In-orbit satellite intelligent fault diagnosis system based on deep learning

Abstract

The invention discloses an in-orbit satellite intelligent fault diagnosis system based on deep learning, and relates to the technical field of in-orbit satellite intelligent fault diagnosis, and the system comprises the steps: carrying out the data regularization processing of multi-source heterogeneous telemetry data generated in the in-orbit operation process of a satellite; capturing a weak abnormal mode at the early stage of the fault; identifying a fault mode in the operation process of the on-orbit satellite; and fault emergency levels are divided according to the fault emergency degrees. According to the method, efficient regularization of on-orbit satellite telemetry data, automatic mining of on-satellite lightweight depth features, online incremental learning recognition of unknown fault modes and intelligent hierarchical scheduling of diagnosis tasks are realized; the accuracy, timeliness, autonomy and generalization ability of on-orbit satellite fault diagnosis are improved, and the technical bottlenecks that a traditional method depends on manual rules, is difficult to deal with novel composite faults, lacks deep data mining ability, is difficult to deploy on a satellite and the like are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of on-orbit satellite intelligent fault diagnosis technology, and in particular to an on-orbit satellite intelligent fault diagnosis system based on deep learning. Background Technology

[0002] With the rapid development of aerospace technology and the continuous expansion of satellite applications, humanity's reliance on satellites in orbit is deepening. Satellite systems play an irreplaceable role in key areas such as communication, navigation, remote sensing, meteorological monitoring, and national defense security. At the same time, the complexity of satellite missions and the harshness of operating environments have significantly increased, leading to higher risks of failure during satellite operation. Therefore, society's demand for reliable, autonomous, and intelligent operation and maintenance capabilities of satellite systems is becoming increasingly urgent. There is a pressing need for efficient, accurate, and adaptive fault diagnosis methods to ensure long-term stable satellite operation and reduce ground intervention costs.

[0003] However, current mainstream on-orbit satellite fault diagnosis technologies still mainly rely on rule-based expert systems or traditional signal processing methods. These methods generally suffer from problems such as strong dependence on prior knowledge, poor generalization ability, and difficulty in dealing with unknown or complex faults. In addition, most existing systems lack the ability to deeply mine massive amounts of telemetry data, failing to effectively capture weak early-stage fault characteristics, resulting in insufficient diagnostic timeliness and accuracy. At the same time, under the condition of limited on-orbit computing resources, traditional algorithms often cannot achieve lightweight deployment and real-time response, restricting the improvement of satellite autonomous health management capabilities. Summary of the Invention

[0004] In view of the problems existing in the deep learning-based on-orbit intelligent fault diagnosis system for satellites, this invention is proposed.

[0005] Therefore, the problem to be solved by this invention is: how to achieve efficient, autonomous, accurate and adaptive intelligent fault diagnosis of multi-source telemetry data of on-orbit satellites under the condition of limited on-board computing resources, so as to overcome the defects of existing technologies such as reliance on prior rules, weak generalization ability, difficulty in capturing early weak faults, inability to cope with unknown faults, and insufficient real-time performance and autonomy.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide an on-orbit intelligent fault diagnosis system for satellites based on deep learning, which includes a telemetry data preprocessing module for performing data normalization processing on multi-source heterogeneous telemetry data generated during the on-orbit operation of the satellite. The lightweight deep feature extraction module is used to deploy a deep neural network on the embedded processing unit on the satellite itself to automatically mine high-dimensional nonlinear features in telemetry time series data and capture weak anomaly patterns in the early stages of a fault. The adaptive fault identification module is used to dynamically adjust the internal parameters of the deep neural network based on an online incremental learning mechanism to identify fault modes during the operation of the satellite in orbit. The edge and cloud collaborative inference scheduling module is used to classify the urgency of identified fault modes and classify the urgency level of faults based on the urgency level.

[0007] As a preferred embodiment of the deep learning-based on-orbit satellite intelligent fault diagnosis system of the present invention, the telemetry data preprocessing module includes a multi-source data alignment submodule, a telemetry data standardization submodule, a noise suppression submodule, and a feature structure alignment submodule. The multi-source data alignment submodule is used to synchronize the time and unify the dimensions of the telemetry data stream, thereby eliminating the problem of telemetry data stream misalignment. The telemetry data standardization submodule is used to normalize the telemetry data based on the telemetry data of each telemetry data stream, so that telemetry variables of different magnitudes and units of telemetry data are mapped to a unified numerical range. The noise suppression submodule is used to suppress non-fault-related disturbances in the telemetry sequence of telemetry variables using an adaptive multi-scale filtering method, while retaining the signal components of effective telemetry variables in the mapping process; The feature structure alignment submodule is used to segment and reassemble telemetry data according to a preset diagnostic time window, and fill in the signal components of the failed telemetry variables.

