A multi-task processing system and method for multi-modal enhanced satellite telemetry data

Abstract

The present invention discloses a multi-modal enhanced satellite telemetry data multi-task processing system and method. The system includes an input embedding module, a knowledge-driven prompt enhancement module, a fusion perception module, a large language model processing module, a feature enhancement module, a time series model processing module, and a task execution and optimization module. The present invention utilizes the knowledge-driven prompt enhancement module to perform knowledge graph enhancement on initial instructions in natural language mode to generate optimized instructions. The fusion perception module is used to fuse the natural language mode of the optimized instructions with the time series mode of satellite telemetry data. The large language model processing module is used to mine mixed features in the fused data. The time series model processing module is used to derive the final prediction result to achieve prediction and anomaly detection of satellite telemetry data. The present invention realizes efficient processing and accurate analysis of satellite telemetry data under the technical framework of integrating the large language model, the large time series model, and the knowledge graph.

Description

Technical Field

[0001] The present invention belongs to the field of satellite telemetry technology, and in particular relates to a multi-modal enhanced satellite telemetry data multi-task processing system and method. Background Art

[0002] Accurate prediction and timely anomaly detection of satellite telemetry data are key to maintaining stable satellite operations. However, existing satellite telemetry data processing technologies suffer from problems such as single processing tasks, isolated data modalities, and shallow knowledge graph fusion. As a result, the satellite telemetry data prediction accuracy is unable to meet actual application requirements, as shown in the following:

[0003] 1. Insufficient information mining and multimodal capabilities.

[0004] Current satellite telemetry data processing models are mostly limited to single-modal data analysis, or focus on satellite image analysis, or concentrate on anomaly and fault diagnosis of telemetry data of a satellite subsystem. There is a lack of systematic mining and integration of semantic background information such as satellite type and orbital environment. This single-task approach does not fully learn the inherent connections between multi-source data, so that data understanding only stays at the surface level and it is impossible to build a complete satellite operation status cognition system. For example, when processing multi-parameter telemetry data such as voltage, current, and temperature, traditional methods fail to fully consider the coupling relationship between parameters, changes in equipment status, and the impact of satellite mission execution. As a result, the accuracy of data prediction and the comprehensiveness of anomaly detection are not high enough, making it difficult to achieve in-depth operation and control and accurate analysis of the multi-task operation status of satellites.

[0005] 2. Weak technical adaptability and generalization capabilities.

[0006] Existing satellite telemetry data processing methods rely heavily on specific satellite timing models and mission scenario knowledge. This leads to significant limitations in versatility and flexibility when faced with the diverse telemetry data processing requirements and complex, ever-changing timing tasks of different satellite types. Due to significant differences in satellite design architecture, functional characteristics, and operating environments, traditional processing methods struggle to quickly adapt to the multi-parameter telemetry data processing requirements of diverse satellites. The complex model training and application processes severely restrict processing efficiency, making it impossible to meet the urgent need for real-time processing and analysis of massive amounts of telemetry data, and thus unable to support efficient satellite operations management.

[0007] 3. The depth of knowledge graph integration and application efficiency are poor.

[0008] As an important tool for enriching the contextual information of satellite telemetry data, knowledge graphs are still in their infancy in existing processing technologies. Most methods simply overlay knowledge graphs as auxiliary information, lacking a deep integration of satellite telemetry time series features and text features. When constructing knowledge graphs covering areas such as satellite equipment status and energy systems, there are challenges such as high construction costs and difficult updating and maintenance, making it difficult to adapt to the dynamic processing needs of multimodal data. Due to the inability to update the impact of factors such as equipment aging and space environment changes on multiple parameters in real time, and the difficulty in achieving deep integration and analysis of knowledge graph semantic information and telemetry data, it is unable to provide substantial technical support for satellite telemetry data prediction and anomaly detection, which in turn affects the scientific nature and accuracy of satellite operation management decisions. Summary of the Invention

[0009] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a multi-modal enhanced satellite telemetry data multi-task processing system and method, so as to simultaneously realize satellite telemetry data prediction, anomaly detection and missing data completion with one system.

[0010] In order to achieve the above object, the technical solution adopted by the present invention is:

[0011] A multi-modal enhanced satellite telemetry data multi-task processing system includes an input embedding module, a knowledge-driven prompt enhancement module, a fusion perception module, a large language model processing module, a feature enhancement module, a time series model processing module, and a task execution and optimization module.

