An artificial intelligence enhanced extended kalman filtering method and system

Abstract

This invention discloses an AI-enhanced extended Kalman filter method and system, belonging to the field of Kalman filter technology. The method constructs an extended state including velocity and measurement bias, and performs prediction and linearization based on a dynamic model. The forward innovation and its variance are calculated using the minimum process noise of the fixed bias channel, and used for feature normalization. A sliding window feature sequence is constructed using the normalized innovation and velocity increment, and input into a one-dimensional convolutional neural network for multiple random deactivation inferences to obtain anomaly intensity scores and cognitive uncertainty. Based on this result, the process noise of the bias channel is adaptively adjusted to enhance bias tracking capability when anomalies are significant, and maintain estimation smoothness when normal conditions are normal. Finally, the extended Kalman filter update is completed, achieving robust joint estimation of velocity and bias. This invention improves the estimation stability and accuracy under strong noise and transonic conditions, and is applicable to velocity measurement of aircraft and ultra-high-speed motors.

Description

Technical Field

[0001] This invention belongs to the field of Kalman filtering technology, specifically relating to an artificial intelligence-enhanced Extended Kalman Filter (EKF) method and system. Background Technology

[0002] Traditional Extended Kalman Filter (EKF) and its improved methods have been widely used in the field of state estimation. These methods are usually based on the fundamental assumptions that the measurement noise is approximately Gaussian white noise and that the system error can be fully covered by the noise model. However, in complex dynamic environments such as high-speed motion and transonic speeds, actual measurement signals often exhibit strong non-Gaussian and non-stationary characteristics, which poses a serious challenge to the performance of traditional filtering methods.

[0003] Specifically, during the testing of aircraft, hypersonic propulsion systems, and high-dynamic ground vehicles, sensor data is susceptible to various complex interferences. On the one hand, due to aerodynamic disturbances, pressure pulsations, or structural resonances, sudden spikes and clusters of noise with short durations and high amplitudes frequently appear in the measurement signals. This type of noise exhibits significant non-Gaussian statistical characteristics and short-term bursts, severely compromising the stability of filters. On the other hand, when the velocity of a moving body approaches or exceeds the speed of sound, drastic changes in the flow field structure induce strong broadband vibrations and noise modulation phenomena, leading to a significant increase in the measured noise power and a continuous evolution of its spectral characteristics with velocity. Traditional fixed-parameter noise models are unable to accurately describe this time-varying behavior.

[0004] Furthermore, in practical operation, sensors are also affected by factors such as temperature drift, environmental changes, or flow field switching, which can produce slowly changing or even step-like systematic biases. These slowly varying measurement biases cannot be adequately characterized by conventional white noise or simple random walk models. Without explicit processing, this will lead to lag or continuous drift in the estimation results. More complexly, these various disturbances often occur coupled together and are strongly correlated with motion states (such as velocity and Mach number), forming a complex measurement environment where "peak clusters-bias-vibration noise" coexist, placing higher demands on the adaptive capability and robustness of the filter.

[0005] Against this backdrop, traditional extended Kalman filtering methods often exhibit significant shortcomings: during spike cluster interference, the estimated values ​​oscillate violently; when the measurement bias changes abruptly, the filter responds slowly due to inertia, leading to tracking lag and steady-state deviation; in the transonic region, unmodeled high-power vibration noise further reduces the smoothness and reliability of the estimation results. Although some existing adaptive filtering techniques (such as covariance matching and adaptive noise estimation) can improve specific problems to some extent, they mostly focus on adjusting for single types of interference and lack a collaborative sensing and processing mechanism for complex anomaly modes. Therefore, it is difficult to simultaneously achieve effective noise suppression, fast bias tracking, and global smooth estimation under extreme conditions where strong non-Gaussian noise, time-varying bias, and transonic vibration coexist.

