Cross-scene active reasoning intelligent decision-making and consensus fusion navigation method and device

Abstract

The invention provides a cross-scene active reasoning intelligent decision and consensus fusion navigation method and device, and the method comprises the steps: carrying out the preliminary fusion of the information of each navigation subsystem, inputting the information into an evaluation layer, employing a generation model constructed based on a neural network by the evaluation layer, achieving the credibility evaluation of the navigation subsystems, and outputting the likelihood probability; the decision-making layer is realized by adopting a neural network, training is carried out by adopting reinforcement learning, the likelihood probability of each navigation subsystem is input, and a decision is output as a fusion weight for carrying out re-fusion on the preliminary fusion pose; in training, a decision-making layer determines expected free energy based on credibility of a navigation subsystem and accuracy of a fusion track, and network parameters of the decision-making layer are optimized by taking minimization of the expected free energy as a reward target; and carrying out weighted fusion on the initial fusion poses of the navigation subsystems by adopting the optimal fusion weight to obtain a fusion track. By using the method of the invention, higher precision and robustness can be maintained in a dynamic and uncertain environment.

Description

Technical Field

[0001] This invention belongs to the field of multi-source fusion navigation technology, specifically relating to a cross-scenario active reasoning intelligent decision-making and consensus fusion navigation method and device. Background Technology

[0002] Intelligent decision-making in complex environments is a key aspect of enhancing the autonomous navigation capabilities of unmanned platforms. Its main objective is to generate safe, efficient, and robust action strategies by integrating multi-source sensor information and mission objectives in dynamic and uncertain cross-scenario tasks. Due to the vast state space of the navigation environment, the uncertainty of sensor information, and the diversity of mission objectives, achieving high-level intelligent decision-making remains one of the most challenging problems in the field of autonomous navigation systems.

[0003] With the rapid development of artificial intelligence, the current mainstream technical solutions mainly revolve around reinforcement learning frameworks, constructing a decision-making model from environmental perception to action output. The agent optimizes its decision-making strategy by obtaining reward signals through interaction with the environment, determining the optimal navigation action to take under different observation states, and ultimately achieving the goal of completing a specific navigation task. Its main principle is to approximate the state value function or policy function through deep neural networks, continuously adjusting network parameters using optimization algorithms such as gradient descent to maximize long-term cumulative rewards. However, when applied to complex cross-scene navigation in the real world, reinforcement learning-based decision-making schemes still face many unresolved problems. One of the main bottlenecks is that in partially observable real-world scenarios with dynamic interference, due to the limitations of the algorithm framework and decision-making mechanism itself, the intelligent decision-making process still faces several problems, including insufficient assessment of the credibility of perceived information, low efficiency in exploring unknown environments, and weak generalization ability for cross-scene tasks. Therefore, through reasonable framework design and algorithm integration, explicitly modeling and reasoning about the cognitive uncertainty in the decision-making process, and adaptively adjusting the decision-making strategy based on the reasoning results, is key to achieving a higher level of intelligent navigation.

[0004] Current research on improving the uncertainty and generalization ability of navigation decision-making models mostly focuses on network structure optimization or reward function design, rather than proposing a specific technical framework that can endogenously integrate perceived uncertainty into decision-making logic and actively guide exploration behavior based on cognitive science principles. Summary of the Invention

[0005] In view of this, the present invention provides a cross-scenario proactive reasoning intelligent decision-making and consensus fusion navigation method and device, which can maintain higher accuracy and robustness in dynamic and uncertain environments.

[0006] To solve the above-mentioned technical problems, the present invention is implemented as follows.

[0007] A cross-scenario proactive reasoning intelligent decision-making and consensus fusion navigation method includes: Step 1: Collect observation information from multiple navigation sources; Step 2: Select one of the navigation sources as the basic navigation source, and form navigation subsystems in pairs with other navigation sources; perform preliminary fusion of information from each navigation subsystem to obtain a preliminary fused pose; Step 3: Input the initial fusion pose of all navigation subsystems into the evaluation layer. The evaluation layer adopts a generative model based on neural network to evaluate the reliability of the navigation subsystems and output the likelihood probability. Step 4: The decision layer is implemented using a neural network and trained using reinforcement learning. The input is the likelihood probability of each navigation subsystem, and the output decision is the fusion weight for re-fusion of the initial fused pose. During training, the decision layer determines the expected free energy based on the reliability of the navigation subsystems and the accuracy of the fused trajectory. To minimize the expected free energy To achieve the reward objective, optimize the network parameters of the decision-making layer; Step 5: Using the fusion weights determined in Step 4, perform weighted fusion on the preliminary fusion poses of each navigation subsystem to obtain the fusion trajectory.

[0008] Preferably, the error state iterative extended Kalman filter (IESKF) is used to initially fuse the information of each navigation subsystem.

[0009] Preferably, step 1 involves acquiring observation information from multiple navigation sources, including asynchronous raw data streams from GNSS receivers, LiDAR, cameras, and IMUs. In step 2, the IMU is selected as the basic navigation source, and the basic navigation source is initially fused with other navigation sources to obtain the initial fused pose.

