In-orbit satellite remaining service life prediction method based on deep learning

Abstract

The invention discloses an in-orbit satellite remaining service life prediction method based on deep learning, and the method comprises the steps: feature extraction and fusion: obtaining various types of sensor data of an in-orbit satellite, extracting time domain, frequency domain and time-frequency domain features from the sensor data, carrying out the fusion of the time domain, frequency domain and time-frequency domain features through employing a principal component analysis method, and carrying out the prediction of the remaining service life of the in-orbit satellite; fusion features are obtained; health index construction: inputting the fusion features into a pre-trained health index generation model, and outputting a health index curve by the health index generation model; and residual life prediction: inputting the health index curve into a pre-trained life prediction model, and outputting the residual life by the life prediction model. According to the method, through feature extraction and fusion of the data of the multi-source sensor, the information with the highest discrimination capability is reserved, and the accuracy of residual life prediction is improved.

Description

Technical Field

[0001] The present invention relates to the field of satellite engineering technology, and in particular to a method for predicting the remaining useful life of an on-orbit satellite based on deep learning. Background Art

[0002] During the operation of satellites in orbit, they may face complex space environments (such as high-energy particle radiation, temperature fluctuations) and hardware aging, which may lead to system failures. By predicting the remaining life, the potential failure risks of satellites can be identified in advance to ensure that the satellites maintain reliable operation during the mission, thereby avoiding the interruption or failure of critical missions. Life prediction provides data support for satellite health management, enabling operators to take targeted maintenance or adjustment strategies based on the prediction results. For repairable satellites (such as high-value satellites with on-orbit maintenance capabilities), life prediction can help determine the best maintenance time and avoid premature or late maintenance. For non-repairable satellites, life prediction helps to formulate their retirement plans and ensure orderly exit from orbit to reduce space debris. Through life prediction and health management, satellite resources (such as electricity and fuel) can be reasonably allocated and adjusted to help engineers optimize the satellite's task scheduling, so as to complete more tasks at the end of its life. Therefore, satellite remaining life prediction is of great significance. Sensor data effectively reflects the working status of on-orbit satellites and is an important basis for remaining life prediction. How to achieve satellite remaining life prediction based on sensor data is an urgent problem to be solved. Summary of the invention

[0003] In order to solve some or all of the technical problems existing in the above-mentioned prior art, the present invention provides a method for predicting the remaining useful life of an in-orbit satellite based on deep learning.

[0004] The technical solution of the present invention is as follows:

[0005] An embodiment of the present invention provides a method for predicting the remaining useful life of an on-orbit satellite based on deep learning, the method comprising:

[0006] Feature extraction and fusion: Acquire various sensor data of on-orbit satellites, extract time domain, frequency domain and time-frequency domain features from the sensor data, fuse the time domain, frequency domain and time-frequency domain features using principal component analysis to obtain fused features;

[0007] Health index construction: inputting the fusion features into a pre-trained health index generation model, the health index generation model outputs a health index curve, and the health index generation model is trained based on the fusion features and their corresponding health indicators;

[0008] Remaining life prediction: The health index curve is input into a pre-trained life prediction model, and the life prediction model outputs the remaining life. The life prediction model is trained based on the health index curve and its corresponding remaining life.

[0009] In one embodiment of the present invention, the training process of the health indicator generation model is as follows:

[0010] Acquire a training set, where each piece of data in the training set includes a fusion feature and a corresponding health indicator;

[0011] Construct health indicators to generate initial models and set training parameters;

[0012] The health indicator is trained according to the data in the training set to generate an initial model, and the model parameters are updated to obtain a trained health indicator generation model.

[0013] In one embodiment of the present invention, the training process of the remaining life prediction model is as follows:

[0014] Acquire a training set, wherein each data in the training set includes a health indicator curve and a corresponding remaining lifespan;

[0015] Construct the initial model for remaining life prediction and set the training parameters;

[0016] The initial remaining life prediction model is trained according to the data in the training set, and the model parameters are updated to obtain the trained remaining life prediction model.

[0017] In one embodiment of the present invention, the initial model for generating the health index is a convolutional neural network or a multi-layer perceptron;

[0018] The initial model for remaining life prediction is a convolutional neural network or a long short-term memory network.

