A method for full-link hardware deployment and verification of a satellite on-orbit fault diagnosis model

Abstract

The application discloses a kind of satellite on-orbit fault diagnosis model whole link hardware deployment and verification method, it is related to aerospace intelligent computing and system engineering technical field.The application proposes a kind of end-to-end whole link method covering six stages of data preparation, model training, model compression, hardware deployment, ground verification and on-orbit verification.Through multi-source telemetry data standardization processing and fault knowledge base construction, ternary fusion diagnosis model is trained on ground GPU platform, and model light weight is realized in combination with knowledge distillation, structured pruning and physical consistency perception mixed precision quantization.On this basis, the model is deployed to FPGA using automatic compilation tool chain, and three-level radiation hardening is realized by Scrubbing, TMR and ABFT.Meanwhile, four-level verification system of function, performance, reliability and environment is constructed, and on-orbit verification is carried out in low-orbit satellite, data and model closed-loop updating mechanism is established, so as to effectively solve the problem of process fragmentation, precision deviation and model difficult to update.

Description

Technical Field

[0001] This invention relates to the field of aerospace intelligent computing and systems engineering technology, specifically to a method for the full-link hardware deployment and verification of a satellite on-orbit fault diagnosis model. Background Technology

[0002] With the rapid development of low-Earth orbit (LEO) satellite constellations, the number of satellites has expanded from a small number to thousands or even tens of thousands, significantly increasing the complexity of constellation operation. Against this backdrop, the traditional satellite health management model, relying on centralized ground processing, has gradually revealed significant limitations, such as insufficient response timeliness, limited communication bandwidth, and excessive burden of manual analysis, making it difficult to meet the demands of efficient operation of large-scale constellations. Therefore, moving fault diagnosis capabilities to the satellite itself to achieve autonomous on-orbit health management has become an important direction for the current development of intelligent aerospace. Meanwhile, intelligent algorithms such as deep learning have demonstrated good performance in the field of fault diagnosis, providing a technological foundation for achieving high-precision automatic diagnosis. However, effectively deploying the diagnostic models trained on the ground to the onboard hardware platform and ensuring their stable operation in the complex space environment still faces many engineering challenges.

[0003] Existing technologies for hardware deployment of satellite on-orbit fault diagnosis models suffer from several shortcomings: First, there is a lack of a unified end-to-end system methodology across all stages, with a lack of standardized interfaces and processes between data processing, model training, model compression, and hardware deployment, resulting in low engineering implementation efficiency and difficulty in reproducibility. Second, there are significant discrepancies between ground training and on-orbit operation, such as the accuracy loss caused by converting floating-point models to fixed-point models, and the uncertainties introduced by environmental factors such as space radiation and temperature changes, but currently there is a lack of systematic quantification and verification methods. Third, after model deployment, there is a lack of on-orbit feedback and dynamic update mechanisms, making it difficult to adapt to new fault modes and changes in data distribution, leading to a gradual degradation of diagnostic performance. In addition, existing model compression methods do not fully consider the constraints and physical characteristics of onboard hardware resources, making it difficult to achieve a balance between performance and efficiency under limited resources. Finally, the verification system is incomplete, mostly remaining at the functional verification level, lacking a systematic verification process covering performance, reliability, and space environment adaptability, making it difficult to meet the high reliability requirements of aerospace engineering.

[0004] 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

[0005] The purpose of this invention is to provide a full-link hardware deployment and verification method for satellite on-orbit fault diagnosis model, so as to solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for the full-link hardware deployment and verification of a satellite on-orbit fault diagnosis model, comprising the following steps: Historical telemetry data from in-orbit satellites are collected and preprocessed to obtain a standardized dataset. At the same time, a fault knowledge base is constructed and data augmentation is performed to assess data quality. A diagnostic model containing data-driven channels, physical information channels, and knowledge-enhancing channels is trained using a standardized dataset and a fault knowledge base, and a full-precision floating-point model is output under the condition of meeting the preset performance indicators. The full-precision floating-point model is distilled and pruned, and differentiated quantization is performed based on the precision sensitivity of different layers to obtain a quantized model; Based on the quantization model, complete the hardware implementation and introduce fault tolerance processing, perform consistency evaluation of inference results, and obtain deployment results; Based on the deployment results, the functions, performance, reliability, and environment were verified, and a comprehensive evaluation was conducted to obtain the verification results. The validated hardware is run in orbit, performance is evaluated and updates are triggered based on the in-orbit data, and the model training, compression and hardware implementation processes are re-executed based on the updated data.

[0007] Preferably, the preprocessing procedure includes: detecting outliers within a sliding window using the 3-sigma criterion and replacing them with linear interpolation; completing data segments with consecutive missing values ​​not exceeding a preset length using cubic spline interpolation; and uniformly resampling parameters at different sampling rates to the reference frequency. And adopting a robust normalization method Normalize the data.

[0008] Preferably, the training strategy adopts a four-stage step-by-step approach: first, train the physical information channel to minimize the physical equation residuals; then, train the knowledge enhancement channel to learn knowledge graph embeddings; then, train the data-driven channel to optimize temporal feature extraction; and finally, jointly fine-tune all channels and the fusion module.

[0009] Preferably, the quantization adopts a physically consistent perception-based differentiated hybrid precision strategy, wherein the learnable physical parameter layer and physical equation residual calculation layer in the physical information channel adopt INT16 quantization, and the convolutional layer and fully connected layer in the data-driven channel and knowledge-enhancing channel adopt INT8 quantization. Furthermore, the quantization process adopts a combination of post-training quantization initialization and quantization-aware training fine-tuning.

[0010] Preferably, the graded radiation hardening includes three levels: the first level implements periodic readback loading scrubbing for the FPGA configuration memory; the second level implements triple modular redundancy (TMR) hardening for the control state machine and output decision logic; and the third level implements checksum-based algorithm-level fault tolerance (ABFT) in matrix multiplication operations.

[0011] Preferably, the passing standard for the four-level progressive verification is: functional verification requirement hardware-floating-point consistency rate. The performance verification requires an average inference latency of no less than 90% and a power consumption of no more than 5W; the reliability verification requires continuous operation for no less than 72 hours without any abnormalities; and the environmental verification requires passing thermal vacuum, vibration, and electromagnetic compatibility tests.

