Navigation positioning method and system for double-layer adaptive verification

Abstract

The invention relates to the technical field of positioning and navigation, in particular to a navigation and positioning method and system for double-layer self-adaptive verification. The method comprises the following steps: collecting multi-source sensor data and constructing time sequence features, and outputting a positioning result and confidence by using a first neural network; when the confidence coefficient is insufficient, activating the second neural network to deduce the position increment; dynamic boundary hard constraint verification and INS trend consistency verification are executed in sequence; and dynamically adjusting a verification threshold according to the confidence coefficient, carrying out space-time smooth correction on the prediction trajectory by utilizing a verification residual error, and carrying out directional correction on the internal memory state of the network by utilizing physical constraint loss in a system idle time slice. The invention provides a navigation positioning method and system based on double-layer adaptive verification, and aims to solve the problems of positioning precision divergence and lack of physical consistency of neural network prediction in a satellite signal rejection environment.

Description

Technical Field

[0001] This invention relates to the field of positioning and navigation technology, and in particular to a navigation and positioning method and system with dual-layer adaptive verification. Background Technology

[0002] With the widespread application of intelligent mobile vehicles in complex urban canyons, tunnels, underground parking garages, and other environments where GNSS is denied or weakly signaled, the demand for high-precision and robust navigation and positioning capabilities is becoming increasingly urgent. Integrated navigation technology based on deep fusion of IMU and GNSS, due to its complementary characteristics, has become the core path to achieving continuous and reliable positioning. However, existing fusion schemes generally face key bottlenecks when encountering long-term signal loss or sudden changes in dynamic scenes, such as statically fixed prediction models, single-layer passive verification mechanisms, and lack of feedback loops, making it difficult to effectively suppress accumulated errors and maintain long-term trajectory consistency.

[0003] A search revealed that CN113885064B, titled "A Dual-System Single-Frequency BeiDou Inertial Navigation and Positioning Method, Device, and Storage Medium," proposes using inertial navigation data to correct the horizontal and vertical deviations of satellite positioning results, thereby improving positioning continuity. However, this method only employs a one-way substitution correction logic, and its verification process lacks the ability to reverse-influence the internal state of the prediction model. It cannot achieve online evolution of model parameters or memory states, resulting in significant drift due to model mismatch even under continuous GNSS-free conditions. Another search revealed that CN114488233B, titled "Global Satellite Navigation Terminal and Its Navigation and Positioning Method," while introducing robust adaptive filtering and zero-baseline double-difference observations to optimize observation noise modeling, limits its verification mechanism to statistical analysis of observation residuals. It does not consider whether the carrier motion conforms to physical dynamic boundaries, nor does it establish a multi-level verification information fusion mechanism. Furthermore, it does not use the physical residuals generated by verification to drive the active correction of the prediction model, thus making it difficult to cope with the long-term robustness challenges under unknown dynamic disturbances.

[0004] The aforementioned problems indicate that current IMU / GNSS fusion positioning systems generally suffer from drawbacks such as a single verification dimension, open-loop feedback mechanisms, and weak adaptive capabilities. They lack a closed-loop adaptive architecture that can seamlessly integrate "physical constraint perception—multi-level verification triggering—trajectory smoothing correction—model memory evolution." Therefore, a novel navigation and positioning method is urgently needed. When GNSS is unavailable or unreliable, this method should not only promptly identify and correct trajectory anomalies through a multi-level verification mechanism, but also transform residual information generated by violations of physical laws into learning signals, driving the online evolution of the prediction model. This would enable high-precision, highly robust, and long-term autonomous positioning in complex and unknown environments. Summary of the Invention

