Bidirectional prediction intelligent continuing method and system for long-term interruption data of deep space narrow-band Doppler measurement

Abstract

The invention provides a bidirectional prediction intelligent continuing method and system for long-term interruption data of deep space narrowband Doppler measurement. The method comprises the following steps: S1, performing channel decomposition and aggregation of bidirectional prediction data before and after interruption based on time-frequency domain multi-angle characteristics; s2, carrying out unidirectional data modeling and prediction based on multi-channel decomposition and aggregation; s3, bidirectional prediction data variable weight fusion based on a confidence interval and frequency domain decomposition; and S4, polynomial fitting and prediction curve smoothing of the integrated data. According to the invention, adaptive decomposition and channel aggregation are carried out on data segments before and after long-term interruption of deep space exploration based on own characteristics, modeling and prediction are carried out on data of each channel from forward and backward directions, and data fusion and trend smoothing are carried out under the guidance of a confidence interval. The interpolation precision of the method is obviously higher than that of a traditional one-way extrapolation fitting method, and the method has good generalization and stability especially in a long-term deep space acquisition data interruption state.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep space detection and intelligent signal processing, and in particular relates to a bidirectional prediction intelligent connection method and system for long-term interruption data of deep space narrow-band Doppler measurement. Background Art

[0002] One of the important tasks of ground observation stations is to continuously track the phase and Doppler frequency changes of the downlink narrowband signal of deep space probes. The changes in the phase and frequency of downlink signals help to infer the motion parameters such as the position and speed of deep space probes, understand the subtle changes in the deep space environment around the probes, and help the navigation and orbit determination of deep space probes, as well as the research of planetary geodesy and related space sciences.

[0003] However, due to the complex deep space observation environment and the diverse requirements of the mission itself, such as short-term signal shielding, strong radio noise interference, or alternating observations of targets required by the mission, ground station equipment will inevitably experience observation interruptions for a period of time during the detector downlink signal tracking process. Frequent interruptions significantly reduce the availability of collected data, pose a threat to deep space missions, and increase uncertainty in the execution of observation missions. These difficult-to-eliminate interruptions pose technical challenges to current deep space exploration missions. Current technologies, such as fitting interpolation values, have many problems with the prediction and continuation of interrupted data, such as poor stability and poor generalization, and the accuracy of the continuation needs to be further improved. Summary of the invention

[0004] The purpose of the present invention is to provide a two-way predictive intelligent connection method and system for long-term interruption data of deep space narrow-band Doppler measurement, so as to provide high-precision intelligent connection for long-term interruption of observation data caused by continuous strong noise interference, mission observation requirements, etc., thereby improving mission robustness and the availability of collected data.

[0005] In order to solve the above technical problems, the present invention provides a two-way prediction intelligent connection method and system for long-term interrupted data of deep space narrow-band Doppler measurement. When a long-term data observation interruption is detected, high-precision modeling can be performed simultaneously from both positive and negative directions according to the characteristics of the data before and after the interruption, and the missing data can be connected with high precision through steps such as two-way intelligent prediction, data fusion and fitting smoothing. The specific technical solution is as follows:

[0006] A two-way prediction intelligent connection method for long-term interruption data of deep space narrow-band Doppler measurement, comprising the following steps:

[0007] S1, channel decomposition and aggregation of bidirectional prediction data before and after interruption based on multi-angle features in time and frequency domains;

[0008] S2, unidirectional data modeling and prediction based on multi-channel decomposition in step S1;

[0009] S3, variable weight fusion of bidirectional prediction data based on confidence interval and frequency domain decomposition;

[0010] S4, Polynomial fitting of integrated data and smoothing of prediction curves.

[0011] Preferably, in step S1, the channel decomposition based on multi-angle features in the time-frequency domain is specifically:

[0012] The overall decomposition process is divided into two levels.

[0013] Firstly, the time lookback window size is selected, and the data segments before and after the interruption are selected as network input data for prediction and continuation; an STL decomposition model is designed for the data in the selected lookback window to separate its trend item and periodic item.

