A satellite on-orbit health diagnosis method and system based on physical-knowledge-data triple fusion

Abstract

The application discloses a satellite on-orbit health diagnosis method and system based on physical-knowledge-data triple fusion, and relates to the technical field of intelligent health management of spacecraft. The application constructs three parallel information processing paths of a data-driven channel (multi-scale time sequence convolution + self-attention), a physical information channel (a learnable physical constraint layer) and a knowledge enhancement channel (knowledge graph vector embedding + graph attention reasoning), dynamically adjusts the weights of the three channels according to the data quality through an adaptive gating fusion module, and gives priority to data driving when the data is sufficient, and gives priority to physical constraints and knowledge guidance when the data is sparse. Under the unified framework, health state evaluation, degradation trend prediction and fault diagnosis are simultaneously realized, and an interpretable fault causal link is output. The application solves the three dilemmas of insufficient small sample generalization of pure data-driven methods, difficulty of complex degradation covered by pure physical models and non-systematic use of domain knowledge.

Description

Technical Field

[0001] This application relates to the field of intelligent health management technology for spacecraft, specifically to a satellite on-orbit health diagnosis method and system based on the fusion of physical, knowledge, and data elements. Background Technology

[0002] With the rapid development of large-scale low-Earth orbit satellite constellations, such as SpaceX's Starlink and OneWeb, the number of satellites in orbit globally has reached tens of thousands. Satellite systems are evolving towards high-density networking, large-scale deployment, and autonomous operation. Against this backdrop, the traditional operation and maintenance model, which relies on ground control stations to monitor individual satellites and combine this with human experience for health status assessment, has gradually revealed problems such as high cost, slow response, and poor scalability as the system scales up. When the constellation size reaches thousands or even larger, the demand for ground operation and maintenance resources increases linearly with the number of satellites, making it difficult to meet the actual needs of efficient operation of large-scale constellations. Therefore, there is an urgent need to develop on-orbit intelligent health management and diagnostic technologies with autonomous sensing, analysis, and decision-making capabilities to improve the intelligence level and reliability of satellite system operation.

[0003] Existing satellite health diagnosis methods mainly include data-driven methods, physical model methods, and methods that fuse the two, but they still have significant shortcomings: First, pure data-driven methods have limited generalization ability under small sample conditions. Since real fault samples of satellites in orbit are extremely scarce and unevenly distributed, deep learning-based models are prone to overfitting to a few fault types and have insufficient ability to identify unknown fault modes, resulting in a high false alarm rate. Secondly, although the pure physics model method has strong interpretability, it is difficult to accurately characterize the complex nonlinear degradation process under the coupling of multiple physical fields such as heat, electricity, and structure in the satellite system. The accuracy of the model is limited by parameter uncertainty and model simplification. Furthermore, the expert knowledge accumulated in the aerospace field over a long period of time, such as Failure Mode and Effects Analysis (FMECA), Fault Tree Analysis (FTA), and Fault Propagation Path, has not been effectively systematized and integrated into the diagnostic model, resulting in a lack of interpretability and engineering credibility of the diagnostic results. Furthermore, existing dual-channel fusion methods of "data + physics" (such as physical information neural networks) only utilize two types of information sources and fail to incorporate structured domain knowledge as an independent information dimension into reasoning, which has significant limitations in complex fault propagation analysis and cross-system correlation diagnosis scenarios.

[0004] Therefore, there is an urgent need for a satellite on-orbit health diagnosis method that can integrate three information sources: physical mechanism constraints, domain expert knowledge, and data-driven learning, and can adaptively adjust the contribution ratio of each information source according to data quality and application scenario.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this application is to provide a satellite on-orbit health diagnosis method and system based on the ternary fusion of physics, knowledge and data, so as to solve the problems in the background art mentioned above.

[0007] To achieve the above objectives, this application provides the following technical solution: a satellite on-orbit health diagnosis method based on the ternary fusion of physics, knowledge, and data, comprising the following steps: We acquire multi-source telemetry time-series data from satellites in orbit, and perform outlier detection and removal, missing value imputation, time series alignment and normalization on the multi-source telemetry time-series data to obtain standardized telemetry data. At the same time, we construct a three-layer satellite fault knowledge graph that includes component-level failure modes, subsystem-level fault propagation paths and whole-satellite fault coupling relationships. For standardized telemetry data, a temporal feature extraction network containing multi-scale temporal convolutional structure and self-attention structure is established. Temporal features at different time scales are extracted through multiple parallel causal dilated convolutional branches, and the correlation between telemetry parameters is represented by the self-attention structure to obtain data-driven feature vectors. The physical equations of at least two key subsystems of the satellite are discretized and embedded into a neural network to form a learnable constraint layer. Standardized telemetry data is input into the learnable constraint layer. By adaptively calibrating the key physical parameters in the physical equations and calculating the weighted error of the physical equation residuals, the physical information feature vector is obtained. Vectorize and embed the satellite fault knowledge graph and establish a graph attention reasoning structure. Use data-driven feature vectors as queries to perform reasoning in the knowledge graph embedding space to obtain knowledge-enhanced feature vectors and corresponding fault causal links. Data-driven feature vectors, physical information feature vectors, and knowledge-enhanced feature vectors are concatenated and input into a gating network to output a normalized weight vector. The three types of features are then weighted and fused based on the normalized weight vector to obtain a fused feature vector. The fused feature vector is then input into a multi-task output structure to complete health status assessment, degradation trend prediction, and fault diagnosis.

