Method for predicting residual service life of control moment gyroscope based on reconstruction error

Abstract

The invention discloses a control moment gyroscope remaining service life prediction method based on reconstruction errors, and belongs to the technical field of spacecraft health management. The method comprises the following steps: collecting a multi-dimensional sensor signal and carrying out wavelet denoising and normalization; performing singular value decomposition on the data to extract features; training an auto-encoder model by using the normal state features, and calculating a reconstruction error to quantify an abnormal degree; based on the reconstruction error, determining a threshold value and constructing a health index, and performing index moving average and adaptive Kalman filtering processing; and calculating an estimated degradation speed and combining theoretical design information to carry out weighted fusion correction to obtain a predicted remaining service life. According to the method, key information of multi-dimensional sensor data can be fully mined, accurate identification of an abnormal state and accurate prediction of the residual service life are realized, the accuracy and reliability of a control moment gyroscope health monitoring system are remarkably improved, and a powerful guarantee is provided for safe and stable operation of a spacecraft attitude control system.

Description

Technical Field

[0001] This invention belongs to the field of spacecraft health management technology, specifically relating to a method for health status perception and identification and remaining useful life (RUL) prediction of control moment gyroscopes (CMGs) on spacecraft. Background Technology

[0002] Control moment gyroscopes (CMGs) are the core actuators of spacecraft attitude control systems. They provide precise control torque to the spacecraft through changes in angular momentum generated by a high-speed rotating flywheel, thereby enabling flexible adjustment and stable maintenance of the spacecraft's attitude. Considering the long-term and complex nature of spacecraft on-orbit missions, CMGs typically have a design life of over 10 years. However, during long-term on-orbit operation, the performance of CMGs inevitably undergoes gradual degradation due to the combined effects of multiple complex factors, including mechanical component wear, alternating extreme temperatures in space, orbital vibration interference, and the corrosive effects of the vacuum environment. In severe cases, this can directly lead to malfunctions, posing a fatal threat to the accuracy of spacecraft attitude control and even the successful execution of the entire on-orbit mission.

[0003] Since the health status of the CMG (Continuous Metal Gearbox) directly determines the reliability and stability of the spacecraft's attitude control system, real-time and accurate health monitoring, and scientific prediction of its remaining service life based on the monitoring data, are of crucial engineering significance for ensuring the safe operation of spacecraft in orbit and improving the quality of mission completion. The CMG is a type of rotating machinery. Currently, there are various technical approaches for monitoring the health of rotating machinery. Among them, vibration signal-based monitoring methods are commonly used. This method achieves fault identification and health status perception by analyzing the spectral characteristics of vibration signals. However, CMGs on spacecraft are usually not equipped with vibration sensors, making the application of this method difficult. Some studies use temperature signal-based monitoring methods, focusing on the temperature change patterns of key CMG components and judging the health status through abnormal temperature fluctuations. However, changes in temperature signals often lag behind the actual fault occurrence process, easily leading to delays in fault diagnosis.

[0004] In addition to the two methods mentioned above, monitoring methods based on current signals identify faults by analyzing the changing characteristics of the CMG drive current. However, since the current signal is strongly coupled with the operating conditions of the CMG, complex operating condition compensation algorithms are needed to eliminate interference, which greatly increases the difficulty of technical implementation. On the other hand, the remaining service life prediction scheme based on statistical methods usually assumes that the degradation process of the CMG follows a specific statistical distribution (such as the Weibull distribution) and completes the estimation of the remaining service life by fitting the degradation curve. This method is difficult to adapt to the complex nonlinear degradation characteristics of the CMG in actual operation, resulting in limited applicability of the prediction results.

[0005] In summary, existing CMG health monitoring and remaining service life prediction technologies still have many significant shortcomings, making it difficult to meet the actual needs of high reliability assurance for spacecraft. In the feature extraction stage, traditional single-dimensional feature extraction methods in the time or frequency domains cannot fully exploit the coupled information contained in multi-dimensional sensor data, making it difficult to construct a feature index system that can comprehensively and accurately characterize the health status of the CMG. At the anomaly detection level, existing methods mostly employ fixed threshold judgment mechanisms. This mechanism lacks adaptability to differences in the structure and performance parameters of different CMG models, as well as different on-orbit operating environments, and is prone to anomaly judgment bias.

