Method for predicting residual service life of control moment gyroscope based on parameter fusion

Abstract

The invention discloses a method for predicting the remaining service life of a control moment gyroscope based on parameter fusion, and relates to the technical field of health management of spacecraft attitude control systems. According to the method, optimal fusion of measurement parameters and implicit parameters is realized through adversarial learning feature extraction and a channel attention mechanism, and the method is particularly suitable for high-precision residual life prediction of the spacecraft control moment gyroscope under variable working conditions. The method comprises four main steps of data preparation and multi-source input, adversarial learning feature extraction, SE-Attention feature fusion and three-stage training optimization. In the data preparation step, original time sequences of voltage and current of a rotor motor are collected through a built-in electrical sensor of a CMG, and key physical parameters of a full-film lubrication friction coefficient, lubricating oil viscosity, a bearing clearance and contact stiffness are inverted based on an electromechanical coupling model; in the adversarial learning step, a game framework of a feature extractor and a working condition discriminator is constructed, and working condition-independent robust feature extraction is realized through a gradient inversion layer; in the SE-Attention feature fusion step, importance weights are adaptively distributed to different feature channels through an extrusion-excitation mechanism; and the three-stage training optimization step adopts a preheating-confrontation-fine tuning strategy to ensure the network convergence. The method solves the problem of domain adaptation, has the technical advantages of strong robustness, high precision, strong generalization, interpretability and self-adaptation, and can realize high-precision RUL prediction under variable working conditions.

Description

Technical Field

[0001] This invention relates to the field of health management technology for spacecraft attitude control systems, specifically to a method for predicting the remaining service life of a control moment gyroscope based on parameter fusion. This method achieves optimal fusion of measured parameters and latent parameters through adversarial learning feature extraction and channel attention mechanisms, making it particularly suitable for high-precision prediction of the remaining service life of spacecraft control moment gyroscopes under varying operating conditions. Background Technology

[0002] Control moment gyroscopes (CMGs) are the core actuators of spacecraft attitude control systems. They provide precise attitude control torque to the spacecraft through changes in the angular momentum of a high-speed flywheel, and are crucial for achieving attitude stability and accurate pointing in complex space missions. With the increasing complexity of deep space exploration missions and the development of commercial spaceflight, spacecraft design lifespans have extended from the traditional 5-8 years to over 10-15 years. Therefore, health status assessments and remaining life predictions of CMGs are essential for ensuring their long-term reliable operation and have become a core guarantee for the success of space missions.

[0003] From an engineering practice perspective, CMG health monitoring and life prediction face numerous technical challenges. First, due to spacecraft launch costs and payload limitations, on-orbit CMGs cannot carry traditional mechanical monitoring sensors such as vibration and acoustic emission sensors. They can only rely on electrical telemetry signals such as motor current, voltage, and speed for condition monitoring. However, these signals have a complex nonlinear mapping relationship with the mechanical health status, leading to significant uncertainty when directly used for health assessment. Simultaneously, while multi-source information fusion is considered an effective path to improve prediction accuracy, it is difficult to efficiently integrate information of different types, time scales, and confidence levels. Traditional weighted averaging or voting fusion methods neither distinguish the importance and reliability of information sources nor properly handle information redundancy or conflicts. Second, CMG degradation is the result of the coupling effects of multiple physical domains, including mechanical, electrical, and thermal aspects. Taking a bearing system as an example, its degradation encompasses multiple stages, including lubricant decomposition, metal surface fatigue wear, and cage deformation and damage. Each stage has distinct physical characteristics, making it difficult for traditional single-parameter monitoring to comprehensively capture changes in health status.

[0004] Furthermore, the varying demands on CMG output torque across different mission phases of a spacecraft lead to frequent fluctuations in operating parameters such as frame rotation speed, flywheel rotation speed, and load torque. This makes it difficult for health assessment methods to distinguish between signal fluctuations caused by changes in operating conditions and signal anomalies caused by equipment degradation, easily resulting in misjudgments. In addition, while existing data-driven methods alleviate the challenges of physical modeling, they suffer from subjectivity issues such as reliance on expert experience in feature extraction and a lack of objective standards. Furthermore, their models have insufficient generalization capabilities, exhibiting significant performance degradation across different operating conditions, and also suffer from the drawback of insufficient utilization of multi-source heterogeneous information. Deep learning, while performing excellently in many fields, struggles to meet the high reliability requirements of aerospace due to the scarcity of operational data for aerospace-grade equipment, the "black box" nature of its models, and its sensitivity to data distribution. To address these technical challenges, a systematic solution is still lacking, particularly in areas such as how to effectively integrate physical mechanism knowledge with data-driven methods, how to handle small-sample learning problems, and how to achieve interpretable predictions. Further research and exploration are needed in these areas.

