Micro-pnt multi-source data fusion navigation method and system based on spatiotemporal neural network

Abstract

The application discloses a micro PNT multi-source data fusion navigation method and system based on a space-time neural network.Under satellite navigation denial conditions, a global time axis is established by using a high-stable-frequency clock, and multi-source data such as inertial, visual, geomagnetic and barometric data are synchronously collected.Through extracting scene features and combining prior templates, the environment scene is intelligently determined and the observation quality is evaluated, and then the navigation mode and network configuration are dynamically scheduled.Combining the structured geographic prior features of the local map, the multi-source sequence is sent into a multi-modal space-time cross-perception navigation network, the navigation states such as position, speed and attitude and the model confidence are directly output, and the standardization packaging and integrity monitoring are carried out.The application realizes continuous space-time navigation with high precision, high reliability, self-adaptation and long-time anti-drift capability in a complex environment.

Description

Technical Field

[0001] This invention relates to the technical field of autonomous navigation of micro PNTs, and in particular to a micro PNT multi-source data fusion navigation method and system based on spatiotemporal neural networks. Background Technology

[0002] With the rapid development of information and intelligent technologies, positioning, navigation, and timing (PNT) systems have become an important component of national critical infrastructure. Currently, Global Navigation Satellite Systems (GNSS) are the primary means of acquiring high-precision spatiotemporal information. However, due to the extremely weak satellite navigation signals reaching the ground, GNSS systems are inherently vulnerable and easily susceptible to obstruction by complex environments, as well as intentional or unintentional electromagnetic interference and deception. In recent years, interference and deception incidents targeting satellite navigation and timing terminals have become increasingly serious, leading to a growing number of cases of integrated positioning, navigation, and timing failures. In complex scenarios such as urban canyons, dense forests, underground spaces, and indoor environments, satellite signals often face severe attenuation or even complete loss, seriously threatening the safe and reliable operation of miniaturized equipment such as unmanned platforms and robots.

[0003] To address the interference and spoofing issues of satellite navigation signals, existing technologies typically employ interference monitoring and suppression methods. For example, multi-antenna array-based techniques utilize direction-of-arrival estimation and beamforming, along with eigenvalue testing using cyclic correlation matrices and singular value decomposition, to eliminate spoofing interference. Dual-antenna-based techniques employ carrier phase double-difference measurements for spoofing identification. Additionally, there are detection methods based on single-antenna receiver signal characteristics (such as power, carrier-to-noise ratio, and correlation peak anomalies). However, these existing methods have significant drawbacks: multi-antenna array systems are bulky, power-consuming, and expensive, making them difficult to apply to miniaturized platforms with stringent requirements for size, weight, and power consumption; the performance of dual-antenna detection methods is highly dependent on the baseline length between antennas, exhibiting extremely poor performance under short baseline and low positioning accuracy conditions; and single-antenna feature detection methods often struggle to effectively identify covert and slowly changing intelligent spoofing strategies. Furthermore, none of the aforementioned methods can provide continuous navigation capabilities when satellite signals are completely suppressed or denied.

[0004] Given the unreliability of satellite navigation in complex electromagnetic environments, multi-source sensor fusion technologies (such as inertial, visual, geomagnetic, and barometric fusion) that do not rely on satellite navigation have become a research hotspot in the field of micro PNTs. However, existing multi-source fusion navigation methods are mostly based on the traditional Kalman filter framework, which has many limitations. First, traditional methods rely on strict linearized mathematical models and Gaussian noise assumptions, making it difficult to fully explore and characterize the deep nonlinear characteristics of various heterogeneous sensors in complex dynamic environments. Second, traditional fusion strategies usually use fixed observation weights or simple threshold switching, lacking the ability to intelligently perceive and adaptively evaluate the current environmental scene. When the vehicle moves between different scenes (such as from an open outdoor area to an indoor or underground space), visual features may become invalid due to sudden changes in illumination, and geomagnetic observations may be distorted due to local anomalies. Traditional methods cannot dynamically schedule navigation modes, which can easily lead to filter divergence and rapid accumulation of errors.

[0005] Furthermore, while existing deep learning-based navigation algorithms have improved feature extraction capabilities to some extent, they often simply stitch together data from various sensors, lacking rigorous temporal and spatial synchronization and cross-sensory modeling of multimodal heterogeneous data. Simultaneously, existing methods rarely integrate structured prior information such as local map tiles, road topology, and geomagnetic fingerprints into the neural network's computation process, resulting in a lack of absolute geographical benchmark constraints for state recursion under long-endurance conditions. More critically, existing black-box neural network models generally lack quantitative evaluation of output confidence and real-time monitoring mechanisms for anomalies, failing to meet the stringent system integrity requirements of high-reliability upper-layer applications.

[0006] In summary, overcoming the hardware limitations of existing satellite navigation anti-interference methods and breaking through the bottlenecks of poor adaptability, rapid error accumulation, and lack of integrity monitoring in traditional multi-source fusion models under complex scenarios, and proposing a micro-PNT multi-source fusion method that can intelligently perceive the scene, dynamically schedule and configure, and deeply integrate multimodal spatiotemporal features under satellite navigation denial conditions, has become an urgent technical challenge to be solved in the current navigation field. Summary of the Invention

[0007] In view of this, the present invention provides a micro PNT multi-source data fusion navigation method based on spatiotemporal neural networks. The purpose is to provide a micro PNT multi-source data fusion navigation method based on spatiotemporal neural networks. By establishing global high-precision time synchronization, intelligently sensing the environment and scene, dynamically scheduling the navigation working mode, introducing structured geographical prior constraints, and constructing a multimodal spatiotemporal cross-sensing network, and introducing a multi-layer redundant integrity monitoring and closed-loop feedback mechanism at the output end, the method can provide high-precision, high-reliability, adaptive, and long-endurance anti-drift continuous spatiotemporal navigation information for miniaturized platforms in complex and ever-changing environments.

[0008] To achieve the above objectives, the present invention provides a micro-PNT multi-source data fusion navigation method based on spatiotemporal neural networks, comprising the following steps: S1: Under the condition that the satellite navigation receiver does not participate in the operation, a global time axis of the system is established with a high stable frequency clock, and the microelectromechanical system inertial measurement unit, vision sensor, magnetometer and barometer are synchronously triggered to complete the acquisition of raw observation data. A unified timestamp is applied to the acceleration, angular velocity, image frame, geomagnetic field strength, geomagnetic direction and barometric altitude data output by various heterogeneous sensors to form a multi-source raw observation sequence with time correspondence. S2: Extract image texture density, number and distribution stability of visual feature points, geomagnetic fluctuation amplitude, degree of geomagnetic anomaly, barometric altitude change trend, and inertial motion excitation features from the multi-source original observation sequence as scene feature quantities. Call the scene prior template corresponding to the current coarse location index or initial task binding information, compare the scene feature quantities with the scene prior template item by item, determine the current environmental scene type, and output the observation quality scores of visual observation availability, geomagnetic observation stability, inertial data continuity, barometric altitude reliability, and multi-source data synchronization consistency. S3: Based on the environmental scene type and observation quality score, and combined with the current task stage's requirements for navigation accuracy, state update frequency, real-time performance, and endurance, query the preset strategy library to determine the current navigation working mode. The navigation working modes include visual enhancement mode, geomagnetic assisted mode, inertial-dominated mode, multimodal joint mode, and low-power maintenance mode. Generate input configuration parameters for the multimodal spatiotemporal cross-sensing navigation network, such as the multi-source data time window length, inertial sequence sampling density, visual feature point filtering threshold, geomagnetic observation input frequency, barometric altitude input switch, and model inference frequency. At the same time, issue corresponding sampling rate control commands. S4: Based on the initial task binding information or the current position estimate output from the previous calculation cycle, retrieve and load the local map tiles of the corresponding task area, extract the geographical environment feature information of terrain elevation, road boundaries, building outlines, traversable areas, geomagnetic fingerprints, and indoor spatial structures, and convert the geographical environment feature information into structured prior inputs of elevation prior features, road passage topology features, building boundary features, geomagnetic fingerprint features, regional traversability features, and spatial coding features; S5: The multi-source original observation sequence, input configuration parameters, and structured prior input are sent into the multimodal spatiotemporal cross-sensing navigation network. The multimodal spatiotemporal cross-sensing navigation network is composed of a multimodal original feature encoding layer, a heterogeneous feature cross-sensing layer, a long-range temporal state recursion layer, and a navigation state regression output layer cascaded together. It directly completes the regression mapping from multi-source observation data to navigation state quantities, outputs the device's position, velocity, attitude, and time-related state information, and simultaneously outputs the model confidence and abnormal observation response index. S6: The position, velocity, attitude, time-related state information, model confidence, current navigation working mode label, and observation source availability status are encapsulated and formatted according to the standard protocol defined by the upper-layer application platform. The standardized spatiotemporal information stream is output in real time through a serial bus or network interface. Real-time monitoring is performed on the working conditions of decreased model confidence, increased abnormal observation response, continuous output state abrupt change, and decreased consistency of multi-source data synchronization. Alarm information is generated and reported to the upper-layer system.

