptation mechanism to dynamic changes in the environment.

[0004] Furthermore, traditional particle filtering frameworks are limited by sample degradation and high-dimensional search efficiency, making them prone to local optima in unstructured scenarios, significantly reducing the robustness of navigation systems. Achieving spatiotemporal consistency, quantized state estimation, and closed-loop adaptive control through multi-source heterogeneous data fusion has become a technical bottleneck hindering the development of highly reliable autonomous navigation systems. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides an adaptive navigation system with multi-source sensor fusion, which solves the technical problems of reduced navigation reliability caused by inaccurate spatiotemporal references of multimodal sensor data, divergence of state estimation, and insufficient cross-modal correlation in dynamic and complex environments.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: an adaptive navigation system with multi-source sensor fusion, comprising: The multi-source sensor spatiotemporal synchronization module is used to align timestamps and convert spatial coordinates of raw data from lidar, visual cameras, inertial measurement units, and satellite positioning sensors, and output a standardized data stream that is spatiotemporally aligned. A quantum particle filter positioning estimation module receives the standardized data stream and generates a pose estimation result in a dynamic environment through quantum state encoding and annealing optimization; A spatiotemporal meta-learning controller module dynamically calculates the fusion weight coefficients of each sensor based on the pose estimation results and environmental characteristic parameters; A cross-modal quantum fusion module performs quantum entanglement feature-level fusion on the lidar point cloud, visual image, and inertial data according to the fusion weight coefficient, and outputs an enhanced navigation state; An adaptive navigation control module generates anti-interference control instructions based on the enhanced navigation state and feeds back the instructions to the spatiotemporal meta-learning controller module in real time to trigger weight update.

[0007] Preferably, the spatial coordinate transformation in the multi-source sensor spatiotemporal synchronization module is implemented by Lie group SE (3), and its transformation matrix is ​​defined as: in: G m (m=1,2,...,6) is the generator of SE(3) Lie algebra, corresponding to the six degrees of freedom of spatial motion (translation x, y, z and rotation R x , R y , R z ); ξ m are the calibration parameters obtained by optimizing the characteristic points of the calibration plate; exp(·) represents the Lie group exponential mapping operation.

[0008] Preferably, the quantum state encoding in the quantum particle filter positioning estimation module satisfies: |ψ i >=α i |0>+β i |1>, and |α i | 2 +|β i | 2 =1; in: |ψ i > is the quantum superposition state of the i-th particle, |0> represents the position credible quantum state, and its amplitude α i squared|α i | 2 Characterizes the position confidence probability, |1> represents the credible quantum state of the motion state, and its amplitude β i squared|β i | 2 Characterize the confidence probability of the motion state.

[0009] Preferably, the physical constraints in the spatiotemporal meta-learning controller module are implemented through a loss function: is the mean square error loss of sensor weight prediction, is the time derivative of the velocity vector, calculated by automatic differentiation, a IMU is the acceleration measurement value output by the inertial measurement unit, g is the gravitational acceleration vector, and λ is the weight coefficient of the physical constraint term, with a value range of [0.1, 1.0].

[0010] Preferably, the quantum entanglement channel in the cross-modal quantum fusion module is constructed by a controlled NOT gate, which satisfies: in: |P> is the quantum state mapped by the lidar point cloud P; |E> is the quantum state mapped by the event camera data stream E; f(P) is the point cloud geometric feature extraction function, and the output is a binary feature vector; Represents a bitwise exclusive OR operation.

[0011] Preferably, the fault-tolerant triggering condition of the adaptive navigation control module is: ε=1-Tr(ρ 2 )<∈; in: ρ is the density matrix of the quantum fusion state, which is calculated by ρ=|Ψ><Ψ|; Tr(ρ 2 ) is the purity of the density matrix; ∈ is the preset quantum purity threshold, which ranges from [0.2, 0.5].

[0012] Preferably, the quantum state encoding is implemented through a quantum-classical hybrid computing architecture, specifically: On edge computing devices, quantum bit operations are simulated through GPU tensor cores to encode particle states as quantum state amplitudes.

[0013] Preferably, the physical kinematic equation constraints include: The motion acceleration equation of the inertial measurement unit is used as a hard constraint and injected into the loss function of the meta-learning network through the Lagrange multiplier method.