[0010] Preferably, the input to the decision layer is the likelihood probability of each navigation subsystem output by the evaluation layer, and the output of the decision layer is the fusion weight; The reward is constructed using the expected free energy; the smaller the expected free energy, the higher the reward. The expected free energy is characterized by an accuracy term. and ambiguity characterization terms Composition; among which, The accuracy characterization item For time windows The Euclidean error between the fused trajectory and the true trajectory calculated based on the fusion weights of the decision-based calculation; The ambiguity characterization term The covariance matrix of the IESKF filter is multiplied by the likelihood probability and then applied in a time window. The inner summation yields:

[0011] in, The total number of navigation subsystems, for Time navigation subsystem The covariance matrix corresponding to the IESKF filter, For the output of the evaluation layer Time navigation subsystem The corresponding likelihood probability.

[0012] Preferably, the action space of the decision-making layer is set as follows: The weights of each navigation subsystem are set to range from [0,1] with a step size of 0.1, and the sum of the weights of all navigation subsystems is 1. Under the constraints of the weight range, step size, and weight sum, the fusion weight combination is determined, and a discrete action space composed of different fusion weight combinations is constructed as the action space of the decision layer.

[0013] Preferably, the generative model employs an LSTM network.

[0014] Preferably, the method further includes: employing an active reasoning variational free energy minimization mechanism to dynamically optimize the evaluation capability of the generative model, including: Based on the preliminary fused pose of each navigation subsystem, the prior probability is determined. ; Based on the prior probability The likelihood probability determined in step 3 The true posterior probability is calculated using the Bayesian inference process formula. ; The mean and variance are calculated based on the node parameters of the last layer of the generative model. A Gaussian distribution is then constructed using these mean and variance as the approximate posterior probability. ; Using the approximate posterior probability and true posterior probability Calculate variational free energy ; To minimize variational free energy As an optimization objective, the network parameters of the generative model are adjusted to improve the ability to assess the credibility of navigation sources.

[0015] Preferably, the prior probability The determination method is as follows: the initial fused pose of each navigation subsystem is modeled using the credible prior modeling DoO to obtain the prior probability. .

[0016] Preferably, the method further includes: performing feedback correction on the parameters of the algorithm used for preliminary fusion based on the fusion trajectory obtained in step 5.

[0017] The present invention also provides a cross-scenario proactive reasoning intelligent decision-making and consensus fusion navigation device, including a perception layer, an evaluation layer, a decision layer, a fusion trajectory generation module and a feedback correction module; The perception layer is used to acquire observation information from multiple navigation sources; select one navigation source as the basic navigation source, and form navigation subsystems in pairs with other navigation sources; perform preliminary fusion of information from each navigation subsystem to obtain a preliminary fused pose; The evaluation layer is used to evaluate the reliability of the navigation subsystem. The evaluation layer is implemented using a generative model based on a neural network. It takes the preliminary fused pose of each navigation subsystem as input and outputs the likelihood probability of each navigation subsystem. The decision layer is implemented using a neural network and trained using reinforcement learning. Its input is the likelihood probability of each navigation subsystem, and its output is the fusion weights for re-fusion of the initial fused pose. During training, the decision layer determines the expected free energy based on the reliability of the navigation subsystems and the accuracy of the fused trajectory. To minimize the expected free energy To achieve the reward objective, optimize the network parameters of the decision-making layer; The fusion trajectory generation module is used to perform weighted fusion of the preliminary fusion poses of each navigation subsystem using the optimal fusion weights determined by the decision layer, in order to obtain the fusion trajectory. The feedback correction module connects the perception layer and the evaluation layer, and determines the prior probability based on the preliminary fused pose of each navigation subsystem. Based on the prior probability Likelihood probability The true posterior probability is calculated using the Bayesian inference process formula. Using the approximate posterior probability and true posterior probability Calculate variational free energy To minimize variational free energy As an optimization objective, the network parameters of the generative model are adjusted to improve the ability to assess the credibility of navigation sources.

[0018] Beneficial effects: First, by explicitly integrating cognitive uncertainty into decision-making, the intelligence and long-term benefits of decision-making are improved by balancing "exploration-exploitation". By introducing the minimization of expected free energy as a decision criterion, its form naturally includes two aspects: "risk (difference from expected goals)" and "ambiguity (information gain)". This allows the system to not only consider current accuracy when selecting fusion weights, but also to actively choose strategies that can reduce the uncertainty of future environments, thus achieving better long-term navigation performance.

[0019] Second, it achieves endogenous and dynamic evaluation of perceived uncertainty, improving the system's real-time robustness. By introducing a variational free energy minimization mechanism based on active reasoning, the generative model is optimized, thereby quantifying the credibility of each navigation source into a dynamically changing continuous value. When a navigation source signal fails or its performance degrades, its free energy increases sharply, and the system can immediately and automatically reduce its credibility weight, effectively preventing unreliable information from contaminating the fusion results, thus significantly improving navigation robustness under dynamic interference and cross-scenario environments.