[0005] The purpose of this invention is to solve or at least alleviate the significant deficiencies in the adaptive capability of the prediction model and the closed-loop feedback of the verification mechanism in the existing technical solutions, which makes it difficult to effectively suppress positioning drift and maintain long-term stability. This invention provides a navigation and positioning method and system with dual-layer adaptive verification.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a navigation and positioning method with dual-layer adaptive verification, comprising the following steps: Step A: Acquire inertial measurement unit (IMU) data and GNSS (GNSS) satellite navigation data, and perform spatiotemporal synchronization preprocessing; Step B: Input the preprocessed data into the first neural network and output the fusion localization result at the current moment and the confidence scalar representing the reliability of the result in parallel; Step C: When the confidence scalar is lower than the preset safety threshold or GNSS data is missing, the second neural network is triggered to infer the position increment based on IMU data and generate a predicted trajectory; Step D: Construct the inertial navigation INS physical projection trajectory as a benchmark, and calculate the instantaneous consistency residual between the predicted trajectory and the INS physical projection trajectory; Step E: Dynamically adjust the verification threshold according to the confidence scalar. If the instantaneous consistency residual is less than the verification threshold, output the predicted trajectory; if it is greater than the verification threshold, use the residual to correct the predicted trajectory and then output it.

[0007] To further realize the present invention, the following technical solutions may be preferred: Preferably, in step B, the first neural network includes a shared feature extraction layer, a localization regression head, and a confidence evaluation head; the confidence evaluation head reuses the intermediate hidden states of the shared feature extraction layer and calculates the output value between 0 and 1 as a confidence scalar through mapping. This value reflects the degree of matching between the current input features and the distribution of training data and the stability of the internal state of the network.

[0008] Preferably, in step C, triggering the second neural network specifically includes: The input tensor is constructed using only the force and angular velocity sequences from the IMU; The second neural network uses long short-term memory units to process the input tensor and outputs the relative position increment of the current time relative to the previous time. The relative position increment is accumulated to the absolute position at the previous moment to generate the predicted trajectory.

[0009] Preferably, in step D, calculating the instantaneous consistency residual specifically includes: The inertial navigation algorithm is used to perform double integration on the IMU data to calculate the physical displacement vector within a short time window; Calculate the cumulative displacement vector of the predicted trajectory output by the second neural network within the same time window; Calculate the difference in magnitude and the angle between the directions of the two displacement vectors mentioned above, and combine them to obtain the instantaneous consistency residual.

[0010] Preferably, the method further includes a first-level dynamic boundary verification step, located between step C and step D: Calculate the instantaneous acceleration and turning rate values ​​implied in the position increment output by the second neural network; The implicit value is compared with the vehicle's preset physical limit parameters; If the physical limit parameters are exceeded, it is determined to be a physical violation, and the position increment is forcibly truncated to generate a primary correction increment, which is then used to replace the original position increment in step D.

[0011] Preferably, in step E, the logic for dynamically adjusting the verification threshold is as follows: establish a positive correlation between the confidence scalar and the verification threshold; when the confidence scalar decreases, reduce the verification threshold, thereby tightening the tolerance for instantaneous consistency residuals and enhancing the weight of dependence on physical laws.

[0012] Preferably, in step E, the correction of the predicted trajectory using the residual specifically includes spatiotemporal smoothing correction: Establish a spatiotemporal sliding window to cache the consistent residual sequence of historical moments; By combining the current real-time consistent residuals with the historical residual sequences, the final compensation vector is calculated using a weighted moving average logic. The final compensation vector is superimposed on the predicted trajectory to eliminate trajectory abrupt changes.

[0013] Preferably, the method further includes an online feedback learning step: The instantaneous consistency residual calculated in step D is transformed into a physical constraint loss that characterizes the degree of violation of physical laws. The processor load status of the monitoring system is used to generate a suppression signal when an idle time slice of the system is detected; The suppression signal is applied inversely to the memory unit of the second neural network to directionally correct the internal state vector of the memory unit, thereby suppressing the divergence trend of the predicted trajectory at the next moment.

[0014] A navigation and positioning system with dual-layer adaptive verification, comprising: The data acquisition and preprocessing module is used to acquire sensor data and construct feature vectors. The dual-channel neural network module includes a first neural network and a second neural network, which are used to output the fused localization result and confidence scalar, as well as the inferred predicted trajectory, respectively. The inertial navigation consistency verification module is used to solve the INS physical projection trajectory and calculate the instantaneous consistency residual between it and the predicted trajectory. The dynamic verification and output control module is used to dynamically adjust the threshold based on the confidence scalar and determine the output strategy based on the residual comparison results.

[0015] Preferably, it further includes: An online feedback learning controller is used to convert instantaneous consistency residuals into physical constraint losses and use these losses to directionally correct the internal state of the second neural network during system idle time slices.