[0006] It is worth noting that the performance degradation process of CMGs generally exhibits complex characteristics of strong nonlinearity and multi-factor coupling. Traditional linear modeling methods are unable to accurately characterize their degradation patterns, resulting in insufficient accuracy in health status assessment. In terms of remaining service life prediction, existing methods mostly rely on a single degradation model for prediction, failing to effectively integrate the complementary information of differences between different models during the design phase and actual on-orbit degradation data. This leads to prediction accuracy that is difficult to meet the requirements of engineering applications. At the same time, existing technologies have poor robustness to complex space noise, and are prone to false alarms and missed alarms under noise interference, seriously affecting the reliability of the monitoring system and thus threatening the safety of spacecraft attitude control.

[0007] Therefore, with the increasing reliability requirements of spacecraft attitude control systems, there is an urgent need to develop a CMG health management method with efficient feature extraction capabilities, accurate anomaly identification performance, and high-precision remaining service life prediction. By fully exploring the value of multi-dimensional sensor data, accurately adapting to nonlinear degradation characteristics, and integrating multi-source information to improve prediction accuracy, this method can ultimately achieve precise control over the entire lifecycle of CMG health status, providing solid technical support for the reliable operation of spacecraft attitude control systems. Summary of the Invention

[0008] This invention addresses the shortcomings of existing control moment gyroscope (CMG) remaining service life prediction technologies, such as insufficient feature extraction effectiveness, poor anomaly detection adaptability, weak nonlinear pattern recognition capability, limited prediction accuracy, and poor robustness. It proposes a CMG remaining service life prediction method based on singular value decomposition (SVD) feature extraction and an autoencoder. This method integrates core technologies such as efficient SVD feature extraction, accurate anomaly detection using an autoencoder, and weighted degradation rate prediction to fully extract key information from CMG multidimensional sensor data. This enables accurate identification of abnormal states and precise prediction of remaining service life, ultimately improving the accuracy and reliability of the CMG health monitoring system and providing strong support for the safe and stable operation of spacecraft attitude control systems.

[0009] The remaining useful life prediction method proposed in this invention follows a complete technical process of "data acquisition - feature extraction - anomaly detection - health assessment - life prediction," with each step closely linked to form a closed-loop management system. First, multidimensional data from the CMG's operation is collected via sensors, covering signals closely related to its health status, such as vibration, temperature, and current. Based on this, the collected raw data is windowed, and SVD decomposition is performed on each window to extract the row vectors of the singular value matrix as core features, achieving dimensionality reduction of high-dimensional data while preserving key information. Subsequently, the autoencoder model is trained using the SVD features of the CMG under normal conditions. The root mean square error (RMSE) between the input data and the reconstructed model data is calculated to quantify the degree of data deviation from the normal state, thus accurately representing the degree of anomaly.

[0010] After determining the anomaly threshold based on the statistical characteristics of reconstruction errors from normal data, a health index (HI = 1 - normalized anomaly ratio) is defined by calculating the proportion of errors exceeding the threshold, thus achieving a quantitative description of the health status. To improve the stability and accuracy of the health index, multiple processing methods, including exponential moving average (EMA) smoothing and adaptive Kalman filtering (AKF), are applied to the health index series to effectively suppress noise interference and correct abnormal fluctuations. Finally, considering different theoretical designs and actual operating conditions, the estimated degradation rate is calculated and then corrected based on theoretical design information to achieve accurate prediction of the remaining useful life (RUL), ensuring that the prediction results are consistent with the actual degradation characteristics of the CMG.

[0011] This invention extracts implicit feature information from the original multidimensional data through SVD, calculates the health index and degradation trend based on the difference between the reconstruction error of the autoencoder network and the normal state, and combines theoretical design information to correct the estimated lifespan, thereby achieving accurate and stable health assessment and remaining service life prediction for CMG.