[0005] In summary, given the limitations of existing methods in addressing the severe interference from operating conditions, difficulties in fusing multi-source information, and poor interpretability in CMG health status identification and lifespan prediction, this invention proposes a control moment gyroscope-based method for predicting remaining lifespan based on parameter fusion. This method removes operating condition interference from measurement data and fuses it with intrinsic system state parameters using an attention mechanism. By introducing prior physical knowledge, interpretability is enhanced, providing reliable technical support for CMG health monitoring and lifespan prediction. Summary of the Invention

[0006] This invention addresses the problems of severe operating condition coupling interference, difficulty in multi-source information fusion, insufficient domain adaptability, and limited prediction accuracy in existing CMG (Continuous Usage) prediction technologies by providing a parameter fusion-based method for predicting the remaining usability (RUL) of a control moment gyroscope (CMG). This method achieves high-precision RUL prediction for CMGs under varying operating conditions through a core framework combining adversarial learning feature extraction and attention mechanism feature fusion.

[0007] The method for predicting the remaining service life of a control moment gyroscope based on parameter fusion comprises four main steps: data preparation and multi-source input, adversarial learning feature extraction, SE-Attention feature fusion, and three-stage training optimization. The data preparation and multi-source input step is primarily responsible for acquiring and preprocessing two types of heterogeneous data sources: measurement parameters and latent parameters. The adversarial learning feature extraction step is primarily responsible for extracting robust features independent of operating conditions through game-like training between the feature extractor and the adversarial discriminator. The SE-Attention feature fusion step is primarily responsible for adaptively assigning importance weights to different feature channels to achieve optimal fusion of measurement parameter features and latent parameters. The three-stage training optimization step is primarily responsible for ensuring network convergence and prediction accuracy through a phased training strategy.

[0008] The features of this invention are:

[0009] (1) By introducing the concept of adversarial learning, a game framework is designed between the Feature Extractor (FE) and the Operating Condition Discriminator (AD). The goal of the FE is to learn feature representations that can accurately predict the RUL (Roughly Indicative of Operating Conditions) but cannot be identified by the AD, while the goal of the AD is to identify the current operating condition as accurately as possible based on the features extracted by the FE. Through this minima-maximum game, the FE is forced to learn robust features decoupled from the operating condition.

[0010] (2) By introducing a channel attention mechanism, specifically, a squeeze-and-stimulation attention module is introduced.

[0011] ExcitationAttention (SE-Attention) automatically learns the importance weights for each feature channel. Specifically, it first compresses the features of each channel into a scalar through global average pooling, then learns the non-linear relationship between channels through two fully connected layers, and finally obtains weight coefficients between 0 and 1 through the sigmoid activation function. This mechanism can adaptively highlight the feature channels (usually latent parameter channels) that are most important for lifetime prediction, while suppressing the influence of redundant or noisy channels.

[0012] (3) Through the multi-layer nonlinear transformation of deep neural networks, two types of heterogeneous data are mapped to a common feature space. In this space, features from different sources are comparable and complementary. At the same time, by introducing residual connections and batch normalization techniques, the gradient vanishing and internal covariate shift problems of deep networks are effectively solved, ensuring that the network can fully learn the intrinsic relationship between the two types of data.

[0013] (4) By introducing a three-stage training strategy of warm-up-adversarial-fine-tuning, this progressive training strategy effectively avoids the divergence problem of adversarial training and ensures the stability and repeatability of the algorithm. Specifically, in the warm-up stage, the gradient inversion layer is frozen and FE and AD are trained independently to ensure that both networks have basic feature extraction and working condition discrimination capabilities.

[0014] During the adversarial phase, a gradient inversion layer is activated for joint adversarial training, and stable game convergence is achieved by dynamically adjusting the adversarial loss weights. During the fine-tuning phase, the adversarial loss weights are fixed, and only the RUL prediction loss is used for the entire network.

[0015] The network was finely adjusted to further improve prediction accuracy. Attached Figure Description

[0016] Figure 1 This is a flowchart of a method for predicting the remaining service life of a control torque gyroscope based on parameter fusion.

[0017] Figure 2 Flowchart for data preparation and multi-source input

[0018] Figure 3 This is a schematic diagram of an adversarial learning network structure.

[0019] Figure 4 This is a schematic diagram of the SE-Attention feature fusion mechanism.

