Spacecraft high-dimensional coupling data fault diagnosis method supporting various system configurations

Abstract

A spacecraft high-dimensional coupling data fault diagnosis method supporting system configuration diversity comprises the steps that firstly, telemetry parameters of all components of a certain spacecraft attitude and orbit control system are preprocessed; then, in combination with a stacked auto-encoder, a universal two-dimensional fault pattern in a fixed format is extracted; and finally, aiming at the two-dimensional fault pattern, realizing system-level high-precision fault diagnosis through a deep convolutional neural network. According to the method, a two-dimensional fault pattern in a fixed format is extracted from high-dimensional coupling telemetry data through a stacked auto-encoder, and fault diagnosis is carried out by using a deep convolutional neural network, so that the problem that the high-dimensional coupling telemetry data of all components of the attitude and orbit control system are difficult to consider when fault diagnosis is carried out in an existing method is solved; and the universality between different system configurations is poor.

Description

Technical Field

[0001] This invention relates to a method for fault diagnosis of high-dimensional coupled data in spacecraft that supports diverse system configurations, belonging to the field of aerospace. Background Technology

[0002] With the rapid development of spacecraft constellations in the aerospace field, limited ground stations are struggling to support timely fault diagnosis and data interpretation for hundreds or thousands of spacecraft operating in orbit. This can easily lead to missed or misdiagnosed faults on the satellites, resulting in major accidents such as service interruptions or complete satellite failures. To ensure the safe, reliable, and stable operation of large-scale spacecraft in orbit, they must possess intelligent autonomous fault diagnosis capabilities to alleviate the pressure on ground telemetry and control.

[0003] As a typical high-reliability closed-loop control system, spacecraft exhibit closed-loop propagation characteristics in their faults. Existing fault diagnosis methods typically analyze only a single component or a certain type of component, making it difficult to consider the high-dimensional coupled telemetry data of all spacecraft components. Furthermore, due to the diverse system configurations of different types of spacecraft based on their on-orbit mission requirements, the dimensionality of system telemetry parameters for different spacecraft is difficult to determine, making it impossible to adopt a unified approach for universal fault diagnosis across various spacecraft types. Summary of the Invention

[0004] The technical problem solved by this invention is to overcome the shortcomings of existing technologies and provide a fault diagnosis method for high-dimensional coupled telemetry data of spacecraft with diverse system configurations. This method extracts fixed-format two-dimensional fault patterns from high-dimensional coupled telemetry data using a stacked autoencoder and utilizes a deep convolutional neural network for fault diagnosis. This solves the problems of existing methods, such as difficulty in considering high-dimensional coupled telemetry data from all components of the attitude and orbit control system during fault diagnosis, and poor universality between different system configurations.

[0005] The technical solution of this invention is: A method for fault diagnosis of high-dimensional coupled data in spacecraft that supports diverse system configurations, comprising the following steps: (1) Collection M The original telemetry sequences of all telemetry points of the spacecraft under each fault mode. The original telemetry sequence of each telemetry point is generated by the corresponding telemetry point at time point 1, time point 2, ..., time point 3. N The collected telemetry data are arranged from earliest to latest according to the collection time; and the original telemetry sequences of all telemetry points under each fault mode are preprocessed to obtain... M All telemetry points of the spacecraft under each fault mode at time point 1, time point 2, ..., time point N Fault telemetry data; (2) Based on the stacked autoencoder, obtain the fault pattern sequence of the spacecraft in each fault mode; the fault pattern sequence includes time point 1, time point 2, ..., time point ... of the spacecraft in the corresponding fault mode. N Fault diagram; (3) Input the sequence of fault patterns of spacecraft under all fault modes into a deep convolutional neural network to complete the training of the deep convolutional neural network; (4) After processing the telemetry data of the spacecraft to be fault detected in steps (1) and (2), the data is input into the trained deep convolutional neural network. The deep convolutional neural network performs fault detection on the spacecraft and outputs the fault mode of the spacecraft.

[0006] Furthermore, the specific steps for preprocessing the original telemetry sequences of all telemetry sites under each fault mode in step (1) are as follows: (1.1) Interpolate and compensate the telemetry sequences of all telemetry sites under each fault mode to obtain the interpolated compensated telemetry sequence corresponding to each telemetry site; (1.2) Based on the Wright criterion, outlier removal is performed on all interpolated compensation data under each fault mode to obtain the outlier removal telemetry sequence corresponding to each telemetry point; (1.3) Remove outliers from all telemetry sites under each fault mode and perform mapping processing on the telemetry sequence to obtain the fault telemetry sequence for all telemetry sites; the fault telemetry sequence includes the corresponding telemetry site at time point 1, time point 2, ..., time point N Fault telemetry data.