[0014] Preferably, the correlation strength of the quantum entanglement channel is dynamically controlled by the density matrix purity, specifically: the fusion ratio of the laser radar and visual data is adjusted according to the quantum state purity EE: F fused =ε·F LiDAR +(1-ε)·F Event ; Among them F LiDAR With F Event are the feature vectors of the lidar and event camera, respectively.

[0015] An adaptive navigation method based on multi-source sensor fusion includes the following steps: S1, time synchronization and spatial coordinate unification of multimodal sensor data; S2, realize pose estimation in dynamic environment through quantum particle filtering; S3, dynamically assign sensor fusion weights based on environmental characteristics and physical constraints; S4. Using quantum entanglement mechanism to achieve cross-modal feature-level fusion; S5. Generate anti-interference control instructions and feedback environmental mutation signals to trigger weight updates.

[0016] The present invention provides an adaptive navigation system based on multi-source sensor fusion. It has the following beneficial effects: 1. The present invention uses Lie group SE (3) unified calibration and dynamic time warping algorithm to eliminate the spatiotemporal deviation between multimodal sensors such as lidar, vision, and inertial sensors, and can achieve sub-pixel spatiotemporal synchronization, providing highly consistent data input for subsequent fusion, and solving the feature misalignment problem caused by sensor heterogeneity in traditional methods.

[0017] 2. The present invention is based on a pose estimation framework based on quantum particle filtering. It suppresses particle degradation through quantum state encoding and annealing optimization, and combines the quantum tunneling effect to enhance the ability to escape from local optimal solutions. It can achieve high-confidence pose estimation in unstructured environments and overcome the divergence risk of traditional particle filtering in complex dynamic scenes.

[0018] 3. The present invention uses the quantum entanglement mechanism to construct a feature-level fusion channel for lidar, vision and inertial data, and realizes the essential correlation of cross-modal features through controlled NOT gate operations. It can solve the information redundancy and conflict problems caused by modal differences in traditional methods, and significantly improve the integrity of environmental perception in complex scenarios.

[0019] 4. The present invention combines model predictive control with a feedback mechanism triggered by quantum purity to adjust the fusion weights and control strategies in real time. It can autonomously activate redundant modules and smoothly switch control modes when sensors fail or the environment changes suddenly, ensuring the system's continued stable operation under extreme working conditions, breaking through the traditional open-loop control's reliance on preset parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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.

[0022] Example: Please see the attached Figure 1 The embodiment of the present invention provides an adaptive navigation system with multi-source sensor fusion, comprising: The multi-source sensor spatiotemporal synchronization module is used to align timestamps and convert spatial coordinates of raw data from lidar, visual cameras, inertial measurement units, and satellite positioning sensors, and output a standardized data stream that is spatiotemporally aligned. The multi-source sensor spatiotemporal synchronization module is used to eliminate temporal and spatial deviations between lidar, visual cameras, inertial measurement units, and satellite positioning sensors. This module first establishes a sensor clock deviation model and calibrates the hardware clock using a bidirectional precision time protocol, exemplified by the IEEE 1588 PTP protocol. The master clock node generates a global time reference based on the inertial measurement unit's crystal oscillator signal, and the slave clock nodes calculate clock offsets using a delayed request-response mechanism, satisfying the following relationship: where t send With t recv are the timestamps of sending and receiving clock synchronization signals, Δt k represents the deviation of the kth sensor relative to the master clock. This deviation is dynamically compensated by the Kalman filter, achieving sub-microsecond synchronization accuracy.

[0023] For the time alignment problem of asynchronous data streams, such as the synchronization between the 10Hz scanning data of the lidar and the 1MHz event stream of the event camera, the dynamic time warping algorithm is further adopted. First, the cost matrix C is constructed. i,j , whose elements are defined as the lidar point cloud p i With event stream j The space-time distance: C i,j =||p i -e j ||2+λ·|t i -t j |; Where λλ is the time dimension weight coefficient, and the preferred value range is 0.5≤λ≤1.5. Then the optimal alignment path W is solved by dynamic programming. * , its objective function is: Where γ controls the path smoothness and Length(W) represents the path length penalty. The algorithm can achieve sub-pixel alignment accuracy under non-uniform sampling rates.