[0020] Figure 5 This is a schematic diagram of the three-stage training strategy process. Detailed Implementation

[0021] The remaining service life prediction method for a control torque gyroscope based on parameter fusion, provided by the present invention, will be described in detail below with reference to the accompanying drawings.

[0022] This invention provides a method for predicting the remaining useful life (RUL) of a control moment gyroscope (CMG) based on parameter fusion. Overall, this method achieves high-precision RUL prediction for CMGs under varying operating conditions through four main steps: data preparation and multi-source input, adversarial learning feature extraction, SE-Attention feature fusion, and three-stage training optimization. Based on multi-source heterogeneous data, the method decouples the influence of operating conditions through adversarial learning and achieves optimal feature fusion using an attention mechanism, providing reliable lifetime prediction support for spacecraft health management. The overall flowchart of the parameter fusion-based control moment gyroscope RUL prediction method is shown below. Figure 1 As shown.

[0023] 1. Data Preparation and Multi-Source Input

[0024] The data preparation and multi-source input steps are used to acquire and preprocess two types of heterogeneous data sources: measurement parameters and latent parameters, providing high-quality input for subsequent feature extraction and fusion. These steps mainly include four sub-steps: measurement parameter acquisition, latent parameter identification, data preprocessing, and label construction. A schematic diagram of the data preparation and multi-source input process is shown below. Figure 2 As shown.

[0025] 1.1. Measurement Parameter Acquisition

[0026] The parameter acquisition step involves collecting operating status signals, including the original time series of rotor motor voltage U and rotor motor current I, through the electrical sensors built into the CMG. The acquired data is digitally stored, including timestamp information and sensor accuracy identifiers, to ensure the temporal integrity of the data and the reliability of the measurement.

[0027] 1.2. Latent Parameter Identification

[0028] The implicit parameter identification step is based on the physical mechanism model of CMG, and uses identification methods such as nonlinear least squares method and filtering algorithms to inversely derive key physical parameters characterizing the health status of CMG from electrical telemetry data. The implicit parameters include the full-film lubrication friction coefficient μ. EHL Lubricating oil viscosity ν, motor torque coefficient K e These parameters, extracted from the raw measurement data through electromechanical coupling models and parameter identification algorithms, have clear physical meaning and health indication functions.

[0029] 1.3. Data Preprocessing

[0030] The data preprocessing step employs multi-scale wavelet transform to denoise the original signal. By selecting the db4 wavelet basis function and a 5-level decomposition structure, measurement noise and environmental interference are effectively suppressed. Then, Z-score normalization is performed, calculated using the following formula:

[0031]

[0032] Where μ is the sequence mean and σ is the sequence standard deviation, ensuring that data of different dimensions are comparable.

[0033] 1.4. Tag Construction

[0034] The label construction steps described above construct two types of supervisory labels: RUL label y and operating condition label OC. The RUL label represents the remaining time from the current moment until the device fails, using the difference between the actual time to failure and the current time. The operating condition label is represented by the rotational speed of the outer frame of the CMG, providing a basis for operating condition discrimination in adversarial training.

[0035] 2. Adversarial learning feature extraction

[0036] The adversarial learning feature extraction stage extracts robust features decoupled from the working conditions by constructing a game framework between the feature extractor and the working condition discriminator. It mainly includes four sub-steps: network structure design, feature extractor training, working condition discriminator training, and gradient inversion optimization. A schematic diagram of the adversarial learning network structure is shown below. Figure 3 As shown.

[0037] 2.1 Network Structure Design

[0038] The network architecture design steps described above construct a dual-network architecture comprising a Feature Extractor (FE) and a Condition Discriminator (AD). The FE network uses a ResNet-based one-dimensional convolutional neural network structure, including an initial convolutional layer (64 kernels, kernel size 7) and three residual network blocks (64, 128, and 256 channels respectively). Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The AD network employs a composite structure, including a sequence discrimination layer (three fully connected layers with 128, 64, and 1 nodes respectively), a learnable upsampling network (three transposed convolutional layers for learning temporal patterns), and an adaptive pooling layer, capable of processing variable-length sequences and outputting condition discrimination results matching the length of the input sequence.

[0039] 2.2. Feature Extractor Training

[0040] The feature extractor training step prioritizes RUL prediction accuracy, updating network parameters by minimizing prediction error. The loss function uses Mean Squared Error (MSE) loss, defined as: Where n is the batch sample size, y i This represents the true remaining useful life (RUL) value for the i-th sample. Let be the RUL prediction value of the model for the i-th sample.

