A method for online prediction of satellite operational status time series based on multi-expert dynamic fusion

Abstract

This invention discloses an online prediction method for satellite operational status time series based on multi-expert dynamic fusion, belonging to the field of online prediction and intelligent information processing technology for satellite operational status time series. To address the problems of insufficient adaptability to various changing patterns in satellite operational status time series prediction, difficulty in dynamically adjusting fixed fusion methods, and high computational burden in overall online updates, this invention constructs multiple offline expert prediction models, utilizes a gating network to generate fusion weights, dynamically weights and fuses multiple candidate operational status prediction results, and corrects the gating network parameters after the actual future operational status time series arrives. This reduces the computational load of overall online updates, improves continuous prediction adaptability, and is applicable to fields such as online prediction of satellite operational status time series, satellite telemetry and control status data processing, satellite operational status monitoring, and satellite operational control.

Description

Technical Field

[0001] This invention belongs to the field of online prediction of satellite operational status time series and intelligent information processing technology, specifically involving online prediction technology of satellite operational status time series based on multi-expert dynamic fusion. Background Technology

[0002] In existing technologies, satellite operational status time series are typically acquired continuously by satellite telemetry, tracking, and command (TT&C) systems or satellite status acquisition systems. The acquired data reflects the satellite's operational status variables such as position, velocity, acceleration, and attitude angle at different times. With the increase in the sampling frequency and status dimensions of TT&C data, online prediction of future operational status based on historical operational status time series has become an important part of satellite status monitoring, operational control, and risk warning technologies.

[0003] Currently, common methods for predicting satellite operational status time series fall into two categories. One is based on establishing a fixed prediction model according to pre-defined change patterns, such as fitting historical status data according to uniform, uniform acceleration, uniform precipitate, or periodic change patterns, and then using this to predict future status. The other category employs data-driven models such as Long Short-Term Memory networks and encoder-decoder networks to directly learn the mapping relationship between historical input windows and future status sequences. Some solutions also use multiple prediction models to output prediction results in parallel, and then use averaging, fixed weights, or model selection methods to form the final prediction result, thereby improving the prediction process's adaptability to different change patterns.

[0004] However, satellite operational status time series often exhibit characteristics such as smooth changes, periodic fluctuations, local disturbances, and state mode transitions during actual acquisition. Fixed prediction models rely on the assumption of a single change pattern, and when the current operational status does not conform to this assumption, the prediction error is prone to increase. Single data-driven models train by mixing samples from multiple operational statuses, making them susceptible to interference from different mode features. Conventional multi-model fusion methods often use fixed weights or simple averaging, making it difficult to dynamically adjust the contribution of each prediction model based on the operational status time series at the current prediction time. If all prediction models are updated directly after the actual future operational status time series arrives, it will increase the online computational burden and affect the real-time performance of the continuous prediction process.

[0005] Therefore, an online prediction scheme for satellite operational status time series needs to be proposed. In the offline stage, multiple expert prediction models are trained based on the changing patterns of different typical time series, and their parameters are fixed after training. At the current prediction time, a gating network is used to generate fusion weights for each offline expert prediction model based on the current operational status time series, and the prediction results of multiple candidate operational states are dynamically weighted and fused. After the actual future operational status time series arrives, the gating network is corrected only based on the prediction error, so that the fusion weights at subsequent prediction times can be adjusted according to the changes in the current operational status. Summary of the Invention

[0006] To address the shortcomings of existing technologies in satellite operational status time series prediction—namely, insufficient adaptability to various changing patterns, difficulty in dynamic adjustment of fixed fusion methods, and a high computational burden for overall online updates—this invention proposes the following solution: A method for online prediction of satellite operational status time series based on multi-expert dynamic fusion, the method comprising: S1. Obtain historical operational status time series data of the satellite collected by the satellite telemetry and control system or the satellite status acquisition system. The historical operational status time series data of the satellite is used to characterize at least one operational status quantity of the satellite in the historical input window, such as position, velocity, acceleration and attitude angle. Training sample sets corresponding to multiple expert prediction models are constructed according to different typical time series change patterns. S2. Train each expert prediction model based on the training sample set corresponding to the multiple expert prediction models, and fix the parameters of each expert prediction model after training to obtain multiple offline expert prediction models. S3. At the current prediction time, obtain the time series of the current operating status of the satellite within the historical input window; S4. Input the current operating state time series into multiple offline expert prediction models respectively to obtain multiple candidate operating state prediction results; S5. Input the current running state time series into the gating network, extract the temporal hidden features of the current running state time series, weight the temporal hidden features by importance to generate context features, and generate the fusion weights corresponding to the multiple offline expert prediction models based on the context features. S6. Dynamically weight and fuse the prediction results of the multiple candidate operating states according to the fusion weight to obtain the final future operating state time series prediction result corresponding to the current prediction time. S7. When the actual future operating state time series corresponding to the current prediction time arrives, a gated network is constructed to correct the loss based on the prediction error between the final future operating state time series prediction result and the actual future operating state time series. S8. Adjust the parameters of the gated network according to the gated network correction loss to obtain an updated gated network, and use the updated gated network to generate the fusion weights corresponding to the multiple offline expert prediction models at the next prediction time, wherein the parameters of the multiple offline expert prediction models remain unchanged.

