Satellite clock bias dual-stream prediction model construction method, satellite clock bias prediction method and system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-08
- Publication Date
- 2026-08-11
AI Technical Summary
[0006]本发明的目的在于提供一种基于模型反演与梯度敏感性分析的卫星钟差双流预测模型构建方法、卫星钟差预测方法及系统，以解决现有技术中存在的递归外推误差累积、统一高分辨率建模导致的冗余计算较大以及历史时间贡献分布缺乏量化刻画的问题
[0034]1. 本发明通过在完整历史视野上执行梯度敏感性诊断，并将归一化敏感度分布直接转化为分流规则，使快流输入区间、压缩处理区间和慢流视野长度的确定由可计算规则完成，而非依赖人工经验设定。
Smart Images
Figure CN122548146A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of satellite navigation and intelligent timing prediction technology, and particularly relates to a method for constructing a satellite clock bias dual-stream prediction model based on model inversion and gradient sensitivity analysis, a satellite clock bias prediction method and system. Background Technology
[0002] Satellite clock bias (SCB) prediction is a fundamental technology in satellite navigation for precise positioning, timing, and continuous service. Even in application scenarios with network latency, link interruptions, and limited edge terminal resources, it is still necessary to predict satellite clock biases for multiple future moments based on historical satellite clock bias sequences.
[0003] In existing technologies, one type of method uses recursive extrapolation to gradually generate predictions for multiple future time points. This type of method is prone to error accumulation as the prediction horizon increases. Another type of method directly models the entire long historical sequence using a uniform high-resolution structure. While this type of method can avoid some of the recursive error problem, it usually assumes that each historical time segment has a similar contribution to the future prediction result, making it difficult to reflect the differences between distant periodic anchor points, near-terminal local perturbations, and intermediate low-contribution intervals.
[0004] Satellite clock bias sequences typically contain both long-term trends and near-term high-frequency fluctuations. When using existing techniques to model the entire long-history sequence with a uniform structure, additional computational resources are often consumed in low-contribution historical intervals, and there is a lack of clear description of the basis for the division of labor between the two streams.
[0005] Based on the aforementioned existing technologies, there is an urgent need in this field to provide a technical solution that can first diagnose the time contribution distribution of historical satellite clock error sequences, then determine the dual-stream division rules based on the diagnostic results, and construct a dual-stream prediction model accordingly. Summary of the Invention
[0006] The purpose of this invention is to provide a method for constructing a satellite clock bias dual-stream prediction model based on model inversion and gradient sensitivity analysis, a satellite clock bias prediction method and system, to solve the problems of recursive extrapolation error accumulation, large redundant calculations caused by unified high-resolution modeling, and lack of quantitative characterization of historical time contribution distribution in the existing technology.
[0007] To achieve the above objectives, the present invention provides the following technical solution:
[0008] A method for constructing a satellite clock bias dual-stream prediction model based on model inversion and gradient sensitivity analysis includes the following steps:
[0009] S1, acquire multiple sets of historical satellite clock error sequences as training samples, construct and train a full-view detection model; when the full-view detection model uses a complete historical sequence corresponding to any candidate historical input length, it does not need to pre-split the input sequence according to the dual-stream division rule, and directly learns the mapping relationship between the historical sequence and the target prediction sequence.
[0010] S2, freeze the network parameters of the full-view detection model after training in step S1, and obtain the normalized sensitivity distribution of the future prediction output at each historical moment by performing model inversion and gradient backpropagation on the input historical sequence.
[0011] S3, determine the dual-stream division of labor rule according to the normalized sensitivity distribution, wherein the dual-stream division of labor rule includes at least the slow stream field of view length, the fast stream input interval, and the compression processing interval;
[0012] S4. Construct a satellite clock bias dual-stream prediction model based on the dual-stream division of labor rule. The satellite clock bias dual-stream prediction model includes a slow-stream branch, a fast-stream branch, and a fusion decoder. The slow-stream branch is used to extract long-period trend features from historical inputs within the slow-stream field of view. The fast-stream branch is used to extract short-term disturbance features from historical inputs within the fast-stream input interval. The fusion decoder is used to fuse the outputs of the two branches to generate satellite clock bias predictions for multiple future moments.
[0013] S5, the satellite clock bias dual-stream prediction model is trained using the training samples, and the satellite clock bias dual-stream prediction model is obtained after training.
[0014] The dual-flow clock bias prediction model designed in this invention locks in long-term trends through slow flow and captures short-term jumps through fast flow. It actively eliminates redundant data and adaptively adjusts weights, which improves accuracy while significantly reducing computing power overhead. It is especially suitable for edge deployment scenarios with extreme computing power constraints, such as deep space exploration, and has good physical interpretability.
[0015] Preferably, in a specific implementation, in step S1, the historical satellite clock error sequence is first-order differencing is performed to weaken the low-frequency drift term; and the input sequence after differencing is reversibly instance normalized to reduce the statistical distribution shift caused by different observation conditions.