[0016] The beneficial effects of this invention are: This invention employs a two-layer verification system comprised of hard constraints on dynamic boundaries and consistency verification of inertial navigation trends, achieving comprehensive verification from instantaneous physical feasibility to the rationality of short-term motion trends. The first layer of verification directly intercepts abnormal predictions that violate the basic laws of vehicle motion, while the second layer, through dynamic comparison with a purely physical integral benchmark, identifies trend drifts that, while not exceeding the limits, have deviated from the reasonable trajectory. These two layers of verification complement each other, jointly constructing a highly reliable defense system and effectively solving the problems of single verification dimensions and passive responses in existing technologies.

[0017] Meanwhile, this invention establishes a "verification-learning" closed loop. The system not only uses the verification results to correct the current output trajectory, but more importantly, it transforms the physical residuals generated by the verification into learning signals, which are then used to directionally correct the memory state of the prediction model when the system is idle. This mechanism enables the prediction and extrapolation channel to self-adjust online based on violations of physical laws exposed during operation, gradually optimizing its internal state representation. Thus, even in scenarios with long-term absence of global navigation satellite system signals, it continuously suppresses accumulated errors and maintains high-precision positioning capabilities, fundamentally overcoming the inherent defects of traditional static models. Attached Figure Description

[0018] Figure 1 This is a structural framework diagram of the system of the present invention.

[0019] Figure 2 This is the main flowchart of the method of the present invention.

[0020] Figure 3 This is a comparison diagram of the planar trajectory tracking effect in a long-distance tunnel scenario according to the present invention.

[0021] Figure 4 This is a graph showing the evolution of positioning error over time during GNSS signal interruption according to the present invention.

[0022] Figure 5 This is an adaptive linkage diagram between the confidence scalar and the dynamic verification threshold of the present invention.

[0023] Figure 6 This is a microscopic diagram illustrating the targeted correction of the internal state of a neural network by the online feedback learning controller of this invention. Detailed Implementation

[0024] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0026] This embodiment discloses a navigation and positioning system with dual-layer adaptive verification, such as... Figure 1 As shown, at the physical architecture level, this implementation uses a heterogeneous embedded computing platform with industrial-grade reliability as the core carrier. This platform is not a single central processing unit (CPU) architecture, but rather adopts a composite topology of "main controller + coprocessor + accelerator". The core computing unit uses an automotive-grade system-on-a-chip (SoC), such as the NVIDIA Jetson AGX series or a high-performance industrial processor with equivalent computing power. The SoC integrates a multi-core ARM Cortex-A72 application processor for running complex logic judgments, state machine management, and operating system kernel scheduling; it also integrates an independently running real-time coprocessor (such as an ARM Cortex-R5 core) specifically responsible for high-deterministic sensor data acquisition and low-level driver control, ensuring microsecond-level jitter control in data reading. More importantly, the SoC embeds a high-performance dedicated neural network acceleration unit (NPU) or tensor processing unit (TPU), providing hardware computing power support for real-time inference of dual-channel deep neural networks, so that complex matrix operations no longer consume the valuable resources of a general-purpose CPU.

[0027] To achieve accurate perception of the physical world, the system front end is equipped with a multi-source heterogeneous sensor array. The inertial measurement unit (IMU) uses an industrial-grade six-axis MEMS sensor, such as the ADI ADIS16470 or equivalent, which integrates a three-axis micro-machined accelerometer and a three-axis micro-machined gyroscope. The sensor's sampling frequency is set to 200 Hz, a frequency that can fully capture the instantaneous bumps and steering details of the vehicle without generating excessive data throughput pressure. The IMU is directly connected to the real-time coprocessor via a high-speed SPI bus, using DMA (Direct Memory Access) technology to directly move raw data to the shared memory area, minimizing CPU interrupt load. The Global Navigation Satellite System (GNSS) receiver uses a positioning module that supports multiple constellations and multiple frequencies, capable of simultaneously receiving signals from GPS, BeiDou, Galileo, and other satellites. This module's output frequency is set to 10 Hz, transmitting NMEA format positioning data streams and underlying signal quality indicators, including carrier-to-noise ratio and multipath error envelope, via a UART or USB interface.

