An inter-satellite fault knowledge sharing method based on federated learning
By transmitting model parameters instead of data between satellites and employing a federated learning method for collaborative updating of multi-satellite fault diagnosis models, the problem of the inability to continuously update fault diagnosis models in satellite constellations is solved, thereby improving the accuracy and communication efficiency of satellite constellation fault diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INST OF CONTROL ENG
- Filing Date
- 2026-02-12
- Publication Date
- 2026-06-05
Smart Images
Figure CN122154845A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an inter-satellite fault knowledge sharing method based on federated learning, belonging to the aerospace field. Background Technology
[0002] During constellation operation, whether it's a homogeneous or heterogeneous constellation, when one satellite malfunctions, it needs to expand its fault knowledge on other satellites to ensure incremental updates of the fault diagnosis models deployed on each satellite. However, due to the large number of constellation members, inter-satellite communication constraints make it difficult to centrally train the fault diagnosis models based on all data when updating them. This results in the inability to continuously update the fault diagnosis models, affecting the accuracy of subsequent constellation fault diagnoses. Summary of the Invention
[0003] The technical problem solved by this invention is to overcome the shortcomings of existing technologies and provide an inter-satellite fault knowledge sharing method based on federated learning. This method reduces the amount of information exchanged between satellites by transmitting parameters of various models instead of fault data between satellites. Simultaneously, it achieves collaborative updating of multi-satellite fault diagnosis models based on the transmitted model parameters, solving the problem that existing fault diagnosis models cannot be continuously updated.
[0004] The technical solution of this invention is: An inter-satellite fault knowledge sharing method based on federated learning, comprising the following steps: (1) In the satellite constellation, the computing power and system configuration of each satellite are comprehensively evaluated, and the satellite with the best comprehensive evaluation result is determined as the master star, and the other satellites in the satellite constellation are slave stars; execute step (2). (2) Configure a global model on the primary star and a local model on each slave star; execute step (3); (3) Determine if any slave satellite in the satellite constellation has malfunctioned. If any slave satellite malfunctions, proceed to step (4); otherwise, proceed to step (6). (4) Each slave satellite preprocesses its local fault data to obtain feature data and divides the feature data into training data and validation data; the slave satellite uses the training data to train the local model deployed on itself and obtains the local model weights accordingly; all slave satellites send their local model weights and validation data to the master satellite; execute step (5). (5) The master satellite updates the global model based on the local model weights sent by the slave satellites to obtain the initial updated global model; the master satellite trains the initial updated global model based on its own local fault data and the verification data sent by each slave satellite to obtain the updated global model; execute step (7). (6) Determine whether the primary star has malfunctioned. If the primary star has malfunctioned, the primary star performs global model update training based on its own local fault data to obtain the updated global model and execute step (7); otherwise, execute step (8). (7) The master satellite sends the updated global model to each slave satellite, and the slave satellite updates its own deployed local model based on the updated global model; execute step (8); (8) Repeat steps (3) to (7) until the satellite constellation stops working.
[0005] Furthermore, in step (2), the global model includes an input layer, a first fully connected layer with 16 neurons, a second fully connected layer with 64 neurons, a third fully connected layer with 32 neurons, a fourth fully connected layer with 128 neurons, a fifth fully connected layer with 64 neurons, and an output layer. The local models are categorized into small-scale, medium-scale, and large-scale local models. The small-scale local model includes an input layer, a first fully connected layer with 16 neurons, and an output layer. The medium-scale local model includes an input layer, a second fully connected layer with 64 neurons, a third fully connected layer with 32 neurons, and an output layer. The large-scale local model includes an input layer, a fourth fully connected layer with 128 neurons, a fifth fully connected layer with 64 neurons, and an output layer. The type of local model deployed on satellite is determined by a comprehensive evaluation of the satellite's computing power and system configuration. Local models with an excellent comprehensive evaluation result are classified as large-scale local models, those with a medium comprehensive evaluation result are classified as medium-scale local models, and those with a general comprehensive evaluation result are classified as small-scale local models.
