State estimation method, system, device and storage medium based on model mismatch compensation network

Abstract

The state estimation method based on model mismatch compensation network provided by this invention includes: constructing a state-space model containing kinetic model mismatch terms and measurement model mismatch terms; obtaining a compensation filter prediction equation and a compensation filter update equation based on the state-space model; obtaining kinetic model mismatch terms through a pre-trained kinetic compensation network and obtaining measurement model mismatch terms through a pre-trained measurement compensation network; correcting the compensation filter prediction equation and the compensation filter update equation based on the kinetic model mismatch terms and the measurement model mismatch terms; and estimating the target state based on the corrected compensation filter prediction equation and the corrected compensation filter update equation. This invention directly corrects the system model mismatch terms through kinetic compensation network and measurement compensation network, effectively solving the problem of decreased filtering estimation accuracy caused by model inaccuracy, and significantly improving the state estimation accuracy and robustness in model mismatch scenarios. This system has the same beneficial effects.

Description

Technical Field

[0001] This invention relates to the field of positioning technology, and in particular to a state estimation method, system, device, and storage medium based on a model mismatch compensation network. Background Technology

[0002] The Kalman filter, as an optimal recursive estimation algorithm, is widely used in many important fields such as robotics, financial analysis, biomedicine, and weather reconstruction. Based on a "prediction-update" recursive framework, this algorithm achieves real-time and accurate estimation of the state of a dynamic system by fusing information from the system's dynamics model and measurement model. However, the estimation performance of the Kalman filter is highly dependent on the accuracy of the dynamics and measurement models. In practical applications, factors such as model simplification and insufficient understanding of target motion patterns and sensor states often lead to a mismatch between the constructed model and the real system. For example, in target tracking tasks, it is often difficult to predict the timing and manner of the target's movement, and the actual dynamics model of the target's motion often fails to strictly conform to standard motion models such as uniform velocity and uniform turning. For the measurement model, problems such as sensor calibration deviations, installation position offsets, clock synchronization errors, or hardware failures can cause systemic deviations in the measurement model. These model inaccuracies directly lead to reduced state estimation accuracy, and in severe cases, can even cause filter divergence. Therefore, model mismatch has become a core bottleneck restricting the application of the Kalman filter.

[0003] To mitigate the impact of model mismatch, traditional adaptive filtering methods, such as strong tracking filtering and Sage-Husa filtering, dynamically adjust the second-order covariance of state and measurement noise through heuristic rules. While this enhances the adaptability to inaccurate models, it lacks rigorous theoretical support and its stability is difficult to guarantee. On the other hand, variational Bayes-based methods model the state covariance and measurement noise covariance as inverse Wissaud distributions and use the VB method to adaptively iterate the distribution parameters, providing a solid theoretical foundation for adaptive adjustment. However, they inevitably introduce variational approximation errors. Neural network-assisted filtering methods, such as KalmanNet and Split-KalmanNet, directly learn key parameters (such as Kalman gain and covariance matrix) in Kalman filtering through deep neural networks and integrate them into the filtering framework. These methods reduce dependence on training data by incorporating state-space model information and make up for the deficiencies of pure deep learning in uncertainty quantification. However, their essence is still to adjust the covariance to expand the uncertainty quantification range, thereby tolerating model errors. They do not fundamentally correct the mismatch between the dynamic model and the measurement model. Therefore, when the model mismatch is severe, their performance improvement is limited, and the bias in state estimation will still accumulate over time.

[0004] Therefore, how to provide a state estimation method and system based on model mismatch compensation network that can directly correct the mismatch terms in the system model, so as to improve the state estimation accuracy in model mismatch scenarios, is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This invention addresses the technical problem of decreased accuracy in Kalman filter state estimation using traditional methods and existing neural network-assisted methods in scenarios where there is a mismatch in the dynamics model and / or measurement model. The objective of this invention is to provide a state estimation method that can directly learn and compensate for the model's own mismatch terms, thereby fundamentally improving the accuracy and robustness of state estimation.

