A spacecraft telemetry timing data generation method, device, medium and product

Abstract

The application discloses a spacecraft telemetry time series data generation method and device, medium and product, and relates to the field of telemetry time series data generation. The method comprises the following steps: inputting a spacecraft telemetry time series training data set into a spacecraft time series data generation model for training; the spacecraft time series data generation model comprises a data embedding and reconstruction submodel and a diffusion transformer submodel; the data embedding and reconstruction submodel is used for introducing a group of preset learnable vectors into the spacecraft telemetry time series training data and then mapping the spacecraft telemetry time series training data into a low-dimensional latent space; the diffusion transformer submodel is used for realizing data reconstruction generation; random noise is input into the trained spacecraft time series data generation model, low-dimensional data is generated through an inverse process of the diffusion transformer submodel, and the low-dimensional data is restored into new spacecraft telemetry time series data through the data embedding and reconstruction submodel. The application can quickly generate high-quality spacecraft telemetry time series data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of telemetry time series data generation, and particularly relates to a spacecraft telemetry time series data generation method and device, medium and product. BACKGROUND

[0002] In the field of aerospace, due to the consideration of data acquisition cost and other factors, only a small amount of sample data can be obtained, which is particularly prominent in the task of predicting the remaining service life of a spacecraft. For some high-precision key components of a spacecraft, full-life experiments not only require high costs, but also these devices often have very high reliability, and it takes a long time to obtain their full-life cycle data. At the same time, the data acquisition has very strict requirements on the performance of the sensor. The above factors limit the preparation of the remaining service life data set of high-precision key components of a spacecraft. At the same time, due to the consideration of equipment secrecy, the equipment manufacturer will only provide a part of the data to the researchers, which results in that the size of the available data set for time series data modeling is very small. When a deep learning model that performs well on other large data sets is applied to such a small data set, underfitting often occurs, and the performance is very poor. Therefore, in order to use the deep learning method on a smaller data set, data generation is a common means.

[0003] Time series data generation is a process of generating new time series data from existing time series data. Unlike other data generation, the newly generated time series data not only has the same distribution as the original data, but also has a long-time dependence relationship between the generated data. This requires the model to capture the distribution of features at each time point and the way the distribution changes over time. For multi-dimensional time series data, the relationship between dimensions is also a factor worth considering. Therefore, compared with other data type generation, time series data generation has certain challenges.

[0004] Most of the current popular generation models are implemented based on generative adversarial networks and variational autoencoders. Both of them perform very well in the field of image generation, and many works have applied them to time series data generation and achieved good results. However, there are still some problems in the time series data generation models based on the two. Due to the unique structure of the generative adversarial network, the problem of gradient disappearance and mode collapse will occur in the training process, which makes it difficult to train the generative model based on the generative adversarial network. And usually a high-quality generative model based on generative adversarial network contains a large number of parameters. Using it to generate multi-dimensional time series data requires a lot of time, and such a model is only suitable for generating some short sequence data and cannot solve the problem of long sequence generation. The time series generation model based on variational autoencoder has a shorter time to generate new sequences due to its encoder-decoder structure, but it will over-smooth the data distribution during the training process, resulting in blurred data samples, so it is difficult to generate high-quality data. SUMMARY

[0005] The purpose of the present application is to provide a spacecraft telemetry time series data generation method, device, medium and product, which can quickly generate high-quality spacecraft telemetry time series data.

[0006] To achieve the above purpose, the present application provides the following scheme:

[0007] In a first aspect, the present application provides a spacecraft telemetry time series data generation method, comprising:

[0008] Obtaining a set of spacecraft telemetry time series training data;

[0009] Inputting the set of spacecraft telemetry time series training data into a spacecraft time series data generation model for training; the spacecraft time series data generation model comprises a data embedding and reconstruction sub-model and a diffusion transformer sub-model; the data embedding and reconstruction sub-model is used to introduce a set of preset learnable vectors to the spacecraft telemetry time series training data and then map them to a low-dimensional latent space; the diffusion transformer sub-model is used to add noise to the data in the low-dimensional latent space and perform inverse diffusion denoising to realize data reconstruction generation; the preset learnable vector is used to simulate the characteristics of the telemetry time series data collected by the sensor;

[0010] Sampling a random noise, inputting the random noise into the trained spacecraft time series data generation model, generating low-dimensional data through the inverse process of the diffusion transformer sub-model, and then restoring the low-dimensional data to new spacecraft telemetry time series data through the data embedding and reconstruction sub-model.