[0012] The input embedding module is used to perform normalization and segmentation processing on satellite telemetry data.

[0013] The knowledge-driven prompt enhancement module is used to perform knowledge graph enhancement on the initial instructions in the natural language modality to generate optimized instructions.

[0014] The fusion perception module is used to fuse the natural language modality of the optimized instruction and the time series modality of the satellite telemetry data.

[0015] The large language model processing module is used to mine the mixed features in the fused data. The large language model processing module of the present invention is mainly implemented based on the large language model (LLM).

[0016] The feature enhancement module is used to perform dimension mapping and enhancement on the mixed features.

[0017] The time series model processing module is used to derive the final prediction result based on the features enhanced by the feature enhancement module, thereby realizing the prediction and anomaly detection of satellite telemetry data. The time series model processing module of the present invention is mainly implemented based on the large time series model (LTSM).

[0018] The task execution and optimization module is used to identify the task intent according to the instructions, execute one or more of the satellite telemetry data prediction, satellite telemetry data completion and satellite anomaly detection tasks, and perform system training optimization.

[0019] Accordingly, the present invention also provides a multi-task processing method for satellite telemetry data with multi-modal enhancement, comprising the following steps:

[0020] Step 1: Normalize and segment the satellite telemetry data, perform knowledge graph enhancement on the initial instructions in the natural language modality, and generate optimized instructions.

[0021] Step 2: Fusion the natural language modality of the optimized command and the time series modality of the satellite telemetry data.

[0022] Step 3: Use the large language model processing module to supplement the temporal features of the satellite telemetry data according to the prompt words of the optimized instructions, and mine the mixed features in the fused data.

[0023] Step 4: Dimension mapping and enhancement of the mixed features.

[0024] In step 5, the time series model processing module is used to derive the final prediction results based on the enhanced features to achieve prediction and anomaly detection of satellite telemetry data.

[0025] Step 6: Determine the task intent based on the prompt word in the user input instruction, and execute one or more of the satellite telemetry data prediction task, satellite telemetry data completion task, and satellite anomaly detection task.

[0026] Compared with existing technologies, this invention achieves efficient processing and accurate analysis of satellite telemetry data under the technical framework of integrating large language models (LLMs), large time series models (LTSMs), and knowledge graphs. The specific advantages are as follows:

[0027] First, we achieve deep fusion analysis of multimodal data. We deeply integrate multi-parameter satellite telemetry data with natural language descriptions and knowledge graph information, covering the entire process from data preprocessing and feature fusion to model prediction. This breaks the limitations of existing technologies that rely on a single data modality, deeply explores the complex relationships between data, and provides rich and accurate information for satellite telemetry data prediction and anomaly detection, comprehensively and accurately reflecting the satellite's operational status.

[0028] Second, we enhance multitasking capabilities. Through the task execution and optimization module, we intelligently and efficiently handle diverse tasks such as parameter prediction, data completion, and anomaly detection. We flexibly adjust model parameters and loss functions based on task characteristics, changing the traditional single-task approach and ensuring high accuracy and efficiency across all tasks.

[0029] Third, we will deepen the integration and application of knowledge graphs. By using methods such as knowledge graph retrieval and instruction optimization, we can decouple knowledge information from model processing, achieving "plug-and-play" knowledge graphs. This addresses the inadequate integration of knowledge graphs with satellite telemetry data in existing technologies. This allows for rapid adjustment and expansion of system functionality based on satellite type, mission scenario, and knowledge update requirements, enhancing system flexibility and scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is an architecture diagram of the satellite telemetry data multi-task processing system of the present invention.

[0031] Figure 2 This is a structural diagram of the knowledge-driven prompt enhancement module of the present invention.

[0032] Figure 3 This is the structural diagram of the fusion perception module of the present invention. DETAILED DESCRIPTION

[0033] The embodiments of the present invention are described in detail below with reference to the accompanying drawings and examples.

[0034] The present invention aims to provide a multi-task processing system for satellite telemetry data that integrates a large language model (LLM), a large time series model (LTSM), and a knowledge graph to achieve efficient and accurate processing of satellite telemetry data, and comprehensively improve the performance and versatility of the system in tasks such as satellite telemetry data prediction and anomaly detection.