[0006] The purpose of this invention is to provide a state estimation method based on a model mismatch compensation network;

[0007] The technical solution provided by this invention is as follows:

[0008] The state estimation method based on model mismatch compensation network includes the following steps:

[0009] A state-space model containing mismatch terms of the dynamic model and the measurement model is constructed, and the compensation filter prediction equation and the compensation filter update equation are obtained based on the state-space model.

[0010] The mismatch terms of the dynamic model are obtained through a pre-trained dynamic compensation network, and the mismatch terms of the measurement model are obtained through a pre-trained measurement compensation network.

[0011] The compensation filter prediction equation and the compensation filter update equation are modified based on the mismatch terms of the dynamic model and the measurement model;

[0012] Target state estimation is performed based on the modified compensation filter prediction equation and the modified compensation filter update equation.

[0013] Preferably, the expression for the state-space model is:

[0014]

[0015]

[0016] in, It is the target state at time t; For the preset dynamic model; The true state at time t-1 Mismatch term in the dynamic model For process noise The sensor measurement value at time t Preset measurement model; For the measurement model mismatch term, For measuring noise.

[0017] Preferably, the kinetic model mismatch term includes a kinetic mismatch term mean vector and a kinetic mismatch term covariance, and the measurement model mismatch term includes a measurement mismatch term mean vector and a measurement mismatch term covariance.

[0018] The compensation filter prediction equation is as follows:

[0019]

[0020]

[0021] in, The predicted state mean vector at time t; This is the linearization matrix of the dynamic model; This is the mean vector of the states after the update at time t-1; This is the mean vector of the dynamic mismatch term; Predict the state covariance at time t; The state covariance after update at time t-1; for The transpose of the matrix; is the covariance of the kinetic mismatch term; Q is the process noise covariance.

[0022] It should be noted that the compensation filter prediction equation and the compensation filter update equation are based on the following assumptions. Posterior probability density at time 1 Follows the mean vector The covariance matrix is The Gaussian distribution is obtained under the condition that it is independent of process noise and measurement noise.

[0023] Preferably, the compensation filter update equation is:

[0024]

[0025]

[0026]

[0027]

[0028] Among them, the Let be the mean vector of the states after the update at time t; Kalman gain; This is the linearization matrix of the measurement model; The value of the mismatch term is measured at time t; This is the mean vector of the measurement mismatch terms; for The transpose of the matrix; Let be the state covariance after the update at time t; For the new information covariance; R is the measurement mismatch covariance; R is the measurement noise covariance.

[0029] Preferably, the training process of the pre-trained dynamics compensation network and the pre-trained measurement compensation network includes:

[0030] Obtain a labeled dataset containing the target observation sequence and the corresponding real state sequence;

[0031] Based on the labeled dataset, a differential sequence feature combination is constructed as the input features for the dynamic compensation network and the measurement compensation network;

[0032] The loss functions of the dynamic compensation network and the measurement compensation network are multi-task weighted loss functions;

[0033] The multi-task weighted loss function is:

[0034]

[0035] in, The mean error of the state estimation; Let be the covariance error matrix, where for The transpose of the matrix; is the weight hyperparameter; tr() is the matrix trace.

[0036] Preferably, both the dynamic compensation network and the measurement compensation network include:

[0037] The feature extraction component, consisting of a linear layer, a ReLU activation function, and a gated recurrent unit, is used to extract temporal features from time-series data.

[0038] The shared feature component, consisting of a linear layer, a ReLU activation function, and a random deactivation layer, is used to map the temporal features to the state space;

[0039] The output component consists of two linear layers, which output the mean vector and covariance of the mismatch terms, respectively.

[0040] Preferably, the step of correcting the compensation filter prediction equation and the compensation filter update equation based on the mismatch terms of the dynamic model and the measurement model includes:

[0041] The mean vector of the dynamic mismatch term output by the dynamic compensation network and the covariance of the dynamic mismatch term output by the dynamic compensation network are used as the mean vector of the dynamic mismatch term and the covariance of the dynamic mismatch term in the compensation filter prediction equation and the compensation filter update equation, respectively.

