Navigation method and system for hierarchical multi-layer multi-source fusion instant access

Abstract

The present invention discloses a hierarchical multi-layer multi-source fusion instant access navigation method and system. The method comprises: evaluating the performance of multiple sensors for collecting navigation information in real time, and making decisions on the navigation information combination scheme in an intelligent manner based on the evaluated performance; using a nonlinear fusion method of an adaptive factor graph, allocating weights based on the navigation information combination scheme to achieve adaptive fusion, obtaining a fusion result, and navigating based on the fusion result. The present invention solves the technical problem of inaccurate navigation accuracy caused by the failure of one of the sensors.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-source information fusion of combined navigation and intelligent navigation, and in particular to a hierarchical multi-layer multi-source fusion instant access navigation method and system. Background Art

[0002] Currently, multi-source fusion navigation has become a focus of attention and research hotspot in academia and industry. Due to various error sources in the signal propagation process, interference from the external environment, and uncertainty in the motion process, a single sensor cannot continuously provide robust and accurate navigation information in complex and dynamic scenarios.

[0003] The navigation information sources of the commonly used positioning methods at present include inertial navigation devices based on dead reckoning, global satellite positioning systems and wireless positioning technologies based on signal reception, images and laser radars based on environmental perception, etc., which form a multi-source fusion navigation system that can cover most of the scenarios required for positioning, such as indoors, on water, in the air, in urban canyons, underground, etc. Therefore, in the process of cross-scenario, multi-task, multi-terrain, and long-term driving, high-precision and high-reliability navigation positioning can be achieved by carrying a variety of navigation information sources with complementary information.

[0004] For the fusion method of ten or more navigation information sources in large-scale complex dynamic scenes, the amount of calculation will explode with the increase in the number of navigation information sources. For application scenarios that require real-time positioning, the delay in analysis will cause a "cliff-like" drop in navigation accuracy. Navigation fusion methods are divided into loose coupling and tight coupling. Loose coupling is mainly based on federal Kalman filtering. When the reference navigation source is abnormal, it will affect the performance of the overall positioning system and has poor fault tolerance. For the tightly coupled method, the excessive number of sensor types leads to a significant increase in the amount of calculation when introducing new sensors. At the same time, the increase in the dimension of the Jacobian matrix during nonlinear optimization also makes it difficult to solve.

[0005] When a sensor fails for a short time or a long time, the accuracy of the sensors will not match each other, and the unreliable sensor information will affect the entire fusion framework. As the environment and movement mode of the autonomous unmanned system are constantly changing, the optimal combined navigation mode should be selected in real time. Therefore, in order to meet the navigation needs of instant access under cross-scenario conditions, it is necessary to design a high-precision, high-robust multi-source fusion navigation software to achieve hierarchical management of access sensors, be compatible with various sensor interfaces, and complete the combination and instant access of different types of sensors during operation.

[0006] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0007] The embodiment of the present invention provides a hierarchical, multi-layer, multi-source fusion instant access navigation method and system to at least solve the technical problem of inaccurate navigation accuracy caused by failure of one of the sensors.

[0008] According to one aspect of an embodiment of the present invention, there is provided a hierarchical, multi-layer, multi-source fusion, instant access navigation method, comprising: real-time evaluation of the performance of multiple sensors for collecting navigation information, and making decisions on a navigation information combination scheme in an intelligent manner based on the evaluated performance; using a nonlinear fusion method of an adaptive factor graph, allocating weights based on the navigation information combination scheme to achieve adaptive fusion, obtaining a fusion result, and performing navigation based on the fusion result.

[0009] In an embodiment of the present invention, the performance of multiple sensors used to collect navigation information is evaluated in real time, and a decision on a navigation information combination scheme is made in an intelligent manner based on the evaluated performance; and a nonlinear fusion method of an adaptive factor graph is adopted to allocate weights based on the navigation information combination scheme to achieve adaptive fusion, obtain a fusion result, and navigate based on the fusion result, thereby solving the technical problem of inaccurate navigation accuracy caused by the failure of one of the sensors, thereby achieving the technical effect of improving navigation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0011] Figure 1 is a flowchart of a hierarchical multi-layer multi-source fusion instant access navigation method according to a first embodiment of the present invention;

[0012] Figure 2 is a flowchart of a hierarchical multi-layer multi-source fusion instant access navigation method according to a second embodiment of the present invention;

[0013] Figure 3 is a flowchart of a hierarchical multi-layer multi-source fusion instant access navigation method according to a third embodiment of the present invention;

[0014] Figure 3A is a step-by-step flow chart of step S2 according to an embodiment of the present invention;

[0015] Figure 3B is a step-by-step flow chart of step S3 according to an embodiment of the present invention;

[0016] Figure 3C is a step-by-step flow chart of step S5 according to an embodiment of the present invention;

[0017] Figure 3D is a step-by-step flow chart of step S6 according to an embodiment of the present invention;

[0018] Figure 3E is a step-by-step flow chart of step S63 according to an embodiment of the present invention;

[0019] Figure 3F is a step-by-step flow chart of step S631 according to an embodiment of the present invention;

[0020] Figure 3G is a step-by-step flow chart of step S632 according to an embodiment of the present invention;

[0021] Figure 4 is a structural schematic diagram of a hierarchical, multi-layer, multi-source fusion instant access navigation system according to an embodiment of the present invention;

[0022] Figure 5 It is a structural schematic diagram of another hierarchical multi-layer multi-source fusion instant access navigation system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. 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 creative work should fall within the scope of protection of the present invention.

[0024] Example 1

[0025] According to an embodiment of the present invention, a hierarchical multi-layer multi-source fusion instant access navigation method is provided, such as Figure 1 As shown, the method includes: step S102, evaluating the performance of multiple sensors used to collect navigation information in real time, and making decisions on the navigation information combination scheme in an intelligent manner based on the evaluated performance; step S104, using a nonlinear fusion method of an adaptive factor graph, allocating weights based on the navigation information combination scheme to achieve adaptive fusion, obtaining a fusion result, and navigating based on the fusion result.

[0026] In an exemplary embodiment, the performance of multiple sensors for collecting navigation information is evaluated in real time, and a decision on a navigation information combination scheme is made in an intelligent manner based on the evaluated performance, including: constructing an abnormality rating criterion to describe the degree of influence of the uncertainty of the external environment on at least one of the multiple sensors in view of the uncertainty of the carrier maneuver and the randomness of the dynamic interference of the external environment; calculating the observability of each combined navigation sub-filter state variable based on the observability analysis of the linear time-varying system to quantify the observable capability of at least one of the multiple sensors; determining a reinforcement learning method based on the influence degree and the observability, wherein the reinforcement learning method includes a Q-learning algorithm and an asynchronous dominant actor-judge algorithm; constructing a reinforcement learning network according to the determined reinforcement learning method, training the reinforcement learning network in a complex environment and motion, and using the reinforcement learning network to make a decision on the navigation information combination scheme to obtain the optimal navigation information combination scheme.

