Sensor dynamic zero correction method and system based on deep Kalman filtering

Abstract

The invention provides a sensor dynamic zero point correction method based on deep Kalman filtering, and relates to the technical field of sensor zero point tracking. The method comprises the following steps: acquiring voltage signal zero point output time sequence data of a sensor; the obtained data serve as input of a deep learning network model based on Kalman filtering, the model processes the data, noise and zero drift interference are removed from the output data, dynamic zero correction is achieved, and a real measurement value of the sensor is obtained. The deep learning network model based on Kalman filtering comprises an RNN recurrent neural network module, a Pyro probability programming module and a variational inference module. According to the invention, the external interference is simulated by using the environmental vibration to carry out the experiment, and the result shows that the deep Kalman filter model provides a new idea for the dynamic zero correction of the force transducer.

Description

Technical Field

[0001] The present invention belongs to the technical field of sensor zero point tracking, and relates to a sensor dynamic zero point correction method, and in particular to a sensor dynamic zero point correction method and system based on deep Kalman filtering, which can perform dynamic zero point correction on the sensor under the interference of vibration noise. Background Art

[0002] With the development and exploration of human beings in harsh scenes such as deep sea and deep space, as well as the development of humanoid machine applications, higher requirements are placed on the accuracy and reliability of sensors. As one of the commonly used sensors, the study of zero drift and dynamic zero correction methods of strain gauge force sensors is of great significance to the improvement of measurement accuracy and application in harsh environments.

[0003] In recent years, many new technologies have been applied to the design of force sensors, such as optical fiber (PCF) technology, capacitance technology, MEMS technology, etc. However, these new technologies are still based on the deformation principle of elastic bodies. Compared with traditional strain gauge force sensors, sensors based on these new technologies only replace the strain bridge measurement element with a new measurement unit, and it is still difficult to overcome the problem of poor dynamic characteristics. In particular, the measurement range of force sensors based on these new technologies is generally small, and they cannot be applied to fields such as automobiles, aviation, aerospace, deep sea and robots that require large-range force sensors.

[0004] Zero drift exists in any sensor. In order to improve the comprehensive performance of the sensor, it is necessary to track and compensate the zero drift of the force sensor accordingly. Compensation methods can be divided into two categories: hardware compensation method and software compensation method. The hardware compensation method mainly adopts self-compensating strain gauges, differential circuits, and replacement of constant current excitation sources. Sensor software zero drift compensation mainly finds out the factors affecting zero drift, and establishes relevant compensation models through comprehensive analysis of zero drift data under different environments. There is currently a lack of systematic research on the dynamic zero correction technology of force sensors, and there are few reports on force sensors that have successfully applied dynamic zero correction technology. There are even fewer related studies on dynamic zero tracking of sensors in various aspects through methods such as neural networks, and most of them are still in the laboratory stage, making it difficult to achieve specific industrial applications. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies in the prior art and to provide a method and system for dynamic zero-point correction of sensors based on deep Kalman filtering. The present invention is a method and system for dynamic zero-point correction of sensors that combines deep learning related algorithms with digital processing technologies such as filtering to explore dynamic zero-point correction technology of sensors. The method can provide a certain theoretical basis for the development of high-precision and high-performance sensors such as force sensors and solve related technical problems.

[0006] The technical solution adopted by the present invention to solve the above problems is as follows:

[0007] A sensor dynamic zero point correction method based on deep Kalman filtering includes:

[0008] The zero-point output time series data of the sensor's voltage signal is obtained, which includes the zero-point drift caused by vibration noise interference. The obtained data is used as the input of a deep learning network model based on Kalman filtering. The model processes the data, and the output data has removed noise and zero-point drift interference, realizing dynamic zero-point correction and obtaining the true measurement value of the sensor.

