Error calibration method and system for intelligent converged terminal

Abstract

The application discloses an error calibration method and system for an intelligent fusion terminal, and relates to the technical field of error calibration. The method comprises the following steps: collecting and synchronizing original observation data of multiple sensors of the intelligent fusion terminal and environment state parameters; generating a fusion reference true value dynamically based on the consistency of the multiple sensor data; intelligently selecting a calibration task from a pre-defined strategy library according to the terminal motion state and the environment parameters; estimating the specific value of the error parameter based on the reference benchmark and the physical constraints according to the selected calibration task, and performing real-time compensation and updating by using an online learning model with the dynamic benchmark as a supervisory signal; performing multi-dimensional confidence evaluation on the estimation result, and arbitrating whether to adopt the parameter to update the system according to the evaluation result. The application realizes error calibration of the intelligent fusion terminal which is independent of external equipment, adaptive to dynamic environment and highly reliable, and significantly improves the perception accuracy and system robustness of the intelligent fusion terminal.

Description

Technical Field

[0001] This application relates to the field of error calibration technology, and in particular to an error calibration method and system for intelligent fusion terminals. Background Technology

[0002] Intelligent fusion terminals perceive their environment and self-state by integrating multiple heterogeneous sensors (such as cameras, lidar, inertial measurement units, global navigation satellite systems, millimeter-wave radar, etc.). Multi-sensor fusion methods aim to provide more robust and accurate pose, velocity, and environmental perception results than single-sensor methods through information complementarity. However, inherent errors in the sensors within the terminal, as well as time-varying errors caused by environmental factors such as temperature, vibration, and electromagnetic interference, can severely degrade the performance of intelligent fusion terminals.

[0003] Traditional terminal sensor error calibration methods are mainly divided into two categories: laboratory calibration and online estimation. Laboratory calibration is typically performed in a controlled environment using high-precision turntables, guide rails, and other equipment to obtain accurate calibration parameters; however, this method cannot adapt to environmental changes after the deployment of intelligent fusion terminals and the aging and drift of sensor parameters over time. Online estimation methods, such as sensor calibration based on Kalman filtering, usually estimate error parameters as part of the state vector, but suffer from slow convergence, specific requirements for motion excitation, and susceptibility to abnormal observations. More importantly, most existing online methods rely on reference sources, making them inoperable in environments where reference sources fail or are unavailable.

[0004] Furthermore, existing calibration schemes are mostly designed for single sensors or fixed sensor pairs, failing to fully utilize the rich multi-sensor redundancy and complementary information inherent in intelligent fusion terminals. When multiple sensors simultaneously observe the same physical quantity, the consistency between their observations makes internal self-calibration possible, but currently there is a lack of systematic methods to utilize this internal consistency to construct dynamic and reliable calibration benchmarks.

[0005] Therefore, there is an urgent need for an error calibration method that can operate continuously online without relying on external high-precision references, make full use of internal sensor redundancy, adapt to dynamic environments, and ensure the perception and state estimation accuracy of intelligent fusion terminals throughout their entire lifecycle and under all operating conditions. Summary of the Invention

[0006] To address the shortcomings of existing technologies in sensor error calibration, such as reliance on external references, poor environmental adaptability, and insufficient utilization of multi-sensor collaboration, this application provides an error calibration method and system for intelligent fusion terminals that does not rely on external devices, dynamically generates benchmarks based on the internal consistency of multiple sensors, and enables intelligent collaboration and online learning.

[0007] This application provides an error calibration method for intelligent fusion terminals, comprising:

[0008] Step S10: Collect and synchronize raw observation data from multiple sensors on the target intelligent fusion terminal, and obtain the current environmental status parameters; Step S20: Based on the observation data of multiple sensors after synchronization, calculate the internal consistency metric between the sensors and identify the subset of sensors with the greatest consistency, and dynamically generate the fusion benchmark truth value and its corresponding uncertainty metric. Step S30: Based on the current real-time motion state and environmental state parameters of the target intelligent fusion terminal, select a calibration task from the predefined calibration strategy library. The calibration task includes the target sensor, the type of error parameter to be calibrated, and the reference benchmark. Step S40: Based on the reference benchmark and physical constraints, and according to the selected calibration task, estimate the specific value of the error parameter to be calibrated, and simultaneously use an online learning model for real-time compensation. The online learning model is based on supervised learning and updating of the fused benchmark true value. Step S50: Perform a multi-dimensional confidence assessment on the specific value of the error parameter to be calibrated estimated in step S40, and arbitrate whether to adopt the specific value of the error parameter to be calibrated based on the assessment results, so as to update the system parameter library.