[0014] Secondly, for the periodic terms after STL decomposition, they are converted into the frequency domain through discrete Fourier transform, and three frequency bands of high, medium and low are delineated in the spectrum, and then independent data channels are established.

[0015] Again, for each independent data channel designed, the high, medium and low frequency data are converted to the time domain through inverse Fourier transform for channel aggregation to complete the channel separation and analysis of the periodic part of the input data.

[0016] Preferably, in step S2, the one-way data modeling and prediction based on multi-channel decomposition is specifically as follows:

[0017] S21, simple network modeling of trend items: The trend items of the input data after STL decomposition are modeled and predicted by the GRBF network. The network output adopts the iterative input prediction method, that is, The data pred(t-1) predicted at time t-1 will be used as the input for the prediction data at the next moment.

[0018] S22, multi-channel aggregation and network analysis modeling of the periodic term part: the periodic term part of the data obtained by STL decomposition, after discrete Fourier transform to frequency domain, the entire normalized spectrum is divided into three intervals of high frequency [0.7 ~ 1], medium frequency [0.4 ~ 0.7], and low frequency [0 ~ 0.4], that is, the input periodic term data is divided into three independent channels of high, medium and low frequency according to the spectrum characteristics. Then, after the high, medium and low frequency data information is converted back to the time domain through the inverse Fourier transform (IDFT), the LSTM model is established for the data of each independent channel respectively, and prediction is performed separately.

[0019] S23, and the sum of the data output of the periodic term is determined by the weighted fusion of the outputs of each individual channel. The weighting coefficient is determined by the proportion of the amplitude of each component in the spectrum diagram.

[0020] S24, when it is detected that the data observation is interrupted at the set time, forward and reverse one-way predictions are performed simultaneously; wherein the total output result of the forward and reverse one-way predictions is d F (t) and d I (t) is determined by adding the prediction results of the period term and the trend term in each prediction direction.

[0021] Preferably, in step S3, the variable weight fusion of bidirectional prediction data based on confidence interval is specifically:

[0022] First, based on the calculation of forward and reverse data predictions, the confidence intervals of the forward data prediction and the reverse data prediction are calculated respectively.

[0023] Secondly, find the data segments with intersections in the two-way prediction confidence intervals to carry out data fusion. The data fusion weight is based on the area ratio of each intersection interval, that is, d'(t) = α × d F (t)+(1-α)×d I (t), Among them, S F With S I It is the area occupied by the forward prediction and the backward prediction in the intersection of the confidence intervals.

[0024] Preferably, in step S4, the polynomial fitting and prediction curve smoothing are specifically as follows:

[0025] First, multiple prediction data are integrated. Specifically, the prediction data used for the continuation consists of the following three parts: 1) The data before the intersection of the confidence intervals uses the forward prediction result d F (t), 2) The data after the confidence intervals have an intersection uses the reverse prediction result d I (t), 3) and the part with intersection in the middle uses the fused data result d'(t). The above three segments of data are integrated as the preliminary result of data continuation.

[0026] Secondly, polynomial fitting is applied to the above three data segments to perform smoothing processing to obtain the final continuation data segment.

[0027] The present invention also discloses a bidirectional prediction intelligent connection system for long-term interruption data of deep space narrow-band Doppler measurement, which is used to execute the above method and includes the following modules:

[0028] Channel decomposition module: channel decomposition and aggregation of bidirectional prediction data before and after interruption based on multi-angle features in time and frequency domains;

[0029] One-way data modeling and prediction module; One-way data modeling and prediction with multi-channel decomposition;

[0030] Data fusion module: variable weight fusion of bidirectional prediction data based on confidence interval and frequency domain decomposition;

[0031] Followed by the data acquisition module: polynomial fitting of integrated data and smoothing of prediction curves.