[0016] Preferably, in a specific embodiment, in step S2, the full-view detection model is placed in an evaluation state, and the input history sequence is set to a differentiable state; a scalar response function is constructed with the future prediction output; the scalar response function is backpropagated with respect to the input history sequence; the original sensitivity sequence is obtained based on the statistical values of the input gradient at each time step, and the original sensitivity sequence is normalized to obtain the normalized sensitivity distribution.
[0017] Preferably, in a specific implementation, in step S3, a normalized sensitivity distribution is calculated for at least two different candidate historical input lengths; the global mean and standard deviation are calculated for the normalized sensitivity sequence corresponding to each candidate historical input length, and then... Determine an adaptive high-sensitivity threshold; under the constraint of the adaptive high-sensitivity threshold, extract the near-end continuous high-sensitivity segments located at the end of the candidate historical input length and close to the current time, and the far-end continuous high-sensitivity segments located far from the current time; determine the continuous intermediate segments located between the far-end continuous high-sensitivity segments and the near-end continuous high-sensitivity segments, and whose average sensitivity is lower than the average sensitivity of the aforementioned two high-sensitivity segments, as continuous low-sensitivity segments; calculate the separation degree based on the average sensitivity of the far-end continuous high-sensitivity segments, the near-end continuous high-sensitivity segments, and the continuous low-sensitivity segments, and select the candidate historical input length with the largest separation degree as the slow-flow field of view length; determine the near-end continuous high-sensitivity segments as the fast-flow input interval, and determine the continuous low-sensitivity segments as the compression processing interval.
[0018] Preferably, in step S4, the slow-flow branch includes: performing downsampling or smooth compression processing on the compression processing interval within the slow-flow field of view to obtain a compressed low-resolution segment, and combining the low-resolution segment with the uncompressed segment within the slow-flow field of view in chronological order to form a slow-flow input; wherein, the downsampling step size is dynamically determined based on the ratio of the compression processing interval length to the fast-flow input interval length; and performing long-period feature extraction on the slow-flow input to obtain long-period trend features.
[0019] Preferably, the slow-flow branch further includes: mapping the slow-flow input to the frequency domain and dividing it into three channels—high-frequency, mid-frequency, and low-frequency—based on the spectral amplitude distribution; performing two-dimensional reconstruction after inverse transformation of the three channels back to the time domain; and extracting the reconstructed features using two-dimensional multi-scale convolution to obtain the long-period trend features.
[0020] Preferably, in step S4, the fast-flow branch includes: extracting the original resolution data corresponding to the fast-flow input interval from the input history sequence as the near-end fast-flow input; and performing temporal feature extraction on the near-end fast-flow input to obtain short-term perturbation features.
[0021] Preferably, the temporal feature extraction is achieved through one-dimensional convolution, a lightweight bottleneck layer, or a depthwise separable one-dimensional convolution.
[0022] Preferably, step S5 includes: fusing the long-period trend features extracted from the slow-flow branch and the short-term disturbance features extracted from the fast-flow branch in step S4 to obtain fused features; the fused decoder preferably adopts a direct multi-step prediction decoder, which outputs satellite clock bias prediction results for multiple future times through one forward propagation of the fused features; and updating the network parameters of the satellite clock bias dual-flow prediction model based on the error between the multiple future time satellite clock bias prediction results and the target prediction sequence.
[0023] This invention also discloses a satellite clock bias prediction method, which includes the following steps:
[0024] P1, obtain the historical satellite clock bias sequence before the time to be predicted, and call the pre-set dual-stream division of labor rules and the parameters of the pre-trained satellite clock bias dual-stream prediction model. The dual-stream division of labor rules and the parameters of the satellite clock bias dual-stream prediction model can be encapsulated into a prediction deployment package.
[0025] P2, according to the dual-stream division of labor rule, the original resolution data corresponding to the fast stream input interval is extracted from the historical satellite clock error sequence as the fast stream input, and the downsampling step size is dynamically determined according to the ratio of the slow stream field of view length and the length of the compression processing interval to the length of the fast stream input interval. Downsampling or smooth compression processing is performed on the compression processing interval to construct the slow stream input.
[0026] P3, input the slow stream input and the fast stream input to the satellite clock bias dual-stream prediction model determined by the parameters of the satellite clock bias dual-stream prediction model, wherein the satellite clock bias dual-stream prediction model includes a slow stream branch, a fast stream branch and a fusion decoder;
[0027] P4 extracts long-term trend features and short-term disturbance features using the slow-flow branch and the fast-flow branch respectively, and outputs satellite clock error prediction results for multiple future moments through a fusion decoder.