[0009] Further, in step S20, the dynamic generation of the fusion benchmark truth value specifically includes: Step S21: Calculate the Mahalanobis distance between the observations of any two sensors on the target intelligent fusion terminal in the same observation dimension, and convert the distance values ​​into a consistency metric, constructing a system of size S21. Consistency matrix ,in For the number of sensors; Step S22, convert the consistency matrix Treating it as an adjacency matrix, we construct an undirected graph to abstractly represent sensor networks and their consistency relationships. And apply the maximum clique search algorithm to find undirected graphs. The largest group in The sensors within this largest cluster constitute the largest consistent subset of sensors at the current moment; Step S23, for the largest cluster Each sensor in Assign fusion weights , By sensor reliability factor and based on current environmental characteristics Dynamic factors obtained by querying the environment confidence table Joint decision, that is , This indicates that the variables at both ends of the symbol are proportional, and all fusion weights are normalized; Step S24: Use a robust fusion algorithm to analyze the largest clique. The observations from all sensors are weighted and fused, and the fused state estimate is output as the dynamic baseline truth, along with the corresponding covariance matrix.

[0010] Furthermore, in step S30, the motion state includes a stationary state, a uniform linear motion state, a uniformly accelerated motion state, a pure rotational motion state, and an unconstrained general motion state. The calibration strategy library is a rule mapping table that maps different combinations of motion states and environmental state parameters to specific calibration tasks. The selection principles for the reference datum include: In static and uniform motion scenarios, the physical constraints of the sensor are specified as the reference benchmark; in complex scenarios or scenarios without external reference, the fusion benchmark truth value generated in step S20 is specified as the reference benchmark.

[0011] Furthermore, step S40 includes two parallel and complementary error processing paths: Path 1, Explicit Parameter Estimation: Based on the calibration task decided in step S30, a constrained optimization problem (such as nonlinear least squares) is constructed using the specified reference benchmark and inviolable physical constraints (such as zero angular velocity at rest and constant gravity vector). By calling the optimization solver, the error parameters of the target sensor (such as IMU zero bias vector and camera-LiDAR extrinsic parameters) are explicitly estimated.

[0012] Path Two, Implicit Online Learning: Simultaneously, an online learning model is run to perform real-time forward compensation on the target sensor data. This model takes the original observation data sequence, historical state information, and current environmental state parameters as input and directly outputs the error compensation value. The training and updating of the online learning model are continuously performed online based on the dynamic baseline truth value generated in step S20 as a supervision signal; this allows the model to adaptively learn and compensate for complex, time-varying nonlinear errors (such as temperature drift and vibration-induced noise characteristic changes) that are difficult to describe using parameterized models.

[0013] The two paths are not completely independent; they coordinate through a system parameter library. The real-time compensation output of the online learning model can serve as a preprocessing step for the raw sensor data in Path 1, enabling the parameter estimation algorithm to operate on data that more closely approximates ideal characteristics, thereby improving estimation accuracy and convergence speed. When Path 1 produces a new parameter estimate with high confidence and it is adopted by the system, this parameter can be used to reset or fine-tune the corresponding online learning model. For example, the newly estimated IMU zero bias can be used as the initial value of the bias of the network output layer, allowing the network to learn more refined residual compensation based on it. This prevents the model from continuously learning on an erroneous basis and achieves knowledge transfer.

[0014] Further, in step S50, the multi-dimensional confidence assessment includes: Step S51: Calculate the specific value of the error parameter to be calibrated estimated in step S40 within a length of... Standard deviation within the sliding window Further calculate the time-domain consistency score. : ,in This represents the preset scaling parameter used to adjust the score function with respect to the standard deviation. Sensitivity is determined by experience; Step S52: After calibrating the sensor using the specific values ​​of the error parameters to be calibrated estimated in step S40, calculate the Mahalanobis distance between the newly observed sensor data and the fused reference true value. Further calculate the spatial consistency score. : ,in This represents a preset threshold parameter used to control the impact of Mahalanobis distance on consistency scores. The effect on the rate of decay; Step S53: Check whether each component of the specific value of the error parameter to be calibrated estimated in step S40 is within the preset physically feasible range. ]Inside, and Let these represent the upper and lower limits of the range, respectively. Further calculations are performed to determine the proportion of parameter components within the physically feasible range, which serves as the physical plausibility score. ; Step S54: Extract the final residual of the optimization algorithm used in step S40 to estimate the specific values ​​of the error parameters to be calibrated, and determine whether the residual is less than the threshold. The convergence score is... ,otherwise ; Step S55: Calculate the overall confidence score. : ,in Indicates the evaluation dimension index. , , and These respectively indicate temporal consistency, spatial consistency, physical plausibility, and convergence. This indicates the weight of each evaluation dimension; if Greater than the acceptance threshold Then the specific value of the error parameter to be calibrated estimated in step S40 is adopted.

