Unsupervised state estimation method based on model and data driving technology

Abstract

The invention discloses an unsupervised state estimation method based on a model and a data driving technology. The unsupervised state estimation method comprises the following steps: S1, performing prior prediction and posterior estimation on a state in a dynamic system through a state estimation framework; s2, separating task-independent global parameters from task-specific local parameters through a meta-learning strategy of parameter partition; and S3, a soft constraint mechanism is constructed through physical information loss adjustment, and domain knowledge is introduced to ensure the rationality of state conversion. According to the method, the robustness, generalization and interpretability of state estimation are remarkably improved, and an efficient and reliable solution is provided for state estimation of a complex dynamic system.

Description

Technical Field

[0001] The present invention relates to the technical field of state estimation of dynamic systems, and particularly to an unsupervised state estimation method combining model-based and data-driven technologies. Background Art

[0002] State estimation is one of the core tasks in the analysis and control of dynamic systems, and is widely used in fields such as autonomous driving, robot control, and sensor networks. Traditional state estimation methods usually rely on accurate physical models, such as Kalman filters and their extensions, which depend on the accuracy of the system transition model and the measurement model. However, in practical applications, dynamic systems often have highly nonlinear, non-Gaussian noise, and model uncertainties, which make model-based methods show significant limitations in complex scenarios. At the same time, data-driven methods capture system dynamics by learning historical data and can well adapt to complex nonlinear systems when there is sufficient data. However, data-driven methods are highly dependent on labeled data and lack physical interpretability, which often limits their practical application capabilities in dynamic environments. In recent years, hybrid strategies that combine data-driven and model-based methods have received extensive attention. Such methods attempt to strike a balance between robustness and flexibility. However, they still face problems such as low fusion efficiency and high dependence on data, and further optimization is urgently needed. Summary of the Invention

[0003] Aiming at the deficiencies in the prior art, the purpose of the present invention is to provide an unsupervised state estimation method combining model-based and data-driven technologies, so that the state transition of the model is more in line with the actual physical laws, effectively improving the adaptability of the model in unknown scenarios, significantly enhancing the robustness, generalization, and interpretability of state estimation, and providing an efficient and reliable solution for the state estimation of complex dynamic systems. To achieve the above object and other advantages according to the present invention, an unsupervised state estimation method combining model-based and data-driven technologies is provided, including: S1. Perform prior prediction and posterior estimation on the state in the dynamic system through a state estimation framework; S2. Separate task-irrelevant global parameters and task-specific local parameters through a meta-learning strategy with parameter partitioning; S3. Construct a soft constraint mechanism through physical information loss adjustment, and introduce domain knowledge to ensure the rationality of state transition.

[0004] Preferably, the present application is implemented through a strategy method that combines data-driven prior prediction and model-based posterior estimation, including using a mixture density network to parameterize a Gaussian mixture model to capture complex nonlinear and non-Gaussian dynamics in the prior prediction stage and improve the modeling ability of multimodal state distributions.

[0005] Preferably, in the data-driven prior prediction, through the meta-learning strategy of parameter partitioning, the task-agnostic global parameters are separated from the task-specific local parameters, improving the generalization ability of the model in unknown scenarios.

[0006] Preferably, in this application, a physically-guided loss function is introduced simultaneously. By combining soft constraints with domain knowledge, the state transition of the model conforms more to the basic physical laws, thereby improving the reliability and data efficiency of the model.

[0007] Preferably, through the combination of multiple strategies, not only the need for a large amount of labeled data is reduced, but also the performance and interpretability of the model in dynamic environments are significantly enhanced, providing a robust, efficient, and flexible solution for the state estimation of dynamic systems. Detailed implementation manners

[0008] Next, in combination with the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. Embodiment 1

[0009] An unsupervised state estimation method combining model-based and data-driven technologies, comprising: First, parameterize the Gaussian mixture model through a mixture density network to achieve prior prediction of the system state and capture complex non-linear and non-Gaussian dynamic characteristics. Secondly, in combination with the current measurement data, update the posterior distribution through Bayesian inference and introduce UD decomposition to improve numerical stability. Then, adopt the meta-learning strategy of parameter partitioning to separate the task-agnostic global parameters from the task-specific local parameters and enhance the generalization ability of the data-driven model. In addition, integrate multiple loss functions through a joint unsupervised learning architecture to enable the model to effectively learn the state estimation task under unlabeled conditions. Finally, construct a soft constraint mechanism through physical information loss adjustment, introduce domain knowledge to ensure the rationality of state transition, and enhance the interpretability of the model.