[0019] In one embodiment of the present invention, the time domain features are multiple or all of the following statistical features:

[0020] Maximum value, maximum absolute value, minimum value, mean value, peak-to-peak value, standard deviation, absolute mean value, RMS value, RMS amplitude, standard deviation, kurtosis, skewness, margin index, waveform index, pulse index, peak index.

[0021] In one embodiment of the present invention, the frequency domain features are extracted from the time domain signal by fast Fourier transform, and the frequency domain features include centroid frequency, frequency root mean square and / or frequency variance.

[0022] In one embodiment of the present invention, the time-frequency domain features are extracted using an empirical mode decomposition method, and the time-frequency domain features are intrinsic mode function components of the signal.

[0023] In one embodiment of the present invention, before training the life prediction model, the health indicator curve is segmented into different health stages, each health stage corresponds to a different health indicator curve change rate, the remaining life curve is segmented according to the segmented health indicator curve, and then the correspondence between the health indicator curve and the remaining life is determined, and based on the correspondence between the segmented health indicator curve and the remaining life, a training set of the life prediction model is constructed.

[0024] In one embodiment of the present invention, the health indicator curve segmentation uses a bottom-up algorithm for segmented fitting, and based on the fitting results, it is divided into two health stages with the initial failure moment as the turning point. The remaining life before the initial failure moment remains unchanged, and the remaining life after the initial failure moment decreases with time.

[0025] In one embodiment of the present invention, the satellite sensor data includes the following or all of them: shaft temperature, shell temperature, current, voltage, friction torque and rotation speed of the satellite.

[0026] The main advantages of the technical solution of the present invention are as follows:

[0027] The deep learning-based remaining useful life prediction method for in-orbit satellites of the present invention extracts and fuses the features of multi-source sensor data, retains the most discriminative information as the input of the health indicator generation model, and obtains health indicators that can accurately reflect the status of the satellite, making the health status assessment of the satellite more accurate and improving the accuracy of RUL prediction. Through accurate RUL prediction, on the one hand, potential satellite failures can be discovered in advance, which facilitates the adoption of preventive maintenance measures and reduces the probability of failures. On the other hand, satellite maintenance and task scheduling can be reasonably planned to avoid premature or late maintenance, thereby extending the service life of the satellite and reducing maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0029] Figure 1 This is a flowchart of a method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to an embodiment of the present invention;

[0030] Figure 2 It is a flowchart of the overall implementation process of the method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to an embodiment of the present invention;

[0031] Figure 3A flowchart of extracting time-frequency domain features by empirical mode decomposition in a method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to an embodiment of the present invention;

[0032] Figure 4 A flowchart of the training and prediction process of a health indicator generation model in a method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to an embodiment of the present invention;

[0033] Figure 5 A flowchart of segmenting a health index curve by a bottom-up segmentation algorithm in a method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to an embodiment of the present invention;

[0034] Figure 6 A schematic diagram of the remaining service life label setting in the method for predicting the remaining service life of an on-orbit satellite based on deep learning according to an embodiment of the present invention;

[0035] Figure 7 This is a flow chart of the training and prediction process of the life prediction model in the method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0037] The technical solution provided by the embodiments of the present invention is described in detail below with reference to the accompanying drawings.

[0038] At present, remaining useful life prediction is one of the main tasks of prognostic and health management (PHM), which aims to mine degradation trends from condition monitoring data to predict remaining useful life (RUL). RUL prediction is generally divided into direct method and indirect method. The direct method directly uses condition monitoring data to build a degradation model to obtain RUL, while the indirect method is to construct a health indicator (HI) that effectively reflects the degradation state from the monitoring data, and further obtain RUL based on HI. According to its own characteristics, HI can be further divided into physical health indicators and virtual health indicators. Physical health indicators are directly related to physical characteristics, such as one of the time domain, frequency domain, and time-frequency domain characteristics of monitoring data. Virtual health indicators do not have actual meaning and are often obtained by fusion of multiple features.

