Ion electric propulsion on-orbit fault digital feature extraction method

Abstract

The invention relates to the technical field of propulsion system fault diagnosis, in particular to an ion electric propulsion on-orbit fault digital feature extraction method, which comprises the following steps of: acquiring ion electric propulsion operation parameters under ground test and on-orbit working conditions, constructing a convolutional neural network to respectively extract time domain features and frequency domain features of the operation parameters, and extracting the time domain features and the frequency domain features of the operation parameters; fusing to form a fault feature vector, and identifying through a fault classifier; a feature extraction and classification network is pre-trained by utilizing digital feature knowledge of source domain data (ground test data), target domain and source domain features are aligned by utilizing a transfer learning method, and a feature extraction module is finely adjusted by utilizing target domain data, so that the engineering adaptability of ion electric propulsion on-orbit fault feature extraction is effectively improved; the problems of scarcity of on-orbit fault data, high noise and the like of the ion electric propulsion system are solved, and the accuracy and robustness of on-orbit health state evaluation of the ion electric propulsion system are enhanced.

Description

Technical Field

[0001] This application relates to the field of propulsion system fault diagnosis technology, and more specifically, to a method for extracting digital features of on-orbit faults in ion electric propulsion. Background Technology

[0002] Ion electric propulsion systems, with their high specific impulse, have become the core power source for modern spacecraft, and their operational status directly affects the spacecraft's attitude and orbit control missions and service life. However, under on-orbit conditions, ion electric propulsion systems operate in a complex environment with multi-physics coupling for extended periods, which may lead to malfunctions due to component aging, arc discharge, and other issues, resulting in performance degradation or even failure.

[0003] Currently, some fault diagnosis methods for ion electric propulsion based on deep learning or machine learning have emerged. However, in practical engineering applications, they suffer from technical problems such as small sample sizes and high noise levels. Fault feature extraction based on deep learning or machine learning methods typically relies on a large amount of labeled fault data for ion electric propulsion. However, due to the scarcity of on-orbit test data for aerospace products, the number of on-orbit fault samples for ion electric propulsion systems is small, making it difficult to meet the training requirements of deep learning models. Affected by the space environment, the collected on-orbit operational data of ion electric propulsion has high noise levels, and the on-orbit acquisition frequency is often lower than that of ground tests. This makes it difficult to extract digital features of on-orbit faults, resulting in distribution differences between the data and those obtained under ground test conditions. Consequently, fault diagnosis models trained directly from ground test data are difficult to apply to on-orbit missions.

[0004] In existing technologies, transfer learning has been applied to fields such as image recognition and speech processing to address the problem of scarce data in specialized fields. However, there has been no systematic research on transfer learning methods for fault diagnosis of ion electric propulsion systems. Summary of the Invention

[0005] This application provides a method for extracting digital features of on-orbit faults in ion electric propulsion. By making full use of ground test data through transfer learning technology, it solves the problems of scarce on-orbit fault data and high noise, thereby improving the accuracy and adaptability of fault digital feature extraction.

[0006] To achieve the above objectives, this application provides a method for extracting digital features of on-orbit faults in ion electric propulsion, based on transfer learning technology, including the following steps: Step 1: Under the working state of the ion electric propulsion system, collect various operating parameters of the system, including operating electrical parameters, gas supply parameters and temperature parameters, and record fault information that occurs during the process. Step 2: Preprocess the working parameter sequences collected from the ground test and in-orbit operation of ion electric propulsion respectively, extract data segments with the same time length, and perform center normalization on the data within the segments; Step 3: Extract time-domain features from the preprocessed collected data. Use statistical methods to extract the statistical characteristics of each parameter and concatenate the calculation results into a one-dimensional time-domain information vector. Step 4: Construct a fault digital feature extraction module. Use a one-dimensional convolutional neural network to extract time-domain features from the time-domain information vector in Step 3 to obtain a one-dimensional time-domain feature vector. Step 5: Extract time-frequency features from the preprocessed collected data, using Morlet as the wavelet basis function to obtain the time-frequency spectra of each parameter; Step 6: Using a two-dimensional convolutional neural network, with the time-frequency spectrum from Step 5 as input, extract time-frequency features and concatenate the time-frequency features of each parameter into a one-dimensional time-frequency feature vector; Step 7: Fuse time-domain features and frequency-domain features to obtain the fault digital feature vector; Step 8: The fault digital features extracted from the ground test of the ion electric propulsion system are regarded as source domain features, and the health status digital features extracted in orbit are regarded as target domain features; Step 9: Construct a fault classification module based on a fully connected neural network, pre-train the fault digital feature extraction module and the fault classification module in the source domain, and obtain a fault digital feature extraction module and a fault classification module with generalization ability; Step 10: In the target domain, the MMD loss function is introduced to evaluate the difference in the distribution of features extracted by the fault digital feature extraction module from on-orbit data and ground data. Then, the fault classification module is fine-tuned to align the feature distributions of the source domain and the target domain, thereby realizing on-orbit fault digital feature extraction based on transfer learning for ion electric propulsion.

