A multidimensional prediction method for enhanced satellite internet prediction

Abstract

The present invention belongs to the field of satellite telemetry technology, and specifically relates to a multidimensional prediction method suitable for enhancing satellite internet predictions. The method comprises the following steps: Step 1: selecting a remote sensing satellite and a period of time, acquiring remote sensing data from the satellite at regular intervals within this period of time as a data point, and dividing the acquired data points into a training set, a validation set, and a test set in chronological order; Step 2: normalizing the acquired satellite parameters using a minimum-maximum normalization formula; Step 3: calculating the Pearson correlation coefficient between the satellite parameters to measure the correlation between the satellite parameters; and Step 4: inputting the parameter combination into a two-layer LSTM model for model training. The two-layer LSTM model comprises a first layer containing 64 memory cells and a second layer being a hidden layer containing 64 memory cells. The present invention can significantly improve the accuracy of satellite telemetry data predictions.

Description

Technical Field

[0001] The present invention belongs to the field of satellite telemetry technology, and in particular relates to a multidimensional prediction method suitable for enhancing satellite Internet prediction. Background Art

[0002] Due to the increasing demand for satellite communications, navigation, remote sensing observation, etc. worldwide, the tasks that satellites need to undertake are becoming increasingly heavy. As the number of satellites in orbit, their design lifespan and model types continue to increase, increased monitoring of satellite status and performance is essential. Satellite telemetry parameters are important indicators for measuring satellite status. Through real-time monitoring and prediction of satellite parameters, we can understand the operating status of satellites, such as performance indicators, energy utilization, orbital deviation, etc., thereby realizing functions such as fault warning and prevention, operating status monitoring, resource management and planning, and performance optimization, further improving the stability of satellite operations and the efficiency of mission execution on satellites.

[0003] In the field of satellite telemetry, deep learning can be used to help predict satellite telemetry parameters. As a machine learning technology that can process large amounts of complex data, deep learning can automatically learn and optimize models. By learning from past historical data and current status, it can predict future data parameters and help the system make decisions.

[0004] However, the characteristics of deep learning also require a large amount of real data to help it pre-train. The performance of untrained models often fails to meet the requirements. At the same time, satellite telemetry data involves multiple sensors and multiple observation variables, and has the characteristics of multi-dimensionality and long sequences. Therefore, how to accurately predict satellite telemetry data remains a difficult problem. Summary of the Invention

[0005] The purpose of the present invention is to provide a multidimensional prediction method suitable for enhancing satellite Internet prediction, which can greatly improve the accuracy of satellite telemetry data prediction.

[0006] The technical solutions adopted by the present invention are as follows:

[0007] A multidimensional prediction method for enhancing satellite internet prediction includes the following steps:

[0008] Step 1: Select a remote sensing satellite and a period of time. During this period of time, obtain the satellite's remote sensing data at a certain interval as a data point. The obtained data points are divided into training set, validation set and test set in chronological order.

[0009] Step 2: Normalize the acquired satellite parameters using the minimum-maximum normalization formula. The minimum-maximum normalization formula is:

[0010]

[0011] where x norm Indicates that the normalized parameters are satellite parameters, min(x) indicates the minimum value of the satellite parameters, and max(x) indicates the maximum value of the satellite parameters.

[0012] Step 3: Calculate the Pearson correlation coefficient between satellite parameters to measure the correlation between satellite parameters. The calculation formula of the Pearson correlation coefficient is:

[0013]

[0014] where ρ X,Y represents the Pearson correlation coefficient between satellite parameters, cov(X,Y) represents the covariance between variable X and variable Y, σ X and σ Y Represents the standard deviation of variable X and variable Y respectively;

[0015] Step 4: Input the parameter combination into the two-layer LSTM model, and the LSTM model trains the model. The first layer of the two-layer LSTM model contains 64 memory units, and the second layer is a hidden layer, which contains 64 memory units.

[0016] Step 5: After the model training is completed, it is used to predict satellite remote sensing data.

[0017] Furthermore, the training process of the two-layer LSTM model in step 4 is as follows:

[0018] Step 401: Initialize the weight parameters, memory cell C0, and hidden state H0 of the LSTM model. The above parameters will be continuously updated during the training process.