[0009] As a further improvement of the present invention: Optionally, step S1 further includes: Using a high-stability clock signal output from a temperature-compensated crystal oscillator or a microelectromechanical system (MEMS) temperature-compensated crystal oscillator as the system time reference source, a unified global clock pulse is generated by frequency division. This global clock pulse serves as a synchronization trigger signal, driving the MEMS inertial measurement unit to collect acceleration and angular velocity data, driving the vision sensor to collect image frames or visual feature point data obtained from image frame front-end processing, driving the magnetometer to collect geomagnetic field strength and geomagnetic direction data, and driving the barometer to collect barometric altitude data according to a preset sampling period. The acceleration, angular velocity, image frames or visual feature points, geomagnetic field strength, geomagnetic direction, and barometric altitude data are marked with a unified timestamp based on their arrival time on the system bus and their correspondence with the global time axis, forming a multi-source original observation sequence with a time correspondence.

[0010] Optionally, step S2 further includes: Scene features are extracted from the multi-source raw observation sequence by channel. The extraction of scene features is divided into four parallel processing paths: visual, geomagnetic, barometric, and inertial. In the visual path, image frames are grayscaled and gradient operators are applied. The edge density within a unit area is counted as the image texture density, and the number of visual feature points is counted using a corner detection operator. The distribution stability is calculated by combining the positional differences of visual feature points between adjacent image frames. In the geomagnetic path, the standard deviation within a sliding window of the geomagnetic field strength time series data is calculated as the geomagnetic fluctuation amplitude, and the deviation between the current observation value and the geomagnetic reference value in the scene prior template is used as the degree of geomagnetic anomaly. In the barometric path, the rate of change of barometric altitude data within a window is calculated as the barometric altitude change trend. In the inertial path, the energy value within a window of acceleration and angular velocity data is calculated as the inertial motion excitation feature. The scene prior template corresponding to the coarse location index or initial task binding information is invoked. The scene prior template provides reference feature intervals for forest, city, desert, underground space, and indoor environmental scene types respectively. The extracted scene features are then matched item by item with the scene prior template using a weighted average to obtain a comprehensive discrimination score for the current environmental scene type. The comprehensive discrimination score is calculated using the following formula: ; in, Indicates the current observation of the environmental scene type. The comprehensive judgment score, It draws inspiration from five environmental scenarios: forest, city, desert, underground space, and indoors. This represents the total number of scene features involved in the discrimination; exp represents the natural exponential function. Indicates the currently extracted number of... The values ​​of each scene feature quantity ; and These respectively represent the environment / scene type in the scene prior template. The corresponding number Reference mean and reference standard deviation of each scene feature; Indicates the first Discrimination weights for scene features ; take The environment scene type that achieves the maximum value is taken as the current environment scene type; After determining the current environmental scene type, observation quality scores are calculated according to five dimensions: visual, geomagnetic, inertial, barometric pressure, and multi-source synchronization consistency. Visual observation availability is obtained by normalizing and weighting the image texture density, the number of visual feature points, and the stability of their distribution. Geomagnetic observation stability is obtained by weighting the geomagnetic fluctuation amplitude and the inverse of the geomagnetic anomaly degree. Inertial data continuity is obtained by the statistical stability of the output residuals of the inertial measurement unit at adjacent sampling times. Barometric pressure reliability is obtained by the degree of deviation of the barometric pressure change trend from the prior interval of the scene. Multi-source data synchronization consistency is obtained by comprehensively measuring the timestamp deviation of each sensor and the number of missing markers. All observation quality scores are normalized to the range of 0 to 1.

[0011] Optionally, step S3 further includes: A preset strategy library is constructed, indexed by a combination of environmental scene type and observation quality score, with visual enhancement mode, geomagnetic-assisted mode, inertial-dominated mode, multimodal joint mode, and low-power preservation mode as output entries. The visual enhancement mode is suitable for scenarios where visual observation availability is higher than a set threshold; the geomagnetic-assisted mode is suitable for scenarios where geomagnetic observation stability is higher than a set threshold and visual observation availability is lower than a set threshold; the inertial-dominated mode is suitable for scenarios where both visual observation availability and geomagnetic observation stability are lower than a set threshold; the multimodal joint mode is suitable for scenarios where multiple types of observation sources are higher than a set threshold; and the low-power preservation mode is suitable for mission phases where navigation accuracy requirements are reduced and endurance is prioritized. After querying the preset strategy library to determine the current navigation working mode, input configuration parameters for the multimodal spatiotemporal cross-sensing navigation network are generated based on the current navigation working mode. The input configuration parameters employ an adaptive adjustment strategy coupled with the observation quality score. Taking the model inference frequency as an example, the calculation formula is as follows: ; in, Indicates time Inference frequency of multimodal spatiotemporal cross-sensing navigation network; This indicates the baseline inference frequency set by the current navigation operating mode; Indicates time The comprehensive observation quality score is obtained by normalizing and weighting visual observation availability, geomagnetic observation stability, inertial data continuity, barometric altitude reliability, and multi-source data synchronization consistency. Indicates time The intensity of the carrier motion excitation is obtained by normalizing the mean values ​​of the magnitudes of acceleration and angular velocity within the window; This represents the lower limit preservation coefficient of inference frequency, a fixed constant ranging from 0 to 1; The motion excitation sensitivity coefficient is represented by a non-negative fixed constant. The multi-source data time window length, inertial sequence sampling density, visual feature point screening threshold, geomagnetic observation input frequency, and barometric altitude input switch are all adaptively generated according to the same coupling strategy as the model inference frequency. The above input configuration parameters are sent to the input organization stage of the multimodal spatiotemporal cross-sensing navigation network, and at the same time, corresponding sampling rate control commands are generated and sent to the sensor synchronous acquisition stage, so that the data input scale, model inference frequency and system resource usage of each sensor are dynamically matched with the comprehensive observation quality score according to the current navigation working mode.

[0012] Optionally, step S4 further includes: Using the initial task binding information or the current position estimate output from the previous solution cycle as the query index, local map tiles corresponding to the task area are retrieved from the map feature library. These local map tiles contain geographical environmental features such as terrain elevation, road boundaries, building outlines, traversable areas, geomagnetic fingerprints, and indoor spatial structures. The terrain elevation information is resampled according to a preset grid resolution to obtain prior elevation features. The road boundaries and traversable area information are abstracted into a directed graph representation of nodes and connecting edges to obtain road channel topology features. The building outline information is encoded according to location coordinates and closed polygon vertex sequences to obtain building boundary features. The geomagnetic fingerprint is statistically analyzed according to spatial grid cells to obtain geomagnetic fingerprint features. The traversable area information is encoded using a binary mask to obtain area traversability features. Finally, a unified spatial coordinate encoding is performed on all features according to the current location candidate area to obtain spatially encoded features corresponding to the candidate area. The aforementioned prior features of elevation, road topology, building boundary features, geomagnetic fingerprint features, regional accessibility features, and spatial coding features are compiled into structured prior inputs and simultaneously fed into the multimodal spatiotemporal cross-sensing navigation network along with the multi-source original observation sequence, serving as auxiliary inputs and constraint information for the model during the state regression process.

[0013] Optionally, step S5 further includes: A multimodal spatiotemporal cross-sensing navigation network is constructed, which consists of a multimodal original feature encoding layer, a heterogeneous feature cross-sensing layer, a long-range temporal state recursion layer, and a navigation state regression output layer cascaded together. The multimodal original feature encoding layer sets up five parallel sub-encoding paths for inertial, visual, geomagnetic, barometric, and structured prior inputs respectively. Each parallel sub-encoding path is composed of a one-dimensional convolution operator, a batch normalization operator, a nonlinear activation operator, and a downsampling pooling operator cascaded in sequence. Each input is independently feature-encoded according to its physical dimensions, and at the end of the path, the outputs of each path are projected to a unified dimension through a linear mapping operator to obtain a multimodal feature vector sequence with a consistent shape. The heterogeneous feature cross-sensing layer performs cross-modal association modeling on the multimodal feature vector sequence. It applies linear transformations of the query projection matrix, key projection matrix, and value projection matrix to the feature vector of each modality, obtaining the query vector, key vector, and value vector. Then, the normalized result of the inner product between the query vector and all modal key vectors is used as the association response weight. The value vectors of all modalities are then weighted and summed to obtain the fused feature. The formula for calculating the cross-modal association response weight is as follows: ; in, Indicates time Modal For modes Cross-modal correlation response weights; This represents the set of modes composed of inertial, visual, geomagnetic, air pressure, and structured prior inputs. ; Indicates time Modal The query vector obtained after linear transformation of the query projection matrix; , Representing time respectively Modal , The key vector obtained after linear transformation of the key projection matrix; This represents the number of dimensions of the key vector. This indicates the transpose operation. Used to balance the inner product scale; and Representing time respectively Modal , The corresponding observation quality score is input by modal assignment based on the visual observation availability, geomagnetic observation stability, inertial data continuity, barometric altitude reliability, and multi-source data synchronization consistency. The long-range temporal state recursive layer recursively processes the fused feature sequence output by the heterogeneous feature cross-sensing layer. A gated recurrent unit, consisting of four parallel computation branches—input gate operator, forget gate operator, output gate operator, and candidate state generation operator—updates the hidden state vector and unit state vector time-by-time. This modeling addresses the long-term dynamic dependence of the carrier's motion state, the evolution of historical trajectories, and the implicit changes in inertial error accumulated over time. The hidden state update process of the gated recurrent unit is uniformly represented as follows: ; in, and Representing time respectively With time The hidden state vector; Indicates time The unit state vector; Indicates time The fused feature vector output by the heterogeneous feature cross-sensing layer; This represents a gated loop update function consisting of a cascaded input gate operator, a forget gate operator, an output gate operator, and a candidate state generation operator. The navigation state regression output layer concatenates the hidden state vector with the fused feature vector output by the heterogeneous feature cross-sensing layer along the channel dimension, and then regresses the time step by regressing through a cascaded mapping of fully connected operators and nonlinear activation operators. The system outputs position vector, velocity vector, attitude quaternion, and time-related state information, and simultaneously outputs model confidence and anomaly observation response index. The model confidence is directly predicted by the parallel uncertainty prediction branch of the regression output layer based on the hidden state vector. The anomaly observation response index is calculated by the deviation between the entropy value of the cross-modal correlation response weight on the modality set and the historical mean. To improve the stability of the multimodal spatiotemporal cross-sensing navigation network under satellite navigation denial conditions, the multimodal spatiotemporal cross-sensing navigation network simultaneously introduces a multi-task joint loss during the training phase, which includes position regression loss, velocity regression loss, attitude regression loss, geographic feature constraint consistency loss, and confidence calibration loss. The geographic feature constraint consistency loss is constructed by comparing the feasibility of the location obtained by the model regression on regional accessibility features and building boundary features.