[0024] The spatial coordinate unification is achieved through the Lie group SE(3) transformation, and the transformation matrix from each sensor coordinate system to the base coordinate system is defined as: Among them G m is the generator of SE(3) Lie algebra, corresponding to translation x, y, z and rotation R x , Ry , R z Six degrees of freedom, ξ m is the calibration parameter. The calibration process is obtained by optimizing the joint observation data of multiple sensors on the checkerboard calibration plate, and the objective function is: in To calibrate the known coordinates of the plate corners in the global coordinate system, the optimization algorithm preferably uses the Levenberg-Marquardt nonlinear least squares method, which can effectively handle the coupling error of the sensor external parameters.

[0025] For devices with inherent latency, such as mechanical scanning LiDAR, a motion estimation model is introduced to compensate for timestamp interpolation. Assuming that the sensor's motion in the time interval [t0, t1] is a uniform acceleration model, the pose corresponding to the actual sampling time t can be calculated using the following formula: Where v(t0) and a(t0) are the initial velocity and acceleration, respectively, interpolated from the inertial measurement unit data. This compensation method can eliminate millisecond-level time deviations caused by the sensor's internal scanning mechanism.

[0026] The resulting standardized data stream meets spatiotemporal consistency requirements: LiDAR point clouds, visual image feature points, inertial measurement unit (IMU) data, and satellite positioning information are aligned to the same reference frame in the temporal dimension, and spatial coordinates are unified to a base coordinate system centered on the IMU. This module uses these technical approaches to overcome the fusion bottleneck caused by the heterogeneity of multi-source sensors, providing reliable input for subsequent positioning and navigation modules.

[0027] The quantum particle filter positioning estimation module receives standardized data streams and generates pose estimation results in dynamic environments through quantum state encoding and annealing optimization; The quantum particle filter positioning estimation module is used to achieve highly robust pose estimation in dynamic environments. This module first encodes the particle state in the classical particle filter framework into a quantum superposition state. For example, it simulates quantum effects through a quantum-classical hybrid computing architecture, which can suppress particle degradation and explore the state space in parallel. In specific implementation, the quantum state of the i-th particle is defined as: |ψ i >=α i |0>+β i |1>, and |α i | 2 +|β i | 2 =1; Where |0> represents the position confidence quantum state, and its amplitude α i squared|αi | 2 Characterizes the credible probability of the particle position hypothesis; |1> represents the confidence quantum state of the motion state, β i squared|β i | 2 Characterizing the credible probability of velocity and posture. This encoding method introduces quantum superposition properties, allowing a single particle to simultaneously represent the credibility distribution of multi-dimensional states.

[0028] The quantum evolution of particle states is achieved by constructing a Hamiltonian model. The system Hamiltonian HH is defined as consisting of the observation likelihood term and the inter-particle correlation term: where p(z t |x i ) is the sensor observation data z t In particle state x i Likelihood probability under , γ is the correlation strength coefficient, J ij Represents the association weight between particles: Where σ is the particle distribution scale parameter, preferably within the range of 1-2 times the sensor measurement noise variance. This Hamiltonian model optimizes the particle distribution through a quantum annealing process, giving higher weight to particles with higher likelihood probabilities while suppressing the divergence of isolated particles.

[0029] The quantum annealing optimization process is achieved by simulating the quantum tunneling effect. First, the annealing schedule function s(t) is initialized, and its value range decreases from 1 to 0. For example, a linear annealing strategy is used: Where T max is the maximum annealing time. In each annealing iteration, the quantum state evolves according to the time-dependent Schrödinger equation: Solving the ground state |ψ by quantum Monte Carlo simulation final >, specifically using the path integral method to calculate the time evolution path of the quantum state, which can achieve global optimization of particle weights.

[0030] The resampling process is completed through quantum projection measurement. final >Perform projection operation and calculate the normalized weight of each particle: Then according to the weight Importance resampling is performed to retain high-weight particles and remove low-weight particles. This process allows some low-weight particles to be retained with a certain probability through the quantum tunneling effect, thus avoiding the sample depletion problem caused by classical resampling.

[0031] The pose estimation results are output through weighted averaging: where x i =[p i ,v i ,q i ] contains position, velocity, and attitude quaternions. Through quantized particle state representation and annealing optimization, this module can effectively handle non-Gaussian noise and multimodal distribution problems, providing high-confidence position input for subsequent modules.