[0007] Furthermore, in step 1, the operating electrical parameters include the voltage and current of the ion thrust grid, the voltage and current of the acceleration grid, the voltage and current of the discharge chamber anode, and the contact voltage and current of the cathode and neutralizer; the gas supply parameters include the flow rate of each working fluid in the storage and supply unit; and the temperature parameters include the temperature of the front shell of the ion thruster, the neutralizer, and the pole shoes, as well as the temperature of each module in the power processing unit.

[0008] Furthermore, in step 3, time-domain analysis is performed based on the preprocessed working electrical parameters, gas supply parameters, and temperature parameters to calculate the time-domain information of the maximum and minimum values, skewness, kurtosis, waveform factor, peak factor, and impulse factor of each parameter in the data segment, and the calculation results are spliced ​​into a one-dimensional vector form.

[0009] Furthermore, in step 5, frequency domain analysis is performed based on the preprocessed working electrical parameters, gas supply parameters, and temperature parameters. First, the working parameter signal is subjected to single-step differential processing. Then, based on the differential signal, wavelet transform is used to obtain the time-frequency spectrum of each parameter in the data segment.

[0010] Furthermore, in step 9, the selected fault data segments are labeled to train the fault digital feature extraction and fault classification modules. The fault classification corresponding to each data segment is expressed in the form of a confidence vector. The vector dimension is equal to the number of fault categories. When the data record indicates that a fault has occurred, the corresponding element in the vector is recorded as 1, and the others are recorded as 0. If all elements in the vector are 0, it is considered that the electric propulsion system has not experienced a fault.

[0011] Furthermore, in step 9, the ion electric propulsion fault classification module is constructed based on a fully connected neural network. The network receives the fault digital feature vector fused in step 7 as input and outputs the fault classification vector of the current data segment.

[0012] Furthermore, in step 10, the MMD loss function is introduced to evaluate the distribution difference of features extracted by the fault digital feature extraction module from on-orbit data and ground data on the target domain features, and the weight parameters of the convolutional neural network in the fault digital feature extraction module are fine-tuned in combination with the fault classification loss.

[0013] This application provides a method for extracting digital features of on-orbit faults in ion electric propulsion, which has the following advantages: This application utilizes digital feature knowledge from source domain data to pre-train feature extraction and classification networks, employs transfer learning methods to align features between the target domain and source domain, and fine-tunes the feature extraction module using target domain data. This effectively improves the engineering adaptability of on-orbit fault feature extraction for ion electric propulsion, solves the problems of scarce on-orbit fault data and high noise in ion electric propulsion systems, and enhances the accuracy and robustness of on-orbit health status assessment for ion electric propulsion systems. Attached Figure Description

[0014] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application.

[0015] In the attached diagram: Figure 1 This is a flowchart of the on-orbit fault digital feature extraction method for ion electric propulsion provided in the embodiments of this application; Figure 2 This is a schematic diagram of the time-domain and time-frequency feature extraction process provided according to the embodiments of this application; Figure 3 This is a schematic diagram of the training of the fault digital feature extraction module provided in the embodiments of this application. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] In this application, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," "horizontal," "lateral," and "longitudinal" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are primarily for the purpose of better describing this application and its embodiments, and are not intended to limit the indicated device, element, or component to having a specific orientation, or to be constructed and operated in a specific orientation.

[0019] Furthermore, in addition to indicating location or positional relationship, some of the aforementioned terms may also have other meanings. For example, the term "above" may also be used in some cases to indicate a certain dependency or connection relationship. Those skilled in the art can understand the specific meaning of these terms in this application based on the specific circumstances.

[0020] In addition, the term "multiple" should mean two or more.