[0042] The mean vector and covariance of the dynamic mismatch term output by the dynamic compensation network, and the mean vector and covariance of the measurement mismatch term output by the measurement compensation network are used as the corresponding mean vector and covariance of the measurement mismatch term in the compensation filter update equation.

[0043] The second objective of this invention is to provide a state estimation system based on a model mismatch compensation network;

[0044] The technical solution provided by this invention is as follows:

[0045] State estimation systems based on model mismatch compensation networks include:

[0046] The construction module is used to construct a state-space model containing mismatch terms of the dynamic model and the measurement model, and to obtain the compensation filter prediction equation and the compensation filter update equation based on the state-space model.

[0047] The mismatch term determination module is used to obtain the mismatch term of the dynamic model through a pre-trained dynamic compensation network and to obtain the mismatch term of the measurement model through a pre-trained measurement compensation network.

[0048] The correction module is used to correct the compensation filter prediction equation and the compensation filter update equation based on the mismatch terms of the dynamic model and the measurement model.

[0049] The state estimation module is used to estimate the target state based on the modified compensation filter prediction equation and the modified compensation filter update equation.

[0050] The third objective of this invention is to provide a computer device;

[0051] The technical solution provided by this invention is as follows:

[0052] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any one of the state estimation methods based on a model mismatch compensation network.

[0053] A fourth objective of this invention is to provide a computer-readable storage medium;

[0054] The technical solution provided by this invention is as follows:

[0055] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any one of the state estimation methods based on a model mismatch compensation network.

[0056] This invention provides a state estimation method based on a model mismatch compensation network, comprising the following steps: constructing a state-space model containing kinetic model mismatch terms and measurement model mismatch terms; obtaining a compensation filter prediction equation and a compensation filter update equation based on the state-space model; obtaining kinetic model mismatch terms through a pre-trained kinetic compensation network and measurement model mismatch terms through a pre-trained measurement compensation network; correcting the compensation filter prediction equation and the compensation filter update equation based on the kinetic model mismatch terms and the measurement model mismatch terms; and estimating the target state based on the corrected compensation filter prediction equation and the corrected compensation filter update equation. This invention directly corrects the system model mismatch terms through kinetic compensation networks and measurement compensation networks, effectively solving the problem of decreased filtering estimation accuracy caused by model inaccuracy, and significantly improving the state estimation accuracy and robustness in model mismatch scenarios.

[0057] The present invention also provides a state estimation system based on a model mismatch compensation network. Since this system and the state estimation method based on a model mismatch compensation network solve the same technical problem and belong to the same technical concept, they should have the same beneficial effects, and will not be described in detail here. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a flowchart of a state estimation method based on a model mismatch compensation network in an embodiment of the present invention;

[0060] Figure 2 This is a structural diagram of the dynamic compensation network and the measurement compensation network in an embodiment of the present invention;

[0061] Figure 3 This is a comparison chart of the state estimation accuracy between the prior art and the method described in this embodiment of the invention;

[0062] Figure 4 This is a schematic diagram of the structure of a state estimation system based on a model mismatch compensation network in an embodiment of the present invention;

[0063] Figure 5 This is an internal structural diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0064] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0065] like Figure 1 As shown, this embodiment of the invention provides a state estimation method based on a model mismatch compensation network, including the following steps:

[0066] S1. Construct a state-space model containing mismatch terms of the dynamic model and the measurement model, and obtain the compensation filter prediction equation and the compensation filter update equation based on the state-space model.

[0067] In step S1, the state-space model is a discrete-time state-space model with mismatch terms in the dynamic model and the measurement model.

[0068] Preferably, the expression for the state-space model is:

[0069]

[0070]

[0071] in, Let t be the target state; For the preset dynamic model; The true state at time t-1 Mismatch term in the dynamic model For process noise The sensor measurement value at time t For the preset measurement model; For the measurement model mismatch term, For measuring noise.

[0072] It should be noted that, and These represent the dynamic model and measurement model, respectively, constructed based on the understanding of the real system. Process noise. and It is zero-mean white noise, and the two are independent of each other.