[0032] In the above implementation, the quantum-classical hybrid computing architecture is preferably implemented on the edge computing device, and the quantum state amplitude and Hamiltonian evolution are calculated in parallel by the GPU tensor core, which can meet the real-time requirements. The number of quantum bits required for particle state encoding is determined by the particle dimension. For example, for 6-DOF posture estimation, each particle corresponds to 12 quantum bits (3 degrees of freedom each for position and velocity, and 4 dimensions for attitude quaternion). This module is directly connected to the output data stream of the spatiotemporal synchronization module, and receives the multi-source sensor observation value z after spatiotemporal alignment. t Used to calculate the likelihood probability and pass the pose estimation results to the meta-learning controller module to drive the weight allocation.

[0033] The spatiotemporal meta-learning controller module dynamically calculates the fusion weight coefficients of each sensor based on the pose estimation results and environmental characteristic parameters; The spatiotemporal meta-learning controller module dynamically adjusts the multi-sensor fusion weights based on environmental characteristics and ensures that the fusion process complies with the laws of physical motion. This module first receives pose estimation results from the quantum particle filter localization module and simultaneously collects environmental observation data from multiple sensors. It then generates weight coefficients through feature extraction and a meta-learning network, exemplarily combining physical kinematic constraints to enhance the robustness of the fusion strategy.

[0034] Environmental feature extraction includes point cloud density entropy, image texture entropy, and satellite signal quality indicators. For lidar point cloud data, the space is divided into voxel grids and the Shannon entropy is calculated: where p v is the percentage of point clouds within the vth voxel, and V is the total number of grids. This entropy value reflects the complexity of the environment structure. A high entropy value indicates that there are many occluders or sparse geometric features. For visual images, the texture entropy is calculated by the gray-level co-occurrence matrix: S image =-∑ i,j P(i,j)logP(i,j); Where P(i,j) is the co-occurrence probability of grayscale values ​​i and j, reflecting the richness of image texture. Satellite signal quality index N GPSA weighted combination of carrier-to-noise ratio and number of visible satellites is preferred.

[0035] The meta-learning network uses the aforementioned feature vector f c =[D point ,S image ,N GPS ] as input and output the fusion weight w of each sensor s The network structure exemplarily includes three fully connected layers: Input layer: The dimension is the same as the length of the feature vector, and the activation function is ReLU; Hidden layer: The number of neurons is preferably 2-4 times the dimension of the input layer, and batch normalization and dropout regularization are used; Output layer: Softmax function normalizes weights to satisfy ∑w s =1.

[0036] Physical constraints are injected into the network training process by modifying the loss function. The total loss function is defined as: in is the mean square error between the weighted predicted value and the true value, is the rate of change of network output, a IMU is the measured acceleration of the inertial measurement unit, g is the gravitational acceleration vector, and λ is the constraint weight coefficient. This constraint forces the motion state output by the network to conform to Newton's laws of motion and can suppress abnormal weight assignments caused by sensor noise or environmental interference.

[0037] The dynamic weight allocation process is executed in real time and includes the following steps: 1. Obtain multi-sensor data from the spatiotemporal synchronization module and extract environmental features f c ; 2. Input the feature vector into the pre-trained meta-learning network to obtain the initial weight 3. Estimated value based on current pose Calculate physical constraints and fine-tune network parameters through back propagation; 4. Output optimized weights To the cross-modal quantum fusion module.

[0038] The network training adopts an incremental learning strategy, preferably using the Adam optimizer to update parameters, and the learning rate decays exponentially with the training rounds: η t =η0·exp(-κt); Where η0 is the initial learning rate, κ is the decay coefficient, and t is the current training round. Training data is dynamically updated using a sliding window mechanism, with the window size preferably set to 100-200 frames of historical data to balance computational efficiency and environmental adaptability.

[0039] This module interacts bidirectionally with the quantum particle filter module: it receives pose estimation results for physical constraint calculations and feeds weight coefficients into the particle filter's observation likelihood function, illustratively adjusting the contribution of different sensors to particle weight updates. In abnormal situations such as satellite signal loss, the network enforces an increase in inertial navigation weights through constraints, maintaining the system's basic positioning capabilities.

[0040] In the above embodiment, the gradient calculation of the physical constraint term is realized by automatic differentiation. Specifically, Expressed as the time difference of the velocity vector output by the network: Where Δt is the data update period. This differential operation is embedded in the computational graph, allowing the constraint terms to participate in end-to-end backpropagation, ensuring that the network simultaneously optimizes weight prediction accuracy and conformity to physical laws.

