Training method and using method of electron content prediction model and related device

Abstract

The embodiment of the invention discloses a training method, a using method and a related device of an electron content prediction model, and the method comprises the steps: carrying out the feature extraction of the total electron content data, geomagnetic activity index data and solar activity index data of an ionized layer collected by a Beidou satellite navigation system through a random forest algorithm; obtaining key feature data corresponding to the total electron content data; adopting a sliding window mechanism, taking the key feature data under a preset time window length as input time sequence data, taking the total electron content data at the last moment of the time window length as target output, and constructing a training sample pair; and performing prediction training of total electron content data according to the training sample pair and a preset deep learning model to obtain a trained electron content prediction model. In this way, the nonlinear dynamic characteristics of the ionosphere delay are effectively captured by the deep learning model, the prediction precision of the total electron content is improved, and the correction precision of the ionosphere delay is improved.

Description

Technical Field

[0001] The present invention relates to the field of satellite navigation technology, and in particular to a training method, a use method and a related device of an electronic content prediction model. Background Art

[0002] The BeiDou Navigation Satellite System (BDS) is a global satellite navigation system independently developed by China. One of its core functions is to provide users with high-precision positioning, navigation, and timing (PNT) services. However, ionospheric delay is a major source of error affecting the BDS's positioning accuracy, particularly in high-precision applications such as unmanned driving, aerospace, and precision agriculture.

[0003] The ionosphere is a region of Earth's atmosphere containing a large number of free electrons and ions. Electromagnetic wave signals passing through the ionosphere are delayed by the free electrons. This delay manifests itself in nonlinearity, regional variations, and dynamic complexity. Because the Beidou system's positioning accuracy relies on the propagation time of satellite signals, ionospheric delay can result in positioning errors ranging from several to more than ten meters. Therefore, precise correction for this delay is a key technology for improving Beidou system accuracy.

[0004] Because ionospheric delay exhibits complex nonlinear dynamic characteristics, traditional empirical models (such as the Klobuchar model and the IRI model) are unable to accurately capture these changes, resulting in a decrease in the correction accuracy of ionospheric delay. Therefore, the current correction accuracy of ionospheric delay is insufficient and needs to be improved. Summary of the Invention

[0005] The main purpose of the present invention is to provide a training method, a use method and related devices for an electron content prediction model, which can solve the lack of means in the prior art to improve the correction accuracy of ionospheric delay.

[0006] To achieve the above objectives, the present invention provides, in a first aspect, a method for training an electron content prediction model, the method comprising:

[0007] Acquiring historical observation data collected by the Beidou satellite navigation system, the historical observation data including at least total electron content data of the ionosphere, geomagnetic activity index data, and solar activity index data at various historical moments, the total electron content data being used to reflect ionospheric delay;

[0008] Performing feature extraction on the total electron content data, geomagnetic activity index data, and solar activity index data of the ionosphere using a random forest algorithm to obtain key feature data corresponding to the total electron content data;

[0009] A sliding window mechanism is used to construct training sample pairs by taking the key feature data within a preset time window length as the input time series data and the total electron content data at the last moment of the time window length as the target output;

[0010] Prediction training of total electron content data is performed based on the training sample pairs and a preset deep learning model to obtain a trained electron content prediction model.

[0011] In one feasible implementation, the obtaining of historical observation data collected by the Beidou satellite navigation system further includes:

[0012] The historical observation data are preprocessed to obtain preprocessed historical observation data, wherein the preprocessing at least includes outlier removal and missing value interpolation.

[0013] In one feasible implementation, performing prediction training of total electron content data based on the training sample pairs and a preset deep learning model to obtain a trained electron content prediction model includes:

[0014] Use the preset bidirectional long short-term memory network to process the input time series data to obtain the spliced ​​hidden state vector;

[0015] Calculating the attention weight of the concatenated hidden state vector based on a preset attention mechanism;

[0016] Generate a context vector using the attention weight, the concatenated hidden state vector, and a preset context vector algorithm;

[0017] Inputting the context vector into a preset fully connected layer and activation function to predict the total electron content data at the last moment, thereby obtaining the total electron content prediction data at the last moment;

[0018] Determining whether the deep learning model has converged based on the total electron content data at the last moment, the total electron content prediction data, and a preset loss function;

[0019] If the deep learning model converges, the converged deep learning model is used as the electron content prediction model.

[0020] In one feasible implementation, the bidirectional long short-term memory network includes at least a forward network calculation, and the forward network calculation includes the following expression:

[0021]

[0022] Where x t :Input time series data X=[x1,x2,...,x t ]’s eigenvector at time step t;

[0023] The activation value of the input gate at time t;

[0024] σ: Sigmoid activation function;

[0025] W i f : weight matrix of input gate;

[0026] Bias of the input gate;

[0027] f t f : activation value of the forget gate;

[0028] The activation value of the output gate;

[0029] The cell state at time step t;

[0030] tanh: hyperbolic tangent activation function;

[0031] The weight matrix of the cell state;

[0032] The forward hidden state at time step t;

[0033] The forward hidden state at time step t-1.