[0007] Furthermore, the different typical time series change patterns mentioned in S1 include uniform change patterns, uniform acceleration change patterns, uniform jump change patterns, and periodic change patterns, and the multiple expert prediction models include expert prediction models corresponding to the uniform change patterns, uniform acceleration change patterns, uniform jump change patterns, and periodic change patterns, respectively.

[0008] Furthermore, the construction of the training sample set corresponding to multiple expert prediction models in S1 includes: dividing the time series samples of each type of satellite operation status into windows according to historical input windows and future prediction windows; using the continuous operation status quantities in the historical input window as expert training input sequences and the continuous operation status quantities in the future prediction window as expert training output sequences; and forming the training sample set of the corresponding expert prediction model by the expert training input sequences and the expert training output sequences.

[0009] Furthermore, the offline expert prediction model described in S2 adopts a sequence-to-sequence prediction structure, which includes an encoder and a decoder. The encoder is used to receive the expert training input sequence and extract the encoded features, and the decoder is used to gradually obtain the future running state prediction sequence based on the encoded features.

[0010] Furthermore, the step S2 of training each expert prediction model separately includes: inputting each type of training sample set into the corresponding expert prediction model, training the corresponding expert prediction model based on the mean square error between the training prediction sequence and the real training output sequence, and fixing the parameters of the corresponding expert prediction model after training is completed.

[0011] Furthermore, the gating network in S5 includes a long short-term memory network layer, an attention weight calculation layer, and a normalized weight generation layer. The long short-term memory network layer is used to extract the temporal hidden features of the current running state time series. The attention weight calculation layer is used to weight the temporal hidden features by importance to generate context features. The normalized weight generation layer is used to obtain the fusion weights corresponding to the multiple offline expert prediction models based on the context features.

[0012] Further, S5, generating the fusion weights corresponding to the multiple offline expert prediction models, includes: the gating network generating weight scores corresponding to each offline expert prediction model based on the current operating state time series, and normalizing the weight scores to obtain the fusion weights corresponding to each offline expert prediction model, and the sum of each fusion weight is one.

[0013] Furthermore, S6 dynamically weights and fuses the multiple candidate operating state prediction results according to the fusion weights, including: weighting and summing the candidate operating state prediction results obtained by each offline expert prediction model according to the fusion weights corresponding to each offline expert prediction model, to obtain the final future operating state time series prediction result.

[0014] Further, S8 adjusts the parameters of the gated network according to the gated network correction loss, including: performing gradient updates on the gated network according to the gated network correction loss to obtain an updated gated network, and preventing the multiple offline expert prediction models from participating in the gradient update of the gated network correction loss.

[0015] Based on the same inventive concept, the present invention also proposes a computer storage medium storing a computer program, which, when executed by a processor, implements the above-described method.

[0016] Compared with the prior art, the present invention has the following beneficial effects: By constructing training sample sets corresponding to multiple expert prediction models based on different typical time series change patterns, and training multiple offline expert prediction models respectively, each offline expert prediction model can form corresponding prediction capabilities for uniform change patterns, uniform acceleration change patterns, uniform precipitation change patterns, and periodic change patterns, thereby reducing the impact of mismatch between the single model change pattern assumption and the satellite operation status time series on the prediction results.

[0017] By fixing the parameters of each expert prediction model after training, and inputting the current running state time series into multiple offline expert prediction models at the current prediction time, multiple offline expert prediction models can output multiple candidate running state prediction results based on the same historical input window, thereby providing prediction basis under different change patterns for subsequent dynamic fusion.

[0018] By inputting the current running state time series into a gating network, and having the gating network generate fusion weights corresponding to multiple offline expert prediction models, the fusion weights can be adjusted according to the temporal characteristics of the current running state time series, thereby improving the problem that fixed weights or simple average fusion cannot adapt to the changes in the state at the current prediction time.

[0019] By dynamically weighting and fusing the prediction results of multiple candidate operating states according to the fusion weight, the offline expert prediction model that is more consistent with the current operating state time series can make a corresponding contribution to the final future operating state time series prediction results, thereby reducing the risk of model selection bias caused by hard selection of a single prediction model.