[0041] The cross-modal quantum fusion module performs quantum entanglement feature-level fusion on the lidar point cloud, visual image, and inertial data according to the fusion weight coefficient, and outputs the enhanced navigation state; The cross-modal quantum fusion module is used to achieve feature-level fusion of heterogeneous sensor data, such as lidar, visual cameras, and inertial measurement units. This module first maps data from different modalities into quantum state space, establishes cross-modal correlations through quantum entanglement channels, and exemplarily implements feature fusion using controlled quantum gate operations. This module can address the problem of modal differences that are difficult to handle with traditional methods.

[0042] Quantum state mapping: For the lidar point cloud data P, its three-dimensional coordinates and reflection intensity are encoded into quantum states: Where (x n ,y n , z n ) is the point cloud coordinate, r n is the reflection intensity, k is the wave number parameter, which is preferably inversely proportional to the laser wavelength. The event camera data stream E is mapped to the quantum state: Where (u m , v m ) is the pixel coordinate, t m The event trigger timestamp, p m is the polarity (+1 / -1), and w is the time frequency parameter.

[0043] Quantum entanglement channel construction: Apply a controlled NOT (CNOT) operation to the lidar state |P> and the event state |E>: Among them, f(P) is the point cloud geometric feature extraction function, which extracts the edge point set as an example τ is the reflection intensity gradient threshold, Represents a bitwise exclusive-or operation. This operation associates the qubits of the event flow |E> with the edge features of the point cloud, forming an inseparable entangled state.

[0044] Quantum purity calculation: The fused quantum state density matrix is: ρ=|Ψ><Ψ|, where The purity index ε is calculated by trace square: ε=1-Tr(ρ 2 ); When ε is lower than the preset threshold ∈, it indicates that the cross-modal correlation is weakened and the fusion strategy needs to be adjusted.

[0045] Dynamic fusion ratio adjustment: The fusion weight of lidar and event data is adjusted in real time according to the purity ε: F fused =ε·F LiDAR +(1-ε)·F Event ; Among them F LiDAR is the point cloud feature vector, which exemplarily includes the edge point density and curvature distribution; F Event is the event stream feature vector, preferably a spatiotemporal gradient histogram. This mechanism can automatically enhance the contribution of highly reliable modes when the environment changes suddenly.

[0046] Inertial data fusion: Inertial measurement unit data a IMU Injection fusion process through quantum phase encoding: in is the accumulated acceleration phase. Perform quantum exchange operation on |I> and |Ψ>: This operation fuses the high-frequency characteristics of inertial data with the spatial characteristics of vision-lidar, outputting an enhanced navigation state |Ψ final >.

[0047] In the above implementation, the quantum gate operation is preferably simulated and implemented on a quantum-classical hybrid computing architecture, and the quantum state evolution is calculated in parallel by the GPU tensor core. The module receives the fusion weight w sent by the spatiotemporal meta-learning controller. s , which is encoded as quantum rotation gate parameters: Apply R to the corresponding quantum bit y (θ k) rotates, making the weight coefficient directly affect the entanglement strength. This module forms a closed loop with the adaptive navigation control module. When it detects that ε < ∈, it sends a redundant activation signal to the control module to trigger the backup sensor data fusion.

[0048] The time complexity of the quantum fusion process is reduced by sparse matrix optimization. For example, the point cloud and event data are downsampled to a voxel grid, with the grid size preferably being 2-3 times the spatial resolution of the sensor. Before the fusion result is transmitted to the adaptive navigation module, the classical eigenvector is obtained by quantum state projection: F out =<Ψ final |M|Ψ final > Where M is a feature extraction operator, and its matrix elements are determined by a pre-trained feature selection model. This module achieves the essential correlation of cross-modal data through quantum entanglement, resolving the feature misalignment problem caused by modal differences in traditional methods.

[0049] The adaptive navigation control module generates anti-interference control instructions based on the enhanced navigation state and feeds them back to the spatiotemporal meta-learning controller module in real time to trigger weight updates.

[0050] The adaptive navigation control module generates anti-interference control commands based on the fused navigation state and provides real-time feedback on environmental change signals to trigger system parameter updates. This module first establishes a model predictive control framework and generates an optimal control sequence through rolling horizon optimization. It also incorporates quantum fusion purity metrics to implement a dynamic fault-tolerance mechanism capable of handling abnormal operating conditions such as sensor failure or environmental changes.