[0008] Preferably, the multi-source telemetry data processing adopts an anomaly detection mechanism based on sliding window statistics. The sliding window length is set to 100 to 500 sampling points. The anomaly judgment threshold is determined by calculating the mean and standard deviation within the window. Data points exceeding the threshold range are marked and replaced by interpolation using neighboring normal data. At the same time, missing data segments are filtered in conjunction with the continuous missing length judgment rule, thereby ensuring the continuity of the time series and the stability of data quality.

[0009] Preferably, the multi-scale temporal convolutional structure contains 2 to 6 parallel branches, and the expansion rate of each branch increases exponentially to construct feature expression paths at different time scales. By performing hierarchical modeling of second-level, minute-level and hour-level change information, and by channel splicing and point-by-point convolution to achieve feature fusion, the ability to uniformly characterize long-term degradation trends and short-term fluctuation behaviors is enhanced.

[0010] Preferably, the self-attention structure uses a multi-head mechanism to construct a multi-subspace feature interaction mode. By mapping the input features to query vectors, key vectors and value vectors, the correlation weights between different time steps and different parameters are calculated, thereby dynamically strengthening the expression of the coupling relationship between key telemetry variables and improving the ability to extract implicit correlation structures in multivariate time series data.

[0011] Preferably, the physical equation discretization process uses an explicit numerical integration method to convert the continuous-time model into a discrete form. At the same time, the key parameters of the thermal control subsystem, power supply subsystem, and attitude control subsystem are set as trainable variables, and the parameters are adaptively corrected through the error backpropagation mechanism, thereby improving the model's ability to fit complex nonlinear degradation processes.

[0012] Preferably, the physical constraint error is constructed using a multi-system joint weighting method, which aggregates the residuals of different physical equations according to their importance by assigning weights. By introducing a subsystem weight adjustment mechanism, the constraint effect of key subsystems is strengthened, and global consistency constraints are achieved by combining sample-level error statistics, thereby improving the degree of matching between the model output results and physical laws.

[0013] Preferably, the knowledge graph embedding process adopts a vector representation method based on translation relationship to encode entities and relationships in low-dimensional vectors. By introducing a negative sampling mechanism, a contrastive learning target is constructed, thereby enhancing the semantic distinction between different fault modes. A stable embedding space structure is obtained through multiple rounds of iterative optimization.

[0014] Preferably, the graph attention inference structure uses a multi-head attention mechanism to weighted aggregate the features of neighboring nodes, dynamically allocates the contribution ratio of neighboring nodes to the representation of the current node using attention coefficients, and combines nonlinear mapping to update node features, thereby characterizing the difference in influence intensity between different nodes in the fault propagation path.

[0015] Preferably, the gated network adopts a two-layer fully connected structure to construct a feature weight generation mechanism. By performing nonlinear mapping on the concatenation results of the three types of features and normalizing them, a weight vector is obtained. This weight vector automatically adjusts the contribution ratio of the three types of information sources as the input data distribution changes, thereby improving the model's adaptability to changes in data distribution and unknown fault modes.

[0016] A satellite on-orbit health diagnosis system based on the physics-knowledge-data ternary fusion includes a multi-source data processing module, a temporal feature extraction module, a physical constraint modeling module, a knowledge graph reasoning module, and a ternary fusion diagnosis module; The multi-source data processing module acquires multi-source telemetry time-series data from the satellite in orbit, performs outlier detection and removal, missing value interpolation, time series alignment and normalization on the multi-source telemetry time-series data to obtain standardized telemetry data; at the same time, it constructs a three-layer satellite fault knowledge graph that includes component-level failure modes, subsystem-level fault propagation paths and whole-satellite fault coupling relationships. The temporal feature extraction module establishes a temporal feature extraction network containing multi-scale temporal convolutional structures and self-attention structures for standardized telemetry data. It extracts temporal features at different time scales through multiple parallel causal dilated convolutional branches and uses the self-attention structure to characterize the correlation between telemetry parameters, thereby obtaining data-driven feature vectors. The physical constraint modeling module discretizes the physical equations of at least two key subsystems of the satellite and embeds them into a neural network to form a learnable constraint layer. Standardized telemetry data is input into the learnable constraint layer. By adaptively calibrating the key physical parameters in the physical equations and calculating the weighted error of the physical equation residuals, a physical information feature vector is obtained. The knowledge graph reasoning module embeds the satellite fault knowledge graph into vectors and establishes a graph attention reasoning structure. It uses data-driven feature vectors as queries to perform reasoning in the knowledge graph embedding space, and obtains knowledge-enhanced feature vectors and corresponding fault causal links. The ternary fusion diagnostic module concatenates data-driven feature vectors, physical information feature vectors, and knowledge-enhanced feature vectors and inputs them into a gating network. It outputs a normalized weight vector and then performs weighted fusion of the three types of features based on the normalized weight vector to obtain a fused feature vector. The fused feature vector is then input into a multi-task output structure to complete health status assessment, degradation trend prediction, and fault diagnosis.