[0014] Optionally, step S6 further includes: The position, velocity, attitude, time-related state information and model confidence are mapped, byte order aligned, checksum calculated and frame header and tail encapsulated according to the standard protocol defined by the upper-layer application platform to obtain a standardized spatiotemporal information stream, which is then output to the unmanned platform and robot platform in real time through a serial bus or network interface according to a preset output cycle. During the output of the standardized spatiotemporal information flow, the system integrity is monitored in real time. This system integrity monitoring includes four parallel decision channels: The first decision channel compares the sliding window mean of the model confidence with a set lower limit; when the sliding window mean of the model confidence is lower than the set lower limit, a model confidence decrease alarm is triggered. The second decision channel compares the sliding window mean of the abnormal observation response index with a set upper limit; when the sliding window mean of the abnormal observation response index is higher than the set upper limit, an abnormal observation response increase alarm is triggered. The third decision channel compares the difference magnitude between the position, velocity, and attitude state vectors of adjacent output cycles with a kinematic constraint threshold; when the difference magnitude is higher than the kinematic constraint threshold, an alarm is triggered. The system triggers a continuous output state change alarm; the fourth determination channel compares the sliding window mean of the multi-source data synchronization consistency score with a set lower limit, and triggers a multi-source data synchronization consistency decrease alarm when the sliding window mean of the multi-source data synchronization consistency score is lower than the set lower limit; when any determination channel triggers an alarm, the alarm type, alarm time, corresponding navigation working mode label, and observation source availability status are encapsulated into alarm information and reported to the upper layer system, and simultaneously fed back to the navigation working mode determination link, triggering a query of the preset strategy library to re-determine the navigation working mode, thereby forming a multi-layer redundancy guarantee mechanism between the multimodal spatiotemporal cross-sensing navigation network output layer, system integrity monitoring layer, and navigation working mode scheduling layer.

[0015] This invention also discloses a micro-PNT multi-source data fusion navigation system based on a spatiotemporal neural network, comprising: Sensor spatiotemporal reference module: When the satellite navigation receiver is not involved, a global time axis of the system is established with a high-stability clock. The microelectromechanical system inertial measurement unit, visual sensor, magnetometer and barometer are synchronously triggered to complete the acquisition of raw observation data. A unified timestamp is applied to the acceleration, angular velocity, image frame, geomagnetic field strength, geomagnetic direction and barometric altitude data output by various heterogeneous sensors to form a multi-source raw observation sequence with time correspondence. The environmental scene recognition module extracts image texture density, number and distribution stability of visual feature points, geomagnetic fluctuation amplitude, geomagnetic anomaly degree, barometric altitude change trend, and inertial motion excitation features from the multi-source original observation sequence as scene feature quantities. It calls the scene prior template corresponding to the current coarse location index or initial task binding information, compares the scene feature quantities with the scene prior template item by item, determines the current environmental scene type, and outputs observation quality scores for visual observation availability, geomagnetic observation stability, inertial data continuity, barometric altitude reliability, and multi-source data synchronization consistency. The mode management and scheduling module, based on the environmental scene type and observation quality score, and combined with the current task stage's requirements for navigation accuracy, status update frequency, real-time performance, and endurance, queries a preset strategy library to determine the current navigation working mode. The navigation working modes include visual enhancement mode, geomagnetic-assisted mode, inertial-dominated mode, multimodal joint mode, and low-power hold-up mode. It also generates input configuration parameters for the multimodal spatiotemporal cross-sensing navigation network, including the multi-source data time window length, inertial sequence sampling density, visual feature point filtering threshold, geomagnetic observation input frequency, barometric altitude input switch, and model inference frequency, while simultaneously issuing corresponding sampling rate control commands. Geographic feature constraint module: Based on the initial task binding information or the current location estimate output from the previous calculation cycle, retrieve and load the local map tiles of the corresponding task area, extract the geographic environmental feature information of terrain elevation, road boundaries, building outlines, traversable areas, geomagnetic fingerprints, and indoor spatial structures, and convert the geographic environmental feature information into structured prior inputs of elevation prior features, road passage topology features, building boundary features, geomagnetic fingerprint features, regional traversability features, and spatial coding features; Multimodal spatiotemporal cross-sensing navigation solution module: The multi-source original observation sequence, input configuration parameters, and structured prior input are fed into the multimodal spatiotemporal cross-sensing navigation network. The multimodal spatiotemporal cross-sensing navigation network is composed of a multimodal original feature encoding layer, a heterogeneous feature cross-sensing layer, a long-range temporal state recursion layer, and a navigation state regression output layer cascaded together. It directly completes the regression mapping from multi-source observation data to navigation state quantities, outputs the device's position, velocity, attitude, and time-related state information, and simultaneously outputs model confidence and abnormal observation response indicators. Interface status output module: The position, velocity, attitude, time-related status information, model confidence, current navigation working mode label, and observation source availability status are encapsulated and formatted according to the standard protocol defined by the upper-layer application platform. The standardized spatiotemporal information stream is output in real time through a serial bus or network interface. The module also monitors in real time the working conditions such as decreased model confidence, increased abnormal observation response, continuous output state abrupt change, and decreased synchronization consistency of multi-source data, generates alarm information, and reports it to the upper-layer system.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects: This invention establishes a deep adaptive coupling mechanism between low-level data processing and high-level mode scheduling, significantly improving the system's survivability and resource utilization efficiency in complex environments. Traditional methods often employ fixed sampling and computation frequencies, while this invention extracts the physical characteristics of multi-source heterogeneous sensors and combines them with scene prior templates to not only accurately determine the current environmental scene but also quantify the quality scores of each observation source. More importantly, this invention innovatively couples these observation quality scores directly with the intensity of carrier motion excitation, constructing an adaptive adjustment strategy that dynamically generates input configuration parameters such as the length of the multi-source data time window, sequence sampling density, and model inference frequency. This improvement enables the system to intelligently switch between visual enhancement, geomagnetic assistance, or low-power maintenance modes based on environmental conditions and motion status, maximizing the allocation of system computing power and power consumption while ensuring navigation accuracy.

[0017] Secondly, this invention substantially improves the internal structure and loss function of the multimodal spatiotemporal cross-sensing navigation network, effectively overcoming the "black box" blindness and long-term drift problems of conventional deep learning models in multi-source fusion. In the heterogeneous feature cross-sensing layer, this invention does not simply apply the conventional attention mechanism, but innovatively embeds the observation quality score obtained from the front end directly into the calculation formula of the cross-modal correlation response weight. This algorithm improvement enables the network to have clear physical guidance when performing feature interactions, forcing the model to suppress contaminated or malfunctioning sensor modalities at the bottom feature dimension and enhance the weight of high-confidence modalities. In addition, during the network training phase, this invention introduces structured prior inputs based on local map tile extraction and specifically constructs a geographic feature constraint consistency loss. This improvement transforms absolute geographic spatial constraints such as regional navigability and building boundaries into penalty terms for network regression, forcing the model to output a navigation state that conforms to the real-world physical topology, fundamentally curbing the unbounded accumulation of errors in the long-range temporal recursion process.

