Adaptive filtering method based on sliding window innovation evaluation and joint constraint

Abstract

The application relates to an adaptive filtering method based on sliding window innovation evaluation and joint constraint. The method comprises the following steps: calculating the innovation covariance through parallel calculation theory and the empirical innovation covariance estimated based on the sliding window, dynamically quantifying the mismatch degree of the two, and adaptively amplifying the GNSS measurement noise covariance according to the mismatch degree, so as to suppress the weight of abnormal signals in fusion. The second module is a state prediction covariance joint constraint unit, which prevents excessive convergence of filtering by introducing a dynamic forgetting factor and applying a minimum boundary constraint related to the position state to the prediction covariance matrix, thereby ensuring that the LiDAR maintains effective state correction capability within the continuous GNSS update interval. Through the synergistic effect of the above mechanisms, the method realizes early perception and active inhibition of advanced spoofing attacks, and significantly improves the positioning safety and robustness of the multi-sensor fusion system in a complex dynamic environment.

Description

Technical Field

[0001] This application relates to the field of high-precision navigation and information security technology for autonomous vehicles, and in particular to an adaptive filtering method based on sliding window information evaluation and joint constraints. Background Technology

[0002] Loosely coupled multi-sensor fusion systems based on GNSS, INS, and LiDAR have become a core technology supporting high-level autonomous driving (L3 and above) due to their ability to provide continuous, reliable, and high-precision positioning capabilities. However, GNSS signals, due to their openness and inherent low-power characteristics, are highly vulnerable to malicious spoofing attacks. Attackers can induce target receivers to generate incorrect positioning and timing information by transmitting forged satellite signals that are highly similar to real signals.

[0003] Currently, the industry generally uses integrity monitoring methods based on the chi-square test for innovation as the fundamental anomaly detection approach. The core principle of this method is to test whether the statistical characteristics of the filter innovation sequence deviate from the pre-set Gaussian white noise assumption. However, with the evolution of attack techniques, more covert deception models have emerged, such as "progressive soft attacks." These attacks do not inject large deviations instantaneously, but rather gradually induce a shift in a smooth and slow manner through carefully designed attack parameters, causing the energy of the malicious signal to be submerged in the system's inherent measurement or process noise. In this situation, the traditional chi-square test fails to detect instantaneous and significant innovation abrupt changes, rendering its detection mechanism ineffective and allowing the attack to remain hidden and ultimately successfully guide the vehicle off its intended path.

[0004] Therefore, the core problem facing existing technologies is that when faced with such elaborate GNSS spoofing attacks coupled with the dynamic characteristics of the system, traditional defense methods based on single-point innovation detection have the defects of perception lag and insufficient suppression capabilities, making it difficult to ensure the navigation safety of autonomous vehicles in complex and dynamic environments. Summary of the Invention

[0005] Therefore, it is necessary to provide an adaptive filtering method based on sliding window information evaluation and joint constraints that can achieve early detection and effective suppression of deception attacks, addressing the aforementioned technical problems.

[0006] An adaptive filtering method based on sliding window innovation evaluation and joint constraints, the method comprising: At each GNSS measurement update time, the data processing flow is started to obtain the current state prediction covariance matrix, measurement matrix, GNSS basic measurement noise covariance matrix output by the Kalman filter, and the GNSS historical information sequence stored within a sliding window of a preset time length. Based on the current state predicted covariance matrix, measurement matrix, and GNSS basic measurement noise covariance matrix, calculate the current theoretical innovation covariance matrix; simultaneously, estimate the empirical innovation covariance matrix based on the GNSS historical innovation sequence. The mismatch metric between the theoretical and empirical information covariance matrices is quantified to generate a real-time risk assessment index for GNSS signal anomalies. A dynamic adaptive threshold is set, and the obtained risk assessment index is compared with the dynamic adaptive threshold. If the risk assessment index exceeds the dynamic adaptive threshold, it is determined that there is a risk of GNSS signal anomaly. The noise adjustment coefficient is calculated, and the measurement noise covariance matrix corresponding to GNSS is adaptively increased based on the noise adjustment coefficient to obtain the adjusted GNSS measurement noise covariance matrix. If the risk assessment index does not exceed the dynamic adaptive threshold, the basic GNSS measurement noise covariance matrix is ​​kept unchanged and used as the adjusted GNSS measurement noise covariance matrix. During the Kalman filter time update stage, the real-time estimates of the uncertainties of GNSS and LiDAR at each location dimension at the current moment are obtained. Based on the real-time uncertainty estimates, the dynamic minimum constraint factor is calculated, and a diagonal minimum constraint factor matrix is ​​constructed. A forgetting factor greater than 1 is determined, and combined with the obtained minimum constraint factor matrix, the previous time step state estimation covariance matrix and the process noise covariance matrix, a joint constraint is applied to the state prediction covariance matrix to obtain the state prediction covariance matrix after applying the joint constraint. Based on the adjusted GNSS measurement noise covariance matrix and the state prediction covariance matrix after applying joint constraints, a Kalman filter measurement update step is performed to fuse GNSS, INS, and LiDAR data and output the positioning result.