[0073] In practical applications, the compensation filter prediction equation and compensation filter update equation are obtained based on the state-space model. Specifically, based on the initial state-space model, the preset dynamic model, the preset measurement model, and the Kalman wave "prediction-update" framework, the initial compensation filter prediction equation and compensation filter update equation are derived.

[0074] Preferably, the mismatch term in the kinetic model includes the mean vector of the kinetic mismatch term and the covariance of the kinetic mismatch term, and the mismatch term in the measurement model includes the mean vector of the measurement mismatch term and the covariance of the measurement mismatch term;

[0075] The compensation filter prediction equation is:

[0076]

[0077]

[0078] in, The predicted state mean vector at time t; This is the linearization matrix of the dynamic model; This is the mean vector of the states after the update at time t-1; This is the mean vector of the dynamic mismatch term; Predict the state covariance at time t; The updated state covariance at time t-1; for The transpose of the matrix; is the covariance of the kinetic mismatch term; Q is the process noise covariance.

[0079] Preferably, the compensation filter update equation is:

[0080]

[0081]

[0082]

[0083]

[0084] Among them, the Let be the mean vector of the states after the update at time t; Kalman gain; This is the linearization matrix of the measurement model; The value of the mismatch term is measured at time t; This is the mean vector of the measurement mismatch terms; for The transpose of the matrix; Let be the state covariance after the update at time t; For the new information covariance; R is the measurement mismatch covariance; R is the measurement noise covariance.

[0085] It should be noted that the compensation filter prediction equation and compensation filter update equation in step S1 are the initial compensation equations, and the mean vector of the dynamic mismatch term, the covariance of the dynamic mismatch term, the mean vector of the measurement mismatch term, and the covariance of the measurement mismatch term are obtained in the subsequent step S2.

[0086] S2. Obtain the mismatch term of the dynamic model through the pre-trained dynamic compensation network, and obtain the mismatch term of the measurement model through the pre-trained measurement compensation network;

[0087] It should be noted that the dynamic compensation network and the measurement compensation network are neural network models, specifically, they can be deep neural networks (DNN), multilayer perceptron neural networks (LSTM), recurrent neural networks (RNN), etc.

[0088] As one specific implementation, the dynamic compensation network and the measurement compensation network are DNN neural networks.

[0089] Preferably, such as Figure 2 As shown, both the dynamic compensation network and the measurement compensation network include:

[0090] The feature extraction component, consisting of a linear layer, a ReLU activation function, and a gated recurrent unit, is used to extract temporal features from time-series data.

[0091] The shared feature component, consisting of a linear layer, a ReLU activation function, and a random deactivation layer, is used to map temporal features to the state space;

[0092] The output component consists of two linear layers, which output the mean vector and covariance of the mismatch terms, respectively.

[0093] Preferably, the training process of the pre-trained dynamics compensation network and the pre-trained measurement compensation network includes:

[0094] Obtain a labeled dataset containing the target observation sequence and the corresponding real state sequence;

[0095] Based on the labeled dataset, a combination of differential sequence features is constructed as the input features for the dynamics compensation network and the measurement compensation network;

[0096] The loss functions of the dynamic compensation network and the measurement compensation network are multi-task weighted loss functions;

[0097] The multi-task weighted loss function is:

[0098]

[0099] in, The mean error of the state estimation; Let be the covariance error matrix, where for The transpose of the matrix; is the weight hyperparameter; tr() is the matrix trace.

[0100] In practical applications, the initialized dynamics compensation network and the initialized measurement compensation network are trained using supervised learning. First, a labeled dataset is formed by acquiring the observation sequence of the target system and the corresponding state truth sequence. Based on the standard dataset, a difference sequence feature combination is constructed as the input features for the dynamics compensation network and the measurement compensation network. The difference sequence feature combination includes:

[0101] F1:

[0102] F2:

[0103] F3:

[0104] F4: - -

[0105] A differential sequence feature combination is constructed based on prior knowledge, wherein features F1 and F3 are used to characterize the uncertainty of state estimation, and features F2 and F4 are used to characterize information about the state and measurement evolution process. Preferably, the differential sequence features composed of F1 and F2 are used as input to the dynamic compensation network to estimate the mismatch term of the dynamic model; and the differential sequence features composed of F3 and F4 are used as input to the measurement compensation network to estimate the mismatch term of the measurement model.