[0018] Finally, the various innovations of this invention are highly synergistic in logic and function, forming an inseparable closed-loop overall solution that achieves multi-layered redundancy protection from underlying hardware synchronization to upper-layer integrity monitoring. This invention breaks the traditional disconnect between sensor data acquisition, network model inference, and upper-layer integrity monitoring. Hardware-level synchronization with a unified timestamp at the front end provides a precise time reference for the network's long-term time-series recursion; the network output layer not only regresses the navigation state but also synchronously outputs the model confidence predicted by the hidden state vector and the abnormal observation response index calculated from the cross-modal weight entropy value; these deep indicators within the network are then directly integrated into the multi-channel judgment of the system integrity monitoring module. Once an alarm is triggered, the monitoring module immediately feeds back the status to the front-end navigation working mode scheduling stage, triggering a re-query of the policy library and parameter reconstruction. This tightly linked mechanism of "perception assessment - dynamic configuration - cross-fusion - integrity monitoring - feedback reconstruction" makes this invention no longer a static solution algorithm but an intelligent navigation system with a high degree of self-diagnosis and self-repair capabilities, greatly enhancing the creativity and engineering application value of the overall solution. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a micro-PNT multi-source data fusion navigation method based on a spatiotemporal neural network according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a micro PNT multi-source data fusion navigation system based on a spatiotemporal neural network according to an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0021] Example 1: A micro-PNT multi-source data fusion navigation method based on spatiotemporal neural networks, such as Figure 1 As shown, it includes the following steps: S1: Without the satellite navigation receiver participating, a high-frequency stable clock is used to establish the system's global time axis. This synchronously triggers the microelectromechanical system (MEMS) inertial measurement unit, visual sensor, magnetometer, and barometer to acquire raw observation data. A unified timestamp is applied to the acceleration, angular velocity, image frames, geomagnetic field strength, geomagnetic direction, and barometric altitude data output from various heterogeneous sensors, forming a multi-source raw observation sequence with time correspondence, including: Using a high-stability clock signal output from a temperature-compensated crystal oscillator or a microelectromechanical system (MEMS) temperature-compensated crystal oscillator as the system time reference source, a unified global clock pulse is generated by frequency division. This global clock pulse serves as a synchronization trigger signal, driving the MEMS inertial measurement unit to collect acceleration and angular velocity data, driving the vision sensor to collect image frames or visual feature point data obtained from image frame front-end processing, driving the magnetometer to collect geomagnetic field strength and geomagnetic direction data, and driving the barometer to collect barometric altitude data according to a preset sampling period. The acceleration, angular velocity, image frames or visual feature points, geomagnetic field strength, geomagnetic direction, and barometric altitude data are marked with a unified timestamp based on their arrival time on the system bus and their correspondence with the global time axis, forming a multi-source original observation sequence with a time correspondence.

[0022] In this embodiment, as an auxiliary technical feature, a timestamp anomaly repair mechanism is introduced: when the deviation between the arrival time of a certain sensor data and the trigger time of the global clock pulse is detected to be greater than a preset tolerance. In this embodiment Set to 100 microseconds, mark the data entry as timestamp abnormal; for timestamp abnormal data, use the nearest neighbor global clock pulse alignment strategy for repair, that is, force the timestamp to be aligned with the trigger time of the nearest global clock pulse, and record the repair operation in the auxiliary flag bit of the data entry; when the number of timestamp abnormalities of the same sensor exceeds 3 within 5 consecutive global clock cycles, the sensor is included in the list of observation sources to be downweighted, and participates in the calculation of multi-source data synchronization consistency score with a discount factor of 0.5, thereby improving the reliability of multi-source raw observation sequence under sensor jitter or link delay conditions.

[0023] S2: Extract image texture density, number and distribution stability of visual feature points, geomagnetic fluctuation amplitude, geomagnetic anomaly degree, barometric altitude change trend, and inertial motion excitation features from the multi-source raw observation sequence as scene feature quantities. Call the scene prior template corresponding to the current coarse location index or initial task binding information, compare the scene feature quantities with the scene prior template item by item, determine the current environmental scene type, and output the observation quality scores for visual observation availability, geomagnetic observation stability, inertial data continuity, barometric altitude reliability, and multi-source data synchronization consistency, including: Scene features are extracted from the multi-source raw observation sequence by channel. The extraction of scene features is divided into four parallel processing paths: visual, geomagnetic, barometric, and inertial. In the visual path, image frames are grayscaled and gradient operators are applied. The edge density within a unit area is counted as the image texture density, and the number of visual feature points is counted using a corner detection operator. The distribution stability is calculated by combining the positional differences of visual feature points between adjacent image frames. In the geomagnetic path, the standard deviation within a sliding window of the geomagnetic field strength time series data is calculated as the geomagnetic fluctuation amplitude, and the deviation between the current observation value and the geomagnetic reference value in the scene prior template is used as the degree of geomagnetic anomaly. In the barometric path, the rate of change of barometric altitude data within a window is calculated as the barometric altitude change trend. In the inertial path, the energy value within a window of acceleration and angular velocity data is calculated as the inertial motion excitation feature. In this embodiment, scene features are extracted from the multi-source original observation sequence by channel. The scene feature extraction is performed within each sliding analysis window with a length of 1 second, specifically as follows: On the visual path, for image frames The image is converted to grayscale, and the horizontal gradient is calculated using the Sobel gradient operator. gradient in the vertical direction The image texture density is calculated as the ratio of the number of pixels with a gradient magnitude greater than a threshold of 40 within a given unit area to the total number of pixels in that area. The unit region size is set to 32×32 pixels; the total number of visual feature points in the current image frame is counted using the FAST-9 corner detection operator as the visual feature point count. For adjacent image frames, the KLT optical flow operator is used to track the positions of feature points, and the reciprocal of the standard deviation of the change in feature point positions is calculated as the distribution stability. ; Time series data of geomagnetic field intensity along the geomagnetic path The sample standard deviation is calculated as the geomagnetic fluctuation amplitude within a sliding window of 1 second. ; to the current observation value Geomagnetic reference value in the scene prior template absolute difference between As the degree of geomagnetic anomaly ; Along the pressure path, the pressure altitude A least-squares linear fit is performed within a 1-second sliding window, and the slope obtained from the fit is used as the trend of pressure altitude change. ; On the inertial path, for the acceleration vector With angular velocity vector The mean of the sum of squared magnitudes is calculated within a 1-second sliding window and used as the characteristic of inertial motion excitation. ; The scene prior template corresponding to the coarse location index or initial task binding information is invoked. The scene prior template provides reference feature intervals for forest, city, desert, underground space, and indoor environmental scene types respectively. The extracted scene features are then matched item by item with the scene prior template using a weighted average to obtain a comprehensive discrimination score for the current environmental scene type. The comprehensive discrimination score is calculated using the following formula: ; in, Indicates the current observation of the environmental scene type. The comprehensive judgment score, It is derived from five environmental scene types: forest, city, desert, underground space, and indoor; exp represents the natural exponential function. This represents the total number of scene features involved in the discrimination, as shown in this embodiment. ; Indicates the currently extracted number of... The values ​​of each scene feature quantity ; and These respectively represent the environment / scene type in the scene prior template. The corresponding number Reference mean and reference standard deviation of each scene feature; Indicates the first The discrimination weights of scene features in this embodiment to The values ​​were set sequentially to 0.18, 0.15, 0.12, 0.18, 0.12, 0.10, and 0.15. ; take The environment scene type that achieves the maximum value is taken as the current environment scene type; In this embodiment, after determining the current environmental scene type, the observation quality score is calculated according to five dimensions: visual, geomagnetic, inertial, air pressure, and multi-source synchronization consistency. Visual observation availability is further categorized into: The stability of geomagnetic observations was obtained by normalizing and weighting image texture density, the number of visual feature points, and distribution stability with weights of 0.3, 0.4, and 0.3, respectively. It is obtained by weighting the inverse of the geomagnetic fluctuation amplitude and the inverse of the geomagnetic anomaly degree with a weight of 0.5; inertial data continuity The barometric altitude confidence level is obtained by normalizing the inverse of the sample variance of the residuals output by the inertial measurement unit at 10 consecutive sampling times; It is obtained by normalizing the deviation of the barometric altitude change trend from the center point of the prior interval of the scenario; multi-source data synchronization consistency. The five observation quality scores are all normalized to the interval between 0 and 1, obtained by normalizing the product of the root mean square of the timestamp deviation of each sensor and the number of missing markers.

[0024] S3: Based on the environmental scene type and observation quality score, and considering the current task stage's requirements for navigation accuracy, state update frequency, real-time performance, and endurance, query the preset strategy library to determine the current navigation working mode. The navigation working modes include visual enhancement mode, geomagnetic-assisted mode, inertial-dominated mode, multimodal joint mode, and low-power maintenance mode. Simultaneously, generate input configuration parameters for the multimodal spatiotemporal cross-sensing navigation network, including the multi-source data time window length, inertial sequence sampling density, visual feature point filtering threshold, geomagnetic observation input frequency, barometric altitude input switch, and model inference frequency. Also, issue corresponding sampling rate control commands, including: A preset strategy library is constructed, indexed by a combination of environmental scene type and observation quality score, with visual enhancement mode, geomagnetic assist mode, inertial-dominated mode, multimodal joint mode, and low-power maintenance mode as output entries. In this embodiment, the specific scheduling rule is: when visual observation availability... Enter visual enhancement mode at that time; local magnetic observation stability and Enter geomagnetic-assisted mode at this time; when and When it enters the inertia-dominated mode; when , , When all three parameters are greater than 0.6, the system enters multimodal joint mode; when the task phase is marked as cruise standby and the remaining battery life percentage is lower than the set threshold, the system enters low power consumption hold mode. The input configuration parameters employ an adaptive adjustment strategy coupled with the observation quality score. Taking the model inference frequency as an example, the calculation formula is as follows: ; in, Indicates time Inference frequency of multimodal spatiotemporal cross-sensing navigation network; This represents the reference inference frequency set by the current navigation operating mode. In this embodiment, it corresponds to the visual enhancement mode, geomagnetic assist mode, inertial-dominated mode, multimodal joint mode, and low-power hold-up mode. Set to 50Hz, 30Hz, 100Hz, 80Hz, and 10Hz respectively; Indicates time The comprehensive observation quality score is obtained by normalizing and weighting the visual observation availability, geomagnetic observation stability, inertial data continuity, barometric altitude reliability, and multi-source data synchronization consistency with weights of 0.25, 0.20, 0.25, 0.10, and 0.20, respectively. Indicates time The excitation intensity of the carrier motion is obtained by normalizing the mean values ​​of the magnitudes of acceleration and angular velocity within a 1-second window to the interval of 0 to 1 using the Sigmoid function; This represents the lower bound preservation coefficient for inference frequency, in this embodiment. Set to 0.4; This represents the motion excitation sensitivity coefficient, in this embodiment... The value is set to 0.5; the multi-source data time window length, inertial sequence sampling density, visual feature point filtering threshold, geomagnetic observation input frequency, and barometric altitude input switch are all adaptively generated according to the same coupling strategy as the model inference frequency; The above input configuration parameters are sent to the input organization stage of the multimodal spatiotemporal cross-sensing navigation network, and at the same time, corresponding sampling rate control commands are generated and sent to the sensor synchronous acquisition stage, so that the data input scale, model inference frequency and system resource usage of each sensor are dynamically matched with the comprehensive observation quality score according to the current navigation working mode.