[0017] The technical effects and advantages provided by this application in the above technical solution are as follows: Beneficial Effect 1: This application constructs a collaborative modeling framework integrating three information sources: physical mechanism constraints, fault knowledge graphs, and data-driven learning. This framework enables information from different sources to play a complementary role within a unified system, thereby forming a more complete diagnostic logic chain. This approach overcomes the limitations of a single information source, simultaneously considering data representation capabilities, mechanism constraint capabilities, and empirical reasoning capabilities. This results in more comprehensive and reliable diagnostic results and improves the accuracy of identifying the state of complex systems.

[0018] Benefit 2: This application introduces an adaptive gating fusion mechanism, enabling the model to automatically adjust the weight allocation of each information channel based on input data quality and changes in operating conditions. When data samples are insufficient or exhibit abnormal distributions, it enhances the influence of physical constraints and knowledge reasoning, thereby maintaining stable diagnostic output. This mechanism effectively improves the model's adaptability under various operating conditions and enhances its ability to identify unknown fault scenarios.

[0019] Benefit 3: This application outputs the causal link of the fault through a knowledge-enhanced channel, enabling the diagnostic results to have a clear reasoning path and reflect the inherent logical relationship of fault occurrence and propagation. This approach not only improves the understandability of the diagnostic results but also assists engineers in quickly locating the source of the problem, thereby enhancing the credibility and decision support capabilities during system application.

[0020] Benefit 4: This application achieves health status assessment, degradation trend prediction, and fault diagnosis functions within a unified framework, enabling multiple tasks to share the same feature representation system and form a collaborative optimization relationship, thereby avoiding computational redundancy caused by repeated modeling. This structure can improve overall computational efficiency while ensuring consistency between results from different tasks, providing continuous and stable health management capabilities for satellite on-orbit operation. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0022] Figure 1 This is a flowchart of this application.

[0023] Figure 2 This is a module diagram of this application. Detailed Implementation

[0024] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0025] This application provides, as follows: Figure 1 The satellite on-orbit health diagnosis method based on the physics-knowledge-data ternary fusion, as shown, includes the following steps: We acquire multi-source telemetry time-series data from satellites in orbit, and perform outlier detection and removal, missing value imputation, time series alignment and normalization on the multi-source telemetry time-series data to obtain standardized telemetry data. At the same time, we construct a three-layer satellite fault knowledge graph that includes component-level failure modes, subsystem-level fault propagation paths and whole-satellite fault coupling relationships. For standardized telemetry data, a temporal feature extraction network containing multi-scale temporal convolutional structure and self-attention structure is established. Temporal features at different time scales are extracted through multiple parallel causal dilated convolutional branches, and the correlation between telemetry parameters is represented by the self-attention structure to obtain data-driven feature vectors. The physical equations of at least two key subsystems of the satellite are discretized and embedded into a neural network to form a learnable constraint layer. Standardized telemetry data is input into the learnable constraint layer. By adaptively calibrating the key physical parameters in the physical equations and calculating the weighted error of the physical equation residuals, the physical information feature vector is obtained. Vectorize and embed the satellite fault knowledge graph and establish a graph attention reasoning structure. Use data-driven feature vectors as queries to perform reasoning in the knowledge graph embedding space to obtain knowledge-enhanced feature vectors and corresponding fault causal links. Data-driven feature vectors, physical information feature vectors, and knowledge-enhanced feature vectors are concatenated and input into a gating network to output a normalized weight vector. The three types of features are then weighted and fused based on the normalized weight vector to obtain a fused feature vector. The fused feature vector is then input into a multi-task output structure to complete health status assessment, degradation trend prediction, and fault diagnosis.

[0026] Specific Implementation Method 1: This implementation method takes a certain low-Earth orbit communication constellation as the application object. The constellation has 30 satellites in orbit, each with a mass of about 250 kg, a design life of 5 years, and an orbital altitude of 550 km. It belongs to a typical large-scale low-Earth orbit satellite constellation operation scenario. This type of constellation is characterized by a large number of satellites, a complex operating environment, a long on-orbit time, and frequent changes in operating conditions. Therefore, it puts forward requirements for high real-time performance, high reliability, and high automation level for the on-orbit health management system.

[0027] In this embodiment, the power supply subsystem is selected as the key diagnostic object for explanation because, as the "energy core" of the satellite, it directly affects the overall power supply stability of the satellite, and its health status plays a decisive role in the satellite's mission execution. Furthermore, this subsystem exhibits typical nonlinear degradation characteristics and multi-factor coupling properties, fully demonstrating the advantages of the proposed method in the health diagnosis of complex systems. First, telemetry data was continuously collected from 30 in-orbit satellites for two years. This time span covers typical operating condition changes of the satellites from initial in-orbit operation to the mid-term degradation stage, ensuring that the data possesses sufficient temporal evolution characteristics. The telemetry parameters collected from each satellite include five key parameters: bus voltage, battery charging current, battery discharging current, battery temperature, and solar array output current. Bus voltage reflects the system's power supply stability, charging and discharging current reflects the battery's operating state, temperature reflects the impact of the thermal environment on battery performance, and solar array output current reflects the energy input. These parameters, from input and state to output, constitute... The complete energy flow closed loop of the power subsystem was established; the sampling frequency of each parameter was set to 1Hz or 0.5Hz, which were reasonable values ​​determined after comprehensively considering the satellite telemetry bandwidth, storage resources and dynamic response characteristics, so as to avoid data redundancy while ensuring information integrity; in the data preprocessing stage, the outlier detection method based on the 3σ criterion was first adopted, and combined with the sliding window strategy (window length W_ano=200) to perform dynamic statistics on the local data distribution, thereby improving the adaptability of anomaly detection to non-stationary data. For the detected outliers, linear interpolation replacement was performed on the nearest normal data points to ensure the continuity of the time series.