[0007] The aforementioned adaptive filtering method based on sliding window innovation evaluation and joint constraints dynamically quantifies the mismatch between theoretical innovation covariance and empirical innovation covariance estimated by sliding window through parallel computation. Based on this, it adaptively amplifies the GNSS measurement noise covariance to suppress the weight of anomalous signals in the fusion process. Furthermore, it can capture the cumulative changes in innovation statistical characteristics caused by progressive spoofing attacks, enabling early identification of attacks and allowing sufficient time for defensive responses. Then, by introducing a dynamic forgetting factor and imposing minimum boundary constraints related to the position state on the predicted covariance matrix, it jointly prevents excessive convergence of the filter, thereby ensuring that LiDAR maintains effective state correction capabilities within continuous GNSS update intervals. Finally, through the synergistic effect of the above mechanisms, early detection and active suppression of advanced spoofing attacks are achieved, significantly improving the positioning security and robustness of the multi-sensor fusion system in complex dynamic environments without affecting normal navigation accuracy. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating an adaptive filtering method based on sliding window information evaluation and joint constraints in one embodiment. Figure 2 This is a flowchart of online evaluation of innovation mismatch and measurement noise adjustment in one embodiment; Figure 3 This is a flowchart of the joint constraints of state prediction covariance in one embodiment. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0010] In one embodiment, such as Figure 1 As shown, an adaptive filtering method based on sliding window innovation evaluation and joint constraints is provided, including the following steps: Step 102: At each GNSS measurement update time, start the data processing flow to obtain the current state prediction covariance matrix, measurement matrix, GNSS basic measurement noise covariance matrix output by the Kalman filter, and the GNSS historical information sequence stored within a sliding window of a preset time length.

[0011] GNSS measurement updates are typically performed at a fixed frequency, such as 1 Hz, meaning an update occurs once per second. The current state prediction covariance matrix is ​​the uncertainty description matrix for system state prediction in Kalman filtering. The measurement matrix establishes the mapping relationship between the system state and the measured values. The GNSS fundamental measurement noise covariance matrix is ​​the statistical characteristic matrix of the measurement noise of the GNSS sensor itself. A sliding window refers to a preset time window of fixed length used to store the historical GNSS information sequence within that time period. The information sequence is the sequence of differences between measured and predicted values, stored in an information buffer. The length of the sliding window... L It can be set according to the actual application scenario.

[0012] Step 104: Calculate the current theoretical innovation covariance matrix based on the current state prediction covariance matrix, measurement matrix, and GNSS basic measurement noise covariance matrix; simultaneously, estimate the empirical innovation covariance matrix based on the GNSS historical innovation sequence.

[0013] The theoretical innovation covariance matrix is ​​calculated based on the system model parameters obtained through Kalman filtering, reflecting the expected uncertainty level of innovations under the condition that the system model is accurate and anomalies-free. The empirical innovation covariance matrix is ​​estimated based on the historical GNSS innovation sequences observed within a sliding window, reflecting the statistical characteristics of the innovation sequences in real time during actual system operation. By computing these two covariance matrices in parallel, a comparative basis can be provided for subsequent mismatch quantification, thereby enabling the perception of GNSS signal anomalies. Specifically, the calculation of the theoretical innovation covariance matrix is ​​based on the basic formula of Kalman filtering, while the estimation of the empirical innovation covariance matrix uses an unbiased estimation method, calculated using historical innovation data within a sliding window.

[0014] Step 106: Perform mismatch quantification on the theoretical information covariance matrix and the empirical information covariance matrix to generate a real-time risk assessment index for GNSS signal anomalies.