[0009] In the above technical solution, further, the deep learning network model based on Kalman filtering includes an RNN recurrent neural network module, a Pyro probability programming module, and a variational inference module; wherein:

[0010] The RNN recurrent neural network module is used to process time series data; it has a feedback mechanism that enables the network to dynamically adjust weights and continuously circulate information when processing sequence data; its output is affected by the current input, the historical state of the network, and the future state, and sends the initial hidden state and the observation sequence to the variational inference module;

[0011] The Pyro probabilistic programming module includes two classes: Emitter and Gated Transition. The Emitter class is used to define the conditional distribution of observed variables given latent variables. The GatedTransition class is used to accept initial latent variables and parameterize Gaussian latent transition probabilities.

[0012] The variational inference module includes a Combiner class, which is used to receive the hidden state sent by the RNN recurrent neural network module, define the variational distribution of the latent variables, and combine the previous latent variables and the hidden state of the RNN to generate the parameters of the distribution.

[0013] Furthermore, the Pyro probabilistic programming framework is used to define probabilistic models and perform variational inference, including the model's forward propagation method and variational inference method (guide), as well as parameter initialization and other functions required for model training.

[0014] Furthermore, in the Pyro probabilistic programming module, the Emitter class is used to define a given latent variable z t In the case of observed variables x t In this model, a neural network with a multi-layer perceptron structure is used to parameterize the Gaussian observation likelihood p(x t |z t), using linear transformations and nonlinear functions (ReLU activation function) to learn the parameters of the mean vector that maps latent variables to observed values, thereby improving the expressiveness of the model.

[0015] Furthermore, in the Pyro probabilistic programming module, the Gated Transition class should be based on the potential variable z at the given previous moment. t-1 , outputs the mean and variance vector of the latent variable at the current moment. Its structure is designed as a variant of a gated recurrent unit (GRU), including gating functions, proposed mean and variance calculations, etc. This design enables the model to better consider uncertainty in the inference process, improving the generalization ability and inference accuracy of the model.

[0016] Furthermore, the variational inference module also includes a Combiner class that defines a latent variable z t The variational distribution q(z t |z t-1 ,x :t ), the previous latent variable z t-1 and the hidden state h(x) of the RNN :t ) are combined to generate the parameters of the distribution. The idea of ​​Kalman filtering is adopted, and the information of observed data and state transition is combined to obtain a more accurate posterior estimate by updating the mean and variance of the potential state at the current moment.

[0017] Furthermore, the Kalman filter idea is integrated into the deep learning model, and the above modules are integrated through the defined DKF class to form a complete deep Kalman filter model (DKF model).

[0018] A sensor dynamic zero point correction system based on deep Kalman filtering is used to implement the method described in any one of the above items.

[0019] The beneficial effects of the present invention are:

[0020] Based on Kalman filtering, the present invention uses deep learning-related algorithms for optimization design, studies a deep Kalman filter model suitable for the dynamic zero correction problem of sensors, and verifies the feasibility of dynamic zero correction of the model under the interference of sensor vibration noise through experiments. Compared with the prior art, the present invention has the following important advances:

[0021] 1) Fill the gap in technology. Currently, dynamic zero-point tracking is generally achieved through Kalman filtering and its extension or hardware circuits, and there is no dynamic zero-point correction technology in the context of deep learning. Compared with hardware correction technology, the method of the present invention has the characteristics of high tracking accuracy, strong versatility, and simple design. The experimental results show that this method can reduce the interference of noise in the environment to the zero point to 3% of the total interference amplitude, meeting the application needs in production, scientific research and harsh environments.

[0022] 2) Through the introduction of RNN recurrent neural network, Kalman filter, and variational inference methods, the technology can better process time series data and take uncertainty into account during the inference process. The accuracy and robustness of this combined model are improved, which is especially important for processing time series data with complex dynamic characteristics.

[0023] 3) The present invention experimentally tests the designed cylindrical force sensor and obtains the output zero-point related data of the sensor under vibration noise interference. Compared with the common mean filter, median filter, and standard Kalman filter algorithms, the deep Kalman filter model proposed in the present invention has a better data filtering effect and can be applied to the dynamic zero-point correction requirements of the force sensor. The data analysis results verify the good performance of the DKF model.