[0006] Therefore, engineering practice urgently needs a new state estimation framework that can explicitly model measurement biases, intelligently identify composite anomaly patterns, and dynamically adjust filter parameters accordingly, in order to improve robustness, adaptability, and overall estimation accuracy in complex and harsh measurement environments. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides an AI-enhanced extended Kalman filter method and system. By introducing an AI anomaly perception mechanism without directly interfering with the filter state update structure, it can significantly improve the estimation robustness and stability under strong noise interference and transonic conditions, reduce speed estimation errors, and is applicable to application scenarios such as ultra-high-speed linear motors, aircraft speed measurement, and vehicle power systems.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] An AI-enhanced extended Kalman filter method includes:

[0010] Step 1: Based on the extended state vector including velocity and measurement bias and the nonlinear dynamic model, perform prior state prediction, and linearize the state function to obtain the Jacobian matrix;

[0011] Step 2: Fix the bias channel process noise to the minimum value, calculate the forward prediction covariance corresponding to the prior state, and then obtain the measurement innovation and its forward covariance.

[0012] Step 3: Calculate the normalized innovation feature using the measured innovation and its forward covariance, calculate the velocity increment feature by combining the internal velocity estimation of the filter, and combine the normalized innovation feature and velocity increment feature into an AI feature matrix through a sliding time window.

[0013] Step 4: Input the AI ​​feature matrix into a one-dimensional convolutional neural network, perform multiple forward inferences under random deactivation conditions, and output the score representing the intensity of the anomaly and the corresponding model cognitive uncertainty.

[0014] Step 5: Based on the anomaly score and the corresponding model cognitive uncertainty, dynamically calculate and limit the adaptive process noise value of the bias channel at the current moment;

[0015] Step 6: Using the latest noise covariance matrix containing the adaptive process noise values, recalculate the state prediction covariance; then calculate the Kalman gain, update the state estimate and covariance matrix, and finally output a joint estimate of velocity and measurement bias.

[0016] Furthermore, in step 1, linearizing the state function specifically means: for the nonlinear velocity dynamics equation, performing a first-order Taylor expansion at the current velocity state estimate to obtain the state transition Jacobian matrix, where the cross-derivative term between the velocity state and the bias state is zero.

[0017] Furthermore, in step 2, the noise variance corresponding to the measurement innovation is configured as a time variable that varies with the system velocity or Mach number, in order to characterize the phenomenon of increased measurement noise power caused by flow field instability in the transsonic region; the measurement noise includes spike cluster noise with continuous high amplitude burst characteristics.

[0018] Furthermore, in step 3, the length of the sliding time window is between 32 and 128 sampling points, which is used to construct a two-row, multi-column feature matrix, wherein the first row is the normalized innovation sequence arranged in chronological order, and the second row is the velocity increment sequence at the corresponding time.

[0019] Furthermore, in step 4, the number of input data channels of the one-dimensional convolutional neural network is 2, corresponding to the normalized innovation sequence and the speed increment sequence, respectively; its network structure includes at least two one-dimensional convolutional layers, a random deactivation layer and a fully connected layer, wherein the features extracted by the convolutional layers are passed to the fully connected layer after average pooling in the time dimension.

[0020] Furthermore, in step 4, the number of forward inferences is between 4 and 16; the anomaly score is obtained by calculating the average of the output results of multiple inferences; and the model cognitive uncertainty is obtained by calculating the variance of the output results of multiple inferences and then performing linear scaling and numerical truncation.

[0021] Furthermore, in step 5, the specific method for dynamically calculating and limiting the amplitude is as follows: based on the preset minimum value of the bias process noise, an adjustment amount is added, which is composed of the difference between the preset maximum and minimum noise values, the anomaly score, and the product of the difference between the preset maximum and minimum noise values, and the cognitive uncertainty. The calculation result is then limited to the preset minimum and maximum values.

[0022] Furthermore, in step 5, the adaptive adjustment of the bias channel process noise also includes verifying the validity of the adaptive process noise value. If the product of the abnormal score and the cognitive uncertainty is less than a preset threshold, the previously valid bias process noise value remains unchanged.

[0023] Furthermore, the function of the one-dimensional convolutional neural network is limited to outputting the anomaly score and the model's cognitive uncertainty to drive the adaptive adjustment of the bias channel process noise, without directly modifying the measurement data, innovation value, or the state update equation of the extended Kalman filter.