[0015] Mismatch quantification refers to measuring the degree of difference between the theoretical innovation covariance matrix and the empirical innovation covariance matrix through specific mathematical operations. When a GNSS signal is subjected to a progressive spoofing attack, the malicious attack causes the statistical distribution of the actual innovation sequence to gradually change, resulting in the empirical innovation covariance matrix systematically deviating from the theoretical innovation covariance matrix. The degree of difference between the two reflects the anomaly risk of the GNSS signal. In this step, the trace of the difference between the two covariance matrices is used to quantify the mismatch. The trace operation can comprehensively reflect the overall difference in variance of all diagonal elements of the matrix, i.e., each component, and is computationally efficient. Using this mismatch as a real-time risk assessment indicator for GNSS signal anomalies can intuitively reflect the severity of the signal anomalies.

[0016] Step 108: Set a dynamic adaptive threshold, compare the obtained risk assessment index with the dynamic adaptive threshold. If the risk assessment index exceeds the dynamic adaptive threshold, it is determined that there is a risk of GNSS signal anomaly. Calculate the noise adjustment coefficient, and adaptively increase the GNSS measurement noise covariance matrix based on the noise adjustment coefficient to obtain the adjusted GNSS measurement noise covariance matrix. If the risk assessment index does not exceed the dynamic adaptive threshold, keep the basic GNSS measurement noise covariance matrix unchanged and use it as the adjusted GNSS measurement noise covariance matrix.

[0017] The dynamic adaptive threshold is a benchmark value used to determine whether risk assessment indicators are abnormal. It is not a fixed value but can be dynamically adjusted according to the system's operating status to adapt to changes in system noise under different operating conditions. The initial threshold can be determined based on the allowable false alarm rate using the chi-square distribution quantiles. During the filtering process, iterative smoothing updates are performed based on the statistical stationarity of recent innovation sequences to reduce misjudgments. When the risk assessment indicator exceeds this dynamic adaptive threshold, it indicates that the difference between the theoretical and empirical innovation covariance matrices exceeds the normal range, indicating a risk of GNSS signal anomalies. At this point, a noise adjustment coefficient is calculated. This coefficient is used to adaptively adjust the GNSS measurement noise covariance matrix. Increasing this matrix aims to proactively reduce the weight of abnormal GNSS measurements in the fusion process, thereby suppressing the influence of deceptive signals at the source of the error propagation path. When the risk assessment indicator does not exceed the threshold, it indicates that the GNSS signal is normal, and there is no need to adjust the measurement noise covariance matrix; its base value remains unchanged.

[0018] Step 110: In the Kalman filter time update stage, obtain the real-time uncertainty estimates of GNSS and LiDAR in each location dimension at the current time, calculate the dynamic minimum constraint factor based on the real-time uncertainty estimates, and construct the minimum constraint factor matrix in diagonal form.

[0019] The time update phase of Kalman filtering refers to the process of predicting the system state at the current moment based on the state estimate and system model from the previous moment. Real-time uncertainty estimates for GNSS and LiDAR are real-time assessments of the positioning accuracy of the two sensors at each position dimension at the current moment. These estimates can be obtained through the quality indicators output by the sensors themselves or from external environmental perception modules. The dynamic minimum constraint factor is designed for the position state parameters in the state vector—eastward and northward positions. It is not a fixed value but is dynamically adjusted according to the ratio of the real-time uncertainty estimates of GNSS and LiDAR. The design logic is as follows: when the environment is conducive to deception (i.e., GNSS appears reliable while LiDAR reliability decreases), the factor approaches 1, and the constraint effect weakens; when LiDAR reliability relatively increases or GNSS becomes uncertain, the factor increases, thereby forcibly increasing the lower bound of the position state variance. The minimum constraint factor matrix is ​​a diagonal matrix, with the dynamic minimum constraint factor placed only on the diagonal elements corresponding to the eastward and northward positions. The diagonal elements corresponding to the other state parameters are 1. This construction method can accurately constrain the predicted covariance of the position state without affecting the calculation of other state parameters.

[0020] Step 112: Determine a forgetting factor greater than 1, and combine the obtained minimum constraint factor matrix, the previous time step state estimation covariance matrix, and the process noise covariance matrix to apply joint constraints to the state prediction covariance matrix, thereby obtaining the state prediction covariance matrix after applying joint constraints.