[0027] In an exemplary embodiment, based on the degree of influence and the observability, a reinforcement learning method is determined, including: in a simple environment, defining a state type, using a Q-learning algorithm, and discretizing the continuous variables of the multiple sensors in segments by setting thresholds, and normalizing them to achieve the action of controlling at least one of the multiple sensors, wherein the state type includes at least one of the following: auxiliary sensor uncertainty, auxiliary sensor abnormality, speed, GNSS status, number of satellites, and satellite signals, the action is switching the sensor, and the immediate benefit is the inverse of the sum of squares of the position error at the current moment; in a complex environment and motion scene, an asynchronous advantage actor-judge algorithm is used to interact with the environment in multiple threads to train a neural network model so that the neural network model converges and outputs an accurate estimation effect, wherein the state element input to the neural network model is the continuous variable, the action output by the neural network model is the fusion weight of the auxiliary sensor among the multiple sensors, and the immediate benefit is the inverse of the sum of squares of the position error at the current moment.

[0028] In an exemplary embodiment, in a simple environment, a state type is defined, and a Q-learning algorithm is used to discretize the continuous variables of the multiple sensors in segments by setting thresholds, and normalize them to achieve the action of controlling at least one of the multiple sensors, including: receiving at least one of the following input parameters: number of iterations, state set, action set, step size, attenuation factor, and exploration rate; based on the input parameters, using an ∈-greedy method, based on the current state in the state set, selecting a current action, and executing the current action to obtain a new state and reward; based on the new state and the reward, updating the value function and state, and outputting an estimate of the true value of all states and actions and the switching of auxiliary sensors based on the value function and state.

[0029] In an exemplary embodiment, in a complex environment and motion scene, an asynchronous advantage actor-critic algorithm A3C is used to interact with the environment in multiple threads to learn and train a neural network model so that the neural network model converges and outputs an accurate estimation effect, including: receiving at least one of the following input parameters: neural network parameters of the global part, the number of globally shared iterations, the global maximum number of iterations, the A3C neural network structure of the current thread, the maximum length of a single iteration time series in the thread, the state feature dimension, the action set, the step size, and the attenuation factor; selecting an action based on the determined navigation information combination scheme and executing the action to obtain a reward and a new state; based on the reward and the new state, calculating the action value function at each moment, accumulating the local gradient update of the Actor, and accumulating the local gradient update of the Critic, and updating the parameters of the neural network model; using the updated neural network model, based on the input parameters, outputting the neural network parameters of the global part and the fusion weights of the auxiliary sensors among the multiple sensors.

[0030] In an exemplary embodiment, a nonlinear fusion method of an adaptive factor graph is adopted, weights are allocated based on the navigation information combination scheme to achieve adaptive fusion, a fusion result is obtained, and navigation based on the fusion result includes: seeking an optimal state estimation through least squares estimation:

[0031]

[0032] Where X = (X1, ..., X s ) is the state vector, and the adaptive information matrix is ​​Ω ij , l represents the measurement set, e ij = e(x i , x j , z ij ) is the error vector, z ij For measurement information, is the transpose of the error vector, i is node i, and j is node j.

[0033] In an exemplary embodiment, before making a decision on the navigation information combination scheme in an intelligent manner based on the evaluated performance, the method also includes: constructing a combined navigation model to determine the state equation and measurement equation of the multi-source fusion navigation system; based on the state equation and the measurement equation, using Kalman filtering in the sub-filter to perform state estimation to obtain accurate navigation information; performing local filtering based on the accurate navigation information to obtain the optimal estimate of the local filtering, using the residual and the observed prediction error covariance as abnormality input parameters, and using the prediction error covariance and the state prediction posterior error covariance as observability input parameters.

[0034] In an exemplary embodiment, before determining the state equation and measurement equation of the multi-source fusion navigation system, the method also includes: the sensor MCU of the sensor layer writes the solved navigation information into the real-time message queue through the formulated data frame specification, and the fusion layer MCU, as a receiving node, completes the message transmission by collecting the navigation information data in the message queue, wherein the real-time message queue is used to collect and send the internal sensor data of the multiple sensors; the sensor MCU completes the solution of the navigation information, and at the same time sends the task ID to the fusion layer MCU in the polling request interface, after multiple communication cycles, the sensor MCU sends the navigation information after receiving the polling request, the polling request interface is cleared, and the message is transmitted from the sensor layer to the fusion layer; the instant access sensor that meets the hardware and interface standards completes the raw data collection and navigation information solution, and after the system's instant access manager determines that the instant access sensor is connected, it allocates the corresponding hardware resources and automatically adjusts the device-related parameters to ensure that the navigation information of the sensor layer is transmitted to the fusion layer after the connection is completed.

[0035] In an exemplary embodiment, the method further includes: identifying the structure in the fusion result according to the grammatical rules includes: dividing the fusion result into multiple partitions based on the grammatical rules, and determining the structure of the multiple partitions of the fusion result; determining each refined symbol value in the refined symbol information in the fusion result based on the average value of the number of predetermined blocks in the multiple partitions, and using a quantization step to calibrate the each refined symbol value, wherein the quantization step depends on the reference quantization step adopted by the predetermined space vector associated with the fusion result; associating each refined symbol value after the quantization step with the multiple partitions respectively, and linearly combining the average values ​​of the refined symbol values ​​after the quantization step of two adjacent partitions in the multiple partitions; and refining the structure of the two adjacent partitions based on the average value of the refined symbol values ​​of the two adjacent partitions after the linear combination.

[0036] The present embodiment provides a hierarchical multi-level adaptive multi-source fusion navigation method, which aims to solve the problem of large-scale high-precision navigation and positioning, and proposes a hierarchical adaptive multi-source fusion filtering architecture. The architecture includes a sensor layer composed of multiple sensors, a data transmission layer, a fusion layer, and a navigation data application layer. This hierarchical framework can realize the hardware architecture and instant access software interface specifications of an open all-source navigation system.

[0037] In addition, this embodiment also proposes a message collection and sending mechanism based on sensor classification. This mechanism forms a message collection and sending mechanism of corresponding priority by reasonably classifying the sensors connected to the system. After reasonable allocation of hardware computing resources, the three mechanisms can realize asynchronous transmission of sensor data, application decoupling and current limiting and peak elimination, avoiding transmission congestion and signal stacking when connecting to multiple sensors in complex, dynamic and large-scale scenarios, ensuring the fault tolerance performance of the system, and realizing the strong robustness and safe and reliable positioning of the navigation and positioning algorithm.