[0015] This application also provides an error calibration system for intelligent fusion terminals, comprising: Data acquisition and preprocessing module: used to acquire and synchronize raw observation data from multiple sensors on the target intelligent fusion terminal, and obtain current environmental status parameters; Dynamic benchmark generation module: Based on the observation data of multiple synchronized sensors, it dynamically generates the fusion benchmark truth value and its corresponding uncertainty metric by calculating the internal consistency metric between sensors and identifying the subset of sensors with the greatest consistency. Calibration strategy decision module: used to select a calibration task from a predefined calibration strategy library based on the current real-time motion state and environmental state parameters of the target intelligent fusion terminal. The calibration task includes the target sensor, the type of error parameter to be calibrated, and the reference benchmark. Parameter estimation and online compensation module: Based on the reference benchmark and physical constraints, and according to the selected calibration task, it estimates the specific values ​​of the error parameters to be calibrated, and performs real-time compensation using an online learning model. The online learning model is supervised learning and updated based on the fused benchmark true value. Confidence assessment and arbitration module: used to conduct multi-dimensional confidence assessment of the specific values ​​of the error parameters to be calibrated estimated in step S40, and to arbitrate whether to adopt the specific values ​​of the error parameters to be calibrated based on the assessment results, so as to update the system parameter library.

[0016] This application discloses the following technical effects: This application provides an error calibration method and system for intelligent fusion terminals. By utilizing the internal consistency of multiple sensors to generate dynamic fusion reference true values, it eliminates the dependence on external high-precision reference sources, enabling intelligent fusion terminals to maintain high-precision calibration in any environment without external signals, greatly improving the system's autonomy and environmental robustness.

[0017] Secondly, the method proposed in this application, through a context-aware calibration strategy decision module, automatically selects the optimal calibration strategy and sensor pairing based on real-time operating conditions, which significantly improves the efficiency and resource utilization of the calibration process. Furthermore, it adopts a dual-track parallel architecture of explicit parameter estimation and implicit online learning, which can accurately calibrate modeled errors and compensate for non-modeled, time-varying complex errors in real time through data-driven models, enabling the system to have lifelong learning and adaptability.

[0018] In addition, this method introduces a rigorous multi-dimensional confidence assessment and arbitration mechanism, which can effectively identify and shield erroneous results caused by sensor momentary failures, data anomalies, or estimation non-convergence, providing key protection for applications with high safety requirements such as autonomous driving. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating an error calibration method for a smart fusion terminal, provided as an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of an error calibration system for a smart fusion terminal provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] Example 1: This application provides an error calibration method for intelligent fusion terminals, such as... Figure 1 As shown, the method includes: Step S10: Collect and synchronize raw observation data from multiple sensors on the target intelligent fusion terminal, and obtain the current environmental status parameters.

[0023] In this embodiment, multiple sensors on the target intelligent fusion terminal form a sensor array, continuously outputting raw observation data through a dedicated driver interface for the sensor array. These sensors are divided into two categories according to their functions: The first type of sensor is the pose and perception sensor, used to acquire the intelligent fusion terminal's own motion state and environmental geometric information, including: Vision sensors: monocular, binocular, and multi-view cameras, outputting RAW or compressed image data, transmitted via MIPI CSI-2, GMSL2, or USB3 Vision interfaces; LiDAR: Mechanical and solid-state LiDAR, outputting three-dimensional point cloud data and intensity information, transmitted via gigabit vehicle Ethernet; Inertial Measurement Unit (IMU): 6-axis IMU, outputting three-axis acceleration, three-axis angular velocity, and three-axis magnetic field data, which are transmitted at high frequency (200Hz-1kHz) via SPI, I2C, or UART interfaces; Global Navigation Satellite System (GNSS) receiver: Outputs latitude, longitude, altitude, velocity, UTC time, and positioning status word, transmitted via UART or CAN FD interface; Odometer: Obtains speed and incremental displacement information from motor encoders or wheel speed sensors.