[0034] In one possible implementation, the attention weight includes the following expression:

[0035]

[0036] in:

[0037] h t : The concatenated hidden state vector at the t-th time step;

[0038] αt: attention weight at time step t;

[0039] W α : attention weight matrix;

[0040] T: total number of time steps;

[0041] exp: exponential function;

[0042] h k : The concatenated hidden state vector at the k-th time step.

[0043] In one possible implementation, the context vector algorithm includes the following expression:

[0044]

[0045] in:

[0046] C: context vector;

[0047] αt: attention weight at time step t;

[0048] h t : The concatenated hidden state vector at the t-th time step;

[0049] T: Total number of time steps.

[0050] To achieve the above-mentioned object, the second aspect of the present invention provides a method for using an electron content prediction model, the method comprising:

[0051] Obtain target observation data within the current time window length collected by the Beidou satellite navigation system;

[0052] Inputting the target observation data into an electron content prediction model to predict total electron content data to obtain target total electron content data at the current moment, wherein the electron content prediction model is trained based on the training method described in the first aspect and any feasible implementation manner;

[0053] The target total electron content data and a preset ionospheric delay algorithm are used to obtain an ionospheric delay time, and the ionospheric delay time is used to correct the positioning result of the Beidou satellite navigation system.

[0054] In one possible implementation, the ionospheric delay algorithm includes the following mathematical expression:

[0055]

[0056] in:

[0057] ΔT: ionospheric delay time;

[0058] TEC: target total electron content data;

[0059] f: signal frequency.

[0060] To achieve the above-mentioned object, the present invention provides, in a third aspect, a training device for an electron content prediction model, the device comprising:

[0061] A first acquisition module is configured to acquire historical observation data collected by the BeiDou satellite navigation system, wherein the historical observation data includes at least total electron content data of the ionosphere, geomagnetic activity index data, and solar activity index data at each historical moment, wherein the total electron content data is used to reflect ionospheric delay;

[0062] Feature extraction module: used to extract features from the total electron content data, geomagnetic activity index data and solar activity index data of the ionosphere using a random forest algorithm to obtain key feature data corresponding to the total electron content data;

[0063] Sample construction module: used to construct training sample pairs using a sliding window mechanism, taking key feature data within a preset time window length as input time series data and the total electron content data at the last moment of the time window length as the target output;

[0064] Model training module: used to perform prediction training of total electron content data based on the training sample pairs and the preset deep learning model to obtain a trained electron content prediction model.

[0065] To achieve the above-mentioned object, the present invention provides, in a fourth aspect, a device for using an electron content prediction model, the device comprising:

[0066] The second acquisition module is used to obtain the target observation data within the current time window length collected by the Beidou satellite navigation system;

[0067] A data prediction module is configured to input the target observation data into the electron content prediction model to predict the total electron content data, thereby obtaining the target total electron content data at the current moment, wherein the electron content prediction model is trained based on the training method described in the first aspect and any feasible implementation manner;

[0068] Delay determination module: used to use the target total electron content data and a preset ionospheric delay algorithm to obtain the ionospheric delay time, and the ionospheric delay time is used to correct the positioning result of the Beidou satellite navigation system.

[0069] To achieve the above-mentioned objectives, the fifth aspect of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps shown in the first aspect, the second aspect and any feasible implementation method.

[0070] To achieve the above-mentioned objectives, the sixth aspect of the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps shown in the first aspect, the second aspect and any feasible implementation method.

[0071] The embodiments of the present invention have the following beneficial effects:

[0072] The present invention provides a training method for an electron content prediction model, the method comprising: obtaining historical observation data collected by the Beidou satellite navigation system, the historical observation data at least including total electron content data, geomagnetic activity index data, and solar activity index data of the ionosphere at each historical moment, the total electron content data being used to reflect ionospheric delay; performing feature extraction on the total electron content data, geomagnetic activity index data, and solar activity index data of the ionosphere using a random forest algorithm to obtain key feature data corresponding to the total electron content data; adopting a sliding window mechanism, taking the key feature data under a preset time window length as input time series data, and the total electron content data at the last moment of the time window length as the target output, to construct a training sample pair; performing prediction training on the total electron content data based on the training sample pair and a preset deep learning model to obtain a trained electron content prediction model. The above method realizes the training of the ionospheric electron content prediction model, and obtains the electron content prediction model, so that the complex nonlinear dynamic characteristics of the ionospheric delay due to solar activity, geomagnetic activity, and regional differences can be effectively captured by the deep learning model, thereby improving the prediction accuracy of the total electron content and the correction accuracy of the ionospheric delay. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0074] in:

[0075] Figure 1 Flowchart of a training method for an electron content prediction model according to an embodiment of the present invention;

[0076] Figure 2 This is a flow chart of a method for using an electron content prediction model according to an embodiment of the present invention;

[0077] Figure 3 This is a structural block diagram of a training device for an electron content prediction model according to an embodiment of the present invention;

[0078] Figure 4 This is a structural block diagram of a device for using an electron content prediction model according to an embodiment of the present invention;