[0006] Furthermore, the specific steps for each slave satellite to preprocess its own local fault data in step (3) are as follows: The first step is to remove faulty data and normalize it to obtain preprocessed data. The second step is to extract features from the preprocessed data to obtain feature data.
[0007] Furthermore, in step (4), dividing the feature data into training data and validation data specifically involves selecting 80% of the feature data as training data and the remaining 20% as validation data.
[0008] Furthermore, the specific steps for updating the global model based on the local model weights sent from the satellite in step (5) are as follows: (5.1) Based on the weights of the local models transmitted from the satellites that have deployed small-scale local models, calculate the small-scale aggregate weights using the following formula:
[0009] in, For small-scale aggregation weights; For the first i Local model weights transferred from the star, each with a small-scale local model deployed; For the first i The amount of local fault data possessed by a satellite with a small-scale local model deployed; This is the sum of local fault data possessed by all slave satellites that have deployed small-scale local models; The number of slave stars with small-scale local models deployed; (5.2) Based on the weights of the local models transmitted from all satellites that have deployed medium-scale local models, calculate the medium-scale aggregate weights using the following formula:
[0010] in, For medium-scale aggregation weights; For the first i Local model weights transferred from a star, with a medium-scale local model deployed; For the first i The amount of local fault data possessed by a satellite with a medium-sized local model deployed; This is the sum of local fault data possessed by all slave satellites deployed with medium-scale local models; The number of slave stars with medium-scale local models deployed; (5.3) Based on the weights of all local models transmitted from satellites that have deployed large-scale local models, calculate the large-scale aggregate weights using the following formula:
[0011] in, For large-scale aggregated weights; For the first i Local model weights transferred from a star, with a large-scale local model deployed; For the first i The amount of local fault data possessed by a satellite with a large-scale local model deployed; This is the sum of local fault data possessed by all slave satellites deployed with large-scale local models; The number of slave stars with large-scale local models deployed; (5.4) Based on small-scale aggregated weights, update the weights of the first fully connected layer of the global model; based on medium-scale aggregated weights, update the weights of the second and fourth fully connected layers of the global model; based on large-scale aggregated weights, update the weights of the fourth and fifth fully connected layers of the global model; after the weights of the first, second, third, fourth and fifth fully connected layers are updated, the initial updated global model is obtained.
[0012] Furthermore, in step (5), the primary satellite trains the initial global update model based on its own local fault data and the verification data sent by each satellite as follows: the primary satellite trains the initial global update model based on its own local fault data and evaluates the initial global update model during the training process using the verification data sent by all satellites; when the verification accuracy of the initial global update model in the verification data is greater than a preset threshold or the number of training iterations of the initial global update model is greater than a preset number, the initial global update model stops training; otherwise, the initial global update model continues to iterate and train; after the initial global update model is trained, the updated global model is obtained.
[0013] Furthermore, in step (7), the process of updating the local model deployed by the satellite based on the updated global model is as follows: if the local model deployed by the satellite is a small-scale local model, the satellite updates the weights of the first fully connected layer in the small-scale local model using the updated global model; if the local model deployed by the satellite is a medium-scale local model, the satellite updates the weights of the local first fully connected layer in the medium-scale local model using the updated global model; if the local model deployed by the satellite is a small-scale local model, the satellite updates the weights of the local second fully connected layer and the local third fully connected layer in the small-scale local model using the updated global model; if the local model deployed by the satellite is a large-scale local model, the satellite updates the weights of the local fourth fully connected layer and the local fifth fully connected layer in the large-scale local model using the updated global model weights.
[0014] Secondly, the present invention also proposes a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the above-described method for sharing inter-satellite fault knowledge based on federated learning.