[0025] S4: Based on the initial task binding information or the current location estimate output from the previous calculation cycle, retrieve and load the local map tiles of the corresponding task area, extract the geographical environment feature information of terrain elevation, road boundaries, building outlines, navigable areas, geomagnetic fingerprints, and indoor spatial structures, and convert the geographical environment feature information into structured prior inputs of elevation prior features, road passage topology features, building boundary features, geomagnetic fingerprint features, regional navigability features, and spatial coding features, including: Using the initial task binding information or the current position estimate output from the previous calculation cycle as the query index, local map tiles corresponding to the task area are retrieved from the map feature library. In this embodiment, the local map tiles are organized in a geographic coordinate grid, and each tile covers a 500m × 500m area centered on the current position. The local map tiles contain geographic environmental feature information such as terrain elevation, road boundaries, building outlines, passable areas, geomagnetic fingerprints, and indoor space structures. In this embodiment, the terrain elevation information is resampled at a preset grid resolution of 1 meter × 1 meter to obtain elevation prior features, which are represented as a two-dimensional tensor of size 500 × 500. The road boundary and passable area information are abstracted into a directed graph representation of nodes and connecting edges to obtain road passage topology features. The maximum number of nodes in the directed graph is set to 256, and the maximum number of connecting edges is set to 512. Each node is described using 32-dimensional position encoding and 8-dimensional attribute encoding. The building outline information is encoded according to position coordinates and closed polygon vertex sequences. Each building retains a maximum of 16 vertices, and a maximum of [number missing] vertices are recorded. Sixty-four buildings were recorded to obtain building boundary features. The geomagnetic fingerprint was then analyzed using a 5m x 5m spatial grid to obtain reference intensity and reference direction, resulting in geomagnetic fingerprint features. These features were represented as a 100 x 100 x 4 three-dimensional tensor, with four channels corresponding to the reference intensity and three directional components. The passable area information was encoded using a binary mask to obtain area passability features, which were represented as a 500 x 500 binary tensor. A unified spatial coordinate encoding was performed on all features according to the candidate areas at the current location, resulting in 64-dimensional spatial encoded features. The aforementioned prior features of elevation, road topology, building boundary features, geomagnetic fingerprint features, regional accessibility features, and spatial coding features are compiled into structured prior inputs and simultaneously fed into the multimodal spatiotemporal cross-sensing navigation network along with the multi-source original observation sequence, serving as auxiliary inputs and constraint information for the model during the state regression process.

[0026] S5: The multi-source original observation sequence, input configuration parameters, and structured prior input are fed into a multimodal spatiotemporal cross-sensing navigation network. This network consists of a cascaded multimodal original feature encoding layer, a heterogeneous feature cross-sensing layer, a long-range temporal state recursion layer, and a navigation state regression output layer. It directly completes the regression mapping from multi-source observation data to navigation state quantities, outputting the device's position, velocity, attitude, and time-related state information, and simultaneously outputting model confidence and abnormal observation response indicators, including: A multimodal spatiotemporal cross-sensing navigation network is constructed, which consists of a multimodal original feature encoding layer, a heterogeneous feature cross-sensing layer, a long-range temporal state recursion layer, and a navigation state regression output layer cascaded together. In this embodiment, the multimodal original feature encoding layer sets up five parallel sub-encoding paths for inertial, visual, geomagnetic, barometric, and structured prior inputs respectively. Each parallel sub-encoding path consists of a one-dimensional convolution operator, a batch normalization operator, a ReLU nonlinear activation operator, and a max pooling downsampling operator cascaded sequentially. The specific configuration is as follows: the inertial sub-encoding path has an input dimension of 6×200, and sequentially passes through a one-dimensional convolution operator with a kernel size of 7 and an output channel of 32, a one-dimensional convolution operator with a kernel size of 5 and an output channel of 64, and a one-dimensional convolution operator with a kernel size of 3 and an output channel of 128. Each convolution operator is followed by a batch normalization, ReLU, and a max pooling operator with a pooling window of 2; the visual sub-encoding... The input to each pathway is a sequence of visual feature points, which are sequentially processed by a three-layer one-dimensional convolution operator with a kernel size of 3 and output channels of 32, 64, and 128. The geomagnetic sub-encoding pathway has an input dimension of 4×40, adopts the same structure as the inertial pathway but with half the number of output channels. The barometric sub-encoding pathway has an input dimension of 1×20 and uses a two-layer one-dimensional convolution operator. The structured prior pathway uses two-dimensional convolution branches to process the elevation prior features and regional navigability features, graph convolution branches to process the road channel topology features, and two-dimensional convolution branches to process the geomagnetic fingerprint features. The outputs of each branch are flattened and then concatenated. At the end of each pathway, a linear mapping operator projects the outputs of each pathway onto a unified dimension of 128, resulting in a multimodal feature vector sequence with a consistent shape. In this embodiment, the heterogeneous feature cross-sensing layer performs cross-modal association modeling on the multimodal feature vector sequence, and applies a query projection matrix to the feature vector of each modality. Key projection matrix and value projection matrix A linear transformation is performed, with the dimensions of the query projection matrix, key projection matrix, and value projection matrix all set to 128×128, resulting in a query vector. Key vector AND value vector The normalized result of the inner product between the query vector and all modal key vectors is used as the association response weight. The value vectors of all modalities are weighted and summed to obtain the fusion feature. The formula for calculating the cross-modal association response weight is as follows: ; in, Indicates time Modal For modes Cross-modal correlation response weights; This represents the set of modes composed of inertial, visual, geomagnetic, air pressure, and structured prior inputs. ; Indicates time Modal The query vector obtained after linear transformation of the query projection matrix; , Representing time respectively Modal , The key vector obtained after linear transformation of the key projection matrix; This represents the number of dimensions of the key vector. This indicates the transpose operation. Used to balance the inner product scale; and Representing time respectively Modal , The corresponding observation quality score is input by the visual observation availability, geomagnetic observation stability, inertial data continuity, barometric altitude reliability, and multi-source data synchronization consistency according to modal attribution; in this embodiment, the heterogeneous feature cross-sensing layer adopts a 4-head parallel attention structure, each head has a dimension of 32, and after parallel calculation, it is restored to 128 dimensions through a linear mapping layer; In this embodiment, the long-range temporal state recursion layer consists of two stacked gated recurrent units, with each layer's hidden state vector dimension set to 256. It recursively processes the fused feature sequence output by the heterogeneous feature cross-perception layer. The hidden state update process of the gated recurrent unit is uniformly represented as follows: ; in, and Representing time respectively With time The hidden state vector; Indicates time The unit state vector; Indicates time The fused feature vector output by the heterogeneous feature cross-sensing layer; This represents a gated loop update function consisting of a cascaded input gate operator, a forget gate operator, an output gate operator, and a candidate state generation operator. The four gate operators are respectively implemented by cascading a fully connected matrix of dimension 384×256 and a Tanh activation function. In this embodiment, the navigation state regression output layer will hide the state vector. The fused feature vectors output from the heterogeneous feature cross-sensing layer are concatenated along the channel dimension to obtain a 640-dimensional feature vector. This vector is then cascaded through two fully connected operators (256-dimensional and 128-dimensional) and a LeakyReLU nonlinear activation operator, and output through four parallel regression branches to determine the time steps. The model confidence is obtained by predicting the hidden state vector using the Softplus activation function based on the hidden state vector by the parallel uncertainty prediction branch, and the output is a 3D variance vector; the anomaly observation response index is calculated by the degree of deviation between the entropy value of the cross-modal correlation response weight on the modal set and the historical moving mean. In this embodiment, the multimodal spatiotemporal cross-sensing navigation network employs a multi-task joint loss function during the training phase. The expression is: ; in, , These represent the position regression loss and velocity regression loss, respectively, both calculated using smoothed L1 loss. The attitude regression loss is represented by the quaternion geodesic distance loss calculation. The geographical feature constraint consistency loss is represented by the sum of the difference between the binary mask value of the location obtained from the model regression on the regional accessibility feature and 1, and the negative values ​​of the directed distances between the location and each closed polygon in the building boundary feature. The confidence level calibration loss is calculated using the negative log-likelihood loss. , , , , The weighting coefficients for each loss term are set to 1.0, 0.5, 1.0, 0.3, and 0.2 respectively in this embodiment. In this embodiment, the Adam optimizer is used during the training phase, and the initial learning rate is set to... The weight decay coefficient is set to The Cosine annealing learning rate scheduling strategy is used to anneal the learning rate to a certain value within 200 training rounds. The batch size was set to 64. The training dataset was derived from measured multi-source observation sequences and corresponding high-precision reference trajectories collected in five typical environments: forest, city, desert, underground space, and indoor. The reference trajectories were provided by total station, motion capture system, or post-processing high-precision integrated navigation system.