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] Current methods for extracting fault features in ion electric propulsion systems mainly suffer from the following problems: 1. On-orbit fault data is relatively scarce, resulting in a small number of on-orbit fault samples for ion electric propulsion systems, which is insufficient to meet the training requirements of deep learning models; 2. Due to limitations in on-board resources and the influence of the on-orbit environment, the sampling frequency and accuracy of on-orbit operational data for ion electric propulsion are limited, and the data is easily affected by electromagnetic interference, environmental changes, and other factors, resulting in high acquisition noise and blurred digital features; 3. Fault features of ion electric propulsion systems have strong nonlinearity and time-varying characteristics, and the complexity of the space operating environment will exacerbate these characteristics, leading to differences in the distribution of on-orbit data and ground data, resulting in poor on-orbit adaptability of fault diagnosis methods based on ground data.

[0023] This application provides a method for extracting digital features of on-orbit faults in ion electric propulsion, such as... Figure 1 As shown, based on ground-based and on-orbit data from ion electric propulsion tests, a fault digital feature extraction module based on convolutional neural networks is constructed. This module extracts the time-domain and time-frequency features of the test data, fuses them, and inputs them into a neural network-based fault classification module. Using ground-based test data as the source domain and on-orbit test data as the target domain, a transfer learning method is employed to align the features of the target and source domains. This addresses the problems of scarce and noisy on-orbit fault data, improving the accuracy and adaptability of fault digital feature extraction. The specific steps include: Step 1: Under the working state of the ion electric propulsion system, collect various operating parameters of the system, including operating electrical parameters, gas supply parameters and temperature parameters, and record fault information that occurs during the process. The operating electrical parameters include the voltage and current of the ion thruster grid, the voltage and current of the accelerating grid, the voltage and current of the discharge chamber anode, and the contact voltage and current of the cathode and neutralizer; the gas supply parameters include the flow rate of each working fluid in the storage and supply unit; the temperature parameters include the temperature of the ion thruster front shell, neutralizer and pole shoes, and the temperature of each module in the power processing unit. Step 2: Preprocess the working parameter sequences collected from the ground test and in-orbit operation of ion electric propulsion respectively, extract data segments with the same time length, and perform center normalization on the data within the segments; Specifically, the ion electric propulsion data involved in this application embodiment includes on-orbit operating parameters and ground test operating parameters, corresponding to target domain data and source domain data, respectively. Due to on-board resource limitations, the dimensionality and sampling frequency of on-orbit data are usually lower than those of ground tests. For both on-orbit and ground tests, a fixed duration T is used to extract fault data segments. For operating parameters of the same category, they are recorded in the following manner, where t i Let n be the number of samples within a segment, representing different sampling times within a duration T, as shown in formula (1): The operation of an ion electric propulsion system involves working parameters with different dimensions and distribution patterns, including current, voltage, temperature, and pressure. To eliminate the interference of these differences on the accuracy of the deep learning algorithm, each parameter is normalized by centering the average time value of each parameter in the data segment as the center, and the mean of each feature parameter is zeroed out. As shown in formula (2): The variance of each parameter within the data segment after zero-mean normalization is calculated and then normalized. As shown in formulas (3) and (4), the preprocessed parameter sequence is shown in formula (5): Step 3: Extract time-domain features from the preprocessed collected data. Use statistical methods to extract the statistical characteristics of each parameter and concatenate the calculation results into a one-dimensional time-domain information vector. Step 4: Construct a fault digital feature extraction module. Use a one-dimensional convolutional neural network to extract time-domain features from the time-domain information vector in Step 3 to obtain a one-dimensional time-domain feature vector. Specifically, based on the preprocessed working electrical parameters, gas supply parameters, and temperature parameters, time-domain analysis is performed to calculate the time-domain information of the maximum and minimum values, skewness, kurtosis, waveform factor, peak factor, and impulse factor of each parameter in the data segment, and the calculation results are concatenated to form a one-dimensional vector. More specifically, such as Figure 2 As shown, for the preprocessed source and target domain data, time-domain feature information is extracted. For each type of data in the data segment, such as discharge voltage and accelerating current, the maximum value, minimum value, standard deviation, root mean square, skewness, kurtosis, waveform factor, peak factor, impulse factor, and margin factor of the parameter sequence within the time segment are calculated as the time-domain features of that indicator. Calculations are performed for each indicator, and the time-domain features of each parameter are concatenated into a vector as the time-domain feature. , where x skew The parameter sequence represents the skewness; n represents the number of samples within time segment T, and x represents the skewness. std The