[0032] The present invention provides a two-way prediction intelligent connection method and system for long-term interruption data of deep space narrow-band Doppler measurement, which realizes high-precision one-way prediction by adaptively decomposing the data in the lookback window and setting up independent channels; and based on this, a data fusion scheme based on confidence interval is designed for forward and reverse one-way prediction data, which successfully suppresses the error accumulation in the multi-step prediction connection process. The present invention has the advantages of high connection accuracy and good generalization in the connection of long-term interruption data. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The present invention is a flowchart of a two-way prediction intelligent connection method for long-term interruption data of deep space narrow-band Doppler measurement in a preferred embodiment of the present invention.

[0034] Figure 2 A diagram of multi-level adaptive decomposition and channel aggregation of input data for a preferred embodiment of the present invention.

[0035] Figure 3 Input data independent channel modeling and prediction diagram for the preferred embodiment of the present invention.

[0036] Figure 4a ,4b is a preferred embodiment of the present invention, which is a single-channel long short-term memory artificial neural network (Long Short-Term Memory, LSTM) modeling based on frequency domain decomposition and a periodic item data prediction and fusion diagram.

[0037] Figure 5a ,5b is the Gradient Radial Basis Function neural network (GRBF) modeling and iterative prediction diagram of the trend item data of the preferred embodiment of the present invention.

[0038] Figure 6 This is a diagram of bidirectional data fusion and data integration based on confidence intervals according to a preferred embodiment of the present invention.

[0039] Figure 7 This is a diagram showing the cumulative error suppression effect of the bidirectional prediction intelligent connection method for long-term interruption data prediction in the preferred embodiment of the present invention. Figure 8The block diagram of a two-way prediction intelligent connection system for long-term interruption data of deep space narrow-band Doppler measurement is a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] like Figure 1 As shown, this embodiment is a two-way prediction intelligent connection method for deep space narrow-band Doppler measurement of long-term interruption data, and the specific implementation steps are as follows:

[0042] S1, channel decomposition of bidirectional prediction data before and after interruption based on multi-angle features in time and frequency domains;

[0043] S2, one-way data modeling and prediction based on decomposition mechanism;

[0044] S3, variable weight fusion of bidirectional prediction data based on confidence interval and frequency domain decomposition;

[0045] S4, Polynomial fitting of integrated data and smoothing of prediction curves.

[0046] In step S1, during the data preprocessing and modeling process, after a long data interruption is detected, the width of the backtracking window is first determined. On this basis, a segment of data is selected from both sides of the connecting data as network input.

[0047] Secondly, multi-level adaptive feature decomposition is performed on the bidirectional input data to divide the input data channels. The first level is STL decomposition, which decomposes the input data into trend terms and periodic terms. The second level of decomposition converts the more complex features of the periodic terms into the frequency domain, and then performs adaptive high-frequency, medium-frequency, and low-frequency channel division according to the different performance of the periodic frequency.

[0048] Finally, the high, medium and low frequency data are converted to the frequency domain again through inverse discrete Fourier transform to complete the channel separation and aggregation of the periodic part of the input data.

[0049] Step S2, after the adaptive decomposition and channel aggregation are completed, a prediction network is designed for each independent channel to perform unidirectional prediction. When the high, medium and low frequency channels are fused, their weights are determined according to their respective amplitude proportions in the spectrum.

[0050] Specific:

[0051] S21, for the decomposed trend item, a simple model based on GRBF is designed to model and iteratively predict it;

[0052] S22, for the periodic part of the input data, LSTM networks are designed for modeling and iterative prediction according to the three channels of high, medium and low frequencies;

[0053] S23, the overall output of the periodic term of the input data is obtained by weighting the prediction results of the three channels of high, medium and low frequencies, and the respective weights are determined by the amplitude ratio of the high, medium and low frequency components in the spectrum;

[0054] S24, when the set time data observation interruption is detected, forward and reverse one-way predictions are performed simultaneously; wherein the forward and reverse one-way prediction output results d F (t) and d I (t) is determined by adding the prediction results of the period term and the trend term in each prediction direction.

