Distribution drift satellite telemetry data prediction method based on sliding window design

Abstract

The invention discloses a distributed drift satellite telemetry data prediction method based on sliding window design, and belongs to the technical field of time series data prediction. According to the method, the problems that the sliding window span is not determined from the time sequence characteristics of the time sequence data and the time sequence data has distribution offset in the existing method, so that the data prediction accuracy is poor are solved. According to the Shannon sampling theorem, the method for designing the sliding window span based on the Shaplet features of the time sequence data is designed, the time sequence features of the data can be effectively extracted, and the prediction precision is improved. According to the data normalization method designed by the invention, only the data distribution information of the current sliding window and the previous m steps of sliding windows needs to be considered, so that the problem of overall data interference is effectively solved, more data distribution information can be comprehensively considered, the problem of time sequence data distribution drift is effectively solved, and the prediction effect is enhanced. The method can be applied to time series data prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of time series data prediction, and particularly relates to a distribution drift satellite telemetry data prediction method based on a sliding window design. Background Art

[0002] Time series data prediction refers to the process of predicting future data points by analyzing the time series characteristics in historical data using methods such as statistics, machine learning, or deep learning. Time series data refers to ordered observational data recorded over time, such as temperature changes. The goal of prediction is to make a reasonable estimate of the data value at a certain future moment or within a certain period based on the characteristics and trends of past data.

[0003] Time series data prediction based on a deep neural network (DNN) is a method of using a deep learning model to predict the future values of time series data. This method constructs a deep neural network model to automatically learn the complex patterns and characteristics in time series data, thereby achieving more accurate prediction. This method first needs to preprocess the data and then divide the data set in order to better utilize the deep neural network for training. Setting a sliding window for time series data and then moving the sliding window according to the sliding step to extract data is a common and effective method for constructing a data set. However, there is currently no systematic and effective method for selecting the span of the sliding window. Existing research does not target the division of the sliding window by considering the time series characteristics of the time series data itself. Such blind division may consume a large amount of time and effort, and even lead to distorted prediction data. Statistical characteristics such as the mean and variance in time series data usually change over time, which leads to the problem of distribution shift. This change in time distribution seriously affects the accuracy of data prediction. Summary of the Invention

[0004] The purpose of the present invention is to solve the problem of poor data prediction accuracy caused by the fact that the existing method does not determine the span of the sliding window based on the time series characteristics of the time series data itself and the distribution shift of the time series data, and to propose a distribution drift satellite telemetry data prediction method based on a sliding window design.

[0005] The technical solution adopted by the present invention to solve the above technical problems is: a distribution drift satellite telemetry data prediction method based on a sliding window design, and the method specifically includes the following steps:

[0006] Step 1: Obtain the historical time series temperature data of the satellite thruster fuel nozzle, and preprocess the obtained data to obtain the preprocessed time series temperature data;

[0007] Step 2: Obtain the period T of the preprocessed time series temperature data;

[0008] Step 3. Determine the window length \(T\) according to the period \(T\). x , according to the window length \(T\) x and the preprocessed time series temperature data, train the prediction model to obtain a trained prediction model;

[0009] Step 4. Use the trained prediction model to predict future time series temperature data.

[0010] Furthermore, the preprocessing method is to remove missing values and outliers.

[0011] Furthermore, the specific process of Step 2 is as follows:

[0012] Perform a fast Fourier transform on the preprocessed time series temperature data, and obtain the period \(T\) of the time series temperature data according to the transformation result.

[0013] Furthermore, the prediction model is a time domain convolutional network TCN.

[0014] Furthermore, the window length \(T\) x is between \(2T\) and \(3T\).

[0015] Furthermore, in Step 3, according to the window length \(T\) x and the preprocessed time series temperature data, train the prediction model, and the specific process is as follows:

[0016] Step 3-1. For the preprocessed time series temperature data \(x\), use a sliding window to divide the time series temperature data \(x\) to obtain a training data set, and denote the \(k\)-th group of data in the training data set as \(x\) k , and the length of each group of data is \(T\) x ;

[0017] Step 3-2. Initialize the number of data groups \(k = 1\);

[0018] Step 3-3. Perform a normalization operation on the \(k\)-th group of data:

[0019]

[0020] where and are the affine parameters of the \(k\)-th group of data, \(x\) kt represents the \(t\)-th data in the \(k\)-th group, \(E\) t [\(x\) kt represents the mean of all data in the \(k\)-th group, \(Var[x\) kt represents the standard deviation of all data in the \(k\)-th group, \(\epsilon\) is a constant, represents the \(t\)-th data in the \(k\)-th group after the normalization operation;