[0024] The second category is environmental status sensors, used to quantify the physical environment in which the intelligent fusion terminal is located, including: Temperature sensor: an external probe that monitors the temperature of key components in the environment where the intelligent fusion terminal is located; Triaxial magnetometer: Integrated into the IMU, it measures the ambient magnetic field vector; Ambient light sensor: measures ambient light intensity.

[0025] Vibration sensor: Monitors the vibration spectrum of a structure.

[0026] First, to establish a unified spatiotemporal reference system based on the raw observation data acquired by the aforementioned sensor array, this step adopts a hierarchical synchronization strategy to achieve hardware-level absolute time synchronization and software-level relative timestamp alignment: In an IEEE 802.1AS Ethernet backbone network, a global PTP master clock is deployed; all PTP-enabled sensors access the network as slave clocks and synchronize their local clocks with the master clock by exchanging synchronization messages, achieving sub-microsecond absolute time alignment. For devices that do not support hardware synchronization, when their data packets arrive at the host driver, the operating system immediately assigns a high-precision, monotonically increasing host receive timestamp in kernel mode. The clock source of this timestamp should be synchronized with the PTP master clock. Through calibration experiments, the fixed hardware delay of each sensor (the total time from the occurrence of the physical event to the timestamp of the data packet) is measured and compensated. This delay value will be stored in the system parameter library for correcting the timestamp during the preprocessing stage.

[0027] Secondly, the time-synchronized sensor observation data is preprocessed, including: Based on the pre-calibrated extrinsic parameters (transformation matrix from sensor coordinate system to terminal body coordinate system) stored in the system parameter library, all sensor observation data after time synchronization are transformed into a unified terminal body coordinate system. For scanning sensors (mechanical lidar), the actual acquisition time of different points in the collected point cloud data is different within a frame. By using the synchronized IMU data corresponding to the scanning cycle of the frame, the local pose of each laser point cloud at the start of the scanning of the frame is calculated through inertial navigation interpolation. This unifies the entire frame point cloud into the same coordinate system and eliminates the point cloud distortion caused by the movement of the terminal itself.

[0028] From the data collected by environmental state sensors, a set of standardized environmental state parameters are extracted and calculated to form an environmental state feature vector. : ,in It represents the temperature, and is the direct reading from the temperature sensor; The total magnetic field strength is represented by the square root of the triaxial ambient magnetic field vector. It represents the ambient light intensity and is a direct reading from the ambient light sensor; The root mean square value of vibration energy is extracted by frequency domain analysis of the IMU accelerometer signal (after removing gravity) within a short time window; these environmental state parameters are normalized to the [0,1] interval.

[0029] Step S20: Based on the observation data of multiple synchronized sensors, calculate the internal consistency metric between sensors and identify the subset of sensors with the highest consistency, and dynamically generate the fusion benchmark truth value and its corresponding uncertainty metric.

[0030] In this embodiment, the state variables to be estimated are first extracted from the multi-sensor observation data that has achieved spatiotemporal synchronization and coordinate unification, output in step S10. A dynamic fusion benchmark generation instance is maintained for each state variable, and the input to each instance is... The current observations of this state quantity by each sensor and its corresponding, pre-calibrated observation noise covariance matrix ; For the current moment, calculate all sensor pairs Consistency of observations on the same state variable, where and This invention preferably uses the reciprocal of the standardized Mahalanobis distance as a consistency measure within the sensors, comprehensively considering the differences in observations and the observation uncertainties of each sensor. The calculation process of the consistency measure includes the following detailed steps: Calculate the observation residuals : ,in and They represent the first and the One sensor; Calculate joint uncertainty : ,in and They represent the first and the The observation noise covariance matrix of each sensor; Calculate the first and the Mahalanobis distance between sensors The square of: ,in This represents the matrix transpose operation. Represents the inverse of a matrix; Mapping the square of the Mahalanobis distance to a value between 0 and 1 yields the consistency coefficient. : ,in It is an adjustable scaling parameter, set empirically. After iterating through all sensors, a size of [size missing] is obtained. Consistency matrix The matrix elements are consistency coefficients, and the diagonal elements are all 1.