[0008] As a preferred embodiment of the deep learning-based on-orbit satellite intelligent fault diagnosis system of the present invention, the lightweight deep feature extraction module includes a network structure compression submodule, a quantization encoding submodule, a knowledge distillation integration submodule, and a temporal anomaly perception submodule. The network structure compression submodule is used to perform channel pruning, layer fusion and parameter sparsification on the deep neural network to generate a compact network topology that adapts to the computing power and storage capacity of the spaceborne embedded processing unit. The quantization encoding submodule is used to convert the compressed network weights and activation values ​​from floating-point representation to low-bit integer format, and combine fixed-point arithmetic strategy to optimize the inference process, thereby reducing computational energy consumption and memory usage. The knowledge distillation integration submodule is used to supervise the training of the lightweight target network through the reference network, and transfer the temporal dependencies and fault representation capabilities learned by the reference network to the target network. The time-series anomaly perception submodule is used to perform sliding window feature mapping on the normalized telemetry time-series data based on a lightweight network to extract potential anomaly patterns in multi-scale time contexts.

[0009] As a preferred embodiment of the deep learning-based on-orbit satellite intelligent fault diagnosis system of the present invention, the adaptive fault identification module includes an incremental sample screening submodule, a parameter local update submodule, a fault category expansion submodule, and a model stability maintenance submodule. The incremental sample screening submodule is used to perform confidence evaluation and distribution offset detection on the abnormal feature vectors output by the lightweight deep feature extraction module, and to screen out new samples that have diagnostic value and deviate from the known fault categories. The parameter local update submodule is used to perform gradient adjustment on the connection weights of local neurons in the deep neural network that are associated with the current fault mode based on the newly added samples after screening. The fault category extension submodule is used to automatically initialize new fault category nodes and integrate them into the existing classification structure when a new abnormal pattern is detected. The model stability maintenance submodule is used to suppress the network's forgetting effect on old failure modes by applying regularization constraints after each incremental update.

[0010] As a preferred embodiment of the deep learning-based on-orbit satellite intelligent fault diagnosis system of the present invention, the edge and cloud collaborative reasoning and scheduling module includes a fault urgency assessment submodule, a resource status perception submodule, a hierarchical decision execution submodule, and a task distribution control submodule. The fault urgency assessment submodule is used to comprehensively analyze the fault characterization index of the current fault based on the fault mode type, anomaly confidence level and the degree of deviation of telemetry parameters from the threshold output by the adaptive fault identification module, and generate an urgency score. The resource status awareness submodule is used to monitor the onboard resource load status of the satellite embedded processing unit and obtain the satellite-to-ground communication quality indicators of the satellite-to-ground communication link. The hierarchical decision execution submodule is used to classify faults into three urgency levels: high, medium, and low, based on the fault urgency score and resource status information. For high-level faults, onboard local real-time diagnosis and response are triggered; for medium-level faults, lightweight backhaul and ground-assisted confirmation are initiated; for low-level faults, data is cached and uploaded to the ground cloud for in-depth analysis when appropriate. The task distribution control submodule is used to dynamically generate inference task instructions based on the urgency level of the fault, schedule the on-board lightweight deep feature extraction and fault identification processes to be executed in orbit, and upload the encapsulated telemetry feature package to the ground data center.

[0011] As a preferred embodiment of the deep learning-based on-orbit intelligent fault diagnosis system for satellites described in this invention, the telemetry feature packets generated by the edge and cloud collaborative reasoning scheduling module are used to upload the task and the deep analysis results returned from the ground. Together with the incremental sample screening submodule and the parameter local update submodule in the adaptive fault identification module, a satellite-ground diagnostic feedback fusion submodule is derived. The satellite-ground diagnostic feedback fusion submodule is used to receive the deep analysis results returned by the ground data center, perform semantic alignment and confidence fusion with the on-board fault identification results, and trigger local parameter updates or fault category expansion when a new or high-confidence fault mode is detected.

[0012] As a preferred embodiment of the deep learning-based on-orbit satellite intelligent fault diagnosis system of the present invention, the fault urgency assessment submodule includes an introduced fault urgency score calculation formula to quantify the urgency score, the specific formula being:

[0013] in, Indicates the urgency score of the fault. This indicates the preset hazard weight corresponding to the current failure mode type. This represents the anomaly confidence level output by the adaptive fault identification module. This represents the maximum normalized deviation of key telemetry parameters from their normal thresholds. Indicates the weighting coefficient for the failure mode type. This represents the weighting coefficient for anomaly confidence. This represents the parameter deviation weighting coefficient.

[0014] Secondly, embodiments of the present invention provide an intelligent fault diagnosis method for on-orbit satellites based on deep learning, comprising: performing data normalization processing on multi-source heterogeneous telemetry data generated during the on-orbit operation of the satellite; deploying a deep neural network on the embedded processing unit on the satellite itself to automatically mine high-dimensional nonlinear features in the telemetry time series data and capture weak anomaly patterns in the early stages of faults; dynamically adjusting the internal parameters of the deep neural network based on an online incremental learning mechanism to identify fault modes during the operation of the on-orbit satellite; classifying the fault urgency of the identified fault modes and classifying the fault urgency level according to the fault urgency.

[0015] Thirdly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the steps of the above-described deep learning-based on-orbit intelligent fault diagnosis system for satellites.

[0016] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any of the steps of the above-described deep learning-based on-orbit intelligent fault diagnosis system for satellites.