[0041] 2.3. Training of the Operating Condition Discriminator

[0042] The training steps of the operating condition discriminator aim to optimize the prediction and fitting accuracy of the continuous operating condition curve, updating the network parameters by minimizing the regression prediction error. The loss function uses Mean Squared Error (MSE) loss, defined as: Where n is the batch sample size, z i For the i-th sample, the true continuous value of the operating condition is given. This is the model's prediction result for the continuous working condition value of the i-th sample, ensuring that AD has accurate continuous working condition fitting and prediction capabilities.

[0043] 2.4. Gradient Reversal Optimization

[0044] The gradient reversal optimization step inserts a gradient reversal layer (GRL) between FE and AD to achieve adversarial training. The GRL acts as an identity transformation during forward propagation: f out =f in Reverse the gradient during backpropagation: Where λ is the adversarial weight coefficient. This mechanism makes FE and AD form a minima game: AD tries to identify the working condition based on features, while FE tries to extract features that cannot identify the working condition, and finally obtains a robust feature representation that is independent of the working condition.

[0045] 3. SE-Attention Feature Fusion

[0046] The SE-Attention feature fusion stage adaptively assigns importance weights to different feature channels through a squeeze-excitation mechanism, achieving optimal fusion of measurement parameter features and latent parameters. It mainly includes four sub-steps: feature concatenation, channel compression, weight learning, and feature recalibration. Specifically, firstly, the features of each channel are compressed into scalars using global average pooling. Then, the non-linear relationship between channels is learned through two fully connected layers with a dimensionality reduction ratio of 16. Finally, weight coefficients between 0 and 1 are obtained using the Sigmoid activation function. Further, the measurement parameter features and latent parameters are directly concatenated using a direct concatenation method and then input into the SE module. A schematic diagram of the SE-Attention feature fusion mechanism is shown below. Figure 4 As shown.

[0047] 3.1. Feature splicing

[0048] The feature stitching step involves extracting the measurement parameter features f from the FE. FE ∈R 125 With latent parameter features f HP ∈R 3 By concatenating the features along the feature dimension, we obtain the fused feature f. fused ∈R 128 The splicing operation preserves complete information about both features, providing a basis for subsequent attention selection.

[0049] 3.2. Channel Compression

[0050] The channel compression step described above uses global average pooling to compress the features of each channel into a scalar, calculated as follows:

[0051]

[0052] Where u c Let z be the feature map of the c-th channel, where L is the length of the one-dimensional feature, and z is the length of the feature map. cThis is a compressed scalar value used to achieve global information aggregation of feature channels.

[0053] 3.3. Weight Learning

[0054] The weight learning step learns the nonlinear relationship between channels through two fully connected layers. The network structure is: R 128 →R 16 →R 128 The first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the Sigmoid activation function, resulting in weight coefficients s∈R between 0 and 1. 128 This bottleneck structure can both learn complex channel dependencies and control model complexity.

[0055] 3.4. Feature Recalibration

[0056] The feature recalibration step involves multiplying the learned weight coefficients with the original features channel by channel to obtain the recalibrated features. This mechanism can adaptively highlight important feature channels and suppress redundant or noisy channels to achieve optimal feature fusion results.

[0057] 4. Three-stage training optimization

[0058] The three-stage training optimization phase employs a progressive training strategy to ensure network convergence and prediction accuracy, primarily comprising three sub-steps: a warm-up phase, an adversarial training phase, and a global fine-tuning phase. The flowchart of the three-stage training strategy is shown below. Figure 5 As shown.

[0059] 4.1. Preheating Stage

[0060] The aforementioned warm-up phase freezes the gradient inversion layer before training the FE and AD networks. First, the gradient inversion layer is frozen, establishing a collaborative relationship between FE and AD. The entire system is trained for 50 epochs, enabling FE to acquire basic feature extraction capabilities and AD to acquire basic situation discrimination capabilities. This warm-up strategy provides a suitable initial network for subsequent adversarial games, avoiding performance fluctuations in the early training stages.

[0061] 4.2. Combat Training Phase

[0062] The adversarial training phase activates the gradient inversion layer to jointly train the FE, AD, and FusionPredictor (FP). The total loss function consists of three parts:

[0063] L total =L RUL +αL disc

[0064] Where L RULTo predict the loss for RUL, L disc The operating condition discrimination loss is represented by α, which is the adversarial weight, initially 0.1 and gradually increased to 0.5 during training. A dynamic weight adjustment strategy is adopted to ensure the stability and convergence of the adversarial training.