[0021] The role of the forgetting factor is to proactively inject a small amount of additional uncertainty into the state estimation during the recursive calculation of the state prediction covariance. This prevents the filter from over-relying on historical best estimates and losing sensitivity to current measurements. It can be set to a fixed value slightly greater than 1, or designed as a simple function related to the vehicle's motion state. For example, the forgetting factor can be slightly increased when a large lateral acceleration or angular velocity is detected to quickly respond to model uncertainties caused by maneuvers. The previous-time state estimation covariance matrix is ​​the uncertainty description matrix of the system state estimation at the previous time step, and the process noise covariance matrix is ​​the statistical characteristic matrix of the system process noise. Joint constraints include dynamic minimum boundary constraints and forgetting factor constraints. By combining the minimum constraint factor matrix, the forgetting factor, the previous-time state estimation covariance matrix, and the process noise covariance matrix, the state prediction covariance matrix is ​​adjusted. This prevents excessive filter convergence, ensures that LiDAR maintains effective state correction capability within continuous GNSS update intervals, and avoids ignoring the LiDAR's correction function due to filter overconfidence.

[0022] Step 114: Based on the obtained adjusted GNSS measurement noise covariance matrix and the obtained state prediction covariance matrix after applying joint constraints, perform the measurement update step of Kalman filtering, fuse GNSS, INS and LiDAR data, and output the positioning result.

[0023] The measurement update step of Kalman filtering involves using the adjusted measurement noise covariance matrix and the state prediction covariance matrix after applying joint constraints, combined with measurement data from GNSS, INS (Inertial Navigation System), and LiDAR (Light Detection and Ranging) to optimally estimate the system state. INS offers high positioning accuracy over short periods and is unaffected by external interference, while LiDAR provides absolute observation information. Through multi-sensor data fusion, the advantages of each sensor can be combined, suppressing the influence of GNSS spoofing signals while ensuring the continuity and high accuracy of the positioning results. The output positioning results provide reliable location information for the navigation decisions of autonomous vehicles, ensuring vehicle safety in complex dynamic environments.

[0024] The improvement of the adaptive filtering method based on sliding window innovation evaluation and joint constraints lies in the introduction of two core mechanisms that work together: Mechanism 1: Online Evaluation of Innovation Mismatch and Adaptive Adjustment of Measurement Noise Based on Sliding Window: This mechanism aims to establish an anomaly detection capability based on statistical trends that transcends single-point detection. Its working principle involves maintaining two statistical paths in parallel: one based on the current internal model parameters of the filter (predicted covariance). and measurement noise The theoretical innovation covariance calculated It represents the expected level of uncertainty of the system when the model is accurate and anomalies-free; the second is the empirical innovation covariance calculated based on the innovation sequence actually observed within a fixed-length sliding window. It reflects the actual operational statistical characteristics of the system in real time. When a GNSS signal is subjected to a progressive spoofing attack, the introduced bias gradually alters the statistical distribution of the actual innovation sequence, leading to changes in the empirical covariance. It began to systematically deviate from the theoretical covariance. This invention quantifies this mismatch (e.g., by calculating the difference between their traces). This allows it to detect abnormal trends when deception bias is still in its early, small accumulation stage. Once the mismatch exceeds a dynamic threshold set based on the system's steady-state characteristics, a defensive response is triggered: the GNSS measurement noise covariance matrix is ​​adaptively amplified proportionally. Within the framework of Kalman filtering, this is increased... The physical meaning is to proactively inform the fusion algorithm that "the credibility of the current GNSS measurement has decreased", so that in subsequent state updates, the fusion weight of the suspicious GNSS measurement value will be automatically reduced, and more reliance will be placed on dead reckoning from INS and absolute observation information from LiDAR, thereby suppressing the deception signal from the source of the error propagation path.

[0025] Mechanism 2: Joint constraint mechanism for state prediction covariance: This mechanism aims to address the problem of "loss of sensor calibration capability" caused by excessive convergence of the filter. After a long period of stable operation, the Kalman filter's estimation of the system state tends to be very accurate, as evidenced by the state prediction covariance matrix. The variance corresponding to the midpoint state becomes extremely small. However, excessively small prediction uncertainty means that the filter is overly confident in its own state prediction. At this point, if LiDAR observes a measurement that is slightly deviated from the current prediction due to deception, the filter will almost ignore its correction effect because the "news" of the observation is very small (relative to the extremely small prediction uncertainty), meaning that LiDAR loses its correction capability. This is a key factor in the eventual success of many incremental attacks.