[0020] After the actual future operational status time series arrives, a gated network correction loss is constructed based on the prediction error between the final future operational status time series prediction result and the actual future operational status time series. The gated network parameters are then adjusted based on the gated network correction loss, so that the fusion weights of subsequent prediction times can be corrected according to the actual operational status time series that have arrived. This improves the adaptability to changes in satellite operational status during continuous online prediction.

[0021] By keeping the parameters of multiple offline expert prediction models unchanged during the adjustment of the gated network parameters, the online correction process only applies to the gated network and does not synchronously update all offline expert prediction models. This reduces the computational burden of the overall online update and maintains the prediction stability of the offline expert prediction models corresponding to typical change patterns.

[0022] This invention features the generation of candidate operational status prediction results based on multiple offline expert prediction models, dynamic allocation of fusion weights based on a gating network, and correction of gating network parameters based on real future operational status time series. It can improve the problems of insufficient adaptability of a single model, difficulty in dynamic adjustment of fixed fusion, and large computational burden of overall online update in satellite operational status time series prediction. It is applicable to fields such as online prediction of satellite operational status time series, satellite telemetry and control status data processing, satellite operational status monitoring, and satellite operation control. Attached Figure Description

[0023] Figure 1 This is a flowchart of the online prediction method for satellite operation status time series based on multi-expert dynamic fusion as described in the implementation method; Figure 2 This is a schematic diagram of the offline expert prediction model structure described in the implementation method; Figure 3 This is a schematic diagram of the online dynamic fusion prediction and gating network correction structure described in the implementation method; Figure 4 This is a graph showing the time series prediction results of the operating status as described in the implementation method. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.

[0025] Implementation Method 1 like Figure 1As shown, an online prediction method for satellite operational status time series based on multi-expert dynamic fusion is proposed, the method comprising: S1. Obtain historical operational status time series data of the satellite collected by the satellite telemetry and control system or the satellite status acquisition system. The historical operational status time series data of the satellite is used to characterize at least one operational status quantity of the satellite in the historical input window, such as position, velocity, acceleration and attitude angle. Training sample sets corresponding to multiple expert prediction models are constructed according to different typical time series change patterns. S2. Train each expert prediction model based on the training sample set corresponding to the multiple expert prediction models, and fix the parameters of each expert prediction model after training to obtain multiple offline expert prediction models. S3. At the current prediction time, obtain the time series of the current operating status of the satellite within the historical input window; S4. Input the current operating state time series into multiple offline expert prediction models respectively to obtain multiple candidate operating state prediction results; S5. Input the current running state time series into the gated network, extract the temporal hidden features of the current running state time series, weight the temporal hidden features by importance to generate context features, and generate the fusion weights corresponding to the multiple offline expert prediction models based on the context features. S6. Dynamically weight and fuse the prediction results of the multiple candidate operating states according to the fusion weight to obtain the final future operating state time series prediction result corresponding to the current prediction time. S7. When the actual future operating state time series corresponding to the current prediction time arrives, a gated network is constructed to correct the loss based on the prediction error between the final future operating state time series prediction result and the actual future operating state time series. S8. Adjust the parameters of the gated network according to the gated network correction loss to obtain an updated gated network, and use the updated gated network to generate the fusion weights corresponding to the multiple offline expert prediction models at the next prediction time, wherein the parameters of the multiple offline expert prediction models remain unchanged.

[0026] Furthermore, the different typical time series change patterns mentioned in S1 include uniform change patterns, uniform acceleration change patterns, uniform jump change patterns, and periodic change patterns, and the multiple expert prediction models include expert prediction models corresponding to the uniform change patterns, uniform acceleration change patterns, uniform jump change patterns, and periodic change patterns, respectively.

[0027] By assigning multiple expert prediction models to different typical time series change patterns, each expert prediction model can form candidate prediction capabilities for different operational state change characteristics, thereby providing a prediction basis under multiple change patterns for subsequent dynamic fusion.

[0028] Furthermore, the construction of the training sample set corresponding to multiple expert prediction models in S1 includes: dividing the time series samples of each type of satellite operation status into windows according to historical input windows and future prediction windows; using the continuous operation status quantities in the historical input window as expert training input sequences and the continuous operation status quantities in the future prediction window as expert training output sequences; and forming the training sample set of the corresponding expert prediction model by the expert training input sequences and the expert training output sequences.

[0029] By constructing expert training input sequences and expert training output sequences according to historical input windows and future prediction windows, the training process of each expert prediction model can be kept consistent with the subsequent online prediction process, thereby improving the adaptability of offline training results to the input data at the current prediction time.