[0024] On the other hand, the present invention provides an artificial intelligence-enhanced extended Kalman filter system, comprising:

[0025] The state prediction module is used to predict prior states based on an extended state vector containing velocity and measurement bias and a nonlinear dynamic model, and to linearize the state function to obtain the Jacobian matrix.

[0026] The auxiliary variable calculation module is used to fix the bias channel process noise to a minimum, calculate the forward prediction covariance corresponding to the prior state, and then obtain the measurement innovation and its forward covariance.

[0027] The feature construction module is used to calculate the normalized innovation feature using the measured innovation and its forward covariance, calculate the velocity increment feature by combining the internal velocity estimation of the filter, and combine the normalized innovation feature and the velocity increment feature into an AI feature matrix through a sliding time window.

[0028] The one-dimensional convolutional neural network detection module is used to input the AI ​​feature matrix into the one-dimensional convolutional neural network, perform multiple forward inferences under random deactivation conditions, and output a score representing the intensity of the anomaly and the corresponding model cognitive uncertainty.

[0029] The bias and process noise adaptive adjustment module is used to dynamically calculate and limit the adaptive process noise value of the bias channel at the current moment based on the anomaly score and the corresponding model cognitive uncertainty.

[0030] An extended Kalman filter update module is used to recalculate the state prediction covariance using the latest noise covariance matrix containing the adaptive process noise values; then, the Kalman gain is calculated, the state estimate and covariance matrix are updated, and finally, a joint estimate of velocity and measurement bias is output.

[0031] Thirdly, the present invention provides an electronic device comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned artificial intelligence-enhanced extended Kalman filter method.

[0032] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned artificial intelligence-enhanced extended Kalman filtering method.

[0033] The beneficial effects of this invention are as follows:

[0034] First, by explicitly introducing the measurement bias state and constructing an extended state vector, the method can directly model and track slowly varying systematic errors caused by sensor drift, environmental changes, etc., fundamentally improving the filter's ability to characterize real complex errors.

[0035] Secondly, this design innovatively utilizes the standardized measurement residuals (the difference between actual measurements and those predicted based on the system model) generated by the extended Kalman filter itself, along with velocity increments, to construct time-series features. A lightweight one-dimensional convolutional neural network combined with Monte Carlo random deactivation technology is then employed for online inference. This design enables the artificial intelligence module to effectively perceive and quantify the intensity of complex anomaly patterns such as "peak cluster noise" and "bias mutations," while simultaneously assessing the confidence level of its own judgments, thus achieving intelligent identification of complex interference.

[0036] Furthermore, by dynamically mapping the anomaly intensity score and cognitive uncertainty output by the artificial intelligence module to the process noise of the bias channel, adaptive adjustment of the filter parameters is achieved. Under strong interference or significant bias changes, the noise variance is automatically increased to improve tracking speed and flexibility; under stable conditions, a small noise variance is maintained to ensure the smoothness and stability of the estimation results. This "perception-adjustment" mechanism significantly enhances the robustness of the standard extended Kalman filter to non-Gaussian, time-varying interference without altering its core update structure.

[0037] Ultimately, the combined application of the above technologies enables the method of this invention to simultaneously achieve excellent noise suppression, fast bias tracking capability, and stable and smooth estimation across the entire speed range in harsh measurement environments with spike cluster noise, transonic vibration, and slowly varying bias. This significantly reduces the overall error and fluctuation of speed estimation, providing a reliable technical solution for high-precision speed measurement of aircraft, hypersonic equipment, and vehicle power systems. Attached Figure Description

[0038] Figure 1 This is a flowchart of an artificial intelligence-enhanced extended Kalman filter method according to the present invention;

[0039] Figure 2 A sequence diagram for constructing forward-looking statistics and artificial intelligence features;

[0040] Figure 3 This is a diagram of a one-dimensional convolutional neural network structure.