[0019] Step 402: First, at each time step t, the LSTM model receives input X t and the hidden state H of the previous time step t-1 , then calculate the input gate I t 、Forget Gate F t , output gate O t and candidate memory cells The calculation formulas are:

[0020] I t =σ(X t W xi +H t-1 W hi +b i )

[0021] F t =σ(X tW xf +H t-1 W hf +b f )

[0022] O t =σ(X t W xo +H t-1 W ho +b o )

[0023]

[0024] Among them, W xi 、W xf 、W xo 、W hi 、W hf 、W ho 、b i 、b f 、b o 、W xc 、W hc is the weight parameter of the LSTM model.

[0025] Then, update the memory cell C t and hidden state H t , where the input gate I is used t and forget gate F t To update memory cells C t , using output gate O t and updated memory cells C t To calculate the hidden state H of the current time step t , and update the process as follows:

[0026]

[0027] H t =O t ⊙tanh(C t )

[0028] Among them, ⊙ represents element-wise multiplication;

[0029] Step 403: Hide the state H t Perform linear transformation and output the prediction results of the LSTM model;

[0030] Step 404: Calculate the loss between the prediction result of the current time step and the actual label using the mean absolute error and the mean square error;

[0031] Step 405: Calculate the gradient of each parameter through the back-propagation algorithm, and then use the Adam optimization algorithm to update the parameters of the LSTM model to minimize the loss function;

[0032] Step 406: Repeat the above steps to traverse all time steps and samples until the predetermined number of training iterations is reached or the convergence condition is met;

[0033] Step 407: During the training process, use the validation set to monitor the performance of the model and adjust the hyperparameters as needed to improve performance;

[0034] Step 408: After training is completed, the generalization ability and performance of the model are evaluated using the test set.

[0035] Furthermore, the modeling method for the LSTM model label generation problem in step 4 includes the following steps:

[0036] Step 4a: When the model predicts, the telemetry values ​​of the parameters at the previous 30 moments are used to predict the telemetry value of the next moment. The telemetry values ​​at the previous 30 moments are used as data, and the telemetry value at the 31st moment is used as the label.

[0037] Step 4b: Generate labels using a sliding window approach. The sliding window size is set to 30. The data within the sliding window is used as the input of the LSTM model, and the first data after the sliding window is used as the label of the data within the sliding window.

[0038] Step 4c: Whenever the sliding window completes the label generation operation, the sliding window moves backward by 1 bit;

[0039] Step 4d: Repeat step 4c until the data after the sliding window is the last bit of the data set, and the sliding window stops moving.

[0040] Furthermore, the formula for the mean absolute error in step 404 is:

[0041]

[0042] Where n is the number of samples, n is the number of samples, y i is the true value of the i-th sample, is the predicted value of the i-th sample.

[0043] Furthermore, the calculation formula of the mean square error in step 404 is:

[0044]

[0045] Where n is the number of samples, y i is the true value of the i-th sample, is the predicted value of the i-th sample.

[0046] Furthermore, the update formula of the Adam optimization algorithm in step 405 is:

[0047] m t =β1m t-1 +(1-β1)g t

[0048] v t =β2m t-1 +(1-β2)(g t ) 2

[0049] m in the formula t is the gradient first moment estimate, β1 and β2 are hyper parameters, v t is the second moment estimate, g t is the gradient at time step t; the Adam optimization algorithm also makes a deviation correction in the normal parameter update process. The deviation correction formula is:

[0050]

[0051] In the formula is the bias-corrected gradient first moment estimate, is the bias-corrected second moment estimate, β1 and β2 are hyperparameters;

[0052] The formula for the Adam algorithm to update the LSTM model parameters is:

[0053]

[0054] Where α is the learning rate of the neural network model, and ε is the constant to prevent zero error. is the bias-corrected gradient first moment estimate, is the bias-corrected second moment estimate.

[0055] The technical effects achieved by the present invention are:

[0056] The multidimensional prediction method for enhancing satellite Internet prediction of the present invention reduces errors and enhances prediction accuracy when processing large amounts of time series data through an effective LSTM training process, provides stronger generalization capability and efficiency for satellite Internet prediction, and greatly improves the accuracy of satellite telemetry data prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a flow chart of the present invention;

[0058] Figure 2 It is a structural diagram of the LSTM model of the present invention;

[0059] Figure 3 It is an architectural diagram of the double-layer LSTM model of the present invention;

[0060] Figure 4 is a graph of the Pearson coefficients of various parameter combinations of the present invention;

[0061] Figure 5 is the mean absolute error MAE graph of different parameter combinations of the present invention;

[0062] Figure 6 It is the mean square error MSE diagram of different parameter combinations of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is described in detail below with reference to the following examples. It should be understood that the following text is only used to describe one or more specific embodiments of the present invention and does not strictly limit the scope of protection of the present invention.