[0011] In a second aspect, the present application provides a computer device, comprising: a memory, a processor to store a computer program on the memory and executable on the processor, and the processor executes the computer program to implement the steps of the method.

[0012] In a third aspect, the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the method.

[0013] In a fourth aspect, the present application provides a computer program product, comprising a computer program, and the computer program is executed by a processor to implement the steps of the method.

[0014] According to the specific embodiments of the present application, the following technical effects are disclosed.

[0015] The space time series data generation model is composed of a data embedding and reconstruction sub-model and a diffusion transformer sub-model. The data embedding and reconstruction sub-model is used to map the spacecraft telemetry time series training data into a low-dimensional latent space. The diffusion transformer sub-model is used to add noise to the data in the low-dimensional latent space and perform inverse diffusion denoising to realize data reconstruction generation. The data embedding and reconstruction sub-model solves the problem that the existing time series data generation model cannot quickly generate long time series data. The low-dimensional representation of the new data is generated through the diffusion transformer sub-model, and then the spacecraft telemetry time series new data is restored through the data embedding and reconstruction sub-model. By introducing a preset learnable vector, the generated time series data has the structural characteristics of the original data, thereby improving the quality of the generated data. The present application realizes the deep learning of long time series data generation. Based on this, researchers can quickly construct a spacecraft key component telemetry time series data set according to a small amount of spacecraft telemetry data, and then carry out many meaningful deep research work. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0017] Figure 1 The flowchart of the spacecraft telemetry time series data generation method of the present application is shown in the figure.

[0018] Figure 2 The overall flowchart of the method in the present application example is shown in the figure.

[0019] Figure 3 The overall architecture of the space timing data generation model of the present application;

[0020] Figure 4 The structural diagram of the variational autoencoder;

[0021] Figure 5 The structural diagram of the diffusion transformer sub-model;

[0022] Figure 6 The internal structure diagram of the computer device. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] The purpose of the present application is to provide a spacecraft telemetry timing data generation method, device, medium and product. On the basis of using the spacecraft telemetry data received by the ground, the spacecraft key component telemetry data set is expanded by processing and modeling, and the spacecraft key component telemetry time series data set can be quickly constructed according to a small amount of spacecraft telemetry data, and then many meaningful deep research work can be carried out.

[0025] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0026] As shown in Figure 1 and Figure 2 The present application provides a spacecraft telemetry timing data generation method, comprising:

[0027] Step 100, acquiring a spacecraft telemetry timing training data set.

[0028] Step 100, specifically comprising:

[0029] (1) acquiring spacecraft real telemetry time series data collected by a sensor.

[0030] (2) using K- nearest neighbor method to fill data for the spacecraft real telemetry time series data; specifically, in the acquisition process, some problems in the acquisition and transmission process will cause data loss. In view of this problem of data loss, K- nearest neighbor method is used for filling, and further, the average value of k nearest neighbors in the data set is used for filling.

[0031] (3) Linear normalization is performed on the filled spacecraft real telemetry time series data to obtain homodimensional telemetry time series data; specifically, the dimensions of each data in the spacecraft real telemetry time series data are quite different, and linear normalization is used to process the original data to scale all dimensions of the data to one dimension, facilitating subsequent processing.

[0032] Specifically, X is set to be norm for the normalized data, X is the original data, X max and X min are the maximum and minimum values of the original data set, respectively, and the calculation formula of normalization is as follows:

[0033]

[0034] The original data is converted to data within a certain range through the normalization operation, thereby eliminating the influence of the data dimension.