[0030] Furthermore, the offline expert prediction model described in S2 adopts a sequence-to-sequence prediction structure, which includes an encoder and a decoder. The encoder is used to receive the expert training input sequence and extract the encoded features, and the decoder is used to gradually obtain the future running state prediction sequence based on the encoded features.

[0031] By employing a sequence-to-sequence prediction structure that includes an encoder and a decoder, the offline expert prediction model is able to extract temporal coding features from the expert training input sequence and gradually obtain the future running state prediction sequence, thereby enhancing the continuous output capability of candidate running state prediction results.

[0032] Furthermore, the step S2 of training each expert prediction model separately includes: inputting each type of training sample set into the corresponding expert prediction model, training the corresponding expert prediction model based on the mean square error between the training prediction sequence and the real training output sequence, and fixing the parameters of the corresponding expert prediction model after training is completed.

[0033] By training each expert prediction model based on the mean square error between the training prediction sequence and the real training output sequence, and fixing the parameters of the corresponding expert prediction model after training, the offline expert prediction model can maintain its ability to predict the corresponding typical change patterns, thereby providing a stable source of candidate predictions for the subsequent dynamic adjustment of the fusion weights of the gating network.

[0034] Furthermore, the gating network in S5 includes a long short-term memory network layer, an attention weight calculation layer, and a normalized weight generation layer. The long short-term memory network layer is used to extract the temporal hidden features of the current running state time series. The attention weight calculation layer is used to weight the temporal hidden features by importance to generate context features. The normalized weight generation layer is used to obtain the fusion weights corresponding to the multiple offline expert prediction models based on the context features.

[0035] By setting up a long short-term memory network layer, an attention weight calculation layer, and a normalized weight generation layer, the gating network can extract temporal hidden features from the current running state time series and form contextual features through importance weighting. This enables the fusion weights corresponding to multiple offline expert prediction models to be generated according to the running state change features at the current prediction time.

[0036] Further, S5, generating the fusion weights corresponding to the multiple offline expert prediction models, includes: the gating network generating weight scores corresponding to each offline expert prediction model based on the current operating state time series, and normalizing the weight scores to obtain the fusion weights corresponding to each offline expert prediction model, and the sum of each fusion weight is one.

[0037] By normalizing the weight scores of each offline expert prediction model, multiple fusion weights can form a unified weight allocation relationship, which facilitates the dynamic determination of the contribution of each candidate operating state prediction result to the final prediction result based on the current operating state time series.

[0038] Furthermore, S6 dynamically weights and fuses the multiple candidate operating state prediction results according to the fusion weights, including: weighting and summing the candidate operating state prediction results obtained by each offline expert prediction model according to the fusion weights corresponding to each offline expert prediction model, to obtain the final future operating state time series prediction result.

[0039] By weighting and summing the prediction results of multiple candidate operating states according to the fusion weights, the candidate prediction results of different offline expert prediction models can participate in the fusion according to the weight allocation of the current prediction time, thereby obtaining the final future operating state time series prediction result that is adapted to the current operating state time series.

[0040] Further, S8 adjusts the parameters of the gated network according to the gated network correction loss, including: performing gradient updates on the gated network according to the gated network correction loss to obtain an updated gated network, and preventing the multiple offline expert prediction models from participating in the gradient update of the gated network correction loss.

[0041] By updating the gradient of the gated network based on the gated network correction loss, and by excluding multiple offline expert prediction models from participating in the gradient update of the gated network correction loss, the online correction process is concentrated on the gated network. This allows the fusion weights for the next prediction time to be adjusted based on the actual future operating state time series that have been reached, while maintaining the stability of the offline expert prediction model parameters.

[0042] The method described in this embodiment can be executed by a processor calling a computer program, which can be stored in a computer storage medium. When the computer program is executed by the processor, the above-described online prediction method for satellite operational status time series based on multi-expert dynamic fusion can be implemented.

[0043] Implementation Method 2 Combination Figure 1 This embodiment provides an online prediction method for satellite operational status time series based on multi-expert dynamic fusion. This method is applicable to satellite operational status time series data continuously acquired by a satellite telemetry, tracking, and command system or a satellite status acquisition system. This satellite operational status time series data characterizes at least one operational status variable among the satellite's position, velocity, acceleration, and attitude angle at different times. The overall process of this method includes: acquiring historical satellite operational status time series data; constructing training sample sets corresponding to multiple expert prediction models; offline expert prediction model training; acquiring the operational status time series at the current prediction time; generating prediction results for multiple candidate operational states; generating fusion weights for a gating network; dynamically weighting and fusing the candidate operational state prediction results; and correcting the gating network parameters based on the actual future operational status time series.