[0028] Subsequently, to address potential missing data in the telemetry data, a linear interpolation method was used to fill in the gaps, and the maximum allowable length of consecutive missing data was set to L_miss=20. This threshold was set based on the stability of the telemetry link and the dynamic characteristics of the system to avoid misleading subsequent modeling due to long-term missing data. For multi-source heterogeneous sampling data, all parameters were aligned to a 1Hz reference frequency through a unified resampling method, thereby eliminating the time scale inconsistency caused by different sampling frequencies.

[0029] In the normalization stage, the Min-Max normalization method was adopted to map each parameter to a unified numerical range, effectively eliminating dimensional differences and improving the stability of model training. This resulted in a standardized dataset with approximately 189 million data points, a scale sufficient to support effective training of deep learning models and enhance their generalization ability. Simultaneously, a fault knowledge graph for the power supply subsystem was constructed. Based on engineering experience and historical data, this graph systematically integrates fault mechanism information of the power supply subsystem, including 87 failure modes and 42 fault propagation paths. These failure modes cover typical problems such as battery aging, increased internal resistance, abnormal voltage, and abnormal temperature. The fault propagation paths describe the causal relationships and evolution paths between different faults, providing structured support for subsequent knowledge reasoning. After completing the data and knowledge foundation construction, a data-driven channel was further established for… Multi-scale temporal features are extracted from standardized telemetry data. Specifically, K_branch=4 parallel temporal convolutional branches are set, with each branch using different dilation rate sequences: {1, 2, 4, 8}, {1, 4, 16, 64}, {1, 8, 64, 512}, and {1, 16, 256, 4096}. This design allows each branch to focus on short-term fluctuations, medium-term trends, and long-term degradation features, thereby achieving comprehensive modeling of information across multiple time scales. The number of convolutional layers is set to L=4, the kernel size K_conv=3, and the number of output channels C_out=64. This parameter combination ensures the model's expressive power while controlling computational complexity. The output features of each branch are concatenated along the channel dimension and then fused and dimensionality reduced using 1×1 convolutions to obtain a unified 256-dimensional feature representation. This process not only achieves multi-scale information fusion but also reduces redundant features.

[0030] Building upon this, a self-attention structure is introduced to further enhance the model's ability to model complex relationships between parameters. The self-attention module is configured as a 4-layer Transformer encoding structure, with each layer containing 8 attention heads and a hidden dimension of 256. It captures dependencies in different subspaces in parallel through a multi-head mechanism. Its core computation process satisfies Attention(Q, K, V) = softmax(QK^T / ... sqrt(d_k))·V……, where Q, K, and V are the query, key, and value matrices obtained by linear transformation of the input features, respectively. This mechanism can dynamically allocate weights between different time steps and different parameters, thereby strengthening key features and suppressing irrelevant information. Through the synergistic effect of the above multi-scale convolution and self-attention structure, deep modeling of telemetry data in the time and feature dimensions is achieved, and the final output is a data-driven feature vector F_data. This feature vector not only contains dynamic information on the operating status of the power subsystem, but also implies potential degradation trends and anomaly patterns, providing high-quality input for subsequent physical constraint fusion and knowledge reasoning. At the same time, this data-driven channel has good scalability and can be applied to different satellite platforms and different subsystems. Only the types of input parameters and network scale need to be adjusted to complete the migration application, thus demonstrating the universality and practical value of this implementation method in engineering applications.

[0031] Specific Implementation Method Two: In this implementation method, based on the obtained data-driven feature vector F_data, a physical information channel and a knowledge enhancement channel are further constructed to achieve multi-source information modeling and deep semantic reasoning of the satellite power subsystem's operating status. The physical information channel is used to introduce system mechanism constraints to ensure that the model output conforms to objective physical laws, while the knowledge enhancement channel is used to introduce expert experience and fault causal relationships, thereby improving the interpretability and engineering credibility of the diagnostic results. Firstly, in the physical information channel, considering the dynamic behavior characteristics of the power subsystem, a second-order Thevenin equivalent circuit model is selected to model the battery charging and discharging process. This model can better characterize the battery's polarization effect and dynamic response characteristics. Its state equations are expressed as: dV_1 / dt = -V_1 / (R_1·C_1) + I / C_1……,dV_2 / dt = -V_2 / (R_2·C_2) + I / C_2……,V_t = OCV(SOC) - V_1 - V_2 - I·R_0……, where V_1 and V_2 represent the polarization voltages of the two RC branches, reflecting the dynamic response characteristics of the battery at different time scales, R_0 represents the ohmic internal resistance of the battery, R_1, C_1 and R_2, C_2 represent the resistance and capacitance parameters of the two polarization circuits, I is the charging and discharging current, and OCV(SOC) represents the nonlinear mapping relationship between the open-circuit voltage and the state of charge. The above model introduces two time constants to simultaneously model the short-term dynamics and medium-to-long-term polarization behavior of the battery, which has higher accuracy and stronger expressive power compared with the first-order model.