[0051] Model predictive control: Construct an objective function to minimize the trajectory tracking error and the control amount: Where Q is the state error weight matrix, R is the control quantity weight matrix, is the predicted state at time t+k, x ref,k is the reference trajectory. The optimization problem is solved by quadratic programming, and the effective set algorithm is used to iteratively calculate the optimal control sequence.

[0052] Quantum purity trigger mechanism: Receive the quantum purity ε output by the cross-modal quantum fusion module, and determine it as an environmental mutation when ε < ∈. The trigger condition is defined as: ε=1-Tr(ρ 2 )<∈; Where ρ is the quantum fusion state density matrix and ∈ is the preset threshold. This condition triggers the redundant sensor activation strategy, exemplarily starting the combined navigation mode of the UWB positioning module and the visual odometry, and adjusting the weight update frequency of the spatiotemporal meta-learning controller.

[0053] Control instruction generation: The first item of the optimal control sequence Converted into actuator instructions. For a wheeled mobile platform, the linear velocity and angular velocity instructions are decomposed into left and right wheel speeds: Where v is the linear velocity, w is the angular velocity, d is the wheelbase, and r is the wheel radius. This command is sent to the motor controller via the CAN bus to achieve precise trajectory tracking.

[0054] Feedback closed-loop design: real-time monitoring of control errors when ||e t When ||2>δ, a feedback signal f is sent to the spatiotemporal meta-learning controller b : This signal triggers the incremental training of the meta-learning network and dynamically adjusts the sensor fusion weights w s , forming a "control-perception-learning" closed loop.

[0055] Fault-tolerant recovery strategy: During the activation of redundant sensors, the fusion weights are gradually adjusted to avoid sudden changes in instructions: Where γ is the smoothing coefficient, and the preferred value satisfies 0<γ<1, w backup is the preset redundant sensor weight. This process continues until the quantum purity ε≥∈, after which the normal fusion mode is restored.

[0056] Control parameter adaptation: Dynamically adjust the prediction time domain TT according to the complexity of the environment. The adjustment rule is defined as: T = T base +η·D point ; Where T base is the basic time domain length, η is the scaling factor, D point is the point cloud density entropy. This mechanism extends the optimization time domain in complex environments to improve stability, and shortens the time domain in open environments to reduce the computational load.

[0057] In the above embodiment, the constraints of the quadratic programming solver include the motor torque limit and the platform kinematic constraints: min ≤u k ≤u max ; ||v k ||≤v max ,||w k ||≤w max ; where v max With w maxare the maximum linear and angular velocities. Constraint processing is achieved through projection method to ensure the physical feasibility of control instructions.

[0058] The data interface between the module and the quantum fusion module transmits quantum purity ε in real time, forming a bidirectional coupling with the feedback channel of the spatiotemporal meta-learning controller. When a persistent low-purity state is detected, a system-level restart protocol is triggered, which reinitializes the particle filter and meta-learning network parameters and switches to a safe obstacle avoidance mode.

[0059] Before the control command is output, high-frequency noise is eliminated through a low-pass filter: u out =u * *h(t); Where h(t) is an exponential decay filter whose time constant is preferably matched to the platform inertia parameters. This module uses the above technical means to achieve stable navigation control in dynamic environments and ensure the system's autonomous recovery capability under abnormal conditions.

[0060] Another embodiment of the present invention provides an adaptive navigation method using multi-source sensor fusion, comprising the following steps: S1, time synchronization and spatial coordinate unification of multimodal sensor data; S2, realize pose estimation in dynamic environment through quantum particle filtering; S3, dynamically assign sensor fusion weights based on environmental characteristics and physical constraints; S4. Using quantum entanglement mechanism to achieve cross-modal feature-level fusion; S5. Generate anti-interference control instructions and feedback environmental mutation signals to trigger weight updates.

[0061] 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. An adaptive navigation system based on multi-source sensor fusion, characterized in that: include: The multi-source sensor spatiotemporal synchronization module is used to align timestamps and convert spatial coordinates of raw data from lidar, visual cameras, inertial measurement units, and satellite positioning sensors, and output a standardized data stream that is spatiotemporally aligned. A quantum particle filter positioning estimation module receives the standardized data stream and generates a pose estimation result in a dynamic environment through quantum state encoding and annealing optimization; A spatiotemporal meta-learning controller module dynamically calculates the fusion weight coefficients of each sensor based on the pose estimation results and environmental characteristic parameters; A cross-modal quantum fusion module performs quantum entanglement feature-level fusion on the lidar point cloud, visual image, and inertial data according to the fusion weight coefficient, and outputs an enhanced navigation state; An adaptive navigation control module generates anti-interference control instructions based on the enhanced navigation state and feeds back the instructions to the spatiotemporal meta-learning controller module in real time to trigger weight update.