[0012] The features of this invention are:

[0013] (1) By applying SVD feature extraction technology to CMG multidimensional sensor data processing, we can achieve dimensionality reduction of high-dimensional data and preservation of core structural information, break through the limitation of insufficient information utilization of traditional single-dimensional feature extraction methods, and provide high-quality feature support for health status identification.

[0014] (2) By learning the unsupervised learning characteristics of the autoencoder model, the normal data distribution features are learned to achieve adaptive identification of abnormal states, which breaks through the adaptability shortcomings of the traditional fixed threshold detection method and improves the accuracy of abnormal detection in different CMG models and complex operating environments.

[0015] (3) Through a multi-fusion mechanism of EMA smoothing and AKF filtering, noise suppression and fluctuation correction of the health index sequence are achieved.

[0016] It meets the core requirements of industrial scenarios for the stability and accuracy of CMG health assessment. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for predicting the remaining service life of a control torque gyroscope based on reconstruction error.

[0018] Figure 2 A schematic diagram of the data acquisition and preprocessing process.

[0019] Figure 3 This is a schematic diagram of the SVD feature extraction process.

[0020] Figure 4 This is a schematic diagram of the anomaly detection process for an autoencoder.

[0021] Figure 5 Diagram of the health index calculation process

[0022] Figure 6 This is a schematic diagram of the remaining useful life prediction process. Detailed Implementation

[0023] The following description, in conjunction with the accompanying drawings, details a method for predicting the remaining service life of a control moment gyroscope based on reconstruction error, provided by the present invention.

[0024] This invention provides a method for predicting the remaining service life of a control moment gyroscope (CMG) based on reconstruction error. Overall, this method achieves full lifecycle health status monitoring and remaining service life prediction for a CMG through five main stages: data acquisition and preprocessing, SVD feature extraction, autoencoder anomaly detection, health index calculation and correction, and remaining service life prediction. The method is based on multidimensional sensor data, constructs a health index model through feature dimensionality reduction and anomaly detection, and corrects the estimated degradation rate by combining theoretical design information to achieve accurate prediction of the remaining service life. The overall flowchart of the method is shown below. Figure 1 As shown.

[0025] 1. Data Acquisition and Preprocessing

[0026] The data acquisition and preprocessing stage is used to obtain raw data on the CMG's operating status and perform noise reduction and standardization processing to provide high-quality input for subsequent feature extraction. It mainly includes four steps: multi-dimensional sensor data acquisition, noise reduction processing, filtering processing, and normalization processing. A schematic diagram of the data acquisition and preprocessing process is shown below. Figure 2 As shown.

[0027] 1.1. Multidimensional sensor data acquisition

[0028] The multi-dimensional sensor data acquisition step involves collecting operating status signals, including high-speed rotor current, high-speed rotor voltage, and high-speed rotor speed, using a sensor array deployed on the CMG. The acquired data is stored in time-series format, including sensor channel identifiers and timestamp information.

[0029] 1.2. Noise Reduction Processing

[0030] The denoising process described employs wavelet transform to eliminate environmental and measurement noise. By selecting appropriate wavelet basis functions and decomposition levels (default is db4 wavelet and 5-level decomposition), the original signal is decomposed using wavelet transform, retaining effective components in high-frequency detail coefficients, suppressing noise interference components, and reconstructing the denoised signal.

[0031] 1.3. Normalization Process

[0032] The normalization process maps the filtered signal to the [-1, 1] interval, and the calculation formula is as follows:

[0033]

[0034] Where x is the original signal value, and min(x) and max(x) are the minimum and maximum values ​​of the signal sequence, respectively. Normalization can eliminate the dimensional differences between signals from different sensors, providing data with a uniform scale for subsequent feature extraction.

[0035] 2. SVD Feature Extraction

[0036] The SVD feature extraction stage extracts key features characterizing the health status of CMG from preprocessed data through windowing and singular value decomposition. It mainly includes four steps: windowing, singular value decomposition, feature vector construction, and feature standardization. A schematic diagram of the SVD feature extraction process is shown below. Figure 3 As shown.