[0044] The specific process of this invention is as follows.

[0045] Step 1: Obtain time series data of the satellite's historical operational status.

[0046] The historical operational status time series data of the satellite, acquired by the satellite telemetry, tracking, and command system or the satellite status acquisition system, is denoted as:

[0047] in, Indicates the first The time series values ​​at each time point. When the time series is a multidimensional sequence. Representing a state vector, for example:

[0048] in, This indicates the state dimension of the time series. Historical satellite operational state time series data is used as training samples to build offline expert models and also for subsequently generating the current operational state time series. Historical input window length. The permitted range is 10 to 200, with a prediction step size. The permitted range is 1 to 50. In this embodiment, we take... , .

[0049] Step 2: Construct training samples for multiple expert prediction models.

[0050] Training samples for multiple expert prediction models are constructed based on different time series variation patterns. These expert prediction models include those corresponding to the uniform velocity model, the uniform acceleration model, the uniform precipitate model, and the periodic model. The basic time series expressions for each model are as follows: Uniform velocity model:

[0051] Uniform acceleration model:

[0052] Uniform quiescent model:

[0053] Periodic model:

[0054] in, Indicates the first Time series values ​​at each moment; , , , , , , , , , , , and These are model parameters; This represents the disturbance term. The above model is used to generate or label time series samples of satellite operational status corresponding to different variation patterns.

[0055] For each type of satellite operational status time series sample, a window is divided to form an expert training input sequence and an expert training output sequence. The expert training input sequence is represented as follows:

[0056] The expert training output sequence is represented as:

[0057] in, Indicates the length of the history input window. Indicates the length of the future prediction window. This represents the current time in the training samples. Each training sample pair is represented as:

[0058] The permissible range for the division ratio of training set, validation set and test set is 6:2:2 to 9:0.5:0.5. In this embodiment, it is divided according to the ratio of 8:1:1.

[0059] Step 3: Train each expert prediction model separately to obtain multiple offline expert prediction models.

[0060] Input the training input sequence for each type of expert into the corresponding expert prediction model to obtain the training prediction sequence:

[0061] in, Indicates the first An expert prediction model, Indicates the first The expert prediction model outputs a future predicted sequence. The training loss of the expert model is calculated based on the error between the trained predicted sequence and the actual training output sequence.

[0062] in, Indicates the first The training loss of an expert prediction model. Indicates the first The actual time series value at each moment. Indicates the first The expert prediction model in the first... The predicted value at each time point, This indicates the prediction step size.

[0063] The expert prediction model employs a sequence-to-sequence prediction structure, comprising an encoder and a decoder. The encoder receives historical input sequences and extracts coded features, while the decoder progressively outputs future prediction sequences based on these coded features. The encoder mapping can be represented as:

[0064] The decoder mapping can be represented as:

[0065] in, This represents the encoded features output by the encoder. Indicates encoder mapping, Indicates the first Decoder mapping corresponding to an expert prediction model.

[0066] In this embodiment, both the encoder and decoder include one-dimensional convolutional layers and long short-term memory (LSM) network layers. The number of convolutional kernels in the one-dimensional convolutional layers is allowed to range from 4 to 64, and the kernel size is allowed to range from 2 to 7. The hidden layer dimension of the LSM network is allowed to range from 32 to 512. The number of training epochs is allowed to range from 20 to 300. The learning rate is allowed to range from 0.0001 to 0.01. In this embodiment, the number of convolutional kernels is 8, the kernel size is 3, the hidden layer dimension is 128, the learning rate is 0.001, and the Adam optimization algorithm is used. After training, the parameters of each expert prediction model are fixed to obtain multiple offline expert prediction models.

[0067] Step 4: Obtain the current running status time series.

[0068] At the current prediction time, obtain the satellite's operational status time series within the historical input window to form the current operational status time series:

[0069] in, Indicates the current prediction time. This indicates the length of the historical input window. The current running state time series is simultaneously input into multiple offline expert prediction models and gating networks to ensure that the candidate running state prediction results and fusion weights are generated based on the same input window.

[0070] Step 5: Generate multiple candidate running state prediction results.

[0071] The current running state time series is input into multiple trained offline expert prediction models to obtain multiple candidate running state prediction results:

[0072] in, This indicates the number of offline expert prediction models. The permitted range is 2 to 10; in this embodiment, . Indicates the first The candidate predicted sequences output by several offline expert prediction models. Different expert models correspond to different time series change patterns, therefore, multiple candidate operational state prediction results can provide future sequence estimates from the perspective of different evolutionary modes.

[0073] Step 6: Generate fusion weights through a gating network.