[0028] This invention also discloses a satellite clock bias dual-stream prediction model construction system based on model inversion and gradient sensitivity analysis, comprising the following modules:
[0029] Training sample construction and model calling module: used to obtain the historical satellite clock difference sequence before the time to be predicted, and call the parameters of the pre-set dual-stream division of labor rules and the pre-trained satellite clock difference dual-stream prediction model;
[0030] Dual-stream division of labor rule module: used to extract the original resolution data corresponding to the fast stream input interval from the historical satellite clock difference sequence as the fast stream input according to the dual-stream division of labor rule, and construct the slow stream input according to the slow stream field of view length and compression processing rules;
[0031] Input module: used to input the slow-flow input and the fast-flow input into the satellite clock bias dual-flow prediction model;
[0032] Prediction output module: It is used to extract long-term trend features and short-term disturbance features by using the slow flow branch and the fast flow branch respectively, and output the satellite clock difference prediction results for multiple future times through the fusion decoder.
[0033] Compared with the prior art, the present invention has at least the following beneficial effects:
[0034] 1. This invention performs gradient sensitivity diagnosis over a complete historical view and directly transforms the normalized sensitivity distribution into a flow splitting rule, enabling the determination of the fast flow input interval, compression processing interval, and slow flow view length to be accomplished by computable rules rather than relying on manual experience.
[0035] 2. This invention identifies distant continuous high-sensitivity segments, near continuous high-sensitivity segments, and continuous low-sensitivity segments according to an adaptive high-sensitivity threshold, and selects the candidate historical input length with the highest separation, which is beneficial for objectively determining the boundary relationship between the long-period anchor point region, the near-end perturbation region, and the intermediate low-contribution region.
[0036] 3. This invention dynamically determines the downsampling step size based on the ratio of the length of the compressed processing interval to the length of the fast-flow input interval, and performs redundancy reduction processing only on the compressed processing interval, which is beneficial to reduce the computational overhead caused by the intermediate low-contribution interval while retaining the far-end periodic information.
[0037] 4. This invention enables the prediction side to complete closed-loop inference without repeatedly performing probe model training and sensitivity diagnosis by jointly calling the dual-stream division rules and trained model parameters on the prediction side. It also reduces the impact of the gradual accumulation of errors in recursive extrapolation by direct multi-step prediction. In a preferred embodiment, the dual-stream division rules and model parameters can be encapsulated as a prediction deployment package. Attached Figure Description
[0038] Figure 1 This is a flowchart of the overall process of constructing a satellite clock bias dual-stream prediction model and a satellite clock bias prediction method based on model inversion and gradient sensitivity analysis in a preferred embodiment of the present invention. It shows the overall process from full-view detection model training, sensitivity diagnosis, diversion rule determination to dual-stream prediction output.
[0039] Figure 2 The flowchart for normalized sensitivity diagnosis and diversion rule determination in a preferred embodiment of the present invention illustrates the process of identifying key intervals and generating dual-flow diversion rules based on sensitivity distribution.
[0040] Figure 3This is a schematic diagram of the structure of the asymmetric dual-stream prediction model in a preferred embodiment of the present invention, showing the compositional relationship between the slow-stream branch, the fast-stream branch, and the fusion decoder.
[0041] Figure 4 The flowchart of the feature processing of the slow flow branch and the fast flow branch in a preferred embodiment of the present invention shows the process of performing long-term trend modeling and short-term disturbance modeling for the two branches and completing feature fusion.
[0042] Figure 5 This is a schematic diagram illustrating the evolution of the normalized sensitivity distribution under different candidate historical input lengths in a preferred embodiment of the present invention. It shows the changes in the historical time contribution distribution under different input lengths and its supporting role in determining the diversion rule.
[0043] Figure 6 This is a comparison chart of the prediction accuracy of the preferred embodiment of the present invention and the comparative method in the satellite clock error prediction task. It shows the comparison results of LSTM, LSTM-Attention, Transformer and the method of the present invention in terms of RMSE and RE under 12h and 24h prediction settings.
[0044] Figure 7 This is a comparison chart of the model complexity of the preferred embodiment of the present invention and the comparative method in the satellite clock error prediction task. It shows the comparison results of LSTM, LSTM-Attention, Transformer and the method of the present invention in terms of parameter quantity, computational quantity, inference latency and peak memory.
[0045] Figure 8 This is a block diagram of a satellite clock error prediction system according to a preferred embodiment of the present invention. Detailed Implementation
[0046] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.
[0047] This invention belongs to the field of satellite navigation and intelligent timing prediction technology. First, it acquires historical satellite clock bias sequences and constructs a full-view detection model. Based on this, it obtains a normalized sensitivity distribution through model inversion and gradient backpropagation, and determines the dual-stream division of labor rules according to the sensitivity distribution. Subsequently, it constructs and trains a satellite clock bias dual-stream prediction model, and finally outputs satellite clock bias prediction results for multiple future times.