[0031] Secondly, from all sensors, find a set of sensors with the most consistent internal observations; this is abstracted as a graph theory problem and solved using the maximum clique search algorithm: Taking each sensor as a vertex, if two sensors and Consistency coefficient between Greater than the preset threshold (In this example, the value is 0.7), then an undirected edge is added between them, thereby constructing an undirected graph. Used to abstractly represent sensor networks and their consistency relationships, where and Let represent the set of vertices and the set of edges of the graph, respectively. The maximum clique search algorithm is used to find a complete subgraph (i.e., a clique) in the graph such that any two vertices in the subgraph are connected by an edge (meaning that all sensors in the subset are at the same height pairwise), and this clique has the maximum number of vertices; this is the maximum clique. This is the largest consistent subset of sensors at the current moment.

[0032] Obtain the largest consistent subset of sensors Then, the sensor observations within the subset are weighted and fused: For subsets Each sensor in Its fusion weight The calculation formula is as follows:

[0033] in, Indicates sensor The reliability factor, obtained offline based on its model and historical performance data, is a constant; Indicates sensor The dynamic environmental adaptation factor is a vector of the current environmental state. The function is obtained by querying a pre-stored environment-performance lookup table, which reflects the impact of current environmental conditions such as temperature and vibration on the sensor's expected noise level. The Lubang fusion algorithm is adopted, and this implementation uses the covariance intersection algorithm to target the maximum clique. The observations in the data are fused to obtain the fused baseline true value. and its corresponding uncertainty measure : Fusion benchmark truth and its covariance matrix Solving through iterative optimization yields the following relationship:

[0034] This result ensures that the uncertainty after fusion is an upper bound of the true error covariance under any possible sensor error correlation, thereby enhancing the robustness of the system.

[0035] Step S30: Based on the current real-time motion state and environmental state parameters of the target intelligent fusion terminal, select a calibration task from a predefined calibration strategy library. The calibration task includes the target sensor, the type of error parameter to be calibrated, and a reference benchmark.

[0036] In this embodiment, the motion feature vector of the target intelligent fusion terminal at the current moment is first extracted from the fusion results of sensors such as the inertial measurement unit and the odometer. This vector describes the current real-time operating status of the terminal and includes: Motion intensity: Calculate the magnitude of acceleration and angular velocity; Motion richness: Analyze the variance or spectrum of acceleration and angular velocity within a past time window to determine whether the motion has sufficiently excited each axis; Motion mode: Based on speed and acceleration characteristics, identify the current stationary state, uniform linear motion state, uniformly accelerated motion state, pure rotational motion state, or unconstrained general motion state of the intelligent fusion terminal.

[0037] The target intelligent fusion terminal directly uses the standardized environmental feature vector output in step S10 to determine the current environmental state parameters. ; The predefined calibration policy library is a structured collection. Each policy is an "IF-THEN" rule that defines the conditions and task content that trigger a calibration task, including the following dimensions: Triggering condition: A condition related to motion state Environmental conditions Boolean expressions.

[0038] Target sensor: Specifies one or a group of sensors that need to be calibrated.

[0039] The type of error parameter to be calibrated: clearly define the error term that needs to be estimated, for example: Internal parameters: camera focal length, distortion coefficient; IMU accelerometer, scale factor, non-orthogonal error; External parameters: The installation rotation matrix and translation vector of the sensor relative to the terminal body coordinate system; Time parameter: Fixed time delay between the sensor and other sensors.

[0040] Reference benchmark: Specifies the source of truth that the calibration process relies on.

[0041] Task attributes include the expected time, computational complexity, and priority of the calibration task.

[0042] The selection principles for the reference benchmark include: in static and uniform motion scenarios, the physical constraints of the sensor are specified as the reference benchmark; in scenarios without external reference, the true value of the dynamic fusion benchmark generated in step S20 is specified as the reference benchmark.

[0043] Establish a finite state machine, where each state represents the dominant motion mode of the intelligent fusion terminal, and the state transitions are determined by... Trigger; In each state, activate the corresponding subset of calibration strategies, evaluate the triggering conditions of each strategy in the currently activated subset of strategies, generate the final calibration task based on the principle that "calibration tasks that handle critical faults (such as complete sensor failure) have the highest priority", encapsulate it as calibration quality, and send it to the execution module of step S40.

[0044] Step S40: Based on the reference benchmark and physical constraints, and according to the selected calibration task, estimate the specific value of the error parameter to be calibrated, and simultaneously use an online learning model for real-time compensation. The online learning model performs supervised learning and updates based on the fused benchmark true value.