[0007] Furthermore, in step (1.3), the mapping process involves mapping all elements in the outlier removal telemetry sequence to a preset normalized value range; the preset normalized value range is the normalized value range of the telemetry data measured when the corresponding telemetry site does not experience a fault.

[0008] Furthermore, the specific process for obtaining the spacecraft's fault pattern sequence for each fault mode in step (2) is as follows: (2.1) Input the fault telemetry data of all telemetry points in the first fault mode at time point 1 into the stacked autoencoder; (2.2) The stacked autoencoder combines the fault telemetry data of all telemetry points at time point 1 into a set of input sequences, and uses a deep encoder to extract features; after feature extraction is completed, the preset intermediate hidden layer in the stacked encoder converts the extracted features into a fault pattern in a fixed format, thereby obtaining the fault pattern of the spacecraft at time point 1 in the first fault mode; the expression of the overall process is as follows:

[0009] in, P For fault diagrams, x Given the input sequence, and These are the equivalent weights and equivalent biases of the deep encoder, respectively. For activation functions; (2.3) Obtain the spacecraft in the first fault mode at time point 2, time point 3, ..., time point N The fault diagram is obtained, and the specific process is the same as steps (2.1) to (2.2); using the spacecraft in the first fault mode at time point 1, time point 2, ..., time point N The fault patterns form the fault pattern sequence of the spacecraft in the first fault mode; (2.4) Obtain the second fault mode, the third fault mode, ..., the third fault mode. M The sequence of fault patterns of a spacecraft under each fault mode is as follows: the specific process is the same as steps (2.1) to (2.3).

[0010] Furthermore, the deep convolutional neural network used in step (3) is a two-layer CNN network; the two-layer CNN network includes an input layer, a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, and a fully connected layer.

[0011] Furthermore, during the training of the deep convolutional neural network, the input layer assembles a set of high-dimensional input data from the spacecraft fault images at each time point. The first convolutional layer uses a first set of shared weight matrices to scan the local region of each fault image in the high-dimensional input data, extracting regional features through convolution operations, and thus outputting the first set of convolutional feature maps. The first max-pooling layer reduces the size of the first set of convolutional feature maps by retaining the maximum value of the preset first pooling region, thus extracting spatial features and outputting the first set of spatial feature maps. The second convolutional layer uses a second set of shared weight matrices to scan the local region of the first set of spatial feature maps, extracting regional features through convolution operations, and thus outputting the second set of convolutional feature maps. The second max-pooling layer reduces the size of the second set of convolutional feature maps by retaining the maximum value of the preset second pooling region, thus extracting spatial features and outputting the second set of spatial feature maps. The fully connected layer is used to fuse the second set of spatial feature maps to generate a dataset for fault classification. M The deep convolutional neural network is trained by using a sequence of spacecraft failure patterns across all failure modes to generate 3D feature vectors.

[0012] Furthermore, the convolutional layer outputs feature maps through convolution operations. The calculation formula is:

[0013] in, For input to the convolutional layer, The feature map output by the convolutional layer. Let be the weight matrix of the convolution kernel. For the bias of the convolutional layer; Max pooling layers are based on the feature maps output by convolutional layers. The formula for further calculating the spatial feature map is as follows:

[0014] in, For spatial feature maps, k and j These are the output feature maps of the convolutional layer. The row and column numbers.

[0015] Furthermore, the loss function used during training of deep convolutional neural networks is the cross-entropy loss function. The calculation formula is as follows:

[0016] in, For switch signals, when the network outputs... i Categories and c When consistent, Take 1, otherwise, Set to 0; For network output i Category c The predicted probability.

[0017] The advantages of this invention compared to the prior art are: (1) This invention extracts fixed-format two-dimensional fault patterns from high-dimensional coupled telemetry data by stacking autoencoders. It can compress the data volume and reduce data storage costs while preserving the differences between high-dimensional coupled data. It also provides fixed-format fault feature data for subsequent diagnostic networks, avoids the problem of repeated construction of deep diagnostic networks, and improves the universality of diagnostic algorithms under different system configurations.