[0027] S6: The position, velocity, attitude, time-related state information, model confidence, current navigation working mode label, and observation source availability status are encapsulated and formatted according to the standard protocol defined by the upper-layer application platform. A standardized spatiotemporal information stream is output in real time via a serial bus or network interface. Real-time monitoring is performed for conditions such as decreased model confidence, increased abnormal observation response, continuous output state abrupt changes, and decreased consistency of multi-source data synchronization. Alarm information is generated and reported to the upper-layer system, including: The position, velocity, attitude, time-related state information and model confidence are mapped, byte order aligned, checksum calculated and frame header and tail encapsulated according to the standard protocol defined by the upper-layer application platform to obtain a standardized spatiotemporal information stream, which is then output to the unmanned platform, individual soldier terminal or robot platform in real time through a serial bus or network interface according to a preset output cycle. In this embodiment, system integrity is monitored in real time. The system integrity monitoring includes four parallel decision channels: the first decision channel compares the sliding window mean of the model confidence with a set lower limit. In comparison, the sliding window length in this embodiment is set to 50 output cycles. Set to 0.6 when the sliding window mean of the model confidence is lower than The model confidence level decreases and an alarm is triggered; the sliding window mean of the response index of the second judgment channel to abnormal observations is compared with the set upper limit. Comparison, Set to 0.4, when the sliding window mean of the anomaly observation response index is higher than The system triggers an alarm for an abnormal observation response increase; the third judgment channel compares the difference magnitude between the position, velocity, and attitude state vectors of adjacent output cycles with kinematic constraint thresholds. The position difference magnitude threshold is set to 5 meters / output cycle, the velocity difference magnitude threshold is set to 3 meters / second / output cycle, and the attitude difference magnitude threshold is set to 10 degrees / output cycle. When any difference magnitude exceeds the corresponding threshold, a continuous output state change alarm is triggered; the fourth judgment channel compares the sliding window mean of the multi-source data synchronization consistency score with a set lower limit. Comparison, Set to 0.5 when the sliding window mean of the multi-source data synchronization consistency score is lower than 0.5. This triggers an alarm indicating a decrease in the consistency of multi-source data synchronization. In this embodiment, when any determination channel triggers an alarm, the alarm type, alarm time, corresponding navigation working mode label, and observation source availability status are encapsulated into alarm information and reported to the upper-layer system. At the same time, the alarm is fed back to the navigation working mode determination stage, triggering a query of the preset strategy library to re-determine the navigation working mode. In this embodiment, as an auxiliary technical feature, a hierarchical alarm and navigation mode rollback coordination mechanism is introduced: when a single judgment channel triggers an alarm, the mode management and scheduling link rolls back the current navigation working mode by one level to a direction with lower resource consumption and higher fault tolerance. The rollback order is as follows: multimodal joint mode, visual enhancement mode or geomagnetic assisted mode, inertial-dominated mode, and low-power maintenance mode. When the same alarm is triggered repeatedly within three consecutive sliding windows or when two or more judgment channels trigger alarms simultaneously, the alarm level is upgraded to a critical alarm, and the system is forced to switch to the inertial-dominated mode. At the same time, a task path adjustment suggestion is sent to the upper-layer system via broadcast message. When the critical alarm lasts for more than 30 seconds without recovery, the continuous output state is maintained at the most recent valid output value, and the model confidence is forcibly decayed to below 0. The upper-layer system then takes over the navigation task. This forms a multi-layered redundancy guarantee mechanism between the multimodal spatiotemporal cross-sensing navigation network output layer, the system integrity monitoring layer, and the navigation working mode scheduling layer.

[0028] Example 2: This invention also discloses a micro-PNT multi-source data fusion navigation system based on a spatiotemporal neural network, such as... Figure 2 As shown, it includes the following modules: Sensor spatiotemporal reference module: When the satellite navigation receiver is not involved, a global time axis of the system is established with a high-stability clock. The microelectromechanical system inertial measurement unit, visual sensor, magnetometer and barometer are synchronously triggered to complete the acquisition of raw observation data. A unified timestamp is applied to the acceleration, angular velocity, image frame, geomagnetic field strength, geomagnetic direction and barometric altitude data output by various heterogeneous sensors to form a multi-source raw observation sequence with time correspondence. The environmental scene recognition module extracts image texture density, number and distribution stability of visual feature points, geomagnetic fluctuation amplitude, geomagnetic anomaly degree, barometric altitude change trend, and inertial motion excitation features from the multi-source original observation sequence as scene feature quantities. It calls the scene prior template corresponding to the current coarse location index or initial task binding information, compares the scene feature quantities with the scene prior template item by item, determines the current environmental scene type, and outputs observation quality scores for visual observation availability, geomagnetic observation stability, inertial data continuity, barometric altitude reliability, and multi-source data synchronization consistency. The mode management and scheduling module, based on the environmental scene type and observation quality score, and combined with the current task stage's requirements for navigation accuracy, status update frequency, real-time performance, and endurance, queries a preset strategy library to determine the current navigation working mode. The navigation working modes include visual enhancement mode, geomagnetic-assisted mode, inertial-dominated mode, multimodal joint mode, and low-power hold-up mode. It also generates input configuration parameters for the multimodal spatiotemporal cross-sensing navigation network, including the multi-source data time window length, inertial sequence sampling density, visual feature point filtering threshold, geomagnetic observation input frequency, barometric altitude input switch, and model inference frequency, while simultaneously issuing corresponding sampling rate control commands. Geographic feature constraint module: Based on the initial task binding information or the current location estimate output from the previous calculation cycle, retrieve and load the local map tiles of the corresponding task area, extract the geographic environmental feature information of terrain elevation, road boundaries, building outlines, traversable areas, geomagnetic fingerprints, and indoor spatial structures, and convert the geographic environmental feature information into structured prior inputs of elevation prior features, road passage topology features, building boundary features, geomagnetic fingerprint features, regional traversability features, and spatial coding features; Multimodal spatiotemporal cross-sensing navigation solution module: The multi-source original observation sequence, input configuration parameters, and structured prior input are fed into the multimodal spatiotemporal cross-sensing navigation network. The multimodal spatiotemporal cross-sensing navigation network is composed of a multimodal original feature encoding layer, a heterogeneous feature cross-sensing layer, a long-range temporal state recursion layer, and a navigation state regression output layer cascaded together. It directly completes the regression mapping from multi-source observation data to navigation state quantities, outputs the device's position, velocity, attitude, and time-related state information, and simultaneously outputs model confidence and abnormal observation response indicators. Interface status output module: The position, velocity, attitude, time-related status information, model confidence, current navigation working mode label, and observation source availability status are encapsulated and formatted according to the standard protocol defined by the upper-layer application platform. The standardized spatiotemporal information stream is output in real time through a serial bus or network interface. The module also monitors in real time the working conditions such as decreased model confidence, increased abnormal observation response, continuous output state abrupt change, and decreased synchronization consistency of multi-source data, generates alarm information, and reports it to the upper-layer system.