[0045] In this embodiment, the calibration task instruction output in step S30 is used as input. The target sensor, the set of error parameters to be calibrated, the specified reference benchmark, and the physical constraints in the instruction are analyzed. Based on the above information, this step runs two parallel and complementary error processing paths: Path 1: Explicit Parameter Estimation. This path involves constructing and solving a model-based optimization problem to explicitly estimate the specific values ​​of the error parameters. This includes the following steps: Based on the set of error parameters to be calibrated The type calls a predefined sensor error model. For example, for an IMU, the error model is:

[0046]

[0047] in, and This represents the raw measured angular velocity and linear acceleration actually read from the IMU sensor, including errors. and This represents the true angular velocity and linear acceleration, which is obtained through calibration and compensation from the erroneous measurements. and The truth value recovered from it; and This is the scale factor matrix. and It is a non-orthogonal matrix. and Zero bias, and For noise, the error parameter to be estimated Includes the model , , , , and .

[0048] Within the sliding window Collect the observation sequences of the target sensor. Sequence corresponding to the reference benchmark (If the reference benchmark is a dynamic fusion benchmark, then) The output from step S20; if it is a physical constraint, then To construct a nonlinear least squares problem for the constraint equations (such as zero velocity):

[0049] in, This means that when the following expression reaches its minimum value, The value of ; Represents the observation error function. and Representing time respectively The raw observation data from the target sensor and the corresponding true values ​​obtained from the reference standard. Represents the square of the Mahalanobis distance. It is an observation information matrix used to weight error components with different reliability or different units; Indicates the first A physical constraint error function It is also the square of the Mahalanobis distance. This represents the constraint information matrix, used to indicate the confidence level of our physical constraints.

[0050] The Levenberg-Marquardt optimization algorithm is used to solve this problem, and the solver outputs the optimal parameter estimates. And its estimated uncertainty (obtained through the Hessian matrix in the solution process); this The data will be passed to step S50 for confidence assessment.

[0051] Path 2: Implicit Online Learning. This path runs in parallel with Path 1, operating a neural network model. This enables real-time, feedforward compensation of target sensor errors and the ability to learn complex errors that are difficult to parameterize, including the following steps: An independent neural network is built for each sensor requiring compensation (e.g., a 3-layer fully connected network for IMU temperature drift compensation); during the inference phase, the model... Running in real time via forward propagation: Input the current raw sensor data Environmental characteristics and system status; outputs the real-time compensation amount for the current sensor reading. Output after performing the compensation action : .

[0052] The model is based on the fusion baseline truth value generated in step S20. Conduct supervised learning, when Available and its uncertainty When the value is below the threshold, the current input features are compared with the corresponding supervision signal. Pairing is performed to form training samples, which are then stored in a fixed-size first-in-first-out replay buffer. The loss function is defined as the difference between the model's compensated output and the dynamic baseline. The system periodically samples small batches of training samples randomly from the buffer, performs a step of stochastic gradient descent, and updates the model weights.

[0053] The two paths are not completely independent; they coordinate through a system parameter library. The real-time compensation output of the online learning model can serve as a preprocessing step for the raw sensor data in Path 1, enabling the parameter estimation algorithm to operate on data that more closely approximates ideal characteristics, thereby improving estimation accuracy and convergence speed. When Path 1 produces a new parameter estimate with high confidence and it is adopted by the system, this parameter can be used to reset or fine-tune the corresponding online learning model. For example, the newly estimated IMU zero bias can be used as the initial value of the bias of the network output layer, allowing the network to learn more refined residual compensation based on it. This prevents the model from continuously learning on an erroneous basis and achieves knowledge transfer.

[0054] Step S50: Perform a multi-dimensional confidence assessment on the specific value of the error parameter to be calibrated estimated in step S40, and arbitrate whether to adopt the specific value of the error parameter to be calibrated based on the assessment results, so as to update the system parameter library.

