Mars sand and dust activity prediction method, device, equipment and medium

Abstract

The invention relates to a Mars dust activity prediction method and device, equipment and a medium. A training set is obtained; constructing a sand and dust activity prediction pre-training model, wherein the sand and dust activity prediction pre-training model comprises an encoder and a decoder; the encoder comprises a plurality of ConvGRU layers of which the depths are deepened layer by layer, the output of the previous ConvGRU layer is the input of the next ConvGRU layer, and finally, the encoder outputs a middle hidden layer; the decoder comprises a plurality of ConvGRU layers of which the depths are reduced layer by layer, and the middle hidden layer is used as the input of each ConvGRU layer in the decoder; the number of ConvGRU layers of the decoder is less than that of ConvGRU layers of the encoder; training the sand and dust activity prediction pre-training model through the training set to obtain a trained sand and dust activity prediction model; and predicting future Mars sand and dust through the trained sand and dust activity prediction model. The prediction model provided by the invention is simple in structure and high in calculation and reasoning speed.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, device, equipment and medium for predicting dust activity on Mars. Background Art

[0002] As the most important atmospheric phenomenon on Mars, dust storms affect atmospheric circulation through radiation heating and redistribution of atmospheric circulation. The hazards of dust storms to Mars exploration include reducing the solar radiation received by the lander, reducing the power generation efficiency of solar panels, and reducing the quality of Mars atmospheric images. In addition, dust storms will change the temperature and dynamics of the Martian atmosphere by absorbing solar radiation. Dust in the Martian atmosphere scatters, absorbs and reradiates solar radiation, thus affecting the thermal structure and large-scale atmospheric circulation of the Martian atmosphere. During global-scale dust storms, the temperature of the middle atmosphere can rise by more than 40K compared with the same season without a lot of dust, and the surface temperature and air pressure will also be strongly affected. Dust storms also contribute to atmospheric tides and planetary wave excitation. The annual variation of Martian dust has certain regularities, but the daily variation of dust is more complicated. Due to the harmful effects of Martian dust on Mars exploration and its profound impact on the Martian atmosphere, it is of great significance to understand the activity of Martian dust.

[0003] However, traditional statistical analysis cannot accurately capture the changing patterns of dust, and therefore cannot predict multi-scale dust activities (from dust storms to global dust storms). Dust storms on Mars are divided into local dust storms, regional dust storms, and global dust storms. Local and regional dust storms can occur and be observed throughout the year, and in some years, multiple regional dust storms may break out and merge into global dust storms. On the other hand, global dust storms are concentrated near the perihelion of Mars, when fast-moving dust in the atmosphere absorbs sunlight and is heated, causing bulges. Large-scale, long-term dust storms have an important impact on atmospheric visibility, thermal structure, and circulation. Shirley considered variables related to the orbital motion of Mars and predicted the next global dust storm on Mars. Montabon et al. discussed the results based on knowledge of the dynamic functions of the Martian atmosphere and observations of past dust activities, but none of them could accurately predict the occurrence of dust storms. Due to the lack of observational data, predictions of short-term dust activity on Mars are even more lackluster, resulting in a lack of in-depth understanding of dust lifting mechanisms, radiation / dynamic feedback, and dust interactions. This scarcity makes it difficult to predict dust activity through traditional meteorological means. Summary of the invention

[0004] Based on this, it is necessary to provide a Mars dust activity prediction method, device, equipment and medium with a simple structure, fast calculation speed and the ability to effectively predict dust activities in response to the above technical problems.

[0005] A method for predicting dust activity on Mars, the method comprising:

[0006] Obtaining a training set, wherein the training set is two-dimensional time series data;

[0007] Constructing a dust activity prediction pre-training model, the dust activity prediction pre-training model includes an encoder and a decoder; the encoder includes a plurality of ConvGRU layers with increasing depth, and the output of the previous ConvGRU layer is the input of the next ConvGRU layer, and finally the encoder outputs an intermediate hidden layer; the decoder includes a plurality of ConvGRU layers with decreasing depth, and the intermediate hidden layer is used as the input of each ConvGRU layer in the decoder; the number of ConvGRU layers of the decoder is less than the number of ConvGRU layers of the encoder;

[0008] Inputting the two-dimensional time series data into each ConvGRU layer in the encoder in sequence, training the dust activity prediction pre-training model, and obtaining a trained dust activity prediction model;

[0009] The future dust on Mars is predicted by the trained dust activity prediction model.