[0017] The beneficial effects of this invention are as follows: By constructing four core modules—telemetry data preprocessing, lightweight deep feature extraction, adaptive fault identification, and edge-cloud collaborative reasoning scheduling—and further refining them into multiple functional sub-modules, this invention achieves efficient regularization of on-orbit satellite telemetry data, automatic mining of lightweight deep features on satellite, online incremental learning and identification of unknown fault modes, and intelligent hierarchical scheduling of diagnostic tasks. The system can not only capture weak anomalies in the early stages of faults in real time on resource-constrained embedded platforms, but also dynamically allocate satellite-ground computing resources according to the urgency of the fault, and continuously optimize model performance through a satellite-ground diagnostic feedback fusion mechanism. This significantly improves the accuracy, timeliness, autonomy, and generalization ability of on-orbit satellite fault diagnosis, effectively solving the technical bottlenecks of traditional methods such as reliance on manual rules, difficulty in dealing with new complex faults, lack of deep data mining capabilities, and difficulties in on-orbit deployment. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a schematic diagram of an on-orbit intelligent fault diagnosis system for satellites based on deep learning, provided as an embodiment of the present invention.

[0019] Figure 2 This is a flowchart of a method for an on-orbit intelligent fault diagnosis system for satellites based on deep learning, provided as an embodiment of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0024] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0025] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0026] Example Reference Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides an on-orbit intelligent fault diagnosis system for satellites based on deep learning, comprising: S1: Telemetry data preprocessing module, used to perform data normalization processing on multi-source heterogeneous telemetry data generated during satellite operation.

[0027] The telemetry data preprocessing module includes a multi-source data alignment submodule, a telemetry data standardization submodule, a noise suppression submodule, and a feature structure alignment submodule. The multi-source data alignment submodule is used to synchronize the time and unify the dimensions of telemetry data streams, eliminating the problem of telemetry data stream misalignment. The telemetry data standardization submodule is used to normalize the telemetry data based on the telemetry data from each telemetry data stream, so that telemetry variables of different magnitudes and units are mapped to a unified numerical range. The noise suppression submodule is used to suppress non-fault-related disturbances in the telemetry sequence of telemetry variables using an adaptive multi-scale filtering method, while preserving the signal components of effective telemetry variables in the mapping process; The feature structure alignment submodule is used to segment and reassemble telemetry data according to a preset diagnostic time window, filling in the signal components of failed telemetry variables.

[0028] Furthermore, the telemetry data preprocessing module systematically processes the multi-source heterogeneous telemetry data generated during satellite operation, sequentially completing four key steps: time alignment, numerical normalization, noise suppression, and structural reconstruction. First, the multi-source data alignment submodule synchronizes the time of telemetry data streams from different sensors with different sampling frequencies and timestamps, unifying the characteristic dimensions of each data stream to a consistent format, eliminating data misalignment caused by asynchronous acquisition or interface differences. Subsequently, the telemetry data standardization submodule normalizes the raw data based on the historical statistical characteristics or physical dimensions of each telemetry parameter, compressing telemetry variables of different magnitudes and units to the same value. The numerical range provides a scale-consistent input for subsequent deep learning models. Next, the noise suppression submodule uses an adaptive multi-scale filtering method to effectively filter out non-fault-related noise introduced by transmission interference, sensor drift, or environmental disturbances in the telemetry sequence, while retaining effective signal components that reflect the true system state. Finally, the feature structure alignment submodule segments and reassembles the processed telemetry data according to a preset diagnostic time window, and interpolates or fills in missing or invalid data points caused by communication interruption or sensor failure, outputting telemetry data blocks with complete structure, continuous temporal sequence, and regular format, providing high-quality, directly inference-ready input data for the lightweight deep feature extraction module.

[0029] Furthermore, the telemetry data preprocessing module performs progressively refined processing of multi-source heterogeneous telemetry data generated during satellite operation in a pipeline manner: First, addressing the issues of inconsistent sampling rates, timestamp misalignments, and feature dimension differences in telemetry data streams from multiple subsystems such as attitude control, power management, and thermal control due to different sensor types, the multi-source data alignment submodule synchronizes each data stream on a unified time axis through interpolation or resampling, and maps all variables to a predefined standardized feature space to ensure consistency in subsequent processing; based on this, the telemetry data standardization submodule, according to the physical meaning and historical operating range of each type of telemetry parameter, uses min-max normalization or Z-score standardization methods to convert the original values ​​into a dimensionless and evenly distributed unified scale, avoiding high... The magnitude of parameters dominates model training, masking weak but crucial anomalous signals. Subsequently, the noise suppression submodule introduces an adaptive multi-scale filtering method, dynamically adjusting the filtering intensity based on the local characteristics of the data. This effectively suppresses non-fault disturbances such as high-frequency electronic noise, communication jitter, and slow drift, while preserving subtle trend changes related to the actual fault evolution to the greatest extent. Finally, the feature structure alignment submodule slices the continuous telemetry sequence according to a fixed-length diagnostic time window. For missing values ​​within the window caused by link interruption or temporary sensor failure, a linear interpolation or moving average filling strategy based on nearby valid points is used to repair them. Ultimately, a batch of telemetry data blocks with regular structure, complete time sequence, standardized values, and controllable noise is output, providing a stable, reliable, and high-quality input that can be directly used for inference for the onboard lightweight deep feature extraction module.

[0030] S2: Lightweight deep feature extraction module, used to deploy deep neural networks on the embedded processing unit on the satellite itself to automatically mine high-dimensional nonlinear features in telemetry time series data and capture weak anomaly patterns in the early stages of a fault.