[0010] Further, the fusion of data-driven and model-driven is achieved through a state estimation framework based on prior prediction and posterior estimation. Specifically, first input historical data into the mixture density network model. The model extracts features from the input data through a multi-layer neural network and generates Gaussian mixture model parameters describing the state distribution. The parameters include the weights, means, and covariance matrices of the components.

[0011] ; Among them, represents the weight of the th Gaussian distribution, and represent the mean and covariance matrix of the prior of the th Gaussian distribution respectively. These parameters are used to construct the prior probability distribution of the state, specifying the range of possible states at each time step. In principle, the mixture density network can represent any conditional probability distribution and is very useful when the modeled phenomenon cannot be well represented by a simpler distribution.

[0012] For the covariance matrix of the state prior prediction, this method uses the UD decomposition method for parameterization implementation. The UD decomposition, , decomposes a positive semi - definite symmetric matrix into an upper triangular matrix with all diagonal elements equal to 1 and a diagonal matrix with all diagonal elements greater than or equal to zero, which is a variant of the matrix LU decomposition. Through this decomposition method, the system only needs to check the sign of the D elements, and by ensuring the non - negativity of the D elements, the positive semi - definiteness of the covariance is forced. Further, the UD form of the Kalman filter is borrowed to perform posterior update processing on the covariance factor. This method provides an efficient and numerically stable solution for the parameterization of the covariance, ensuring the symmetry and positive semi - definiteness of the covariance while usually having high efficiency.

[0013] Subsequently, when the system obtains new measurement values, the measurement values are combined with the prior probability distribution, and the posterior distribution is calculated through Bayesian inference. In this process, the measurement values are used to update the estimation result of the state, thereby correcting the prediction deviation and obtaining a more accurate posterior distribution of the state.

[0014] ; where ; where represents the weight of the th posterior Gaussian distribution, and represent the mean and covariance matrix of the th posterior Gaussian distribution respectively. The whole process is based on the Bayesian estimation framework, and finally the mean and covariance of the posterior state estimation are obtained. This iterative inference process from prior to posterior, without assuming the specific distribution characteristics of the measurement noise, only assuming its mean is zero, can effectively handle complex non - linear and non - Gaussian environments, thus improving the robustness and adaptability of the model.

[0015] Furthermore, in order to improve the adaptability of the fusion-driven model in diverse scenarios, the present invention introduces a meta-learning strategy based on parameter partitioning. Specifically, during the model training process, the hidden state parameters are divided into two parts: global parameters and local parameters. The global parameters are used to capture the general dynamic characteristics of the dynamic system, such as the changing trends determined by physical laws or invariant laws. The global parameters are optimized through cross-task meta-learning to ensure their strong generalization ability. On this basis, the local parameters are adaptively adjusted according to the dynamic characteristics of specific tasks. For example, in different environments or conditions, the local parameters can be quickly adjusted according to the real-time input to reflect the specific requirements of the current scenario.

[0016] ; Among them, represents the hidden state parameters, and represent the global parameters and local parameters respectively.

[0017] The meta-learning strategy of parameter partitioning realizes the dynamic update of global and local parameters by combining with a recurrent neural network. Among them, the global parameters provide basic stability, and the local parameters ensure flexible response to environmental changes. Finally, this two-layer optimization mechanism enables the model to not only retain the stability of global characteristics but also perform excellently in quickly adapting to task requirements.

[0018] Furthermore, a soft constraint mechanism is constructed by adjusting the physical information loss, introducing domain knowledge to ensure the rationality of state transitions, and enhancing the interpretability of the model. Specifically, an existing physical model is used to model the state transition process. This model can provide a basic prediction of state evolution. Even if it is not completely accurate or has simplified assumptions, it can still be used as a reliable prior guidance. In each model update, the difference between the state estimation value and the physical model prediction value is calculated, and this difference is added to the total loss as part of the physical loss to guide the model to gradually approach the physical laws during the optimization process.