[0032] To adapt to the uncertainties and aging effects of parameters in complex on-orbit environments, this implementation sets R_0, R_1, C_1, R_2, and C_2 as learnable parameters, with initial values ​​of 0.05 ohm, 0.03 ohm, 1000 F, 0.01 ohm, and 5000 F, respectively. These initial values ​​are derived from ground calibration tests and engineering experience. During model training, adaptive optimization is performed through a backpropagation mechanism, thereby achieving online calibration of the battery's actual operating state. Through this "parameter learnability" approach, the traditional fixed-parameter physical model is extended into a semi-physical model that can be dynamically updated with data, effectively solving the model parameter drift problem. During training, the physical equation residuals were calculated by comparing the model's predicted output with the actual telemetry voltage. These residuals were then used as the optimization objective for iterative updates. After 50 epochs of training, the physical equation residuals decreased from the initial 0.82V to 0.06V, indicating that the model could accurately fit the actual system behavior. Furthermore, the deviations between the calibration results of each physical parameter and the laboratory calibration values ​​were less than 8%, verifying the reliability and accuracy of this method in engineering applications. Further, to integrate physical constraints into the overall model training process, this implementation defines the physical constraint loss function as: L_physics = (1 / N)·sum_{i=1}^{N} sum_{s=1}^{S} w_s·||f_s(x_i, theta_s)||^2……, where f_s represents the residual function of the s-th physical equation, theta_s represents the corresponding set of learnable physical parameters, w_s represents the importance weights of different subsystems, N represents the number of samples, and S represents the number of physical equations. By weighted summation of the residuals of each physical equation, a unified modeling of multiple physical constraints is achieved. This loss function plays a regularization role in the overall training objective, enabling the model output to approximate the data distribution while satisfying physical consistency, thereby significantly improving the model's generalization ability under unknown conditions. Finally, the physical information feature vector F_physics is obtained. This feature vector not only contains physical consistency information that data-driven features cannot capture, but also reflects the energy flow and state evolution laws within the system.

[0033] In the knowledge enhancement channel, the fault knowledge graph is first vectorized and embedded, mapping the originally discrete symbolic knowledge to a continuous vector space for fusion with the deep learning model. Specifically, the TransE model is used for pre-training. Its basic idea is to represent the semantic relationships between entities through vector translation relationships; that is, for any triple (h, r, t), its embedding satisfies h + r ≈ t. The corresponding loss function is defined as: L_KG = sum_{(h, r, t) in E} max(0, ||h+rt||2 - ||h_neg+r-t_neg||2 + ... The embedding dimension D_kg is set to 128, the margin is set to 2.0, and the embedding space is trained for 200 epochs to effectively express the semantic relationships between fault modes. After knowledge embedding is completed, a graph attention network is further constructed to realize deep reasoning of structured knowledge. The network is set to a 2-layer structure (L_gat=2), and each layer contains 4 attention heads (H_gat=4). The multi-head mechanism enhances the ability to model different relational patterns. The node update process is as follows: alpha_{ij}^{(l)} = softmax_j(LeakyReLU(a^T·[W^{(l)}h_i^{(l)} || W^{(l)}h_j^{(l)}]))……,h_i^{(l+1)} = sigma(sum_{j in N(i)} alpha_{ij}^{(l)}·W^{(l)}h_j^{(l)})……, where alpha_{ij}^{(l)} represents the attention weight of node i to neighbor node j. This weight is used to dynamically allocate the contribution of different neighbor nodes, thereby propagating fault information in the graph structure. This process essentially simulates the propagation path of faults in the system, enabling the model to automatically learn the causal relationships between faults.

[0034] Building upon this foundation, to achieve deep integration of telemetry data and knowledge graphs, this implementation introduces a cross-attention mechanism. Data-driven feature vectors are used as query vectors to perform information retrieval within the knowledge embedding space. The calculation process is as follows: alpha_{query, j} = softmax_j(Q_query·K_j^T / sqrt(d_k))……,F_knowledge =sum_j alpha_{query,j}·V_j……, where Q_query is the query vector obtained from telemetry feature mapping, and K_j and V_j are the key and value vectors of the knowledge graph nodes, respectively. This mechanism can dynamically select the most relevant knowledge nodes according to the current system state and perform weighted fusion, thereby achieving the effect of "on-demand knowledge retrieval". The final output knowledge-enhanced feature vector F_knowledge not only contains the probability distribution information of the fault mode, but also outputs the corresponding fault propagation link, providing a clear causal explanation path for the diagnostic results. For example, when a voltage abnormality is detected, it can be inferred from the knowledge graph that it may be caused by battery aging or temperature abnormality, and further infer its propagation range, thereby significantly improving the interpretability and engineering usability of the diagnostic system. Through the synergistic effect of the above physical information channel and knowledge enhancement channel, the transformation from "data fitting" to "mechanism constraint + empirical reasoning" is realized, so that the model can still maintain stable output when facing data scarcity or distribution deviation scenarios, providing high-quality, multi-dimensional feature input for the subsequent ternary fusion step.