[0026] Therefore, this invention designs two methods to... Perform joint constraints: 1. Dynamic Minimum Boundary Constraint: Define a dynamic minimum constraint factor for the position and state parameters. This factor is not a fixed value, but is dynamically adjusted based on the ratio of the real-time uncertainty estimates of GNSS and LiDAR. Its design logic is as follows: when the environment favors deception (i.e., GNSS appears reliable while LiDAR reliability decreases), the factor approaches 1, weakening its constraint effect; when LiDAR reliability relatively improves or GNSS becomes uncertain, the factor increases, thereby forcibly raising the lower bound of the position state variance and maintaining the filter's "acceptance window" for LiDAR measurements.

[0027] 2. Introduction of the forgetting factor: In the recursive calculation of the state prediction covariance, the forgetting factor is introduced for the covariance matrix of the previous time step. Multiply by a forgetting factor slightly greater than 1 This is equivalent to proactively injecting a small amount of additional uncertainty into the state estimate at each time update, preventing the filter from becoming overly reliant on historical best estimates and ensuring that it maintains the necessary sensitivity to the latest information at the current moment.

[0028] In one embodiment, the current theoretical innovation covariance matrix is ​​calculated based on the current state prediction covariance matrix, the measurement matrix, and the GNSS basic measurement noise covariance matrix, including: Based on the current state-predicted covariance matrix, measurement matrix, and GNSS basic measurement noise covariance matrix, the current theoretical innovation covariance matrix is ​​calculated as follows: ; in, Represents the measurement matrix. This represents the current state prediction covariance matrix. Represents the GNSS basic measurement noise covariance matrix, with superscript... T This indicates the transpose operation.

[0029] Specifically, this calculation method can accurately obtain the expected information covariance characteristics based on the system model parameters, providing a reliable theoretical benchmark for comparison with the empirical information covariance matrix, ensuring the accuracy of subsequent mismatch measurement, and thus achieving accurate perception of GNSS signal anomalies.

[0030] In one embodiment, estimating the empirical innovation covariance matrix based on GNSS historical innovation sequences includes: The empirical innovation covariance matrix is ​​estimated based on the GNSS historical innovation sequence: ; in, L This indicates that the most recent time is read from the information buffer. k Indicates the time sequence number. Indicates the first i GNSS information sequence at time 1, superscript TThis indicates the transpose operation.

[0031] Specifically, this sliding window-based estimation method can track changes in the statistical properties of the innovation sequence in real time. When a GNSS signal is subjected to a progressive spoofing attack, the statistical properties of the innovation sequence will gradually deviate from the normal state, and the empirical innovation covariance matrix will also change accordingly. This allows for the capture of abnormal trends by comparing the empirical innovation covariance matrix with the theoretical innovation covariance matrix. Compared to single-point innovation detection, the sliding window approach can accumulate statistical information about the innovation, avoiding misjudgments caused by accidental fluctuations in single-point innovations. It can also detect subtle, cumulative abnormal changes earlier, improving the foresight and accuracy of anomaly detection.

[0032] In one embodiment, a mismatch metric is performed between the theoretical innovation covariance matrix and the empirical innovation covariance matrix to generate a real-time risk assessment index for GNSS signal anomalies, including: The mismatch metric between the theoretical and empirical innovation covariance matrices is quantified to generate a real-time risk assessment index for GNSS signal anomalies: ; in, Represents the empirical information covariance matrix. This represents the covariance matrix of the current theoretical information.

[0033] Specifically, the generated mismatch index serves as a real-time risk assessment indicator for GNSS signal anomalies, and its value directly reflects the degree of GNSS signal anomaly. When When the difference is small, it indicates that the difference between the theoretical and empirical information covariance matrices is small, and the GNSS signal is normal; when When the value increases and exceeds the dynamic adaptive threshold, it indicates that the GNSS signal may be subject to a spoofing attack. This indicator provides a clear quantitative basis for subsequent anomaly judgment and noise adjustment, ensuring the objectivity and accuracy of anomaly detection.

[0034] In one embodiment, the dynamic adaptive threshold is set as follows: ; in, Forgetting factor, For safety reasons, k Indicates the time sequence number. This represents the covariance matrix of the current theoretical information.