[0074] The current running state time series is input into the gating network. The gating network consists of a long short-term memory network layer, an attention weight calculation layer, and a normalized weight generation layer.

[0075] Long Short-Term Memory (LSTM) network layers are used to extract temporal hidden features of the current running state time series, denoted as:

[0076] in, Indicates the first entry in the history input window The temporal hidden features corresponding to each time point.

[0077] The attention weight calculation layer weights the hidden features at different historical moments to generate contextual features:

[0078] in, Representing contextual features, Indicates the first Attention weights corresponding to each historical moment, satisfying the following:

[0079] Normalized weight generation layer based on context features Output the fusion weights corresponding to multiple offline expert prediction models:

[0080] in, Indicates the first The fusion weights correspond to the offline expert prediction models. The fusion weights satisfy:

[0081] In this embodiment, the fusion weights are generated using a normalization function:

[0082] in: , Indicates the first The weight score corresponding to each expert model and This represents the trainable parameters of the normalized weight generation layer. The allowed range for the hidden layer dimension of the gated network is 32 to 512, the allowed range for the hidden layer dimension of the attention layer is 16 to 256, and the allowed range for the online update learning rate is 0.0001 to 0.1. In this embodiment, the hidden layer dimension of the gated network is 128, the hidden layer dimension of the attention layer is 64, and the online update learning rate is 0.001.

[0083] Step 7: Dynamically weight and fuse the prediction results of multiple candidate operating states.

[0084] Based on the fusion weights output by the gating network, the candidate operating state prediction results output by multiple offline expert prediction models are weighted and fused to obtain the final future operating state time series prediction result corresponding to the current prediction time:

[0085] in, This represents the final future time series prediction result output by this invention. Indicates the first The fusion weights of offline expert prediction models Indicates the first The fusion process involves the output of candidate predictions from several offline expert prediction models. Instead of simply averaging the results from multiple models, the gating network dynamically determines the contribution of each expert model based on the changing characteristics of the current input sequence.

[0086] Step 8: Update the gated network parameters based on the true future sequence.

[0087] When the real future time series Upon arrival, calculate the error between the final predicted result and the actual future time series:

[0088] in, This indicates the loss from online updates to the gating network. Represents the true future time series value. This represents the fusion prediction result. The gated network parameters are updated based on this loss:

[0089] in, Indicates the parameters of the gating network. Indicates the learning rate. This represents the gradient of the loss function with respect to the gating network parameters. In this step, the parameters of multiple offline expert prediction models remain unchanged; only the gating network parameters are updated to obtain the updated gating network. At the next prediction time, the updated gating network is used to generate the fusion weights corresponding to the multiple offline expert prediction models. This reduces the computational cost of online updates and enables the system to quickly adjust the fusion weights of different expert models based on newly arrived data.

[0090] In one specific embodiment, the current operational status time series of an unknown type of three-dimensional satellite is used as the current operational status time series. This time series is composed of... , , It consists of three state components and does not belong to any single type of model, including uniform velocity, uniform acceleration, uniform precipitate, and periodicity. Instead, it is composed of trend changes, periodic fluctuations, and local abrupt changes. Its expression is as follows:

[0091] in , The local mutation function is:

[0092] The above parameters can form a composite time series that simultaneously exhibits smooth trends, local periodic fluctuations, and abrupt enhancement characteristics, which can be used to verify the online prediction capability of this embodiment for time series of unknown satellite operational status.

[0093] In this embodiment, the offline expert prediction models include four types: expert prediction models corresponding to uniform velocity variation patterns, expert prediction models corresponding to uniform acceleration variation patterns, expert prediction models corresponding to uniform precipitous variation patterns, and expert prediction models corresponding to periodic variation patterns. Each type of expert prediction model is trained based on time series samples of the corresponding satellite operational status. The training samples are divided into a training set, a validation set, and a test set in an 8:1:1 ratio. This ratio ensures a sufficient number of training samples while retaining a certain number of validation and test samples for monitoring the model's training status and prediction generalization performance. If the training set ratio is too low, the expert prediction model will struggle to fully learn the corresponding time series variation patterns; if the training set ratio is too high, the validation and test sets will be insufficient, making it difficult to evaluate the model's generalization performance.

[0094] The historical input window length ranges from 30 to 100, and the prediction step size ranges from 5 to 30. In this embodiment, the historical input window length is... Predicting step size .when When the input sequence is too small, it contains insufficient historical information, making it difficult to reflect trend changes, cyclical fluctuations, and states before and after abrupt changes; when... When the value is too large, the amount of redundant information in the input sequence increases, which increases the computational load and reduces the efficiency of online updates. This approach balances the effectiveness of historical feature extraction with computational efficiency. Prediction step size. If the time frame is too small, only prediction results within a short time range can be obtained; When the value is too large, prediction errors tend to accumulate. Take... At the same time, it can meet the needs of short-term multi-step prediction and effectively control error accumulation.