[0048] like Figure 1 As shown in the figure, this embodiment presents a method for constructing a satellite clock bias dual-stream prediction model based on model inversion and gradient sensitivity analysis. The steps are as follows:
[0049] Step S1: Obtain multiple sets of historical satellite clock error sequences as training samples to construct and train a full-view detection model; specifically:
[0050] Multiple sets of historical satellite clock bias sequences and their corresponding target prediction sequences are obtained to form supervised learning samples. The historical satellite clock bias sequences serve as the input sequences, and the target prediction sequences serve as the predicted targets for multiple future time points.
[0051] In this embodiment, the original historical satellite clock bias sequence is... First, perform the first-order difference to obtain:
[0052]
[0053] This is to reduce the impact of long-term drift terms on the modeling process.
[0054] in, This represents the sequence value of the original historical satellite clock bias sequence at time t. This represents the sequence value of the original historical satellite clock error sequence at time t+1. This represents the first-order difference value corresponding to time t. By performing this first-order difference, the impact of the long-term drift term on the modeling process can be reduced.
[0055] Furthermore, invertible instance normalization is performed on the differencing input sequence. Let the input sequence instance be... Its mean is The variance is The normalization result is:
[0056]
[0057] Network output normalized predicted value Then, the dimensions can be recovered through inverse transformation:
[0058]
[0059] in, For numerically stable terms, This represents the normalized predicted value output by the network. This represents the prediction result after inverse normalization.
[0060] To avoid introducing a dual-stream bias in the diagnostic phase, this embodiment first constructs a full-view detection model. The full-view detection model refers to a reference model that maintains the complete historical sequence range corresponding to the candidate historical input length during the diagnostic phase, and does not pre-divide the historical sequence based on the fast-stream input interval or the compressed processing interval. It is used to learn the overall mapping relationship between historical satellite clock error sequences and future predicted sequences.
[0061] In this embodiment, the full-view detection model can employ a network structure with adaptive frequency domain decomposition and time-frequency feature aggregation capabilities. The structure may include an adaptive frequency domain decomposition module, a period extraction module, and a multi-scale feature aggregation module. For example, a detection network containing high-frequency, mid-frequency, and low-frequency component decomposition, period extraction, and time-frequency transformation blocks can be used to achieve full-view modeling. More specifically, the adaptive frequency domain decomposition module can divide the high-frequency, mid-frequency, and low-frequency components based on the spectral amplitude distribution, and then map each frequency component back to the time domain for aggregation modeling. However, this invention does not limit it to using a single network structure, as long as it can form stable time-dependent learning results over a complete historical view.
[0062] Step S2: Freeze the model parameters and obtain the normalized sensitivity distribution of future predictions at each historical moment by inverting and backpropagating the input historical sequence; specifically:
[0063] After the full-view detection model is trained, it is placed in the evaluation state, all network parameters are frozen, and the input historical sequence is set to a differentiable state. Let the future prediction output of the i-th sample be... Where H is the prediction horizon length and C is the number of channels.
[0064] In this embodiment, a scalar response function is constructed by summing all prediction steps and channel outputs:
[0065]
[0066] in, Let i represent the scalar response function corresponding to the i-th sample. H represents the prediction output of the i-th sample at the h-th prediction step and the c-th channel, where H represents the prediction horizon length and C represents the number of channels.
[0067] Backpropagation is performed on the gradient of the input sequence at time t, and the original sensitivity is defined:
[0068]
[0069] Where N is the number of samples participating in the statistics. Represents the original sensitivity at time t. Let represent the gradient of the scalar response function of the i-th sample with respect to the input history sequence at time t for each channel input. Let L1 norm be denoted. To facilitate comparisons between different candidate input lengths, the original sensitivity is normalized along the time dimension, resulting in:
[0070]
[0071] in, This represents the normalization sensitivity at time t. Indicates the original sensitivity at each time point The original sensitivity sequence is composed of These represent the minimum and maximum values in the original sensitivity sequence, respectively.
[0072] This allows us to obtain the normalized sensitivity distribution of future prediction outputs at each historical moment.
[0073] Step S3: Determine the dual-stream division of labor rules based on the normalized sensitivity distribution;
[0074] This step transforms the sensitivity diagnostic results into executable model structure rules. Specifically:
[0075] like Figure 2 As shown, this embodiment first obtains the normalized sensitivity distribution of historical sequences to future predicted outputs based on a full-field-of-view detection model. Then, based on this sensitivity distribution, it identifies continuous high-sensitivity segments at the far end, continuous high-sensitivity segments at the near end, and continuous low-sensitivity segments, thereby determining the slow-flow field-of-view length, the fast-flow input interval, and the compression processing interval. Based on this, to further determine the optimal splitting boundary under different candidate historical input lengths, the normalized sensitivity distributions corresponding to multiple candidate historical input lengths can be compared and analyzed.
[0076] In this embodiment, at least two different candidate historical input lengths are pre-set, and the corresponding normalized sensitivity distribution is calculated for each candidate historical input length. Preferably, the candidate historical input length can span short-term local scales, partial periodic scales, full periodic scales, and extended periodic scales. In the example scenario of BeiDou satellites and 30-second sampling, the candidate historical input lengths can correspond to 2 hours, 12 hours, 24 hours, and 48 hours, but this set of durations is only a preferred example and does not constitute the only limitation of the present invention.