[0041] Figure 4 A schematic diagram of the extended Kalman filter update process;

[0042] Figure 5 This is a schematic diagram showing the comparison of velocity estimation and error under strong disturbance conditions according to the present invention. Detailed Implementation

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0044] like Figure 1 As shown, this invention provides an artificial intelligence-enhanced extended Kalman filter method, the method comprising:

[0045] Step S1: State Prediction and Linearization. An extended state vector is constructed, incorporating velocity and measurement bias. The velocity is an internal velocity estimate from the filter, and the measurement bias characterizes systematic errors caused by sensor drift, environmental changes, and abnormal disturbances. Based on the extended state vector and the nonlinear dynamic model, a priori prediction of the current state is performed to obtain the predicted state at the next sampling time. Simultaneously, the nonlinear state transition relationship is linearized at the current estimated state to obtain a linear approximation model of the state transition.

[0046] Step S2: Calculation of Forward Measurement Residuals. Before executing AI inference, the process noise of the channel corresponding to the measurement bias is fixed to a preset minimum value. Based on this, the forward prediction covariance corresponding to the predicted state is calculated. Furthermore, using the forward prediction covariance, the difference between the predicted and actual measured values, i.e., the measurement residual, and the forward covariance corresponding to the measurement residual, are calculated. The forward prediction covariance and measurement residual statistics are only used for subsequent anomaly feature construction and do not participate in the final update process of the extended Kalman filter.

[0047] Step S3: AI Feature Construction. Based on the measurement residuals and their forward covariance, the measurement residuals are normalized to obtain normalized measurement residual features reflecting the relative degree of measurement deviation. Simultaneously, based on the velocity estimation results within the filter, the velocity change between adjacent sampling times is calculated as the velocity increment feature. Using a sliding time window approach, the normalized measurement residual features from multiple consecutive times are combined with the velocity increment feature to construct a feature matrix for artificial intelligence inference.

[0048] Step S4: CNN Anomaly Detection. The feature matrix is ​​input into a one-dimensional convolutional neural network, and multiple forward inferences are performed while keeping the random deactivation mechanism enabled. Based on the results of multiple inferences, an anomaly intensity score is obtained to characterize the degree of anomaly within the current time window, and a cognitive uncertainty is obtained to reflect the confidence level of the neural network's own judgment, thereby achieving intelligent perception of complex noise and anomaly patterns.

[0049] Step S5: Adaptive Adjustment of Bias Channel Noise. Based on the anomaly intensity score and cognitive uncertainty, the process noise of the channel corresponding to the measurement bias is adaptively adjusted, and upper and lower limits are imposed on the adjustment results. When the anomaly intensity is large and the cognitive uncertainty is low, the process noise level of the bias channel is increased to enhance the tracking ability of bias changes; when the anomaly intensity is small or the cognitive uncertainty is high, the process noise of the bias channel is kept small to avoid interference from random noise on the estimation results and maintain the smoothness of the system estimation.

[0050] Step S6: Extended Kalman Filter Update. Based on the latest noise covariance model including the process noise of the adaptive bias channel, the state prediction covariance is recalculated, and the Kalman gain is further calculated to update the prior state, obtaining the velocity estimate and measurement bias estimate at the current sampling time. At the same time, the corresponding covariance matrix is ​​updated, completing one AI-enhanced extended Kalman filter iteration.

[0051] Specifically, in step S1:

[0052] First, define the variables: z represents the true velocity of the object being measured at the k-th sampling time. k This is the velocity measurement value obtained at the k-th sampling time. This represents the posterior velocity estimate within the filter at the k-th sampling time. The predicted velocity value at the (k+1)th sampling time is obtained from the system model. This is the estimated measurement bias value at the k-th sampling time, used to characterize the slowly varying systematic errors caused by sensor drift, environmental changes, and spike cluster noise.

[0053] Then, assuming at discrete sampling time k, an extended state vector containing velocity and measurement bias is constructed:

[0054] ,

[0055] A discrete-time nonlinear dynamic model is established using velocity as the fundamental physical quantity. The evolution of the actual velocity satisfies:

[0056] ,

[0057] in, For discrete sampling periods; a cmd (t k ) is at time t k c1 is the equivalent acceleration; c2 is the linear damping coefficient; c2 is the air drag coefficient related to the square of the velocity; m is the equivalent inertial mass of the system.