[0029] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0030] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0031] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A micro-PNT multi-source data fusion navigation method based on spatiotemporal neural networks, characterized in that, Includes the following steps: S1: Under satellite navigation denial conditions, a global time axis of the system is established with a high-stability clock, and the microelectromechanical system inertial measurement unit, vision sensor, magnetometer and barometer are synchronously triggered to complete the acquisition of raw observation data. A unified timestamp is applied to the acceleration, angular velocity, image frame, geomagnetic field strength, geomagnetic direction and barometric altitude data output by various heterogeneous sensors to form a multi-source raw observation sequence with time correspondence. S2: Extract image texture density, number and distribution stability of visual feature points, geomagnetic fluctuation amplitude, degree of geomagnetic anomaly, barometric altitude change trend, and inertial motion excitation features from the multi-source original observation sequence as scene feature quantities. Call the scene prior template corresponding to the current coarse location index or initial task binding information, compare the scene feature quantities with the scene prior template item by item, determine the current environmental scene type, and output the observation quality scores of visual observation availability, geomagnetic observation stability, inertial data continuity, barometric altitude reliability, and multi-source data synchronization consistency. S3: Based on the environmental scene type and observation quality score, and combined with the current task stage's requirements for navigation accuracy, state update frequency, real-time performance, and endurance, query the preset strategy library to determine the current navigation working mode. The navigation working modes include visual enhancement mode, geomagnetic assisted mode, inertial-dominated mode, multimodal joint mode, and low-power maintenance mode. Generate input configuration parameters for the multimodal spatiotemporal cross-sensing navigation network, such as the multi-source data time window length, inertial sequence sampling density, visual feature point filtering threshold, geomagnetic observation input frequency, barometric altitude input switch, and model inference frequency. At the same time, issue corresponding sampling rate control commands. S4: Based on the initial task binding information or the current position estimate output from the previous calculation cycle, retrieve and load the local map tiles of the corresponding task area, extract the geographical environment feature information of terrain elevation, road boundaries, building outlines, traversable areas, geomagnetic fingerprints, and indoor spatial structures, and convert the geographical environment feature information into structured prior inputs of elevation prior features, road passage topology features, building boundary features, geomagnetic fingerprint features, regional traversability features, and spatial coding features; S5: The multi-source original observation sequence, input configuration parameters, and structured prior input are sent into the multimodal spatiotemporal cross-sensing navigation network. The multimodal spatiotemporal cross-sensing navigation network is composed of a multimodal original feature encoding layer, a heterogeneous feature cross-sensing layer, a long-range temporal state recursion layer, and a navigation state regression output layer cascaded together. It directly completes the regression mapping from multi-source observation data to navigation state quantities, outputs the device's position, velocity, attitude, and time-related state information, and simultaneously outputs the model confidence and abnormal observation response index. S6: The position, velocity, attitude, time-related status information, model confidence, current navigation working mode label, and observation source availability status are encapsulated and formatted according to the standard protocol defined by the upper-layer application platform. The standardized spatiotemporal information stream is output in real time through a serial bus or network interface. Real-time monitoring is performed on the working conditions of decreased model confidence, increased abnormal observation response, continuous output state abrupt change, and decreased consistency of multi-source data synchronization. Alarm information is generated and reported to the upper-layer system to adjust the task path or replan the navigation strategy.

2. The micro-PNT multi-source data fusion navigation method based on spatiotemporal neural network according to claim 1, characterized in that, Step S1 includes: Using a high-stability clock signal output from a temperature-compensated crystal oscillator or a microelectromechanical system (MEMS) temperature-compensated crystal oscillator as the system time reference source, a unified global clock pulse is generated by frequency division. This global clock pulse serves as a synchronization trigger signal, driving the MEMS inertial measurement unit to collect acceleration and angular velocity data, driving the vision sensor to collect image frames or visual feature point data obtained from image frame front-end processing, driving the magnetometer to collect geomagnetic field strength and geomagnetic direction data, and driving the barometer to collect barometric altitude data according to a preset sampling period. The acceleration, angular velocity, image frames or visual feature points, geomagnetic field strength, geomagnetic direction, and barometric altitude data are marked with a unified timestamp based on their arrival time on the system bus and their correspondence with the global time axis, forming a multi-source original observation sequence with a time correspondence.

3. The micro-PNT multi-source data fusion navigation method based on spatiotemporal neural network according to claim 2, characterized in that, Step S2 includes: Scene features are extracted from the multi-source raw observation sequence by channel. The extraction of scene features is divided into four parallel processing paths: visual, geomagnetic, barometric, and inertial. In the visual path, image frames are grayscaled and gradient operators are applied. The edge density within a unit area is counted as the image texture density, and the number of visual feature points is counted using a corner detection operator. The distribution stability is calculated by combining the positional differences of visual feature points between adjacent image frames. In the geomagnetic path, the standard deviation within a sliding window of the geomagnetic field strength time series data is calculated as the geomagnetic fluctuation amplitude, and the deviation between the current observation value and the geomagnetic reference value in the scene prior template is used as the degree of geomagnetic anomaly. In the barometric path, the rate of change of barometric altitude data within a window is calculated as the barometric altitude change trend. In the inertial path, the energy value within a window of acceleration and angular velocity data is calculated as the inertial motion excitation feature. The scene prior template corresponding to the coarse location index or initial task binding information is invoked. The scene prior template provides reference feature intervals for forest, city, desert, underground space, and indoor environmental scene types respectively. The extracted scene features are then matched item by item with the scene prior template using a weighted average to obtain a comprehensive discrimination score for the current environmental scene type. The comprehensive discrimination score is calculated using the following formula: ； in, Indicates the current observation of the environmental scene type. The comprehensive judgment score, It is derived from five environmental scene types: forest, city, desert, underground space, and indoor; exp represents the natural exponential function. This represents the total number of scene features involved in the discrimination. Indicates the currently extracted number of... The values ​​of each scene feature quantity ; and These respectively represent the environment / scene type in the scene prior template. The corresponding number Reference mean and reference standard deviation of each scene feature; Indicates the first Discrimination weights for scene features ; take The environment scene type that achieves the maximum value is taken as the current environment scene type; After determining the current environmental scene type, observation quality scores are calculated according to five dimensions: visual, geomagnetic, inertial, barometric pressure, and multi-source synchronization consistency. Visual observation availability is obtained by normalizing and weighting the image texture density, the number of visual feature points, and the stability of their distribution. Geomagnetic observation stability is obtained by weighting the geomagnetic fluctuation amplitude and the inverse of the geomagnetic anomaly degree. Inertial data continuity is obtained by the statistical stability of the output residuals of the inertial measurement unit at adjacent sampling times. Barometric pressure reliability is obtained by the degree of deviation of the barometric pressure change trend from the prior interval of the scene. Multi-source data synchronization consistency is obtained by comprehensively measuring the timestamp deviation of each sensor and the number of missing markers. All observation quality scores are normalized to the range of 0 to 1.

4. The micro-PNT multi-source data fusion navigation method based on spatiotemporal neural network according to claim 3, characterized in that, Step S3 includes: A preset strategy library is constructed, indexed by a combination of environmental scene type and observation quality score, with visual enhancement mode, geomagnetic-assisted mode, inertial-dominated mode, multimodal joint mode, and low-power preservation mode as output entries. The visual enhancement mode is suitable for scenarios where visual observation availability is higher than a set threshold; the geomagnetic-assisted mode is suitable for scenarios where geomagnetic observation stability is higher than a set threshold and visual observation availability is lower than a set threshold; the inertial-dominated mode is suitable for scenarios where both visual observation availability and geomagnetic observation stability are lower than a set threshold; the multimodal joint mode is suitable for scenarios where multiple types of observation sources are higher than a set threshold; and the low-power preservation mode is suitable for mission phases where navigation accuracy requirements are reduced and endurance is prioritized. After querying the preset strategy library to determine the current navigation working mode, input configuration parameters for the multimodal spatiotemporal cross-sensing navigation network are generated based on the current navigation working mode. The input configuration parameters employ an adaptive adjustment strategy coupled with the observation quality score. Taking the model inference frequency as an example, the calculation formula is as follows: ； in, Indicates time Inference frequency of multimodal spatiotemporal cross-sensing navigation network; This indicates the baseline inference frequency set by the current navigation operating mode; Indicates time The comprehensive observation quality score is obtained by normalizing and weighting visual observation availability, geomagnetic observation stability, inertial data continuity, barometric altitude reliability, and multi-source data synchronization consistency. Indicates time The intensity of the carrier motion excitation is obtained by normalizing the mean values ​​of the magnitudes of acceleration and angular velocity within the window; This represents the lower limit preservation coefficient of inference frequency, a fixed constant ranging from 0 to 1; The motion excitation sensitivity coefficient is represented by a non-negative fixed constant. The multi-source data time window length, inertial sequence sampling density, visual feature point screening threshold, geomagnetic observation input frequency, and barometric altitude input switch are all adaptively generated according to the same coupling strategy as the model inference frequency. The above input configuration parameters are sent to the input organization stage of the multimodal spatiotemporal cross-sensing navigation network, and at the same time, corresponding sampling rate control commands are generated and sent to the sensor synchronous acquisition stage, so that the data input scale, model inference frequency and system resource usage of each sensor are dynamically matched with the comprehensive observation quality score according to the current navigation working mode.

5. The micro-PNT multi-source data fusion navigation method based on spatiotemporal neural network according to claim 4, characterized in that, Step S4 includes: Using the initial task binding information or the current position estimate output from the previous calculation cycle as the query index, local map tiles corresponding to the task area are retrieved from the map feature library. These local map tiles contain geographical environmental features such as terrain elevation, road boundaries, building outlines, traversable areas, geomagnetic fingerprints, and indoor spatial structures. The terrain elevation is resampled according to a preset grid resolution to obtain prior elevation features. The road boundaries and traversable area information are abstracted into a directed graph representation of nodes and connecting edges to obtain road channel topology features. The building outlines are encoded according to their location coordinates and closed polygon vertex sequences to obtain building boundary features. The geomagnetic fingerprint is statistically analyzed using spatial grid cells to determine its reference intensity and reference direction, resulting in geomagnetic fingerprint features. The traversable area information is encoded using a binary mask to obtain area traversability features. Finally, a unified spatial coordinate encoding is performed on all features according to the current location candidate area to obtain spatially encoded features corresponding to the candidate area. The aforementioned prior features of elevation, road topology, building boundary features, geomagnetic fingerprint features, regional accessibility features, and spatial coding features are compiled into structured prior inputs and simultaneously fed into the multimodal spatiotemporal cross-sensing navigation network along with the multi-source original observation sequence, serving as auxiliary inputs and constraint information for the model during the state regression process.