[0028] In the process of multi-source sensor data fusion, the accuracy of time synchronization directly determines the upper limit of the fusion system's performance. This implementation abandons the traditional soft synchronization scheme and instead adopts a hardware-level hard synchronization mechanism. The pulses per second (PPS) signal output by the GNSS receiver is physically connected to the external interrupt pin of the SoC. Whenever the rising edge of the PPS pulse arrives, the hardware interrupt service routine is immediately triggered, latching the current high-precision real-time clock (RTC) count value and resetting the microsecond-level counter inside the system. Although the IMU's data acquisition is triggered by an independent timer, its timestamps are uniformly mapped to this global time base.

[0029] The core of the system includes a sensor data preprocessing module, a dual-channel neural network inference engine, a physical consistency verification module, and an online feedback learning controller.

[0030] Sensor data preprocessing module: This module receives high-frequency data (sampling rate ≥100Hz) from the MEMS inertial measurement unit (IMU) and low-frequency data from the Global Navigation Satellite System (GNSS). It performs spatiotemporal alignment, converting the IMU's specific force and angular velocity sequences into sliding time window tensors.

[0031] Dual-channel neural network inference engine: The first neural network (main network) adopts a lightweight CNN architecture to output fused positioning results when GNSS is available, and outputs a confidence scalar based on "information entropy" in parallel.

[0032] The second neural network (auxiliary network) employs a gated recurrent unit (GRU) architecture. This network is activated only when the confidence level is below a threshold or GNSS is lost, and is specifically designed to infer the relative position increment of the carrier based on pure IMU data.

[0033] Physical consistency verification module: It has a built-in strapdown inertial navigation (SINS) solution algorithm to calculate the physical trajectory in a short period of time and use it as a "physical reference" to verify the relative position increment output by the auxiliary network. Example 2

[0034] like Figure 2 As shown, this embodiment discloses a navigation and positioning method based on the two-layer adaptive verification of the system in Embodiment 1.

[0035] Upon system power-on, the initialization process executes silently in the background. This includes not only configuring sensor registers but also loading neural network model weights and allocating video memory. The memory management unit allocates a contiguous physical memory region in high-speed DRAM to construct a spatiotemporal sliding window with a fixed time span. This window uses a circular buffer structure and is managed through the efficient movement of double-ended pointers to achieve first-in, first-out (FIFO) data processing. The window internally stores all the carrier's state information from the past ten seconds, including 3D position coordinates, 3D velocity vectors, 3D attitude angles (represented using quaternions to avoid gimbal lock), and the corresponding system confidence scalar for each frame. This historical background information provides the necessary temporal dimension support for subsequent trajectory smoothing and online learning.

[0036] The data preprocessing module runs continuously, cleaning and standardizing the received raw sensor data. For low-frequency signals like GNSS, the system doesn't directly use the most recent frame during preprocessing. Instead, it employs high-order Lagrange interpolation, using the IMU's high-frequency sampling time as a reference, to align and resample the GNSS positioning results on the time axis. For IMU data, the system loads its zero-bias compensation model and scaling factor correction matrix, deducting sensor static errors in real time. The system also deducts the inherent temperature drift effect of different batches of MEMS devices based on readings from the built-in temperature sensor, compensating with polynomial equations. After error correction, the force data is rotated and projected from the body frame to the local horizontal coordinate system (ENU frame) using the current attitude matrix, and the local gravitational acceleration component is deducted to extract the motion acceleration representing the vehicle's true motion. Finally, the preprocessed inertial features and satellite observation features are stitched together and normalized to form a standard multidimensional temporal feature vector.