[0020] Third, it provides a more interpretable and general intelligent navigation framework that aligns with cognitive principles, enhancing the system's generalization ability. The constructed three-layer active reasoning framework of "perception-evaluation-decision" endows the system with a reasoning ability similar to "common sense," enabling it to handle complex, unseen scenarios based on first principles. This reduces reliance on large amounts of scenario data, thus demonstrating stronger cross-scenario and cross-task generalization capabilities. Furthermore, the decision-making process, based on probabilistic reasoning and explicit optimization objectives, offers greater interpretability compared to "black box" neural networks. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the cross-scenario proactive reasoning intelligent decision-making and consensus fusion navigation method of the present invention.

[0022] Figure 2 This is a schematic diagram of observation information from multiple navigation sources.

[0023] Figure 3 The flowchart for the IESKF filter.

[0024] Figure 4 This is a structural diagram of the active reasoning framework.

[0025] Figure 5 This is a block diagram of the cross-scenario proactive reasoning intelligent decision-making and consensus fusion navigation device of the present invention. Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] This invention addresses the shortcomings of existing intelligent navigation decision-making methods, such as the lack of assessment of perceived uncertainty, insufficient modeling of cognitive uncertainty, and weak cross-scenario generalization ability. It provides a cross-scenario proactive reasoning intelligent decision-making and consensus fusion navigation scheme. The core idea is to use proactive reasoning as a theoretical framework. Based on acquiring multi-source observation information at the perception layer, the reliability of navigation sources is dynamically assessed through a generative model at the evaluation layer. Furthermore, at the decision layer, the optimal fusion weights are selected by minimizing the expected free energy. Finally, adaptive error correction of local subsystems is achieved in the feedback loop. This method realizes a closed loop of "perception-evaluation-decision-correction" in a hierarchical structure, enabling multi-source navigation systems to maintain higher accuracy and robustness in dynamic and uncertain environments.

[0028] To achieve the above solution, such as Figure 1 As shown, this invention constructs a perception layer based on the fusion of multiple sensors (GNSS, LiDAR, and cameras) and an inertial navigation system (INS), providing a unified and reliable underlying state estimate for decision-making. It proposes an evaluation layer based on an active reasoning generative model, which dynamically optimizes the credibility of each navigation source (represented by likelihood probability) by calculating variational free energy, achieving an endogeneous evaluation of perceptual uncertainty. Finally, it proposes a decision layer based on a reinforcement learning fusion framework, which adaptively allocates the fusion weights of each navigation source by minimizing the expected free energy, ultimately generating the final trajectory. This invention explicitly integrates cognitive uncertainty into the decision-making logic to balance exploration and utilization, providing unmanned platforms with a general intelligent navigation decision-making framework that conforms to the principles of cognitive science and can proactively cope with uncertainty.

[0029] To achieve the above solution, the following will be combined with Figure 1 The embodiments of the present invention will be further described in detail below. The cross-scenario proactive reasoning intelligent decision-making and consensus fusion navigation method of the present invention includes the following steps: Step 1: The perception layer collects observation information from multiple navigation sources.

[0030] Step 11: Receive and parse the raw data from each sensor.

[0031] In this embodiment, the multiple navigation sources include GNSS, LiDAR, camera, and IMU. Therefore, asynchronous raw data streams from the GNSS (Global Navigation Satellite System) receiver, LiDAR, camera, and IMU are received simultaneously, and their data frames are parsed, such as... Figure 2 As shown.

[0032] GNSS: Extract latitude and longitude With speed value ; LiDAR: Receiving point cloud data ; Camera: Receives image frames ; IMU: Extracting triaxial angular increments and triaxial acceleration .

[0033] Step 12: Perform time synchronization and coordinate system one.

[0034] Within the ROS (Robot Operating System) framework, INS time is used as the reference. Let the INS output time be... Other sensors at adjacent times and The observed values ​​are respectively , Then, by linear interpolation, the value at time t is obtained. Observations:

[0035] Align all sensor data to the INS timeline according to the above formula.

[0036] Regarding the coordinate system, to ensure that multi-source observation information can be fused under a unified physical meaning, the output results of each sensor need to be unified to the same reference coordinate system. The East-North-Up (ENU) coordinate system is chosen as the unified coordinate system, and all observations from GNSS, LiDAR, camera, and INS need to be transformed to this coordinate system. The IMU coordinate system serves as an intermediate bridge for the data from each sensor. During the system calibration phase, the extrinsic parameters (rotation matrix and translation vector) of each sensor coordinate system relative to the IMU coordinate system are obtained. Thus, the raw observations output by sensor s (s∈{GNSS, LiDAR,Camera}) are... The coordinates can be transformed to the IMU coordinate system using the following relationship:

[0037] in, It is the rotation matrix from the sensor coordinate system to the IMU coordinate system. It is a translation vector.

[0038] Subsequently, based on the real-time attitude and position calculated by INS, a transformation matrix from the IMU coordinate system to the ENU coordinate system is constructed to achieve the unification of observation data from all sensors.

[0039] The conversion process varies depending on the sensor used. GNSS: The raw output of GNSS is latitude, longitude, and elevation. First, its three-dimensional coordinates in the geocentric-earth-fixed (ECEF) coordinate system are calculated using the WGS84 ellipsoid model. :

[0040] Where N is a radius of curvature parameter of the Earth determined by latitude, representing the radius of curvature of the zonal circle.