[0064] First, let's introduce the LSTM model. LSTM is a special recursive neural network consisting of an input gate, a forget gate, an output gate, and memory cells for recording additional information. LSTM considers time and nonlinear relationships while learning the rules between complex data, thereby achieving more stable prediction results. Satellite data prediction requires processing data with a large time span, so the LSTM model is well suited for predicting satellite telemetry data. Deep learning technology is now widely used in the aerospace field, and many models and algorithms have emerged in this field to solve practical application problems.

[0065] like Figure 1 As shown, a multidimensional prediction method suitable for enhancing satellite Internet prediction includes the following steps:

[0066] Step 1: Select a remote sensing satellite and a period of time. During this period of time, obtain the satellite's remote sensing data at regular intervals as a data point. The obtained data points are divided into a training set, a validation set, and a test set in chronological order.

[0067] Step 2: Normalize the acquired satellite parameters using the minimum-maximum normalization formula to prevent certain features from dominating the calculation process. The minimum-maximum normalization formula is:

[0068]

[0069] where x norm Indicates that the normalized parameters are satellite parameters, min(x) indicates the minimum value of the satellite parameters, and max(x) indicates the maximum value of the satellite parameters.

[0070] Step 3: Calculate the Pearson correlation coefficient between satellite parameters to measure the correlation between satellite parameters. The calculation formula of the Pearson correlation coefficient is as follows:

[0071]

[0072] where ρ X,Y represents the Pearson correlation coefficient between satellite parameters, cov(X,Y) represents the covariance between variable X and variable Y, σ X and σ Y represent the standard deviation of variables X and Y respectively.

[0073] Step 4: Input the parameter combination into the two-layer LSTM model. The LSTM model trains the model. The first layer of the two-layer LSTM model contains 64 memory units, and the second layer is a hidden layer, which contains 64 memory units.

[0074] The training process of the two-layer LSTM model in step 4 is as follows:

[0075] Step 401: Before starting training, the weight parameters of the LSTM model (such as W xi 、W xf 、W xo 、W hi 、W hf 、W ho 、b i 、b f 、b o 、W xc 、W hc 、b c etc.) and memory cell C0 and hidden state H0. The above parameters will be gradually updated during the training process.

[0076] Step 402: At each time step t, the LSTM model receives input X t and the hidden state H of the previous time step t-1 , then calculate the input gate I t 、Forget Gate F t , output gate O t and candidate memory cells The calculation formulas are as follows:

[0077] I t =σ(X t W xi +H t-1 W hi +b i )

[0078] F t =σ(Xt W xf +H t-1 W hf +b f )

[0079] O t =σ(X t W xo +H t-1 W ho +b o )

[0080]

[0081] Finally update memory cell C t and hidden state H t , where the input gate I t and forget gate F t To update memory cells C t , using output gate O t and updated memory cells C t To calculate the hidden state H of the current time step t , and its update process is modeled as follows, where ⊙ represents element-wise multiplication:

[0082]

[0083] H t =O t ⊙tanh(C t )

[0084] Step 403: Hide the state H t Perform linear transformation and output the prediction results of the LSTM model

[0085] Step 404: Calculate the loss between the prediction result of the current time step and the actual label using the mean absolute error (MAE) and mean square error (MSE).

[0086] Step 405: Calculate the gradient of each parameter through the back-propagation algorithm, and then use the Adam optimization algorithm to update the parameters of the LSTM model to minimize the loss function.

[0087] Step 406: Repeat the above steps to traverse all time steps and samples until a predetermined number of training iterations is reached or convergence conditions are met.

[0088] Step 407: During the training process, the validation set can be used to monitor the performance of the model and adjust the hyperparameters as needed to improve performance.

[0089] Step 408: After training is completed, the test set is used to evaluate the generalization ability and performance of the model.

[0090] Step 5: Once the model training is complete, it can be deployed in real-world applications for prediction of satellite remote sensing data.

[0091] The modeling method for the LSTM model label generation problem in step 4 includes the following steps:

[0092] Step 4a: When the model predicts, the telemetry values ​​of the parameters of the previous 30 moments are used to predict the telemetry value of the next moment. Therefore, during training, the telemetry values ​​of the first 30 moments are used as data, and the telemetry value of the 31st moment is used as the label.