[0037] The deep fusion and initial confidence assessment stage is the first hurdle in the system's perception. Feature vectors are directly fed into the first neural network residing in the NPU, i.e., the main fusion channel. This network employs a hybrid architecture of "Convolutional Neural Network (CNN) extracting spatial features + Long Short-Term Memory Network (LSTM) extracting temporal features." The CNN layer captures local abrupt changes in sensor data through sliding convolutional kernels; the LSTM layer, through its gating mechanism, memorizes and models the vehicle's motion trend over a period of time. The network's output is designed with a multi-task parallel structure. One branch is the localization regression head, which directly outputs the fused 3D position and velocity at the current moment; the other branch is the confidence assessment head, a major innovation of this implementation. The confidence assessment head does not directly output the position, but instead analyzes the distribution characteristics of the hidden state vectors in the penultimate layer of the network. By calculating the entropy or discreteness of the hidden state vector, the network can perceive the uncertainty of its current prediction: when the input features are within the network training manifold (such as an open road), the hidden state distribution converges. The greater the discreteness of the hidden state distribution, the greater the uncertainty of the hidden state representation, and the closer the output scalar confidence score is to one. When the vehicle enters a tunnel or experiences severe skidding, and the input features exceed the training manifold, the hidden state distribution diverges, and the confidence scalar score drops rapidly.

[0038] The system decision-making logic makes a dynamic branch judgment based on the scalar confidence level of the above output. When all quality indicators of the GNSS signal are normal and the confidence scalar is higher than a safety threshold (e.g., 0.8), it is considered that the current environment is a credible full-source observation environment, and the output result of the main fusion channel is directly adopted. When the vehicle enters the satellite signal blind zone, or when there is a signal but the multipath effect is strong and the confidence level drops sharply, the pseudo signal generation module is activated and the system seamlessly switches to the prediction and inference mode.

[0039] In predictive simulation mode, the second neural network (predictive simulation channel) is activated. This second neural network is designed specifically for pure inertial simulation, with all received features originating solely from the IMU. To ensure simulation accuracy in the absence of an absolute position reference, the network employs a gated recurrent unit (GRU) architecture. The GRU unit accurately retains the vehicle's long-term motion inertia in memory through update and reset gates, while also rapidly responding to the driver's instantaneous actions. The network output is the position increment relative to the previous moment, i.e., the pseudo-GNSS observation value.

[0040] The biggest risk of a pure data model is generating unrealistic trajectories that violate physical laws. To address this, a hard constraint check of the dynamic boundary is introduced as the first line of defense. This check is not statistical but based on the laws of rigid body dynamics. A pre-loaded table of vehicle physical limit parameters is stored in the simulation memory, including the vehicle's maximum longitudinal acceleration, maximum lateral acceleration, and maximum steering rate on surfaces with different adhesion coefficients. In each prediction cycle, the differential algorithm infers the implicit instantaneous acceleration and steering rate from the predicted position increment. If the inferred lateral acceleration exceeds the boundary (e.g., the car moves half a lane laterally after 100 milliseconds, with an implicit acceleration reaching ten times the force of gravity), the prediction is deemed invalid, and the amplitude is truncated. The vector projection algorithm maintains the velocity direction and forcibly compresses the acceleration amplitude to the boundary, generating a primary correction increment. The hard constraint ensures that regardless of how the neural network diverges, the output is constrained within the reachable domain of the vehicle's physical capabilities.

[0041] In addition, the inertial navigation consistency verification module runs independently on the real-time CPU, forming a consistency verification mechanism for the system. Its operation follows a classic strapdown inertial navigation algorithm (SINS), without using any neural networks, and integrates the IMU data twice. Although SINS also has a cumulative error that increases quadratically over time, it exhibits high relative accuracy and physical determinism regarding motion trends within a short time window (e.g., 1 second). The system calculates the physically extrapolated linear displacement vector within the short time window and compares it with the accumulated result from the initial correction output of the neural network. The comparison indicators include magnitude (i.e., distance) and direction (i.e., angle). The deviation vectors of the two are weighted and combined to form the instantaneous consistency residual.

[0042] In this stage, dynamic threshold adjustment plays a crucial role. The system aims to ensure that higher confidence levels in the main fusion channel indicate a relatively simple environment or lower sensor noise, requiring more stringent trend verification and a smaller dynamic threshold. Conversely, lower confidence levels suggest a more challenging environment or higher vehicle maneuverability, necessitating a larger threshold and more relaxed trend verification. The system establishes a linear or non-linear mapping between confidence levels and thresholds, enabling self-scaling of the verification scale. If the instantaneous consistency residual exceeds the dynamic threshold, the system considers the prediction and extrapolation channel to have experienced trend drift, requiring intervention.