[0021] Step 3-4: Use the k-th group of data after normalization as the input of the Temporal Convolutional Network (TCN), and adopt the parameters γ k and β k to perform inverse normalization on the data predicted and output by the Temporal Convolutional Network (TCN), and calculate the loss function based on the actual data at the prediction moment of the Temporal Convolutional Network (TCN) and the result of the inverse normalization operation;

[0022] Adjust the parameters of the Temporal Convolutional Network (TCN) backward according to the loss function;

[0023] Step 3-5: Let the number of data groups k = k + 1, and return to execute Step 3-3.

[0024] Furthermore, the mean E t [x kt and the standard deviation Var[x kt are respectively:

[0025]

[0026] Furthermore, the calculation methods of the parameters γ k and β k are:

[0027]

[0028] where ω 0 , ω 1 , …, ω m are weighting coefficients.

[0029] Furthermore, the process of performing inverse normalization on the data predicted and output by the Temporal Convolutional Network (TCN) by adopting the parameters γ k and β k is specifically as follows:

[0030] Input the k-th group of data after normalization into the Temporal Convolutional Network (TCN), and denote the predicted output of the Temporal Convolutional Network (TCN) as Then adopt the parameters γ k and β k to perform inverse normalization on :

[0031]

[0032] where represents the result of the inverse normalization operation on .

[0033] Furthermore, the specific process of Step 4 is:

[0034] Step 4-1: Collect the historical T of the satellite thruster fuel nozzlex The temperature data at each moment is pre - processed, and the pre - processed time - series temperature data is used as the first group of data;

[0035] Step Four Two: Initialize the number of groups \(k' = 1\);

[0036] Step Four Three: Perform normalization on the \(k'\) - th group of data:

[0037]

[0038] Among them, and are the affine parameters of the \(k'\) - th group of data, \(x'\) k′t represents the \(t\) - th data in the \(k'\) - th group, \(E\) t [x'\) k′t represents the mean of all data in the \(k'\) - th group, \(Var[x'\) k′t represents the standard deviation of all data in the \(k'\) - th group, \(\epsilon\) is a constant, represents the \(t\) - th data in the \(k'\) - th group after normalization;

[0039] According to and to calculate \(\gamma\) k′ and \(\beta\) k′ ;

[0040]

[0041] Step Four Four: Use the \(k'\) - th group of data after normalization as the input of the trained temporal convolutional network TCN, and perform inverse normalization on the predicted output k′ and \(\beta\) k′ of the trained temporal convolutional network TCN; Perform inverse normalization operation;

[0042]

[0043] Among them, represents the result of the inverse normalization operation, that is, the final prediction result of the temporal convolutional network TCN;

[0044] Take and the temperature data at the previous \(T\) x - 1 moment as the \((k'+1)\) - th window data;

[0045] Step Four Five: Let \(k'=k'+1\), and return to execute Step Four Three.

[0046] The beneficial effects of the present invention are:

[0047] The present invention designs a method for dividing a sliding window based on the characteristics of time-series data according to the Shannon sampling theorem. Compared with the traditional sliding window design method, the present invention uses the characteristics of Shapelet of time-series data to design the sliding window, which can effectively extract the time-series characteristics of the data and improve the prediction accuracy. Through the data normalization method designed by the present invention, only the data distribution information of the current sliding window and the previous m-step sliding windows needs to be considered, which not only effectively solves the problem of overall data interference, but also comprehensively considers more data distribution information, effectively solves the problem of time-series data distribution drift, and enhances the prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a schematic diagram of a sliding window;

[0049] Figure 2 is the preprocessed time-series data;

[0050] 1e6 represents 10 to the power of 6;

[0051] Figure 3 is a frequency domain graph;

[0052] Figure 4 is a change curve graph of the loss function loss during training;

[0053] Figure 5 is the coefficient of determination R 2 during training;

[0054] Figure 6 is a prediction effect graph;

[0055] Figure 7 is a prediction error graph;

[0056] Figure 8 is a comparison graph of the loss function loss;

[0057] Figure 9 is the coefficient of determination R 2 comparison graph. DETAILED DESCRIPTION OF THE INVENTION

[0058] DETAILED DESCRIPTION OF THE INVENTION I: A method for predicting satellite telemetry data with distribution drift based on sliding window design according to this embodiment specifically includes the following steps:

[0059] Step 1: Obtain the historical time-series temperature data of the satellite thruster fuel nozzle, and preprocess the obtained data to obtain the preprocessed time-series temperature data;

[0060] Step 2: Obtain the period T of the preprocessed time-series temperature data;

[0061] Step 3: Determine the window length T according to the period Tx , according to the window length T x and the preprocessed time-series temperature data, train the prediction model to obtain a trained prediction model;

[0062] Step 4: Use the trained prediction model to predict future time-series temperature data.