[0039] In recent years, due to the widespread application of deep learning, a large number of RUL prediction methods based on deep learning have emerged. Convolution Neural Network (CNN) and Long Short Term Memory (LSTM) have powerful feature extraction and time series prediction capabilities, and are two commonly used network models for building HI and realizing RUL prediction. In addition, feature fusion technologies such as principal component analysis (PCA) are the main auxiliary methods for realizing RUL prediction, which can effectively obtain the main features of the signal and reduce the complexity of the problem. At present, direct RUL prediction and indirect RUL prediction methods based on deep learning have been applied in many fields such as aircraft engines, rolling bearings, lithium batteries, circuit systems, etc., and have achieved good results.

[0040] However, for the special object of on-orbit satellites, the sensor data is diverse and huge in quantity, and there is little research on how to achieve RUL prediction of on-orbit satellites. Therefore, the purpose of the present invention is to form a complete set of on-orbit satellite RUL prediction methods based on deep learning methods to make up for the application defects of RUL prediction in the satellite field.

[0041] The embodiment of the present invention provides a method for predicting the remaining useful life of an on-orbit satellite based on deep learning, as shown in the attached Figure 1 and attached Figure 2 As shown, including:

[0042] S1, feature extraction and fusion: obtain various sensor data of on-orbit satellites, extract time domain, frequency domain and time-frequency domain features from the sensor data, and use principal component analysis to fuse the time domain, frequency domain and time-frequency domain features to obtain fused features.

[0043] The data collected by the satellite's various sensors (such as temperature sensors, current, voltage, friction torque, etc.) are diverse and complex, and a single data feature may not be able to fully reflect the health of the satellite. Therefore, this step can capture subtle changes in the satellite's working status through multi-angle feature extraction in the time domain, frequency domain, and time-frequency domain. Through principal component analysis (PCA), multiple features are fused to reduce redundancy and retain the most discriminative information.

[0044] Therefore, the feature extraction and fusion stage can effectively integrate the data features from different sensors to generate a comprehensive and information-rich feature set, providing accurate input data for the subsequent health indicator (HI) construction. Through feature fusion, the data dimension is reduced, the computational complexity is reduced, and the sensitivity of the prediction model to satellite state changes is enhanced.

[0045] S2, health indicator construction: the fusion features are input into a pre-trained health indicator generation model, and the health indicator generation model outputs a health indicator curve. The health indicator generation model is trained based on the fusion features and their corresponding health indicators.

[0046] It is understandable that the health index generation model needs to be pre-trained before use, and the training set should include the input and corresponding output of the model, that is, the fusion feature and its corresponding health index HI. HI is an important quantitative indicator reflecting the overall health status of the satellite, which can comprehensively express the working status and degradation trend of the satellite. In an embodiment of the present invention, the data in the training set is derived from historical satellite data, generally including various data of the entire life cycle of the satellite, that is, including satellite sensor data at different time points from the commissioning to the end of the use of the satellite and the remaining life value at the corresponding moment. The health index is an intermediate indicator for measuring the health status of the satellite. It is set by technical personnel in this field based on experience and is an empirical value. In one possible implementation, as the remaining life of the satellite decreases, the health index continues to increase. Of course, vice versa is also possible, as long as the relationship between the health index and the remaining life can be reflected.

[0047] After determining the health indicators at different times, the sensor data, fusion features, health indicators, and remaining life span all correspond to time. According to the corresponding relationship between the fusion features and the health indicators, a training set including fusion features and health indicators is determined. During the training process, the deep learning model can extract implicit features that are closely related to the health status of the satellite from complex fusions by learning the potential patterns in historical data, and generate a continuous indicator curve that reflects the health status of the satellite. In an embodiment of the present invention, a health indicator refers to a value at a certain moment. During the prediction and training process, data at a single moment is often difficult to meet the requirements. Therefore, the input of the health indicator generation model is the fusion features at multiple different moments in a continuous period of time, and the corresponding output is the health indicators at multiple different moments in the period. The health indicators at multiple different moments can form a health curve.

[0048] S3, remaining life prediction: the health indicator curve is input into a pre-trained life prediction model, and the life prediction model outputs the remaining life. The life prediction model is trained based on the health indicator curve and its corresponding remaining life.

[0049] It is understandable that the life prediction model needs to be pre-trained before use, and the training set should include the model input and corresponding output, that is, based on the health index curve and its corresponding remaining life. By inputting the HI curve and its corresponding RUL during the training process, the model can learn the complex relationship between the HI curve and its remaining life, thereby providing accurate RUL prediction in practical applications.