[0058] Based on the EKF principle, a measurement bias state is introduced to construct the extended state equation:

[0059] ,

[0060] Where: w v,k with w b,k These are the process noises acting on the velocity channel and the bias channel, respectively.

[0061] The corresponding measurement equation is:

[0062] ,

[0063] Among them, z k This represents the measured velocity value at the k-th sampling time. n represents the actual measurement bias at the k-th sampling time; k For measuring noise.

[0064] The measurement equation can be written in matrix form:

[0065] , .

[0066] Furthermore, based on the aforementioned state equation, prior state prediction and linearization are performed, resulting in:

[0067] , ,

[0068] For the state transition function in Performing first-order linearization at this point yields the state Jacobian matrix:

[0069] .

[0070] In step S2:

[0071] The process noise of the fixed bias channel is minimized. Construct the forward prediction covariance matrix:

[0072] ,

[0073] in The variance of the speed channel process noise can be designed as a function of speed or Mach number, with the superscript T indicating transpose.

[0074] Calculate the measurement residuals using the aforementioned forward prediction covariance. and its variance:

[0075] , ,

[0076] Among them, y k+1 Characterize the difference between model predictions and actual measurements; For y k+1 The variance estimate is used for subsequent normalization calculations; R k+1 To measure the noise variance, it can be designed as a function of velocity or Mach number. The forward prediction covariance and measurement residual statistics will be used for anomaly feature construction.

[0077] In step S3:

[0078] Based on the measurement residual information obtained in step S2, two types of features are constructed for the artificial intelligence input:

[0079] (1) Characteristics of normalized measurement residuals

[0080] ,

[0081] Where ε is a minimal constant to prevent division by zero.

[0082] (2) Speed ​​increment

[0083] ,

[0084] Using a sliding time window of length W, the features are combined into a two-dimensional feature matrix:

[0085] .

[0086] The contents of steps 2 and 3 are summarized as follows: Figure 2In this process, under the condition that the process noise of the bias channel is fixed to the minimum value, the forward prediction covariance is constructed sequentially, the measurement residual and its variance between the predicted measurement value and the actual measurement value are calculated, and the measurement residual is normalized. At the same time, the velocity increment features of adjacent time moments are calculated, and the normalized measurement residual features and velocity increment features are fused by using a sliding time window method, and finally the input feature matrix for the artificial intelligence anomaly perception module is generated.

[0087] In step S4:

[0088] The feature matrix Inputting the data into a one-dimensional convolutional neural network, performing M forward inference iterations while maintaining random deactivation, yields a provisional score:

[0089] i = 1, … , M,

[0090] in, Let i be the score for the i-th time. Indicates input The convolutional neural network process, where M is the number of inferences.

[0091] The convolutional neural network structure used is as follows: Figure 3 As shown. The network is... The feature matrix is ​​used as input. This feature matrix is ​​first input to a one-dimensional convolutional feature extraction layer. A convolutional kernel positioned along the time axis models the local correlation of the multi-channel features to extract high-dimensional temporal features reflecting the dynamic evolution of the signal. After nonlinear activation, the convolutional features are input to a fully connected layer to output a discrimination result related to the system's operating state. During the network inference phase, the random deactivation mechanism in the convolutional neural network is kept enabled, and multiple forward inference operations are performed on the same feature matrix. Based on the distribution of the output results obtained from multiple forward inferences, an anomaly strength score characterizing the degree of anomaly within the current time window and a cognitive uncertainty index reflecting the reliability of the discrimination result by the convolutional neural network are calculated. This enables intelligent perception and robust evaluation of complex noisy environments and anomaly patterns within the convolutional neural network structure.

[0092] Based on the results of multiple inferences, the following calculations are performed:

[0093] (1) Final anomaly intensity score

[0094] ,

[0095] (2) Cognitive uncertainty

[0096] ,

[0097] Where α is the scaling factor, used to scale the variance to the range [0, 1]; This is an operation to truncate a value to the specified interval [0, 1]. Indicates the result of multiple inferences The sample variance.