[0012] Preferably, the feedback loop includes: a data feedback path for supplementing the training dataset with on-orbit telemetry data; a knowledge feedback path for recording new fault modes into the fault knowledge base; a model iteration path for re-executing the model training, compression, and hardware deployment process to generate a new model when the triggering conditions are met; and an on-orbit update path for uploading the new bitstream to the satellite via the software upload channel to complete the inference engine replacement.

[0013] Preferably, a model freshness index is also defined. When this indicator falls below a preset threshold, a preventative model update is performed, and this indicator, together with the on-orbit diagnostic consistency rate, constitutes the basis for model update decisions.

[0014] Preferably, the processing flow of the automatic compilation toolchain includes: parsing the ONNX model file to extract the computation graph structure; converting it into an intermediate representation and performing operator fusion and constant folding optimization; performing hardware mapping and scheduling according to FPGA resource constraints; generating synthesizable hardware description language code; and calling the FPGA vendor's toolchain to complete synthesis, placement and routing to generate a bitstream file.

[0015] Preferably, a multi-task joint optimization approach is adopted during model training. The model is trained collaboratively through three tasks: health status assessment, degradation trend prediction, and fault diagnosis. A weighted loss function is used to balance and optimize each task, thereby improving the model's comprehensive diagnostic performance and generalization ability under multiple operating conditions.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention proposes an end-to-end, end-to-end methodology covering six stages: data preparation, model training, lightweight compression, hardware deployment, ground validation, and on-orbit validation. By standardizing the input-output relationships and data formats for each stage, a standardized process is established throughout the entire model lifecycle. This method effectively solves the problems of fragmented processes and inconsistent interfaces in existing technologies, enabling a closed-loop linkage between data flow, model flow, and validation process. This significantly improves engineering implementation efficiency and system reproducibility, providing a scalable engineering implementation path for satellite on-orbit intelligent fault diagnosis systems.

[0017] This invention proposes a physically consistent, differentiated hybrid precision quantization strategy. Based on the sensitivity of different structures in the model to numerical precision, different bit widths are used for quantization of the physically constrained layer and the data-driven layer. While ensuring the computational accuracy and stability of the physical model, the strategy fully leverages the storage and computational efficiency advantages of low-bit-width quantization to maximize model compression efficiency. This method overcomes the limitations of traditional unified quantization strategies in complex hybrid models, improving the feasibility of model deployment and performance on resource-constrained hardware platforms.

[0018] This invention constructs a four-level progressive ground verification system comprising functional verification, performance verification, reliability verification, and environmental verification, and introduces a comprehensive scoring mechanism to uniformly and quantitatively evaluate the verification results. This system not only enables systematic evaluation of hardware inference systems from multiple dimensions, but also ensures that the verification results are measurable, comparable, and traceable, effectively addressing the problem of incomplete verification processes in existing technologies, thereby significantly improving the engineering reliability and quality assurance level after the model is deployed in hardware.

[0019] This invention establishes a closed-loop feedback mechanism from on-orbit operational data to model updates. Through four paths—data feedback, knowledge feedback, model iteration, and on-orbit updates—it achieves continuous optimization and dynamic evolution of the model. Simultaneously, a model freshness index is introduced as an auxiliary criterion, which, together with the on-orbit diagnostic consistency rate, constitutes the basis for update decisions, enabling the model to adjust and upgrade in a timely manner according to changes in the on-orbit environment. This mechanism effectively overcomes the problem of difficulty in updating traditional models after deployment, giving the spaceborne intelligent diagnostic system long-term adaptive and continuous evolution capabilities. Attached Figure Description

[0020] 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 invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0021] Figure 1This is a flowchart illustrating the end-to-end hardware deployment and verification method for a satellite on-orbit fault diagnosis model according to the present invention. Detailed Implementation

[0022] 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.

[0023] This invention provides, for example Figure 1 The following are the specific steps of the end-to-end hardware deployment and verification method for a satellite on-orbit fault diagnosis model: In satellite on-orbit fault diagnosis applications, the acquisition and standardized processing of multi-source telemetry data constitute the foundation of the entire intelligent diagnostic system. To ensure the sufficiency and representativeness of the data, it is necessary to continuously collect historical telemetry data from multiple on-orbit satellites over a long period, typically covering several years of operational cycles, to ensure that it includes normal operating conditions as well as various fault conditions. The collected data should cover the satellite's key subsystems, including the attitude control subsystem, power supply subsystem, thermal control subsystem, and payload subsystem, thereby forming a multi-dimensional, multi-source, and strongly coupled time-series data set.

[0024] Because telemetry parameters from different subsystems vary significantly in sampling frequency, dimensional range, and noise characteristics, raw data often contains outliers, missing values, and time asynchrony issues. Therefore, a unified data preprocessing pipeline needs to be constructed. Firstly, for outlier handling, a 3-sigma criterion based on statistical distribution is adopted. Local statistical features are calculated within a sliding window, and anomaly detection is achieved by identifying data points that deviate from the mean by more than three standard deviations. This method effectively removes outliers caused by sudden sensor interference, communication errors, or equipment jitter. After anomalies are detected, they are replaced by linear interpolation using adjacent normal data points to ensure data continuity.

[0025] Regarding missing value handling, different strategies are adopted based on the consecutive length of the missing data. For short, consecutively missing data segments, cubic spline interpolation is used for smooth completion to restore the data's trend. For longer, consecutively missing data segments, they are directly marked as unusable and removed to avoid introducing excessive errors. This strategy balances data integrity and data reliability.

[0026] During time alignment, telemetry parameters with different sampling frequencies are mapped to a unified time reference. High-frequency data is downsampled and mean-filtered to reduce noise, while low-frequency data is expanded using a zero-order hold method, thereby constructing a multivariate time-series data matrix with a unified time axis.

[0027] In the data normalization stage, to enhance robustness to outliers, a quantile-based normalization method is adopted, with the following formula: in: Represents the original data value; This represents the median; Indicates the 25th percentile; This represents the 75th percentile; the formula uses the interquartile range to scale the data, making the data distribution more stable. Compared to the traditional mean-standard deviation normalization method, this method is insensitive to extreme values ​​and can effectively suppress the impact of outliers on the overall distribution, making it particularly suitable for long-tailed distributions and abrupt changes in satellite telemetry data.