[0055] Step S3, after completing the unidirectional prediction of the data, calculate the confidence intervals of each prediction direction. For the intersection of the intervals, perform weighted fusion according to the area proportion of the confidence intervals of each direction in the intersection. Specifically:

[0056] S31, after the one-way prediction is completed, the confidence intervals of the forward data prediction and the reverse data prediction are calculated respectively;

[0057] S32, take the intersection of the positive and negative data prediction confidence intervals for data fusion, and use the fused data to calibrate the predicted data. The fusion rule is d'(t) = α × d F (t)+(1-α)×d I (t); where the fusion factor S F With S I It is the area occupied by the forward prediction and the backward prediction in the intersection of the confidence intervals.

[0058] Step S4, integrating the three segments of forward prediction data, bidirectional interval fusion data and reverse prediction data to form a rough continuation data segment. Perform polynomial fitting on the rough continuation data segment, smooth its numerical trend, and use it as the final continuation data segment. Specifically:

[0059] S41, prediction data integration; the prediction data for subsequent use consists of the following three parts: 1) the data before the intersection of the confidence intervals is obtained by forward prediction d F (t); 2) The intersection of the confidence intervals is composed of the fused data d'(t); 3) The data after the intersection of the confidence intervals is composed of the reverse prediction d I(t) constitute;

[0060] S42, applying polynomial fitting to the three sorted data to smooth the curve, and outputting it as the final connection data.

[0061] For step S1, Figure 2 A multi-level adaptive decomposition and channel aggregation of input data is described.

[0062] The first level decomposition is the STL decomposition as shown in equations (1) and (2), which is used to divide the input data into trend items X Trend With the periodic term X Period . Avgpool represents the average pooling operation on the input data, and Padding represents the padding operation on the input data sequence X during the pooling process. x is a specific value of the input data sequence, N is the average pooling data length, and x k is a specific data in the pooling sequence.

[0063]

[0064] X period =XX trend (2)

[0065] For the periodic term of the data, due to its complex characteristics, a secondary frequency domain channel decomposition mechanism is designed for it. The specific steps are as follows:

[0066] First, the periodic term is transformed into the frequency domain by the discrete Fourier transform of the N-point sequence shown in equation (3). DFT is the number of Fourier transform points, x w [n] is the corresponding Fourier transform coefficient, e is the natural logarithm base, j is the imaginary unit, n is the time domain sequence number, and k is the frequency domain sequence number.

[0067]

[0068] Secondly, the entire transformed spectrum is normalized and divided into frequency domain intervals, including low frequency [0-0.4], medium frequency [0.4-0.7], and high frequency [0.7-1]. The data in the frequency domain are aggregated according to this divided interval and converted back to the time domain through inverse Fourier transform to divide the data channels. The specific mathematical expression is shown in formula (4), where D HF ,D MF ,D LF Represent high, medium and low frequency data channels respectively.

[0069] D HF =IDFT{Spec[0.7:1]}, DMF =IDFT{Spec[0.4:0.7]}, D LF =X Peridical -D HF -D MF (4)

[0070] For step S2, Figure 3 Model and predict independent channels of input data. After adaptively decomposing and aggregating the input data, model each independent channel data.

[0071] Among them, for periodic item data, its channel adaptive decomposition, aggregation and independent prediction process are as follows Figure 4a As shown, the high, medium and low frequency independent channels after the periodic term is decomposed again are designed as follows Figure 4b The LSTM network shown in the figure performs iterative multi-step prediction. At the output, weighted fusion is performed according to the amplitude proportion of each component in the spectrum.

[0072] Among them, for trend item data, the design is as follows Figure 5a The GRBF network shown in the figure performs iterative prediction. In the GRBF network, in order to better perceive the gradient changes of the input data and deduce the data change trend at future time points, a difference layer is added before the GRBF overall network to perform differential processing on the input data. The data prediction rule is established as shown in formula (5). The iterative prediction method is used in the data prediction process, that is, the data pred(t-1) obtained by the prediction at time t-1 will be used as the input of the prediction data at the next moment.