[0035] Specifically, a dynamically updated decision threshold is set. Its initial value Based on the allowable false alarm rate (e.g.) The threshold is determined using the chi-square distribution quantiles. During operation, the threshold is adaptively and smoothly updated. ,in The forgetting factor (e.g., 0.95). The safety factor is set to 1.5. The dynamic adaptive threshold can slowly track changes in the steady-state noise level of the system. When the system noise changes slowly, the threshold can be adjusted accordingly to ensure that normal innovation fluctuations and abnormal innovation changes can be accurately distinguished under different noise environments, thereby improving the adaptability and reliability of anomaly detection and reducing false positives and false negatives.

[0036] In one embodiment, the noise adjustment factor is calculated as follows: ; in, To adjust the gain, As a real-time risk assessment indicator, This is a dynamically adaptive threshold.

[0037] Specifically, comparison and .like If a statistical mismatch occurs, then the adjustment factor is calculated. ,in To adjust the gain (e.g., 0.1), and Limit to the preset maximum value (e.g., within 3.0). Then, output the adjusted GNSS measurement noise covariance matrix: .like Then output That is, no adjustment is made.

[0038] In one embodiment, the adjusted GNSS measurement noise covariance matrix is ​​obtained by adaptively increasing the GNSS measurement noise covariance matrix based on the noise adjustment coefficient, including: Based on the adaptive increase of the noise adjustment coefficient, the GNSS measurement noise covariance matrix is ​​obtained as follows: ; in, This is the noise adjustment factor. This is the GNSS basic measurement noise covariance matrix.

[0039] Specifically, in the Kalman filter framework, the size of the measurement noise covariance matrix directly affects the fusion weight of the measurement values ​​in state updates; the larger the measurement noise covariance matrix, the smaller the corresponding measurement weight. When a risk of GNSS signal anomalies is determined, this formula is used to calculate the GNSS basic measurement noise covariance matrix. Multiply by noise adjustment factor The adjusted measurement noise covariance matrix is ​​obtained. This enables adaptive amplification of the GNSS measurement noise covariance matrix. This operation proactively informs the fusion algorithm that the reliability of the current GNSS measurement has decreased, thus automatically reducing the fusion weight of the suspicious GNSS measurement in subsequent state updates. Instead, it relies more on INS dead reckoning and LiDAR absolute observation information, suppressing spoofing signals at the source of the error propagation path and effectively preventing positioning errors caused by spoofing attacks. When the GNSS signal is normal, This does not affect the normal measurement weight allocation, ensuring that the normal navigation performance of the system is not affected.

[0040] In one embodiment, the dynamic minimum constraint factor is calculated based on real-time uncertainty estimation as follows: ; in, This represents the estimated GNSS horizontal positioning uncertainty at the current moment for the sensor uncertainty assessment unit. This represents the estimated LiDAR horizontal positioning uncertainty of the sensor uncertainty assessment unit.

[0041] Specifically, the engineering meaning of this formula is: when GNSS performance is relatively more reliable ( Smaller) and LiDAR reliability decreases ( When the value is large, the ratio is small. When the value approaches 1, the constraint effect is weak; conversely, when the reliability of GNSS decreases or LiDAR becomes more reliable, the ratio increases. Increasing this value forces a stronger lower bound constraint, reserving room for LiDAR correction. Northbound position factor. The calculation is similar.

[0042] In one embodiment, the diagonal constraint matrix is ​​constructed as follows: The diagonal constraint matrix places dynamic constraint factors only on the diagonal elements corresponding to the east and north positions, and the remaining elements are 1.

[0043] In one embodiment, a forgetting factor greater than 1 is determined, and combined with the obtained minimum constraint factor matrix, the previous time-state estimation covariance matrix, and the process noise covariance matrix, a joint constraint is applied to the state prediction covariance matrix to obtain the state prediction covariance matrix after applying the joint constraint, including: Construct a diagonal form of the minimum constraint factor matrix Its construction method is as follows: for the parameter indices in the state vector corresponding to the east and north positions. Its diagonal elements For other state parameters, ;in, For the first The dynamic minimum constraint factor of each position state parameter at time k; During the time update process, the covariance matrix of the state at the previous time step is estimated. Apply the constraint factor matrix and a forgetting factor greater than 1 And combined with the process noise covariance matrix The state prediction covariance matrix after joint constraints is calculated. : , in, Here is the state transition matrix. This represents the sum of the positional and state-related elements in the diagonal of a matrix, and a predefined minimum boundary value matrix. The operation retrieves the maximum value of the corresponding element.