[0035] Specific Implementation Method 3: In this implementation method, based on the data-driven feature vector F_data, physical information feature vector F_physics, and knowledge-enhanced feature vector F_knowledge already obtained, the three types of features are uniformly modeled and deeply fused to achieve comprehensive judgment and predictive analysis of the operating status of the satellite power subsystem. Considering that the contribution of different information sources to the diagnostic results varies significantly under different operating conditions—for example, when data is sufficient and the operating conditions are stable, data-driven features have stronger representational capabilities, while when data is scarce or unknown fault modes occur, physical constraints and knowledge reasoning are more reliable—this implementation method introduces an adaptive gating mechanism to dynamically weight the three types of features, thereby avoiding the performance degradation problem caused by fixed-weight fusion.

[0036] Specifically, a gated network is first constructed, whose input is the concatenation result of three types of feature vectors, i.e., G = [F_data; F_physics; F_knowledge]..., where the concatenation operation is performed on the feature dimension, enabling information from different sources to interact in a unified space. Then, the concatenated features are input to a fully connected layer and mapped using a non-linear activation function, calculated as h_gate = ReLU(W_1·G + b_1)..., where W_1 and b_1 are learnable parameters. The ReLU function is used to introduce non-linear expressive power and suppress negative interference, thereby extracting higher-order interaction relationships between the three types of features. Further, the intermediate representation h_gate is input to the second fully connected layer and normalized using the softmax function to obtain a three-dimensional weight vector w = softmax(W_2·h_gate + b_2)..., where w = (w_d, w_p, w_k) represent the weight coefficients of the data-driven channel, the physical information channel, and the knowledge-enhancing channel, respectively. This weight vector satisfies non-negativity and normalization constraints, thereby ensuring the stability and interpretability of the fusion process. After obtaining the weight vector, the three types of features are weighted and fused to obtain the final fused feature vector F_fused = w_d·F_data + w_p·F_physics + w_k·F_knowledge... This fusion process essentially realizes the adaptive adjustment of the contribution of different information sources, enabling the model to automatically select the optimal information combination according to the characteristics of the input data. In actual testing, statistical analysis of the weight distribution under different operating conditions revealed that when the system is operating normally and the data quality is high, w_d is approximately 0.58, w_p is approximately 0.25, and w_k is approximately 0.17, indicating that the model tends to rely more on data-driven features for judgment. However, when the system encounters unseen fault types or the data distribution shifts, w_d decreases to 0.31, w_p increases to 0.38, and w_k increases to 0.31, indicating that the model automatically enhances the role of physical constraints and knowledge reasoning, thereby compensating for the shortcomings of the data-driven method in unknown scenarios. This phenomenon verifies that the gating mechanism has good adaptability under different operating conditions and can effectively improve the robustness of the model.

[0037] Based on the fused feature vector F_fused, this implementation further constructs a multi-task output structure to simultaneously complete three tasks: health status assessment, degradation trend prediction, and fault diagnosis. The health status assessment task outputs discrete level labels through a fully connected layer to describe the current operating status of the system. The degradation trend prediction task outputs the changing trends of key parameters over a future period through regression to provide early warning of potential faults. The fault diagnosis task outputs the probability distribution of fault types through classification and combines the causal links provided by the knowledge enhancement channel to achieve interpretable diagnosis, thus forming a complete "detection-prediction-diagnosis" closed-loop system. To achieve multi-task collaborative optimization, this implementation design uses a multi-objective loss function L_total = alpha·L_health + beta·L_pred + gamma·L_diag + delta·L_physics + epsilon·L_knowledge + zeta·L_reg… where L_health represents the health status assessment loss, L_pred represents the degradation prediction error, L_diag represents the fault diagnosis loss, L_physics represents the physical constraint loss, L_knowledge represents the knowledge consistency loss, and L_reg represents the regularization term. These loss terms are weighted and combined using weight parameters to maintain the overall model's balance while ensuring the performance of each task. By introducing physical constraint loss and knowledge consistency loss, the model focuses not only on data fitting but also on physical rationality and knowledge consistency during optimization, effectively avoiding overfitting.

[0038] Regarding the training strategy, this implementation adopts a four-stage step-by-step training method to address potential convergence instability issues during multi-channel model training. Specifically, the four stages are: first, pre-training the physical information channel to gradually converge the learnable physical parameters and satisfy basic physical laws; second, training the knowledge enhancement channel to fully learn knowledge graph embedding and graph reasoning capabilities; third, training the data-driven channel to fully exploit the temporal features in telemetry data; and fourth, jointly fine-tuning the entire model, achieving collaborative work between channels through end-to-end optimization. This step-by-step training strategy effectively avoids gradient interference between different information sources and improves the overall performance of the model. After model training, a test dataset containing 12 fault modes was selected for performance evaluation. The results show that, under known fault scenarios, the diagnostic accuracy of this implementation method reaches 96.3%, significantly higher than the 89.5% of the pure data-driven method and the 93.1% of the data and physical fusion method. Under unknown conditions, the F1 score reaches 91.2%, the false alarm rate is only 1.7%, and the missed alarm rate is 2.4%, all of which are significantly lower than the comparison methods, indicating that this method has stronger generalization ability and stability in complex environments. In particular, in terms of the false alarm rate, compared with the 8.3% of the pure data-driven method and the 4.2% of the dual-channel method, this method significantly reduces the false alarm probability, thereby reducing unnecessary operation and maintenance intervention costs. In summary, by introducing an adaptive gating fusion mechanism, a multi-task collaborative optimization strategy, and a phased training method, this implementation method achieves efficient fusion and utilization of multi-source information, which not only significantly improves diagnostic accuracy but also enhances the robustness and interpretability of the model under complex conditions, fully demonstrating the engineering application value and promotion prospects of this application in the field of satellite on-orbit health management.