standard deviation of the parameter series, x kurt Indicates the kurtosis of the parameter sequence; x wave The waveform factor representing the parameter sequence; x rms Let x represent the root mean square of the parameter sequence. peak The peak factor of the parameter sequence; x pulse The pulse factor represents the parameter sequence; the time-domain information sequence of a single parameter is shown in formula (12), and the time-domain information sequences of all parameters are concatenated into the form of formula (13): Using the above time-domain information sequence as input, 1D-CNN is used to extract features from the time-domain information. The output of the convolutional neural network is the time-domain feature vector, as shown in formula (14): ; Step 5: Extract time-frequency features from the preprocessed collected data, using Morlet as the wavelet basis function to obtain the time-frequency spectra of each parameter; Step 6: Using a two-dimensional convolutional neural network, with the time-frequency spectrum from Step 5 as input, extract time-frequency features and concatenate the time-frequency features of each parameter into a one-dimensional time-frequency feature vector; Specifically, such as Figure 2 As shown, frequency domain analysis is performed based on the preprocessed working electrical parameters, gas supply parameters, and temperature parameters. First, the working parameter signal is subjected to single-step differential processing. Then, based on the differential signal, wavelet transform (Mollet wavelet basis function) is used to obtain the time-frequency spectrum of each parameter in the data segment. On this basis, a 2D-CNN is used to extract features from the time-frequency spectrum. The convolutional neural network outputs a one-dimensional vector, which is the frequency domain feature vector, as shown in formula (15). ; Step 7: Fuse the time-domain features and frequency-domain features to obtain the fault digital feature vector, as shown in formula (16): ; Step 8: The fault digital features extracted from the ground test of the ion electric propulsion system are regarded as source domain features, and the health status digital features extracted in orbit are regarded as target domain features; Step 9: Construct a fault classification module based on a fully connected neural network, pre-train the fault digital feature extraction module and the fault classification module in the source domain, and obtain a fault digital feature extraction module and a fault classification module with generalization ability; Specifically, selected fault data segments are labeled to train the fault digital feature extraction and fault classification modules. The fault classification corresponding to each data segment is expressed in the form of a confidence vector. The vector dimension is equal to the number of fault categories. When a data record indicates that a fault has occurred, the corresponding element in the vector is marked as 1, and others are marked as 0. If all elements in the vector are 0, it is considered that the electric propulsion system has not experienced a fault. The ion electric propulsion fault classification module is built based on a fully connected neural network. The network receives the fault digital feature vector fused in step 7 as input and outputs the fault classification vector of the current data segment. More specifically, the fault classification module is built on a fully connected neural network. The network input layer receives and fuses the fault digital feature vectors output by the time-domain feature extraction module and the frequency-domain feature extraction module. The network has four hidden layers, and the number of neurons in the output layer is consistent with the number of fault types in ion electric propulsion. The softmax function is used to map the output results to the [0, 1] interval as the confidence level for fault classification. , Where h1, h2, h3, and h4 are the output vectors of the four hidden layers; δ is the nonlinear activation function; W1, W2, W3, and W4 are the hidden layer weight matrices, and b1, b2, b3, and b4 are the corresponding biases. State is the output result, i.e., the confidence vector for different fault occurrences. In the formula, ρ i Characterizing the confidence level of different types of product failures; Step 10: In the target domain, the MMD loss function is introduced to evaluate the difference in the distribution of features extracted by the fault digital feature extraction module from on-orbit data and ground data. The weight parameters of the convolutional neural network in the fault digital feature extraction module are fine-tuned in combination with the fault classification loss to align the feature distributions of the source domain and the target domain, thereby realizing on-orbit fault digital feature extraction of ion electric propulsion based on transfer learning. Specifically, such as Figure 3 As shown, in the source domain, i.e., in the ion electric propulsion ground test data, the entire fault feature extraction and fault diagnosis network is trained using cross-entropy loss, as shown in formula (18): ; The weight parameters in the fault digital feature extraction module and the fault classification module are updated based on the loss results to complete the pre-training. During transfer learning, MMD loss is introduced to evaluate the distribution difference of features extracted by the fault digital feature extraction module from on-orbit and ground data under the same fault, in order to align the features of the source domain and the target domain, and only the weight parameters of the fault feature extraction module are updated. The MMD loss function is shown in formula (19), and the total loss function is shown in formula (20). , where f si f represents an element in the source domain feature vector. ti κ represents the elements in the target domain vector; m and n are the dimensions of the source domain feature vector and the target domain feature vector, respectively; λ is a hyperparameter that can be used to adjust the impact of MMD loss and classification loss on training.