[0015] Thirdly, the present invention also proposes a processor, characterized in that the processor is used to run a program, wherein the program executes the above-described inter-satellite fault knowledge sharing method based on federated learning.
[0016] Fourthly, the present invention also proposes a storage medium comprising a stored program, wherein, when the program is executed, the device where the storage medium is located executes the aforementioned inter-satellite fault knowledge sharing method based on federated learning.
[0017] The advantages of this invention compared to the prior art are: (1) This invention reduces the amount of information exchanged between satellites by transmitting the parameters of each model between satellites instead of directly transmitting fault data. At the same time, it enables the collaborative updating of multi-satellite fault diagnosis models based on the transmitted model parameters, thereby realizing the continuous updating of fault diagnosis models on each satellite and improving the accuracy of subsequent satellite constellation fault diagnosis.
[0018] (2) The model deployed on each satellite in this invention is trained multiple times using the local fault dataset, so that the trained model parameters are closer to the optimal parameters, further reducing the number of subsequent model communication rounds and improving communication efficiency. Attached Figure Description
[0019] Figure 1 This is a flowchart of an inter-satellite fault knowledge sharing method based on federated learning according to the present invention. Figure 2 This is a schematic diagram of model deployment in an inter-satellite fault knowledge sharing method based on federated learning according to the present invention. Detailed Implementation
[0020] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings.
[0021] like Figure 1 As shown, this invention provides a method for inter-satellite fault knowledge sharing based on federated learning, with the following steps: (1) In the satellite constellation, the computing power and system configuration of each satellite are comprehensively evaluated, and the satellite with the best comprehensive evaluation result is determined as the master star, and the other satellites in the satellite constellation are slave stars; execute step (2). (2) Configure a global model on the primary star and a local model on each slave star; execute step (3); The global model in step (2) includes an input layer, a first fully connected layer with 16 neurons, a second fully connected layer with 64 neurons, a third fully connected layer with 32 neurons, a fourth fully connected layer with 128 neurons, a fifth fully connected layer with 64 neurons, and an output layer. The local models are categorized into small-scale, medium-scale, and large-scale local models. The small-scale local model includes an input layer, a first fully connected layer with 16 neurons, and an output layer. The medium-scale local model includes an input layer, a second fully connected layer with 64 neurons, a third fully connected layer with 32 neurons, and an output layer. The large-scale local model includes an input layer, a fourth fully connected layer with 128 neurons, a fifth fully connected layer with 64 neurons, and an output layer. The type of local model deployed on satellite is determined by a comprehensive evaluation of the satellite's computing power and system configuration. Local models with an excellent comprehensive evaluation result are classified as large-scale local models, those with a medium comprehensive evaluation result are classified as medium-scale local models, and those with a general comprehensive evaluation result are classified as small-scale local models.
[0022] (3) Determine if any slave satellite in the satellite constellation has malfunctioned. If any slave satellite malfunctions, proceed to step (4); otherwise, proceed to step (6). The specific steps for each slave satellite to preprocess its own local fault data in step (3) are as follows: The first step is to remove faulty data and normalize it to obtain preprocessed data. The second step is to extract features from the preprocessed data to obtain feature data.
[0023] (4) Each slave satellite preprocesses its local fault data to obtain feature data and divides the feature data into training data and validation data; the slave satellite uses the training data to train the local model deployed on itself and obtains the local model weights accordingly; all slave satellites send their local model weights and validation data to the master satellite; execute step (5). In step (4), dividing the feature data into training data and validation data specifically involves selecting 80% of the feature data as training data and the remaining 20% as validation data.