[0079] Figure 5 4 is a structural block diagram of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0080] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0081] It should be noted that ionospheric delay presents complex nonlinear dynamic characteristics due to solar activity, geomagnetic activity and regional differences. Traditional empirical models (such as Klobuchar model and IRI model) cannot accurately capture these changes. The present invention adopts a bidirectional long short-term memory network (BiLSTM) and a gated recurrent unit (GRU), combined with an attention mechanism (Attention), to effectively capture the time series characteristics of ionospheric delay and significantly improve prediction accuracy. Under extreme conditions such as magnetic storms, traditional methods (such as dual-frequency methods and empirical models) cannot adapt to the drastically changing ionospheric environment, resulting in a significant decrease in correction accuracy. The present invention integrates solar activity indices (such as F10.7, Dst), geomagnetic indices and historical TEC data through the multi-source data fusion capability of the deep learning model, so that the model has a strong adaptability to the delay characteristics of the magnetic storm period and significantly improves the correction performance. The ionospheric delay has significant differences in high latitudes, mid-latitudes and low latitudes, and traditional methods are difficult to effectively correct for different regions. The present invention uses a random forest feature selection algorithm to screen key influencing factors based on regional characteristics, and combines deep learning models to specifically model the ionospheric characteristics of different regions, thereby improving regional adaptability and prediction accuracy. Actual ionospheric delay data often contain outliers and missing values. Traditional methods have high requirements for data preprocessing and limited processing capabilities. The present invention uses the adaptive ability of deep learning, combined with data interpolation methods and anomaly detection algorithms, to effectively handle outliers and missing values, ensuring the robustness of model training and prediction. The Beidou system has high real-time requirements for ionospheric delay correction, and complex traditional correction methods may not be able to meet real-time positioning requirements. The present invention combines a lightweight deep learning model (GRU) and an attention mechanism to optimize the calculation process, reduce model complexity, and significantly improve calculation efficiency while ensuring correction accuracy, meeting the real-time requirements of the Beidou system. Traditional methods cannot fully utilize the correlation between multi-source data (such as solar activity index, geomagnetic index and historical TEC data) in ionospheric delay correction, resulting in limited correction accuracy. The present invention uses deep learning multi-source data fusion technology to explore the correlation between different data sources, thereby more comprehensively characterizing the ionospheric delay characteristics and further improving the correction effect. For details, please refer to the following content.

[0082] See also Figure 1 , Figure 1This is a flow chart of a method for training an electronic content prediction model in an embodiment of the present invention. This method can be applied to both a terminal and a server. This embodiment uses the server as an example. The terminal can be a desktop terminal or a mobile terminal. The mobile terminal can be at least one of a mobile phone, a tablet computer, a laptop computer, etc. The server can be implemented as an independent server or a server cluster consisting of multiple servers, such as Figure 1 The training method shown includes the following steps:

[0083] 101. Acquire historical observation data collected by the Beidou satellite navigation system, the historical observation data including at least total electron content data of the ionosphere, geomagnetic activity index data, and solar activity index data at various historical moments, the total electron content data being used to reflect ionospheric delay;

[0084] It is understandable that in order to train a deep learning model that can predict the total electron content (TEC) of the ionosphere, this application obtains historical observation data collected by the Beidou satellite navigation system. The historical observation data at least includes the total electron content data (TEC) of the ionosphere at each historical moment, geomagnetic activity index data and solar activity index data. The total electron content data is used to reflect the ionospheric delay. Specifically, historical observation data is collected from the Beidou satellite navigation system, including ionospheric total electron content (TEC) data, geomagnetic activity index (such as Dst, AE, etc.), solar activity index (such as F10.7, sunspot number R), pseudorange, carrier phase, satellite position, etc. By obtaining training samples from the above data, the deep learning model can learn that the ionospheric delay presents complex nonlinear dynamic characteristics due to solar activity, geomagnetic activity and regional differences.

[0085] In one feasible implementation, the obtaining of historical observation data collected by the Beidou satellite navigation system further includes:

[0086] The historical observation data are preprocessed to obtain preprocessed historical observation data, wherein the preprocessing at least includes outlier removal and missing value interpolation.

[0087] It is understood that in order to improve data quality, the historical observation data can also be preprocessed to obtain preprocessed historical observation data. The preprocessing includes at least outlier removal and missing value interpolation. For example, the collected data is cleaned, including removing outliers (e.g., using the triple standard deviation method) and interpolating missing values ​​(e.g., bilinear interpolation). The time resolution is unified to 1 hour to construct a time series data set.

[0088] 102. Perform feature extraction on the total electron content data, geomagnetic activity index data, and solar activity index data of the ionosphere using a random forest algorithm to obtain key feature data corresponding to the total electron content data;

[0089] 103. Using a sliding window mechanism, the key feature data within a preset time window length is used as the input time series data, and the total electron content data at the last moment of the time window length is used as the target output to construct a training sample pair;

[0090] Furthermore, in order to improve the effectiveness of features and the accuracy of model prediction and avoid interference of non-critical features on the model, key input features (such as F10.7 index, Dst index, etc.) are screened by random forest. Specifically, the total electron content data, geomagnetic activity index data and solar activity index data of the ionosphere are feature extracted by the random forest algorithm to obtain key feature data corresponding to the total electron content data.