[0106] To train the network, construct a multi-task weighted loss function containing the mean vector and covariance matrix of the compensation term, and define... The error is the state mean vector error. The state covariance matrix error is represented by the following loss function:

[0107]

[0108] in, The mean error of the state estimation; Let be the covariance error matrix, where for The transpose of the matrix; This is a weight hyperparameter, and its value range is typically... Its core function is to balance the contribution weights of mean vector estimation and covariance matrix estimation to the total loss. Preferably, =0.95; tr() is the matrix trace.

[0109] The initial dynamics compensation network and the initial measurement compensation network are initially trained using the time backpropagation algorithm. When the preset training target is reached, the pre-trained dynamics compensation network and the pre-trained measurement compensation network are obtained.

[0110] S3. Based on the mismatch terms of the dynamic model and the measurement model, the prediction equation and the update equation of the compensation filter are corrected.

[0111] Preferably, the compensation filter prediction equation and the compensation filter update equation are corrected based on the mismatch terms of the dynamic model and the measurement model, including:

[0112] The mean vector of dynamic mismatch terms and the covariance of dynamic mismatch terms output by the dynamic compensation network are used as the mean vector of dynamic mismatch terms and the covariance of dynamic mismatch terms in the compensation filter prediction equation and the compensation filter update equation, respectively.

[0113] The mean vector and covariance of the dynamic mismatch term output by the dynamic compensation network, and the mean vector and covariance of the measurement mismatch term output by the measurement compensation network are used as the corresponding mean vector and covariance of the measurement mismatch term in the compensation filter update equation.

[0114] It should be noted that the compensation filter prediction equation and compensation filter update equation in step S1 are the initial compensation equations. Step S3 is to replace the mean vector of the dynamic mismatch term, the covariance of the dynamic mismatch term, and the mean vector of the measurement mismatch term and the covariance of the measurement mismatch term in the compensation filter prediction equation and compensation filter update equation in step S1.

[0115] S4. Target state estimation is performed based on the modified compensation filter prediction equation and the modified compensation filter update equation.

[0116] Compared to existing technologies, this invention designs a compensation network oriented towards model mismatch. Based on an inaccurate state-space model, it learns the mismatch terms of the dynamic and measurement models in the Kalman filter through a dynamic compensation network and a measurement compensation network, respectively. This directly corrects the corresponding mismatch term parameters in the prediction-update equation, thereby directly correcting the inaccuracies of the dynamic and measurement models. Furthermore, this invention fully integrates prior knowledge into the dynamic and measurement compensation networks, focusing on correcting the mismatch parts in the model and compensation equations without reconstructing the complex global dynamics of the system, thus reducing the solution space for network learning.

[0117] To verify the effectiveness of the method of the present invention, experiments were conducted using the Michigan NCLT dataset, with the goal of locating the estimated position of the mobile robot from the raw measurements at any given time.

[0118] like Figure 3 As shown in the figure, the state estimation accuracy of four methods, including this method, is compared with the ground truth. The methods compared include:

[0119] This method: proposed;

[0120] Comparison Method 1: Split-KalmanNet, an improved neural network filtering method that learns state and measurement covariance separately, [Reference: Choi G, Park J, Shlezinger N, et al. Split-KalmanNet: Arobust model-based deep learning approach for state estimation[J]. IEEE transactions on vehicular technology, 2023, 72(9): 12326-12331.];

[0121] Comparison Method 2: Extended Kalman Filter (EKF), using a preset CV model and observation model, [Reference: Sunahara Y, Yamashita K. An approximate method of state estimation for non-linear dynamical systems with state-dependent noise[J]. International Journal of Control, 1970, 11(6): 957-972.];

[0122] Comparison Method 3: KalmanNet, a neural network filtering method that learns Kalman gain, [Reference: Revach G, Shlezinger N, Ni X, et al. KalmanNet: Neural network aided Kalman filtering for partially known dynamics[J]. IEEE Transactions on Signal Processing, 2022, 70: 1532-1547.].