[0041] Then, using the set reference point Using the origin as the coordinate system, the coordinates are transformed from ECEF to ENU to obtain the final coordinates of the GNSS observations in the ENU system.

[0042] LiDAR: Each point in the point cloud is represented in the LiDAR coordinate system as follows: Then, it is transformed to the IMU coordinate system through the extrinsic parameter matrix, and then unified to the ENU coordinate system through the pose information of the INS.

[0043] Camera: The camera pose is obtained by solving the image frames output by the camera through visual odometry or SLAM algorithms. The pose is first transformed to the IMU coordinate system through an extrinsic parameter matrix, and then unified to the ENU coordinate system through the pose information of the INS, thus obtaining the visual observation in the ENU.

[0044] Step 2: Initial fusion of multi-navigation source information in the perception layer.

[0045] In this step, INS is selected as the basic navigation source and paired with other navigation sources to form navigation subsystems; the information of each navigation subsystem is initially fused to obtain the initial fused pose.

[0046] In a preferred scheme, to ensure the real-time performance and numerical stability of the fusion algorithm, an error state iterative extended Kalman filter (IESKF) is used to fuse the states recursively derived from INS with external observations. The basic process is as follows: Figure 3 As shown, IESKF uses the attitude, velocity, and position obtained from INS inertial mechanical orchestration as the nominal state prediction of the system, and performs filtering updates in the error state space, thereby avoiding numerical problems caused by directly operating on nonlinear quantities such as quaternions. Simultaneously, it uses the position and velocity calculated by GNSS, the incremental pose calculated by LiDAR through point cloud matching, and the pose and velocity calculated by the camera through visual odometry as observation values, and updates them with each observation.

[0047]

[0048] in, Kalman gain; For at any time The received observation vectors. For example, the position and velocity given by GNSS at this moment, or the pose calculated by LiDAR / visual odometry at this moment; For the first The predicted observations are calculated from the nominal state of the next iteration. This nominal state is at time [time value missing]. It is obtained through iterative correction based on the prediction; In order to be in The Jacobian matrix obtained from the linearized observation model; For at any time The In the next iteration, the current error state is estimated.

[0049] Finally, a preliminary fused pose containing position, velocity, and attitude is obtained and sent to the subsequent evaluation layer for reliability assessment.

[0050] Step 3: The evaluation layer dynamically evaluates the credibility of the navigation source.

[0051] In this step, the preliminary fused poses of all navigation subsystems are input into a generative model constructed based on a neural network to assess the reliability of the navigation source corresponding to each navigation subsystem and obtain the likelihood probability. Furthermore, during the training process of the generative model, an active inference variational free energy minimization mechanism is employed to optimize the evaluation capability of the generative model.

[0052] Before introducing the specific implementation method of this step, the theoretical framework of the present invention regarding active reasoning will be described first.

[0053] (1) Constructing an active reasoning generative model Functionally, this invention corresponds to the "Bayesian brain" hypothesis, which describes a probabilistic reasoning system that continuously updates its beliefs about environmental states by integrating new observational information. In the active reasoning framework, the generative model is a cognitive structure used by the agent to simulate the external world. It describes how potential states generate observational results and updates beliefs and guides behavior accordingly. Figure 4 The structure of the active reasoning framework is clearly illustrated. The lower left side shows its underlying implementation logic, which, starting from Bayesian principles, models potential external states by constructing generative models and continuously refining estimates using variational reasoning methods, thereby achieving an integrated perception-action reasoning loop. The upper right side corresponds to the cognitive motivation of active reasoning, based on the free energy principle. Within the subject-environment boundary defined by the Markov blanket, the agent continuously verifies the correctness of its internal model through its own perception and actions, achieving "self-verification" and "self-creation." This Bayesian reasoning process follows the following theorem: (1) On the right side of the formula, It refers to the probability obtained based on past experience and analysis under the current world model. It does not rely on the probability distribution of observed data and represents the prior probability. It represents the likelihood probability, which is the probability of obtaining a specific observation result if a certain state is true within the model. This is model evidence, also known as marginal likelihood, which represents the degree of agreement between observed results and the overall world model. On the left side of the formula, It is the true posterior probability, representing the belief in the optimal adjustment after obtaining new observation data.

[0054] Therefore, the specific implementation of belief renewal includes the following three components: 1) Prior modeling: Prior beliefs (i.e., prior probability) represents the initial estimate of the credibility of the navigation source. This invention employs the Credibility Prior (DoO) method to model the initial fused pose of each navigation subsystem. Credibility prior modeling can be constructed with reference to reference 1 (K. Shen, M. Wang, M. Fu, Y. Yang, and Z. Yin, “Observability analysis and adaptive information fusion for integrated navigation of unmanned groundvehicles,” IEEE Transactions on Industrial Electronics, vol. 67, DOI 10.1109 / TIE.2019.2946564, no. 9, pp. 7659–7668, 2020.). In practical multi-source fusion navigation tasks, information from different sensors often exhibits inconsistencies and differences in credibility. Therefore, a method for real-time evaluation of multiple sensors is used to construct a generative model to adaptively assign weights to different navigation information sources.