[0093] The parameter telemetry value is the remote sensing parameter data at the current moment;

[0094] Step 4b: This technical solution uses a sliding window method to generate labels. The size of the sliding window is set to 30. The data in the sliding window is used as the input of the LSTM model, and the first data after the sliding window is used as the label of the data in the sliding window.

[0095] Step 4c: Whenever the sliding window completes the label generation operation, the sliding window moves backward by 1 bit;

[0096] Step 4d: Repeat step 4c until the data after the sliding window is the last bit of the data set, and the sliding window stops moving.

[0097] The method for calculating the mean absolute error in step 404 is to average the absolute differences of each sample. The smaller the value obtained, the better the prediction ability of the model.

[0098] The formula is as follows:

[0099]

[0100] Where n is the number of samples, n is the number of samples, y i is the true value of the i-th sample, is the predicted value of the i-th sample.

[0101] The method for calculating the mean square error in step 404 is to first add up the squares of the differences between the predicted value and the true value of each sample, and then calculate the average value. The smaller the final value, the better the prediction ability of the model.

[0102] The formula is as follows:

[0103]

[0104] Where n is the number of samples, y i is the true value of the i-th sample, is the predicted value of the i-th sample;

[0105] After calculating the loss function to get the difference between the predicted value and the true value, this technical solution uses the stochastic gradient descent method to back-propagate and train our deep neural network.

[0106] Since random sampling of samples by stochastic gradient descent may lead to instability in the training process, this technical solution combines the optimization algorithm Adam. Adam is an optimization algorithm for adaptive learning rate. Its basic idea is to combine the advantages of adaptive gradient (AdaGrad) and RMSProp. The updated step size is obtained by estimating the first moment of the gradient (the average value of the gradient) and the second moment (the variance of the centerless gradient). This algorithm can not only ensure the stability of parameter updates, but also adaptively adjust the learning rate.

[0107] The update formula of the Adam optimization algorithm in step 405 is:

[0108] m t =β1m t-1 +(1-β1)g t

[0109] v t =β2m t-1 +(1-β2)(g t ) 2

[0110] m in the formula t is the gradient first moment estimate, β1 and β2 are hyper parameters, v t is the second moment estimate, g t is the gradient at time step t. The Adam optimization algorithm also makes the following deviation corrections during the normal parameter update process:

[0111]

[0112] In the formula is the bias-corrected gradient first moment estimate, is the bias-corrected second moment estimate, and β1 and β2 are hyperparameters.

[0113] The formula for the Adam algorithm to update the LSTM model parameters is as follows:

[0114]

[0115] Where α is the learning rate of the neural network model, and ε is the constant to prevent zero error. is the bias-corrected gradient first moment estimate, is the bias-corrected second moment estimate.

[0116] Figure 3 and Figure 4 These are the mean absolute error (MAE) and mean square error (MSE) obtained by inputting three different parameter combinations into the LSTM model. These three sets of experiments show that combining highly correlated parameters as model input results in lower prediction errors, demonstrating the effectiveness of this method.

[0117] In summary, this technical solution reduces errors and enhances prediction accuracy when processing large amounts of time series data through an effective LSTM training process, providing stronger generalization capabilities and efficiency for satellite Internet predictions and greatly improving the accuracy of satellite telemetry data predictions.

[0118] The foregoing is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained herein shall, unless otherwise specified or limited, be implemented in accordance with conventional means in the art.

Claims

1. A multi-dimensional prediction method suitable for enhancing satellite Internet prediction, characterized by: The following steps are involved: Step 1: Select a remote sensing satellite and a period of time. During this period of time, obtain the satellite's remote sensing data at a certain interval as a data point. The obtained data points are divided into training set, validation set and test set in chronological order. Step 2: Normalize the acquired satellite parameters using the minimum-maximum normalization formula. The minimum-maximum normalization formula is: Among them, x norm Indicates that the normalized values ​​are satellite parameters, min(x) indicates the minimum value of the satellite parameters, and max(x) indicates the maximum value of the satellite parameters; Step 3: Calculate the Pearson correlation coefficient between satellite parameters to measure the correlation between satellite parameters. The calculation formula of the Pearson correlation coefficient is: where ρ X,Y represents the Pearson correlation coefficient between satellite parameters, cov(X,Y) represents the covariance between variable X and variable Y, σ X and σ Y Represents the standard deviation of variable X and variable Y respectively; Step 4: Input the parameter combination into the two-layer LSTM model, and the LSTM model trains the model. The first layer of the two-layer LSTM model contains 64 memory units, and the second layer is a hidden layer, which contains 64 memory units. Step 5: After the model training is completed, it is used to predict satellite remote sensing data; The modeling method for the LSTM model label generation problem in step 4 includes the following steps: Step 4a: When the model predicts, the telemetry values ​​of the parameters at the previous 30 moments are used to predict the telemetry value of the next moment. The telemetry values ​​at the previous 30 moments are used as data, and the telemetry value at the 31st moment is used as the label. Step 4b: Generate labels using a sliding window approach. The sliding window size is set to 30. The data within the sliding window is used as the input of the LSTM model, and the first data after the sliding window is used as the label of the data within the sliding window. Step 4c: Whenever the sliding window completes the label generation operation, the sliding window moves backward by 1 bit; Step 4d: Repeat step 4c until the data after the sliding window is the last bit of the data set, and the sliding window stops moving.