[0038] Example 2

[0039] This embodiment solves the problem of fusing ten or more sensors for large-scale high-precision navigation and positioning, realizes instant access to sensor combinations, and ensures the accuracy and reliability of multi-source navigation information fusion.

[0040] In this embodiment, a hierarchical adaptive multi-source fusion filtering architecture is firstly proposed. The architecture includes a sensor layer composed of multiple sensors, a data transmission layer, a fusion layer and an application layer.

[0041] The sensor layer includes internal sensors composed of MEMS gyroscopes, MEMS accelerometers, magnetometers, satellite receivers, etc., expandable sensors composed of visual odometers, laser radars, etc., external instant access sensors composed of laser rangefinders, millimeter wave rangefinders, etc., and embedded sensor microcontroller units (MCUs) integrated in each sensor. Different sensors transmit data to their microcontroller units (MCUs) through data buses such as SPI, I2C, CAN, or serial ports such as CSI according to the protocols supported by their hardware. In an exemplary embodiment, each navigation information source in the sensor layer can receive observation data, pre-process the observation data through the processor at the sensor end, and output positioning information.

[0042] The data transmission layer includes a wired and wireless transmission combination consisting of "Body Area Network / Local Area Network + Serial Port". According to the data transmission capabilities of different sensors, there are different specific connection methods such as Bluetooth, WIFI, USB serial port, etc. The data transmission layer can also connect to the sensor layer and the fusion layer to achieve communication with the host.

[0043] The fusion layer includes a local fusion layer and a centralized fusion layer, which respectively perform combined filtering of observation information and optimal navigation information combination decision. The local fusion layer in the fusion layer can perform Kalman filtering on the observation information of inertial information and auxiliary information sources, and send the optimized results to the centralized fusion layer. The centralized fusion layer in the fusion layer evaluates the performance of the sensor in real time, selects the corresponding reinforcement learning algorithm according to the complexity of the environment, obtains the optimal navigation information combination scheme through reinforcement learning, and feeds the combination results back to the interaction layer to realize the selection of sensors.

[0044] The application layer includes navigation information application devices such as handheld terminals, water equipment, road unmanned vehicles, and aerial drones to achieve navigation and positioning in different application environments. Furthermore, the application layer can receive multi-sensor fusion information realized by adaptive factor graphs, and send the fusion results to the application layer to achieve navigation and positioning indoors, on the ground, in the air, or on the water.

[0045] The multi-source fusion navigation method based on the sensor hierarchical message collection and sending mechanism provided in this embodiment includes the following steps:

[0046] Step S201, classifying sensors.

[0047] According to the different sensors connected, they can be reasonably classified into three levels: internal sensors, expandable sensors, and external instant access sensors according to the access structure of sensors and hardware systems. Internal sensors include: MEMS gyroscopes, MEMS accelerometers, magnetometers, satellite receivers, etc., expandable sensors include: visual odometers, laser radars, etc., and external instant access sensors include: laser ranging sensors, millimeter wave ranging sensors, etc.

[0048] Step S202, receiving observation data and outputting positioning information.

[0049] Each navigation information source receives observation data, pre-processes the observation data through the embedded processor at the sensor end, and outputs positioning information.

[0050] Step S203: output the positioning information to the data transmission layer.

[0051] Output the positioning information to the data transmission layer, and through the stratification of sensor priority levels in step S202, build a real-time message queue corresponding to the core sensor layer, an event polling mechanism corresponding to the extensible sensor layer, and an external instant access message collection and transmission mechanism corresponding to the external sensor layer at the software level, so as to reasonably allocate hardware computing resources and avoid transmission congestion and signal stacking when multiple sensor data are being transmitted.

[0052] Step S204: input the data into the local fusion layer.

[0053] Through different message collection and sending mechanisms, the sensor layer data is input into the local fusion layer to complete the distributed transmission of sensor data.

[0054] Step S205: perform local filtering.

[0055] The local fusion layer selects the inertial information source as the reference navigation source, combines it with the auxiliary navigation source in pairs for local filtering, and then inputs each local filtering solution into the centralized adaptive fusion layer.

[0056] Step S206: perform real-time evaluation.

[0057] The centralized fusion layer first evaluates the sensor performance in real time and makes decisions on the navigation information combination scheme in an intelligent way.

[0058] Step S207, performing adaptive fusion.

[0059] The centralized fusion layer adopts the nonlinear fusion method of adaptive factor graph, and realizes adaptive fusion by allocating weights through the centralized fusion module, and finally sends the fusion result to the application layer.

[0060] This embodiment divides the access method of the sensor information source combined navigation mode into three levels: internal sensors, expandable sensors, and external instant access sensors, which correspond to three message collection and sending mechanisms respectively. The three mechanisms can realize asynchronous transmission, application decoupling, and current limiting and peak elimination of sensor data, avoiding transmission congestion and signal stacking when connecting to multiple sensors in complex, dynamic, and large-scale scenarios, and can reasonably allocate limited hardware resources, ensure the fault tolerance performance of the system, and realize the strong robustness and safe and reliable positioning of the navigation and positioning algorithm.

[0061] In addition, the hierarchical multi-level framework adopted in this embodiment can realize instant access to navigation information sources, and the navigation information sources are independent of each other, and the hardware architecture and instant access software interface specifications of the open all-source navigation system are realized in terms of hardware and interface.

[0062] Example 3

[0063] The present invention provides a multi-source fusion navigation method based on a sensor hierarchical message collection and transmission mechanism, such as Figure 3 As shown, the following steps S1-S7 are included.

[0064] Step S1, classifying sensors.

[0065] According to the different sensors connected, they can be reasonably classified into three levels: internal sensors, expandable sensors, and external instant access sensors according to the access structure of sensors and hardware systems. Internal sensors include: MEMS gyroscopes, MEMS accelerometers, magnetometers, satellite receivers, etc., expandable sensors include: visual odometers, laser radars, etc., and external instant access sensors include: laser ranging sensors, millimeter wave ranging sensors, etc.

[0066] Step S2, receiving and processing observation data, and outputting positioning information.

[0067] Each navigation information source receives observation data, pre-processes the observation data through the embedded processor at the sensor end, and outputs positioning information.

[0068] like Figure 3A As shown, step S2 includes the following sub-steps S21-S27.

[0069] Step S21: The inertial sensor collects and outputs data.