[0028] In terms of constructing the fault knowledge base, a structured set of failure modes is established by integrating data from multiple sources, including FMECA analysis reports, on-orbit fault records, and ground test data. Each failure mode includes not only basic attribute information but also associated telemetry parameters and fault propagation paths, providing knowledge constraints to support subsequent models.

[0029] To address the scarcity of real-world fault samples, a data augmentation strategy is introduced. The augmentation process employs a two-path approach: firstly, simulation augmentation is performed through physical modeling, using system dynamics models to simulate fault conditions of varying degrees; secondly, generative adversarial networks (GANs) learn the data distribution and generate synthetic samples. In evaluating the quality of the generated samples, a maximum mean difference index is introduced, expressed as follows: in: Represents a real sample; Indicates the generation of samples; and These represent the number of real samples and the number of generated samples, respectively. This represents the feature mapping function.

[0030] This formula measures the difference between the distribution of the real samples and the distribution of the generated samples. A small MMD value indicates that the generated samples are statistically close to the real samples; a large value indicates that the generated samples deviate from the real distribution. In practical applications, a threshold is set to filter the generated data, retaining only samples that meet the distribution consistency requirements, thereby ensuring the effectiveness of data augmentation.

[0031] In the dataset production stage, all preprocessed data is uniformly stored as a standard format dataset and combined with a fault knowledge base to form a complete data foundation. To quantify data quality, a comprehensive evaluation index is introduced, the expression of which is as follows: in: Indicates the missing rate. Indicates the anomaly rate. This indicates parameter coverage.

[0032] This formula comprehensively evaluates data quality from three dimensions: First, the completeness dimension, through... The first dimension is to measure data missing information; the second is the reliability dimension, through... The third dimension is coverage, measured by the proportion of outlier data. Measure the extent to which data covers the system state.

[0033] Multiplying these three factors together forms an overall quality metric, which can avoid the bias caused by a single metric. For example, even if there are few missing data points, if the proportion of outliers is high or the parameter coverage is insufficient, the final quality metric will still be low, indicating that the data needs further optimization.

[0034] Through the aforementioned multi-source data collection and standardization process, a high-quality, structured, and quantifiable data system was constructed. The normalization formula ensures stable data distribution, while the MMD metric enhances the consistency of data distribution. The indicators enable closed-loop control of overall data quality. This process provides a reliable data foundation for subsequent model training and on-orbit deployment, effectively improving the accuracy, stability, and engineering feasibility of the fault diagnosis model.

[0035] In satellite on-orbit fault diagnosis missions, the model training phase directly determines the accuracy and reliability of subsequent deployments. Therefore, it is necessary to build a high-performance computing environment and design a model structure suitable for complex multi-source time-series data. The training process typically relies on a ground-based high-performance computing platform, using multi-GPU parallel computing to improve training efficiency. The computing resource configuration must meet the needs of large-scale time-series data processing, and the GPU memory capacity should support batch training of complex models. Simultaneously, mixed-precision training techniques (combining FP16 and FP32) are used to reduce memory usage and improve computational efficiency. At the software level, mainstream deep learning frameworks are adopted to ensure model building flexibility and subsequent deployment compatibility, and to support model export to a universal intermediate representation format, thereby providing a standardized interface for hardware implementation.

[0036] In terms of model structure design, a three-channel fusion architecture is adopted to fully utilize data-driven information, physical prior information, and knowledge reasoning capabilities. The data-driven channel is primarily used to extract temporal features from telemetry data, capturing dynamic changes at different time scales through a multi-scale temporal convolutional network. Simultaneously, a self-attention mechanism enhances the ability to focus on key temporal segments, thereby improving the ability to identify complex patterns. The physical information channel discretizes the system's physical equations and embeds learnable parameters, enabling the model to not only rely on data during training but also follow the system's physical laws, thus improving the model's generalization ability under unknown conditions. The knowledge enhancement channel introduces a knowledge graph, encoding existing fault modes and their relationships into a graph structure, and utilizes a graph attention network for reasoning, thereby enhancing the model's understanding of complex fault propagation paths.

[0037] The outputs of the three information processing channels are integrated through a fusion mechanism. Adaptive weight allocation is introduced during the fusion process, enabling the model to dynamically adjust the contribution ratio of each channel based on different input data. This structure achieves synergistic optimization between data-driven approaches, physical constraints, and knowledge reasoning, thereby improving overall diagnostic performance.

[0038] In terms of training objective design, the model performs multi-task learning simultaneously to achieve a more comprehensive state awareness capability. The training objectives include three aspects: health status assessment, degradation trend prediction, and fault diagnosis and identification. To coordinate the optimization directions among multiple tasks, a unified total loss function is introduced, the expression of which is as follows: in: It is the health status classification loss, used to assess the accuracy of system operating status classification; It is the degradation trend prediction loss, used to measure the error in predicting future states; It is fault diagnosis loss, used to assess the ability to identify fault types; It is the physical constraint loss, used to constrain the model output to conform to physical laws; It is knowledge reasoning loss, used to enhance the reasoning ability of knowledge graphs; Regularization terms are used to prevent the model from overfitting. This represents the weighting coefficient of each loss term.

[0039] This formula achieves collaborative optimization across different tasks by weighting and combining multiple loss terms. During actual training, the weight coefficients can be automatically adjusted using an uncertainty-based weighting strategy, enabling the model to achieve dynamic equilibrium across different tasks and preventing any single task from dominating the training process.

[0040] In terms of training strategy, a phased, progressive training approach is adopted to reduce the training difficulty of complex models and improve convergence stability. In the initial stage, the physical information channel is optimized first, allowing the model to learn the physical laws of the system and establish basic constraints. Then, the knowledge enhancement channel is trained, enabling the model to possess basic fault reasoning capabilities. Next, the data-driven channel is trained to fully explore the statistical features in the data. Finally, the joint training stage is entered, where all parameters are optimized as a whole to bring the model to its optimal state. This "constraint-first, learning-later" training path effectively avoids the model getting trapped in local optima and improves the overall convergence quality.

[0041] In terms of algorithm optimization, an adaptive optimization method is used to update the model parameters, and a learning rate scheduling strategy is combined to gradually reduce the learning rate, thereby achieving more refined parameter adjustments in the later stages of training. The initial learning rate is set within a reasonable range to balance convergence speed and stability, while a periodic change strategy is used to avoid oscillations during training.