[0077] like Figure 5 As shown, with the change in the length of the candidate historical input, the sensitivity distribution of different time positions in the historical sequence to the future prediction results exhibits different evolutionary characteristics. Based on the separation between the far-end high-sensitivity region, the near-end high-sensitivity region, and the middle low-sensitivity region, this invention selects the candidate historical input length with the largest separation degree and generates the corresponding dual-stream division of labor rule accordingly.
[0078] Suppose there are K candidate historical input lengths, and the number of historical sampling points corresponding to the k-th candidate historical input length is T. k ,in Let the normalized sensitivity sequence corresponding to the length of the k-th candidate historical input be . ,in, This represents the normalized sensitivity corresponding to the t-th historical sampling point with the k-th candidate historical input length. We can first calculate its global mean. and global standard deviation :
[0079]
[0080] The adaptive high-sensitivity threshold corresponding to the length of the k-th candidate historical input is determined according to the following formula:
[0081]
[0082] After obtaining the adaptive high-sensitivity threshold, the traffic splitting rule can be determined as follows:
[0083] 1. Identify distant, continuous, highly sensitive segments The distant continuous highly sensitive segment is located on the side of the candidate historical input length that is far from the current time, and satisfies the following conditions: A continuous time segment.
[0084] 2. Identify proximal contiguous highly sensitive fragments The proximal continuous high-sensitivity segment is a continuous time range located at the end of the candidate historical input length and immediately adjacent to the time to be predicted, satisfying the following conditions: A continuous time segment.
[0085] 3. Identify consecutive low-sensitivity segments The continuous low-sensitivity segment is a continuous intermediate segment located between the distal continuous high-sensitivity segment and the proximal continuous high-sensitivity segment, and whose average sensitivity is lower than the average sensitivity of the two aforementioned high-sensitivity segments.
[0086] Furthermore, the separation degree of the length of the k-th candidate historical input can be defined as:
[0087]
[0088] in, This represents the far-end continuous highly sensitive segment corresponding to the length of the k-th candidate historical input. This represents the near-end contiguous highly sensitive segment corresponding to the length of the k-th candidate historical input. This represents the continuous low-sensitivity segment corresponding to the length of the k-th candidate historical input. This represents the average sensitivity on the corresponding segment. The separation degree D is selected. kThe longest candidate historical input length is used as the slow-flow field of view length; the corresponding consecutive high-sensitivity segments near the front are defined as the fast-flow input interval; and the corresponding consecutive low-sensitivity segments are defined as the compression processing interval. Thus, the dual-flow division rule no longer relies on manually preset fixed boundaries, but is directly generated from the segmented shape of the normalized sensitivity distribution.
[0089] Step S4: Construct a satellite clock bias dual-stream prediction model based on the dual-stream division of labor rule: the slow-stream branch is used for long-term trend modeling, the fast-stream branch is used for short-term disturbance modeling, and the fusion decoder fuses the outputs of the two branches to generate multi-step predictions; specifically:
[0090] A satellite clock bias dual-stream prediction model is constructed based on the dual-stream division of labor rules. The satellite clock bias dual-stream prediction model includes a slow stream branch, a fast stream branch, and a fusion decoder.
[0091] like Figure 3 As shown, the satellite clock bias dual-stream prediction model includes a slow-stream branch, a fast-stream branch, and a fusion decoder. The slow-stream branch extracts long-term trend features, the fast-stream branch extracts near-term short-term disturbance features, and the fusion decoder fuses the features from the two branches and outputs prediction results for multiple future times. Further, as... Figure 4 As shown, in the slow flow branch, downsampling or smooth compression processing is performed on the compression processing interval within the slow flow field of view, and long-period feature extraction is performed on the constructed slow flow input; in the fast flow branch, temporal feature extraction is performed on the original resolution data corresponding to the fast flow input interval to obtain short-time perturbation features, and then the features of the two branches are fused and the prediction result is output.
[0092] 1) Slow-flow branch
[0093] The slow-flow branch is used to perform long-cycle trend modeling on historical inputs within the slow-flow field of view and to perform redundancy reduction processing on the compressed processing interval.
[0094] Let the original sequence corresponding to the compression processing interval be... Its length is Let the length of the fast current input interval be L. f In one implementation, the downsampling step size is first dynamically determined according to the following formula:
[0095]
[0096] Then, downsampling or smoothing compression is performed only on the compressed region to obtain the compressed low-resolution segment:
[0097]
[0098] Where AvgPool represents the average pooling operation, kernel_size represents the average pooling window size, and stride represents the sliding step size of the pooling window; in this embodiment, both kernel_size and stride are set to M.