[0035] As attached Figure 1 As shown, the system of the present invention mainly includes an input embedding module, a knowledge-driven prompt enhancement module, a fusion perception module, a large language model processing module, a feature enhancement module, a time series model processing module, and a task execution and optimization module. Each part is described in detail as follows:

[0036] 1. Input embed module.

[0037] The input embedding module is primarily used to normalize and segment incoming satellite telemetry data. Based on processing requirements, it processes satellite telemetry data containing information such as time, voltage, current, temperature, device status, and power consumption, providing standardized data and valid command input for subsequent steps. Its specific implementation is as follows:

[0038] First, the satellite telemetry data, including time data (ground time and onboard time accurate to seconds and milliseconds), voltage parameters (such as bus voltage, battery pack voltage, lithium-ion single cell voltage1-7, etc.), current parameters (charging current, discharging current, etc.), temperature-related data (power controller internal temperature, +5V temperature reference voltage, etc.), and power information (battery pack current power, charging power, etc.), are normalized as non-stationary time series to make the data mean zero and the standard deviation unity.

[0039] Next, the satellite time series data from multiple sources are merged into a univariate sequence pool, and the univariate sequence time points for training are selected according to the normal distribution, and then divided into several continuous overlapping or non-overlapping fragments.

[0040] 2. Knowledge-driven prompt enhancement module.

[0041] The knowledge-driven prompt enhancement module integrates information retrieval and optimized instruction generation based on knowledge retrieval enhancement technology, as well as prompt generation based on knowledge graphs, providing semantic enhancement support for multi-parameter satellite telemetry data processing. It is mainly used to perform knowledge graph enhancement on initial instructions in natural language modalities and generate optimized instructions, providing rich background knowledge and accurate instruction guidance for subsequent data processing. Its specific implementation method is as follows:

[0042] First, a knowledge-driven augmentation model is used to extract keywords from the initial instructions. Knowledge from specialized external knowledge databases closely related to satellite telemetry data, such as the satellite power system knowledge base and the satellite thermal control system database, is organized into a graph structure. Specifically, this invention improves the graph structure using GraphRAG technology.

[0043] Then, according to the initial instructions of the natural language mode input by the user for satellite telemetry data processing, a graph search is performed, and the retrieved information is deeply integrated with the initial instructions to generate optimized instructions, as shown in the attached figure. Figure 2 Specifically, in the embodiment of the present invention, information is retrieved from an external knowledge database and combined with the aforementioned keywords to generate optimized instructions.

[0044] The present invention improves GraphRAG technology including:

[0045] Before building the knowledge graph, identify the entity type according to the scenario and save it;

[0046] When building the knowledge graph, entities are extracted based on type and duplicate entities are removed;

[0047] Add descriptions of entities to the knowledge graph;

[0048] Locate information in the knowledge graph based on keyword matching during retrieval.

[0049] Ultimately, by defining standardized data processing rules, we ensure that the generated optimization instructions meet standardization and applicability requirements in terms of data format, semantic expression, etc., providing strong support for subsequent in-depth data processing by fusion perception modules, large language model processing modules, etc.

[0050] 3. Fusion perception module.

[0051] Based on the input embedding module and the knowledge-driven prompt enhancement module, the natural language modality of the optimized instructions and the time series modality of the satellite telemetry data are deeply integrated through the fusion perception module.

[0052] The specific implementation method is as follows Figure 3 As shown, the optimized instruction embeddings are first mean-pooled to generate a global hint representation of the initial semantic vector. For example, for a long semantic vector (e, 768), where e represents the length of the natural language text and 768 represents the dimension of the vector embedding, the resulting global hint representation is (1, 768), equivalent to representing the entire semantics with a single, unique vector. This is then repeatedly expanded to align with the length of the time series features of the satellite telemetry data. Multi-scale convolution operations are then performed on the time series features of the satellite telemetry data (such as the time-varying characteristics of voltage and current) using convolution kernels of different scales to extract multi-level features. The expanded global hint representation is concatenated with the multi-level features, and learnable fusion weights are introduced. A weighted residual connection is then performed through linear layer fusion to generate a hybrid feature. Finally, this hybrid feature is concatenated with the initial semantic vector to generate the fused data. This module effectively preserves the integrity of satellite time series data while capturing semantic information, improving the model's ability to represent complex data patterns.