[0091] Among them, feature extraction and training data construction use the random forest algorithm to sort the input data (including TEC and activity index) by importance and extract the most relevant features, such as key feature data F10.7, sunspot number R and Dst index.

[0092] Then, the training samples are constructed by adopting a sliding window mechanism, and the key feature data under the preset time window length is used as the input time series data, and the total electron content data at the last moment of the time window length is used as the target output. For example, the time window length is 7 days, and the training sample pairs are constructed as follows: using a sliding window mechanism, the time series of 7 consecutive days (that is, the time series data set constructed with a time resolution of 1 hour as mentioned above) is used as input, and the TEC value of the last day is used as the target output to form a sample pair.

[0093] 104. Perform prediction training on the total electron content data based on the training sample pairs and a preset deep learning model to obtain a trained electron content prediction model.

[0094] After obtaining the training sample pairs, prediction training of the total electron content data can be performed based on the training sample pairs and the preset deep learning model to obtain a trained electron content prediction model.

[0095] The deep learning model may include a sequentially connected bidirectional long short-term memory network (BiLSTM), an attention layer, a context vector layer, a fully connected layer, and an output layer. The bidirectional long short-term memory network (BiLSTM) includes forward network calculations, backward network calculations, and hidden feature concatenation. The attention layer includes an attention mechanism, the context vector layer includes a context vector algorithm, and the output layer includes an activation function. By inputting key feature data as input time series data into the sequentially connected bidirectional long short-term memory network (BiLSTM), the attention layer, the context vector layer, the fully connected layer, and the output layer, the predicted data of the total electron content at the last moment is obtained. The convergence of the model is determined by comparing the predicted data with the actual data. When convergence occurs, the training is considered complete, and the deep learning model has learned that the ionospheric delay exhibits complex nonlinear dynamic characteristics due to solar activity, geomagnetic activity, and regional differences, thereby obtaining an electron content prediction model.

[0096] In one possible implementation, step 104 includes steps A01 to A06:

[0097] A01. Use the preset bidirectional long short-term memory network to process the input time series data to obtain the concatenated hidden state vector;

[0098] First, the deep learning model is trained using a bidirectional long short-term memory network (BiLSTM) to process the input time series data. The input is: time series X = [x1, x2, ..., x t ], where x t is the eigenvector of the tth hour.

[0099] Exemplarily, the bidirectional long short-term memory network includes at least a forward network calculation, and the forward network calculation includes the following expression:

[0100]

[0101] It should be noted that the reverse network is similar. and reverse h t b The hidden layer state splicing is obtained by splicing the hidden state vector

[0102] Where x t :Input time series data X=[x1,x2,...,x t ]’s eigenvector at time step t;

[0103] The activation value of the input gate at time t determines the current input x t Cell status the degree of update;

[0104] σ: Sigmoid activation function, which limits the output value to between 0 and 1;

[0105] W i f : The weight matrix of the input gate, used to map the forward hidden state of the previous time step and the current input x t ;

[0106] Bias of the input gate;

[0107] f t f : The activation value of the forget gate determines the cell state at the previous time step How much information is retained?

[0108] The activation value of the output gate determines the cell state at the current time step Hidden state contribution; The cell state at the tth time step is the cell state at the previous time step The combined result of the input of the current time step;

[0109] tanh: Hyperbolic tangent activation function, which limits the input to [-1,1][-1,1][-1,1];

[0110] The weight matrix of the cell state;

[0111] The forward hidden state at time step t contains the output information of that time step;

[0112] The forward hidden state at time step t-1;

[0113] h t : The hidden state after splicing, containing bidirectional information.

[0114] A02. Calculating the attention weight of the concatenated hidden state vector based on a preset attention mechanism;

[0115] Get the concatenated hidden state vector h t After that, it can be optimized through the Attention mechanism, and the attention weight of the spliced ​​hidden state vector can be calculated through the preset attention mechanism of the attention layer.

[0116] For example, the hidden state vector H = [h1,h2,...,h T ] Calculate the attention weight, which includes the following expression:

[0117]

[0118] in:

[0119] h t : The concatenated hidden state vector at the t-th time step;

[0120] α t : The attention weight of the tth time step, indicating the importance of this time step;

[0121] W α : Attention weight matrix, used to map the hidden state h t ;

[0122] T: total number of time steps;

[0123] exp: exponential function, used to convert weight values ​​into positive numbers and normalize them to probabilities.

[0124] A03. Generate a context vector using the attention weight, the concatenated hidden state vector, and a preset context vector algorithm;

[0125] By obtaining the attention weight and the concatenated hidden state vector, a context vector can be generated. The context vector is generated using the attention weight, the concatenated hidden state vector, and a preset context vector algorithm. For example, the context vector algorithm includes the following expression:

[0126]

[0127] in:

[0128] C: context vector, which is the hidden state h of all time steps t The weighted sum of t Decision. The context vector contains the model’s comprehensive understanding of the input sequence;

[0129] αt: attention weight at time step t;

[0130] h t : The concatenated hidden state vector at the t-th time step;

[0131] T: Total number of time steps.