[0099] Let the historical segment within the slow-flow field of view, excluding the compression processing interval, be denoted as . Then you can Combining them in chronological order yields the slow-flow input:
[0100] Therefore, the downsampling step size M is not a fixed hyperparameter, but is dynamically given by the interval length relationship obtained from sensitivity diagnosis. It should be noted that average pooling is only a preferred implementation method, and the redundancy reduction processing can also be implemented using other compression or smoothing operators.
[0101] Subsequently, long-period feature extraction is performed on the slow-flow input. In this embodiment, the slow-flow input can be mapped to the frequency domain and divided into three channels—high frequency, mid frequency, and low frequency—based on the spectral amplitude distribution. After each channel is inversely transformed back to the time domain, it is reconstructed in two dimensions, and then long-period trend features are extracted through two-dimensional multi-scale convolution. It should be noted that the above-mentioned frequency domain three-channel partitioning and two-dimensional multi-scale convolution are only preferred implementation forms and do not constitute a limitation on the core solution of this invention.
[0102] 2) Fast Flow Branch
[0103] The fast flow branch is used to perform short-term perturbation modeling on the raw resolution data within the fast flow input interval.
[0104] In this embodiment, temporal features are extracted from the original resolution data corresponding to the fast-flow input interval to obtain short-time perturbation features. In a preferred implementation, one-dimensional convolution can be used:
[0105]
[0106] in, For convolution parameters, This is the activation function.
[0107] In this embodiment, the temporal feature extraction can be achieved by a lightweight bottleneck layer or a depthwise separable one-dimensional convolution to further reduce the number of parameters and computational cost. For example, a lightweight bottleneck layer may include a one-dimensional 1×1 dimensionality-reduced convolution, a depthwise separable one-dimensional convolution, a one-dimensional 1×1 dimensionality-upgrading convolution, batch normalization, and residual connections.
[0108] 3) Fusion decoder
[0109] Long-term trend characteristics With short-term disturbance characteristics The components are then fused to obtain fused features. In a preferred implementation, the fusion method can be a post-stitching projection:
[0110]
[0111] Then, the satellite clock bias prediction results for multiple future time periods are output through a direct multi-step prediction decoder:
[0112]
[0113] in, These are the projection matrix and the bias term, respectively. Besides the projection after stitching, the fusion method can also employ other feature fusion methods that do not deviate from the core idea of this invention.
[0114] Step S5: Train the satellite clock bias dual-stream prediction model using the training samples to obtain the trained satellite clock bias dual-stream prediction model; specifically:
[0115] The dual-stream prediction model is trained using the training samples constructed in step S1, and the model parameters are updated based on the error between the prediction results and the target prediction sequence. When the training loss function converges to below a preset loss threshold, the decrease in the validation set loss within a preset number of consecutive rounds is less than a preset tolerance, or the preset maximum number of training rounds is reached, the training of the satellite clock bias dual-stream prediction model is considered complete, resulting in a trained satellite clock bias dual-stream prediction model. Because a direct multi-step prediction method is used, the model can output prediction results for multiple future times simultaneously through a single forward propagation, thereby reducing the impact of error accumulation in recursive extrapolation.
[0116] like Figure 1 This embodiment discloses a satellite clock bias prediction method, which includes the following steps:
[0117] P1, obtain the historical satellite clock bias sequence before the time to be predicted, and call the parameters of the pre-set dual-stream division of labor rules and the pre-trained satellite clock bias dual-stream prediction model mentioned above;
[0118] P2, according to the dual-stream division of labor rule, extract the original resolution data corresponding to the fast stream input interval from the historical satellite clock difference sequence as the fast stream input, and construct the slow stream input according to the slow stream field of view length and compression processing rule;
[0119] P3, input the slow flow input and the fast flow input into the satellite clock bias dual-flow prediction model;
[0120] P4 extracts long-term trend features and short-term disturbance features using the slow-flow branch and the fast-flow branch respectively, and outputs satellite clock error prediction results for multiple future moments through a fusion decoder.
[0121] Other aspects of this embodiment can be found in the above embodiment of the method for constructing a satellite clock bias dual-stream prediction model based on model inversion and gradient sensitivity analysis.
[0122] To verify the effectiveness of the method of this invention, experimental samples were constructed using historical satellite clock bias sequence data, and the method of this invention was compared with LSTM, LSTM-Attention, and Transformer under the same experimental conditions. Evaluation metrics included prediction accuracy metrics RMSE and RE, and model complexity metrics such as number of parameters, computational cost, inference latency, and peak memory usage.