[0098] This allows AI anomaly detection and processing to achieve the following:

[0099] s k+1 The larger the value, the more severe the spike cluster / anomaly is considered by the CNN;

[0100] u k+1 The smaller the value, the more confident CNN is in its own judgment.

[0101] In step S5:

[0102] Set the minimum value of the process noise variance for the bias channel ( ) and maximum value ( ):

[0103] ,

[0104] The bias channel process noise variance is adaptively adjusted based on outlier scores and cognitive uncertainty.

[0105] ,

[0106] This invention incorporates the analysis results of the operating state within the current time window from a convolutional neural network into the process noise adjustment mechanism of the measurement bias channel. Specifically, the convolutional neural network outputs anomaly intensity information and corresponding cognitive uncertainty information based on multi-channel temporal features, and fuses and maps these two information into a process noise adjustment factor for the bias channel, thereby adaptively adjusting the rate of change of the measurement bias state.

[0107] When the anomaly intensity is high and the cognitive uncertainty is low ( Larger and When smaller, The value will be close to 1), indicating that the system is significantly disturbed and the neural network has a high confidence in the anomaly judgment. At this time, the process noise variance of the bias channel is close to the preset maximum level, thus allowing the bias state to change rapidly, so as to effectively absorb the measurement offset error caused by external interference, model mismatch or system mutation.

[0108] Conversely, when the anomaly intensity is low and the cognitive uncertainty is high ( Smaller and When it is large, The value will be close to 0), indicating that the system is running relatively smoothly and the neural network has limited reliability in judging anomalies. At this time, the process noise variance of the bias channel is close to the preset minimum level, thereby suppressing the rapid fluctuation of the bias state and avoiding unnecessary bias adjustments to the filter due to accidental noise or uncertain judgments.

[0109] Through the above methods, the present invention achieves adaptive adjustment of the bias channel based on the anomaly perception results of the convolutional neural network, thereby improving the stability and robustness of the filtering process in complex noise environments.

[0110] The adaptive adjustment of the bias channel process noise also includes verifying the validity of the adaptive process noise value. If the product of the abnormal score and the cognitive uncertainty is less than a preset threshold, the previously valid bias process noise value remains unchanged.

[0111] In step S6:

[0112] Construct the updated process noise covariance matrix:

[0113] ,

[0114] And complete the extended Kalman filter update:

[0115] ,

[0116] ,

[0117] ,

[0118] ,

[0119] like Figure 4As shown, this invention provides an enhanced Kalman filter processing flow, which includes steps such as prior prediction, innovation calculation, gain calculation, and state and covariance update. First, the state information of the previous time step is extrapolated based on the system state transition model to obtain the prior state estimate and corresponding prior covariance for the current time step, used to characterize the prediction result of the system before the introduction of new measurement information. Then, the actual measured value is compared with the predicted measured value obtained based on the prior state to calculate the measurement innovation, reflecting the deviation between the prediction result and the actual observation. Simultaneously, the innovation covariance is calculated by combining the measurement model and the prior covariance to characterize the uncertainty level of the innovation. Based on this, the Kalman gain is calculated according to the prior covariance, the measurement model, and the innovation covariance to determine the weight of the measurement information in the state correction. Next, the prior state is corrected using the Kalman gain to obtain a posterior state estimate that integrates the predicted and measurement information, and the state covariance is updated synchronously to reflect the convergence process of uncertainty. Finally, the posterior estimate result containing the system's main state and the extended measurement bias state is output, providing initial conditions for subsequent filtering iterations. Through the above steps, this invention achieves recursive estimation of system state and provides a unified filtering framework for introducing artificial intelligence-assisted adjustment mechanisms.

[0120] like Figure 5 As shown, this paper presents a comparison of the velocity estimation performance of the traditional Extended Kalman Filter (EKF) method and the Artificial Intelligence Enhanced Extended Kalman Filter (AI-EKF) method proposed in this invention under complex measurement environments with strong spike cluster noise and transonic vibrations.