[0035] (4) The time window method is used to divide the homodimensional telemetry time series data to obtain a plurality of telemetry time period data; any of the telemetry time period data is taken as a spacecraft telemetry time sequence training data to obtain a spacecraft telemetry time sequence training data set. Specifically, the homodimensional telemetry time series data is divided into a series of window data according to a specified window size, i.e., a plurality of telemetry time period data are obtained. The determination of the specified window size and time step is determined according to the specific data situation.

[0036] Step 200, inputting the spacecraft telemetry time sequence training data set into a space time sequence data generation model for training; the space time sequence data generation model includes a data embedding and reconstruction sub-model and a diffusion transformer sub-model; the data embedding and reconstruction sub-model is used to introduce a group of preset learnable vectors (corresponding to sensor embedding in Figure 2 into the spacecraft telemetry time sequence training data (corresponding to data embedding in Figure 2 ), and then mapped into a low-dimensional latent space; the diffusion transformer sub-model is used to add noise (corresponding to sampling random noise in Figure 2 ) to the data in the low-dimensional latent space, and to inverse diffusion denoising (corresponding to noise inverse diffusion in Figure 2 ), so as to realize data reconstruction generation; the preset learnable vector is used to simulate the characteristics of the telemetry time sequence data collected by the sensor.

[0037] Specifically, a multi-dimensional time sequence data generation model is designed according to the characteristics of the spacecraft telemetry time sequence data. For example, Figure 3As shown, the data embedding and reconstruction sub-model includes a variational autoencoder and a sensor embedding module.

[0038] The variational autoencoder includes an encoder, a low-dimensional latent space (i.e., hidden space) and a decoder arranged in sequence, and the structure is as shown in the figure. Figure 4 The variational autoencoder is used to map the original telemetry time series data into a low-dimensional, canonical latent space, sample on the space, generate a low-dimensional representation of new time series data using the inverse diffusion process of DDPM (Denoising Diffusion Probabilistic Models), and then restore it to real time series data, thereby realizing embedding and reconstruction of data. Therefore, in the training process of this sub-model, when measuring the data reconstruction loss and the KL divergence loss of the latent space in the loss function, since the sub-model is mainly used for data dimension reduction, the reconstruction loss needs to account for a larger proportion in the total loss.

[0039] The sensor embedding module is used to introduce a set of preset learnable vectors, which are superimposed with the spacecraft telemetry time series training data and input to the encoder. Specifically, a learnable vector simulation sensor, i.e., a preset learnable vector, is created according to the dimension of the input data set. The value of the preset learnable vector is superimposed with the value of the original data, and then input to the variational autoencoder. The sensor embedding module is trained at the same time as the variational autoencoder is trained, with the purpose of simulating the characteristics of the data acquisition sensor, so that the generated new data has similar structural characteristics as the original data, improving the quality of the generated data. In other words, the present application simulates the sensor through a learnable embedding vector, models the structural information of high-dimensional time series data, adds significant structural features to the input data, and thus improves the quality of the generated data.

[0040] Figure 3 In the diffusion process of the DDPM, z0, z1, z2 represent the changes of data in the diffusion process, from a complete data z0 to a random noise z n Its inverse diffusion process is from a random noise z n to a complete data z0.

[0041] In order to make the generated time series data contain significant time series information, the present application designs a new diffusion transformer structure to replace the U-Net in the prior art to perform partial parameter evaluation work, i.e., a diffusion transformer sub-model is obtained, and the structure diagram is as shown in the figure. Figure 5Specifically, during the parameter estimation part of the diffusion model training process, parameter estimation is performed by adding an attention prior diffusion transformer. The structure of the diffusion transformer submodel is similar to that of the traditional transformer model, but in order to adapt to the processing task of time series data, global time hard coding is added on the basis of the original position coding. The processing method of global time hard coding is to encode the timestamp information of time series data. Generally speaking, the timestamp information of time series data is a string, which needs to be split and decomposed into year, month, day, hour, minute and second existing forms respectively, and the time information of each unit is encoded, and finally unified into a whole coding. And in order to prevent overfitting from occurring in small data sets, a convolution operation is added to the self-attention mechanism to enable it to obtain the local structure of the data, that is, a Gaussian prior is added to the attention mechanism of the transformer.