[0031] The lightweight deep feature extraction module includes a network structure compression submodule, a quantization encoding submodule, a knowledge distillation integration submodule, and a temporal anomaly perception submodule. The network structure compression submodule is used to perform channel pruning, layer fusion and parameter sparsification on deep neural networks to generate a compact network topology that adapts to the computing power and storage capacity of the spaceborne embedded processing unit. The quantization encoding submodule is used to convert the compressed network weights and activation values ​​from floating-point representation to low-bit integer format, and combines fixed-point arithmetic strategy to optimize the inference process, reduce computational energy consumption and memory usage. The knowledge distillation integration submodule is used to supervise the training of the lightweight target network through the reference network, and transfer the temporal dependencies and fault representation capabilities learned by the reference network to the target network. The temporal anomaly perception submodule is used to perform sliding window feature mapping on the normalized telemetry time series data based on a lightweight network to extract potential anomaly patterns in multi-scale temporal context.

[0032] Furthermore, the lightweight deep feature extraction module addresses the strict resource constraints of the onboard embedded processing unit by constructing an efficient, compact, and highly representative on-orbit feature extraction mechanism through four collaborative sub-modules: First, the network structure compression sub-module structurally simplifies the original deep neural network by employing channel pruning to remove redundant convolutional channels, layer fusion to merge functionally similar network layers, and introducing a parameter sparsity strategy to significantly reduce the number of model parameters and computational complexity, generating a compact network topology that matches the computing power and storage capacity of the satellite embedded platform. Based on this, the quantization encoding sub-module converts the floating-point weights and activation values ​​in the compressed network into 8-bit or lower precision integer representations and reconstructs the inference process using fixed-point arithmetic rules, significantly reducing memory bandwidth requirements and power consumption while ensuring... Numerical stability is maintained to preserve diagnostic performance. To further compensate for the loss of expressive power caused by model compression, the knowledge distillation integration submodule introduces a high-precision reference network (usually pre-trained on the ground). The soft labels or intermediate layer features output by the reference network are used as supervision signals to retrain the lightweight target network, thereby effectively transferring the complex temporal dependencies and deep fault semantic information mastered by the reference network to the small network deployed on the satellite. Finally, the temporal anomaly perception submodule, based on the optimized lightweight network, performs sliding window-style segmented inference on the normalized time-series data output by the telemetry data preprocessing module. Through multi-scale temporal context modeling, it automatically captures weak, early, or gradual anomaly patterns hidden in the background of normal fluctuations, providing highly discriminative feature inputs for subsequent adaptive fault identification.

[0033] Furthermore, the lightweight deep feature extraction module, centered on the core objectives of on-orbit usability, high efficiency, and accuracy, constructs a complete technical chain from model design and numerical optimization to knowledge transfer and anomaly detection: First, at the model structure level, the network structure compression submodule systematically slims down the initial deep neural network by identifying and pruning convolutional channels with low contribution to fault diagnosis, merging adjacent network layers with overlapping functions, and resetting a large number of weights to zero to achieve sparsity, thereby generating a lightweight backbone network with low computational cost, few parameters, low memory consumption, and a topology adapted to the hardware characteristics of the onboard embedded processing unit; Next, at the numerical representation level, the quantization encoding submodule performs low-bit integer mapping (such as 8-bit or 4-bit) on all floating-point weights and activation values ​​in the compressed network, and designs corresponding fixed-point arithmetic logic, enabling the entire inference process to be executed efficiently on the onboard chip without relying on a floating-point coprocessor, significantly reducing power consumption and storage overhead, while avoiding the need for a quantization-aware training strategy. Accuracy drops sharply; to compensate for the degradation of representational ability caused by structural simplification and numerical compression, the knowledge distillation integration submodule introduces a high-capacity reference network that has been fully trained on the ground. It uses the soft probability distribution or intermediate high-dimensional features of the network outputting telemetry time-series data as teacher signals to guide the lightweight target network to learn richer and more generalized fault semantic expressions under limited capacity, effectively preserving the ability to model complex dynamic behaviors. Finally, the time-series anomaly perception submodule deploys the above-mentioned optimized lightweight network on the satellite. It uses a sliding window mechanism to perform segment-by-segment forward inference on the preprocessed regular and continuous telemetry time-series data. Combining multi-scale contextual information of short-term local fluctuations and long-term trend evolution, it sensitively extracts weak anomaly features that are difficult to detect by traditional methods, such as slow degradation, intermittent jitter, or multi-parameter coupling deviation, and other early fault signs, forming a high-dimensional feature vector with strong discriminative power. This provides a reliable, real-time, and semantically rich input foundation for the subsequent adaptive fault identification module.

[0034] S3: Adaptive Fault Identification Module, used to dynamically adjust the internal parameters of the deep neural network based on an online incremental learning mechanism to identify fault modes during the operation of on-orbit satellites.