[0050] In summary, the deep learning-based remaining useful life prediction method for an in-orbit satellite provided by an embodiment of the present invention extracts and fuses features of multi-source sensor data, retains the most discriminative information as the input of the health indicator generation model, and obtains health indicators that can accurately reflect the status of the satellite, making the health status assessment of the satellite more accurate and improving the accuracy of RUL prediction. Through accurate RUL prediction, on the one hand, potential satellite failures can be discovered in advance, which facilitates the adoption of preventive maintenance measures and reduces the probability of failures. On the other hand, satellite maintenance and task scheduling can be reasonably planned to avoid premature or late maintenance, thereby extending the service life of the satellite and reducing maintenance costs.

[0051] The working principle of each step of the method for predicting the remaining useful life of an on-orbit satellite based on deep learning provided by an embodiment of the present invention is described in detail below:

[0052] 1. Feature extraction and fusion

[0053] This section will introduce in detail how to extract and fuse features from various forms of sensor data. The initial sensor data may include but is not limited to the satellite shaft temperature, shell temperature, current, voltage, friction torque, speed and other forms of data. Define a certain type of sensor data (such as current) as X = [x 1 ,x 2 ,...,x N ], the data length is N, and it is necessary to extract signals in the time domain, frequency domain, and time-frequency domain. The feature extraction methods of different types of sensor data are consistent.

[0054] (1) Time domain signal extraction

[0055] The time domain features mainly reflect the correlation between the signal and time. The commonly used time domain feature calculation formulas are shown in Table 1. Generally, several or all of the time domain features in Table 1 are selected for subsequent feature fusion. The selected time domain features are defined as s i ∈{1,2,..,15} and s i Different from each other.

[0056] Table 1 Commonly used time domain features

[0057]

[0058]

[0059] (2) Frequency domain feature extraction

[0060] The characteristic changes of some signals in the time domain may not be obvious, resulting in the inability to detect the fault state. However, converting the time domain characteristics into frequency domain characteristics can often reflect the fault state more intuitively. Generally, the fast Fourier transform algorithm is used to obtain the frequency domain characteristics. Since the fast Fourier transform method is relatively mature, the detailed process will not be repeated. The commonly used frequency domain characteristics are shown in Table 2, where r k represents the sampling frequency, S k Represents the power spectrum. Generally, several or all frequency domain features are selected for subsequent fusion, and the selected frequency domain features are defined as p i ∈{1,2,3,4} and p i Different from each other.

[0061] Table 2 Commonly used frequency domain features

[0062]

[0063] (3) Time-frequency domain feature extraction

[0064] The time-frequency domain features can simultaneously analyze the changes of the signal from both the time domain and the frequency domain, so as to make it easier to find abnormal states. The present invention mainly uses the empirical mode decomposition method to extract the time-frequency domain features. Empirical Mode Decomposition (EMD) is an algorithm widely used to process nonlinear and non-stationary time series signals. Using the EMD algorithm, the original signal can be decomposed into multiple intrinsic mode functions (IMF). Each IMF has a different frequency and represents a simple vibration mode compared to the simple harmonic function. Each IMF must meet two basic conditions: 1) The number of extreme points and zero points of the entire signal must be equal to or less than 1; 2) At any point, the average extreme value of the signal is zero.

[0065] For a given signal X = [x 1 ,x 2 ,...,x N ], the EMD algorithm flow chart is as follows Figure 3 As shown. First, find all the local maximum and local minimum values ​​of signal X, and then use the cubic curve interpolation method to get the maximum and minimum values ​​of the envelope curve. Then, calculate the average envelope curve U, and subtract U from the original signal to get a new signal X new Next, determine X new Whether the two conditions of IMF are met. If so, the first IMF component is IMF 1 =X new , the residual signal is calculated as R = X-IMF 1 If not, then X=X newRepeat the above process. Then determine whether the residual signal R is monotonic. If not, X = R and repeat the above operation until R meets the conditions. Finally, obtain all IMFs and residual functions, and decompose the original signal X into X = ∑IMF i +R. Generally, the first few IMFs are selected as time-frequency domain signals, and the selected time-frequency domain signals are defined as t i Different from each other.