[0024] 4) The present invention can achieve effective noise suppression, reduce interference and noise in the sensor output zero point, thereby improving measurement accuracy and stability. In the dynamic zero point correction process, the influence of noise on the sensor output is accurately eliminated, making the measurement results more reliable and accurate, enhancing the robustness of the system, and helping to optimize data processing efficiency and reduce the impact on subsequent data processing and analysis.

[0025] 5) The method of the present invention is easy to integrate with hardware circuits and has a wide range of applications. In the face of high-precision and high-intensity mechanical operations, in order to ensure the stability and accuracy of the force sensor, real-time dynamic zero point correction has a strong practical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a schematic diagram of a deep Kalman filter network model of the present invention;

[0027] Figure 2 A schematic diagram of the technical route of the dynamic zero point correction method of the force sensor of the present invention;

[0028] Figure 3 It is a schematic diagram of the experimental results of the present invention;

[0029] Figure 4 Schematic diagram of the cylindrical force sensor used in the experiment of the present invention. DETAILED DESCRIPTION

[0030] The present invention is further described in detail below with reference to the accompanying drawings and specific examples.

[0031] According to a specific embodiment of the present invention, the core of the dynamic zero-point correction method of the sensor based on deep Kalman filtering of the present invention is to construct a deep learning network model (DKF model) based on Kalman filtering, by obtaining the zero-point output time series data of the sensor's voltage signal, which includes the zero-point drift caused by vibration noise interference; the obtained data is used as the input of the deep learning network model based on Kalman filtering, and the model processes it to remove noise and zero-point drift interference, realize dynamic zero-point correction, and obtain the true measurement value of the sensor.

[0032] The model can be mainly composed of the following parts:

[0033] 1) State space modeling: Similar to Kalman filtering, we first need to define a state space model, including the state transfer equation and the observation equation. The state transfer equation is used to describe how the system state evolves over time, and the observation equation is used to describe how the observation value is associated with the system state.

[0034] 2) State estimation: Using observed data and state space models, state estimation is performed through filtering algorithms. In Kalman filtering, state estimation is achieved by recursively updating the mean and covariance of the state. Therefore, state estimation of latent variables needs to be implemented in deep learning models.

[0035] 3) Uncertainty modeling: Kalman filtering uses the covariance matrix to represent the uncertainty of state estimation. In deep learning models, the probability programming library Pyro can be used to model parameter uncertainty, such as using variational inference to estimate the posterior distribution and obtain confidence intervals for parameters.

[0036] 4) Model evaluation: The deep learning model can use mean square error or other evaluation indicators similar to Kalman filtering to evaluate the performance of the model, and adjust the model structure or hyperparameters as needed. During the training process of the model, the model parameters are optimized by maximizing ELBO so that the model fits the data better. Therefore, the model includes an RNN recurrent neural network module, a Pyro probabilistic programming module, and a variational inference module; wherein:

[0037] The RNN recurrent neural network module is used to process the input time series data and send the obtained hidden state to the variational inference module;

[0038] The Pyro probabilistic programming module includes two classes: Emitter and Gated Transition. The Emitter class is used to define the conditional distribution of observed variables given latent variables. The GatedTransition class is used to accept initial latent variables and parameterize Gaussian latent transition probabilities.

[0039] The variational inference module includes a Combiner class, which is used to receive the hidden state sent by the RNN recurrent neural network module, define the variational distribution of the latent variables, and combine the previous latent variables and the hidden state of the RNN to generate the parameters of the distribution.

[0040] Assume that x represents the original observed data, and z represents the hidden variables generated by the model. The main goal of variational inference is to obtain a probability distribution q(z|x) in order to approximate the posterior distribution p(z|x). According to Bayes' theorem, the following derivation can be obtained:

[0041]

[0042] In the above formula, KL(q(z|x)||p(z|x)) represents relative entropy, also known as Kullback-Leibler (KL) divergence. KL divergence is a non-negative value that measures the degree of difference between two probability distributions. In order to minimize the gap between distributions p(z|x) and q(z|x), it is necessary to minimize the KL divergence. Since p(x) is the observed distribution of the original data, it is fixed and unchanged, so the above goal can be achieved by maximizing the L(q) term. This term is called the optimization objective of variational inference, also known as the evidence lower bound (ELBO). This method can make it as close to the target distribution as possible, thereby achieving approximate inference.