[0018] (2) This invention uses a deep convolutional neural network to automatically identify faults in fixed-format system-level fault patterns. It can fully understand the inherent fault characteristics contained in the two-dimensional fault patterns, explore deep-level coupling correlations and abnormal characteristics, achieve high-precision fault diagnosis and location for known fault modes, reduce the misdiagnosis rate and missed diagnosis rate of manual data interpretation, and improve the autonomous fault diagnosis capability of on-orbit spacecraft. Attached Figure Description

[0019] Figure 1This is a flowchart of a method for diagnosing faults in high-dimensional coupled data of spacecraft that supports diverse system configurations, according to the present invention. Figure 2 This is a network structure diagram of the stacked autoencoder used in the fault diagnosis method of the present invention; Figure 3 This is a structural diagram of the deep convolutional neural network used in the fault diagnosis method of the present invention. Detailed Implementation

[0020] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings.

[0021] like Figure 1 As shown, this invention provides a method for fault diagnosis of high-dimensional coupled data in spacecraft that supports diverse system configurations. The steps are as follows: (1) Collection M The original telemetry sequences of all telemetry points of the spacecraft under each fault mode. The original telemetry sequence of each telemetry point is generated by the corresponding telemetry point at time point 1, time point 2, ..., time point 3. N The collected telemetry data are arranged from earliest to latest according to the collection time; and the original telemetry sequences of all telemetry points under each fault mode are preprocessed to obtain... M All telemetry points of the spacecraft under each fault mode at time point 1, time point 2, ..., time point N Fault telemetry data; The specific steps for preprocessing the original telemetry sequences of all telemetry sites under each fault mode in step (1) are as follows: (1.1) Interpolate and compensate the telemetry sequences of all telemetry sites under each fault mode to obtain the interpolated compensated telemetry sequence corresponding to each telemetry site; (1.2) Based on the Wright criterion, outlier removal is performed on all interpolated compensation data under each fault mode to obtain the outlier removal telemetry sequence corresponding to each telemetry point; (1.3) Remove outliers from all telemetry sites under each fault mode and perform mapping processing on the telemetry sequence to obtain the fault telemetry sequence for all telemetry sites; the fault telemetry sequence includes the corresponding telemetry site at time point 1, time point 2, ..., time point N Fault telemetry data.

[0022] In step (1.3), the mapping process involves mapping all elements in the outlier removal telemetry sequence to a preset normalized value range; the preset normalized value range is the normalized value range of the telemetry data measured when the corresponding telemetry site does not experience a fault.

[0023] (2) Based on the stacked autoencoder, obtain the fault pattern sequence of the spacecraft in each fault mode; the fault pattern sequence includes time point 1, time point 2, ..., time point ... of the spacecraft in the corresponding fault mode. N Fault diagram; The specific process for obtaining the spacecraft's fault pattern sequence for each fault mode in step (2) is as follows: (2.1) Input the fault telemetry data of all telemetry points in the first fault mode at time point 1 into the stacked autoencoder; the network structure of the stacked autoencoder is as follows: Figure 2 As shown in Table 1, the network configuration is as follows: D Dimensions of the input data; Table 1 Network configuration for stacked autoencoders

[0024] (2.2) The stacked autoencoder combines the fault telemetry data of all telemetry points at time point 1 into a set of input sequences, and uses a deep encoder to extract features; after feature extraction is completed, the preset intermediate hidden layer in the stacked encoder converts the extracted features into a fault pattern in a fixed format, thereby obtaining the fault pattern of the spacecraft at time point 1 in the first fault mode; the expression of the overall process is as follows:

[0025] in, P For fault diagrams, x Given the input sequence, and These are the equivalent weights and equivalent biases of the deep encoder, respectively. For activation functions; (2.3) Obtain the spacecraft in the first fault mode at time point 2, time point 3, ..., time point N The fault diagram is obtained, and the specific process is the same as steps (2.1) to (2.2); using the spacecraft in the first fault mode at time point 1, time point 2, ..., time point N The fault patterns form the fault pattern sequence of the spacecraft in the first fault mode; (2.4) Obtain the second fault mode, the third fault mode, ..., the third fault mode. M The sequence of fault patterns for a spacecraft under each fault mode is as follows: The specific process is the same as steps (2.1) to (2.3). Based on the above process, this invention extracts fixed-format two-dimensional fault patterns from high-dimensional coupled telemetry data by stacking autoencoders. This can compress the data volume and reduce data storage costs while preserving the differences between high-dimensional coupled data. It also provides fixed-format fault feature data for subsequent diagnostic networks, avoiding the problem of repeated construction of deep diagnostic networks and improving the versatility of diagnostic algorithms under different system configurations.

[0026] (3) Input the sequence of fault patterns of spacecraft under all fault modes into a deep convolutional neural network to complete the training of the deep convolutional neural network; The deep convolutional neural network used in step (3) is a two-layer CNN network; the two-layer CNN network includes an input layer, a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, and a fully connected layer; the network structure of the two-layer CNN network is as follows: Figure 3 As shown in Table 2, the network configuration is as follows.