[0019] ; Among them, represents the total loss, represents the training loss of the hybrid drive model, represents the physical loss, Represents the weight of physical information loss. The core of this method lies in reducing the over - dependence of the model on the data - driven module through physical information regularization, especially in cases where data is scarce or noisy. Specifically, the physical loss term can not only improve the physical consistency of the model prediction results but also effectively narrow the optimization search space, enhancing the training efficiency and stability. By combining data - driven and physical models, the present invention significantly enhances the robustness of the model in high - noise and low - data scenarios while ensuring its good generalization ability for unknown tasks.

[0020] In summary, the present invention proposes an unsupervised state - estimation method that combines model - based Bayesian inference and data - driven learning. This method uses a mixture density network to parameterize the Gaussian mixture model, captures non - linear and non - Gaussian dynamic characteristics in the prior prediction stage, and at the same time combines a physics - guided loss function to incorporate domain knowledge through soft constraints, making the state transition of the model more in line with actual physical laws. In addition, this method adopts a meta - learning strategy of parameter partitioning, separating task - independent global parameters from task - specific local parameters, effectively enhancing the adaptability of the model in unknown scenarios. Through the above innovations, the present invention significantly improves the robustness, generalization, and interpretability of state estimation, providing an efficient and reliable solution for the state estimation of complex dynamic systems.

[0021] The number of devices and the processing scale described here are used to simplify the description of the present invention, and applications, modifications, and variations of the present invention will be obvious to those skilled in the art.

[0022] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the field, additional modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations here.

Claims

1. An unsupervised state estimation method combining model-based and data-driven techniques, characterized in that: The following steps are involved: S1. Perform a priori prediction and a posteriori estimation of the state in a dynamic system through a state estimation framework; S2, through the meta-learning strategy of parameter partitioning, the task-independent global parameters are separated from the task-specific local parameters; S3. Build a soft constraint mechanism through physical information loss adjustment and introduce domain knowledge to ensure the rationality of state transition.

2. The unsupervised state estimation method combining model-based and data-driven techniques as claimed in claim 1, characterized in that: The a priori prediction of the state in the dynamic system in step S1 specifically includes the following steps: S1, input historical data into the mixed density network model; S2. The mixed density network model extracts features from the input data through a multi-layer neural network to generate Gaussian mixture model parameters that describe the state distribution; and the parameters are used to construct a priori probability distribution of the state and clarify the possible state range in each time step.

3. The unsupervised state estimation method combining model-based and data-driven techniques as claimed in claim 2, characterized in that: The parameters include component weights, means, and covariance matrices, and the covariance matrix is ​​parameterized by UD decomposition.

4. The unsupervised state estimation method combining model-based and data-driven techniques as claimed in claim 1, characterized in that: The a posteriori estimation of the state in the dynamic system in step S1 is specifically that when the dynamic system obtains a new measurement value, the measurement value is combined with the prior probability distribution, and the posterior distribution is calculated by Bayesian reasoning, wherein the measurement value is used to update the estimation result of the state, thereby correcting the prediction deviation.

5. The unsupervised state estimation method combining model-based and data-driven techniques as claimed in claim 1, characterized in that: In step S2, the meta-learning strategy of parameter partitioning is combined with a recursive neural network to achieve dynamic update of global and local parameters, specifically: In the data-driven model training process, the hidden state parameters are divided into global parameters and local parameters. The global parameters are used to capture the general dynamic characteristics of the dynamic system, and the global parameters are optimized through cross-task meta-learning. The local parameters are adaptively adjusted according to the dynamic characteristics of a specific task.

6. The unsupervised state estimation method combining model-based and data-driven techniques as claimed in claim 1, characterized in that: The step S2 also includes integrating multiple loss functions through a joint unsupervised learning architecture so that the fusion-driven model can effectively learn the state estimation task under unlabeled conditions. The joint unsupervised learning architecture is composed of a comprehensive loss function consisting of measurement error, state estimation error and maximum likelihood objective.

7. The unsupervised state estimation method combining model-based and data-driven techniques as claimed in claim 1, characterized in that: The step S3 specifically involves modeling the state transition process using an existing physical model, and providing a basic prediction of the state evolution through the built model; In each model update, the difference between the state estimate and the physical model prediction is calculated and added to the total loss as part of the physical loss to guide the model to gradually approach the physical laws during the optimization process.