[0042] During the model validation phase, a comprehensive evaluation of model performance is required using independent test data. Evaluation metrics include key indicators such as fault diagnosis accuracy, ability to identify unknown operating conditions, and false positive and false negative rates. Through multi-dimensional metric constraints, the model is ensured not only to perform well on known data but also to possess strong generalization ability. The testing process must cover a variety of typical operating conditions, including normal operation, known fault modes, and abnormal operating conditions, to comprehensively evaluate model performance.

[0043] Once the model meets the preset performance metrics, the trained model is exported as a standardized full-precision floating-point model. This model includes complete network structure definitions, weight parameters, and input / output format information, providing a foundation for subsequent model compression and hardware deployment. The exported format must be cross-platform compatible to support conversion and deployment across different hardware platforms.

[0044] Through the above training process, an intelligent diagnostic model combining multi-source data-driven approaches, physical constraints, and knowledge reasoning can be constructed. The overall loss function achieves unified optimization across multiple tasks, a phased training strategy improves model convergence stability, and a high-performance computing platform ensures training efficiency. This results in a diagnostic model with high accuracy, robustness, and good generalization ability, laying a solid foundation for subsequent lightweight compression and on-orbit deployment.

[0045] In satellite-based intelligent diagnostic missions, lightweight model compression and quantization are crucial steps in transforming algorithms for deployment on spaceborne hardware. Due to limited spaceborne computing resources, especially on radiation-hardened FPGAs or embedded platforms, storage capacity, computing power, and power consumption are all strictly constrained. Therefore, it is necessary to significantly reduce the model size and computational complexity while maintaining model accuracy as much as possible.

[0046] In the initial stage of model compression, a knowledge distillation mechanism is introduced to transfer knowledge from the teacher model to the student model. The teacher model is a high-performance model trained to full precision, possessing strong expressive power, while the student model is a lightweight model after structural compression. The student model retains the multi-channel fusion framework in its structure, but reduces its network depth and width, thereby significantly reducing the parameter scale.

[0047] The distillation process introduces soft-label learning, enabling the student model to learn not only the true label information but also the probability distribution information output by the teacher model. The distillation loss function is defined as follows: in, It is the original task loss function, used to measure the difference between the model's prediction and the true label; It is the output logits of the teacher model; The student model outputs Iogits; It is a temperature parameter used to smooth the probability distribution; It is the distillation weighting coefficient; It is the Kullback-Leibler divergence, used to measure the difference between two probability distributions.

[0048] The core idea of ​​this formula is: by adjusting the temperature parameter... This smooths the originally sharp probability distribution, enabling the student model to learn the relative relationships between categories, rather than just the final classification result. (Coefficients) By controlling the weight ratio between soft and hard tags, knowledge transfer can be achieved while maintaining task performance.

[0049] After completing the distillation training, a structured pruning method was introduced to further reduce the model size. During pruning, convolutional channels were used as the basic unit, and the importance of each channel was calculated for selection. The importance index was represented by the L1 norm, and its calculation form was as follows: in: Indicates the first The weight vector of each convolutional channel.

[0050] The L1 norm reflects the contribution of a channel to the output features. Channels with smaller L1 norms generally have less impact on the overall model performance and can therefore be removed first. In each round of pruning, the least important channels are removed according to a preset ratio, and model performance is restored through subsequent fine-tuning training. This process is executed iteratively, gradually compressing the model size until the target number of parameters or the accuracy drop exceeds a preset threshold.

[0051] During the quantization phase, a differentiated mixed-precision quantization strategy is adopted to address the varying sensitivity of different model structures to numerical precision. Physically constrained layers require higher precision and therefore employ a higher bit width representation; data-driven parts, on the other hand, can use a lower bit width to improve computational efficiency. The quantization process is optimized by combining post-training quantization with quantization-aware training.

[0052] Post-training quantization first discretizes the model weights, while quantization-aware training simulates quantization errors during training, allowing the model to gradually adapt to a low-precision computing environment. During backpropagation, since the quantization operation is non-differentiable, a pass-through estimator is introduced to approximate the gradient, thus ensuring the continuity of the training process.

[0053] During the compression effect evaluation, the compression process is constrained by comparing the performance differences between the quantized model and the original model. Evaluation metrics include changes in diagnostic accuracy, changes in physical constraint residuals, and changes in prediction error. When the performance loss caused by compression exceeds the set range, the quantization strategy needs to be readjusted or the number of training rounds increased.

[0054] In the final compression output stage, a unified evaluation index is established for the quantization model to measure the compression effect, and its definition is as follows: in: This is the storage size of the full-precision model; This metric quantifies the storage size of the model and reflects the compression ratio. A higher value indicates a more significant compression effect. For example, a compression ratio greater than 4 means the model size has been reduced to less than a quarter of its original size. In practical applications, this metric needs to be balanced with model accuracy to avoid over-compression that could lead to a severe performance degradation.

[0055] In summary, this compression and quantization process achieves a transformation of the model from a high-precision, complex structure to a low-resource, high-efficiency structure through a three-stage collaborative optimization of "knowledge distillation + structured pruning + hybrid precision quantization". The distillation process enables knowledge transfer, the pruning process reduces redundant structures, and the quantization process reduces computational precision requirements. The combination of these three processes not only significantly reduces the model size but also maintains stable diagnostic performance as much as possible.

[0056] The above method introduces a hierarchical optimization strategy to make the model compression process controllable and adjustable, thereby meeting the stringent requirements of onboard hardware deployment for resource utilization, power consumption control and operational stability, and providing a reliable foundation for subsequent hardware implementation and on-orbit operation.

[0057] In the engineering implementation of spaceborne intelligent diagnostic applications, FPGA hardware deployment is a core link connecting the algorithm model with the actual space operating environment. Due to the strict constraints of the spaceborne environment on power consumption, real-time performance, and radiation resistance, it is necessary to build an automated conversion mechanism from model to hardware, combined with dedicated hardware architecture design, to achieve efficient and reliable inference execution capabilities.