[0035] The adaptive fault identification module includes an incremental sample screening submodule, a parameter local update submodule, a fault category expansion submodule, and a model stability maintenance submodule. The incremental sample screening submodule is used to evaluate the confidence and detect the distribution shift of the abnormal feature vectors output by the lightweight deep feature extraction module, and to screen out new samples that have diagnostic value and deviate from the known fault categories. The parameter local update submodule is used to perform gradient adjustment on the connection weights of local neurons in the deep neural network that are associated with the current fault mode based on the newly added samples after filtering. The fault category extension submodule is used to automatically initialize new fault category nodes and integrate them into the existing classification structure when a new abnormal pattern is detected. The model stability maintenance submodule is used to suppress the network's forgetting effect on old failure modes by applying regularization constraints after each incremental update.

[0036] Furthermore, the adaptive fault identification module constructs a continuous learning and dynamic identification mechanism for long-term on-orbit operation scenarios through four closely coordinated sub-modules: First, the incremental sample screening sub-module performs dual discrimination on the abnormal feature vectors output by the lightweight deep feature extraction module. On the one hand, it assesses whether the confidence level of the feature vectors as a certain type of fault determined by the current model is in the fuzzy or low confidence range. On the other hand, it detects whether the distribution of the feature in the overall feature space significantly deviates from the cluster centers of known fault categories, thereby accurately screening out candidate samples that have potential diagnostic value and may represent novel or complex faults. Subsequently, the parameter local update sub-module performs gradient fine-tuning only on the local network regions (such as specific classification heads or some intermediate layer neurons) that are most semantically relevant to the newly added sample, avoiding the high computational overhead and resource consumption caused by retraining the entire network. This enables efficient and targeted model parameter adaptation on the satellite. When the system continuously receives new samples of the same type and confirms that they constitute stable new fault modes, the fault category extension submodule automatically initializes a new fault category node at the end of the existing classification structure, assigns it an independent identifier and initial decision boundary, and seamlessly integrates it into the on-board diagnostic system, enabling the system to have the ability to autonomously recognize unknown faults. At the same time, after each local update, the model stability maintenance submodule introduces regularization constraints based on weight importance (such as limiting large changes in key connections) or an implicit memory mechanism that replays a small number of representative historical samples. This enhances the model's adaptability to new faults while effectively suppressing catastrophic forgetting of learned old fault modes, thereby ensuring the stability, integrity, and continuous evolution capability of the entire fault identification system during long-term on-orbit operation.

[0037] S4: Edge and cloud collaborative reasoning scheduling module, used to classify the severity of identified fault modes and classify the severity level of faults based on the severity level.

[0038] The edge and cloud collaborative reasoning scheduling module includes a fault urgency assessment submodule, a resource status awareness submodule, a hierarchical decision execution submodule, and a task distribution control submodule. The fault urgency assessment submodule is used to comprehensively analyze the fault characterization index of the current fault based on the fault mode type, anomaly confidence and the degree of deviation of telemetry parameters from the threshold output by the adaptive fault identification module, and generate an urgency score. The resource status awareness submodule is used to monitor the onboard resource load status of the satellite embedded processing unit and obtain the satellite-to-ground communication quality indicators of the satellite-to-ground communication link. The hierarchical decision execution submodule is used to classify faults into three urgency levels: high, medium, and low, based on the fault urgency score and resource status information. For high-level faults, onboard local real-time diagnosis and response are triggered; for medium-level faults, lightweight backhaul and ground-assisted confirmation are initiated; for low-level faults, data is cached and uploaded to the ground cloud for in-depth analysis when appropriate. The task distribution control submodule is used to dynamically generate inference task instructions based on the urgency level of the fault, schedule the on-board lightweight deep feature extraction and fault identification processes to be executed in orbit, and upload the encapsulated telemetry feature package to the ground data center.

[0039] Furthermore, the edge and cloud collaborative reasoning scheduling module constructs an intelligent task scheduling mechanism oriented towards on-orbit resource constraints and diagnostic timeliness requirements through four organically linked sub-modules: First, the fault urgency assessment sub-module integrates the fault mode type output by the adaptive fault identification module (such as power anomaly, attitude loss, etc.), the anomaly confidence of the identification result, and the degree to which key telemetry parameters deviate from normal thresholds to form a comprehensive fault characterization index that integrates multi-dimensional criteria, and generates a quantitative urgency score to objectively measure the potential harm and urgency of the current fault; simultaneously, the resource status perception sub-module collects the onboard resource load status of the satellite embedded processing unit in real time (including CPU utilization, memory availability, and energy supply level), and simultaneously obtains the current satellite-to-ground communication link communication quality indicators (such as bandwidth availability, transmission latency, and bit error rate), providing accurate on-orbit and link resource constraints for scheduling decisions; based on this, the hierarchical decision execution sub-module jointly judges the urgency score and resource status information, and assigns the fault to the appropriate execution mechanism. The system is intelligently categorized into three emergency levels: high, medium, and low. For high-level faults (such as severe anomalies that could lead to satellite instability or functional loss), onboard local full-process diagnosis and emergency response are immediately triggered to ensure millisecond-level handling. For medium-level faults (such as anomalies that cause performance degradation but are still controllable), lightweight feature data backhaul is initiated, and the ground data center is requested to assist in confirmation and provide strategy suggestions. For low-level faults (such as occasional minor fluctuations or modes awaiting verification), relevant telemetry segments are temporarily stored and uploaded to the ground cloud for in-depth analysis when the communication window is idle or energy is sufficient. Finally, the task distribution and control submodule dynamically generates specific inference task instructions based on the final determined fault emergency level: for high-level tasks, the onboard lightweight deep feature extraction and adaptive fault identification process is scheduled for closed-loop execution in orbit; for medium and low-level tasks, the pre-processed and feature-extracted telemetry feature packages are encapsulated and uploaded to the ground data center via the satellite-to-ground link, achieving efficient collaboration between rapid edge response and in-depth cloud analysis, thereby maximizing the overall diagnostic efficiency and reliability of the system with limited onboard resources.