[0043] Intervention measures include spatiotemporal smoothing correction and online feedback learning. Spatiotemporal smoothing correction can be seen as a real-time adjustment of the output. The system reviews the historical residual vectors and the current residual vectors within the spatiotemporal smoothing window, and obtains a compensation vector using a weighted moving average. This compensation vector is used to correct the predicted output trajectory. This compensation vector acts like a low-pass filter, removing high-frequency abrupt noise, making the vehicle positioning trajectory smooth from both a visual and control perspective.

[0044] The purpose of online feedback learning is to allow the system to evolve automatically, breaking the constraints of time slices. However, to avoid impacting the computational resources of the main navigation task, this invention designs an ingenious resource-aware scheduling module. The online anti-learning control module is a low-priority background daemon thread that continuously polls the kernel state of the operating system and reads the CPU load registers. The learning task is only executed after the main positioning loop ends and during time slices when the main CPU utilization is below a certain threshold (e.g., 30%). During this period, this invention transforms the calculated instantaneous consistency residual into a physical constraint loss function. This loss function describes which physical laws the neural network prediction results violate. The controller calculates the gradient of this loss relative to the internal memory units of the second neural network and employs an efficient directional state correction strategy. The controller only uses the calculated gradient to make minor numerical adjustments to the hidden state vector of the GRU unit at the current time step. This is equivalent to directly modifying the network's short-term memory, erasing components that lead to incorrect predictions (such as spurious lateral velocity memories). Through this state flushing, the network's deduction at the next time step will be based on a corrected initial state that conforms to physical laws. Example 3

[0045] To verify the effectiveness and advancement of the system and method described in this embodiment in practical engineering applications, a large-scale real-vehicle test was organized in a typical urban traffic environment. The test scenario was selected as an urban canyon area containing long underground tunnels, road sections shaded by overpasses, and dense high-rise buildings, with a total length of about 20 kilometers, covering various typical operating conditions from good to weak to completely interrupted GNSS signals.

[0046] The test vehicle was a modified autonomous driving prototype equipped with the aforementioned high-precision IMU and GNSS receiver, as well as a post-processing differential positioning system (SPAN-CPT) for acquiring ground truth. During the test, the vehicle maintained a speed of approximately 60 km / h and performed driving maneuvers such as lane changes, overtaking, and ramp turns as required.

[0047] The vehicle conducted a GNSS signal denial test for approximately 120 seconds in an underground tunnel about two kilometers long, and three sets of solutions were compared: Comparative Example A: A traditional pure inertial navigation scheme was adopted.

[0048] Comparative Example B: A pure end-to-end deep learning prediction scheme (LSTM model) without physical constraints was adopted.

[0049] This implementation method employs a two-layer adaptive verification and online feedback learning scheme.

[0050] like Figure 3The image shows a comparison of planar trajectory tracking performance in a long-distance tunnel scenario. The black solid line in the image represents the ground truth trajectory.

[0051] Red dashed line (Comparative Example A - Pure Inertial Navigation): Significant deviation from the lane after about 500 meters in the tunnel, and a smooth but huge parabolic deviation that crosses the tunnel wall in the later part of the tunnel, proving that low-cost IMU integration is unreliable under unconstrained conditions.

[0052] Blue dotted line (comparative example B - pure LSTM prediction): Follows well on most road sections. At the marked emergency lane change point in the figure, the trajectory shows a clear non-physical broken line, and a large lateral deviation appears in subsequent road sections, proving that unconstrained deep learning models are prone to imagining risks.

[0053] Solid green line (in this implementation): Basic tracking of the black truth line. At emergency lane change points, the trajectory is smooth and does not exceed the lane boundaries. This is because the hard constraints of the dynamic boundary intercept non-physical broken lines, while the INS consistency check corrects long-term trend deviations, thus keeping the vehicle within the lane-level positioning accuracy range (lateral error less than 1.5 meters).

[0054] like Figure 4 As shown, the evolution of positioning error over time during GNSS signal interruption is further quantified. The horizontal axis represents time (seconds), and the vertical axis represents positioning error (RMSE, meters).

[0055] Curve A (Comparative Example A): exhibits a typical parabolic growth trend, with the error soaring to 45 meters at the end of 120 seconds.