[0070] The inertial sensor communicates with the embedded MCU via the SPI / I2C bus and outputs the status [r INS V INS φ INS ε INS a INS ] is the position, velocity, attitude, gyroscope drift and accelerometer bias error; the observed position (longitude, latitude, altitude) is The speed (north, east, celestial) is measured as The attitude (north, east, sky) is φ INS =[φ E φ N φ Up ], gyroscope drift ε INS =[ε E ε N ε Up ] and the accelerometer bias error is a INS =[a E a H a Up ].

[0071] Step S22: the magnetometer collects and outputs data.

[0072] The magnetometer communicates with the embedded MCU via the SPI / 12C bus and outputs a six-degree-of-freedom attitude estimate. mag φ mag ], where the position Posture φ mag =[φ roll φpitch φ yaw ].

[0073] Step S23: The satellite collects and outputs data.

[0074] The satellite receiver communicates with the embedded MCU via the SPI / I2C bus and outputs the position (longitude, latitude, altitude) measured by GNSS as The GNSS velocity (north, east, celestial) measurements are

[0075] Step S24: the visual camera collects and outputs data.

[0076] The visual camera communicates with the embedded MCU through the CSI interface and outputs a six-degree-of-freedom attitude estimate. VINS φ VINS ], where Posture φ VINS =[φ roll φ pitch φ yaw ].

[0077] Step S25, the laser radar collects and outputs data.

[0078] The laser radar communicates with the embedded MCU through the CAN bus or RS485 serial port and outputs a six-degree-of-freedom attitude estimate. Lidar φ Lidar ], where Posture φ Lidar =[φ roll φ pitch φ yaw ].

[0079] Step S26: the laser ranging sensor collects and outputs data.

[0080] The laser ranging sensor communicates with the embedded MCU via the SPI / I2C bus and outputs a single-degree-of-freedom distance estimate [r Laser ], used to indicate the distance from the obstacle to the sensor.

[0081] Step S27, the millimeter wave radar collects and outputs data.

[0082] The millimeter wave radar communicates with the embedded MCU via the SPI / I2C bus and outputs a single degree of freedom distance estimate [r MWR ], used to indicate the distance from the obstacle to the sensor.

[0083] Step S3, sending the positioning information of sensors at different levels through different message queues.

[0084] The output positioning information reaches the data transmission layer. Through the stratification of sensor priority levels in step S1, a real-time message queue corresponding to the core sensor layer, an event polling mechanism corresponding to the scalable sensor layer, and an external instant access message collection and transmission mechanism corresponding to the external sensor layer are constructed at the software level. Hardware computing resources are reasonably allocated to avoid transmission congestion and signal stacking when multiple sensor data are being transmitted.

[0085] In one embodiment, Figure 3B As shown, step S3 may include the following sub-steps S31-S33.

[0086] Step S31: The real-time message queue collects and sends internal sensor data.

[0087] The sensor MCU writes the solved navigation information into the message queue according to the established data frame specification, and then the fusion layer MCU, as the receiving node, completes the message transmission by collecting the navigation information data in the message queue.

[0088] Step S32: The event polling mechanism is used to transmit and send the scalable sensor data.

[0089] The sensor MCU completes the solution of the navigation message and sends the task ID to the fusion layer MCU in the polling request interface. After several communication cycles, the sensor MCU sends the navigation information after receiving the polling request, and the polling request interface is cleared to complete the transmission of the message from the sensor layer to the fusion layer.

[0090] Step S33: External instant access is used to transmit and send external instant access sensor data.

[0091] The raw data collection and navigation information solution are completed by the instant access sensors that meet the hardware and interface standards. After the system's instant access manager determines that an instant access sensor is connected, it allocates the corresponding hardware resources and automatically adjusts the device-related parameters to ensure that the navigation information of the sensor layer is transmitted to the fusion layer after the connection is completed.

[0092] Step S4, through different message collection and sending mechanisms, the sensor layer data is input into the local fusion layer to complete the distributed transmission of the sensor data.

[0093] Step S5: The local fusion layer uses the inertial information source as a reference navigation source, combines it with other auxiliary navigation sources in pairs to perform local filtering, and then inputs each local filtering solution into the centralized fusion layer.

[0094] In an exemplary embodiment, Figure 3C As shown, step S5 includes the following sub-steps S51-S53.

[0095] Step S51, select the Northeast Sky (EN-Up) geographic coordinate system as the coordinate system for multi-source fusion navigation solution, build an inertial navigation / auxiliary sensor combined navigation model, and give the system state equation and measurement equation. The state equation of the multi-source fusion navigation system is:

[0096] x k+1 =Φ k+1 / k x k +Γ k w k (1)

[0097] In the formula, is the state vector of the inertial navigation source, including position, velocity, attitude, gyroscope drift and accelerometer bias error; w k is the system process noise; Γ k is the process noise distribution matrix; Φ k+1 / k is the system state transfer matrix, which includes position, velocity, attitude and sensor error model.

[0098] The measurement equation of the multi-source fusion navigation system is:

[0099] z k+1 =H k+1 x k+1 +v k+1 (2)

[0100] Where z k+1 is the measurement vector of the auxiliary sensor, corresponding to the positioning information output by each sensor in step S2; k+1 is the measurement (or observation) matrix, v k+1 is the measurement noise vector, v k ~N(0,Q k ) has a mean of 0 and a variance of Q k Gaussian distribution of

[0101] S52, using Kalman filtering in the sub-filter to perform state estimation, and then obtain accurate navigation information, the algorithm is as follows:

[0102]

[0103] Q k is the system process noise variance matrix; R k is the system observation noise matrix; P k+1|k is the prior covariance matrix of the error estimate, P k+1 is the posterior covariance matrix of the error estimate, is x k+1 The optimal estimate of K k+1 is the Kalman filter gain, is the estimated state variable, Φ k+1|kis the state transition matrix from time k to time k+1, is the state transfer matrix transpose, Γ k+1|k is the process noise input matrix, Q k is the process noise covariance matrix, is the transpose of the process noise input matrix, H k+1 is the observation matrix, is the transpose of the observation matrix, R k+1 is the observation noise matrix, z k+1 is the observation sequence at time k+1.

[0104] Step S53, the optimal estimate of the local filter is sent to the factor graph fusion part of the centralized fusion layer, and the residual and the observed forecast error covariance As the abnormality input parameter. The prediction error covariance and the state prediction posterior error covariance As an input for observability.

[0105] Step S6, evaluating the sensor performance in real time and making decisions on the navigation information combination scheme in an intelligent manner.

[0106] In a Figure 3D As shown, step S6 may include the following sub-steps S61-S64

[0107] Step S61, in view of the uncertainty of carrier maneuvering and the randomness of external environment dynamic interference, an abnormality rating criterion is constructed to describe the degree of influence of external environment uncertainty on the sensor:

[0108]

[0109] in, is the residual matrix multiplied by its transposed mean, n, m are constants, is the observed forecast error covariance.