[0042] The training process of the aerospace time series data generation model is divided into two stages, so there are two stages of loss function. The training of the data embedding and reconstruction submodel is the first training stage of the model, and its loss is the loss function of VAE (Variational Auto-Encoders), which is divided into two parts. One part is the reconstruction loss, which is calculated by the distance between the original data and the reconstructed data after the decoder; the other part is the KL divergence between the latent space and the standard Gaussian distribution. The training of the diffusion transformer submodel is the second training stage of the model, and its loss is obtained by calculating the distance between the noise ε added in the training stage and the noise ε estimated by the diffusion transformer θ Due to the use of the time window import method for data processing, the model can be trained in small batches, which can speed up the training speed.

[0043] The training process of the aerospace time series data generation model, specifically includes:

[0044] 1) For the data embedding and reconstruction submodel, the back propagation algorithm and the Adam algorithm are used to minimize the embedding reconstruction loss function until the value of the embedding reconstruction loss function meets the first preset requirement, and the training is stopped; the embedding reconstruction loss function is:

[0045] loss r =α×loss c +β×loss kl .

[0046]

[0047]

[0048] wherein loss r represents the value of the embedding reconstruction loss function, loss c represents the value of the reconstruction error, loss kl represents the KL divergence between the low-dimensional latent space and the standard Gaussian distribution, and a and b represent weight coefficients for measuring the proportion of loss c and loss kl in the loss function, and the values are between 0 and 1; m represents the total number of data dimensions, x i represents the i-th dimension of the data x, x' i represents the i-th dimension of the reconstructed data x' generated by the data embedding and reconstruction submodel, n represents the degree of freedom, and s i represents the standard deviation of the i-th dimension of the data predicted, and m i represents the expectation of the i-th dimension of the data predicted by the model.

[0049] 2) For the diffusion transformer submodel, a noise loss function is minimized until the value of the noise loss function meets a second preset requirement, and the training is stopped. Specifically, during the training process, part of the parameters of the data embedding and reconstruction submodel are frozen, only the network parameters involved in the data generation module are trained, and the loss is ended when it tends to be stable. The noise loss function is:

[0050] loss g = ||e - e θ || 2 .

[0051] wherein loss g represents the noise loss function, e represents the noise added in the training stage, and e θ represents the noise estimated by the diffusion transformer submodel.

[0052] Step 300: A random noise is sampled from a standard Gaussian distribution, and the random noise is input into the trained space-time data generation model. The low-dimensional data is generated through the inverse process of the diffusion transformer submodel, and then the low-dimensional data is restored to new spacecraft telemetry time series data through the data embedding and reconstruction submodel.

[0053] The application is adapted to the task of quickly generating high-dimensional time series data, and maps high-dimensional data into a lower-dimensional canonical latent space through an encoder and a decoder structure, and simulates sensors through a set of learnable embedding vectors to capture the structural features of the original data; then sampling in the latent space, using the inverse diffusion process of DDPM to generate a low-dimensional representation of new time series data, and then restoring it to real time series data. That is, the application designs a spacecraft time series data generation method based on a diffusion model, which generates new data according to a small amount of real spacecraft telemetry data.

[0054] The application solves the problem of generating spacecraft key component telemetry time series data, and has the characteristics of fast data generation and good data quality. Researchers can expand the original spacecraft telemetry time series data set through the method, and then conduct research on other tasks. After expansion, the application can also be used for expansion of time series data sets generated by other similar complex systems.

[0055] In one embodiment, a computer device, which can be a database, has an internal structure diagram as shown in Figure 6 The computer device includes a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The database of the computer device is used to store transactions to be processed. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement a data processing method.

[0056] In one embodiment, a computer device is also provided, which includes a memory, a processor to store a computer program on the memory and run the computer program on the processor, and the processor executes the computer program to implement the steps in the above method embodiments.

[0057] In one embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by the processor to implement the steps in the above method embodiments.

[0058] In an embodiment, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the steps of any of the above method embodiments.