[0053] The specific implementation methods involved in the above process can be described as follows:

[0054] Mean Pooling:

[0055]

[0056] in, is the global hint representation of the initial semantic vector, is the prompt feature at time t, is the total length of instruction embedding after optimization, P is word embedding, W is weight, is the bias term of the fully connected layer, is the normalized exponential function.

[0057] Repeat extension:

[0058]

[0059] in, is the length of the time series characteristics of satellite telemetry data, Indicates a copy operation, which copies the global prompt to a length of , which facilitates the fusion of hidden dimensions, It means copying n copies and merging them into a semantic vector representation of length n.

[0060] Multi-scale convolution:

[0061]

[0062]

[0063]

[0064] in, 、 is the feature extracted by different convolution scales, F It is a multi-level feature obtained by weighted fusion of the feature attention module, and p is a convolution parameter to keep the length unchanged after extraction. is a one-dimensional convolution, is the time series characteristic of satellite telemetry data, 、 are the convolution kernel sizes, is the parameter for vector dimension splicing, is the splicing function.

[0065] Feature fusion:

[0066]

[0067] Among them, C is the initial mixed feature obtained by splicing, Is F The output features obtained after the fully connected layer are and Fusion splicing together to ensure the same length.

[0068] Through learnable fusion weights The initial mixed feature C obtained by splicing is subjected to weighted residual connection optimization fusion effect to obtain the final mixed feature F.

[0069] 4. Large language model processing module.

[0070] The fused data generated by the fusion perception module is fed into the large language model processing module for enhanced processing. This module leverages the large language model's powerful language understanding and generation capabilities, and performs semantic-level analysis and processing to extract more abstract semantic features. Specifically, based on the pre-trained large language model, it uses prompts from optimized instructions to supplement the time series features of satellite telemetry data. This translates the numerical features of the time series (such as trends, periodicity, lags, maximums, and minimums) into natural language semantics, building a bridge between the pre-trained knowledge of the large language model and the time series task. Based on the global prompt representation of the initial semantic vector, mixed features are mined from the fused data. The semantic vector is embedded as a prefix into the input sequence, and a self-attention mechanism guides the large language model to focus on key features within the mixed features.

[0071] Since the large language model has been pre-trained, the model parameters are not updated at this stage to reduce the amount of computation and avoid overfitting.

[0072] 5. Feature enhancement module.

[0073] The mixed features output by the large language model processing module are further processed, discarding the prefix hints (i.e., the hidden features of the hint word) and retaining only the time series features. These time series features are then dimensionally mapped and enhanced to make them more suitable for input into the pre-trained large time series model (LTSM). This module adjusts the dimension and representation of features based on different task requirements, enhancing the model's adaptability to multi-parameter satellite telemetry data.

[0074] 6. Timing model processing module.

[0075] This module is based on a frozen, pre-trained large-scale time series model and includes a time series decoder. It receives the enhanced features output by the feature enhancement module and derives the final prediction results, enabling forecasting and anomaly detection for satellite telemetry data. For forecasting multi-parameter satellite time series data, it can predict the values ​​of satellite parameters (such as voltage, current, and power) at future times. For anomaly detection, it determines whether the current data is abnormal by comparing it with the normal pattern.

[0076] 7. Task execution and optimization module.

[0077] Identify task intent based on instructions, execute corresponding tasks according to the requirements of different tasks (parameter prediction, data completion, anomaly detection), and perform system training and optimization.

[0078] In the parameter prediction task, autoregression is used to generate training targets, the mean squared error between the predicted value and the true value is calculated, and the loss is back-propagated; in the data completion task, the reconstruction error between the mask segment generated item and the true value is calculated; in the anomaly detection task, the mean squared error between the predicted time series and the true time series is calculated as the anomaly score.

[0079] Specifically, the prediction of satellite telemetry data is implemented as follows:

[0080] In the fine-tuning stage, autoregression is used to generate training targets, and the total length of the optimized instruction embedding is , divide the time step L of historical satellite telemetry data into items, the pre-trained large time series model outputs the next item ; Calculate the mean square error (MSE) of each item with the corresponding satellite actual telemetry data and back-propagate the loss, the loss function for:

[0081]

[0082] in, is the predicted value of satellite telemetry data at time t, is the true value of the satellite telemetry data at time t.

[0083] In the inference phase, the prediction results are connected with the input satellite telemetry data time series, and the pre-trained large time series model is repeatedly used to generate the next item until the total length of the predicted item reaches the expected length, thereby realizing the prediction of satellite telemetry data.