[0110] Step S62, based on the observability analysis of the linear time-varying system, the observability of the state variables of each integrated navigation sub-filter is calculated, so as to quantify the observability of the sensor:

[0111]

[0112] In addition to the above indicators, you can also choose observation data features to describe the motion uncertainty and real-time performance of sensors for non-uniform motions such as acceleration and turning, such as acceleration variance, the number of satellites received by GNSS, and satellite positioning uncertainty reflecting the signal quality of satellites.

[0113] In step S63, two types of reinforcement learning methods are used according to the complexity of the environment: Q-learning and asynchronous dominant actor-critic algorithm (A3C). The former simplifies state variables and only controls sensor switches; the latter can process continuous state variables and control sensor fusion weights.

[0114] Define the set of states, actions, and benefits as S, A, and R. At time t, the agent observes state S t ∈S, will choose action A according to strategy π t ∈A, π(a|s) is when the state S t =s, select action A t = a (capital letters represent variables themselves, and lowercase letters represent corresponding instances) conditional probability. The action with a large probability is more likely to be selected by the individual. Q π is the action value function of strategy π, and V is the state value function.

[0115] At the next moment, the agent receives a numerical reward R t+1 ∈R, and enter the new state S t+1 , the agent continues to learn from the environment until the termination state. γ∈[0,1] is the reward decay factor. When it is equal to 0, it is the greedy method, that is, the value is only determined by the current delayed reward. When it is equal to 1, all subsequent state rewards are treated equally with the current reward. Exploration rate∈, when selecting the optimal action during training, there is a certain probability∈ not to choose the action that maximizes the current value, but to choose other actions.

[0116] In an exemplary embodiment, Figure 3E As shown, step S63 may include the following detailed sub-steps S631 - S632 .

[0117] Step S631: In a simple environment, this embodiment defines the state type as auxiliary sensor uncertainty, auxiliary sensor abnormality, speed, GNSS state, number of satellites, and satellite signal. The continuous variables are discretized and normalized by setting thresholds. The control of sensor switches is realized, and the action set is defined as the combination of auxiliary sensor switches. The immediate benefit is the inverse of the sum of squares of the position error at the current moment.

[0118]

[0119] Table 1

[0120] When the state and action sets are small, Q-learning can be used directly to update the strategy by updating the value function to solve the optimal strategy. Figure 3F As shown, step S631 includes the following detailed sub-steps S6311-S6318:

[0121] Step S6311, input: number of iterations T, state set, action set, step size α∈(0,1], decay factor γ, exploration rate ε;

[0122] Step S6312, all states s∈S and actions a∈A(s), initialize the value Q(s, a). For the terminal state, the Q value is 0;

[0123] Step S6313, initialize S to the first state of the current state sequence;

[0124] Step S6314, select action A in the current state S using the ∈-greedy method;

[0125] Use the ∈-greedy method to select action A in the current state S, execute the current action A, and get the new state S′ and reward R;

[0126] Step S6315, update the value function:

[0127] Step S6316, status update: S=S′.

[0128] Step S6317, determine whether the state is a termination state.

[0129] If s′ is a terminal state, the current round of iteration is completed, otherwise return to S6314;

[0130] Step S6318, output: estimate Q(S, A) of the true value of all states and actions, switch of auxiliary sensor;

[0131] Step S632: In complex environments and motion scenes, the asynchronous advantage actor-critic algorithm A3C algorithm is used, combined with value-based and policy-based reinforcement learning methods, to interact with the environment in multiple threads, so that the model can achieve faster convergence and more accurate estimation. Therefore, the state element is the continuous variable in Table 1, the action is the fusion weight of the auxiliary sensor, and the immediate benefit is the same as Q-learning.

[0132] The A3C algorithm includes the functions of the Actor network and the Critic network. The Actor uses a policy-based reinforcement learning method, takes the state s as input, updates the parameter θ of the policy function π(a|s; θ), and then selects an action. The Critic uses a value-based reinforcement learning method, takes the state s as input, and updates the action value function Parameter w of .

[0133] Among them, the model structure can be: a public network connects several worker threads, each thread has the same network structure as the public neural network, each thread will independently interact with the environment to obtain experience data, and calculate the gradient of the neural network loss function in the thread to update the public neural network.

[0134] In an exemplary embodiment, Figure 3G As shown, step S632 includes the following detailed sub-steps S6321-S63212.

[0135] Step S6321, input: the neural network parameters of the global part are θ, w, the number of globally shared iterations T, the global maximum number of iterations T max , the A3C neural network structure of this thread, corresponding parameters θ′, w′, the maximum length of a single iteration time series in the thread t max , state feature dimension, action set, step size α, β, decay factor γ;

[0136] Step S6322, initializing the time sequence t=1, and the number of global shared iteration rounds T=0;

[0137] Step S6323, reset the gradient update amount of Actor and Critic: dθ=0, dw=0;

[0138] Step S6324, synchronize parameters from the public neural network to the neural network of this thread: θ = θ′w = w′; t sart = t, initialization state s t ;

[0139] Step S6325: Based on the strategy π(a t |s t ;θ′) select action a t And execute, get reward r t and the new state s t+1 .

[0140] Step S6326, time step update t←t+1, T←T+1, until s t is the terminal state, or tt sart =t max , otherwise return to S6325;

[0141] Step S6327: Calculate the last time series

[0142] Among them, s t is the current state, and w′ is the network parameter of this thread.

[0143] Step S6328, loop through steps S63281 to S63283, loop (tstart -t+1) times, i is the number of cycles;

[0144] Step S63281, calculate the action value function Q(s, i) = r at each moment i +γQ(s,i+1);

[0145] Here, i is the number of cycles and γ is the decay factor.

[0146] Step S63282: Accumulate the local gradient update of the policy network Actor:

[0147]

[0148] Among them, θ is the neural network parameter of the global part, is the gradient function about θ′, a i It is an action. i is the state, θ′ and w′ are the neural network parameters of this thread.

[0149] Step S63283, local gradient update of the cumulative value network Critic:

[0150]

[0151] Step S6329: Update the model parameters of the global neural network θ=θ-αdθ, w=w-βdw. If T>T max , then return to step S6323;

[0152] Where α, β are the update step sizes.

[0153] Step S63210, output the neural network parameters θ, w of the global part; and the fusion weights of the auxiliary sensors.