[0059] It should be noted that the object information (including but not limited to object device information, object personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0060] It can be understood by those skilled in the art that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, database or other medium used in the embodiments provided by the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in the embodiments provided by the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided by the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0061] Any technical features in the above embodiments can be combined, and for the sake of brevity, not all possible combinations are described above, however, it should be understood that the scope of the specification includes all possible combinations.

[0062] The principles and implementations of the present application are described in the above embodiments, which are only used to help understand the method and the core idea of the present application; meanwhile, for those skilled in the art, the specific implementation and application range can be changed according to the idea of the present application. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for generating spacecraft telemetry timing data, characterized in that, The methods include: Acquire spacecraft telemetry timing training data set; The spacecraft telemetry timing training data set is input into the spacecraft timing data generation model for training; The spacecraft time-series data generation model includes a data embedding and reconstruction sub-model and a diffusion transformer sub-model. The data embedding and reconstruction sub-model is used to introduce a set of preset learnable vectors into the spacecraft telemetry time-series training data and then map them into a low-dimensional latent space. The diffusion transformer sub-model is used to add noise and perform reverse diffusion denoising on the data in the low-dimensional latent space to achieve data reconstruction generation. The preset learnable vectors are used to simulate the characteristics of telemetry time-series data collected by sensors. The data embedding and reconstruction sub-model includes a variational autoencoder and a sensor embedding module; the variational autoencoder includes an encoder, a low-dimensional latent space and a decoder arranged in sequence; the sensor embedding module is used to introduce a set of preset learnable vectors, which are superimposed with spacecraft telemetry time series training data and then input to the encoder. The diffusion transformer sub-model adds global time hard coding to the original position coding of the transformer model, which encodes the timestamp information of the time series data; A random noise is sampled and input into the trained space time series data generation model. Low-dimensional data is generated through the inverse process of the diffusion transformer sub-model. Then, the low-dimensional data is restored into new spacecraft telemetry time series data through the data embedding and reconstruction sub-model.

2. The spacecraft telemetry timing data generation method according to claim 1, characterized in that, Obtain the spacecraft telemetry time series training data set, specifically including: Acquire real telemetry time series data of the spacecraft collected by sensors; The K-nearest neighbor method was used to fill in the actual telemetry time series data of the spacecraft; Linear normalization was performed on the actual telemetry time series data of the spacecraft after data filling to obtain telemetry time series data with the same dimensions. The time window method is used to divide the telemetry time series data of the same dimension into multiple telemetry time periods; any one of the telemetry time periods is used as a spacecraft telemetry time series training data to obtain a spacecraft telemetry time series training data set.

3. The spacecraft telemetry timing data generation method according to claim 1, characterized in that, The training process of the space time-series data generation model specifically includes: For the data embedding and reconstruction sub-model, the backpropagation algorithm and the Adam algorithm are used to minimize the embedding and reconstruction loss function until the value of the embedding and reconstruction loss function meets the first preset requirement and training stops. For the diffusion transformer sub-model, the noise loss function is minimized until the value of the noise loss function meets the second preset requirement, at which point training stops.

4. The spacecraft telemetry timing data generation method according to claim 3, characterized in that, The embedding reconstruction loss function is: ； ； ； in, This represents the value of the embedding reconstruction loss function. This represents the value of the reconstruction error. This represents the KL divergence between the low-dimensional latent space and the standard Gaussian distribution. , Indicates the weighting coefficient. m This represents the total number of data dimensions. This represents the i-th dimension of data x. This represents the reconstructed data generated by the data embedding and reconstructed sub-model. The i-th dimension, where n represents the degrees of freedom. This represents the standard deviation of the data in the i-th dimension obtained from the prediction. This represents the expected value of the data in the i-th dimension predicted by the model.

5. The spacecraft telemetry timing data generation method according to claim 3, characterized in that, The noise loss function is: ； in, Represents the noise loss function. This indicates the noise added during the training phase. This represents the noise estimated by the diffuse transformer submodel.

6. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the method according to any one of claims 1-5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1-5.

8. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1-5.

Images