[0037] 2.1. Windowed processing

[0038] The windowing process divides the preprocessed sensor data into multiple consecutive windows, each containing a fixed number of samples. The default window size is 64 samples, which can be adjusted to 128, 256, etc., depending on the CMG model. The window step size is the same as the window size by default. To ensure the reliability of feature extraction, this step incorporates an adaptive parameter adjustment mechanism to adapt to different data volumes. Specifically, it first calculates the number of windows that can be generated with the current data volume, using the following formula: Where N is the total amount of data, w size(where 'step' is the current step size). If the number of windows calculated is less than 512, the step size is gradually reduced and the number of windows is recalculated until the number of windows meets the requirement of being greater than or equal to 512. In order to balance computational efficiency and prediction accuracy, the adjusted step size is finally determined to be the largest power of 2 that meets the conditions, so as to ensure that a sufficient number of effective windows can be obtained in different data volume scenarios, providing a stable data foundation for subsequent SVD feature extraction.

[0039] 2.2. Singular Value Decomposition

[0040] The singular value decomposition step performs SVD decomposition on each window data matrix X (m×n, where m is the number of sensor channels and n is the window size) to obtain X = U×S×V T The matrix decomposition form is given by: U being a left singular matrix (m×m) with column vectors being left singular vectors; S being a singular value matrix (m×n) with diagonal elements being singular values ​​arranged in descending order; and V being a right singular matrix (n×n) with row vectors being right singular vectors. The decomposition process satisfies... Where k is the effective rank, σ i For the i-th singular value, u i and v i These are the corresponding left and right singular vectors, respectively.

[0041] 2.3. Feature Vector Construction

[0042] The eigenvector construction step extracts the first m singular values ​​from the singular value matrix S to form the eigenvector. Since the singular values ​​σ1≥σ2≥…≥σ m ≥0 reflects the variance contribution of the data in the orthogonal direction. The first m singular values ​​can contain the main information of the data and can effectively characterize the operating status characteristics of CMG.

[0043] 2.4. Feature Standardization

[0044] The feature standardization step uses the Z-score method to standardize the extracted SVD features, with the following formula:

[0045]

[0046] Where μ is the mean of the feature sequence and σ is the standard deviation of the feature sequence. The standardized data follows a distribution with a mean of 0 and a standard deviation of 1, which facilitates the training and convergence of the autoencoder model.

[0047] 3. Self-encoder anomaly detection

[0048] The aforementioned autoencoder anomaly detection stage identifies CMG anomalies by constructing a fully connected autoencoder model. This mainly includes four steps: model structure design, model training, reconstruction error calculation, and anomaly threshold determination. A schematic diagram of the autoencoder anomaly detection process is shown below. Figure 4 As shown.

[0049] 3.1. Model Structure Design

[0050] The model structure design steps described above construct a fully connected autoencoder network containing an encoder and a decoder. The encoder contains three hidden layers with 64, 32, and 16 nodes respectively, mapping the input features to a low-dimensional latent space. The decoder contains three hidden layers with 16, 32, and 64 nodes respectively, reconstructing the latent representation into the original features. The hidden layers use the ReLU activation function, and the output layer uses the Sigmoid activation function to ensure that the output range matches the standardized features.

[0051] 3.2. Model Training

[0052] The model training steps described above use SVD feature data from a CMG in normal operating condition for model training. The training objective is to minimize the mean squared error (MSE) between the input features and the reconstructed features, and the loss function is... Where x i As input features, x′ i To reconstruct features, N is the number of samples; for optimization, the Adam optimizer is used with a learning rate of 0.001, a batch size of 32, and 100 iterations until the loss function converges.

[0053] 3.3. Calculation of Reconstruction Error

[0054] The reconstruction error calculation step involves inputting the SVD features of the test data into the trained autoencoder and calculating the reconstruction error (RMSE) using the following formula:

[0055]

[0056] Where n is the feature dimension, the larger the reconstruction error, the more significantly the data deviates from the normal state.

[0057] 3.4. Determination of Abnormal Thresholds

[0058] The aforementioned abnormal threshold determination step determines the threshold for distinguishing between normal and abnormal data based on the statistical characteristics of the reconstruction error of normal data. Two methods are supported: one is the 3-sigma rule, T = μ + 3σ, where μ is the mean of the reconstruction error of normal data and σ is the standard deviation; the other is the MAD rule, T = median + 3 × MAD, where median is the median and MAD is the absolute median deviation, MAD = median(|x i -median|).