[0121] (a) shows the comparison curves of velocity estimation results. It can be observed that during multiple bursts of peak noise (shaded areas in the figure), the traditional EKF estimation curve exhibits obvious instantaneous jumps and high-frequency oscillations, resulting in severe waveform distortion. In contrast, the estimation curve of the AI-EKF method of this invention maintains good smoothness and continuity, more closely tracks the reference signal, and significantly suppresses the influence of peak interference. (b) shows the corresponding comparison curves of absolute estimation errors. It can be seen that when peak interference occurs, the error of the traditional EKF shows significant pulse-like spikes, with the maximum error increasing sharply. In contrast, the error curve of the method of this invention remains at a low level with gentle fluctuations, and the maximum error value is effectively suppressed. Furthermore, throughout the entire transonic transition phase, the mean and variance of the error of the method of this invention are significantly smaller than those of the traditional method, demonstrating its superior estimation accuracy and robustness under all operating conditions.

[0122] On the other hand, the present invention also provides an artificial intelligence-enhanced extended Kalman filter system, comprising modules capable of performing the steps of the aforementioned method, specifically including:

[0123] The state prediction module is used to predict prior states based on an extended state vector containing velocity and measurement bias and a nonlinear dynamic model, and to linearize the state function to obtain the Jacobian matrix.

[0124] The auxiliary variable calculation module is used to fix the bias channel process noise to a minimum, calculate the forward prediction covariance corresponding to the prior state, and then obtain the measurement innovation and its forward covariance.

[0125] The feature construction module is used to calculate the normalized innovation feature using the measured innovation and its forward covariance, calculate the velocity increment feature by combining the internal velocity estimation of the filter, and combine the normalized innovation feature and the velocity increment feature into an AI feature matrix through a sliding time window.

[0126] The one-dimensional convolutional neural network detection module is used to input the AI ​​feature matrix into the one-dimensional convolutional neural network, perform multiple forward inferences under random deactivation conditions, and output a score representing the intensity of the anomaly and the corresponding model cognitive uncertainty.

[0127] The bias and process noise adaptive adjustment module is used to dynamically calculate and limit the adaptive process noise value of the bias channel at the current moment based on the anomaly score and the corresponding model cognitive uncertainty.

[0128] An extended Kalman filter update module is used to recalculate the state prediction covariance using the latest noise covariance matrix containing the adaptive process noise values; then, the Kalman gain is calculated, the state estimate and covariance matrix are updated, and finally, a joint estimate of velocity and measurement bias is output.

[0129] Thirdly, the present invention provides an electronic device comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned artificial intelligence-enhanced extended Kalman filter method.

[0130] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned artificial intelligence-enhanced extended Kalman filtering method.

[0131] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An artificial intelligence-enhanced extended Kalman filter method, characterized in that, include: Step 1: Based on the extended state vector including velocity and measurement bias and the nonlinear dynamic model, perform prior state prediction, and linearize the state function to obtain the Jacobian matrix; Step 2: Fix the bias channel process noise to the minimum value, calculate the forward prediction covariance corresponding to the prior state, and then obtain the measurement innovation and its forward covariance. Step 3: Calculate the normalized innovation feature using the measured innovation and its forward covariance, calculate the velocity increment feature by combining the internal velocity estimation of the filter, and combine the normalized innovation feature and velocity increment feature into an AI feature matrix through a sliding time window. Step 4: Input the AI ​​feature matrix into a one-dimensional convolutional neural network, perform multiple forward inferences under random deactivation conditions, and output the score representing the intensity of the anomaly and the corresponding model cognitive uncertainty. Step 5: Based on the anomaly score and the corresponding model cognitive uncertainty, dynamically calculate and limit the adaptive process noise value of the bias channel at the current moment; Step 6: Using the latest noise covariance matrix containing the adaptive process noise values, recalculate the state prediction covariance; then calculate the Kalman gain, update the state estimate and covariance matrix, and finally output a joint estimate of velocity and measurement bias.

2. The artificial intelligence-enhanced extended Kalman filter method according to claim 1, characterized in that, In step 1, linearizing the state function specifically means: for the nonlinear velocity dynamics equation, performing a first-order Taylor expansion at the current velocity state estimate to obtain the state transition Jacobian matrix, where the cross-derivative term between the velocity state and the bias state is zero.