[0073]

[0074] in, is the correspondence between the output trend item result and the input sequence, and n is the width of the network backtracking window. The overall output of the one-way prediction is obtained by adding the prediction results of the period item and the trend item.

[0075] Figure 6 It is a two-way data fusion and data integration based on confidence intervals. For step S4, after the forward and reverse one-way data predictions are completed, the confidence intervals of each are calculated. The intersection of the two confidence intervals is found, and the data is integrated into three parts, namely: 1) the data segment before the intersection of the confidence intervals, 2) the data segment with the intersection of the confidence intervals, and 3) the data segment after the intersection of the confidence intervals.

[0076] For the data segments before the confidence intervals intersect, the forward prediction data is used because the forward prediction has higher reliability. For the data segments where the confidence intervals intersect, weighted fusion is performed based on the proportion of the area of ​​each confidence interval in the intersection.

[0077] That is, the fusion rule is as shown in formula (6):

[0078] d'(t)=α×d F (t)+(1-α)×d I (t) (6)

[0079] The calculation of the fusion factor α is shown in formula (7):

[0080]

[0081] Among them, S F With S I It is the area occupied by the forward prediction and the backward prediction in the intersection of the confidence intervals.

[0082] Figure 7 This is a comparison of the cumulative errors of the two-way prediction intelligent connection method and the conventional one-way fitting extrapolation under different connection step conditions after the typical satellite narrowband Doppler frequency measurement is interrupted. From the results, it can be seen that under the conditions of 60, 90, 150, and 300 connection steps, the proposed two-way prediction intelligent connection method is improved by 4.73, 5.46, 8.85, and 9.29 times in terms of cumulative error respectively compared with the conventional one-way fitting extrapolation method. Therefore, the two-way prediction intelligent connection method for long-term interruption data of deep space narrowband Doppler measurement disclosed in the present invention has the characteristics of high connection accuracy, stable performance, and strong generalization ability when dealing with long-term data interruption connection tasks.

[0083] like Figure 8 As shown, this embodiment discloses a two-way prediction intelligent connection system for long-term interruption data of deep space narrow-band Doppler measurement, which is used to execute the above method and includes the following modules:

[0084] Channel decomposition module: channel decomposition and aggregation of bidirectional prediction data before and after interruption based on multi-angle features in time and frequency domains;

[0085] One-way data modeling and prediction module; One-way data modeling and prediction with multi-channel decomposition;

[0086] Data fusion module: variable weight fusion of bidirectional prediction data based on confidence interval and frequency domain decomposition;

[0087] Followed by the data acquisition module: polynomial fitting of integrated data and smoothing of prediction curves.

[0088] For other contents of this embodiment, please refer to the above method embodiment.

[0089] The present invention can significantly improve the reliability of observation tasks by intelligently connecting the interrupted deep space downlink narrowband Doppler trajectory change data collected by on-site equipment with high precision. At the same time, it can help ground stations to recapture information such as the position of aircraft after data interruption, improve the availability of collected signal data, and has very important theoretical and practical significance for the smooth progress of deep space exploration missions. It is an effective technical compensation method under extremely harsh tracking conditions such as extremely low signal-to-noise ratio in deep space.

[0090] In summary, the present invention performs adaptive decomposition and channel aggregation on the data segments before and after the long-term interruption of deep space exploration based on their own characteristics, and models and predicts the data of each channel from the positive and negative directions respectively, and carries out data fusion and trend smoothing under the guidance of the confidence interval. The interpolation accuracy is significantly higher than that of the traditional one-way extrapolation fitting method. Especially when the long-term deep space acquisition data is interrupted, the present invention has good generalization and stability.

[0091] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed by the present invention should be included in the protection scope recorded in the claims.

Claims

1. A two-way prediction intelligent connection method for long-term interruption data of deep space narrowband Doppler measurement, characterized in that: The following steps are involved: S1, channel decomposition and aggregation of bidirectional prediction data before and after interruption based on multi-angle features in time and frequency domains; S2, unidirectional data modeling and prediction based on multi-channel decomposition in step S1; S3, variable weight fusion of bidirectional prediction data based on confidence interval and frequency domain decomposition; S4, Polynomial fitting of integrated data and smoothing of prediction curves.