[0043]

[0044] The first term in ELBO represents the reconstruction loss of the model, which is used to measure the new data reconstructed by the model by generating variables z The difference between the original observed data x. In other words, the first term evaluates the performance of the model in generating data, that is, the degree of match between the data generated by the model and the real data. The second term is the KL divergence, which is a regularization term that measures the approximation between the posterior distribution q(z|x) and the prior distribution p(z). In variational inference, it is usually necessary to select a prior distribution and then optimize the posterior distribution by maximizing the ELBO to make it as close as possible to the true posterior distribution.

[0045] By combining reconstruction loss and KL divergence, ELBO provides a comprehensive objective function that takes into account both the quality of the data generated by the model and the degree of match between the posterior distribution and the prior distribution. In summary, these two considerations can be balanced during model training, thereby improving the generalization ability and inference accuracy of the model.

[0046] Take the force sensor as an example. Figure 1 As shown, the data processing process in this model can be described as follows:

[0047] 1. Data Input

[0048] Original observation data (x_0): Time series data obtained from the force sensor according to the experiment. Here it refers to the zero-point output of the sensor voltage signal, including the zero-point drift caused by vibration noise interference.

[0049] 2. Initial hidden state setting

[0050] Initial hidden state (h_0) and latent variables (z_0): Before calculation, the RNN module and the variational inference module need to initialize their hidden states and latent variables. Here they are set to zero.

[0051] 3.RNN recurrent neural network module

[0052] Observation sequence (x_t) input to RNN: The raw observation data is input to the RNN module. RNN processes these inputs and generates hidden states of the time series.

[0053] Generation of hidden state (h_t): The RNN module processes time series data and outputs a hidden state (h_t), which captures the information of current and past observations.

[0054] 4. Variational Inference Module

[0055] Hidden state (h_t) is input to the variational inference module: The hidden state generated by the RNN is used as input to generate the distribution parameters (mean and variance) of the latent variable (z_t) through the variational inference module.

[0056] Update of latent variable (z_t): The variational inference module combines the hidden state (h_t) and the latent variable at the previous moment (z_{t-1}) to generate a new latent variable (z_t). This latent variable captures the intrinsic state of the force sensor.

[0057] 5. The role of the Combiner class

[0058] Fusion of different information sources: The Combiner class receives the hidden state (h_t) from the RNN module and the latent variable (z_t) from the variational inference module and fuses this information.

[0059] Generate comprehensive estimates: The Combiner class outputs a comprehensive latent variable estimate (z_t^*) that combines time series information (from the RNN (h_t)) and latent state information (from the variational inference (z_t)), i.e., provides a more accurate state estimate.

[0060] 6. Pyro Probabilistic Programming Module

[0061] Generation of conditional observation distribution (x_t|z_t^*): Use the comprehensively estimated latent variable (z_t^*) to generate the conditional distribution of the current observation variable (x_t) through the Emitter class. This distribution reflects the possibility of observing data under the current latent state.

[0062] Modeling of state transition (z_t^*|z_{t-1}^*): The Gated Transition class receives the latent variable of the previous moment (z_{t-1}^*), generates the mean and variance of the latent variable of the current moment (z_t^*), and updates the state.

[0063] 7. Forward Propagation and Correction

[0064] Correcting zero drift: Combining the observed data (x_t) at the current moment and the estimated potential state (z_t^*), the dynamic zero drift of the force sensor is gradually corrected during the forward propagation process.

[0065] Optimization and Posterior Distribution Update: Using ELBO (Evidence Lower Bound) optimization, the posterior distribution is optimized by maximizing ELBO and the mean and variance of the latent variables are updated to obtain a more accurate state estimate.