[0044] Specifically, forgetting factor It can be set to a fixed value slightly greater than 1 (e.g., 1.005), or it can be designed as a simple function related to the vehicle's motion state, such as slightly increasing it when a large lateral acceleration or angular velocity is detected. The value of is used to quickly respond to the model uncertainty brought about by maneuvering.

[0045] Assume the intermediate prediction covariance generated by the standard time update step is The constraint operations are executed in the following order: a. Applying forgetting and factor constraints: First calculate This step is achieved through... and In advance Perform "expansion" and "scaling".

[0046] b. Apply minimum boundary constraints: For the eastward and northward positional variance elements on the diagonal of the covariance matrix, compare them with the preset minimum variance boundary. Compare and take the maximum value. The value of needs to be determined based on the positioning accuracy of the LiDAR used under good conditions; for example, it can be set to . The remaining diagonal and off-diagonal elements remain unchanged.

[0047] The final state prediction covariance matrix obtained after joint constraints is... The output is used for subsequent measurement update steps.

[0048] like Figure 2The flowchart shown is a detailed process for online innovation mismatch assessment and measurement noise adjustment. This process corresponds to steps 102 to 108: First, the innovation is calculated. Then, the theoretical innovation covariance and the estimated empirical innovation covariance are calculated in parallel. The mismatch is quantified to obtain a real-time risk assessment index. This index is compared with a dynamic adaptive threshold. If the index is greater than the threshold, potential deception is detected, the scaling factor is dynamically adjusted, and the measurement variance matrix is ​​updated. If the index is not greater than the threshold, the measurement variance matrix remains unchanged. This flowchart clearly demonstrates the logical sequence of innovation mismatch assessment and noise adjustment, ensuring the orderly and accurate operation of this module.

[0049] like Figure 3 The diagram shows a detailed flowchart of the joint constraint for state prediction covariance. This flowchart corresponds to steps 110 to 112: First, the sensor uncertainty (real-time estimates of GNSS and LiDAR uncertainties) is obtained. Based on this uncertainty, the dynamic constraint factor is calculated. Then, the constraint matrix is ​​constructed and the forgetting factor is determined. Finally, the joint constraint calculation is performed to obtain the state prediction covariance matrix after applying the joint constraints. This flowchart clarifies the operational steps of the joint constraint module, ensuring the standardization and effectiveness of the constraint process.

[0050] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0051] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0052] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An adaptive filtering method based on sliding window innovation evaluation and joint constraints, characterized in that, The method includes: At each GNSS measurement update time, the data processing flow is started to obtain the current state prediction covariance matrix, measurement matrix, GNSS basic measurement noise covariance matrix output by the Kalman filter, and the GNSS historical information sequence stored within a sliding window of a preset time length. Based on the current state predicted covariance matrix, measurement matrix, and GNSS basic measurement noise covariance matrix, calculate the current theoretical innovation covariance matrix; simultaneously, estimate the empirical innovation covariance matrix based on the GNSS historical innovation sequence. The mismatch metric is quantified between the theoretical and empirical information covariance matrices to generate a real-time risk assessment index for GNSS signal anomalies. A dynamic adaptive threshold is set, and the obtained risk assessment index is compared with the dynamic adaptive threshold. If the risk assessment index exceeds the dynamic adaptive threshold, it is determined that there is a risk of GNSS signal anomaly. A noise adjustment coefficient is calculated, and the measurement noise covariance matrix corresponding to GNSS is adaptively increased based on the noise adjustment coefficient to obtain the adjusted GNSS measurement noise covariance matrix. If the risk assessment index does not exceed the dynamic adaptive threshold, the basic GNSS measurement noise covariance matrix is ​​kept unchanged and used as the adjusted GNSS measurement noise covariance matrix. During the Kalman filter time update stage, the real-time estimates of the uncertainties of GNSS and LiDAR at each location dimension at the current moment are obtained. Based on the real-time estimates of the uncertainties, the dynamic minimum constraint factor is calculated, and a diagonal minimum constraint factor matrix is ​​constructed. A forgetting factor greater than 1 is determined, and combined with the obtained minimum constraint factor matrix, the previous time step state estimation covariance matrix and the process noise covariance matrix, a joint constraint is applied to the state prediction covariance matrix to obtain the state prediction covariance matrix after applying the joint constraint. Based on the adjusted GNSS measurement noise covariance matrix and the state prediction covariance matrix after applying joint constraints, a Kalman filter measurement update step is performed to fuse GNSS, INS, and LiDAR data and output the positioning result.