[0095] Offline expert prediction models employ, for example Figure 2The encoder-decoder structure shown represents a sequence-to-sequence prediction model. Both the encoder and decoder contain one-dimensional convolutional layers and long short-term memory (LSTM) network layers. The number of convolutional kernels ranges from 4 to 32, the kernel size ranges from 2 to 5, and the hidden layer dimension ranges from 64 to 256. In this embodiment, the number of convolutional kernels is 8, the kernel size is 3, and the hidden layer dimension is 128. Too few convolutional kernels result in insufficient extraction of local variation features; too many kernels increase the number of model parameters and training time. A kernel size of 3 allows for better extraction of local variation relationships between adjacent time points. A hidden layer dimension of 128 achieves a good balance between model expressive power and computational complexity.

[0096] The offline expert prediction model uses mean squared error as its loss function and Adam as its optimization algorithm. The initial learning rate ranges from 0.0001 to 0.01. In this embodiment, the initial learning rate is set to 0.001. A learning rate that is too small results in slow model convergence; a learning rate that is too large leads to oscillations during training and difficulty in achieving stable convergence. A learning rate of 0.001 allows the model to steadily reduce training error within a shorter number of training epochs. The number of offline training epochs ranges from 50 to 200; in this embodiment, 100 is used. Too few training epochs result in underfitting of the expert prediction model; too many training epochs increase training time and may lead to overfitting. A learning rate of 100 allows the expert prediction model to learn the changing patterns of the corresponding type of time series more fully.

[0097] At the current prediction time, the historical input window of the satellite's current operational status time series will be used. The algorithm simultaneously inputs four offline expert prediction models and a gating network. The four offline expert prediction models output candidate running state prediction results, and the gating network outputs four fusion weights. And satisfy:

[0098] The final time series prediction results for the future operating state are expressed as follows:

[0099] in, Indicates the first Candidate prediction results of an expert model This represents the fused future time series prediction results.

[0100] like Figure 3As shown, the gating network consists of a Long Short-Term Memory (LSTM) network layer, an attention weight calculation layer, and a normalized weight generation layer. The hidden layer dimension of the gating network ranges from 64 to 256, and the hidden layer dimension of the attention layer ranges from 32 to 128. In this embodiment, the hidden layer dimension of the gating network is 128, and the hidden layer dimension of the attention layer is 64. When the hidden layer dimension is small, the gating network's ability to express the time series features of the current running state is insufficient; when the hidden layer dimension is too large, the computational load of the gating network correction increases. Using 128 and 64, the matching relationship between the current running state time series and each expert prediction model can be identified well, while maintaining high parameter correction efficiency.

[0101] After the actual future operating state time series corresponding to the current prediction time arrives, the gating network parameters are corrected based on the prediction error, while keeping the parameters of the four offline expert prediction models unchanged. The online learning rate ranges from 0.0001 to 0.01, and is set to 0.001 in this embodiment. This value ensures that the gating network gradually adjusts the weights of the expert prediction models according to the newly arrived actual future operating state time series, while avoiding drastic fluctuations in weights. The gating network correction loss uses the mean squared error:

[0102] in, Represents the true future time series value. This represents the fused prediction value. This indicates the prediction step size.

[0103] like Figure 4 As shown in the prediction results of this embodiment, the four offline expert prediction models each provide different candidate prediction sequences. The expert prediction model that best matches the current operational status time series changes receives a higher fusion weight. However, since the operational status time series of unknown satellite types simultaneously includes trend changes, periodic fluctuations, and local abrupt changes, a single expert prediction model still cannot fully describe its future changes. After dynamically allocating weights through a gating network, the multi-expert fusion prediction result is closer to the actual time series than the prediction result of any single expert prediction model.

[0104] Therefore, when the historical input window length, prediction stride, number of convolutional kernels (8), kernel size (3), hidden layer dimension (128), attention layer hidden dimension (64), and learning rate (0.001) are set, this implementation can achieve good prediction results while maintaining computational efficiency. This embodiment demonstrates that it is suitable for situations where the current operational status time series does not belong to any single expert prediction model type. It can improve the online prediction accuracy, stability, and interpretability of unknown satellite operational status time series through multi-expert candidate prediction and dynamic fusion of gating networks.

[0105] The embodiments of the present invention have been described above. These embodiments are used to help understand the technical solutions of the present invention and should not be construed as limiting the scope of the technical solutions of the present invention. Equivalent substitutions or conventional modifications made by those skilled in the art to the embodiments without departing from the technical concept of the present invention should all fall within the scope of the technical solutions of the present invention.