[0066] 8. Model Output

[0067] Corrected sensor data: Finally, the corrected sensor data (x_t) processed by the DKF model is output. These data have been corrected by the model, and the noise and zero drift interference have been removed, reflecting the true measurement value of the sensor.

[0068] According to a specific embodiment of the present invention, the preparation and training process of the model may be as follows:

[0069] 1. Data preparation

[0070] First, the code reads time series data from two CSV files and merges them into a complete data set. Then, the data is divided into training set, validation set and test set. In order to ensure the stability and generalization ability of the model, the data is also standardized to have zero mean and unit variance; the data in the two CSV files are the sensor voltage zero-point outputs obtained by interference experiments under different vibration noise frequencies of the force sensor. The data obtained by using different vibration noise interference is to distinguish the training set from the test set;

[0071] 2. Model definition and initialization

[0072] After data preparation, a deep Kalman filter model was defined and related hyperparameters such as input dimension, latent variable dimension and RNN network size were set. For variational inference, an automatic guider (AutoGuide) was used to approximate the posterior distribution and a variational inferencer (SVI) was defined to optimize the model parameters.

[0073] 3. Model training

[0074] The model is trained over multiple iterations, with training and validation losses calculated for each iteration. During training, the model's parameters are gradually optimized to minimize the loss function. In addition, at certain training cycles, the model's predictive power is evaluated, and the model's performance is checked by generating predictions and comparing them with actual data. The training steps here are: iterate num_epochs times, gradually optimize the model parameters, and in each iteration, use the step method to train the model and calculate the training loss. Calculate the loss on the validation set to evaluate the model's performance on unseen data. Evaluate the model at specified cycles (e.g., every 200 cycles), generate predictions, and visualize them.

[0075] 4. Result Evaluation and Visualization

[0076] After a specified period of training, the model will predict the entire time series and calculate evaluation indicators such as the coefficient of determination (R2), mean absolute error (MAE), and root mean square error (RMSE). These indicators help determine the performance of the model on the training set, validation set, and test set. Finally, the model's prediction results are compared with the actual data in the form of charts to intuitively demonstrate the model's effectiveness.

[0077] Figure 2 , 34 provides an example of dynamic zero-point correction of a force sensor based on deep Kalman filtering of the present invention. In this example, a cylindrical force sensor is taken as the research object, and a parametric modeling program is written in Python language based on the secondary development interface of ABAQUS. Its working principle is analyzed through finite element simulation, thereby improving the accuracy and stability of the force sensor. In view of the noise interference problem that is prevalent in the use of sensors, experiments are conducted using environmental vibrations to simulate external interference, and dynamic zero-point correction is performed for the force sensor through the deep Kalman filter model of the present invention. The feasibility of the model is verified by processing and analyzing the experimental data. In this example, a cylindrical force sensor (such as Figure 4 Based on the geometric configuration characteristics and the output zero-point drift caused by environmental vibration simulation noise interference, a dynamic zero-point correction scheme is designed. The technical route is as follows Figure 2 As shown. The following steps are included:

[0078] Step 1: Use Python language to complete the secondary development process of ABAQUS and realize parametric modeling of force sensors, including geometric modeling, automatic assignment of materials and cross-sections, load application, meshing, and job submission.

[0079] Step 2: Determine the optimal patch position of the strain gauge in the sensor based on the finite element simulation results, and calculate the relevant dimensions of the sensor elastic body;

[0080] Step 3: Build an experimental platform and calibrate the force sensor;

[0081] Step 4: Use vibration to simulate noise interference in actual applications, design experiments, and obtain the voltage output zero point;

[0082] Step 5 studies the output zero point change through the data acquisition device and pre-processes the collected data; Step 6 uses the DKF model to train the data set and analyze the results to achieve dynamic zero point correction of the force sensor.