3. The artificial intelligence-enhanced extended Kalman filter method according to claim 1, characterized in that, In step 2, the noise variance corresponding to the measurement innovation is configured as a time variable that varies with the system velocity or Mach number to characterize the phenomenon of increased measurement noise power caused by flow field instability in the transsonic region; the measurement noise includes spike cluster noise with continuous high amplitude burst characteristics.

4. The artificial intelligence-enhanced extended Kalman filter method according to claim 1, characterized in that, In step 3, the length of the sliding time window is between 32 and 128 sampling points, which is used to construct a two-row, multi-column feature matrix, where the first row is the normalized innovation sequence arranged in chronological order, and the second row is the velocity increment sequence at the corresponding time.

5. The artificial intelligence-enhanced extended Kalman filter method according to claim 1, characterized in that, In step 4, the number of input data channels of the one-dimensional convolutional neural network is 2, which correspond to the normalized innovation sequence and the speed increment sequence, respectively; its network structure includes at least two one-dimensional convolutional layers, a random deactivation layer and a fully connected layer, wherein the features extracted by the convolutional layers are passed to the fully connected layer after average pooling in the time dimension.

6. The artificial intelligence-enhanced extended Kalman filter method according to claim 1, characterized in that, In step 4, the number of forward inferences is between 4 and 16; the anomaly score is obtained by calculating the average of the output results of multiple inferences; and the model cognitive uncertainty is obtained by calculating the variance of the output results of multiple inferences and then performing linear scaling and numerical truncation.

7. The artificial intelligence-enhanced extended Kalman filter method according to claim 1, characterized in that, In step 5, the specific method for dynamically calculating and limiting the amplitude is as follows: based on the preset minimum value of the bias process noise, add an adjustment amount consisting of the difference between the preset maximum and minimum noise values, the anomaly score, and the product of one minus the cognitive uncertainty, and limit the calculation result to between the preset minimum and maximum values.

8. The artificial intelligence-enhanced extended Kalman filter method according to claim 1, characterized in that, In step 5, the adaptive adjustment of the bias channel process noise also includes verifying the validity of the adaptive process noise value. If the product of the abnormal score and the cognitive uncertainty is less than a preset threshold, the previously valid bias process noise value remains unchanged.

9. The artificial intelligence-enhanced extended Kalman filter method according to claim 1, characterized in that, The function of the one-dimensional convolutional neural network is limited to outputting the anomaly score and model cognitive uncertainty to drive adaptive adjustment of bias channel process noise, without directly modifying the measurement data, innovation value, or the state update equation of the extended Kalman filter.

10. An artificial intelligence-enhanced extended Kalman filter system, characterized in that, include: The state prediction module is used to predict prior states based on an extended state vector containing velocity and measurement bias and a nonlinear dynamic model, and to linearize the state function to obtain the Jacobian matrix. The auxiliary variable calculation module is used to fix the bias channel process noise to a minimum, calculate the forward prediction covariance corresponding to the prior state, and then obtain the measurement innovation and its forward covariance. The feature construction module is used to calculate the normalized innovation feature using the measured innovation and its forward covariance, calculate the velocity increment feature by combining the internal velocity estimation of the filter, and combine the normalized innovation feature and the velocity increment feature into an AI feature matrix through a sliding time window. The one-dimensional convolutional neural network detection module is used to input the AI ​​feature matrix into the one-dimensional convolutional neural network, perform multiple forward inferences under random deactivation conditions, and output a score representing the intensity of the anomaly and the corresponding model cognitive uncertainty. The bias and process noise adaptive adjustment module is used to dynamically calculate and limit the adaptive process noise value of the bias channel at the current moment based on the anomaly score and the corresponding model cognitive uncertainty. An extended Kalman filter update module is used to recalculate the state prediction covariance using the latest noise covariance matrix containing the adaptive process noise values; then, the Kalman gain is calculated, the state estimate and covariance matrix are updated, and finally, a joint estimate of velocity and measurement bias is output.

Images