[0058] In the initial stage of model deployment, an automated compilation process is established from the quantized model to the FPGA bitstream. This process takes a standardized model format as input and converts various operators, weight parameters, and quantization configurations in the model into a hardware-compatible representation by parsing the computation graph structure. First, the model file is parsed to extract the computation graph topology and operator information for each layer, providing a foundation for subsequent processing. Then, an intermediate representation is introduced to optimize the computation graph structure, including operator fusion, constant folding, and redundant node elimination. By merging operations such as convolution, normalization, and activation functions into a unified computational unit, the number of memory accesses is reduced, and execution efficiency is improved.

[0059] During the hardware mapping phase, computational tasks are scheduled and allocated based on the resource constraints of the target FPGA. Key resources include the number of lookup tables, the number of multiply-accumulate units, on-chip memory capacity, and external memory bandwidth. By rationally designing parallelism and pipeline depth, the computation process achieves optimal throughput performance under limited resource conditions. Data reuse strategies are particularly crucial in this process; by locally caching weights and feature maps, the frequency of external memory access can be significantly reduced, thereby reducing latency and power consumption.

[0060] In the hardware implementation phase, hardware description language code is automatically generated to achieve the collaborative design of computing units, storage units, and control logic. Convolution operations are implemented using a fixed-point multiply-accumulate array, with computing units of different precisions supporting both low-bit-width and high-bit-width operations to meet the computational needs of different types of network layers. Activation functions are implemented using lookup tables, discretizing nonlinear functions into piecewise linear approximations, thereby reducing hardware computational complexity. Attention mechanism computation is performed using fixed-point operations, achieving dynamic weighting of key features.

[0061] The inference engine construction process requires the collaborative design of multiple functional units. The data preprocessing unit is responsible for performing normalization calculations on the input telemetry data to meet the model's input requirements; the weight management unit is responsible for scheduling data between on-chip and off-chip storage; the convolutional computation array, as the core computational unit, improves inference speed through parallel computing; the activation and attention unit provides nonlinear transformation capabilities; and the output processing unit converts the inference results into readable diagnostic information. These functional units are connected through a pipeline structure to form a complete inference execution path.

[0062] In terms of space environment adaptability design, a hierarchical fault-tolerance mechanism is introduced to improve system reliability. First, a periodic readback mechanism is used to verify and repair the configuration storage to prevent long-term configuration errors caused by radiation. Second, redundant design is adopted for critical control logic, using multiple copies of the logic results for voting to reduce the risk of single points of failure. Furthermore, a verification-based fault-tolerance method is introduced at the computational level to verify the consistency of computation results, promptly detect and locate errors, thereby improving computational reliability.

[0063] In terms of interface integration, compatibility with the satellite data bus needs to be achieved. Multiple communication interfaces are supported to enable telemetry data reception and diagnostic result output. Low-speed interfaces are suitable for routine telemetry data exchange, high-speed interfaces are suitable for large data volume transmission scenarios, and discrete interfaces are used for the input and output of critical status signals. All interfaces must have electrical isolation and electromagnetic interference suppression capabilities to ensure stable operation in complex space electromagnetic environments.

[0064] During the hardware output phase, a consistency evaluation of the deployment results is required to verify the degree of matching between the hardware inference results and the original model output. The evaluation metrics are defined as follows: in, It is a hardware consistency metric used to measure the degree of consistency between FPGA hardware inference results and floating-point model inference results. It is the number of samples in which the hardware inference result is consistent with the floating-point model inference result; It represents the total number of samples participating in the test.

[0065] This metric measures the degree to which the hardware implementation preserves the behavior of the original model. A value close to 1 indicates a high degree of consistency between the hardware inference results and the software model; a lower value indicates that quantization errors or hardware implementation deviations have a significant impact on the results. In the consistency determination process, it is necessary not only to ensure the consistency of classification results but also to constrain the differences between probability distributions to ensure statistical consistency of the output results.

[0066] In summary, this hardware deployment process achieves efficient transformation from algorithm models to spaceborne hardware systems through automated compilation, structural optimization, resource scheduling, and fault-tolerant design. Automated processes improve development efficiency, fixed-point computing and parallel architecture reduce resource consumption, multi-level fault-tolerant mechanisms enhance system reliability, and consistency assessment methods ensure deployment accuracy, thus forming a highly reliable intelligent diagnostic execution platform suitable for the space environment.

[0067] After the intelligent diagnostic model completes hardware deployment, it needs to undergo a rigorous ground verification process to systematically evaluate its functional correctness, performance indicators, operational reliability, and environmental adaptability. This verification phase adopts a multi-level progressive structure, dividing the verification process into four levels: functional verification, performance verification, reliability verification, and environmental verification. By proceeding step by step, this approach ensures that the system possesses sufficient engineering reliability before entering on-orbit operation.

[0068] At the functional verification level, the primary focus is on the correctness and consistency of the inference results. The verification process employs two methods: historical data replay and fault injection. Historical data replay involves inputting preprocessed test data into the inference engine in chronological order and recording each inference output. This output is then compared line by line with the original floating-point model output to verify the hardware implementation's ability to reproduce the behavior of the original model. Fault injection, based on an established set of fault modes, artificially introduces feature perturbations into normal telemetry data to simulate various typical fault scenarios. The system's ability to identify fault modes is evaluated by verifying whether the inference output can correctly identify the corresponding fault category. The core evaluation metric for functional verification is the consistency ratio between the hardware and the floating-point model. This metric must meet a preset threshold to ensure the logical correctness of the hardware implementation.

[0069] At the performance verification level, the focus is on evaluating the inference system's operational capabilities under real-time and resource constraints. First, end-to-end latency for a single inference attempt is measured using a large-scale test sample, and the average and high-quantile latency are statistically analyzed to assess the system's response stability under different load conditions. Second, the system's throughput is evaluated by statistically analyzing the number of inference requests that can be processed per unit time. Regarding power consumption, the average and peak power consumption under full load are measured to ensure that it meets the stringent energy consumption limits of the onboard platform. Furthermore, hardware resource utilization is statistically analyzed, including the proportion of logic, computing, and storage resources used, to ensure that resource usage is within a reasonable range and that redundancy is maintained. Through these multi-dimensional indicators, the engineering performance of the inference system can be comprehensively evaluated.