[0083] Specific:

[0084] First, the secondary development of ABAQUS was completed using Python language, and the parametric modeling of the cylindrical force sensor was realized, including geometric modeling, automatic assignment of materials and cross-sections, load application, meshing, and submission of assignments. By analyzing the strain curve of the inner surface of the bottom of the elastomer, it can be found that the thickness of the bottom of the elastomer and the radius of the connecting part have a great influence on the zero-point output and sensitivity of the sensor. That is, these two parts are the key parameters of the elastomer, and they need to be reasonably designed before the sensor is manufactured. Therefore, using the finite element simulation results completed by the modeling program, the optimal patch position of the strain gauge in the sensor is determined, and the relevant dimensions of the sensor elastomer are calculated.

[0085] Then build the experimental platform and complete the sensor calibration. That is, use a high-precision sensor (also called a standard sensor) on the force standard machine to generate a high-precision standard reference force value for the sensor under test. By comparing the output of the sensor under test with the reference calibration value, the sensor under test can be calibrated and adjusted to improve its measurement accuracy. Finally, the calibrated sensor is verified to ensure its measurement accuracy and stability under different working conditions. The specific experimental steps are as follows:

[0086] 1) Use a screw and nut combination of appropriate size to fix the high-precision force sensor to the unidirectional force standard testing machine and connect it to the data acquisition device;

[0087] 2) Load setting is performed through the unidirectional force standard testing machine operating platform to ensure the accuracy and stability of the reference force value, and record it for subsequent use;

[0088] 3) Place the cylindrical force sensor on the gasket, ensure that the relevant equipment is wired correctly, and load it;

[0089] 4) During the force application process, the sensor output value is recorded in real time to ensure that the data acquisition frequency is high enough to capture all changes, and the output mV / V (that is, sensitivity) of the cylindrical force sensor is obtained based on the data analysis results.

[0090] Finally, the motor vibration of the base is used to simulate the vibration and noise interference of the environment in actual applications. The collected data is preprocessed, combined with the DKF model training data set and compared with the commonly used filtering methods. Figure 3 The experimental results shown in the visualization show that the output zero-point fluctuation amplitude under vibration interference is controlled within 3% of the total interference amplitude, the correlation between the training set learning results and the actual data can reach 99.8%, the mean absolute error between the training prediction value and the true value is 0.021, and the root mean square error is 0.027, realizing the dynamic zero-point correction of the force sensor in a vibration environment.

[0091] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0092] The above-described embodiment is only a preferred solution of the present invention, but it is not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. A sensor dynamic zero point correction method based on deep Kalman filtering, characterized in that: include: The zero-point output time series data of the sensor's voltage signal is obtained, which includes the zero-point drift caused by vibration noise interference. The obtained data is used as the input of a deep learning network model based on Kalman filtering. The model processes the data, and the output data has removed noise and zero-point drift interference, realizing dynamic zero-point correction and obtaining the true measurement value of the sensor.

2. The sensor dynamic zero point correction method based on deep Kalman filtering according to claim 1 is characterized in that: The Kalman filter-based deep learning network model includes an RNN recurrent neural network module, a Pyro probability programming module, and a variational inference module; wherein: The RNN recurrent neural network module is used to process the input time series data and send the obtained hidden state to the variational inference module; The Pyro probabilistic programming module includes two classes: Emitter and Gated Transition. The Emitter class is used to define the conditional distribution of observed variables given latent variables. The GatedTransition class is used to accept initial latent variables and parameterize Gaussian latent transition probabilities. The variational inference module includes a Combiner class, which is used to receive the hidden state sent by the RNN recurrent neural network module, define the variational distribution of the latent variables, and combine the previous latent variables and the hidden state of the RNN to generate the parameters of the distribution.

3. The sensor dynamic zero point correction method based on deep Kalman filtering according to claim 2 is characterized in that: The Pyro probability programming module uses the Pyro probability programming framework to define probability models and perform variational inference, including the forward propagation method of the model and variational inference to estimate the posterior distribution and obtain the confidence interval of the parameters.