[0055] In modern control theory, a navigation system can be represented by the following equation:

[0056]

[0057] in, Indicates in The system state vector at time t. Indicates from time to The state transition matrix at time t. This represents the process noise driving matrix, which maps process noise to the state space. Here, is the process noise vector, representing the model error; Indicates in The data vector actually measured by the time sensor; The observation matrix represents the system state. Mapping to observation space The linear mapping matrix. Observation noise is the random error that occurs during sensor measurements.

[0058] In DoO, each local filter uses IESKF as the fusion framework, taking sensor observations and model predictions as inputs, and updates the output with optimal navigation state information through filtering. During the filtering process, "observability" is calculated. To evaluate the reliability of navigation sources, this index not only integrates relevant information from the observability criterion matrix, but also fully considers factors such as dynamic errors and measurement noise in the estimation process, and effectively models the overall reliability of navigation information of the system (see reference 1 for the specific formula derivation).

[0059] 2) Likelihood model: Likelihood function Given a hidden state and Observed under the conditions The probability distribution is determined by the high observation dimensionality and complex model in multi-source fusion navigation. This likelihood is often difficult to obtain directly through explicit modeling. Therefore, this method employs a generative model based on a neural network structure, using the preliminary fused pose as input and outputting the likelihood probability. It is continuously adjusted based on minimizing the variational free energy so that its output can reflect the probability of the observed results.

[0060] 3) Post-hoc update: Posterior belief (posterior probability) reflects the latest estimate of the reliability of the navigation source after receiving new observation information. When the generative model is known (i.e., the prior and likelihood are known), the posterior distribution can be derived using Bayes' rule; however, since the true posterior is generally difficult to analyze, active inference employs variational inference: introducing a computable approximate posterior. And by minimizing the variational free energy Towards the true posterior near.

[0061] Based on this, by minimizing the variational free energy To achieve adaptive updates from prior to posterior, the variational free energy is defined as follows:

[0062]

[0063]

[0064] Formula (2) in, Indicates the actual state of the system; This represents the data that has been observed; This is the current generative model; Represents entropy; It is an approximate posterior distribution, which is selected to approximate the true posterior. Optimizable distribution family; Indicates in The expected value of the joint log probability of the generative model is given below. The last line of the formula is used in actual calculations.

[0065] The first line of the above equation indicates that minimizing the variational free energy means maintaining the "energy" of the generative model and maximizing the posterior entropy as much as possible. In the absence of perceptual observations and precise priors, the hidden state should follow the "maximum entropy principle." The second and third lines provide a detailed expansion of the free energy, the core idea of ​​which is that minimizing the free energy is equivalent to finding the optimal interpretation of the observations, that is, balancing simplicity and accuracy under the "Occam's Razor" principle. Here, the KL divergence is used to measure the difference between the approximate posterior and the prior or the true posterior. By minimizing this difference, a computationally feasible posterior approximate solution can be obtained, ensuring both the accuracy of the observation modeling and effectively controlling the model complexity, thereby achieving real-time reliability assessment in high-dimensional multi-source environments. The resulting posterior belief not only supports the navigation source weight allocation but also provides a theoretical basis for subsequent adaptive decision-making, achieving a balance between data accuracy and model simplicity.

[0066] Based on variational free energy Constructing rewards, variational free energy The smaller the value, the higher the reward. Minimize the variational free energy. To achieve this, the network parameters of the generative model are adjusted to improve the accuracy of the navigation source credibility assessment.

[0067] The evaluation layer minimizes the variational free energy. During this process, the credibility of the navigation source was updated in real time. This process enabled prior beliefs to be updated. This allows it to be combined with new observational evidence and updated into the optimal approximate posterior belief. Through the above optimization, the prior beliefs are continuously revised by new observational evidence, forming a posterior belief that adapts to environmental dynamics.

[0068] Step 3 corresponds to... Figure 1 The evaluation layer and belief alignment feedback channel are used in the process.

[0069] For the evaluation layer, a preliminary fusion pose input generation model (e.g., an LSTM network) is used to generate the navigation source based on the initial fusion of all navigation subsystems, thereby evaluating the reliability of the navigation source and obtaining the likelihood probability of the navigation subsystem. , Each navigation subsystem in the system corresponds to a likelihood probability value. Input to the decision-making level.

[0070] In the belief alignment and feedback pathway, an active reasoning variational free energy minimization mechanism is employed to enhance the interpretability and evaluation capabilities of the generative model. Specifically, this includes the following steps: (1) Construct priors for each navigation subsystem using the credible prior method (DoO). ; (2) Combining the likelihood given by the generative model The true posterior probability is calculated using the Bayesian inference process formula (Formula 1). Formula 1 The marginal likelihood is known.

[0071] (3) Calculate the mean based on the node parameters of the last layer of the generative model. and variance With this average and variance Construct Gaussian distribution As an approximate posterior probability ,in, , This indicates a Gaussian distribution.

[0072] (4) Using approximate posterior probability and true posterior probability Calculate variational free energy The calculation formula is shown in formula (2), marginal likelihood. Known.