[0059] 4. Health Index Calculation and Adjustment

[0060] The health index calculation and correction stage constructs a health index that can characterize the CMG degradation trend through anomaly ratio transformation and multiple smoothing mechanisms. It mainly includes five steps: anomaly ratio calculation, health index definition, exponential moving average smoothing, and adaptive Kalman filtering. A schematic diagram of the health index calculation process is shown below. Figure 5 As shown.

[0061] 4.1. Calculation of Abnormal Ratio

[0062] The aforementioned abnormality ratio calculation step uses a sliding window to statistically analyze the proportion of reconstruction errors exceeding a threshold. The sliding window size is 1% of the total number of windows, with a minimum of 10 windows. The calculation formula is as follows:

[0063]

[0064] Where N >T N represents the number of errors exceeding the threshold within the sliding window. total This represents the total number of windows contained within the sliding window.

[0065] 4.2. Definition of Health Index

[0066] The health index definition step converts the abnormality ratio into a health index: HI = 1 - r, where HI ∈ [0, 1]. HI = 1 indicates complete health, HI = 0 indicates complete failure, and CMG is determined to be ineffective when HI < 0.1.

[0067] 4.3. Exponential Moving Average (EMA) Smoothing

[0068] The EMA smoothing step smooths the health index sequence to reduce noise interference. The calculation formula is as follows:

[0069] HI EMA (t)=α·HI raw (t)+(1-α)·HI EMA (t-1)

[0070] Where α is the smoothing coefficient, the default is 0.1, and it can be adjusted in the range of 0.05 to 0.2. EMA (0) = HI raw (0), HI raw (t) represents the original health index at time t.

[0071] 4.4. Adaptive Kalman Filter (AKF) Processing

[0072] The AKF treatment steps further smooth the health index curve and preserve the degradation trend, including:

[0073] Equation of state: HI k =HI k-1 +w k

[0074] Observation equation: z k =HI k +v k

[0075] Where w k ~N(0,Q) represents process noise, v k ~N(0,R) represents the observation noise, and Q and R are adjusted in real time using an adaptive algorithm. The adaptive algorithm dynamically adjusts the process noise covariance q and the observation noise covariance R based on the prediction residual. When the residual increases, Q is increased and R is decreased to improve tracking capability.

[0076] 5. Remaining useful life prediction

[0077] The remaining useful life prediction stage, by fusing theoretical design information to correct the estimated degradation rate, achieves accurate prediction of the remaining useful life of the CMG. It mainly includes two steps: calculating the estimated degradation rate and weighted fusing theoretical correction prediction. A schematic diagram of the remaining useful life prediction process is shown below. Figure 6 As shown.

[0078] 5.1. Calculation of Estimated Degradation Rate

[0079] The estimated rate of degradation calculation step is based on the actual degradation rate of the current health index curve, and the formula is:

[0080]

[0081] HI initial The initial health index is 1 by default, HI current For the current health index, T current The current operating time is in years. To avoid the impact of short-term fluctuations, the average degradation rate of the most recent 3 months is used as the actual degradation rate.

[0082] 5.2. Calculation of Estimated Degradation Rate

[0083] The weighted fusion theory correction prediction step uses theoretical design information to correct the calculated degradation rate. First, the actual remaining useful life is calculated based on the actual degradation rate, using the following formula:

[0084]

[0085] Then, the theoretical design information is incorporated for correction. This theoretical design information is based on a linear degradation model assuming a design life, and its calculation formula is as follows:

[0086]

[0087] HI initial The initial health index is 1 by default, HI failure The failure threshold is 0.1 by default, T design The design life is given in years. Based on this, a weighted fusion prediction step is used to combine theory with the estimated degradation rate to accurately estimate the remaining useful life. The final remaining useful life is obtained using a weighted fusion method, and the fusion formula is as follows:

[0088]

[0089] Where w is the weight of the theoretical degradation rate, which defaults to 0.8 and can be dynamically adjusted according to the CMG operation stage, with a value of 0.9 in the early stage, 0.8 in the middle stage, and 0.7 in the later stage.