[0063] The present invention designs a method for designing a sliding window according to the local pattern (Shapelet) of time-series data. Shapelet usually only focuses on a local segment of the time series rather than the entire sequence, so it can effectively capture local features without learning the entire time series. In addition, Shapelet is a pattern directly extracted from the time series and usually has strong interpretability, which is very suitable for constructing a sliding window. To solve the influence of the time-series data distribution shift problem on the prediction effect, the present invention designs a Reversible Learnable Affine Normalization with Previous Step (RLAN-PS) method that refers to the previous-step data. The method of the present invention performs normalization with learnable affine transformation parameters within each sliding window input during model training, and the affine transformation parameters of the kth step refer to the currently learned parameters and the affine transformation parameters of the previous m steps (γ k-1 , β k-1 ), (γ k-2 , β k-2 )... (γ k-m , β k-m ) are calculated by weighted calculation, removing the non-stationary information of the original data and considering the distribution changes within the nearby m-step data set. An inverse normalization operation is designed during output to restore the non-stationary information removed from the input data to the output. This method can reduce the influence of distribution changes on the model prediction effect.

[0064] Specific Embodiment 2: The difference between this embodiment and Specific Embodiment 1 is that the preprocessing method is to remove missing values and outliers.

[0065] Other steps and parameters are the same as those in Specific Embodiment 1.

[0066] Specific Embodiment 3: The difference between this embodiment and Specific Embodiment 1 or 2 is that the specific process of Step 2 is as follows:

[0067] Perform a Fast Fourier Transform (FFT) on the preprocessed time-series temperature data, and obtain the period T of the time-series temperature data according to the transformation result.

[0068] Other steps and parameters are the same as those in the first or second specific implementation manner.

[0069] The Fourier transform is a mathematical tool used to transform a signal from the time domain to the frequency domain. According to the characteristics of the frequency domain graph, it is very convenient to obtain the data period. However, the mathematical Fourier transform has a huge amount of calculation and extremely low efficiency. The present invention uses FFT to greatly improve the calculation efficiency of the Fourier transform, making the signal processing task feasible.

[0070] Specific implementation manner four: The difference between this implementation manner and one of the first to third specific implementation manners is that the prediction model is a temporal convolutional network TCN (Temporal Convolutional Network).

[0071] Other steps and parameters are the same as those in one of the first to third specific implementation manners.

[0072] Specific implementation manner five: The difference between this implementation manner and one of the first to fourth specific implementation manners is that the window length T x is between 2T and 3T.

[0073] Other steps and parameters are the same as those in one of the first to fourth specific implementation manners.

[0074] According to the Shannon Sampling Theorem, for a signal with a limited bandwidth, if the highest frequency of the signal is f max , then it is necessary to sample it at a sampling frequency of at least 2f max to maximize the extraction of the temporal characteristics of the data, helping the model understand how the signal changes over time, so as to maintain high accuracy during prediction. If the sampling frequency is lower than this value, it will cause signal reconstruction distortion, that is, the aliasing phenomenon occurs. The present invention takes 2 to 3 times the period as the length of the sliding window, which is equivalent to simulating the high-frequency sampling process in signal processing. Through the long window, 2 to 3 times the sample points can be captured in each iteration, which is equivalent to "sampling" the signal at a higher frequency. This method enables the model to better understand the frequency characteristics of the signal and capture the temporal characteristics.