4. The sensor dynamic zero point correction method based on deep Kalman filtering according to claim 2 is characterized in that: The Emitter class in the Pyro probabilistic programming module is used to define a given latent variable z t In the case of observed variables x t The conditional distribution of the Gaussian observation likelihood p(x t |z t ), using a linear transformation and a nonlinear function, namely the ReLU activation function, to learn the parameters of the mean vector that maps the latent variables to the observed values.

5. The sensor dynamic zero point correction method based on deep Kalman filtering according to claim 2 is characterized in that: The Gated Transition class in the Pyro probabilistic programming module is based on the latent variable z at the previous moment. t-1 , outputs the mean and variance vector of the latent variable at the current moment; its structure is a variant of the gated recurrent unit GRU, including the gating function, the proposed mean and variance calculation.

6. The sensor dynamic zero point correction method based on deep Kalman filtering according to claim 2 is characterized in that: The Combiner class in the variational inference module defines the latent variable z t The variational distribution q(z t |z t-1 ,x :t ), the latent variable z at the previous moment t-1 and the hidden state h(x) of the RNN :t ) are combined to generate the parameters of the distribution, and the idea of ​​Kalman filtering is adopted. The observation data and state transition information are combined to obtain a more accurate posterior estimate by updating the mean and variance of the potential state at the current moment.

7. The sensor dynamic zero point correction method based on deep Kalman filtering according to claim 2 is characterized in that: The model processes the input time series data as follows: Before processing, the RNN recurrent neural network module and the variational inference module need to initialize their hidden states and latent variables to the initial hidden state (h_0) and latent variable (z_0). The original observation data (x_t) is input into the RNN recurrent neural network module to generate the hidden state (h_t) of the time series, which is used as input to generate the distribution parameters of the latent variable (z_t), namely the mean and variance, through the variational inference module, and combined with the hidden state (h_t) and the latent variable (z_{t-1}) at the previous moment to generate a new latent variable (z_t). This latent variable captures the intrinsic state of the sensor. The Combiner class receives the hidden state (h_t) from the RNN recurrent neural network module and the latent variable (z_t) from the variational inference module, fuses this information, and outputs a comprehensive estimated latent variable (z_t^*) to the Pyro probabilistic programming module. The conditional distribution (x_t|z_t^*) of the current observation variable (x_t) is generated through the Emitter class. Gated The Transition class receives the latent variable (z_{t-1}^*) of the previous moment, generates the mean and variance of the latent variable (z_t^*) of the current moment, updates the state, combines the observed data (x_t) of the current moment and the estimated latent state (z_t^*), gradually corrects the dynamic zero drift of the sensor during the forward propagation process, uses ELBO evidence lower bound optimization, optimizes the posterior distribution by maximizing ELBO, updates the mean and variance of the latent variable to obtain a more accurate state estimate, and finally outputs the processed sensor data.

8. The sensor dynamic zero point correction method based on deep Kalman filtering according to claim 2 is characterized in that: The training method of the model is as follows: Firstly, the sensor voltage zero-point output time series data obtained from the interference experiment under different vibration noise frequencies of the sensor are obtained, and they are merged into a complete data set, which is standardized to have zero mean and unit variance, and these data are divided into training set, validation set and test set. Define the deep Kalman filter model and set the hyperparameters, including the input dimension, latent variable dimension, and RNN network size. In order to perform variational inference, use the AutoGuide to approximate the posterior distribution and define a variational inference SVI to optimize the model parameters. The model is trained through multiple iterations, and the training loss and validation loss are calculated in each iteration. During the training process, the model parameters are gradually optimized to minimize the loss function. The training steps are: through loop iteration num_epochs times, the model parameters are gradually optimized. In each iteration, the model is trained using the step method and the training loss is calculated. After the specified training cycle, the model predicts the entire time series and calculates the evaluation indicators to judge the performance of the model on the training set, validation set and test set.

9. A sensor dynamic zero point correction system based on deep Kalman filtering, characterized in that: Used to implement the method according to any one of claims 1 to 8.

Images