[0070] At the reliability verification level, it is necessary to verify the system's stability under long-term operation and abnormal conditions. Continuous operation testing involves running the inference engine under maximum load conditions for an extended period, continuously recording output results and system status to detect any operational anomalies, performance degradation, or system crashes. This process verifies the system's stability and fatigue resistance during long-term operation. In radiation simulation testing, fault injection tools are used to randomly introduce bit flips into hardware storage and registers to simulate single-event effects in a space radiation environment. By verifying whether the fault tolerance mechanism can detect and correct errors in a timely manner, the system's reliability performance under extreme environments is evaluated.

[0071] At the environmental verification level, the hardware system needs to be comprehensively evaluated by simulating the space environment. Thermal vacuum testing verifies the system's operational stability under extreme temperature and low pressure conditions, detecting performance fluctuations during temperature changes through multiple temperature cycles. Vibration testing simulates the mechanical vibration environment generated during launch and orbital operation, verifying structural stability and functional reliability through sinusoidal frequency sweep and random vibration tests. Electromagnetic compatibility testing evaluates the system's immunity to interference in complex electromagnetic environments, ensuring stable operation even in the presence of electromagnetic noise.

[0072] During the comprehensive evaluation phase, a unified ground-based validation scoring index is introduced to quantitatively integrate the results of multi-level validation. Its expression is as follows: in, It is a comprehensive score for multi-level ground validation, used to quantify the overall performance throughout the entire ground validation phase. It is the functional verification score; It is the performance verification score; It is the reliability verification score; It is the environmental verification score; These are the weight coefficients of each validation item. This formula integrates the results from different validation dimensions using a weighted method to form a unified evaluation index. The weight coefficients satisfy a normalization constraint: By allocating weights appropriately, the importance of different verification dimensions can be highlighted according to actual application needs. For example, in aerospace applications, reliability and environmental adaptability typically have higher weights. The comprehensive score not only reflects individual performance but also embodies the overall system capability. When the comprehensive score reaches a preset threshold, it indicates that the system is ready to proceed to the next stage; at the same time, each individual score is required to be no lower than the minimum standard to avoid any weakness in a key capability.

[0073] Through the aforementioned multi-level verification process, a comprehensive assessment from functional correctness to environmental adaptability can be achieved. Functional verification ensures the correctness of the reasoning logic, performance verification guarantees real-time processing capability, reliability verification improves long-term operational stability, environmental verification ensures the system adapts to complex space conditions, and the comprehensive scoring mechanism achieves unified quantification of multi-dimensional evaluation results. This verification system not only enhances the reliability of the system engineering but also provides sufficient risk control basis for subsequent on-orbit operation.

[0074] After the intelligent diagnostic system is deployed to a low-Earth orbit satellite, it needs to be fully validated in the on-orbit operating environment, and a continuous feedback mechanism needs to be established to ensure the long-term effectiveness of the model. The validation and feedback process mainly includes hardware implementation, monitoring of validation content, on-orbit data inversion and iterative updates of the model, and evaluation of model freshness.

[0075] First, during the hardware integration phase, ground-verified onboard inference computing units are integrated into the satellite platform or auxiliary payloads to ensure they can receive real-time telemetry data and output diagnostic results in orbit. Satellite selection must consider orbital inclination coverage of high-radiation regions, such as the South Atlantic Anomaly, while ensuring a design life of at least two years and possessing software injection capabilities to support subsequent model updates and iterative optimizations. During deployment, it is essential to ensure system interface compatibility, power consumption meeting satellite platform constraints, and maintaining redundant paths to handle unforeseen anomalies.

[0076] During the on-orbit verification cycle, the verification duration is typically no less than [number missing]. Months, of which The value range is 6 to 12 months. The verification covers the following core aspects: Diagnostic accuracy verification: The on-orbit inference output is transmitted back to the ground via a communication link and compared with the results of the full-precision floating-point analysis model to calculate the on-orbit consistency rate. It is used to evaluate the consistency between on-orbit inference and ground models.

[0077] Single Event Flip (SEU) Recording: Records the occurrence time, flip location (including configuration memory, BRAM, or register), detection method (Scrubbing, TMR, ABFT), and correction result (success or failure) for each SEU event. Counts the number of SEU events. This is used to evaluate the effectiveness of the fault tolerance mechanism.

[0078] Power consumption and latency monitoring: Monitor inference latency and power consumption curves, assess performance degradation trends caused by radiation or hardware aging, and ensure that the system meets real-time and energy consumption constraints.

[0079] Novel Failure Mode Discovery: When the confidence level of the inference output is below a threshold ( When the value is between 0.6 and 0.8, it is marked as falling under a new type of failure mode, providing data support for subsequent model updates.

[0080] In the on-orbit feedback phase, a data-driven model update closed loop is established, including four feedback paths: Data feedback: Telemetry data collected in orbit, especially data containing new types of faults or extreme operating conditions, will be transmitted back to the ground training dataset. This is used to enhance the coverage of the training set.

[0081] Knowledge Feedback: Update the fault knowledge base based on new fault modes identified manually or through algorithms. It also expands the recognition path to ensure the knowledge base is continuously updated.

[0082] Model iteration: When on-orbit probability Falling below the threshold (85% to 90%) or the cumulative number of newly discovered failure modes exceeds Types 3 to 5 trigger model iteration, re-execute the model lightweight compression and FPGA generation process.

[0083] In-orbit updates: New features are uploaded to the satellite via satellite software injection, completing the model replacement of the inference unit and ensuring the continuity of in-orbit operation and the effectiveness of the model.

[0084] To quantify model freshness, a model freshness index is introduced. This is used to measure the effectiveness of the quantification model, and its calculation formula is as follows: in: It is the current time; It is the time of the model's most recent update; This is the model lifecycle (usually 12 to 24 months).

[0085] This metric reflects the time decay of the model after quantization deployment. When Less than the threshold Even when diagnostic performance has not significantly degraded (to a value of 0.3 to 0.5), preventative model updates are recommended to maintain diagnostic consistency. It is at a high level. This indicator can be used to dynamically determine whether the model needs to be iterated and updated, thus enabling continuous monitoring of model freshness.

[0086] Through the aforementioned on-orbit verification and feedback iteration mechanism, a closed-loop management system is achieved, encompassing ground verification, on-orbit operation, data feedback, and model updates. Each step in this closed loop is rigorously quantified, using consistency rates, event logs, the number of new failure modes, and model freshness metrics to ensure the diagnostic system's long-term reliable operation in complex space environments. This entire system not only improves the model's on-orbit accuracy and availability but also provides a scalable engineering solution for large-scale deployments across multiple constellations.