[0066] (4) Principal component analysis

[0067] After extracting the time domain, frequency domain, and time-frequency domain features, the selected features need to be further fused. The present invention mainly uses the principal component analysis method for feature fusion [3]. The specific steps of the principal component analysis method will not be repeated here. The selected features are All features are normalized and the calculation formula is as follows:

[0068]

[0069] The normalized features are recorded as The normalized features are subjected to principal component analysis. Generally, the two principal components with the largest contribution are selected as the result of principal component analysis, that is, the fused features, recorded as Finally, according to a certain type of sensor data in different historical periods, the fusion features are obtained according to the above process and divided into training sets and test sets.

[0070] 2. HI Curve Construction

[0071] HI is an important indicator reflecting the degradation state of satellites. The deep learning-based method can further extract implicit features and integrate the features. Converted into a one-dimensional HI curve. Usually, it is necessary to build a neural network model for training. The model input is the fusion feature and the output is the HI curve. The choice of model needs to be based on the characteristics of the data set itself. Generally, common model structures such as convolutional neural network and multi-layer perceptron can be used. In addition, the setting of model hyperparameters also needs to be based on the actual data characteristics. Taking CNN as an example, the specific construction process of the HI curve is as follows: Figure 4 shown.

[0072] 3. HI curve segmentation and RUL label setting

[0073] The HI curve reflects the degradation process of the satellite from normal to failure. Under normal circumstances, the satellite is in a normal state and lasts for a long time, showing a relatively stable change trend on the HI curve. As the satellite parts gradually fail in the later period, the HI curve shows a gradual upward trend until it fails completely. Therefore, according to the change of HI, the satellite's full life operation process can be divided into multiple different health stages, based on which timely adjustments and maintenance can be made to the different states of the satellite. In addition, through the division of health states, the time of initial failure can be obtained, which is important for constructing RUL labels. The initial state of the HI curve of the satellite's full life cycle output by the health indicator generation model is a continuous irregular curve. By finding the turning point of the health indicator curve after segmentation, the turning point of the satellite state can be set accordingly, and the remaining life curve can be corrected according to the turning point of the health indicator curve, so that the satellite remaining life curve is more in line with the actual working conditions of the satellite. According to the corrected remaining life curve, the corresponding relationship between the health indicator curve and the remaining life is determined, so that the correspondence between the health indicator curve and the remaining life is more accurate, and the accuracy of the remaining life prediction is improved.

[0074] First, the existing time series segmentation algorithm can be used for HI curve segmentation, such as top-down (TPD) and bottom-up (BUP). The present invention mainly performs HI curve segmentation based on the BUP algorithm. The HI curve is defined as h = [h 1 ,h 2 ,...,h H ], the sequence length is H. The algorithm uses linear approximation to perform segmented fitting. First, the HI curve is divided into H / 2 linear segments to approximate the entire HI curve. These H / 2 segments are the initial approximation of the HI curve. Then, the adjacent segments are merged using the least squares method, and the cost of merging every two adjacent segments is calculated, that is, the error between the two adjacent segments before fitting and the curve after fitting is calculated. Finally, the adjacent segments with the smallest cost are iteratively merged until the number of segments meets the set number of segments. The main calculation process of the BUP algorithm is as follows: Figure 5 shown.

[0075] Normally, the label of RUL is given manually. One of the commonly used methods is that RUL shows a straight-down trend. The second method is to consider that it is in normal working state for a period of time at the beginning, and RUL shows a trend of first being a constant value and then falling straight down. However, the labels obtained based on experience are often inaccurate, especially for the second method. The turning point setting of RUL showing a straight-down trend is very important, which affects the entire prediction process. Therefore, after using the BUP algorithm to segment the HI curve, the initial failure moment of the satellite can be found according to the segmented health status, which can be used as the turning point of the RUL label, thereby constructing a more accurate RUL label and laying the foundation for the next step of RUL prediction. The schematic diagram of constructing RUL according to the divided HI is shown in the figure. Figure 6 shown.