[0132] A04. Input the context vector into a preset fully connected layer and activation function to predict the total electron content data at the last moment, thereby obtaining the total electron content prediction data at the last moment;

[0133] Furthermore, the total electron content is predicted by the context vector, and the predicted value is output. Combined with the context vector C, the TEC value on the 7th day is predicted through the fully connected layer and activation function.

[0134] A05. Determine whether the deep learning model has converged based on the total electron content data at the last moment, the total electron content prediction data, and a preset loss function;

[0135] A06. If the deep learning model converges, the converged deep learning model is used as the electron content prediction model.

[0136] Finally, the model performance is evaluated, the model effect is judged by the loss function (such as RMSE), and the network structure or hyperparameters are adjusted. Specifically, based on the total electron content data at the last moment, the total electron content prediction data and the preset loss function, it is determined whether the deep learning model has converged. If it has converged, the training is considered complete, and the converged deep learning model is used as the electron content prediction model.

[0137] Furthermore, model updating and optimization involves regularly collecting the latest observation data, adding it to the training set, and updating the model parameters using incremental training. Model performance is evaluated, and the model effect is judged using a loss function (such as RMSE), and the network structure or hyperparameters are adjusted.

[0138] It should be noted that the present invention combines BiLSTM and GRU: BiLSTM can fully exploit the bidirectional time-dependent characteristics of ionospheric delay data, while GRU improves computational efficiency by simplifying parameters. The combination of the two significantly improves the model's ability to accurately capture ionospheric delays. The Attention mechanism is also utilized: the Attention mechanism gives higher weights to important features, allowing the model to focus on the input features that are most relevant to delay correction, reducing noise interference and thus improving prediction accuracy. Random forest feature screening is also utilized: by using random forest to screen key input features (such as the F10.7 index, Dst index, etc.), the interference of non-critical features on the model is avoided, and the effectiveness of the features and the accuracy of model predictions are improved.

[0139] Among them, the disadvantages of the traditional method are as follows:

[0140] 1. Traditional empirical models (such as the Klobuchar model and the NeQuick model) have the following disadvantages compared to BiLSTM-Attention:

[0141] (1) The correction accuracy is low and is only suitable for situations where the ionosphere changes relatively smoothly. The performance is significantly reduced under complex conditions such as geomagnetic storms.

[0142] (2) It is difficult to fully utilize the potential laws in large-scale data and cannot dynamically adapt to the real-time changing ionospheric characteristics.

[0143] 2. Physical models (such as the Global Ionosphere Map (GIM)) have the following disadvantages compared to BiLSTM-Attention:

[0144] (1) It relies on widely distributed observation station data, with long calculation and update cycles and poor real-time performance.

[0145] (2) It is not friendly enough to terminal devices with limited resources and cannot be directly applied to dynamic scenarios.

[0146] 3. Machine learning models (such as support vector machines (SVMs) and random forests (RFs)) have the following disadvantages compared to BiLSTM-Attention:

[0147] (1) The modeling capability of time series data is limited, and it is difficult to handle long-term dependence characteristics.

[0148] (2) It cannot effectively capture the complex dynamic characteristics of ionospheric delay and its accuracy is lower than that of deep learning models.

[0149] The present invention provides a training method for an electron content prediction model, the method comprising: obtaining historical observation data collected by the Beidou satellite navigation system, the historical observation data at least including total electron content data, geomagnetic activity index data, and solar activity index data of the ionosphere at each historical moment, the total electron content data being used to reflect ionospheric delay; performing feature extraction on the total electron content data, geomagnetic activity index data, and solar activity index data of the ionosphere using a random forest algorithm to obtain key feature data corresponding to the total electron content data; adopting a sliding window mechanism, taking the key feature data under a preset time window length as input time series data, and the total electron content data at the last moment of the time window length as the target output, to construct a training sample pair; performing prediction training on the total electron content data based on the training sample pair and a preset deep learning model to obtain a trained electron content prediction model. The above method realizes the training of the ionospheric electron content prediction model, and obtains the electron content prediction model, so that the complex nonlinear dynamic characteristics of the ionospheric delay due to solar activity, geomagnetic activity, and regional differences can be effectively captured by the deep learning model, thereby improving the prediction accuracy of the total electron content and the correction accuracy of the ionospheric delay.

[0150] See also Figure 2 , Figure 2 This is a flow chart of a method for using an electronic content prediction model in an embodiment of the present invention. The method can be applied to both a terminal and a server. This embodiment uses the server as an example. The terminal can be a desktop terminal or a mobile terminal. The mobile terminal can be at least one of a mobile phone, a tablet computer, a laptop computer, etc. The server can be implemented as an independent server or a server cluster consisting of multiple servers, such as Figure 2The method of use shown includes the following steps:

[0151] 201. Obtain target observation data within the current time window length collected by the Beidou satellite navigation system;

[0152] 202. Input the target observation data into an electron content prediction model to predict total electron content data, thereby obtaining target total electron content data at the current moment, wherein the electron content prediction model is trained based on a training method for an electron content prediction model provided by the present invention;

[0153] 203. Obtain ionospheric delay time using the target total electron content data and a preset ionospheric delay algorithm, where the ionospheric delay time is used to correct a positioning result of the Beidou satellite navigation system.