[0075] Specific implementation manner six: The difference between this implementation manner and one of the first to fifth specific implementation manners is that in step three, the prediction model is trained according to the window length T x and the preprocessed temporal temperature data. The specific process is as follows:

[0076] Step 3-1: For the preprocessed temporal temperature data x, use a sliding window to divide the temporal temperature data x to obtain a training data set. Denote the kth group of data in the training data set as x k, the length of each group of data is T x ;

[0077] As Figure 1 shown, the partitioning method is as follows: The data from the 1st moment to the T-th moment in the time-series temperature data x is used as the data within the first window. After sliding the sliding window forward by a fixed step size, the sliding window reaches a new position, obtaining the data within the second window, and so on. Multiple window data can be obtained. The multiple window data obtained is used as the training dataset. The predicted label for each window data is the temperature data at the first moment after the window. In the present invention, each window data is referred to as a group of data; x

[0078] Step 3-2: Initialize the number of data groups k = 1;

[0079] Step 3-3: Perform a normalization operation on the k-th group of data:

[0080]

[0081] Among them, and are the affine parameters of the k-th group of data, x kt represents the t-th data in the k-th group, E t [x kt represents the mean of all data in the k-th group, Var[x kt [ represents the standard deviation of all data in the k-th group, ε is a constant, represents the t-th data in the k-th group after the normalization operation;

[0082] and The determination method of and

[0083] is as follows: Calculate the mean and standard deviation of the k-th group of data after the normalization operation, and combine other indicators to determine whether the normalized data is affected by the drift of the original data distribution. Select the k and β k for which the normalized data is not affected by the drift of the original data distribution

[0084] Take the k-th group of data after the normalization operation as the input of the temporal convolutional network TCN, and use the parameters γ

[0085] Step 3-5: Let the number of data groups k = k + 1, and return to execute Step 3-3.​

[0086] The other steps and parameters are the same as those in any one of the first to fifth specific embodiments.

[0087] To ensure the data volume of the training dataset, the present invention can collect the temperature data of the same measurement point during each working process of the satellite thruster fuel nozzle. Under normal working conditions, the distribution laws of the temperature data at the same measurement point are similar. The temperature data collected at the same measurement point can be used to train the same model. For the temperature data at each other measurement point, a prediction model can be established for the data of each measurement point respectively, and then the corresponding prediction model can be trained using the temperature data collected at the corresponding measurement point to avoid the influence of the temperature distribution differences at different measurement points on the temperature prediction. Using the trained prediction model corresponding to each measurement point, the temperature data of the satellite thruster fuel nozzle at each measurement point can be predicted.

[0088] Through the normalization operation, the distribution of the current group of data can be well described, the problem of data distribution drift can be effectively avoided, and at the same time, the previous-step information is comprehensively considered, enriching the composition of the data distribution and being able to effectively extract more data features.

[0089] Specific embodiment seven: The difference between this embodiment and any one of the first to sixth specific embodiments is that the mean E t [x kt and the standard deviation Var[x kt are respectively:

[0090]

[0091] The other steps and parameters are the same as those in any one of the first to sixth specific embodiments.

[0092] Specific embodiment eight: The difference between this embodiment and any one of the first to seventh specific embodiments is that the calculation methods of the parameters γ k and β k are:

[0093]

[0094] Among them, ω 0 , ω 1 , …, ω m are weighting coefficients.

[0095] The other steps and parameters are the same as those in any one of the first to seventh specific embodiments.

[0096] If the number of windows before the current window is less than m, only the parameters of all the windows before the current window need to be weighted, and the weight coefficients can be set according to the actual situation.

[0097] Specific Embodiment Nine: The difference between this embodiment and any one of Embodiments One to Eight is that the parameter γ k and β k are used to perform inverse normalization on the data predicted and output by the Temporal Convolutional Network (TCN). The specific process is as follows:

[0098] Input the data after the k-th group of normalization operations into the Temporal Convolutional Network (TCN), and denote the predicted output of the Temporal Convolutional Network (TCN) as Then, use the parameter γ k and β k to perform inverse normalization on (The purpose is to restore the non-stationary information removed from the input data to the predicted output result):

[0099]

[0100] where, represents the result of the inverse normalization operation on .

[0101] Other steps and parameters are the same as any one of Embodiments One to Eight.

[0102] E t [x kt , Var[x kt , γ k and β k are all parameters stored during the normalization operation.

[0103] Specific Embodiment Ten: The difference between this embodiment and any one of Embodiments One to Nine is that the specific process of Step Four is as follows:

[0104] Step Four One: Collect the temperature data at the historical T x moments of the satellite thruster fuel nozzle (taking the current moment as the reference, collect the temperature data at the current moment and the previous T x -1 moments). After preprocessing the collected data, use the preprocessed time-series temperature data as the first group of data;

[0105] Step Four Two: Initialize the group number k' = 1;

[0106] Step Four Three: Perform normalization on the k'-th group of data:

[0107]

[0108] where, and are the affine parameters of the k'-th group of data, and x' k′t represents the t-th data in the k'-th group, and Et [x′ k′t represents the mean of all data in the k'-th group, and Var[x′ k′t represents the standard deviation of all data in the k'-th group. ε is a constant, represents the t-th data in the k'-th group after the normalization operation;