[0073] (5) To minimize the variational free energy As an optimization objective, the network parameters of the generative model are adjusted to improve its ability to assess the credibility of navigation sources. This enables the generative model to interpret observations and maintain consistency with prior knowledge when judging the credibility of navigation sources, thereby improving the reliability and robustness of multi-source fusion.

[0074] Step 4: The decision-making layer adopts an expected free energy minimization mechanism based on active reasoning, and incorporates decision weights.

[0075] The decision layer is implemented using a neural network and trained using reinforcement learning. The input is the likelihood probability of each navigation subsystem, and the output decision is the fusion weights for re-fusion of the initial fused pose. During training, the decision layer determines the expected free energy based on the reliability of the navigation subsystems and the accuracy of the fused trajectory. To minimize the expected free energy To achieve the reward objective, optimize the network parameters of the decision-making layer.

[0076] (1) First, define the state space and action space of the decision model.

[0077] The core function of the decision-making layer is to make forward-looking decisions. The goal of decision-making is to assign an optimal set of fusion weights to each current state; this weight vector represents the selected action. All possible weight allocation schemes constitute the action space for decision-making. The weights of each subsystem in multi-source fusion navigation are discretized as action variables. Specifically, an action is defined as the weight allocation of the three subsystems GNSS / INS, VINS, and LiDAR / INS in fusion positioning; that is, each action corresponds to a weight vector. ,in The step size is 0.1, and it satisfies... In actual implementation, each action corresponds to a set of fusion weights for the three sensor subsystems. After the decision layer outputs the corresponding Q-value, it selects the action for localization fusion. This strategy not only intuitively reflects the contribution of each sensor combination but also ensures the finiteness and processability of the action space. Following the above method, a discrete action space containing 66 combinations is constructed for the example of the three subsystems of this invention.

[0078] The state space represents the likelihood probabilities of each navigation subsystem output by the evaluation layer.

[0079] (2) Evaluate and select the optimal strategy by minimizing the expected free energy.

[0080] In adjusting the generative model, reducing divergence through perception will bring the free energy closer to surprise; taking action to change the observation sampling to match prior predictions will directly reduce surprise. However, the variational free energy is essentially a function of past and current observation data, helping to infer the current state based on existing information, but not a function of future observations. To guide the agent in realizing its expectations and plans for the future, the concept of "expectation free energy" needs to be introduced, enabling the generative model to have a comprehensive assessment ability of future uncertainty and the value of actions.

[0081] Using minimizing the expected free energy To achieve the optimal action selection, the mathematical expression for the expected free energy is as follows: (3) In the above formula, the tilde "~" represents a future time series, for example... = {s t+1 , s t+2 , s t+3 , …} represents a sequence of future states, õ = {o t+1 , o t+2 , …} represents the future observation sequence, Indicating in strategy Below, the joint predicted distribution of one end of the future state and future observations. Representation Strategy The posterior probability for the next time period. This indicates that if future observations are õ, the distribution of future posteriors will be updated. This represents the target distribution of preference C in relation to future observations under the generative model m (i.e., preference distribution, a proper noun in active reasoning systems).

[0082] Intuitively, the first decomposition represents active reasoning as maximizing information gain and use value. This allows us to measure the value of an active reasoning agent seeking new information and seeking observations that align with its preferences using the same units, similar to the value of balancing "exploration-exploitation" in reinforcement learning. The second decomposition further reveals the internal structure of the expected free energy, generating two terms that calculate "risk" and "expected ambiguity." The former measures the mismatch between the observed outcome and preferences under the current policy, i.e., the risk of failure to achieve the expected goal, while the latter reflects the ambiguity of the observed state, corresponding to the model's expected complexity and expected uncertainty, respectively. By minimizing the expected free energy, the agent can not only continuously optimize its predictions of the future in uncertain environments but also make the most valuable decisions at the current point in time.

[0083] In formula (3) above, the third row of the formula includes two parts. and . For accuracy characterization , Ambiguity characterization term .

[0084] This invention calculates the variational free energy using a more easily computable method based on the meaning expressed by the constituent terms of the variational free energy, specifically including: (1) Calculate the accuracy characterization terms It is a time window The Euclidean error between the fused trajectory and the real trajectory is calculated based on the fusion weights of the decision-based calculation. The error is obtained by summing the fused trajectory points and real trajectory points at the current time and the previous T-1 time points.

[0085] (2) Calculate the ambiguity characterization term The covariance matrix of the IESKF filter is multiplied by the likelihood probability and applied in a time window. The inner summation yields:

[0086] in, The total number of navigation subsystems, for Time navigation subsystem The covariance matrix corresponding to the IESKF filter, For the output of the evaluation layer Time navigation subsystem The corresponding likelihood probability.

[0087] (3) Variational free energy = accuracy characterization term + Ambiguity characterization term .

[0088] During training, rewards are constructed using the expected free energy; the smaller the expected free energy, the higher the reward, with the goal of minimizing the expected free energy. To achieve the reward objective, optimize the network parameters of the decision-making layer.

[0089] When completing the training and applying it in practice, the optimized decision layer is used to determine the optimal fusion weights.

[0090] In an instance with three subsystems, the optimal fusion weight for the decision is: .

[0091] Step 5: Use the optimal fusion weights to perform weighted fusion of the preliminary fusion poses of each navigation subsystem to obtain the fusion trajectory.