[0154] It should be noted that after completing the training of the electron content prediction model, the electron content prediction model can be used to predict the current electron content. Specifically, the target observation data of the current time window length collected by the Beidou satellite navigation system is obtained, and real-time observation data is collected from the Beidou satellite navigation system, including the total electron content (TEC) data of the ionosphere, geomagnetic activity index (such as Dst, AE, etc.) and solar activity index (such as F10.7, sunspot number R), pseudorange, carrier phase, satellite position, etc., to obtain the target observation data in the current time window. The last moment of the current time window is the current moment, so as to predict the total electron content of the ionosphere at the current moment.

[0155] Furthermore, the target observation data is input into the electron content prediction model to predict the total electron content data to obtain the target total electron content data at the current moment. Since the electron content prediction model is trained based on the electron content prediction model training method provided by the present invention, the target total electron content data (TEC value) at the current moment can be predicted, and then the target total electron content data and the preset ionospheric delay algorithm can be used to obtain the ionospheric delay time, wherein the ionospheric delay time is used to correct the positioning result of the Beidou satellite navigation system. Specifically, the current pseudorange, carrier phase, and satellite position can be corrected by the ionospheric delay time, and the predicted ΔT can be applied to the navigation signal to update the user positioning result and reduce the time error caused by the ionospheric delay time.

[0156] Exemplarily, the ionospheric delay algorithm includes the following mathematical expression:

[0157]

[0158] in:

[0159] ΔT: ionospheric delay time, reflecting the signal delay caused by the ionosphere (unit: seconds);

[0160] TEC: target total electron content data, that is, the predicted total electron content (unit: TECU, 1TECU = 10 16 electrons / m 2 );

[0161] f: signal frequency (unit: Hz).

[0162] The present invention provides a method for using an electron content prediction model, the method comprising: obtaining target observation data within a current time window length collected by a Beidou satellite navigation system; inputting the target observation data into an electron content prediction model to predict total electron content data, thereby obtaining target total electron content data at the current moment, wherein the electron content prediction model is trained based on a training method for an electron content prediction model provided by the present invention; and utilizing the target total electron content data and a preset ionospheric delay algorithm to obtain ionospheric delay time, wherein the ionospheric delay time is used to correct the positioning result of the Beidou satellite navigation system. In the above manner, the target total electron content data of the ionosphere at the current moment is predicted using the ionospheric electron content prediction model. Because the deep learning model has pre-learned the correspondence between the key features extracted from the total electron content data, geomagnetic activity index data, and solar activity index data of the ionosphere and the total electron content data, the electron content prediction model effectively captures the complex nonlinear dynamic characteristics of the ionospheric delay, thereby improving the prediction accuracy of the total electron content and the correction accuracy of the ionospheric delay.

[0163] In summary, this paper proposes a deep learning-based ionospheric delay correction method for the Beidou satellite navigation system, aiming to improve positioning accuracy. Methodologically, by collecting ionospheric total electron content (TEC) data and solar and geomagnetic activity index data, a deep learning model (BiLSTM-Attention) is used to perform nonlinear dynamic modeling and real-time prediction of ionospheric delay. The model effectively corrects ionospheric errors in Beidou positioning signals through time series feature extraction and multivariate feature analysis. Product-wise, this method can be embedded in Beidou navigation receivers or transmit ionospheric correction data to user terminals in real time via mobile applications, achieving high-precision positioning. Compared to traditional methods (such as the Klobuchar model and dual-frequency difference method), this method can maintain high accuracy even in complex environments (such as years with magnetic storms or in polar regions), reducing average positioning error by more than 10%, while significantly improving computational efficiency and meeting real-time requirements. This method can be widely applied in fields such as navigation, transportation, and geological exploration, and has significant social and economic value.

[0164] Product Solution: This module collects raw observation data from the Beidou Satellite System (BDS), including pseudoranges, carrier phases, satellite positions, and ionospheric TEC values. It connects to the Beidou receiver via a communication interface and transmits the collected data to the data preprocessing module. This module performs preprocessing on the collected data, including outlier data removal, missing data filling, and time synchronization. It uses the three-sigma method to remove outliers and bilinear interpolation to fill missing data. It connects to the data acquisition module and the deep learning computation module and transmits the processed data to the computation module. It uses the BiLSTM-GRU-Attention model to model and analyze the input data and perform ionospheric delay correction prediction. It connects to the data preprocessing module, receives the processed input data, and transmits the prediction results to the result output module. The ionospheric delay correction values ​​output by the deep learning computation module are fed back to the user terminal or navigation device. It connects to the deep learning computation module and the user terminal to output the correction data via a display or interface. The result output module receives positioning information for real-time navigation or scientific analysis.