[0055] In this embodiment, the specific value of the error parameter to be calibrated is estimated in step S40. The steps for conducting a multi-dimensional confidence assessment include: Step S51, calculate In length Standard deviation within the sliding window Further calculate the time-domain consistency score. : ,in This represents the preset scaling parameter used to adjust the score function with respect to the standard deviation. Sensitivity is determined by experience; Step S52, using After calibrating the sensor, calculate the Mahalanobis distance between the newly observed sensor data and the dynamic fusion reference ground truth. Further calculate the spatial consistency score. : ,in This represents a preset threshold parameter used to control the impact of Mahalanobis distance on consistency scores. The effect on the rate of decay; Step S53, check Whether each component is within the preset physically feasible range [ ]Inside, and Let these represent the upper and lower limits of the range, respectively. Further calculations are performed to determine the proportion of parameter components within the physically feasible range, which serves as the physical plausibility score. ; Step S54, check the estimation Is the final residual of the optimization algorithm less than the threshold? The convergence score is... ,otherwise ; Step S55: Calculate the overall confidence score. : ,in Indicates the evaluation dimension index. , , and These respectively indicate temporal consistency, spatial consistency, physical plausibility, and convergence. This indicates the weight of each evaluation dimension; if Greater than the acceptance threshold If so, then adopt .

[0056] Example 2: The error calibration system for a smart fusion terminal provided in this embodiment of the invention can execute the error calibration method for a smart fusion terminal provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method execution, such as... Figure 2 As shown, it has the following modules: Data acquisition and preprocessing module: used to acquire and synchronize raw observation data from multiple sensors on the target intelligent fusion terminal, and obtain current environmental status parameters; Dynamic benchmark generation module: Based on the observation data of multiple synchronized sensors, it dynamically generates the fusion benchmark truth value and its corresponding uncertainty metric by calculating the internal consistency metric between sensors and identifying the subset of sensors with the greatest consistency. Calibration strategy decision module: used to select a calibration task from a predefined calibration strategy library based on the current real-time motion state and environmental state parameters of the target intelligent fusion terminal. The calibration task includes the target sensor, the type of error parameter to be calibrated, and the reference benchmark. Parameter estimation and online compensation module: Based on the reference benchmark and physical constraints, and according to the selected calibration task, it estimates the specific values ​​of the error parameters to be calibrated, and performs real-time compensation using an online learning model. The online learning model is supervised learning and updated based on the fused benchmark true value. Confidence assessment and arbitration module: used to conduct multi-dimensional confidence assessment of the specific values ​​of the error parameters to be calibrated estimated in step S40, and to arbitrate whether to adopt the specific values ​​of the error parameters to be calibrated based on the assessment results, so as to update the system parameter library.

[0057] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0058] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than in the embodiments and still achieve the desired results.

Claims

1. An error calibration method for intelligent fusion terminals, characterized in that, The method includes: Step S10: Collect and synchronize raw observation data from multiple sensors on the target intelligent fusion terminal, and obtain the current environmental status parameters; Step S20: Based on the observation data from multiple synchronized sensors, calculate the internal consistency metric among the sensors and identify the subset of sensors with the highest consistency, dynamically generating the fusion benchmark truth value and its corresponding uncertainty metric. Calculate the Mahalanobis distance between the observations of any two sensors on the target intelligent fusion terminal in the same observation dimension, and calculate the consistency metric based on the Mahalanobis distance to construct a sensor consistency matrix; The consistency matrix is ​​regarded as the adjacency matrix. An undirected graph is constructed to abstractly represent the sensor network and its consistency relationship. The maximum clique search algorithm is applied to find the maximum clique in the undirected graph. The set of sensors that constitute the maximum clique is determined as the maximum consistent sensor subset. The fusion weight of each sensor is calculated based on the long-term reliability weight and dynamic confidence weight of each sensor in the maximum consistency sensor subset under the current environmental condition. A robust fusion algorithm is used to weight and fuse the observations of each sensor in the maximum consistency sensor subset to generate the fusion benchmark true value and its uncertainty covariance matrix; Step S30: Based on the current real-time motion state and environmental state parameters of the target intelligent fusion terminal, select a calibration task from the predefined calibration strategy library. The calibration task includes the target sensor, the type of error parameter to be calibrated, and the reference benchmark. Step S40: Based on the reference benchmark and physical constraints, and according to the selected calibration task, estimate the specific value of the error parameter to be calibrated, and simultaneously use an online learning model for real-time compensation. The online learning model is based on supervised learning and updating of the fused benchmark true value. Step S50: Perform a multi-dimensional confidence assessment on the specific value of the error parameter to be calibrated estimated in step S40, and arbitrate whether to adopt the specific value of the error parameter to be calibrated based on the assessment results, so as to update the system parameter library.