[0027] Table 2 Network configurations of deep convolutional neural networks

[0028] Furthermore, during the training of the deep convolutional neural network, the input layer assembles a set of high-dimensional input data from the spacecraft fault images at each time point. The first convolutional layer uses a first set of shared weight matrices to scan the local region of each fault image in the high-dimensional input data, extracting regional features through convolution operations, and thus outputting the first set of convolutional feature maps. The first max-pooling layer reduces the size of the first set of convolutional feature maps by retaining the maximum value of the preset first pooling region, thus extracting spatial features and outputting the first set of spatial feature maps. The second convolutional layer uses a second set of shared weight matrices to scan the local region of the first set of spatial feature maps, extracting regional features through convolution operations, and thus outputting the second set of convolutional feature maps. The second max-pooling layer reduces the size of the second set of convolutional feature maps by retaining the maximum value of the preset second pooling region, thus extracting spatial features and outputting the second set of spatial feature maps. The fully connected layer is used to fuse the second set of spatial feature maps to generate a dataset for fault classification. M The deep convolutional neural network is trained by using a sequence of spacecraft failure patterns across all failure modes to generate 3D feature vectors.

[0029] Convolutional layers output feature maps through convolution operations. The calculation formula is:

[0030] in, For input to the convolutional layer, The feature map output by the convolutional layer. Let be the weight matrix of the convolution kernel. For the bias of the convolutional layer; Max pooling layers are based on the feature maps output by convolutional layers. The formula for further calculating the spatial feature map is as follows:

[0031] in, For spatial feature maps, k and j These are the output feature maps of the convolutional layer. The row and column numbers.

[0032] The loss function used during training of deep convolutional neural networks is the cross-entropy loss function. The calculation formula is as follows:

[0033] in, For switch signals, when the network outputs... i Categories and c When consistent, Take 1, otherwise, Set to 0; For network output i Category c The predicted probability; Based on the above process, this invention uses a deep convolutional neural network to automatically identify faults in fixed-format system-level fault patterns. It can fully understand the inherent fault characteristics contained in the two-dimensional fault patterns, explore deep-level coupling correlations and abnormal characteristics, achieve high-precision fault diagnosis and location for known fault modes, reduce the misdiagnosis and missed diagnosis rates of manual data interpretation, and improve the autonomous fault diagnosis capability of on-orbit spacecraft.

[0034] (4) After processing the telemetry data of the spacecraft to be fault detected in steps (1) and (2), the data is input into the trained deep convolutional neural network. The deep convolutional neural network performs fault detection on the spacecraft and outputs the fault mode of the spacecraft.

[0035] The parts of this invention not described in detail are common knowledge to those skilled in the art.

Claims

1. A method for fault diagnosis of high-dimensional coupled data in spacecraft that supports diverse system configurations, characterized in that... The steps are as follows: (1) Collection M The original telemetry sequences of all telemetry points of the spacecraft under each fault mode. The original telemetry sequence of each telemetry point is generated by the corresponding telemetry point at time point 1, time point 2, ..., time point 3. N The collected telemetry data are arranged from earliest to latest according to the collection time; and the original telemetry sequences of all telemetry points under each fault mode are preprocessed to obtain... M All telemetry points of the spacecraft under each fault mode at time point 1, time point 2, ..., time point N Fault telemetry data; (2) Based on the stacked autoencoder, obtain the fault pattern sequence of the spacecraft in each fault mode; The fault pattern sequence includes time point 1, time point 2, ..., time point ... for the spacecraft during the corresponding fault mode. N Fault diagram; (3) Input the sequence of fault patterns of spacecraft under all fault modes into a deep convolutional neural network to complete the training of the deep convolutional neural network; (4) After processing the telemetry data of the spacecraft to be fault detected in steps (1) and (2), the data is input into the trained deep convolutional neural network. The deep convolutional neural network performs fault detection on the spacecraft and outputs the fault mode of the spacecraft.

2. The method for fault diagnosis of high-dimensional coupled data of spacecraft with diverse system configurations according to claim 1, characterized in that: The specific steps for preprocessing the original telemetry sequences of all telemetry sites under each fault mode in step (1) are as follows: (1.1) Interpolate and compensate the telemetry sequences of all telemetry sites under each fault mode to obtain the interpolated compensated telemetry sequence corresponding to each telemetry site; (1.2) Based on the Wright criterion, outlier removal is performed on all interpolated compensation data under each fault mode to obtain the outlier removal telemetry sequence corresponding to each telemetry point; (1.3) Remove outliers from all telemetry sites under each fault mode and perform mapping processing on the telemetry sequence to obtain the fault telemetry sequence for all telemetry sites; the fault telemetry sequence includes the corresponding telemetry site at time point 1, time point 2, ..., time point N Fault telemetry data.