[0056] Curve B (Comparative Example B): It exhibits approximately linear growth but is accompanied by random jumps. The error is about 18 meters at 120 seconds, and there is a significant error step at 80 seconds (the moment of lane change).

[0057] Curve C (in this implementation): exhibits a linear increase with an extremely low slope, and periodic, minute "sawtooth" rebounds. At the end of 120 seconds, the error is only 3.2 meters.

[0058] Figure 4 The data demonstrates the ability to suppress "online feedback learning." The subtle, jagged callbacks characteristic of curve C directly reflect the "state flushing using idle time slices" implemented in this method. Each time the system detects CPU idleness and performs a directional state correction, the error accumulation trend is reset. This effectively breaks the limitation of inertial navigation error diverging over time, reducing the divergence rate by more than 90%.

[0059] like Figure 5 As shown, the adaptive linkage between the confidence scalar and the dynamic verification threshold is illustrated.

[0060] Solid line (confidence level): It drops from 0.95 to about 0.4 the instant it enters the tunnel, and fluctuates slightly when the road surface is bumpy inside the tunnel.

[0061] Dashed line (dynamic threshold): A strict negative correlation mirror image of the solid line. When confidence drops, the threshold relaxes from a stringent 0.2 meters to 0.8 meters; when stable vehicle movement is detected (slight increase in confidence), the threshold tightens again.

[0062] Figure 5 The operating principle of the dynamic threshold determination mechanism is explained. This adaptive strategy resolves the contradiction between missed detections and false alarms caused by fixed thresholds. When the signal is poor or the movement is violent and the module is not confident, the tolerance is relaxed to avoid trajectory interruption due to verification; while in stable conditions, the tolerance is tightened to ensure accuracy.

[0063] like Figure 6 As shown, the microscopic effect of the online feedback learning controller on the targeted correction of the internal state of the neural network is illustrated. It is a time series plot of the state values ​​of a key hidden layer neuron in the predictive inference channel neural network (second network).

[0064] Gray background line (no feedback): The state value gradually drifts over time and eventually saturates, causing the network to lose its sensitivity to new inputs.

[0065] Highlighted colored lines (with feedback - this implementation): During the drift process, a series of high vertical density correction cuts appear, each correction cut corresponding to a correction by the online feedback learning controller.

[0066] Figure 5 This reveals the essence of physically constrained memory state correction from a microscopic perspective. This is not an abstract concept, but a concrete process of eliminating long-term memory noise in LSTM / GRU units. By subconsciously cleaning the neural network at the physical level, it prevents erroneous movement trends from accumulating in the network memory, fundamentally ensuring the physical reliability of long-term inferences.

[0067] The final test results show that the error was only 3.2 meters at the moment the vehicle regained GNSS signal after exiting the tunnel, and all lateral errors were controlled within the lane lines (within 1.5 meters). Within the two-kilometer signal-free blind spot, the vehicle knew which lane it was in. The trajectory smoothness index shows high smoothness and a continuously smooth output curve, fully meeting the first and second-order continuity requirements of the downstream path planning and control module. Further quantitative analysis of the smoothness index curve reveals that, through the introduction of online feedback learning, this method reduces the error cutoff rate in long-term evolution by more than 60% compared to classical methods. Regarding hardware resource usage, by balancing task priority with idle time slice scheduling, the average CPU utilization of the navigation software stack on the Tegra SoC is only 42%, with a peak of no more than 65%, reserving computational power margin for other perception and planning calculations in autonomous driving.

[0068] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A navigation and positioning method with dual-layer adaptive verification, characterized in that, Includes the following steps: Step A: Acquire inertial measurement unit (IMU) data and GNSS (GNSS) satellite navigation data, and perform spatiotemporal synchronization preprocessing; Step B: Input the preprocessed data into the first neural network and output the fusion localization result at the current moment and the confidence scalar representing the reliability of the result in parallel; Step C: When the confidence scalar is lower than a preset safety threshold or GNSS data is missing, the second neural network is activated; the second neural network generates a predicted trajectory by inferring the position increment based solely on IMU data. Step D: Perform a two-layer adaptive verification on the predicted trajectory: First-level dynamic boundary verification: Calculate the instantaneous acceleration and steering rate values ​​implied by the position increment. If they exceed the vehicle's preset physical limit parameters, perform amplitude truncation processing to generate a primary correction increment. Second-layer INS trend consistency verification: Using the inertial navigation algorithm to construct a short-time physical estimation trajectory as a benchmark, calculate the instantaneous consistency residual between the predicted trajectory formed by the accumulation of the primary correction increment and the physical estimation trajectory; Step E: Dynamically adjust the verification threshold according to the confidence scalar. If the instantaneous consistency residual is less than the verification threshold, output the predicted trajectory; if it is greater than the verification threshold, use the residual to perform spatiotemporal smoothing correction on the predicted trajectory and then output it.