[0040] Furthermore, the edge and cloud collaborative inference scheduling module, based on the core principles of on-demand allocation and dynamic collaboration, constructs a closed-loop, adaptive satellite-ground joint diagnostic scheduling system. First, the fault urgency assessment submodule does not judge fault severity in isolation, but deeply integrates multi-dimensional outputs from the adaptive fault identification module. This includes the specific fault mode type identified (e.g., solar panel drive failure, attitude control flywheel anomaly), the model's confidence level in the judgment (reflecting the reliability of the result), and the magnitude and duration of deviations of key telemetry parameters (e.g., voltage, temperature, angular velocity) from their normal operating thresholds. Weighted fusion forms a quantitative score that comprehensively characterizes the current potential risk and evolution speed of the fault. Simultaneously, the resource status awareness submodule continuously monitors the satellite's real-time operating environment. On one hand, it collects the onboard resource load status of the embedded processing unit (e.g., whether the current CPU utilization is close to the upper limit, whether available memory is sufficient, and whether battery power is in low-power mode). On the other hand, it simultaneously acquires the satellite-ground link communication quality indicators (e.g., whether the current downlink bandwidth is occupied by remote sensing data, whether the link latency exceeds the tolerance threshold, and whether the bit error rate has increased). This allows for a comprehensive understanding of the on-orbit execution conditions integrating computing power, energy, and communication. Based on this information, the hierarchical decision-making execution submodule makes joint decisions: when the fault urgency is high and on-board resources allow, it is classified as a high-level fault, immediately activating the local full-link diagnostic process to achieve a millisecond-level closed-loop response from feature extraction to fault confirmation to emergency command generation; if the fault has a certain degree of uncertainty or moderate harm, it is classified as a medium-level fault, and only lightweight feature summaries or key telemetry segments are compressed and transmitted back for review and feedback from the ground expert system, avoiding invalid full uploads; while for... Anomalies that are low-risk, sporadic, or suspected to be caused by noise are classified as low-level and temporarily stored in the onboard buffer. They are then uploaded in batches to the ground data center for offline in-depth analysis during off-peak communication periods or when there is a surplus of energy. Finally, the task distribution and control submodule accurately generates corresponding execution instructions based on this classification result. For high-level tasks, the onboard lightweight deep feature extraction and adaptive fault identification module is invoked to complete end-to-end on-orbit inference. For medium- and low-level tasks, the extracted regularized feature data is packaged, encrypted, and uploaded at an opportune time through the scheduling communication module, ensuring efficient utilization of space-ground resources. The entire process achieves an intelligent decision-making closed loop of what fault it is, when to handle it, where to handle it, and how to handle it, balancing the real-time performance and accuracy of diagnosis with the sustainable operation capability of the system under strictly constrained on-orbit environment.

[0041] Among them, the telemetry feature packets generated by the edge and cloud collaborative reasoning scheduling module are used to upload the task and the deep analysis results returned from the ground. Together with the incremental sample screening submodule and the parameter local update submodule in the adaptive fault identification module, the satellite-ground diagnostic feedback fusion submodule is derived. The satellite-to-ground diagnostic feedback fusion submodule is used to receive the deep analysis results returned by the ground data center, perform semantic alignment and confidence fusion with the on-board fault identification results, and trigger partial parameter updates or fault category expansion when a new or high-confidence fault mode is detected.

[0042] The fault urgency assessment submodule includes an introduction of a fault urgency score calculation formula to quantify the urgency score. The specific formula is as follows:

[0043] in, Indicates the urgency score of the fault. This indicates the preset hazard weight corresponding to the current failure mode type. This represents the anomaly confidence level output by the adaptive fault identification module. This represents the maximum normalized deviation of key telemetry parameters from their normal thresholds. Indicates the weighting coefficient for the failure mode type. This represents the weighting coefficient for anomaly confidence. This represents the parameter deviation weighting coefficient.