[0024] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for extracting digital features of on-orbit faults in ion electric propulsion, characterized in that, Based on transfer learning techniques, the following steps are included: Step 1: Under the working state of the ion electric propulsion system, collect various operating parameters of the system, including operating electrical parameters, gas supply parameters and temperature parameters, and record fault information that occurs during the process. Step 2: Preprocess the working parameter sequences collected from the ground test and in-orbit operation of ion electric propulsion respectively, extract data segments with the same time length, and perform center normalization on the data within the segments; Step 3: Extract time-domain features from the preprocessed collected data. Use statistical methods to extract the statistical characteristics of each parameter and concatenate the calculation results into a one-dimensional time-domain information vector. Step 4: Construct a fault digital feature extraction module. Use a one-dimensional convolutional neural network to extract time-domain features from the time-domain information vector in Step 3 to obtain a one-dimensional time-domain feature vector. Step 5: Extract time-frequency features from the preprocessed collected data, using Morlet as the wavelet basis function, to obtain the time-frequency spectra of each parameter; Step 6: Using a two-dimensional convolutional neural network, with the time-frequency spectrum from Step 5 as input, extract time-frequency features and concatenate the time-frequency features of each parameter into a one-dimensional time-frequency feature vector; Step 7: Fuse time-domain features and frequency-domain features to obtain the fault digital feature vector; Step 8: The fault digital features extracted from the ground test of the ion electric propulsion system are regarded as source domain features, and the health status digital features extracted in orbit are regarded as target domain features; Step 9: Construct a fault classification module based on a fully connected neural network, pre-train the fault digital feature extraction module and the fault classification module in the source domain, and obtain a fault digital feature extraction module and a fault classification module with generalization ability; Step 10: In the target domain, the MMD loss function is introduced to evaluate the difference in the distribution of features extracted by the fault digital feature extraction module from on-orbit data and ground data. Then, the fault classification module is fine-tuned to align the feature distributions of the source domain and the target domain, thereby realizing on-orbit fault digital feature extraction based on transfer learning for ion electric propulsion.

2. The method for extracting digital features of on-orbit faults in ion electric propulsion according to claim 1, characterized in that, In step 1, the operating electrical parameters include the voltage and current of the ion thrust grid, the voltage and current of the acceleration grid, the voltage and current of the discharge chamber anode, and the contact voltage and current of the cathode and neutralizer; the gas supply parameters include the flow rate of each working fluid in the storage and supply unit; the temperature parameters include the temperature of the front shell of the ion thruster, the neutralizer and the pole shoes, and the temperature of each module of the power processing unit.

3. The method for extracting digital features of on-orbit faults in ion electric propulsion according to claim 2, characterized in that, In step 3, time-domain analysis is performed based on the preprocessed working electrical parameters, gas supply parameters, and temperature parameters. The maximum and minimum values, skewness, kurtosis, waveform factor, peak factor, and impulse factor of each parameter in the data segment are calculated in the time domain, and the calculation results are concatenated to form a one-dimensional vector.

4. The method for extracting digital features of on-orbit faults in ion electric propulsion according to claim 3, characterized in that, In step 5, frequency domain analysis is performed based on the preprocessed working electrical parameters, gas supply parameters, and temperature parameters. First, the working parameter signal is subjected to single-step differential processing. Then, based on the differential signal, wavelet transform is used to obtain the time-frequency spectrum of each parameter in the data segment.

5. The method for extracting digital features of on-orbit faults in ion electric propulsion according to claim 4, characterized in that, In step 9, the selected fault data segments are labeled to train the fault digital feature extraction and fault classification modules. The fault classification corresponding to each data segment is expressed in the form of a confidence vector. The vector dimension is equal to the number of fault categories. When the data record indicates that a fault has occurred, the corresponding element in the vector is recorded as 1, and the others are recorded as 0. If all elements in the vector are 0, it is considered that the electric propulsion system has not experienced a fault.

6. The method for extracting digital features of on-orbit faults in ion electric propulsion according to claim 5, characterized in that, In step 9, the ion electric propulsion fault classification module is constructed based on a fully connected neural network. The network receives the fault digital feature vector fused in step 7 as input and outputs the fault classification vector of the current data segment.

7. The method for extracting digital features of on-orbit faults in ion electric propulsion according to claim 6, characterized in that, In step 10, the MMD loss function is introduced to evaluate the distribution difference of features extracted by the fault digital feature extraction module from on-orbit data and ground data on the target domain features, and the weight parameters of the convolutional neural network in the fault digital feature extraction module are fine-tuned in combination with the fault classification loss.

Images