[0087] To verify the engineering feasibility and on-orbit application effectiveness of the technical solution, a practical application scenario based on a low-Earth orbit communication satellite constellation was constructed. This constellation comprises 30 satellites in orbit, each with a mass of approximately 250 kg, a design life of 5 years, an orbital altitude of 550 km, and an orbital inclination of 53°. The objective is to deploy onboard intelligent fault diagnosis capabilities covering the power supply subsystem and attitude control subsystem within this constellation. In terms of data construction, 2.5 years of historical telemetry data were continuously collected from the 30 satellites, with 63 telemetry parameters collected from each satellite: 28 from the power supply subsystem, 25 from the attitude control subsystem, and 10 from the thermal control subsystem. A unified sampling frequency was set as follows: The original dataset contains approximately 1.89 billion data points. Anomaly detection is performed using the 3-sigma method within a window. Under these conditions, the outlier removal rate is 0.8%; cubic spline interpolation is used to handle missing data, with a missing rate of 1.2%, and the interpolation window length is... After time alignment and normalization, a standardized dataset is formed. The data is stored in HDF5 format and is approximately 47GB in size. Data quality is evaluated using the following metrics: in: , representing the missing rate; , representing the anomaly rate; This indicates that the parameter coverage is calculated as follows: The result is greater than 0.90, indicating that the data quality meets the training requirements.

[0088] During the model training phase, four NVIDIA A100 GPUs were used for parallel training to build a three-channel fusion model with a total of approximately 2 million parameters. The phased training strategy is as follows: Phase 1: 50 epochs, physical channel residuals decreased from 0.82 to 0.06.

[0089] Phase 2: 30 epochs, knowledge channel MRR reaches 0.78.

[0090] Phase 3: 100 epochs, data channel training.

[0091] Phase 4: 200 epochs, with joint fine-tuning and a total training time of approximately 48 GPU hours.

[0092] The test set evaluation results are as follows: Fault diagnosis accuracy: Unknown operating condition identification capability: False alarm rate: ; False alarm rate: .

[0093] Export the model as a full-precision model The file is 7.6MB in size.

[0094] In the model compression stage, a student model with 400,000 parameters is constructed, and lightweighting is achieved through distillation and pruning. The distillation loss function is as follows: in: This is the output of the teacher model; This is the output of the student model; It is a temperature parameter; It is the distillation weight.

[0095] After pruning, the number of parameters decreased from 400,000 to 180,000, while maintaining 95.5% accuracy. The quantized model size is 1.8MB. The compression ratio is calculated as follows: This indicates that the model has been compressed to approximately 1 / 4 of its original volume.

[0096] During the hardware deployment phase, an FPGA platform (ZU9EG) was selected, with the following resource utilization rates: LUT: 67%; DSP: 58%; BRAM: 73%. Inference consistency metrics are as follows: The inference performance is as follows: Average latency: -99; Quantile delay: ; Power consumption: .

[0097] During the ground verification phase, the overall score is as follows: in: ; ; ; .

[0098] Weight: Calculation results: Meets the passing criteria.

[0099] During the on-orbit operation phase, the verification period is 6 months, and the on-orbit consistency rate is: Month 1: 95.2%; Second month: 94.8%; Month 3: 94.1%.

[0100] SEU events were recorded 3 times, and all were successfully recovered. Inference performance: ; During model updates, a freshness metric is introduced: in: Monthly calculation: This value is much higher than the threshold of 0.4, indicating that the model is still within its validity period.

[0101] In its fourth month of operation, a new failure mode was detected, triggering the model update mechanism. By supplementing the model with 120 real-world data points and 500 simulation data points, the model was retrained and an on-orbit update was completed. After the update, the on-orbit consistency rate recovered to 95.8%.

[0102] This embodiment demonstrates that the complete process implements a closed-loop control mechanism from data construction, model training, compressed deployment to on-orbit operation and continuous updates. All key indicators meet engineering application requirements and can operate stably for extended periods in complex space environments, while also possessing continuous evolution capabilities.

[0103] This invention proposes an end-to-end, end-to-end methodology covering six stages: data preparation, model training, lightweight compression, hardware deployment, ground validation, and on-orbit validation. By standardizing the input-output relationships and data formats for each stage, a standardized process is established throughout the entire model lifecycle. This method effectively solves the problems of fragmented processes and inconsistent interfaces in existing technologies, enabling a closed-loop linkage between data flow, model flow, and validation process. This significantly improves engineering implementation efficiency and system reproducibility, providing a scalable engineering implementation path for satellite on-orbit intelligent fault diagnosis systems.

[0104] This invention proposes a physically consistent, differentiated hybrid precision quantization strategy. Based on the sensitivity of different structures in the model to numerical precision, different bit widths are used for quantization of the physically constrained layer and the data-driven layer. While ensuring the computational accuracy and stability of the physical model, the strategy fully leverages the storage and computational efficiency advantages of low-bit-width quantization to maximize model compression efficiency. This method overcomes the limitations of traditional unified quantization strategies in complex hybrid models, improving the feasibility of model deployment and performance on resource-constrained hardware platforms.

[0105] This invention constructs a four-level progressive ground verification system comprising functional verification, performance verification, reliability verification, and environmental verification, and introduces a comprehensive scoring mechanism to uniformly and quantitatively evaluate the verification results. This system not only enables systematic evaluation of hardware inference systems from multiple dimensions, but also ensures that the verification results are measurable, comparable, and traceable, effectively addressing the problem of incomplete verification processes in existing technologies, thereby significantly improving the engineering reliability and quality assurance level after the model is deployed in hardware.

[0106] This invention establishes a closed-loop feedback mechanism from on-orbit operational data to model updates. Through four paths—data feedback, knowledge feedback, model iteration, and on-orbit updates—it achieves continuous optimization and dynamic evolution of the model. Simultaneously, a model freshness index is introduced as an auxiliary criterion, which, together with the on-orbit diagnostic consistency rate, constitutes the basis for update decisions, enabling the model to adjust and upgrade in a timely manner according to changes in the on-orbit environment. This mechanism effectively overcomes the problem of difficulty in updating traditional models after deployment, giving the spaceborne intelligent diagnostic system long-term adaptive and continuous evolution capabilities.