2. The error calibration method for a smart fusion terminal as described in claim 1, characterized in that, In step S30, the motion states include a stationary state, a uniform linear motion state, a uniformly accelerated motion state, a pure rotational motion state, and an unconstrained general motion state. The calibration strategy library is a rule mapping table that maps different combinations of motion states and environmental state parameters to specific calibration tasks.

3. The error calibration method for a smart fusion terminal as described in claim 1, characterized in that, In step S30, the selection principles for the reference datum include: In static and uniform motion scenarios, the physical constraints of the sensor are specified as the reference benchmark; in scenarios without external reference, the true value of the fusion benchmark generated in step S20 is specified as the reference benchmark.

4. The error calibration method for a smart fusion terminal as described in claim 1, characterized in that, Step S40 takes the calibration task selected in step S30 as input, identifies the target sensor, the set of error parameters to be calibrated, the specified reference benchmark, and the physical constraints; and runs parallel and complementary error processing paths, including explicit parameter estimation and implicit online learning. The display parameter estimation constructs a constrained optimization problem, which explicitly estimates the specific values ​​of the target sensor's calibration error parameters of the aforementioned type by invoking an optimization algorithm.

5. The error calibration method for a smart fusion terminal as described in claim 4, characterized in that, The implicit online learning utilizes an online learning model for real-time compensation, including the following detailed steps: Based on the type of target sensor and error parameter to be calibrated in the calibration task, a neural network model is configured as an online learning model; the input of the online learning model includes the original observation data sequence of the target sensor, historical state information and current environmental state parameters, and the output is the error parameter compensation value; During the inference phase, the online learning model is run in real time to perform forward compensation on the target sensor observations; During the model update phase, with the fused baseline true value as the supervision objective, the parameters of the online learning model are optimized by online gradient descent by minimizing the residual between the compensated data of the online learning model and the supervision objective.

6. The error calibration method for a smart fusion terminal as described in claim 1, characterized in that, In step S50, the multi-dimensional confidence assessment includes the following detailed steps: Step S51: Calculate the statistical variance of the specific value of the error parameter to be calibrated estimated in step S40 within the sliding time window, and calculate the time-domain consistency score based on the statistical variance. Step S52: Calculate the Mahalanobis distance between the newly observed sensor data and the fusion reference true value after calibrating the sensor using the specific value of the error parameter to be calibrated estimated in step S40, and calculate the spatial consistency score based on the Mahalanobis distance. Step S53: Verify whether each component of the specific value of the error parameter to be calibrated estimated in step S40 is within the range of preset physical feasible parameters, and calculate the proportion of the component within the range of physical feasible parameters as a physical rationality score. Step S54: Extract the iterative residual of the optimization algorithm used in step S40 to estimate the specific value of the error parameter to be calibrated, and determine whether the iterative residual is lower than a predetermined threshold. If yes, the convergence score is 1; otherwise, it is 0. Step S55: Weighted fusion of temporal consistency score, spatial consistency score, physical rationality score, and convergence score to obtain a comprehensive confidence score, and compare the comprehensive confidence score with a preset acceptance threshold to arbitrate whether to adopt the specific value of the error parameter to be calibrated.

7. An error calibration system for intelligent fusion terminals, characterized in that, The system is used to implement the error calibration method for a smart fusion terminal according to any one of claims 1-6, the system comprising: Data acquisition and preprocessing module: used to acquire and synchronize raw observation data from multiple sensors on the target intelligent fusion terminal, and obtain current environmental status parameters; Dynamic benchmark generation module: Based on the observation data of multiple synchronized sensors, it dynamically generates the fusion benchmark truth value and its corresponding uncertainty metric by calculating the internal consistency metric between sensors and identifying the subset of sensors with the greatest consistency. Calibration strategy decision module: used to select a calibration task from a predefined calibration strategy library based on the current real-time motion state and environmental state parameters of the target intelligent fusion terminal. The calibration task includes the target sensor, the type of error parameter to be calibrated, and the reference benchmark. Parameter estimation and online compensation module: Based on the reference benchmark and physical constraints, and according to the selected calibration task, it estimates the specific values ​​of the error parameters to be calibrated, and performs real-time compensation using an online learning model. The online learning model is supervised learning and updated based on the fused benchmark true value. Confidence assessment and arbitration module: used to conduct multi-dimensional confidence assessment of the specific values ​​of the error parameters to be calibrated estimated in step S40, and to arbitrate whether to adopt the specific values ​​of the error parameters to be calibrated based on the assessment results, so as to update the system parameter library.

Images