[0044] Furthermore, the satellite-to-ground diagnostic feedback fusion submodule, acting as a bridge connecting ground-based deep analysis capabilities and the onboard autonomous diagnostic system, is driven by the telemetry feature packet upload task triggered by the edge and cloud collaborative inference scheduling module and its corresponding ground return results. This, along with the incremental sample screening submodule and parameter local update submodule within the adaptive fault identification module, forms a closed-loop feedback mechanism: when the ground data center receives the telemetry feature packets related to medium and low-level faults uploaded from the satellite, it utilizes its powerful computing capabilities and more complete models to perform refined re-analysis of the data, generating deep diagnostic results containing high-confidence fault labels, potential root cause inferences, or novel pattern identifiers, and then transmits these results back to the satellite; the satellite-to-ground diagnostic feedback fusion submodule... Upon receiving the ground feedback, the module first aligns its semantic description (such as fault category name, code, or feature vector) with the fault ontology library currently used on the satellite to ensure consistency between satellite and ground terminology. Subsequently, it weights and fuses the confidence scores of the ground diagnostic results with the local on-board identification results to form a more reliable comprehensive judgment. When the fusion results indicate the existence of a new fault mode not covered by the on-board model, or when the ground analysis gives a significantly higher confidence score for a certain type of anomaly, the submodule will proactively activate the parameter local update submodule in the adaptive fault identification module to fine-tune the relevant network regions, or trigger the fault category extension submodule to dynamically add fault category nodes, thereby achieving knowledge feedback and model evolution from the cloud to the edge. Meanwhile, to support accurate decision-making in the aforementioned collaborative mechanism, the fault urgency assessment submodule introduces a structured scoring logic. It comprehensively considers the inherent hazard of the fault itself (reflected by preset fault mode type weights), the credibility of the current identification results, and the severity of telemetry parameters deviating from the normal state. The three factors are integrated according to configurable weight ratios to generate a unified urgency score, providing an objective, quantitative, and interpretable basis for subsequent hierarchical scheduling, and ensuring that the entire system always allocates satellite-ground resources with the optimal strategy in a complex and ever-changing on-orbit environment.

[0045] In a preferred embodiment, a deep learning-based intelligent fault diagnosis method for on-orbit satellites includes: performing data normalization processing on multi-source heterogeneous telemetry data generated during the satellite's on-orbit operation; deploying a deep neural network on the embedded processing unit on the satellite itself to automatically mine high-dimensional nonlinear features in the telemetry time-series data and capture weak anomaly patterns in the early stages of faults; dynamically adjusting the internal parameters of the deep neural network based on an online incremental learning mechanism to identify fault modes during the on-orbit satellite operation; classifying the fault urgency of the identified fault modes and classifying the fault urgency level according to the fault urgency.

[0046] The above-mentioned unit modules can be embedded in the processor of the computer device in hardware form or independent of it, or they can be stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0047] In one embodiment, a computer device is provided, which may be a terminal. The computer device includes a processor, memory, a communication interface, a display screen, and an input device connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. The communication interface of the computer device is used for wired or wireless communication with external terminals. Wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen of the computer device may be an LCD screen or an e-ink display screen. The input device of the computer device may be a touch layer covering the display screen, or buttons, a trackball, or a touchpad located on the casing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0048] In summary, this invention constructs four core modules—telemetry data preprocessing, lightweight deep feature extraction, adaptive fault identification, and edge-cloud collaborative inference scheduling—and further refines them into multiple functional sub-modules. This enables efficient regularization of on-orbit satellite telemetry data, automatic mining of lightweight on-board deep features, online incremental learning and identification of unknown fault modes, and intelligent hierarchical scheduling of diagnostic tasks. The system can not only capture subtle anomalies in the early stages of faults in real time on resource-constrained embedded platforms, but also dynamically allocate satellite-ground computing resources according to the urgency of the fault. Furthermore, it continuously optimizes model performance through a satellite-ground diagnostic feedback fusion mechanism, thereby significantly improving the accuracy, timeliness, autonomy, and generalization ability of on-orbit satellite fault diagnosis. This effectively solves the technical bottlenecks of traditional methods, such as reliance on manual rules, difficulty in handling novel complex faults, lack of deep data mining capabilities, and difficulties in on-board deployment.

[0049] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based on-orbit intelligent fault diagnosis system for satellites, characterized in that: include, The telemetry data preprocessing module is used to perform data normalization processing on multi-source heterogeneous telemetry data generated during satellite operation in orbit. The lightweight deep feature extraction module is used to deploy a deep neural network on the embedded processing unit on the satellite itself to automatically mine high-dimensional nonlinear features in telemetry time series data and capture weak anomaly patterns in the early stages of a fault. The adaptive fault identification module is used to dynamically adjust the internal parameters of the deep neural network based on an online incremental learning mechanism to identify fault modes during the operation of the satellite in orbit. The edge and cloud collaborative inference scheduling module is used to classify the urgency of identified fault modes and classify the urgency level of faults based on the urgency level.

2. The on-orbit intelligent fault diagnosis system for satellites based on deep learning as described in claim 1, characterized in that: The telemetry data preprocessing module includes a multi-source data alignment submodule, a telemetry data standardization submodule, a noise suppression submodule, and a feature structure alignment submodule. The multi-source data alignment submodule is used to synchronize the time and unify the dimensions of the telemetry data stream, thereby eliminating the problem of telemetry data stream misalignment. The telemetry data standardization submodule is used to normalize the telemetry data based on the telemetry data of each telemetry data stream, so that telemetry variables of different magnitudes and units of telemetry data are mapped to a unified numerical range. The noise suppression submodule is used to suppress non-fault-related disturbances in the telemetry sequence of telemetry variables using an adaptive multi-scale filtering method, while retaining the signal components of effective telemetry variables in the mapping process; The feature structure alignment submodule is used to segment and reassemble telemetry data according to a preset diagnostic time window, and fill in the signal components of the failed telemetry variables.