[0024] (5) The master satellite updates the global model based on the local model weights sent by the slave satellites to obtain the initial updated global model; the master satellite trains the initial updated global model based on its own local fault data and the verification data sent by each slave satellite to obtain the updated global model; execute step (7). The specific steps for updating the global model based on the local model weights sent from the satellite in step (5) are as follows: (5.1) Based on the weights of the local models transmitted from the satellites that have deployed small-scale local models, calculate the small-scale aggregate weights using the following formula:
[0025] in, For small-scale aggregation weights; For the first i Local model weights transferred from the star, each with a small-scale local model deployed; For the first i The amount of local fault data possessed by a satellite with a small-scale local model deployed; This is the sum of local fault data possessed by all slave satellites that have deployed small-scale local models; The number of slave stars with small-scale local models deployed; (5.2) Based on the weights of the local models transmitted from all satellites that have deployed medium-scale local models, calculate the medium-scale aggregate weights using the following formula:
[0026] in, For medium-scale aggregation weights; For the first i Local model weights transferred from a star, with a medium-scale local model deployed; For the first i The amount of local fault data possessed by a satellite with a medium-sized local model deployed; This is the sum of local fault data possessed by all slave satellites deployed with medium-scale local models; The number of slave stars with medium-scale local models deployed; (5.3) Based on the weights of all local models transmitted from satellites that have deployed large-scale local models, calculate the large-scale aggregate weights using the following formula:
[0027] in, For large-scale aggregated weights; For the first i Local model weights transferred from a star, with a large-scale local model deployed; For the first i The amount of local fault data possessed by a satellite with a large-scale local model deployed; This is the sum of local fault data possessed by all slave satellites deployed with large-scale local models; The number of slave stars with large-scale local models deployed; (5.4) Based on small-scale aggregated weights, update the weights of the first fully connected layer of the global model; based on medium-scale aggregated weights, update the weights of the second and fourth fully connected layers of the global model; based on large-scale aggregated weights, update the weights of the fourth and fifth fully connected layers of the global model; after the weights of the first, second, third, fourth and fifth fully connected layers are updated, the initial updated global model is obtained.
[0028] In step (5), the primary satellite trains the initial global update model based on its local fault data and the verification data sent by each satellite as follows: The primary satellite trains the initial global update model based on its local fault data and evaluates the initial global update model during the training process using the verification data sent by all satellites; when the verification accuracy of the initial global update model in the verification data is greater than a preset threshold or the number of training iterations of the initial global update model is greater than a preset number, the initial global update model stops training; otherwise, the initial global update model continues to iterate and train; after the initial global update model is trained, the updated global model is obtained.
[0029] (6) Determine whether the primary star has malfunctioned. If the primary star has malfunctioned, the primary star performs global model update training based on its own local fault data to obtain the updated global model and execute step (7); otherwise, execute step (8). (7) The master satellite sends the updated global model to each slave satellite, and the slave satellite updates its own deployed local model based on the updated global model; execute step (8); In step (7), the process of updating the local model deployed by the satellite based on the updated global model is as follows: if the local model deployed by the satellite is a small-scale local model, the satellite updates the weights of the first fully connected layer in the small-scale local model using the updated global model; if the local model deployed by the satellite is a medium-scale local model, the satellite updates the weights of the local first fully connected layer in the medium-scale local model using the updated global model; if the local model deployed by the satellite is a small-scale local model, the satellite updates the weights of the local second fully connected layer and the local third fully connected layer in the small-scale local model using the updated global model; if the local model deployed by the satellite is a large-scale local model, the satellite updates the weights of the local fourth fully connected layer and the local fifth fully connected layer in the large-scale local model using the updated global model weights.
[0030] (8) Repeat steps (3) to (7) until the satellite constellation stops working.
[0031] Based on the above process, the overall architecture of the present invention is shown below. Figure 2As shown, this invention reduces the amount of information exchanged between satellites by transmitting the parameters of each model between satellites instead of directly transmitting fault data. Simultaneously, it enables collaborative updating of multi-satellite fault diagnosis models based on the transmitted model parameters, achieving continuous updating of the fault diagnosis model on each satellite and thus improving the accuracy of subsequent satellite constellation fault diagnosis. Furthermore, the model deployed on each satellite in this invention is trained multiple times using the local fault dataset, resulting in model parameters that are closer to the optimal parameters, further reducing the number of subsequent model communication rounds and improving communication efficiency.