[0092] In instances with three subsystems, the optimal fusion weights are employed. The position components in the preliminary fused poses of each navigation subsystem are weighted and fused to obtain the fused trajectory. .

[0093] Step 6: Based on the fused trajectory obtained in Step 5, the parameters of the algorithm used in the initial fusion are fed back for correction, thereby correcting the navigation subsystem error.

[0094] After selecting the optimal fusion weights and obtaining the global fusion result (fusion trajectory), the system not only uses this result as the final output but also feeds it back to each local subsystem to achieve error feedback correction. The purpose of this design is to suppress the accumulation of deviations caused by drift or abrupt errors of a single sensor in a complex environment, thereby improving the overall consistency and robustness of the system.

[0095] For the perception layer employing IESKF filters, the globally fused estimation results are fed back to each IESKF filter to adaptively correct the local state and covariance, effectively suppressing drift and improving global consistency. To achieve this adaptive feedback, the reliability of each subsystem is first quantitatively assessed using the Degree of Abnormality (DoA) described in reference 2 (K. Shen, T. Liu, Y. Li, N. Liu and W. Qi, "Robust Multi-Model Estimation for Reliable Relative Navigation Based on Observability and Abnormity Analysis," in IEEE Transactions on Intelligent Transportation Systems, vol. 24, no. 5, pp. 5144-5158, May 2023). Based on this assessment, the global covariance information is redistributed to each local filter according to a reliability-weighted method. Therefore, the entire system forms a closed-loop belief update process: fusion not only integrates multi-source observation information but also continuously self-corrects through reliability-based feedback, thereby further improving the system's stability and robustness under different navigation scenarios.

[0096] Based on the above method, the present invention also provides a cross-scenario proactive reasoning intelligent decision-making and consensus fusion navigation device, such as... Figure 5 As shown, it includes a perception layer, an evaluation layer, a decision-making layer, a fusion trajectory generation module, and a feedback correction module.

[0097] The perception layer is used to acquire observation information from multiple navigation sources; one navigation source is selected as the basic navigation source, and navigation subsystems are formed in pairs with other navigation sources; the information of each navigation subsystem is initially fused to obtain the initial fused pose.

[0098] The evaluation layer is used to assess the reliability of the navigation subsystems. The evaluation layer is implemented using a generative model based on a neural network. It takes the preliminary fused poses of each navigation subsystem as input and outputs the likelihood probability of each navigation subsystem.

[0099] The decision layer, implemented using a neural network and trained with reinforcement learning, takes the likelihood probabilities of each navigation subsystem as input and outputs the fusion weights for further fusion of the initial fused poses. During training, the decision layer determines the expected free energy based on the reliability of the navigation subsystems and the accuracy of the fused trajectory. To minimize the expected free energy To achieve the reward objective, optimize the network parameters of the decision-making layer.

[0100] The fusion trajectory generation module is used to perform weighted fusion of the preliminary fusion poses of each navigation subsystem using the optimal fusion weights determined by the decision layer, in order to obtain the fusion trajectory.

[0101] The feedback correction module connects the perception layer and the evaluation layer, and determines the prior probability based on the preliminary fused pose of each navigation subsystem. Based on the prior probability Likelihood probability The true posterior probability is calculated using the Bayesian inference process formula. Using the approximate posterior probability and true posterior probability Calculate variational free energy To minimize variational free energy As an optimization objective, the network parameters of the generative model are adjusted to improve the ability to assess the credibility of navigation sources.

[0102] In summary, the present invention has the following characteristics: (1) A three-layer cascaded active reasoning fusion navigation framework consisting of a perception layer, an evaluation layer, and a decision-making layer was constructed. This framework organically combines multi-source data preprocessing, dynamic credibility assessment, and adaptive weight decision-making to form a complete closed-loop system based on cognitive principles.

[0103] (2) The reliability of the navigation source is dynamically evaluated by adopting the variational free energy minimization mechanism. This is the core of realizing the endogenous assessment of perceived uncertainty, which is different from the traditional method based on the fixed noise model.

[0104] (3) The expected free energy minimization mechanism is adopted as the decision criterion for fusion weight allocation. By balancing "exploration-exploitation", the long-term benefits are maximized, which is different from the traditional reinforcement learning method that only maximizes the cumulative reward to achieve intelligent decision-making.

[0105] (4) Active Inference (AIF) and Reinforcement Learning (RL) are deeply integrated to form a unified AIF-RL decision-making framework. This framework combines the processing power of RL with the Bayesian inference and free energy minimization principle of AIF to solve decision-making problems in complex navigation environments.