[0154] Step S64: build a reinforcement learning network according to the algorithm flow, and train the network in complex environments and motions such as high speed, with GNSS signal, weak GNSS signal, no GNSS signal, indoor and outdoor switching, etc. In view of the uncertainty problem in scene switching and navigation, the adaptive hierarchical system can optimize the action online in real time, obtain the optimal combination of navigation information sources and the corresponding fusion weight, i.e., credibility, and input the feedback results into the data interaction layer, thereby realizing sensor optimization.

[0155] Step S7: Adopt the nonlinear fusion method of factor graph, allocate weights through centralized fusion module to realize adaptive fusion, and finally send the fusion result to the application layer.

[0156] In an exemplary embodiment, the specific implementation method of step S7 is:

[0157] Define X = (X1, ..., Xs ) is the state vector, s is the sliding window size, X i =(r i ,φ i ) is the global pose of node i, and the pose observation of the sensor between node i and node j is z ij =(r ij ,φ ij ), the adaptive information matrix is ​​Ω ij , r is the displacement, φ is the attitude. The optimal state estimation is sought through least squares estimation:

[0158]

[0159] Where l represents the measurement set and the error vector e ij = e(x i , x j , z ij ) The measurement information z between two nodes ij Constraint, x i is the state of node i, x j is the state of node j. In this embodiment, the information matrix is ​​the inverse matrix of the output weights of the reinforcement learning network, which is adjusted in real time with the credibility of the sensor. On the basis of selecting the optimal sensor combination, the combined navigation result is further optimized.

[0160] Example 4

[0161] According to an embodiment of the present invention, a hierarchical multi-layer multi-source fusion instant access navigation system for implementing the above method is also provided. Figure 4 As shown, the system includes: a fusion layer 42 and a data interaction layer 44.

[0162] The fusion layer 42 is configured to: evaluate the performance of multiple sensors used to collect navigation information in real time, and make decisions on the navigation information combination scheme in an intelligent manner based on the evaluated performance; adopt a nonlinear fusion method of an adaptive factor graph, assign weights based on the navigation information combination scheme to achieve adaptive fusion, obtain a fusion result, and navigate based on the fusion result.

[0163] In an exemplary embodiment, the fusion layer 42 is also configured to construct an abnormality rating criterion to describe the degree of influence of the uncertainty of the external environment on at least one of the multiple sensors in response to the uncertainty of the carrier maneuver and the randomness of the dynamic interference of the external environment; based on the observability analysis of the linear time-varying system, calculate the observability of the state variables of each combined navigation sub-filter to quantify the observable capability of at least one of the multiple sensors; based on the degree of influence and the observability, determine the reinforcement learning method, wherein the reinforcement learning method includes the Q-leaming algorithm and the asynchronous dominant actor-judge algorithm; construct a reinforcement learning network according to the determined reinforcement learning method, train the reinforcement learning network in a complex environment and motion, and use the reinforcement learning network to make decisions on the navigation information combination scheme to obtain the optimal navigation information combination scheme.

[0164] In an exemplary embodiment, in a simple environment, a state type is defined, and a Q-learning algorithm is used to discretize the continuous variables of the multiple sensors in segments by setting thresholds, and normalize them to achieve the action of controlling at least one of the multiple sensors, wherein the state type includes at least one of the following: auxiliary sensor uncertainty, auxiliary sensor abnormality, speed, GNSS status, number of satellites and satellite signals, the action is to switch the sensor, and the immediate benefit is the inverse of the sum of squares of the position error at the current moment; in a complex environment and motion scene, an asynchronous advantage actor-judge algorithm is used to interact and learn with the environment in multiple threads respectively, and train a neural network model so that the neural network model converges and outputs an accurate estimation effect, wherein the state element input to the neural network model is the continuous variable, the action output by the neural network model is the fusion weight of the auxiliary sensor among the multiple sensors, and the immediate benefit is the inverse of the sum of squares of the position error at the current moment.

[0165] In an exemplary embodiment, the fusion layer 42 is further configured to: receive at least one of the following input parameters: number of iterations, state set, action set, step size, decay factor and exploration rate; based on the input parameters, use the ∈-greedy method, based on the current state in the state set, select the current action, and execute the current action to obtain a new state and reward; update the value function and state based on the new state and the reward, and output an estimate of the true value of all states and actions and the switching of auxiliary sensors based on the value function and state.

[0166] In an exemplary embodiment, the fusion layer 42 is further configured to: receive at least one of the following input parameters: neural network parameters of the global part, globally shared iteration rounds, global maximum number of iterations, A3C neural network structure of this thread, maximum length of a single iteration time series within a thread, state feature dimension, action set, step size, and attenuation factor; select an action based on the decided navigation information combination scheme and execute the action to obtain a reward and a new state; based on the reward and the new state, calculate the action value function at each moment, accumulate the local gradient update of the Actor, and accumulate the local gradient update of the Critic, and update the parameters of the neural network model; use the updated neural network model to output the neural network parameters of the global part and the fusion weights of the auxiliary sensors among the multiple sensors based on the input parameters.

[0167] In an exemplary embodiment, the fusion layer 42 is further configured to seek an optimal state estimate by least squares estimation:

[0168]

[0169] Where X = (X1, ..., X s ) is the state vector, and the adaptive information matrix is ​​Ω ij , l represents the measurement set, e ij = e(x i , x j , z ij ) is the error vector, z ij For measurement information, is the transpose of the error vector, i is node i, and j is node j.

[0170] In an exemplary embodiment, the fusion layer 42 is also configured to construct a combined navigation model, determine the state equation and measurement equation of the multi-source fusion navigation system; based on the state equation and the measurement equation, perform state estimation using Kalman filtering in the sub-filter to obtain accurate navigation information; perform local filtering based on the accurate navigation information to obtain the optimal estimate of the local filtering, use the residual and the observed prediction error covariance as the abnormality input parameters, and use the prediction error covariance and the state prediction posterior error covariance as the observability input parameters.

[0171] In an exemplary embodiment, the fusion layer 42 is further configured to: divide the fusion result into multiple partitions based on the grammatical rules, and determine the structure of the multiple partitions of the fusion result; determine each refined symbol value in the refined symbol information within the fusion result based on the average value of the number of predetermined blocks within the multiple partitions, and use a quantization step to calibrate the each refined symbol value, wherein the quantization step depends on the reference quantization step adopted by the predetermined space vector associated with the fusion result; associate each refined symbol value after the quantization step with the multiple partitions, and linearly combine the average values ​​of the refined symbol values ​​after the quantization step of two adjacent partitions in the multiple partitions; and refine the structure of the two adjacent partitions based on the average value of the refined symbol values ​​of the two adjacent partitions after the linear combination.