[0165] See also Figure 3 , Figure 3 FIG. 1 is a structural block diagram of a training device for an electron content prediction model according to an embodiment of the present invention. Figure 3 The apparatus shown comprises:

[0166] A first acquisition module 301 is configured to acquire historical observation data collected by the BeiDou satellite navigation system, wherein the historical observation data includes at least total electron content data of the ionosphere, geomagnetic activity index data, and solar activity index data at various historical moments. The total electron content data is used to reflect ionospheric delay.

[0167] Feature extraction module 302: configured to perform feature extraction on the ionospheric total electron content data, geomagnetic activity index data, and solar activity index data using a random forest algorithm to obtain key feature data corresponding to the total electron content data;

[0168] Sample construction module 303: used to construct training sample pairs using a sliding window mechanism, taking key feature data within a preset time window length as input time series data and the total electron content data at the last moment of the time window length as target output;

[0169] Model training module 304: used to perform prediction training of total electron content data based on the training sample pairs and a preset deep learning model to obtain a trained electron content prediction model.

[0170] It should be noted that Figure 3 The contents of each module in the device shown are Figure 1The contents of each step in the method shown are similar, so we will not elaborate on them here to avoid repetition. For details, please refer to Figure 1 The content of each step in the method shown.

[0171] The present invention provides a training device for an electron content prediction model, which includes: a first acquisition module: used to acquire historical observation data collected by a Beidou satellite navigation system, wherein the historical observation data at least includes total electron content data, geomagnetic activity index data, and solar activity index data of the ionosphere at each historical moment, and the total electron content data is used to reflect ionospheric delay; a feature extraction module: used to perform feature extraction on the total electron content data, geomagnetic activity index data, and solar activity index data of the ionosphere through a random forest algorithm to obtain key feature data corresponding to the total electron content data; a sample construction module: used to adopt a sliding window mechanism, take the key feature data under a preset time window length as input time series data, and take the total electron content data at the last moment of the time window length as target output, to construct a training sample pair; and a model training module: used to perform prediction training of the total electron content data based on the training sample pair and a preset deep learning model to obtain a trained electron content prediction model. The above-mentioned training device is used to train the ionospheric electron content prediction model, and an electron content prediction model is obtained, so that the complex nonlinear dynamic characteristics of ionospheric delay due to solar activity, geomagnetic activity and regional differences can be effectively captured by the deep learning model, thereby improving the prediction accuracy of the total electron content and the correction accuracy of the ionospheric delay.

[0172] See also Figure 4 , Figure 4 FIG. 1 is a structural block diagram of a device for using an electron content prediction model according to an embodiment of the present invention. Figure 4 The apparatus shown comprises:

[0173] The second acquisition module 401 is used to obtain target observation data within the current time window length collected by the Beidou satellite navigation system;

[0174] Data prediction module 402: used to input the target observation data into the electron content prediction model to predict the total electron content data, and obtain the target total electron content data at the current moment, wherein the electron content prediction model is trained based on the electron content prediction model training method provided by the present invention;

[0175] The delay determination module 403 is configured to obtain an ionospheric delay time by using the target total electron content data and a preset ionospheric delay algorithm. The ionospheric delay time is used to correct the positioning result of the BeiDou satellite navigation system.

[0176] It should be noted that Figure 4The contents of each module in the device shown are Figure 2 The contents of each step in the method shown are similar, so we will not elaborate on them here to avoid repetition. For details, please refer to Figure 2 The content of each step in the method shown.

[0177] The present invention provides a device for using an electron content prediction model, the device comprising: a second acquisition module for acquiring target observation data within a current time window length collected by a Beidou satellite navigation system; a data prediction module for inputting the target observation data into the electron content prediction model to predict total electron content data, thereby obtaining target total electron content data at the current moment, wherein the electron content prediction model is trained based on a training method for an electron content prediction model provided by the present invention; and a delay determination module for using the target total electron content data and a preset ionospheric delay algorithm to obtain ionospheric delay time, wherein the ionospheric delay time is used to correct the positioning results of the Beidou satellite navigation system. The device uses the ionospheric electron content prediction model to predict the target total electron content data at the current moment. Because the deep learning model pre-learns the correspondence between key features extracted from the ionospheric total electron content data, geomagnetic activity index data, and solar activity index data and the total electron content data, the electron content prediction model effectively captures the complex nonlinear dynamic characteristics of ionospheric delay, improves the prediction accuracy of total electron content, and improves the correction accuracy of ionospheric delay.

[0178] Figure 5 FIG1 shows an internal structure diagram of a computer device in an embodiment. The computer device can be a terminal or a server. Figure 5 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system and may also store a computer program. When the computer program is executed by the processor, the processor can implement the above method. The internal memory may also store a computer program. When the computer program is executed by the processor, the processor can implement the above method. It will be understood by those skilled in the art that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0179] In one embodiment, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following Figure 1 or Figure 2 Steps shown.

[0180] In one embodiment, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the processor is caused to execute the following Figure 1 or Figure 2 Steps shown.