[0076] IV. RUL Prediction

[0077] After obtaining the HI curve, it is further divided into training set and test set, and a neural network model is constructed for training. The input of the model is the HI curve, and the output is RUL. The label of RUL is set according to the initial fault moment obtained by the BUP algorithm. The selection of model structure and model parameters depends on the characteristics of the data set itself. Various structures such as convolutional neural network and long short-term memory network can be used. The specific process of RUL prediction is as follows Figure 7 As shown, the model structure constructed can be selected arbitrarily.

[0078] It should be noted that the above HI curve segmentation and RUL label setting steps only occur after the health indicator generation model is trained and before the remaining life model is trained. The purpose is to determine the relationship between the health indicator and the remaining life, so as to facilitate the construction of the training set of the life prediction model. The training set contains the correspondence between the health indicator curve and the remaining life. The health indicator curve is the original curve (unsegmented) output by the health indicator generation model, and the remaining life is obtained from the corresponding corrected remaining life curve after segmentation according to the HI curve. After the two models are trained, there is no need to segment the health indicator curve in the process of using satellite sensor data for satellite life prediction.

[0079] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In addition, "front", "back", "left", "right", "upper" and "lower" in this article are all referenced to the placement state shown in the accompanying drawings.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the remaining useful life of an on-orbit satellite based on deep learning, characterized in that: include: Feature extraction and fusion: Acquire various sensor data of on-orbit satellites, extract time domain, frequency domain and time-frequency domain features from the sensor data, fuse the time domain, frequency domain and time-frequency domain features using principal component analysis to obtain fused features; Health index construction: inputting the fusion features into a pre-trained health index generation model, the health index generation model outputs a health index curve, and the health index generation model is trained based on the fusion features and their corresponding health indicators; Remaining life prediction: The health index curve is input into a pre-trained life prediction model, and the life prediction model outputs the remaining life. The life prediction model is trained based on the health index curve and its corresponding remaining life.

2. The method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to claim 1, characterized in that: The training process of the health indicator generation model is as follows: Acquire a training set, where each piece of data in the training set includes a fusion feature and a corresponding health indicator; Construct health indicators to generate initial models and set training parameters; The health indicator is trained according to the data in the training set to generate an initial model, and the model parameters are updated to obtain a trained health indicator generation model.

3. The method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to claim 2, characterized in that: The training process of the remaining life prediction model is as follows: Acquire a training set, wherein each data in the training set includes a health indicator curve and a corresponding remaining lifespan; Construct the initial model for remaining life prediction and set the training parameters; The initial remaining life prediction model is trained according to the data in the training set, and the model parameters are updated to obtain the trained remaining life prediction model.

4. The method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to claim 3 is characterized in that: The initial model for generating the health index is a convolutional neural network or a multi-layer perceptron; The initial model for remaining life prediction is a convolutional neural network or a long short-term memory network.

5. The method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to any one of claims 1 to 4, characterized in that: The time domain features are multiple or all of the following statistical features: Maximum value, maximum absolute value, minimum value, mean value, peak-to-peak value, standard deviation, absolute mean value, RMS value, RMS amplitude, standard deviation, kurtosis, skewness, margin index, waveform index, pulse index, peak index.

6. The method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to claim 5, characterized in that: The frequency domain features are extracted from the time domain signal by fast Fourier transform, and the frequency domain features include centroid frequency, frequency root mean square and / or frequency variance.

7. The method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to claim 6, characterized in that: The time-frequency domain features are extracted using an empirical mode decomposition method, and the time-frequency domain features are intrinsic mode function components of the signal.

8. The method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to claim 7, characterized in that: Before training the life prediction model, the health index curve is segmented into different health stages, each health stage corresponds to a different rate of change of the health index curve, the remaining life curve is segmented according to the segmented health index curve, and then the correspondence between the health index curve and the remaining life is determined, and according to the correspondence between the health index curve and the remaining life, a training set of the life prediction model is constructed.

9. The method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to claim 8, characterized in that: The health index curve segmentation uses a bottom-up algorithm for segmented fitting, and according to the fitting results, it is divided into two health stages with the initial failure moment as a turning point. The remaining life before the initial failure moment remains unchanged, and the remaining life after the initial failure moment decreases with time.

10. The method for predicting the remaining useful life of an on-orbit satellite based on deep learning according to claim 1, characterized in that: The satellite sensor data may include the following or all of the following: shaft temperature, shell temperature, current, voltage, friction torque and rotation speed of the satellite.

Images