Claims

1. A method for online prediction of satellite operational status time series based on multi-expert dynamic fusion, characterized in that, The method includes: S1. Obtain historical operational status time series data of the satellite collected by the satellite telemetry and control system or the satellite status acquisition system. The historical operational status time series data of the satellite is used to characterize at least one operational status quantity of the satellite in the historical input window, such as position, velocity, acceleration and attitude angle. Training sample sets corresponding to multiple expert prediction models are constructed according to different typical time series change patterns. S2. Train each expert prediction model based on the training sample set corresponding to the multiple expert prediction models, and fix the parameters of each expert prediction model after training to obtain multiple offline expert prediction models. S3. At the current prediction time, obtain the time series of the current operating status of the satellite within the historical input window; S4. Input the current operating state time series into multiple offline expert prediction models respectively to obtain multiple candidate operating state prediction results; S5. Input the current running state time series into the gated network, extract the temporal hidden features of the current running state time series, weight the temporal hidden features by importance to generate context features, and generate the fusion weights corresponding to the multiple offline expert prediction models based on the context features. S6. Dynamically weight and fuse the prediction results of the multiple candidate operating states according to the fusion weight to obtain the final future operating state time series prediction result corresponding to the current prediction time. S7. When the actual future operating state time series corresponding to the current prediction time arrives, a gated network is constructed to correct the loss based on the prediction error between the final future operating state time series prediction result and the actual future operating state time series. S8. Adjust the parameters of the gated network according to the gated network correction loss to obtain an updated gated network, and use the updated gated network to generate the fusion weights corresponding to the multiple offline expert prediction models at the next prediction time, wherein the parameters of the multiple offline expert prediction models remain unchanged.

2. The method according to claim 1, characterized in that, The different typical time series variation patterns mentioned in S1 include uniform variation patterns, uniform acceleration variation patterns, uniform jump variation patterns, and periodic variation patterns. The multiple expert prediction models include expert prediction models corresponding to the uniform variation patterns, uniform acceleration variation patterns, uniform jump variation patterns, and periodic variation patterns, respectively.

3. The method according to claim 1, characterized in that, S1 describes the construction of a training sample set corresponding to multiple expert prediction models, which includes: dividing the time series samples of each type of satellite operation status into windows according to historical input windows and future prediction windows; using the continuous operation status quantities in the historical input window as expert training input sequences and the continuous operation status quantities in the future prediction window as expert training output sequences; and forming the training sample set of the corresponding expert prediction model by the expert training input sequences and the expert training output sequences.

4. The method according to claim 1, characterized in that, The offline expert prediction model described in S2 adopts a sequence-to-sequence prediction structure, which includes an encoder and a decoder. The encoder is used to receive the expert training input sequence and extract the encoded features, and the decoder is used to gradually obtain the future running state prediction sequence based on the encoded features.

5. The method according to claim 1, characterized in that, S2 describes training each expert prediction model separately, including: inputting each type of training sample set into the corresponding expert prediction model, training the corresponding expert prediction model based on the mean square error between the training prediction sequence and the real training output sequence, and fixing the parameters of the corresponding expert prediction model after training is completed.

6. The method according to claim 1, characterized in that, The gated network described in S5 includes a long short-term memory network layer, an attention weight calculation layer, and a normalized weight generation layer. The long short-term memory network layer is used to extract the temporal hidden features of the current running state time series. The attention weight calculation layer is used to weight the temporal hidden features by importance to generate context features. The normalized weight generation layer is used to obtain the fusion weights corresponding to the multiple offline expert prediction models based on the context features.

7. The method according to claim 1, characterized in that, S5 The generation of fusion weights corresponding to the multiple offline expert prediction models includes: the gating network generates weight scores corresponding to each offline expert prediction model based on the current operating state time series, and normalizes the weight scores to obtain the fusion weights corresponding to each offline expert prediction model, and the sum of each fusion weight is one.

8. The method according to claim 1, characterized in that, S6 performs dynamic weighted fusion of the multiple candidate operating state prediction results according to the fusion weight, including: weighting and summing the candidate operating state prediction results obtained by each offline expert prediction model according to the fusion weight corresponding to each offline expert prediction model, to obtain the final future operating state time series prediction result.

9. The method according to claim 1, characterized in that, S8 adjusts the parameters of the gated network according to the gated network correction loss, including: performing gradient updates on the gated network according to the gated network correction loss to obtain an updated gated network, and ensuring that the multiple offline expert prediction models do not participate in the gradient update of the gated network correction loss.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 9.

Images