[0084] Specifically, the anomaly detection method for satellite telemetry data is as follows:

[0085] Based on the idea of ​​predictive anomaly detection, the observed satellite telemetry data time series is used to predict the future time series, and the predicted future time series results are used as the standard of the normal mode and compared with the actual received satellite telemetry data. Specifically, let the satellite telemetry data time series be Pre-trained large time series models utilize observed time series Forecasting future time series , the generative model learns the conditional distribution by optimizing the objective function Where m is the length of the prediction time, k is the time step of the observed time series, and the loss function is used Perform back propagation. Calculate the error between the predicted time series and the true time series as the anomaly score , expressed as , set the threshold ,like , then it is believed that The corresponding satellite time point is an abnormal point. represents the difference calculation function, ,for Time to A real time series of moments.

[0086] The satellite telemetry data completion of the present invention is to treat the satellite telemetry data time series to be completed as a special item, and send the unmasked segment as an input item to the input embedding module to construct a mask matrix , indicating whether to mask the eigenvalue of a certain time step, is the batch size for system inference and training, The number of parameter types for each time step of the time series, such as the number of voltage, current and other parameter types at each moment.

[0087] The time series Combined with the mask matrix, a new matrix is ​​obtained , For dot multiplication operation, the occluded part is 0, based on , pre-trained large time series model output mask segment generation term ,by As a prediction item, by calculating The reconstruction error between the true value and the original value is back-propagated as the loss, and the loss function is ,in Represents the mask part, Indicates the number of non-zero elements in the mask part.

[0088] The system training optimization of the present invention is implemented as follows:

[0089] During fine-tuning, the cosine similarity between the prompt word and the time series block is used as a penalty loss and added to the total loss function, that is: .

[0090] in is the main loss function for system training, is the mixed feature of the t-th training time step, and the total length T of the optimized instruction embedding is the total training time step, is the cosine similarity calculation, It is a weight factor that balances the main loss and the regularization term. The goal is to maximize the similarity between the prompt word and the mixed feature and enhance the model's perception of the prompt word and satellite time series features.

[0091] In one specific embodiment of the present invention, a user inputs telemetry data for a satellite, including hourly charge current, discharge current, battery pack voltage, cell voltage, and depth of discharge for the past week, along with a prompt requesting "predict telemetry parameters of the battery pack components of a satellite's power subsystem within the next 24 hours." The telemetry data and initial prompt are fed into the embedding module and the knowledge-driven prompt enhancement module, respectively. The telemetry data is divided into several consecutive segments, and the initial prompt is converted into a more informative prompt. The prompt is then embedded and fed into the fusion perception module along with the time series segments for feature fusion, generating fused data. The fused data is then fed into the large language model processing module for enhancement, and the resulting features are then fed into the feature enhancement module for dimensional alignment. The task execution and optimization module is then called to identify the user's intent, identifying the intended task as parameter prediction. The time series model processing module decodes the time series features to predict hourly charge current, discharge current, battery pack voltage, cell voltage, and depth of discharge data for the next 24 hours.

[0092] Similarly, if the user's intention is to detect satellite anomalies in the next 24 hours, the actual telemetry data at each time point is compared with the previous predicted value. Set the charging current threshold to The threshold value of the discharge current is The threshold values ​​of ampere and battery voltage are Volts, the threshold value of the single cell voltage is Volts, the threshold value of discharge depth is If the difference between the actual charging current and the predicted charging current at a certain time point exceeds amperes, or the difference between the actual discharge current and the predicted discharge current exceeds Amperes, or the difference between the actual battery pack voltage and the predicted battery pack voltage exceeds Volts, the difference between the actual cell voltage and the predicted cell voltage exceeds Volts, and the difference between the actual discharge depth and the predicted discharge depth exceeds %, it is determined that the telemetry data is abnormal at that time point. For example, in the fifth hour in the future, the actual cell voltage is 28.08 volts, and the difference with the predicted cell voltage of 20 volts is greater than Volt, at this time the system determines that the single cell voltage telemetry data at that time point is abnormal, marks it and issues an alarm in time.