[0106] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cross-scenario proactive reasoning intelligent decision-making and consensus fusion navigation method, characterized in that, include: Step 1: Collect observation information from multiple navigation sources; Step 2: Select one of the navigation sources as the basic navigation source, and form navigation subsystems in pairs with the other navigation sources; The information from each navigation subsystem is initially fused to obtain a preliminary fused pose; Step 3: Input the initial fusion pose of all navigation subsystems into the evaluation layer. The evaluation layer adopts a generative model based on neural network to evaluate the reliability of the navigation subsystems and output the likelihood probability. Step 4: The decision layer is implemented using a neural network and trained using reinforcement learning. The input is the likelihood probability of each navigation subsystem, and the output decision is the fusion weight for re-fusion of the initial fused pose. During training, the decision layer determines the expected free energy based on the reliability of the navigation subsystems and the accuracy of the fused trajectory. To minimize the expected free energy To achieve the reward objective, optimize the network parameters of the decision-making layer; Step 5: Using the fusion weights determined in Step 4, perform weighted fusion on the preliminary fusion poses of each navigation subsystem to obtain the fusion trajectory.

2. The method as described in claim 1, characterized in that, An error state iterative extended Kalman filter (IESKF) is used to initially fuse the information from each navigation subsystem.

3. The method as described in claim 1, characterized in that, Step 1 involves acquiring observation information from multiple navigation sources, including asynchronous raw data streams from GNSS receivers, LiDAR, cameras, and IMUs; In step 2, the IMU is selected as the basic navigation source, and the basic navigation source is initially fused with other navigation sources to obtain the initial fused pose.

4. The method as described in claim 2, characterized in that, The input to the decision layer is the likelihood probability of each navigation subsystem output by the evaluation layer, and the output of the decision layer is the fusion weight. The reward is constructed using the expected free energy; the smaller the expected free energy, the higher the reward. The expected free energy is characterized by an accuracy term. and ambiguity characterization terms Composition; among which, The accuracy characterization item For time windows The Euclidean error between the fused trajectory and the true trajectory calculated based on the fusion weights of the decision-based calculation; The ambiguity characterization term The covariance matrix of the IESKF filter is multiplied by the likelihood probability and then applied in a time window. The inner summation yields: in, The total number of navigation subsystems, for Time navigation subsystem The covariance matrix corresponding to the IESKF filter, For the output of the evaluation layer Time navigation subsystem The corresponding likelihood probability.

5. The method as described in claim 4, characterized in that, The action space of the decision-making layer is set as follows: The weights of each navigation subsystem are set to range from [0,1] with a step size of 0.1, and the sum of the weights of all navigation subsystems is 1. Under the constraints of the weight range, step size, and weight sum, the fusion weight combination is determined, and a discrete action space composed of different fusion weight combinations is constructed as the action space of the decision layer.

6. The method as described in claim 1, characterized in that, The generative model uses an LSTM network.

7. The method as described in claim 1, characterized in that, The method further includes: employing an active inference variational free energy minimization mechanism to dynamically optimize the evaluation capability of the generative model, including: Based on the preliminary fused pose of each navigation subsystem, determine the prior probability. ; Based on the prior probability The likelihood probability determined in step 3 The true posterior probability is calculated using the Bayesian inference process formula. ; The mean and variance are calculated based on the node parameters of the last layer of the generative model. A Gaussian distribution is then constructed using these mean and variance as the approximate posterior probability. ; Using the approximate posterior probability and true posterior probability Calculate variational free energy ; To minimize variational free energy As an optimization objective, the network parameters of the generative model are adjusted to improve the ability to assess the credibility of navigation sources.

8. The method as described in claim 7, characterized in that, The prior probability The determination method is as follows: the initial fused pose of each navigation subsystem is modeled using the credible prior modeling DoO to obtain the prior probability. .

9. The method as described in claim 1, characterized in that, The method further includes: based on the fusion trajectory obtained in step 5, performing feedback correction on the parameters of the algorithm used for the initial fusion.

10. A cross-scenario proactive reasoning intelligent decision-making and consensus fusion navigation device, characterized in that, It includes a perception layer, an evaluation layer, a decision-making layer, a fusion trajectory generation module, and a feedback correction module; The perception layer is used to acquire observation information from multiple navigation sources; Selected One of the navigation sources serves as the basic navigation source, and together with the other navigation sources, they form navigation subsystems. The information from each navigation subsystem is initially fused to obtain a preliminary fused pose; The evaluation layer is used to evaluate the reliability of the navigation subsystem. The evaluation layer is implemented using a generative model based on a neural network. It takes the preliminary fused pose of each navigation subsystem as input and outputs the likelihood probability of each navigation subsystem. The decision layer is implemented using a neural network and trained using reinforcement learning. Its input is the likelihood probability of each navigation subsystem, and its output is the fusion weights for re-fusion of the initial fused pose. During training, the decision layer determines the expected free energy based on the reliability of the navigation subsystems and the accuracy of the fused trajectory. To minimize the expected free energy To achieve the reward objective, optimize the network parameters of the decision-making layer; The fusion trajectory generation module is used to perform weighted fusion of the preliminary fusion poses of each navigation subsystem using the optimal fusion weights determined by the decision layer, in order to obtain the fusion trajectory. The feedback correction module connects the perception layer and the evaluation layer, and determines the prior probability based on the preliminary fused pose of each navigation subsystem. Based on the prior probability Likelihood probability The true posterior probability is calculated using the Bayesian inference process formula. Using the approximate posterior probability and true posterior probability Calculate variational free energy To minimize variational free energy As an optimization objective, the network parameters of the generative model are adjusted to improve the ability to assess the credibility of navigation sources.

Images