[0039] This application provides, as follows: Figure 2 The satellite on-orbit health diagnosis system shown is based on the physics-knowledge-data ternary fusion, including a multi-source data processing module, a time-series feature extraction module, a physical constraint modeling module, a knowledge graph reasoning module, and a ternary fusion diagnosis module; The multi-source data processing module acquires multi-source telemetry time-series data from the satellite in orbit, performs outlier detection and removal, missing value interpolation, time series alignment and normalization on the multi-source telemetry time-series data to obtain standardized telemetry data; at the same time, it constructs a three-layer satellite fault knowledge graph that includes component-level failure modes, subsystem-level fault propagation paths and whole-satellite fault coupling relationships. The temporal feature extraction module establishes a temporal feature extraction network containing multi-scale temporal convolutional structures and self-attention structures for standardized telemetry data. It extracts temporal features at different time scales through multiple parallel causal dilated convolutional branches and uses the self-attention structure to characterize the correlation between telemetry parameters, thereby obtaining data-driven feature vectors. The physical constraint modeling module discretizes the physical equations of at least two key subsystems of the satellite and embeds them into a neural network to form a learnable constraint layer. Standardized telemetry data is input into the learnable constraint layer. By adaptively calibrating the key physical parameters in the physical equations and calculating the weighted error of the physical equation residuals, a physical information feature vector is obtained. The knowledge graph reasoning module embeds the satellite fault knowledge graph into vectors and establishes a graph attention reasoning structure. It uses data-driven feature vectors as queries to perform reasoning in the knowledge graph embedding space, and obtains knowledge-enhanced feature vectors and corresponding fault causal links. The ternary fusion diagnostic module concatenates data-driven feature vectors, physical information feature vectors, and knowledge-enhanced feature vectors and inputs them into a gating network. It outputs a normalized weight vector and then performs weighted fusion of the three types of features based on the normalized weight vector to obtain a fused feature vector. The fused feature vector is then input into a multi-task output structure to complete health status assessment, degradation trend prediction, and fault diagnosis.

[0040] The present invention provides a satellite on-orbit health diagnosis method based on the ternary fusion of physics, knowledge, and data, which is implemented through the aforementioned satellite on-orbit health diagnosis system based on the ternary fusion of physics, knowledge, and data. For details of the specific method and process of the satellite on-orbit health diagnosis system based on the ternary fusion of physics, knowledge, and data, please refer to the above-mentioned embodiment of the satellite on-orbit health diagnosis method based on the ternary fusion of physics, knowledge, and data, which will not be repeated here.

[0041] The foregoing has only described certain exemplary embodiments of this application by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of this application. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of this application.

Claims

1. A satellite on-orbit health diagnosis method based on the ternary fusion of physics, knowledge, and data, characterized in that, Includes the following steps: Acquire multi-source telemetry time-series data from satellites in orbit, and perform outlier detection and removal, missing value imputation, time series alignment and normalization on the multi-source telemetry time-series data to obtain standardized telemetry data; Simultaneously, a three-layer satellite fault knowledge graph is constructed, including component-level failure modes, subsystem-level fault propagation paths, and overall satellite-level fault coupling relationships. For standardized telemetry data, a temporal feature extraction network containing multi-scale temporal convolutional structure and self-attention structure is established. Temporal features at different time scales are extracted through multiple parallel causal dilated convolutional branches, and the correlation between telemetry parameters is represented by the self-attention structure to obtain data-driven feature vectors. The physical equations of at least two key subsystems of the satellite are discretized and embedded into a neural network to form a learnable constraint layer. Standardized telemetry data is input into the learnable constraint layer to calculate the weighted error of the physical equation residuals. Based on the weighted error, the key physical parameters in the physical equations are adaptively calibrated through an error backpropagation mechanism. The physical information feature vector is obtained from the learnable constraint layer. Vectorize and embed the satellite fault knowledge graph and establish a graph attention reasoning structure. Use data-driven feature vectors as queries to perform reasoning in the knowledge graph embedding space to obtain knowledge-enhanced feature vectors and corresponding fault causal links. Data-driven feature vectors, physical information feature vectors, and knowledge-enhanced feature vectors are concatenated and input into a gating network to output a normalized weight vector. The three types of features are then weighted and fused based on the normalized weight vector to obtain a fused feature vector. The fused feature vector is then input into a multi-task output structure to complete health status assessment, degradation trend prediction, and fault diagnosis.

2. The satellite on-orbit health diagnosis method based on the ternary fusion of physics, knowledge, and data as described in claim 1, characterized in that, The multi-source telemetry data processing adopts an anomaly detection mechanism based on sliding window statistics. The sliding window length is set to 100 to 500 sampling points. The anomaly judgment threshold is determined by calculating the mean and standard deviation within the window. Data points exceeding the threshold range are marked and replaced by interpolation using neighboring normal data. At the same time, missing data segments are filtered in conjunction with the continuous missing length judgment rule.