2. The method according to claim 1, characterized in that, Based on the current state predicted covariance matrix, measurement matrix, and GNSS basic measurement noise covariance matrix, calculate the current theoretical innovation covariance matrix, including: Based on the current state predicted covariance matrix, measurement matrix, and GNSS basic measurement noise covariance matrix, the current theoretical innovation covariance matrix is ​​calculated as follows: in, Represents the measurement matrix. This represents the current state prediction covariance matrix. Represents the GNSS basic measurement noise covariance matrix, with superscript... T This indicates the transpose operation.

3. The method according to claim 1, characterized in that, The empirical innovation covariance matrix is ​​estimated based on the aforementioned GNSS historical innovation sequence, including: Based on the aforementioned GNSS historical information sequence, the empirical information covariance matrix is ​​estimated as follows: in, L This indicates that the most recent time is read from the information buffer. k Indicates the time sequence number. Indicates the first i GNSS information sequence at time 1, superscript T This indicates the transpose operation.

4. The method according to claim 1, characterized in that, The theoretical and empirical innovation covariance matrices are quantified to measure mismatch, generating real-time risk assessment indicators for GNSS signal anomalies, including: The mismatch metric is performed on the theoretical and empirical information covariance matrices to generate a real-time risk assessment index for GNSS signal anomalies: in, Represents the empirical information covariance matrix. This represents the covariance matrix of the current theoretical information.

5. The method according to claim 1, characterized in that, Set the dynamic adaptive threshold as follows: in, Forgetting factor, For safety reasons, k Indicates the time sequence number. This represents the covariance matrix of the current theoretical information.

6. The method according to claim 1, characterized in that, The noise adjustment factor is calculated as follows: in, To adjust the gain, As a real-time risk assessment indicator, This is a dynamically adaptive threshold.

7. The method according to claim 1, characterized in that, Based on the noise adjustment coefficient, the measurement noise covariance matrix corresponding to GNSS is adaptively increased to obtain the adjusted GNSS measurement noise covariance matrix, including: Based on the aforementioned noise adjustment coefficient, the GNSS measurement noise covariance matrix is ​​adaptively increased, resulting in the adjusted GNSS measurement noise covariance matrix as follows: in, This is the noise adjustment factor. This is the GNSS basic measurement noise covariance matrix.

8. The method according to claim 1, characterized in that, Based on the real-time estimation of the aforementioned uncertainty, the dynamic minimum constraint factor is calculated as follows: in, This represents the estimated GNSS horizontal positioning uncertainty at the current moment for the sensor uncertainty assessment unit. This represents the estimated LiDAR horizontal positioning uncertainty of the sensor uncertainty assessment unit.

9. The method according to claim 8, characterized in that, Construct the diagonal constraint matrix as follows The diagonal constraint matrix places dynamic constraint factors only on the diagonal elements corresponding to the east and north positions, and the remaining elements are 1.

10. The method according to claim 1, characterized in that, A forgetting factor greater than 1 is determined. Combined with the obtained minimum constraint factor matrix, the previous time-state estimation covariance matrix, and the process noise covariance matrix, joint constraints are applied to the state prediction covariance matrix to obtain the jointly constrained state prediction covariance matrix, including: Construct a diagonal form of the minimum constraint factor matrix Its construction method is as follows: for the parameter indices in the state vector corresponding to the east and north positions. Its diagonal elements For other state parameters, ;in, For the first The dynamic minimum constraint factor of each position state parameter at time k; During the time update process, the covariance matrix of the state at the previous time step is estimated. Apply the constraint factor matrix and a forgetting factor greater than 1 And combined with the process noise covariance matrix The state prediction covariance matrix after joint constraints is calculated. : ， in, Here is the state transition matrix. This represents the sum of the positional and state-related elements in the diagonal of a matrix, and a predefined minimum boundary value matrix. The operation retrieves the maximum value of the corresponding element.

Images