[0093] The processing system of the present invention can automatically complete in-depth analysis and processing of multi-parameter satellite telemetry data. By integrating a large language model, a pre-trained large time series model, and a knowledge graph, it can achieve accurate prediction of satellite operating status and anomaly detection. The system deeply integrates multi-source data and knowledge information, leveraging the advantages of the model to explore the patterns behind the data. Through the top-level task execution and optimization module control, the various functional modules work together to complete complex data processing tasks. At the same time, by freezing the backbone model and updating a small number of parameters, the difficulty of model training is reduced and processing efficiency is improved. Not only can it deeply analyze satellite telemetry data to provide a decision-making basis for satellite operation management; it can also effectively process small sample data, improving the versatility and adaptability of the model; it can also quickly process large amounts of data while ensuring processing accuracy, meeting the timeliness requirements of satellite operation management for data processing.

[0094] The system of the present invention has multiple functions for satellite telemetry data prediction, anomaly detection, and missing data completion, enhancing the accuracy of satellite telemetry data prediction and anomaly detection. It can more accurately identify outliers during anomaly detection tasks, effectively improving the efficiency and accuracy of satellite operations management. Furthermore, based on the designed processing flow and modular architecture, the present invention can successfully complete the efficient processing and analysis of multi-parameter satellite telemetry data.

Claims

1. A multi-modal enhanced satellite telemetry data multi-task processing system, characterized in that: It includes input embedding module, knowledge-driven prompt enhancement module, fusion perception module, large language model processing module, feature enhancement module, timing model processing module and task execution and optimization module; The input embedding module is used to normalize and fragment the satellite telemetry data; The knowledge-driven prompt enhancement module is used to perform knowledge graph enhancement on the initial instructions in the natural language modality to generate optimized instructions; The fusion perception module is used to fuse the natural language modality of the optimized instruction and the time series modality of the satellite telemetry data; The large language model processing module is used to mine mixed features in the fused data; The feature enhancement module is used to perform dimension mapping and enhancement on the mixed features; The time series model processing module is used to derive the final prediction result based on the features enhanced by the feature enhancement module, so as to realize the prediction and anomaly detection of satellite telemetry data; The task execution and optimization module is used to identify the task intent according to the instruction, perform one or more of the tasks of satellite telemetry data prediction, satellite telemetry data completion and satellite anomaly detection, and perform system training optimization; The time series model processing module is based on a pre-trained large time series model and derives the final prediction result according to the features enhanced by the feature enhancement module, thereby realizing the prediction and anomaly detection of satellite telemetry data. The prediction of satellite telemetry data is implemented as follows: In the fine-tuning stage, autoregression is used to generate training targets, and the total length of the optimized instruction embedding is , divide the time step L of historical satellite telemetry data into The pre-trained large time series model outputs the next item, and the mean square error between each item and the actual telemetry data of the corresponding satellite is calculated as the loss function for back propagation; In the inference phase, the prediction result is connected with the input satellite telemetry data time series, and the pre-trained large time series model is repeatedly used to generate the next term until the total length of the predicted term reaches the expected length, thereby achieving the prediction of the satellite telemetry data; The anomaly detection is based on the idea of ​​predictive anomaly detection. It uses the observed time series to predict the future time series, and uses the predicted results as the standard of the normal mode to compare with the actual received satellite telemetry data. Assume that the time series of satellite telemetry data is ,in is the true value of satellite telemetry data at time t, and the pre-trained large time series model uses the observed time series Forecasting future time series ;Generative models learn conditional distributions by optimizing objective functions ,in is the predicted value of satellite telemetry data at time t, m is the prediction time length, k is the time step of the observed time series, and back propagation is performed through the loss function; the error between the predicted time series and the true time series is calculated as the anomaly score, which is expressed as , set the threshold ,like , then it is believed that The corresponding satellite time point is an abnormal point. represents the difference calculation function, ,for Time to A real time series of moments.

2. The multi-modal enhanced satellite telemetry data multi-task processing system according to claim 1, characterized in that: The input embedding module normalizes and segments the satellite telemetry data. The implementation method is as follows: Normalize the non-stationary time series of satellite telemetry data so that the mean of the data is zero and the standard deviation is unity; The satellite telemetry data from multiple sources are merged into a univariate sequence pool, and the univariate sequence time points for training are selected according to the normal distribution, and then divided into several continuous overlapping or non-overlapping fragments.