[0123] In one embodiment, such as Figure 4 As shown, a state estimation system based on a model mismatch compensation network includes:

[0124] The module is used to construct a state-space model containing mismatch terms of the dynamic model and the measurement model, and to obtain the compensation filter prediction equation and the compensation filter update equation based on the state-space model.

[0125] The mismatch term determination module is used to obtain the mismatch terms of the dynamic model through a pre-trained dynamic compensation network and the mismatch terms of the measurement model through a pre-trained measurement compensation network.

[0126] The correction module is used to correct the compensation filter prediction equation and the compensation filter update equation based on the mismatch terms of the dynamic model and the measurement model.

[0127] The state estimation module is used to estimate the target state based on the modified compensation filter prediction equation and the modified compensation filter update equation.

[0128] Specific limitations regarding the state estimation system based on model mismatch compensation networks can be found in the limitations of a state estimation method based on model mismatch compensation networks described above, and will not be repeated here. Each module in the aforementioned state estimation system based on model mismatch compensation networks can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0129] like Figure 5 As shown, in one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the following steps:

[0130] A state-space model containing mismatch terms of the dynamic model and the measurement model is constructed, and the compensation filter prediction equation and the compensation filter update equation are obtained based on the state-space model.

[0131] The mismatch terms of the dynamic model are obtained through a pre-trained dynamic compensation network, and the mismatch terms of the measurement model are obtained through a pre-trained measurement compensation network.

[0132] The compensation filter prediction equation and compensation filter update equation are modified based on the mismatch terms of the dynamic model and the measurement model.

[0133] Target state estimation is performed based on the modified compensation filter prediction equation and the modified compensation filter update equation.

[0134] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0135] In one embodiment, a computer-readable storage medium is provided, storing a computer program that, when executed by a processor, performs the following steps:

[0136] A state-space model containing mismatch terms of the dynamic model and the measurement model is constructed, and the compensation filter prediction equation and the compensation filter update equation are obtained based on the state-space model.

[0137] The mismatch terms of the dynamic model are obtained through a pre-trained dynamic compensation network, and the mismatch terms of the measurement model are obtained through a pre-trained measurement compensation network.

[0138] The compensation filter prediction equation and compensation filter update equation are modified based on the mismatch terms of the dynamic model and the measurement model.

[0139] Target state estimation is performed based on the modified compensation filter prediction equation and the modified compensation filter update equation.

[0140] In the embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple modules or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, and can be electrical, mechanical, or other forms.

[0141] Furthermore, in the various embodiments of the present invention, each functional module can be fully integrated into a processor, or each module can be a separate device, or two or more modules can be integrated into a device; each functional module in the various embodiments of the present invention can be implemented in hardware or in the form of hardware plus software functional units.

[0142] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by program instructions and related hardware. The aforementioned program instructions can be stored in a computer-readable storage medium. When the program instructions are executed, they perform the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0143] It should be understood that the use of terms such as "system," "apparatus," "unit," and / or "module" in this application is only applicable to distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0144] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "a," and / or "the" are not specifically singular and may include the plural. Generally, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements. An element defined by the phrase "comprising an..." does not exclude the presence of other identical elements in the process, method, product, or apparatus that includes the element.

[0145] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "a plurality of" or "several" means two or more, unless otherwise explicitly specified.