6. The micro-PNT multi-source data fusion navigation method based on spatiotemporal neural network according to claim 5, characterized in that, Step S5 includes: A multimodal spatiotemporal cross-sensing navigation network is constructed, which consists of a multimodal original feature encoding layer, a heterogeneous feature cross-sensing layer, a long-range temporal state recursion layer, and a navigation state regression output layer cascaded together. The multimodal original feature encoding layer sets up five parallel sub-encoding paths for inertial, visual, geomagnetic, barometric, and structured prior inputs respectively. Each parallel sub-encoding path is composed of a one-dimensional convolution operator, a batch normalization operator, a nonlinear activation operator, and a downsampling pooling operator cascaded in sequence. Each input is independently feature-encoded according to its physical dimensions, and at the end of the path, the outputs of each path are projected to a unified dimension through a linear mapping operator to obtain a multimodal feature vector sequence with a consistent shape. The heterogeneous feature cross-sensing layer performs cross-modal association modeling on the multimodal feature vector sequence. It applies linear transformations of the query projection matrix, key projection matrix, and value projection matrix to the feature vector of each modality, obtaining the query vector, key vector, and value vector. Then, the normalized result of the inner product between the query vector and all modal key vectors is used as the association response weight. The value vectors of all modalities are then weighted and summed to obtain the fused feature. The formula for calculating the cross-modal association response weight is as follows: ； in, Indicates time Modal For modes Cross-modal correlation response weights; This represents the set of modes composed of inertial, visual, geomagnetic, air pressure, and structured prior inputs. ; Indicates time Modal The query vector obtained after linear transformation of the query projection matrix; , Representing time respectively Modal , The key vector obtained after linear transformation of the key projection matrix; This represents the number of dimensions of the key vector. This indicates the transpose operation. Used to balance the inner product scale; and Representing time respectively Modal , The corresponding observation quality score is input by modal assignment based on the visual observation availability, geomagnetic observation stability, inertial data continuity, barometric altitude reliability, and multi-source data synchronization consistency. The long-range temporal state recursive layer recursively processes the fused feature sequence output by the heterogeneous feature cross-sensing layer. A gated recurrent unit, consisting of four parallel computation branches—input gate operator, forget gate operator, output gate operator, and candidate state generation operator—updates the hidden state vector and unit state vector time-by-time. This modeling addresses the long-term dynamic dependence of the carrier's motion state, the evolution of historical trajectories, and the implicit changes in inertial error accumulated over time. The hidden state update process of the gated recurrent unit is uniformly represented as follows: ； in, and Representing time respectively With time The hidden state vector; Indicates time The unit state vector; Indicates time The fused feature vector output by the heterogeneous feature cross-sensing layer; This represents a gated loop update function consisting of a cascaded input gate operator, a forget gate operator, an output gate operator, and a candidate state generation operator. The navigation state regression output layer concatenates the hidden state vector with the fused feature vector output by the heterogeneous feature cross-sensing layer along the channel dimension, and then regresses the time step by regressing through a cascaded mapping of fully connected operators and nonlinear activation operators. The system outputs position vector, velocity vector, attitude quaternion, and time-related state information, and simultaneously outputs model confidence and anomaly observation response index. The model confidence is directly predicted by the parallel uncertainty prediction branch of the regression output layer based on the hidden state vector. The anomaly observation response index is calculated by the deviation between the entropy value of the cross-modal correlation response weight on the modality set and the historical mean. To improve the stability of the multimodal spatiotemporal cross-sensing navigation network under satellite navigation denial conditions, the multimodal spatiotemporal cross-sensing navigation network simultaneously introduces a multi-task joint loss during the training phase, which includes position regression loss, velocity regression loss, attitude regression loss, geographic feature constraint consistency loss, and confidence calibration loss. The geographic feature constraint consistency loss is constructed by comparing the feasibility of the location obtained by the model regression on regional accessibility features and building boundary features.

7. The micro-PNT multi-source data fusion navigation method based on spatiotemporal neural network according to claim 6, characterized in that, Step S6 includes: The position, velocity, attitude, time-related state information and model confidence are mapped, byte order aligned, checksum calculated and frame header and tail encapsulated according to the standard protocol defined by the upper-layer application platform to obtain a standardized spatiotemporal information stream, which is then output to the unmanned platform, individual soldier terminal or robot platform in real time through a serial bus or network interface according to a preset output cycle. During the output of the standardized spatiotemporal information flow, the system integrity is monitored in real time. This system integrity monitoring includes four parallel decision channels: The first decision channel compares the sliding window mean of the model confidence with a set lower limit; when the sliding window mean of the model confidence is lower than the set lower limit, a model confidence decrease alarm is triggered. The second decision channel compares the sliding window mean of the abnormal observation response index with a set upper limit; when the sliding window mean of the abnormal observation response index is higher than the set upper limit, an abnormal observation response increase alarm is triggered. The third decision channel compares the difference magnitude between the position, velocity, and attitude state vectors of adjacent output cycles with a kinematic constraint threshold; when the difference magnitude is higher than the kinematic constraint threshold, an alarm is triggered. The system triggers a continuous output state change alarm; the fourth determination channel compares the sliding window mean of the multi-source data synchronization consistency score with a set lower limit, and triggers a multi-source data synchronization consistency decrease alarm when the sliding window mean of the multi-source data synchronization consistency score is lower than the set lower limit; when any determination channel triggers an alarm, the alarm type, alarm time, corresponding navigation working mode label, and observation source availability status are encapsulated into alarm information and reported to the upper layer system, and simultaneously fed back to the navigation working mode determination link, triggering a query of the preset strategy library to re-determine the navigation working mode, thereby forming a multi-layer redundancy guarantee mechanism between the multimodal spatiotemporal cross-sensing navigation network output layer, system integrity monitoring layer, and navigation working mode scheduling layer.

8. A micro-PNT multi-source data fusion navigation system based on a spatiotemporal neural network, characterized in that, include: Sensor spatiotemporal reference module: Under satellite navigation rejection conditions, a global time axis of the system is established with a high-stability clock, and the microelectromechanical system inertial measurement unit, visual sensor, magnetometer and barometer are synchronously triggered to complete the acquisition of raw observation data. A unified timestamp is applied to the acceleration, angular velocity, image frame, geomagnetic field strength, geomagnetic direction and barometric altitude data output by various heterogeneous sensors to form a multi-source raw observation sequence with time correspondence. The environmental scene recognition module extracts image texture density, number and distribution stability of visual feature points, geomagnetic fluctuation amplitude, geomagnetic anomaly degree, barometric altitude change trend, and inertial motion excitation features from the multi-source original observation sequence as scene feature quantities. It calls the scene prior template corresponding to the current coarse location index or initial task binding information, compares the scene feature quantities with the scene prior template item by item, determines the current environmental scene type, and outputs observation quality scores for visual observation availability, geomagnetic observation stability, inertial data continuity, barometric altitude reliability, and multi-source data synchronization consistency. The mode management and scheduling module, based on the environmental scene type and observation quality score, and combined with the current task stage's requirements for navigation accuracy, status update frequency, real-time performance, and endurance, queries a preset strategy library to determine the current navigation working mode. The navigation working modes include visual enhancement mode, geomagnetic-assisted mode, inertial-dominated mode, multimodal joint mode, and low-power hold-up mode. It also generates input configuration parameters for the multimodal spatiotemporal cross-sensing navigation network, including the multi-source data time window length, inertial sequence sampling density, visual feature point filtering threshold, geomagnetic observation input frequency, barometric altitude input switch, and model inference frequency, while simultaneously issuing corresponding sampling rate control commands. Geographic feature constraint module: Based on the initial task binding information or the current location estimate output from the previous calculation cycle, retrieve and load the local map tiles of the corresponding task area, extract the geographic environmental feature information of terrain elevation, road boundaries, building outlines, traversable areas, geomagnetic fingerprints, and indoor spatial structures, and convert the geographic environmental feature information into structured prior inputs of elevation prior features, road passage topology features, building boundary features, geomagnetic fingerprint features, regional traversability features, and spatial coding features; Multimodal spatiotemporal cross-sensing navigation solution module: The multi-source original observation sequence, input configuration parameters, and structured prior input are fed into the multimodal spatiotemporal cross-sensing navigation network. The multimodal spatiotemporal cross-sensing navigation network is composed of a multimodal original feature encoding layer, a heterogeneous feature cross-sensing layer, a long-range temporal state recursion layer, and a navigation state regression output layer cascaded together. It directly completes the regression mapping from multi-source observation data to navigation state quantities, outputs the device's position, velocity, attitude, and time-related state information, and simultaneously outputs model confidence and abnormal observation response indicators. Interface status output module: The position, velocity, attitude, time-related status information, model confidence, current navigation working mode label, and observation source availability status are encapsulated and formatted according to the standard protocol defined by the upper-layer application platform. The standardized spatiotemporal information stream is output in real time through a serial bus or network interface. The module also monitors in real time the working conditions of decreased model confidence, increased abnormal observation response, continuous output state abrupt change, and decreased synchronization consistency of multi-source data. Alarm information is generated and reported to the upper-layer system to adjust the task path or replan the navigation strategy. To realize the micro PNT multi-source data fusion navigation method based on spatiotemporal neural network as described in any one of claims 1-7.

Images