[0065] 4.1. Global Fine-tuning Phase

[0066] The global fine-tuning phase described above uses a relatively small learning rate of 1×10. -5 End-to-end fine-tuning of the entire network, using only RUL to predict the loss L RUL Update parameters. In this stage, the adversarial weights are fixed, focusing on improving the final prediction accuracy. Training continues for 500 epochs until convergence. Further refinement of network parameters enhances the model's predictive performance and generalization ability.

Claims

1. A method for predicting the remaining service life of a control torque gyroscope based on parameter fusion, characterized in that, The method includes the following steps: a data preparation and multi-source input step, which acquires and preprocesses two types of heterogeneous data sources: measurement parameters and latent parameters; an adversarial learning feature extraction step, which constructs a game framework between the feature extractor and the working condition discriminator to extract robust features independent of working conditions; an SE-Attention feature fusion step, which adaptively assigns importance weights to different feature channels to achieve optimal fusion of measurement parameter features and latent parameters; and a three-stage training optimization step, which ensures network convergence and prediction accuracy through a phased training strategy.

2. The method according to claim 1, characterized in that, The data preparation and multi-source input steps specifically include: a measurement parameter acquisition sub-step, which acquires the raw time series of rotor motor voltage and current through the electrical sensors built into the CMG; a latent parameter identification sub-step, which inversely derives key physical parameters such as the full-film lubrication friction coefficient, lubricating oil viscosity, and motor torque coefficient from the electrical telemetry data based on an electromechanical coupling model; a data preprocessing sub-step, which uses multi-scale wavelet transform for noise reduction and Z-score standardization to ensure data comparability; and a label construction sub-step, which constructs RUL labels and operating condition labels to provide supervisory information for adversarial training.

3. The method according to claim 1, characterized in that, The adversarial learning feature extraction step specifically includes: a network structure design sub-step, constructing a dual-network architecture containing a feature extractor and a working condition discriminator. The feature extractor adopts a one-dimensional convolutional neural network structure based on ResNet, and the working condition discriminator adopts a composite structure containing a sequence discrimination layer, a learnable upsampling network, and an adaptive pooling layer; a feature extractor training sub-step, with RUL prediction accuracy as the main optimization objective, using the mean squared error (MSE) loss function; a working condition discriminator training sub-step, with prediction fitting accuracy of continuous working condition curves as the optimization objective, using the mean squared error (MSE) loss function; and a gradient inversion optimization sub-step, implementing adversarial training between the feature extractor and the working condition discriminator through a gradient inversion layer. The gradient inversion layer acts as an identity transformation during forward propagation and inverts the gradient during backward propagation.

4. The method according to claim 1, characterized in that, The SE-Attention feature fusion step specifically includes: a feature concatenation sub-step, which concatenates the extracted measurement parameter features and latent parameter features along the feature dimension to obtain fused features; a channel compression sub-step, which compresses the features of each channel into scalars through global average pooling; and a weight learning sub-step, which learns the nonlinear relationship between channels through two fully connected layers, with the network structure being R. 128 →R 16 →R 128 The first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the Sigmoid activation function to obtain weight coefficients between 0 and 1. The feature relabeling step multiplies the learned weight coefficients with the original features channel by channel to obtain the relabeled features.

5. The method according to claim 1, characterized in that, The three-stage training optimization steps specifically include: a warm-up stage, freezing the gradient inversion layer to establish a collaborative relationship between the feature extractor and the adaptive generator (AD), and training for 50 epochs to enable the FE to acquire basic feature extraction capabilities and the AD to acquire basic operational condition discrimination capabilities; and an adversarial training stage, activating the gradient inversion layer to jointly train the feature extractor, operational condition discriminator, and fusion predictor, with a total loss function of L. total =L RUL +αL disc Where α is the adversarial weight, initially 0.1 and gradually increased to 0.5; in the global fine-tuning phase, a relatively small learning rate of 1×10 is used. -5 Fine-tune the entire network end-to-end, using only RUL prediction loss to update parameters, and train for 500 epochs until convergence.

6. The method according to claim 1, characterized in that, The method extracts condition-independent features through adversarial learning and achieves optimal fusion of measurement parameters and latent parameters through the SE-Attention mechanism.

7. The method according to claim 1, characterized in that, The feature extractor adopts a one-dimensional convolutional neural network structure based on ResNet, which includes an initial convolutional layer and three residual network blocks. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The working condition discriminator adopts a composite structure, which includes a sequence discrimination layer, a learnable upsampling network, and an adaptive pooling layer. It can process variable-length sequences and output working condition discrimination results that match the length of the input sequence.

Images