[0109] and The determination methods of and

[0110] are as follows: Calculate the mean and standard deviation of the data in the k'-th group after the normalization operation, and combine other indicators to determine whether the normalized data is affected by the drift of the original data distribution. Select the and where the normalized data is not affected by the drift of the original data distribution. Calculate γ k′ and β k′ ;

[0111]

[0112] Step Four: Use the data in the k'-th group after the normalization operation as the input of the trained Time-domain Convolutional Network (TCN). According to γ k′ and β k′ perform inverse normalization on the predicted output of the trained Time-domain Convolutional Network (TCN);

[0113]

[0114] Among them, represents the result of the inverse normalization operation, that is, the final predicted result of the Time-domain Convolutional Network (TCN);

[0115] Use and the temperature data at the previous T x -1 moment as the (k'+1)-th window data (for the first-step prediction, the data in the second window includes the predicted data at 1 moment and the actual data at T x -1 moments. And so on, the data in the third window includes the predicted data at 2 moments and the actual data at T x -2 moments, until the set number of prediction steps is reached and then stop);

[0116] Step Five: Let k' = k' + 1, and return to execute Step Four.

[0117] Other steps and parameters are the same as those in any one of the specific embodiments one to nine.

[0118] Experimental part

[0119] The temperature of the fuel nozzle of the satellite thruster is a set of typical time-series data. Due to the influence of the periodic working mode and external conditions (such as solar radiation, earth shadow, etc.), this time-series data has periodicity to varying degrees. First, the data is preprocessed to remove missing values and outliers. The preprocessed time-series data is as shown in Figure 2 . It can be seen that the time-series characteristics of this data are very complex and have the characteristics of multiple periods. For this characteristic, FFT analysis is carried out, and the obtained frequency-domain diagram is as shown in Figure 3 . It can be seen that the maximum frequency f of this time-series data is max = 0.000094, and the minimum period T is min = 10638. The minimum period is used as the local mode (Shapelet) of the time-series data.

[0120] Take 2 to 3 times the Shapelet length of the time-series data as the length of the sliding window. The sliding window length is set to 30000 (2.82 times the Shapelet length), and a data set is constructed according to this window for training.

[0121] Before training, the data set is subjected to learnable affine normalization with reference to the previous-step data, and inverse normalization is performed after training. Build a classic time-domain convolutional network TCN as a time-series prediction model for training, and record the loss function loss and the coefficient of determination (Coefficient of Determination, R 2 ) during training. The loss function loss used in the simulation training is the mean square error MSE between the predicted value and the actual value, which represents the accuracy of the model training. The closer the MSE is to 0, the better the model training effect; the coefficient of determination R 2 represents the fitting degree of the model training result. The closer R 2 is to 1, the better the model training effect. The recorded loss function loss and the coefficient of determination are as shown in Figure 4 and Figure 5 respectively. According to Figure 4 and Figure 5 , it can be concluded that the model training effect is very good.

[0122] After the model training is completed, the final training results are summarized in Table 1:

[0123] Table 1 Training Results

[0124]

[0125] By analyzing the results in Table 1, it can be concluded that the effects of the training set and the validation set are very good. The obtained prediction results are as shown in Figure 6 . The prediction fitting degree R 2= 0.9915. It can be found from the prediction results that the fitting degree of the prediction results is good, and the prediction error MSE is as Figure 7 shown. The overall mean square error MSE of the prediction is 0.1798. It can be seen from the above results that the prediction accuracy is very high.

[0126] To verify the advantages of the method of the present invention over the traditional method, the sliding window length is set to 110638 (1 times the Shapelet length) and retrained. The comparison of the loss functions loss of the two trainings and the coefficient of determination R 2 comparisons are respectively as Figure 8 and Figure 9 shown. According to Figure 8 and Figure 9 , it can be concluded that when the sliding window length is 2.8 times the Shapelet length, the model training effect is significantly better than when the sliding window length is 1 times the Shapelet length, thus proving the superiority of the method of the present invention by comparison.

[0127] The above examples of the present invention are only for explaining in detail the calculation model and calculation process of the present invention, and are not intended to limit the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made on the basis of the above description. It is impossible to list all the implementation manners here. Any obvious changes or modifications derived from the technical solutions of the present invention still fall within the protection scope of the present invention.