[0146] If a flowchart is used in this application, it is used to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0147] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A state estimation method based on model mismatch compensation network, characterized in that, Includes the following steps: A state-space model containing mismatch terms of the dynamic model and the measurement model is constructed, and the compensation filter prediction equation and the compensation filter update equation are obtained based on the state-space model. The expression for the state-space model is: ； ； in, Let t be the target state; For the preset dynamic model; This represents the true state at time t-1; This is a mismatch term in the dynamic model; This is process noise; The sensor measurement value at time t; For the preset measurement model; For the measurement model mismatch term, For measuring noise; The mismatch term in the dynamic model includes the mean vector of the dynamic mismatch term and the covariance of the dynamic mismatch term; the mismatch term in the measurement model includes the mean vector of the measurement mismatch term and the covariance of the measurement mismatch term. The compensation filter prediction equation is as follows: ； ； in, The predicted state mean vector at time t; This is the linearization matrix of the dynamic model; This is the mean vector of the states after the update at time t-1; This is the mean vector of the dynamic mismatch term; Predict the state covariance at time t; for The transpose of the matrix; The state covariance after update at time t-1; Q is the covariance of the kinetic mismatch term; Q is the process noise covariance. The compensation filter update equation is as follows: ； ； ； ； Among them, the Let be the mean vector of the states after the update at time t; Kalman gain; This is the linearization matrix of the measurement model; The value of the mismatch term is measured at time t; This is the mean vector of the measurement mismatch terms; for transpose matrix Let be the state covariance after the update at time t; For the new information covariance; R is the measurement mismatch covariance; R is the measurement noise covariance. The mismatch terms of the dynamic model are obtained through a pre-trained dynamic compensation network, and the mismatch terms of the measurement model are obtained through a pre-trained measurement compensation network. The training process of the pre-trained dynamics compensation network and the pre-trained measurement compensation network includes: Obtain a labeled dataset containing the target observation sequence and the corresponding real state sequence; Based on the labeled dataset, a differential sequence feature combination is constructed as the input features for the dynamic compensation network and the measurement compensation network; The loss functions of the dynamic compensation network and the measurement compensation network are multi-task weighted loss functions; The multi-task weighted loss function is: ； in, The mean error of the state estimation; Let be the covariance error matrix, where for The transpose of the matrix; Here are the weight hyperparameters; tr() is the matrix trace; The compensation filter prediction equation and the compensation filter update equation are modified based on the mismatch terms of the dynamic model and the measurement model; Target state estimation is performed based on the modified compensation filter prediction equation and the modified compensation filter update equation.

2. The state estimation method according to claim 1, characterized in that, Both the dynamic compensation network and the measurement compensation network include: The feature extraction component, consisting of a linear layer, a ReLU activation function, and a gated recurrent unit, is used to extract temporal features from time-series data. The shared feature component, consisting of a linear layer, a ReLU activation function, and a random deactivation layer, is used to map the temporal features to the state space; The output component consists of two linear layers, which output the mean vector and covariance of the mismatch terms, respectively.

3. The state estimation method according to claim 2, characterized in that, The step of correcting the compensation filter prediction equation and the compensation filter update equation based on the mismatch terms of the dynamic model and the measurement model includes: The mean vector of the dynamic mismatch term output by the dynamic compensation network and the covariance of the dynamic mismatch term output by the dynamic compensation network are used as the mean vector of the dynamic mismatch term and the covariance of the dynamic mismatch term in the compensation filter prediction equation and the compensation filter update equation, respectively. The mean vector and covariance of the dynamic mismatch term output by the dynamic compensation network, and the mean vector and covariance of the measurement mismatch term output by the measurement compensation network are used as the corresponding mean vector and covariance of the measurement mismatch term in the compensation filter update equation.

4. A state estimation system based on a model mismatch compensation network, used to implement the method described in any one of claims 1-3, characterized in that, include: The construction module is used to construct a state-space model containing mismatch terms of the dynamic model and the measurement model, and to obtain the compensation filter prediction equation and the compensation filter update equation based on the state-space model. The mismatch term determination module is used to obtain the mismatch term of the dynamic model through a pre-trained dynamic compensation network and to obtain the mismatch term of the measurement model through a pre-trained measurement compensation network. The correction module is used to correct the compensation filter prediction equation and the compensation filter update equation based on the mismatch terms of the dynamic model and the measurement model. The state estimation module is used to estimate the target state based on the modified compensation filter prediction equation and the modified compensation filter update equation.

5. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the state estimation method based on a model mismatch compensation network as described in any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the state estimation method based on a model mismatch compensation network as described in any one of claims 1-3.

Images