3. The multi-modal enhanced satellite telemetry data multi-task processing system according to claim 1, characterized in that: The knowledge-driven prompt enhancement module performs knowledge graph enhancement on the initial instructions to generate optimized instructions. The implementation method is as follows: Extract keywords from the initial instructions through a knowledge-driven enhancement model, improve and build a graph structure based on GraphRAG technology, retrieve information from an external knowledge database, and combine it with the keywords to generate optimized instructions; The improvements include: Before building the knowledge graph, identify the entity type according to the scenario and save it; When building the knowledge graph, entities are extracted based on type and duplicate entities are removed; Add descriptions of entities to the knowledge graph; Locate information in the knowledge graph based on keyword matching during retrieval.

4. The multi-modal enhanced satellite telemetry data multi-task processing system according to claim 1, characterized in that: The fusion perception module fuses the natural language modality of the optimized command with the time series modality of the satellite telemetry data. The implementation method is as follows: The optimized instruction embeddings are mean-pooled to generate a global hint representation of the initial semantic vector, which is then repeatedly expanded to align with the length of the time series features of the satellite telemetry data. Multi-scale convolution operations are performed on the time series features of the satellite telemetry data using convolution kernels of different scales to extract multi-level features. The expanded global hint representation is spliced ​​with the multi-level features, and a learnable fusion weight is introduced. A weighted residual connection is performed through linear layer fusion to obtain a mixed feature, and the mixed feature is spliced ​​again with the initial semantic vector to obtain the fused data.

5. The multi-modal enhanced satellite telemetry data multi-task processing system according to claim 1, characterized in that: The large language model processing module mines the mixed features in the fused data, and the implementation method is as follows: The fused data is input into the large language model processing module for enhanced processing. The time series features of the satellite telemetry data are supplemented with the help of the prompt words of the optimized instructions, and the mixed features in the fused data are mined based on the initial semantic vector.

6. The multi-modal enhanced satellite telemetry data multi-task processing system according to claim 1, characterized in that: The feature enhancement module performs dimension mapping and enhancement on the mixed features, and the implementation method is as follows: The hidden features of the prompt words are removed, only the time series features are retained, and the time series features are dimensionally mapped and enhanced.

7. The multi-modal enhanced satellite telemetry data multi-task processing system according to claim 1, characterized in that: The satellite telemetry data completion process treats the satellite telemetry data time series to be completed as a special item, and the unmasked segment is sent to the input embedding module as an input item to construct a mask matrix , indicating whether to mask the eigenvalue of a certain time step, is the batch size for system inference and training, The number of parameter types for each time step of the time series; The time series Combined with the mask matrix, a new matrix is ​​obtained , For dot multiplication operation, the occluded part is 0, based on , pre-trained large time series model output mask segment generation term ,by As a prediction item, by calculating The reconstruction error between the true value and the original value is back-propagated as the loss, and the loss function is ,in Represents the mask part, Indicates the number of non-zero elements in the mask part.

8. The multi-modal enhanced satellite telemetry data multi-task processing system according to claim 1, characterized in that: The task execution and optimization module performs system training optimization, and the implementation method is as follows: During fine-tuning, the cosine similarity between the prompt word and the time series block is used as a penalty loss and added to the total loss function, that is: ； in is the main loss function for system training, is the prompt feature at time t, is the mixed feature of the t-th training time step, with the total length T of the optimized instruction embedding as the total training time step, is the cosine similarity calculation, To balance the weight factors of the main loss and the regularization term, the goal is to maximize the similarity between the hint features and the mixed features.

9. A multi-modal enhanced satellite telemetry data multi-task processing method, implemented based on the multi-modal enhanced satellite telemetry data multi-task processing system of claim 1, characterized in that: The steps include: Step 1: Normalize and segment the satellite telemetry data, perform knowledge graph enhancement on the initial instructions in the natural language mode, and generate optimized instructions; Step 2: Fuse the natural language modality of the optimized command with the time series modality of the satellite telemetry data; Step 3: Using the large language model processing module, the temporal features of the satellite telemetry data are supplemented according to the prompt words of the optimized instructions, and the mixed features in the fused data are mined; Step 4: Dimension mapping and enhancement of the mixed features; Step 5: Using the time series model processing module, the final prediction results are derived based on the enhanced features to achieve prediction and anomaly detection of satellite telemetry data; Step 6: Determine the task intent based on the prompt word in the user input instruction, and execute one or more of the satellite telemetry data prediction task, satellite telemetry data completion task, and satellite anomaly detection task.

Images