[0123] Figure 6 This is a comparison chart of prediction accuracy (arithmetic mean of data from five satellites). For example... Figure 6 As shown, based on the arithmetic mean of data from five satellites, the method of this invention maintains superior prediction accuracy under both 12-hour and 24-hour prediction settings. Particularly under the 24-hour long-horizon prediction condition, the average RMSE of the method of this invention is 0.2620 ns, significantly lower than LSTM's 2.6435 ns and LSTM-Attention's 0.8653 ns, and superior to Transformer's 0.3694 ns; the average RE is 1.6580 ns, lower than LSTM's 5.2532 ns, LSTM-Attention's 2.4781 ns, and Transformer's 1.6951 ns. Therefore, the method of this invention can effectively reduce prediction errors in long-horizon satellite clock bias prediction tasks. Figure 7 As shown, while maintaining good prediction accuracy, the method of this invention also has low model complexity. Specifically, the method of this invention has 4.954 M parameters, 0.599 G computational cost, and a peak GPU memory of 51.09 MB. Compared with Transformer, its parameter count, computational cost, and peak GPU memory are reduced by 52.90%, 96.63%, and 90.18%, respectively. The inference latency is 11.96 ms, which is similar to Transformer's 10.58 ms and significantly lower than LSTM and LSTM-Attention's 85.06 ms and 84.75 ms, respectively. This demonstrates that the method of this invention can effectively reduce model redundancy overhead while ensuring prediction accuracy, and has good engineering deployment value.
[0124] like Figure 8 As shown, this embodiment discloses a satellite clock bias prediction system, which includes the following modules:
[0125] Training sample construction and model calling module: used to obtain the historical satellite clock difference sequence before the time to be predicted, and call the parameters of the pre-set dual-stream division of labor rules and the pre-trained satellite clock difference dual-stream prediction model;
[0126] Dual-stream division of labor rule module: used to extract the original resolution data corresponding to the fast stream input interval from the historical satellite clock difference sequence as the fast stream input according to the dual-stream division of labor rule, and construct the slow stream input according to the slow stream field of view length and compression processing rules;
[0127] Input module: used to input the slow-flow input and the fast-flow input into the satellite clock bias dual-flow prediction model;
[0128] Prediction output module: It is used to extract long-term trend features and short-term disturbance features by using the slow flow branch and the fast flow branch respectively, and output the satellite clock difference prediction results for multiple future times through the fusion decoder.
[0129] Other aspects of this embodiment can be found in the above embodiment of the method for constructing a satellite clock bias dual-stream prediction model based on model inversion and gradient sensitivity analysis.
[0130] In summary, this invention is applicable to long-field-of-view satellite clock bias prediction scenarios in satellite navigation systems, especially for satellite clock bias sequences that simultaneously exhibit long-period trends and near-end high-frequency disturbances. Because this invention directly transforms gradient sensitivity diagnostic results from the complete historical field of view into flow splitting rules, and further dynamically determines the downsampling step size in the slow-flow branch based on the relationship between the compressed processing interval length and the fast-flow input interval length, it can achieve long-period information preservation, near-end disturbance modeling, and compression processing of intermediate low-contribution intervals within the same prediction framework.
[0131] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. All equivalent substitutions, improvements, and modifications made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for constructing a satellite clock error double-flow prediction model based on model inversion and gradient sensitivity analysis, characterized in that, Includes the following steps: S1. Obtain multiple sets of historical satellite clock error sequences as training samples to construct and train a full-view detection model; S2, freeze the network parameters of the full-view detection model after training in step S1, and perform model inversion and gradient backpropagation calculation on the input historical sequence to obtain the normalized sensitivity distribution of the future prediction output at each historical moment. S3. Determine the dual-stream division of labor rule based on the normalized sensitivity distribution obtained in step S2. The dual-stream division of labor rule includes at least the slow-stream field of view length, the fast-stream input interval, and the compression processing interval. S4. Based on the dual-stream division of labor rules determined in step S3, a satellite clock bias dual-stream prediction model is constructed. The satellite clock bias dual-stream prediction model includes a slow-stream branch, a fast-stream branch, and a fusion decoder. The slow-stream branch is used to extract long-term trend features from historical inputs within the slow-stream field of view. The fast-stream branch is used to extract short-term disturbance features from historical inputs within the fast-stream input interval. The fusion decoder is used to generate satellite clock bias prediction results for multiple future times based on the outputs of the slow-stream branch and the fast-stream branch. S5. The satellite clock bias dual-stream prediction model constructed in step S4 is trained using the training samples obtained in step S1. The satellite clock bias dual-stream prediction model is obtained after training.
2. The method of claim 1, wherein the method further comprises: In step S1: First-order difference is performed on the historical satellite clock difference sequence, and reversible instance normalization is performed on the input sequence after difference.
3. The method of claim 1, wherein the method further comprises: In step S2, the specific process of obtaining the normalized sensitivity distribution is as follows: the full-view detection model is placed in the evaluation state, and the input history sequence is set to a differentiable state; a scalar response function is constructed with the future prediction output, and the scalar response function is backpropagated with respect to the input history sequence using the input history sequence as the inversion object; the original sensitivity sequence is obtained based on the statistical value of the input gradient at each time step, and the original sensitivity sequence is normalized to obtain the normalized sensitivity distribution.