2. The bidirectional prediction intelligent connection method for long-term interruption data of deep space narrowband Doppler measurement according to claim 1 is characterized in that: Step S1 is specifically as follows: S11, selecting the time backtracking window size, and selecting the data segments before and after the interruption as network input data for prediction and continuation; S12, performing STL decomposition on the input data to decompose it into periodic items and trend items; S13, transforming the periodic term component into the frequency domain; dividing it into three channels of high frequency, medium frequency and low frequency based on the transformed spectrum; S14, after converting the high, medium and low frequency data into the frequency domain through inverse discrete Fourier transform, complete the channel decomposition and aggregation of the periodic term part of the input data.

3. The bidirectional prediction intelligent connection method for long-term interruption data of deep space narrowband Doppler measurement according to claim 2 is characterized in that: In step S12, the STL decomposition is as follows: X period =X-X trend 。 4. The two-way prediction intelligent connection method for long-term interruption data of deep space narrowband Doppler measurement according to claim 2 is characterized in that: In step S13, the periodic term is transformed into the frequency domain by the following discrete Fourier transform:

5. The two-way prediction intelligent connection method for long-term interruption data of deep space narrowband Doppler measurement according to any one of claims 2 to 4, characterized in that: Step S2 is specifically as follows: S21, for the decomposed trend item, a simple model based on GRBF is designed to model and iteratively predict it; S22, for the periodic part of the input data, LSTM networks are designed for modeling and iterative prediction according to the three channels of high, medium and low frequencies; S23, the overall output of the periodic term of the input data is obtained by weighting the prediction results of the three channels of high, medium and low frequencies, and the respective weights are determined by the amplitude ratio of the high, medium and low frequency components in the spectrum; S24, when the set time data observation interruption is detected, forward and reverse one-way predictions are performed simultaneously; wherein the forward and reverse one-way prediction output results d F (t) and d I (t) is determined by adding the prediction results of the period term and the trend term in each prediction direction.

6. The bidirectional prediction intelligent connection method for long-term interruption data of deep space narrowband Doppler measurement according to claim 5 is characterized in that: Step S3 is specifically as follows: S31, after the one-way prediction is completed, the confidence intervals of the forward data prediction and the reverse data prediction are calculated respectively; S32, take the intersection of the positive and negative data prediction confidence intervals for data fusion, and use the fused data to calibrate the predicted data. The fusion rule is d'(t) = α × d F (t)+(1-α)×d I (t); where the fusion factor S F With S I It is the area occupied by the forward prediction and the backward prediction in the intersection of the confidence intervals.

7. The two-way prediction intelligent connection method for long-term interruption data of deep space narrowband Doppler measurement according to claim 6 is characterized in that: Step S4 is specifically as follows: S41, prediction data integration; the prediction data for subsequent use consists of the following three parts: 1) the data before the intersection of the confidence intervals is obtained by forward prediction d F (t); 2) The intersection of the confidence intervals is composed of the fused data d'(t); 3) The data after the intersection of the confidence intervals is composed of the reverse prediction d I (t) constitute; S42, applying polynomial fitting to the sorted three segments of data to smooth the curve, and outputting it as the final connection data.

8. A two-way prediction intelligent connection system for long-term interruption data of deep space narrowband Doppler measurement, used to execute the method according to any one of claims 1 to 7, characterized in that: Includes the following modules: Channel decomposition module: channel decomposition and aggregation of bidirectional prediction data before and after interruption based on multi-angle features in time and frequency domains; One-way data modeling and prediction module; One-way data modeling and prediction with multi-channel decomposition; Data fusion module: variable weight fusion of bidirectional prediction data based on confidence interval and frequency domain decomposition; Followed by the data acquisition module: polynomial fitting of integrated data and smoothing of prediction curves.

Images