[0172] In an exemplary embodiment, the system also includes a data interaction layer 44, which is configured as follows: the sensor MCU of the sensor layer writes the resolved navigation information into the real-time message queue through the established data frame specification, and the fusion layer 42MCU, as a receiving node, completes the message transmission by collecting the navigation information data in the message queue, wherein the real-time message queue is used to collect and send the internal sensor data of the multiple sensors; the sensor MCU completes the resolution of the navigation information, and at the same time sends the task ID to the fusion layer 42MCU in the polling request interface, and after multiple communication cycles, the sensor MCU sends the navigation information after receiving the polling request, the polling request interface is cleared, and the message is transmitted from the sensor layer to the fusion layer 42; the instant access sensor that meets the hardware and interface standards completes the raw data collection and navigation information resolution, and after the system's instant access manager determines that an instant access sensor has been connected, it allocates the corresponding hardware resources and automatically adjusts the device-related parameters to ensure that the sensor layer navigation information is transmitted to the fusion layer 42 after the connection is completed.

[0173] Example 5

[0174] According to an embodiment of the present invention, a hierarchical multi-layer multi-source fusion instant access navigation system for implementing the above method is also provided. Figure 5 As shown, the system includes a sensor layer, a data interaction layer, a fusion layer 42 and an application layer.

[0175] Among them, the sensor layer includes internal sensors composed of MEMS gyroscopes, MEMS accelerometers, magnetometers, satellite receivers, etc., expandable sensors composed of visual odometers, laser radars, etc., external instant access sensors composed of laser rangefinders, millimeter wave rangefinders, etc., and embedded sensor microcontroller units (MCUs) integrated in each sensor. Different sensors transmit data with their microcontroller units (MCUs) through data buses such as SPI, I2C, CAN, etc. or other serial ports according to the protocols supported by their hardware; the data interaction layer 44 includes a wired and wireless transmission combination composed of "body area network / local area network + serial port", and there may be specific different connection modes such as Bluetooth, WIFI, USB serial port, etc. according to the data transmission capabilities of different sensors; the fusion layer 42 includes a local fusion layer 42 and a centralized fusion layer 42, which respectively perform combined filtering of observation information and optimal navigation information combination decision; the application layer includes navigation information application equipment such as handheld terminals, water equipment, unmanned vehicles on the road, and unmanned aerial vehicles in the air, to achieve navigation positioning in different application environments. Each navigation information source in the sensor layer can receive observation data, pre-process the observation data through the processor at the sensor end, and output positioning information. The data exchange layer 44 can connect the sensor layer and the fusion layer 42 to achieve communication with the host.

[0176] The local fusion layer 42 in the fusion layer 42 can perform Kalman filtering on the observation information of the inertial information and the auxiliary information source, and send the optimized results to the centralized fusion layer 42. The centralized fusion layer 42 in the fusion layer 42 evaluates the performance of the sensor in real time, selects the corresponding reinforcement learning algorithm according to the complexity of the environment, obtains the optimal navigation information combination scheme through reinforcement learning, and feeds the combination result back to the interaction layer to realize the selection of the sensor. The application layer can receive the multi-sensor fusion information realized by the adaptive factor graph, and send the fusion result to the application layer to realize the navigation positioning indoors, on the ground, in the air or on the water.

[0177] Through the above structure, this embodiment achieves the following advantages: This embodiment divides the access method of the sensor information source combined navigation mode into three levels: internal sensors, expandable sensors and external instant access sensors, which correspond to three message collection and sending mechanisms respectively. The three mechanisms can realize asynchronous transmission, application decoupling and current limiting and peak elimination of sensor data, avoid transmission congestion and signal stacking when connecting to multiple sensors in complex, dynamic and large-scale scenarios, and can reasonably allocate limited hardware resources, ensure system fault tolerance, and achieve strong navigation robustness and safe and reliable positioning.

[0178] The hierarchical multi-level framework adopted in this embodiment can realize instant access to navigation information sources, and the navigation information sources are independent of each other, and the hardware architecture and instant access software interface specifications of the open all-source navigation system are realized in terms of hardware and interface.

[0179] Example 6

[0180] The embodiment of the present invention further provides a storage medium. Optionally, in this embodiment, the storage medium stores a program, and when the program is running, the methods in embodiments 1 to 3 can be implemented, which will not be repeated here. Optionally, in this embodiment, the storage medium may include but is not limited to: a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and other media that can store program codes.

[0181] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A hierarchical, multi-layer, multi-source fusion instant access navigation method, characterized in that: include: Evaluate the performance of multiple sensors used to collect navigation information in real time, and make decisions on the combination of navigation information in an intelligent way based on the evaluated performance; Adopting a nonlinear fusion method of an adaptive factor graph, allocating weights based on the navigation information combination scheme to achieve adaptive fusion, obtaining a fusion result, and navigating based on the fusion result; Among them, the performance of multiple sensors used to collect navigation information is evaluated in real time, and a decision on a navigation information combination scheme is made in an intelligent manner based on the evaluated performance, including: In view of the uncertainty of the carrier maneuver and the randomness of the dynamic interference of the external environment, an abnormality rating criterion is constructed to describe the influence degree of the external environment uncertainty on at least one sensor of the multiple sensors; Based on the observability analysis of the linear time-varying system, the observability of each combined navigation sub-filter state variable is calculated to quantify the observability of at least one sensor among the multiple sensors; Determining a reinforcement learning method based on the degree of influence and the observability, wherein the reinforcement learning method includes a Q-learning algorithm and an asynchronous advantage actor-critic algorithm; Constructing a reinforcement learning network according to the determined reinforcement learning method, training the reinforcement learning network in a complex environment and motion, and using the reinforcement learning network to make decisions on a navigation information combination scheme to obtain an optimal navigation information combination scheme; Wherein, based on the influence degree and the observability, determining a reinforcement learning method includes: In a simple environment, the state type is defined, and the Q-learning algorithm is used to discretize the continuous variables of the multiple sensors in segments by setting thresholds, and normalize them to achieve the action of controlling at least one of the multiple sensors, wherein the state type includes at least one of the following: auxiliary sensor uncertainty, auxiliary sensor abnormality, speed, GNSS state, number of satellites and satellite signals, the action is to switch the sensor on and off, and the instant benefit is the inverse of the sum of squares of the position errors at the current moment; In complex environments and motion scenarios, an asynchronous advantage actor-judge algorithm is adopted to conduct interactive learning with the environment in multiple threads and train the neural network model so that the neural network model converges and outputs accurate estimation effects, wherein the state element input into the neural network model is the continuous variable, the action output by the neural network model is the fusion weight of the auxiliary sensors among the multiple sensors, and the immediate benefit is the inverse of the sum of squares of the position errors at the current moment.