3. The on-orbit intelligent fault diagnosis system for satellites based on deep learning as described in claim 2, characterized in that: The lightweight deep feature extraction module includes a network structure compression submodule, a quantization encoding submodule, a knowledge distillation integration submodule, and a temporal anomaly perception submodule. The network structure compression submodule is used to perform channel pruning, layer fusion and parameter sparsification on the deep neural network to generate a compact network topology that adapts to the computing power and storage capacity of the spaceborne embedded processing unit. The quantization encoding submodule is used to convert the compressed network weights and activation values ​​from floating-point representation to low-bit integer format, and combine fixed-point arithmetic strategy to optimize the inference process, thereby reducing computational energy consumption and memory usage. The knowledge distillation integration submodule is used to supervise the training of the lightweight target network through the reference network, and transfer the temporal dependencies and fault representation capabilities learned by the reference network to the target network. The time-series anomaly perception submodule is used to perform sliding window feature mapping on the normalized telemetry time-series data based on a lightweight network to extract potential anomaly patterns in multi-scale time contexts.

4. The on-orbit intelligent fault diagnosis system for satellites based on deep learning as described in claim 3, characterized in that: The adaptive fault identification module includes an incremental sample filtering submodule, a parameter local update submodule, a fault category expansion submodule, and a model stability maintenance submodule. The incremental sample screening submodule is used to perform confidence evaluation and distribution offset detection on the abnormal feature vectors output by the lightweight deep feature extraction module, and to screen out new samples that have diagnostic value and deviate from the known fault categories. The parameter local update submodule is used to perform gradient adjustment on the connection weights of local neurons in the deep neural network that are associated with the current fault mode based on the newly added samples after screening. The fault category extension submodule is used to automatically initialize new fault category nodes and integrate them into the existing classification structure when a new abnormal pattern is detected. The model stability maintenance submodule is used to suppress the network's forgetting effect on old failure modes by applying regularization constraints after each incremental update.

5. The on-orbit intelligent fault diagnosis system for satellites based on deep learning as described in claim 4, characterized in that: The edge and cloud collaborative reasoning scheduling module includes a fault urgency assessment submodule, a resource status awareness submodule, a hierarchical decision execution submodule, and a task distribution control submodule. The fault urgency assessment submodule is used to comprehensively analyze the fault characterization index of the current fault based on the fault mode type, anomaly confidence level and the degree of deviation of telemetry parameters from the threshold output by the adaptive fault identification module, and generate an urgency score. The resource status awareness submodule is used to monitor the onboard resource load status of the satellite embedded processing unit and obtain the satellite-to-ground communication quality indicators of the satellite-to-ground communication link. The hierarchical decision execution submodule is used to classify faults into three urgency levels: high, medium, and low, based on the fault urgency score and resource status information. For high-level faults, onboard local real-time diagnosis and response are triggered; for medium-level faults, lightweight data backhaul and ground-assisted confirmation are initiated; for low-level faults, data is cached and uploaded to the ground cloud for in-depth analysis when appropriate. The task distribution control submodule is used to dynamically generate inference task instructions based on the fault urgency level, schedule the on-board lightweight deep feature extraction and fault identification processes to be executed in orbit, and upload the encapsulated telemetry feature package to the ground data center.

6. The on-orbit intelligent fault diagnosis system for satellites based on deep learning as described in claim 5, characterized in that: The telemetry feature packets generated by the edge and cloud collaborative reasoning scheduling module are used to upload the task and the deep analysis results returned from the ground. Together with the incremental sample screening submodule and the parameter local update submodule in the adaptive fault identification module, the satellite-ground diagnostic feedback fusion submodule is derived. The satellite-ground diagnostic feedback fusion submodule is used to receive the deep analysis results returned by the ground data center, perform semantic alignment and confidence fusion with the on-board fault identification results, and trigger partial parameter updates or fault category expansion when a new or high-confidence fault mode is detected.

7. The on-orbit intelligent fault diagnosis system for satellites based on deep learning as described in claim 6, characterized in that: The fault urgency assessment submodule includes an introduced fault urgency score calculation formula to quantify the urgency score. The specific formula is as follows: in, Indicates the urgency score of the fault. This indicates the preset hazard weight corresponding to the current failure mode type. This represents the anomaly confidence level output by the adaptive fault identification module. This represents the maximum normalized deviation of key telemetry parameters from their normal thresholds. Indicates the weighting coefficient for the failure mode type. This represents the weighting coefficient for anomaly confidence. This represents the parameter deviation weighting coefficient.

8. A deep learning-based intelligent fault diagnosis method for on-orbit satellites, based on the deep learning-based intelligent fault diagnosis system for on-orbit satellites as described in any one of claims 1 to 7, characterized in that: include, Data normalization processing is performed on multi-source heterogeneous telemetry data generated during satellite operation in orbit; Deploy deep neural networks on the embedded processing unit on the satellite itself to automatically mine high-dimensional nonlinear features in telemetry time series data and capture weak anomaly patterns in the early stages of a fault. The internal parameters of a deep neural network are dynamically adjusted based on an online incremental learning mechanism to identify fault modes during the operation of an on-orbit satellite. The identified fault modes are classified into different levels of urgency, and the urgency level is further classified based on the level of urgency.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the deep learning-based on-orbit intelligent fault diagnosis system for satellites as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the deep learning-based on-orbit intelligent fault diagnosis system for satellites as described in any one of claims 1 to 7.

Images