2. The method according to claim 1, characterized in that, In step B, the first neural network includes a shared feature extraction layer, a localization regression head, and a confidence evaluation head. The confidence evaluation head reuses the intermediate hidden states of the shared feature extraction layer and calculates a value between 0 and 1 as a confidence scalar by mapping. This value reflects the degree of matching between the current input features and the training data distribution.

3. The method according to claim 1, characterized in that, In step C, activating the second neural network specifically includes: The input tensor is constructed using only the force and angular velocity sequences from the IMU; The second neural network processes the input tensor using a recurrent layer containing gated recurrent units or a long short-term memory network, and outputs the relative position increment of the current time relative to the previous time. The relative position increment is accumulated to the absolute position at the previous moment to generate the predicted trajectory.

4. The method according to claim 1, characterized in that, In step D, calculating the instantaneous consistency residual specifically includes: The physical displacement vector within a set short time window is calculated by double integration of IMU data using a strapdown inertial navigation algorithm. Calculate the cumulative displacement vector of the predicted trajectory output by the second neural network within the same time window; Calculate the difference in magnitude and the angle between the directions of the two displacement vectors mentioned above, and combine them to obtain the instantaneous consistency residual.

5. The method according to claim 1, characterized in that, In step E, the logic for dynamically adjusting the verification threshold is as follows: Establish a positive correlation between the confidence scalar and the verification threshold; when the confidence scalar decreases, reduce the verification threshold, thereby tightening the tolerance for instantaneous consistency residuals and increasing the weight of dependence on the INS physical inference law.

6. The method according to claim 1, characterized in that, In step E, the correction of the predicted trajectory using the residual specifically includes: Establish a spatiotemporal sliding window to cache the consistent residual sequence of historical moments; By combining the current real-time consistent residuals with the historical residual sequences, the final compensation vector is calculated using a weighted moving average logic. The final compensation vector is superimposed on the predicted trajectory to eliminate trajectory abrupt changes.

7. The method according to claim 1, characterized in that, It also includes step F: performing online feedback learning: converting the instantaneous consistency residual into a physical constraint loss that characterizes the degree of violation of physical laws, and when an idle time slice of the system is detected, using the physical constraint loss to generate an inhibition signal, which is then applied in reverse to the memory unit of the second neural network to perform directional correction on the internal state vector of the memory unit.

8. The method according to claim 7, characterized in that, In step F, the directional correction of the internal state vector of the memory cell specifically refers to: Calculate the gradient of the physical constraint loss with respect to the hidden state vector of the second neural network at the current time step; By utilizing the gradient, without updating the weights of the entire network, only the values ​​of the current hidden state are fine-tuned to suppress the divergence trend of the predicted trajectory at the next moment.

9. A navigation and positioning system with dual-layer adaptive verification, used to implement the method described in any one of claims 1-8, characterized in that, include: The data acquisition and preprocessing module is used to acquire sensor data and construct feature vectors. The dual-channel neural network module includes a first neural network and a second neural network, which are used to output the fused localization result and the confidence scalar, and to infer and predict the trajectory when the confidence is insufficient. The two-layer consistency verification module is used to perform dynamic boundary verification, solve the INS physical projection trajectory, and calculate the instantaneous consistency residual. The dynamic verification and output control module is used to dynamically adjust the threshold based on the confidence scalar and determine the output strategy based on the residual comparison results.

10. The system according to claim 9, characterized in that, Also includes: An online feedback learning controller is used to convert instantaneous consistency residuals into physical constraint losses and use these losses to directionally correct the internal memory state of the second neural network when the system processor is detected to be in an idle time slice.

Images