2. The method according to claim 1, characterized in that In a simple environment, a state type is defined, a Q-learning algorithm is used, and continuous variables of the multiple sensors are discretized piecewise by setting thresholds, and normalized to achieve an action of controlling at least one of the multiple sensors, including: Receive at least one of the following input parameters: number of iterations, state set, action set, step size, decay factor, and exploration rate; Based on the input parameters, using the ∈-greedy method, based on the current state in the state set, select the current action, and execute the current action to obtain a new state and reward; The value function and the state are updated based on the new state and the reward, and estimates of the true values ​​of all states and actions and the switching of auxiliary sensors are output based on the value function and the state.

3. The method according to claim 1, characterized in that In complex environments and motion scenes, the asynchronous advantage actor-critic algorithm A3C is used to interact with the environment in multiple threads to train the neural network model so that the neural network model converges and outputs accurate estimation results, including: Receive at least one of the following input parameters: global neural network parameters, globally shared iteration rounds, global maximum iteration number, A3C neural network structure of this thread, maximum length of single iteration time series in the thread, state feature dimension, action set, step size, decay factor; Selecting an action based on the determined navigation information combination scheme and executing the action to obtain a reward and a new state; Based on the reward and the new state, the action value function at each moment is calculated, the local gradient update of the Actor is accumulated, and the local gradient update of the Critic is accumulated, and the parameters of the neural network model are updated; The updated neural network model is used to output the neural network parameters of the global part and the fusion weights of the auxiliary sensors in the multiple sensors based on the input parameters.

4. The method according to claim 1, characterized in that: Adopting a nonlinear fusion method of an adaptive factor graph, allocating weights based on the navigation information combination scheme to achieve adaptive fusion, obtaining a fusion result, and navigating based on the fusion result includes: Seek the optimal state estimate through least squares estimation: Where X = (X1, L, X s ) is the state vector, and the adaptive information matrix is ​​Ω ij , l represents the measurement set, e ij =e(x i ,x j ,z ij ) is the error vector, z ij For measurement information, is the transpose of the error vector, i is node i, and j is node j.

5. The method according to claim 1, characterized in that Before making a decision on a navigation information combination scheme in an intelligent manner based on the evaluated performance, the method further includes: Construct an integrated navigation model and determine the state equations and measurement equations of the multi-source fusion navigation system; Based on the state equation and the measurement equation, a Kalman filter is used in the sub-filter to perform state estimation to obtain accurate navigation information; Based on the accurate navigation information, local filtering is performed to obtain the optimal estimate of the local filtering, the residual and the observed prediction error covariance are used as abnormality input parameters, and the prediction error covariance and the state prediction posterior error covariance are used as observability input parameters.

6. The method according to claim 5, characterized in that Before determining the state equation and the measurement equation of the multi-source fusion navigation system, the method further includes: The sensor MCU of the sensor layer writes the solved navigation information into the real-time message queue through the formulated data frame specification, and the fusion layer MCU, as a receiving node, completes the message transmission by collecting the navigation information data in the message queue, wherein the real-time message queue is used to collect and send the internal sensor data of the multiple sensors; The sensor MCU completes the calculation of the navigation information and sends the task ID to the fusion layer MCU in the polling request interface. After multiple communication cycles, the sensor MCU sends the navigation information after receiving the polling request, and the polling request interface is cleared to complete the transmission of the message from the sensor layer to the fusion layer. The raw data collection and navigation information solution are completed by the instant access sensors that meet the hardware and interface standards. After the system's instant access manager determines that an instant access sensor has been connected, it allocates corresponding hardware resources and automatically adjusts device-related parameters to ensure that the navigation information of the sensor layer is transmitted to the fusion layer after the connection is completed.

7. The method according to claim 1, characterized in that The method further includes: identifying the structure in the fusion result according to the grammatical rule includes: Based on the grammatical rule, the fusion result is divided into a plurality of partitions, and the structure of the plurality of partitions of the fusion result is determined; Determine each refined symbol value in the refined symbol information in the fusion result based on an average value of the number of predetermined blocks in the plurality of partitions, and calibrate each refined symbol value using a quantization step size, wherein the quantization step size depends on a reference quantization step size adopted by a predetermined space vector associated with the fusion result; Associating the refined symbol values ​​after the quantization step with the multiple partitions respectively, and linearly combining the average values ​​of the refined symbol values ​​after the quantization step of two adjacent partitions among the multiple partitions; The structures of the two adjacent partitions are refined based on an average value of the refined symbol values ​​of the two adjacent partitions after linear combination.

8. A hierarchical, multi-layer, multi-source fusion instant access navigation system, characterized in that: The method comprises a fusion layer, wherein the fusion layer is configured to: Evaluate the performance of multiple sensors used to collect navigation information in real time, and make decisions on the combination of navigation information in an intelligent way based on the evaluated performance; Adopting a nonlinear fusion method of an adaptive factor graph, allocating weights based on the navigation information combination scheme to achieve adaptive fusion, obtaining a fusion result, and navigating based on the fusion result; Wherein, the navigation system is further configured to: construct an abnormality rating criterion to describe the degree of influence of the uncertainty of the external environment on at least one of the multiple sensors in view of the uncertainty of the carrier maneuver and the randomness of the dynamic interference of the external environment; calculate the observability of the state variables of each combined navigation sub-filter based on the observability analysis of the linear time-varying system to quantify the observable capability of at least one of the multiple sensors; determine a reinforcement learning method based on the degree of influence and the observability, wherein the reinforcement learning method includes a Q-learning algorithm and an asynchronous dominant actor-judge algorithm; construct a reinforcement learning network according to the determined reinforcement learning method, train the reinforcement learning network in a complex environment and motion, and use the reinforcement learning network to make decisions on the navigation information combination scheme to obtain the optimal navigation information combination scheme; Wherein, based on the influence degree and the observability, a reinforcement learning method is determined, including: in a simple environment, a state type is defined, a Q-learning algorithm is used, and continuous variables of the multiple sensors are discretized in segments by setting thresholds, and normalized to achieve the action of controlling at least one of the multiple sensors, wherein the state type includes at least one of the following: auxiliary sensor uncertainty, auxiliary sensor abnormality, speed, GNSS state, number of satellites and satellite signals, the action is to switch the sensor on and off, and the immediate benefit is the inverse of the sum of squares of the position error at the current moment; In complex environments and motion scenarios, an asynchronous advantage actor-judge algorithm is adopted to conduct interactive learning with the environment in multiple threads and train the neural network model so that the neural network model converges and outputs accurate estimation effects, wherein the state element input into the neural network model is the continuous variable, the action output by the neural network model is the fusion weight of the auxiliary sensors among the multiple sensors, and the immediate benefit is the inverse of the sum of squares of the position errors at the current moment.

Images