[0181] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0182] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0183] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A training method for an electron content prediction model, characterized in that: The method comprises: Acquiring historical observation data collected by the Beidou satellite navigation system, the historical observation data including at least total electron content data of the ionosphere, geomagnetic activity index data, and solar activity index data at various historical moments, the total electron content data being used to reflect ionospheric delay; Performing feature extraction on the total electron content data, geomagnetic activity index data, and solar activity index data of the ionosphere using a random forest algorithm to obtain key feature data corresponding to the total electron content data; A sliding window mechanism is used to construct training sample pairs by taking the key feature data within a preset time window length as the input time series data and the total electron content data at the last moment of the time window length as the target output; Prediction training of total electron content data is performed based on the training sample pairs and a preset deep learning model to obtain a trained electron content prediction model.

2. The method according to claim 1, characterized in that The obtaining of historical observation data collected by the BeiDou satellite navigation system further includes: The historical observation data are preprocessed to obtain preprocessed historical observation data, wherein the preprocessing at least includes outlier removal and missing value interpolation.

3. The method according to claim 1 or 2, characterized in that The method of performing prediction training on the total electron content data based on the training sample pairs and a preset deep learning model to obtain a trained electron content prediction model includes: Use the preset bidirectional long short-term memory network to process the input time series data to obtain the spliced ​​hidden state vector; Calculating the attention weight of the concatenated hidden state vector based on a preset attention mechanism; Generate a context vector using the attention weight, the concatenated hidden state vector, and a preset context vector algorithm; Inputting the context vector into a preset fully connected layer and activation function to predict the total electron content data at the last moment, thereby obtaining the total electron content prediction data at the last moment; Determining whether the deep learning model has converged based on the total electron content data at the last moment, the total electron content prediction data, and a preset loss function; If the deep learning model converges, the converged deep learning model is used as the electron content prediction model.

4. The method according to claim 3, characterized in that The bidirectional long short-term memory network includes at least a forward network calculation, and the forward network calculation includes the following expression: Where x t :Input time series data X=[x1,x2,...,x t ]’s eigenvector at time step t; The activation value of the input gate at time t; σ: Sigmoid activation function; W i f : weight matrix of input gate; Bias of the input gate; f t f : activation value of the forget gate; The activation value of the output gate; The cell state at time step t; tanh: hyperbolic tangent activation function; The weight matrix of the cell state; The forward hidden state at time step t; The forward hidden state at time step t-1.

5. The method according to claim 3, characterized in that: The attention weight includes the following expression: in: h t : The concatenated hidden state vector at the t-th time step; αt: attention weight at time step t; W α : attention weight matrix; T: total number of time steps; exp: exponential function; h k : The concatenated hidden state vector at the k-th time step.

6. The method according to claim 3, characterized in that: The context vector algorithm includes the following expression: in: C: context vector; αt: attention weight at time step t; h t : The concatenated hidden state vector at the t-th time step; T: Total number of time steps.

7. A method for using an electron content prediction model, characterized in that: The method comprises: Obtain target observation data within the current time window length collected by the Beidou satellite navigation system; Inputting the target observation data into an electron content prediction model to predict total electron content data to obtain target total electron content data at the current moment, wherein the electron content prediction model is trained based on the training method according to any one of claims 1 to 6; The target total electron content data and a preset ionospheric delay algorithm are used to obtain an ionospheric delay time, and the ionospheric delay time is used to correct the positioning result of the Beidou satellite navigation system.

8. The method according to claim 7, characterized in that: The ionospheric delay algorithm includes the following mathematical expression: in: ΔT: ionospheric delay time; TEC: target total electron content data; f: signal frequency.

9. A training device for an electron content prediction model, characterized in that: The device comprises: A first acquisition module is configured to acquire historical observation data collected by the BeiDou satellite navigation system, wherein the historical observation data includes at least total electron content data of the ionosphere, geomagnetic activity index data, and solar activity index data at each historical moment, wherein the total electron content data is used to reflect ionospheric delay; Feature extraction module: used to extract features from the total electron content data, geomagnetic activity index data and solar activity index data of the ionosphere using a random forest algorithm to obtain key feature data corresponding to the total electron content data; Sample construction module: used to construct training sample pairs using a sliding window mechanism, taking key feature data within a preset time window length as input time series data and the total electron content data at the last moment of the time window length as the target output; Model training module: used to perform prediction training of total electron content data based on the training sample pairs and the preset deep learning model to obtain a trained electron content prediction model.

10. A device for using an electron content prediction model, characterized in that: The device comprises: The second acquisition module is used to obtain the target observation data within the current time window length collected by the Beidou satellite navigation system; A data prediction module: configured to input the target observation data into the electron content prediction model to predict the total electron content data, thereby obtaining the target total electron content data at the current moment, wherein the electron content prediction model is trained based on the training method according to any one of claims 1 to 6; Delay determination module: used to use the target total electron content data and a preset ionospheric delay algorithm to obtain the ionospheric delay time, and the ionospheric delay time is used to correct the positioning result of the Beidou satellite navigation system.

Images