Claims

1. A method for predicting the remaining service life of a control moment gyroscope based on reconstruction error, characterized in that: The method for predicting the remaining service life of a control moment gyroscope (CMG) based on reconstruction error extracts hidden feature information from the original multidimensional data through singular value decomposition (SVD), calculates the health index and degradation trend based on the difference between the reconstruction error of the autoencoder network and the normal state, and combines theoretical design information to correct the estimated service life, thereby achieving accurate and stable health assessment and remaining service life prediction of the CMG. The method includes data acquisition and preprocessing, SVD feature extraction, autoencoder anomaly detection, health index calculation and correction, and remaining service life prediction steps.

2. The method for predicting the remaining useful life of CMG according to claim 1, characterized in that: The windowing process in the SVD feature extraction step of the described method uses a fixed window size, with a default window size of 64 samples and a step size of 64 samples. It incorporates an adaptive parameter adjustment mechanism: when the data volume is small, the step size is automatically reduced to ensure that the number of windows is greater than or equal to 512, and the step size is adjusted to the largest power of 2 that satisfies the condition. Further, Singular Value Decomposition (SVD) decomposes the data matrix X into the product of three matrices: X = U × S × V. T , where U is the left singular matrix, S is the singular value matrix, and V is the right singular matrix. The first m singular values ​​are extracted from the singular value matrix S to form an eigenvector, where m is the number of sensor parameter channels.

3. The method for predicting the remaining useful life of CMG according to claim 1, characterized in that: The autoencoder model in the anomaly detection step of the method adopts a fully connected network structure. The encoder contains three hidden layers (with 64, 32, and 16 nodes respectively), and the decoder contains three hidden layers (with 16, 32, and 64 nodes respectively). The hidden layers use the ReLU activation function, and the output layer uses the Sigmoid activation function. The model training objective is to minimize the mean squared error (MSE) between the input features and the reconstructed features, and the Adam optimizer is used for training.

4. The method for predicting the remaining useful life of CMG according to claim 1, characterized in that: The abnormal threshold in the health index calculation step of the method is determined based on the statistical characteristics of the reconstruction error of normal data, and supports two methods: 3-sigma rule: the threshold is the mean of the reconstruction error of normal data + 3 times the standard deviation; MAD rule: the threshold is the median of the reconstruction error of normal data + 3 times the absolute median difference.

5. The method for predicting the remaining useful life of CMG according to claim 1, characterized in that: In the health index calculation step of the method, the sliding window size is 1% of the total number of windows, with a minimum of 10 windows. The health index HI = 1 - abnormality ratio, where HI = 1 indicates complete health, HI = 0 indicates complete failure, and when HI < 0.1, it is determined that CMG has failed.

6. The method for predicting the remaining useful life of CMG according to claim 1, characterized in that: The adaptive Kalman filter (AKF) in the health index calculation step of the method adapts to different degradation modes by adaptively adjusting the filter gain, further smoothing the health index curve and preserving the degradation trend.

7. The method for predicting the remaining useful life of CMG according to claim 1, characterized in that: The estimated degradation rate in the remaining useful life prediction step of the method is calculated based on the estimated degradation rate of the current health index curve, using the following formula: HI initial This is the initial health index, with a default value of 1. current For the current health index, T current The current running time (in years) is used. To avoid the impact of short-term fluctuations, the average degradation rate of the most recent 3 months is used as the estimated degradation rate.

8. The method according to claim 1, characterized in that: The weighted fusion method for correcting the remaining useful life by incorporating theoretical design information in the remaining useful life prediction step of the method is as follows: HI current HI is the current health index. failure This is the failure threshold, with a default value of 0.

1. theory v is the theoretical degradation rate. est To estimate the degradation rate, w is the weight of the theoretical design information, with a default value of 0.

8. It can be dynamically adjusted according to the CMG operation stage, with a value of 0.9 in the early stage, 0.8 in the middle stage, and 0.7 in the later stage.

Images