3. The satellite on-orbit health diagnosis method based on the ternary fusion of physics, knowledge, and data as described in claim 1, is characterized in that... The multi-scale temporal convolutional structure contains 2 to 6 parallel branches. The expansion rate of each branch increases exponentially to construct feature representation paths at different time scales. It performs hierarchical modeling of second-level, minute-level, and hour-level change information, and achieves feature fusion through channel splicing and point-by-point convolution.

4. The satellite on-orbit health diagnosis method based on the ternary fusion of physics, knowledge, and data as described in claim 1, characterized in that, The self-attention structure employs a multi-head mechanism to construct a multi-subspace feature interaction mode. By mapping input features to query vectors, key vectors, and value vectors, it calculates the correlation weights between different time steps and different parameters, dynamically strengthening the expression of the coupling relationship between key telemetry variables.

5. The satellite on-orbit health diagnosis method based on the ternary fusion of physics, knowledge, and data as described in claim 1, characterized in that, The physical equation discretization process uses an explicit numerical integration method to convert the continuous-time model into a discrete form. At the same time, the key parameters of the thermal control subsystem, power supply subsystem, and attitude control subsystem are set as trainable variables, and the parameters are adaptively corrected through the error backpropagation mechanism.

6. The satellite on-orbit health diagnosis method based on the ternary fusion of physics, knowledge, and data as described in claim 1, characterized in that, The physical constraint error is constructed using a multi-subsystem joint weighting method, which aggregates the residuals of different physical equations according to their importance by assigning weights. The constraint effect of key subsystems is strengthened by introducing a subsystem weight adjustment mechanism, and global consistency constraints are achieved by combining sample-level error statistics.

7. The satellite on-orbit health diagnosis method based on the ternary fusion of physics, knowledge, and data as described in claim 1, characterized in that, The knowledge graph embedding process adopts a vector representation method based on translation relationship to encode entities and relations in low-dimensional vectors. By introducing a negative sampling mechanism, a contrastive learning target is constructed to enhance the semantic distinguishability between different fault modes.

8. The satellite on-orbit health diagnosis method based on the ternary fusion of physics, knowledge, and data as described in claim 1, characterized in that, The graph attention inference structure uses a multi-head attention mechanism to weighted aggregate the features of neighboring nodes, dynamically allocates the contribution ratio of neighboring nodes to the representation of the current node using attention coefficients, and combines nonlinear mapping to update node features, thus characterizing the differences in the influence intensity between different nodes in the fault propagation path.

9. The satellite on-orbit health diagnosis method based on the ternary fusion of physics, knowledge, and data as described in claim 1, characterized in that, The gated network adopts a two-layer fully connected structure to construct a feature weight generation mechanism. It obtains a weight vector by performing nonlinear mapping on the concatenation results of three types of features and then normalizing them. This weight vector automatically adjusts the contribution ratio of the three types of information sources as the distribution of input data changes.

10. A satellite on-orbit health diagnosis system based on physical-knowledge-data ternary fusion, used to implement the satellite on-orbit health diagnosis method based on physical-knowledge-data ternary fusion as described in any one of claims 1-9, characterized in that, It includes a multi-source data processing module, a time-series feature extraction module, a physical constraint modeling module, a knowledge graph reasoning module, and a ternary fusion diagnostic module; The multi-source data processing module acquires multi-source telemetry time-series data from the satellite in orbit, performs outlier detection and removal, missing value interpolation, time series alignment and normalization on the multi-source telemetry time-series data, and obtains standardized telemetry data. Simultaneously, a three-layer satellite fault knowledge graph is constructed, including component-level failure modes, subsystem-level fault propagation paths, and overall satellite-level fault coupling relationships. The temporal feature extraction module establishes a temporal feature extraction network containing multi-scale temporal convolutional structures and self-attention structures for standardized telemetry data. It extracts temporal features at different time scales through multiple parallel causal dilated convolutional branches and uses the self-attention structure to characterize the correlation between telemetry parameters, thereby obtaining data-driven feature vectors. The physical constraint modeling module discretizes the physical equations of at least two key subsystems of the satellite and embeds them into a neural network to form a learnable constraint layer. Standardized telemetry data is input into the learnable constraint layer to calculate the weighted error of the physical equation residuals. Based on the weighted error, the key physical parameters in the physical equations are adaptively calibrated through an error backpropagation mechanism. The physical information feature vector is obtained from the learnable constraint layer. The knowledge graph reasoning module embeds the satellite fault knowledge graph into vectors and establishes a graph attention reasoning structure. It uses data-driven feature vectors as queries to perform reasoning in the knowledge graph embedding space, and obtains knowledge-enhanced feature vectors and corresponding fault causal links. The ternary fusion diagnostic module concatenates data-driven feature vectors, physical information feature vectors, and knowledge-enhanced feature vectors and inputs them into a gating network. It outputs a normalized weight vector and then performs weighted fusion of the three types of features based on the normalized weight vector to obtain a fused feature vector. The fused feature vector is then input into a multi-task output structure to complete health status assessment, degradation trend prediction, and fault diagnosis.

Images