[0107] The foregoing has only described certain exemplary embodiments of the present invention 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 the present invention. 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 the present invention.

Claims

1. A method for full-link hardware deployment and verification of a satellite on-orbit fault diagnosis model, characterized in that, Includes the following steps: Historical telemetry data from in-orbit satellites are collected and preprocessed to obtain a standardized dataset. At the same time, a fault knowledge base is constructed and data augmentation is performed to assess data quality. A diagnostic model containing data-driven channels, physical information channels, and knowledge-enhancing channels is trained using a standardized dataset and a fault knowledge base, and a full-precision floating-point model is output under the condition of meeting the preset performance indicators. Among them, the data-driven channel is used to extract the temporal features in telemetry data, capture dynamic changes at different time scales through a multi-scale temporal convolutional network, and enhance the ability to focus on key temporal segments by combining a self-attention mechanism. The physical information channel discretizes the system's physical equations and embeds learnable parameters, enabling the model to not only rely on data during training but also follow the system's physical laws. The knowledge enhancement channel encodes existing fault patterns and their relationships into a graph structure by introducing a knowledge graph, and uses a graph attention network for reasoning. The outputs of the three information processing channels are integrated through a fusion mechanism. Adaptive weight allocation is introduced during the fusion process, enabling the model to dynamically adjust the contribution ratio of each channel according to different input data. The full-precision floating-point model is distilled and pruned, and differentiated quantization is performed based on the precision sensitivity of different layers to obtain a quantized model; Based on the quantization model, the quantization model is converted into an FPGA bitstream file using an automatic compilation toolchain to complete the hardware implementation for the FPGA hardware platform, and fault tolerance processing is introduced to obtain the deployment result. Based on the deployment results, the functions, performance, reliability, and environment were verified, and a comprehensive evaluation was conducted to obtain the verification results. The validated FPGA hardware platform is run in orbit, and performance evaluation and update triggering are performed based on the in-orbit data. A model update feedback loop is established, and the model training, compression, and hardware implementation processes are re-executed based on the updated data.

2. The method of claim 1, wherein the method further comprises: determining the fault of the satellite based on the comparison result. The preprocessing procedure includes: detecting outliers in a sliding window using 3-sigma criterion and replacing them with linear interpolation; using cubic spline interpolation to complete the data segment with continuous missing data not exceeding a preset length; uniformly resampling parameters with different sampling rates to a reference frequency and using a robust normalization method The data is normalized, wherein, is the normalized data value, represents the original data value; represents the median; represents the 25th percentile; represents the 75th percentile.

3. The method of claim 1, wherein the method further comprises: determining the fault of the satellite based on the comparison result. The training strategy adopts a four-stage step-by-step approach: first, train the physical information channel to minimize the physical equation residuals; then, train the knowledge enhancement channel to learn knowledge graph embeddings; then, train the data-driven channel to optimize temporal feature extraction; and finally, jointly fine-tune all channels and the fusion module.

4. The method of claim 1, wherein the method further comprises: determining the fault of the satellite based on the comparison result. The quantization adopts a physical consistency-aware differentiated hybrid precision strategy. In the physical information channel, the learnable physical parameter layer and the physical equation residual calculation layer use INT16 quantization, while the convolutional layer and fully connected layer in the data-driven channel and the knowledge enhancement channel use INT8 quantization. The quantization process combines post-training quantization initialization with quantization-aware training fine-tuning.

5. The method of claim 1, wherein the method further comprises: determining the fault of the satellite based on the fault diagnosis model. Fault tolerance is a graded radiation hardening for the FPGA hardware platform. The graded radiation hardening mechanism includes three levels: the first level implements periodic readback loading and scrubbing for the FPGA configuration memory; the second level implements triple modular redundancy (TMR) hardening for the control state machine and output decision logic; and the third level implements checksum-based algorithm-level fault tolerance (ABFT) in matrix multiplication operations.

6. The method of claim 1, wherein the method further comprises: determining the fault of the satellite based on the comparison result. The passing standard for Level 4 progressive verification is: Functional verification requires hardware-floating-point consistency rate. The performance verification requires an average inference latency of no less than 90% and a power consumption of no more than 5W; the reliability verification requires continuous operation for no less than 72 hours without any abnormalities; and the environmental verification requires passing thermal vacuum, vibration, and electromagnetic compatibility tests.

7. The method of claim 1, wherein, The model update feedback loop includes the data feedback path, the knowledge feedback path, the model iteration path, and the on-orbit update path; The data feedback path is used to supplement the training dataset with on-orbit telemetry data; The knowledge feedback path is used to input new fault modes into the fault knowledge base; the model iteration path is used to re-execute the model training, compression, and hardware deployment process to generate a new model when the triggering conditions are met. The on-orbit update path is used to upload new bitstreams to the satellite via a software uploading channel to complete the inference engine replacement.

8. The method of claim 1, wherein the method further comprises: determining a fault of the satellite based on the comparison result. A model freshness indicator is also defined wherein, is the current time; is the time of the last update of the model; is the model life cycle, is the model freshness indicator, which, when below a preset threshold, triggers a preventive model update, and which, together with the on-orbit diagnosis coincidence rate, forms the basis for the model update decision.

9. The method of claim 1, wherein the method further comprises: determining a fault of the satellite based on the comparison result. The hardware implementation is completed through an automated compilation toolchain, which is used to convert the quantization model into an FPGA bitstream file. Its processing flow includes: parsing the ONNX model file to extract the computation graph structure; converting it into an intermediate representation and performing operator fusion and constant folding optimization; performing hardware mapping and scheduling according to FPGA resource constraints; generating synthesizable hardware description language code; and calling the FPGA vendor's toolchain to complete synthesis, placement and routing to generate the bitstream file.

10. The method of claim 1, wherein the method further comprises: determining a fault in the satellite; and determining a fault in the ground station. The model training process adopts a multi-task joint optimization approach, which involves collaborative training of three tasks: health status assessment, degradation trend prediction, and fault diagnosis. The weighted loss function is used to balance and optimize each task, thereby improving the model's comprehensive diagnostic performance and generalization ability under various operating conditions.

Images