Claims

1. A distributed drift satellite telemetry data prediction method based on sliding window design, characterized in that: The method specifically comprises the following steps: Step 1: Acquire historical time series temperature data of a satellite thruster fuel nozzle, and preprocess the acquired data to obtain preprocessed time series temperature data; Step 2, obtaining the period T of the preprocessed time series temperature data; Step 3: Determine the window length T based on the period T x , according to the window length T x The prediction model is trained using the preprocessed time series temperature data to obtain a trained prediction model; Step 4: Use the trained prediction model to predict future time series temperature data.

2. According to the method of claim 1, the distributed drift satellite telemetry data prediction method based on sliding window design is characterized in that: The preprocessing method is to remove missing values ​​and outliers.

3. The distributed drift satellite telemetry data prediction method based on sliding window design according to claim 1 is characterized in that: The specific process of step 2 is as follows: The preprocessed time series temperature data is subjected to a fast Fourier transform, and the period T of the time series temperature data is obtained according to the transform result.

4. The distributed drift satellite telemetry data prediction method based on sliding window design according to claim 1 is characterized in that: The prediction model is a time domain convolutional network (TCN).

5. The distributed drift satellite telemetry data prediction method based on sliding window design according to claim 1 is characterized in that: The window length T x Between 2T and 3T.

6. The distributed drift satellite telemetry data prediction method based on sliding window design according to claim 1 is characterized in that: In the step 3, according to the window length T x The prediction model is trained with the preprocessed time series temperature data. The specific process is as follows: Step 3: For the preprocessed time series temperature data x, use the sliding window to divide the time series temperature data x to obtain a training data set, and record the kth group of data in the training data set as x k , the length of each set of data is T x ; Step 32, initialize the number of data groups k = 1; Step 3: Normalize the kth group of data: in, and is the affine parameter of the kth group of data, x kt represents the tth data in the kth group, E t [x kt ] represents the mean of all data in the kth group, Var[x kt ] represents the standard deviation of all data in the kth group, ε is a constant, Represents the tth data in the kth group after normalization operation; Step 3 and 4: Use the kth group of data after normalization as the input of the time domain convolutional network TCN, using parameter γ k and β k Perform a denormalization operation on the data output by the time domain convolution network TCN prediction, and calculate the loss function based on the actual data at the time of the time domain convolution network TCN prediction and the denormalization operation result; Reversely adjust the parameters of the time domain convolutional network TCN according to the loss function; Step 35: Set the number of data sets k=k+1, and return to execute step 33.

7. The distributed drift satellite telemetry data prediction method based on sliding window design according to claim 6 is characterized in that: The mean value E t [x kt ] and standard deviation Var[x kt ] are:

8. The distributed drift satellite telemetry data prediction method based on sliding window design according to claim 6 is characterized in that: The parameter γ k and β k The calculation method is: Among them, ω0, ω1, …, ω m is the weighting coefficient.

9. The distributed drift satellite telemetry data prediction method based on sliding window design according to claim 8 is characterized in that: The adopted parameter γ k and β k The data predicted and output by the time domain convolution network TCN is denormalized. The specific process is as follows: The kth group of normalized data is input into the time domain convolutional network TCN, and the predicted output of the time domain convolutional network TCN is recorded as Then use parameter γ k and β k right Perform an inverse normalization operation: in, Express The result of the inverse normalization operation.

10. The distributed drift satellite telemetry data prediction method based on sliding window design according to claim 9 is characterized in that: The specific process of step 4 is as follows: Step 41: Collect the history of satellite thruster fuel nozzles x The temperature data at each moment is preprocessed, and the preprocessed time series temperature data is used as the first group of data; Step 42, initializing the number of groups k′=1; Step 43: Normalize the k'th group of data: in, and is the affine parameter of the k′th group of data, x′ k′t represents the tth data in the k′th group, E t [x′ k′t ] represents the mean of all data in the k′th group, Var[x′ k′t ] represents the standard deviation of all data in the k′th group, ε is a constant, represents the tth data in the k′th group after normalization operation; according to and To calculate γ k′ and β k′ ; Step 4. Take the k′th group of data after normalization as the input of the trained time domain convolutional network TCN, according to γ k′ and β k′ Prediction output of the trained time domain convolutional network TCN Perform inverse normalization operation; in, Represents the result of the inverse normalization operation, that is, the final prediction result of the time-domain convolutional network TCN; Will as well as Previous T x The temperature data at time -1 is taken as the k′+1th window data; Step 45: Set k′=k′+1 and return to execute step 43.

Images