4. The method of claim 1, wherein the method further comprises: In step S4, the specific implementation of the satellite clock bias dual-stream prediction model includes: calculating the corresponding normalized sensitivity sequences for at least two different candidate historical input lengths; calculating the global mean and standard deviation for the normalized sensitivity sequence corresponding to each candidate historical input length, and determining the adaptive high sensitivity threshold based on the global mean and standard deviation; in the normalized sensitivity sequence corresponding to each candidate historical input length, extracting the near-end continuous high-sensitivity segments located at the tail end of the candidate historical input length and close to the current time, and whose sensitivity is continuously higher than the adaptive high sensitivity threshold, and extracting the segments located on the side far from the current time, and whose sensitivity is higher than the adaptive high sensitivity threshold. Continuous high-sensitivity segments at the far end that are continuously higher than the adaptive high-sensitivity threshold; continuous intermediate segments located between the continuous high-sensitivity segments at the far end and the continuous high-sensitivity segments at the near end, and whose average sensitivity is lower than the average sensitivity of the two aforementioned high-sensitivity segments, are identified as continuous low-sensitivity segments; the separation degree is calculated based on the average sensitivity of the continuous high-sensitivity segments at the far end, the continuous high-sensitivity segments at the near end, and the continuous low-sensitivity segments, and the candidate historical input length with the largest separation degree is selected as the slow-flow field of view length; the continuous high-sensitivity segments at the near end are identified as the fast-flow input interval, and the continuous low-sensitivity segments are identified as the compression processing interval.
5. The method of claim 4, wherein the method further comprises: In step S4, the adaptive high-sensitivity threshold is determined according to the following formula: in, This represents the global mean of the corresponding normalized sensitivity sequence. This represents the global standard deviation of the corresponding normalized sensitivity sequence.
6. The method of claim 1, wherein the method further comprises: In step S4, the processing of the slow flow branch includes: dynamically determining the downsampling step size based on the time span of the compression processing interval, so that the downsampling step size is positively correlated with the length of the compression processing interval; performing downsampling or smooth compression processing on the compression processing interval within the slow flow field of view to obtain compressed low-resolution segments; combining the low-resolution segments with the uncompressed segments within the slow flow field of view in chronological order to form the slow flow input; and performing long-period feature extraction on the slow flow input to obtain long-period trend features.
7. The method of claim 6, wherein the method further comprises: In step S4, the extraction of long-period trend features from the slow-flow input specifically includes: mapping the slow-flow input to the frequency domain and dividing it into three channels—high frequency, mid frequency, and low frequency—based on the spectral amplitude distribution; performing two-dimensional reconstruction after inverse transformation of the three channels back to the time domain; and extracting the reconstructed features using two-dimensional multi-scale convolution to obtain the long-period trend features.
8. The method of claim 1, wherein the method further comprises: The fast-flow branch processing in step S4 includes: extracting the original resolution data corresponding to the fast-flow input interval from the input history sequence as the near-end fast-flow input, and performing temporal feature extraction on the near-end fast-flow input to obtain the short-term perturbation features; wherein, the temporal feature extraction is achieved through one-dimensional convolution, lightweight bottleneck layer or depthwise separable one-dimensional convolution.
9. A method for predicting satellite clock error, characterized in that, Includes the following steps: P1, obtain the historical satellite clock bias sequence before the time to be predicted, and call the parameters of the pre-set dual-stream division of labor rules and the pre-trained satellite clock bias dual-stream prediction model constructed by the method described in any one of claims 1-8; P2, according to the dual-stream division of labor rule, extract the original resolution data corresponding to the fast stream input interval from the historical satellite clock difference sequence as the fast stream input, and construct the slow stream input according to the slow stream field of view length and compression processing rule; P3, input the slow flow input and the fast flow input into the satellite clock bias dual-flow prediction model; P4 extracts long-term trend features and short-term disturbance features using the slow-flow branch and the fast-flow branch respectively, and outputs satellite clock error prediction results for multiple future moments through a fusion decoder.
10. A system for predicting satellite clock errors, the system comprising: Includes the following modules: Training sample construction and model calling module: used to obtain the historical satellite clock difference sequence before the time to be predicted, and call the parameters of the pre-set dual-stream division rule and the pre-trained satellite clock difference dual-stream prediction model constructed by the method described in any one of claims 1-8; Dual-stream division of labor rule module: used to extract the original resolution data corresponding to the fast stream input interval from the historical satellite clock difference sequence as the fast stream input according to the dual-stream division of labor rule, and construct the slow stream input according to the slow stream field of view length and compression processing rules; Dual-stream input module: used to input the slow-stream input and the fast-stream input into the satellite clock bias dual-stream prediction model; Prediction output module: It is used to extract long-term trend features and short-term disturbance features by using the slow flow branch and the fast flow branch respectively, and output the satellite clock difference prediction results for multiple future times through the fusion decoder.