[0032] Secondly, the present invention also proposes a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the above-described method for sharing inter-satellite fault knowledge based on federated learning.
[0033] Thirdly, the present invention also proposes a processor, characterized in that the processor is used to run a program, wherein the program executes the above-described inter-satellite fault knowledge sharing method based on federated learning.
[0034] Fourthly, the present invention also proposes a storage medium comprising a stored program, wherein, when the program is executed, the device where the storage medium is located executes the aforementioned inter-satellite fault knowledge sharing method based on federated learning.
[0035] The parts of this invention not described in detail are common knowledge to those skilled in the art.
Claims
1. A method for inter-satellite fault knowledge sharing based on federated learning, characterized in that, The steps are as follows: (1) In the satellite constellation, the computing power and system configuration of each satellite are comprehensively evaluated, and the satellite with the best comprehensive evaluation result is determined as the master star, and the other satellites in the satellite constellation are slave stars; execute step (2). (2) Configure a global model on the primary star and a local model on each slave star; execute step (3); (3) Determine if any slave satellite in the satellite constellation has malfunctioned. If any slave satellite malfunctions, proceed to step (4); otherwise, proceed to step (6). (4) Each slave satellite preprocesses its local fault data to obtain feature data and divides the feature data into training data and validation data; the slave satellite uses the training data to train the local model deployed on itself and obtains the local model weights accordingly; all slave satellites send their local model weights and validation data to the master satellite; execute step (5). (5) The master satellite updates the global model based on the local model weights sent by the slave satellites to obtain the initial updated global model; the master satellite trains the initial updated global model based on its own local fault data and the verification data sent by each slave satellite to obtain the updated global model; execute step (7). (6) Determine whether the primary star has malfunctioned. If the primary star has malfunctioned, the primary star performs global model update training based on its own local fault data to obtain the updated global model and execute step (7); otherwise, execute step (8). (7) The master satellite sends the updated global model to each slave satellite, and the slave satellite updates its own deployed local model based on the updated global model; execute step (8); (8) Repeat steps (3) to (7) until the satellite constellation stops working.
2. The inter-satellite fault knowledge sharing method based on federated learning according to claim 1, characterized in that: The global model in step (2) includes an input layer, a first fully connected layer with 16 neurons, a second fully connected layer with 64 neurons, a third fully connected layer with 32 neurons, a fourth fully connected layer with 128 neurons, a fifth fully connected layer with 64 neurons, and an output layer. The local models are categorized into small-scale, medium-scale, and large-scale local models. The small-scale local model includes an input layer, a first fully connected layer with 16 neurons, and an output layer. The medium-scale local model includes an input layer, a second fully connected layer with 64 neurons, a third fully connected layer with 32 neurons, and an output layer. The large-scale local model includes an input layer, a fourth fully connected layer with 128 neurons, a fifth fully connected layer with 64 neurons, and an output layer. The type of local model deployed on satellite is determined by a comprehensive evaluation of the satellite's computing power and system configuration. Local models with an excellent comprehensive evaluation result are classified as large-scale local models, those with a medium comprehensive evaluation result are classified as medium-scale local models, and those with a general comprehensive evaluation result are classified as small-scale local models.
3. A method for inter-satellite fault knowledge sharing based on federated learning according to claim 1 or 2, characterized in that: The specific steps for each slave satellite to preprocess its own local fault data in step (3) are as follows: The first step is to remove faulty data and normalize it to obtain preprocessed data. The second step is to extract features from the preprocessed data to obtain feature data.
4. A method for inter-satellite fault knowledge sharing based on federated learning according to claim 1 or 2, characterized in that: In step (4), dividing the feature data into training data and validation data specifically involves selecting 80% of the feature data as training data and the remaining 20% as validation data.