2. The multidimensional prediction method for enhancing satellite internet prediction according to claim 1, characterized in that: The training process of the two-layer LSTM model in step 4 is as follows: Step 401: Initialize the weight parameters, memory cell C0, and hidden state H0 of the LSTM model. The above parameters will be continuously updated during the training process. Step 402: First, at each time step t, the LSTM model receives input X t and the hidden state H of the previous time step t-1 , then calculate the input gate I t 、Forget Gate F t , output gate O t and candidate memory cells The value of Then, update the memory cell C t and hidden state H t , where the input gate I is used t and forget gate F t To update memory cells C t , using output gate O t and updated memory cells C t To calculate the hidden state H of the current time step t ; Step 403: Hide the state H t Perform linear transformation and output the prediction results of the LSTM model; Step 404: Calculate the loss between the prediction result of the current time step and the actual label using the mean absolute error and the mean square error; Step 405: Calculate the gradient of each parameter through the back-propagation algorithm, and then use the Adam optimization algorithm to update the parameters of the LSTM model to minimize the loss function; Step 406: Repeat the above steps to traverse all time steps and samples until the predetermined number of training iterations is reached or the convergence condition is met; Step 407: During the training process, use the validation set to monitor the performance of the model and adjust the hyperparameters as needed to improve performance; Step 408: After training is completed, the generalization ability and performance of the model are evaluated using the test set.

3. The multidimensional prediction method for enhancing satellite internet prediction according to claim 2, characterized in that: In step 402, the input gate I t 、Forget Gate F t , output gate O t and candidate memory cells The calculation formulas are: I t =σ(X t W xi +H t-1 W hi +b i ) F t =σ(X t W xf +H t-1 W hf +b f ) O t =σ(X t W xo +H t-1 W ho +b o ) Among them, W xi 、W xf 、W xo 、W hi 、W hf 、W ho 、b i 、b f 、b o 、W xc 、W hc is the weight parameter of the LSTM model.

4. The multidimensional prediction method for enhancing satellite internet prediction according to claim 3, characterized in that: In step 402, the memory cell C t and hidden state H t The update process of is modeled as: H t =O t ⊙tanh(C t ) Here, ⊙ represents element-wise multiplication.

5. The multidimensional prediction method for enhancing satellite internet prediction according to claim 4, characterized in that: The formula for the mean absolute error in step 404 is: Where n is the number of samples, y i is the true value of the i-th sample, is the predicted value of the i-th sample.

6. The multi-dimensional prediction method for enhancing satellite internet prediction according to claim 5, characterized in that: The calculation formula of the mean square error in step 404 is: Where n is the number of samples, y i is the true value of the i-th sample, is the predicted value of the i-th sample.

7. The multi-dimensional prediction method for enhancing satellite internet prediction according to claim 6, characterized in that: The update formula of the Adam optimization algorithm in step 405 is: m t =β1m t-1 +(1-β1)g t v t =β2m t-1 +(1-β2)(g t ) 2 m in the formula t is the gradient first moment estimate, β1 and β2 are hyper parameters, v t is the second moment estimate, g t is the gradient at time step t; In the normal process of updating parameters, the Adam optimization algorithm also makes deviation corrections. The deviation correction formula is: In the formula is the bias-corrected gradient first moment estimate, is the bias-corrected second moment estimate, β1 and β2 are hyperparameters; The formula for the Adam algorithm to update the LSTM model parameters is: Where α is the learning rate of the neural network model, ε is the constant to prevent zero error, is the bias-corrected gradient first moment estimate, is the bias-corrected second moment estimate.

Images