2. The adaptive navigation system of multi-source sensor fusion according to claim 1, characterized in that: The spatial coordinate transformation in the multi-source sensor spatiotemporal synchronization module is implemented by the Lie group SE (3), and its transformation matrix is ​​defined as: in: G m (m=1,2,...,6) are the generators of the SE(3) Lie algebra, corresponding to the six degrees of freedom of spatial motion; ξ m are the calibration parameters obtained by optimizing the characteristic points of the calibration plate; exp(·) represents the Lie group exponential mapping operation.

3. The adaptive navigation system of multi-source sensor fusion according to claim 1, characterized in that: The quantum state encoding in the quantum particle filter positioning estimation module satisfies: |\psi i >=\alpha i |0>+\beta i |1>, and |\alpha i | 2 +|\beta i | 2 =1; in: |ψ i > is the quantum superposition state of the i-th particle, |0> represents the position credible quantum state, and its amplitude α i squared|α i | 2 Characterizes the position confidence probability, |1> represents the credible quantum state of the motion state, and its amplitude β i squared|β i | 2 Characterize the confidence probability of the motion state.

4. The adaptive navigation system of multi-source sensor fusion according to claim 1, characterized in that: The physical constraints in the spatiotemporal meta-learning controller module are implemented through the loss function: is the mean square error loss of sensor weight prediction, is the time derivative of the velocity vector, calculated by automatic differentiation, a IMU is the acceleration measurement value output by the inertial measurement unit, g is the gravitational acceleration vector, and λ is the weight coefficient of the physical constraint term, with a value range of [0.1, 1.0].

5. The adaptive navigation system of multi-source sensor fusion according to claim 1, characterized in that: The quantum entanglement channel in the cross-modal quantum fusion module is constructed by a controlled NOT gate, which satisfies: in: |P> is the quantum state mapped by the lidar point cloud P; |E> is the quantum state mapped by the event camera data stream E; f(P) is the point cloud geometric feature extraction function, and the output is a binary feature vector; ⊕ represents the bitwise exclusive OR operation.

6. The adaptive navigation system of multi-source sensor fusion according to claim 1, characterized in that: The fault-tolerant triggering conditions of the adaptive navigation control module are: ε=1-Tr(ρ 2 )<∈; in: ρ is the density matrix of the quantum fusion state, which is calculated by ρ=|Ψ><Ψ|; Tr(ρ 2 ) is the purity of the density matrix; ∈ is the preset quantum purity threshold.

7. The adaptive navigation system of multi-source sensor fusion according to claim 3, characterized in that: The quantum state encoding is implemented through a quantum-classical hybrid computing architecture, specifically: On edge computing devices, quantum bit operations are simulated through GPU tensor cores to encode particle states as quantum state amplitudes.

8. The adaptive navigation system of multi-source sensor fusion according to claim 4, characterized in that: The physical kinematic equation constraints include: The motion acceleration equation of the inertial measurement unit is used as a hard constraint and injected into the loss function of the meta-learning network through the Lagrange multiplier method.

9. The adaptive navigation system of multi-source sensor fusion according to claim 5, characterized in that: The correlation strength of the quantum entanglement channel is dynamically controlled by the density matrix purity. Specifically, the fusion ratio of the lidar and visual data is adjusted according to the quantum state purity EE: F fused =ε·F LiDAR +(1-e) F Event ; Among them F LiDAR With F Event are the feature vectors of the lidar and event camera, respectively.

10. An adaptive navigation method for multi-source sensor fusion, according to the adaptive navigation system for multi-source sensor fusion according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1, time synchronization and spatial coordinate unification of multimodal sensor data; S2, realize pose estimation in dynamic environment through quantum particle filtering; S3, dynamically assign sensor fusion weights based on environmental characteristics and physical constraints; S4. Using quantum entanglement mechanism to achieve cross-modal feature-level fusion; S5. Generate anti-interference control instructions and feedback environmental mutation signals to trigger weight updates.

Images