3. The method for fault diagnosis of high-dimensional coupled data of spacecraft with diverse system configurations according to claim 1, characterized in that: In step (1.3), the mapping process involves mapping all elements in the outlier removal telemetry sequence to a preset normalized value range; the preset normalized value range is the normalized value range of the telemetry data measured when the corresponding telemetry site does not experience a fault.

4. The method for fault diagnosis of high-dimensional coupled data of spacecraft with diverse system configurations according to claim 1, characterized in that: The specific process for obtaining the spacecraft's fault pattern sequence for each fault mode in step (2) is as follows: (2.1) Input the fault telemetry data of all telemetry points in the first fault mode at time point 1 into the stacked autoencoder; (2.2) The stacked autoencoder combines the fault telemetry data of all telemetry points at time point 1 into a set of input sequences and uses a deep encoder for feature extraction. feature After extraction, the preset intermediate hidden layer in the stacked encoder transforms the extracted features into a fault pattern with a fixed format, thus obtaining the fault pattern of the spacecraft at time point 1 in the first fault mode; the expression for the overall process is as follows: in, P For fault diagrams, x Given the input sequence, and These are the equivalent weights and equivalent biases of the deep encoder, respectively. For activation functions; (2.3) Obtain the spacecraft in the first fault mode at time point 2, time point 3, ..., time point N The fault diagram is obtained, and the specific process is the same as steps (2.1) to (2.2); using the spacecraft in the first fault mode at time point 1, time point 2, ..., time point N The fault patterns form the fault pattern sequence of the spacecraft in the first fault mode; (2.4) Obtain the second fault mode, the third fault mode, ..., the third fault mode. M The sequence of fault patterns of a spacecraft under each fault mode is as follows: the specific process is the same as steps (2.1) to (2.3).

5. The method for fault diagnosis of high-dimensional coupled data of spacecraft with diverse system configurations according to claim 1, characterized in that: The deep convolutional neural network used in step (3) is a two-layer CNN network; the two-layer CNN network includes an input layer, a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, and a fully connected layer.

6. The method for fault diagnosis of high-dimensional coupled data of spacecraft with diverse system configurations according to claim 5, characterized in that: During deep convolutional neural network training, the input layer assembles a set of high-dimensional input data from spacecraft fault images at each time point. The first convolutional layer uses a first set of shared weight matrices to scan local regions of each fault image in the high-dimensional input data, extracting region features through convolution operations, and outputting a first set of convolutional feature maps. The first max-pooling layer reduces the size of the first set of convolutional feature maps by retaining the maximum value of a preset first pooling region, thus extracting spatial features and outputting a first set of spatial feature maps. The second convolutional layer uses a second set of shared weight matrices to scan local regions of the first set of spatial feature maps, extracting region features through convolution operations, and outputting a second set of convolutional feature maps. The second max-pooling layer reduces the size of the second set of convolutional feature maps by retaining the maximum value of a preset second pooling region, thus extracting spatial features and outputting a second set of spatial feature maps. A fully connected layer is used to fuse the second set of spatial feature maps to generate a dataset for fault classification. M The deep convolutional neural network is trained by using a sequence of spacecraft failure patterns across all failure modes to generate 3D feature vectors.

7. The method for fault diagnosis of high-dimensional coupled data of spacecraft supporting diverse system configurations as described in claim 6, characterized in that: Convolutional layers output feature maps through convolution operations. The calculation formula is: in, For input to the convolutional layer, The feature map output by the convolutional layer. The weight matrix of the convolution kernel. For the bias of the convolutional layer; Max pooling layers are based on the feature maps output by convolutional layers. The formula for further calculating the spatial feature map is as follows: in, For spatial feature maps, k and j These are the output feature maps of the convolutional layer. The row and column numbers.

8. A method for fault diagnosis of high-dimensional coupled data of spacecraft with diverse system configurations according to any one of claims 1-7, characterized in that: The loss function used during training of deep convolutional neural networks is the cross-entropy loss function. The calculation formula is as follows: in, For switch signals, when the network outputs... i Categories and c When consistent, Take 1, otherwise, Set to 0; For network output i Category c The predicted probability.

Images