5. The inter-satellite fault knowledge sharing method based on federated learning according to claim 2, characterized in that: The specific steps for updating the global model based on the local model weights sent from the satellite in step (5) are as follows: (5.1) Based on the weights of the local models transmitted from the satellites that have deployed small-scale local models, calculate the small-scale aggregate weights using the following formula: in, For small-scale aggregation weights; For the first i Local model weights transferred from the star, each with a small-scale local model deployed; For the first i The amount of local fault data possessed by a satellite with a small-scale local model deployed; This is the sum of local fault data possessed by all slave satellites that have deployed small-scale local models; The number of slave stars with small-scale local models deployed; (5.2) Based on the weights of the local models transmitted from all satellites that have deployed medium-scale local models, calculate the medium-scale aggregate weights using the following formula: in, For medium-scale aggregation weights; For the first i Local model weights transferred from a star, with a medium-scale local model deployed; For the first i The amount of local fault data possessed by a satellite with a medium-sized local model deployed; This is the sum of local fault data possessed by all slave satellites deployed with medium-scale local models; The number of slave stars with medium-scale local models deployed; (5.3) Based on the weights of all local models transmitted from satellites that have deployed large-scale local models, calculate the large-scale aggregate weights using the following formula: in, For large-scale aggregated weights; For the first i Local model weights transferred from a star, with a large-scale local model deployed; For the first i The amount of local fault data possessed by a satellite with a large-scale local model deployed; This is the sum of local fault data possessed by all slave satellites deployed with large-scale local models; The number of slave stars with large-scale local models deployed; (5.4) Based on small-scale aggregated weights, update the weights of the first fully connected layer of the global model; based on medium-scale aggregated weights, update the weights of the second and fourth fully connected layers of the global model; based on large-scale aggregated weights, update the weights of the fourth and fifth fully connected layers of the global model; after the weights of the first, second, third, fourth and fifth fully connected layers are updated, the initial updated global model is obtained.
6. The inter-satellite fault knowledge sharing method based on federated learning according to claim 2, characterized in that: In step (5), the primary satellite trains the initial global update model based on its local fault data and the verification data sent by each satellite as follows: The primary satellite trains the initial global update model based on its local fault data and evaluates the initial global update model during the training process using the verification data sent by all satellites; when the verification accuracy of the initial global update model in the verification data is greater than a preset threshold or the number of training iterations of the initial global update model is greater than a preset number, the initial global update model stops training; otherwise, the initial global update model continues to iterate and train; after the initial global update model is trained, the updated global model is obtained.
7. The inter-satellite fault knowledge sharing method based on federated learning according to claim 2, characterized in that: In step (7), the process of updating the local model deployed by the satellite based on the updated global model is as follows: if the local model deployed by the satellite is a small-scale local model, the satellite updates the weights of the first fully connected layer in the small-scale local model using the updated global model; if the local model deployed by the satellite is a medium-scale local model, the satellite updates the weights of the local first fully connected layer in the medium-scale local model using the updated global model; if the local model deployed by the satellite is a small-scale local model, the satellite updates the weights of the local second fully connected layer and the local third fully connected layer in the small-scale local model using the updated global model; if the local model deployed by the satellite is a large-scale local model, the satellite updates the weights of the local fourth fully connected layer and the local fifth fully connected layer in the large-scale local model using the updated global model weights.
8. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the inter-satellite fault knowledge sharing method based on federated learning as described in any one of claims 1 to 7.
9. A processor, characterized in that, The processor is used to run a program, wherein the program executes an inter-satellite fault knowledge sharing method based on federated learning as described in any one of claims 1 to 7.
10. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, it controls the device where the storage medium is located to execute the inter-satellite fault knowledge sharing method based on federated learning as described in any one of claims 1 to 7.