[0010] In one of the embodiments, the training set includes a period with global sandstorms and a period without global sandstorms; the data ratio of the period with global sandstorms and the period without global sandstorms is 1:1.

[0011] In one of the embodiments, the training set is preprocessed;

[0012] Dividing the data of the period with global sandstorm and the period without global sandstorm into physical elements;

[0013] The physical factors include: air pressure, temperature, wind speed, wind direction and sky optical thickness;

[0014] The air pressure, the temperature, the wind speed, the wind direction and the sky optical thickness are divided into input dimensions.

[0015] In one embodiment, the input dimensions include predicted time series length, batch size, color channel, length and width.

[0016] In one embodiment, each ConvGRU layer of the encoder has an input channel corresponding to the input dimension.

[0017] In one embodiment, the encoder includes 4 ConvGRU layers, and the decoder includes 3 ConvGRU layers.

[0018] In one embodiment, the hidden state of the decoder at the prediction time is expressed as:

[0019] H t =φ(H t-1 ,y t-1 ,C);

[0020] In the formula, φ represents the nonlinear activation function; H t-1 represents the previous hidden state; y t-1 represents the previous time series; C represents the middle hidden layer.

[0021] A device for predicting dust activity on Mars, comprising:

[0022] A data acquisition module is used to acquire a training set, wherein the training set is two-dimensional time series data;

[0023] A model building module is used to build a pre-trained model for predicting dust activity, wherein the pre-trained model for predicting dust activity includes an encoder and a decoder; the encoder includes a plurality of ConvGRU layers, and the output of the previous ConvGRU layer is the input of the next ConvGRU layer, and finally the encoder outputs an intermediate hidden layer; the decoder includes a plurality of ConvGRU layers, and the intermediate hidden layer is used as the input of each ConvGRU layer in the decoder;

[0024] A pre-training module, used for sequentially inputting the two-dimensional time series data into each ConvGRU layer in the encoder, training the dust activity prediction pre-training model, and obtaining a trained dust activity prediction model;

[0025] The prediction module is used to predict the future Martian dust by using the trained dust activity prediction model.

[0026] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0027] A computer-readable storage medium stores a computer program, which implements the steps of any of the above methods when executed by a processor.

[0028] The above-mentioned Mars dust activity prediction method, device, equipment and medium obtain a training set, which is two-dimensional time series data; construct a dust activity prediction pre-training model, which includes an encoder and a decoder; the encoder includes a plurality of ConvGRU layers with increasing depth, and the output of the previous ConvGRU layer is the input of the next ConvGRU layer, and finally the encoder outputs an intermediate hidden layer; the decoder includes a plurality of ConvGRU layers with decreasing depth, and the intermediate hidden layer is used as the input of each ConvGRU layer in the decoder; the number of ConvGRU layers of the decoder is less than the number of ConvGRU layers of the encoder; the two-dimensional time series data is sequentially input into each ConvGRU layer in the encoder, the dust activity prediction pre-training model is trained, and a trained dust activity prediction model is obtained; and the future Mars dust is predicted by the trained dust activity prediction model.

[0029] The present invention constructs an encoder and decoder framework, and sets a number of ConvGRU layers with increasing depth layer by layer in the encoder to extract key features in the time series; sets a number of ConvGRU layers with decreasing depth layer by layer in the decoder to generate predictions of future dust, and the number of ConvGRU layers of the decoder is less than the number of ConvGRU layers of the encoder, because the decoder usually needs to generate an output smaller than the input matrix, so it may not need as many layers as the encoder. Since the task of the decoder is usually to generate relatively simple outputs, it can have fewer levels, thereby reducing the risk of overfitting. A deeper encoder can help capture the complex features of the input image, while a shallower decoder can map these features to the desired output. Through such a setting, the computational complexity of the model can be reduced and the speed of training and reasoning can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A schematic diagram of a flow chart of a method for predicting dust activity on Mars in one embodiment;

[0031] Figure 2 A schematic diagram of the structure of a dust activity prediction model provided by the present invention in one embodiment;

[0032] Figure 3 A schematic diagram of the influence of different physical element combinations on the accuracy of a dust activity prediction model in one embodiment;

[0033] Figure 4 Schematic diagram of average error variation curves of different models in one embodiment at 12 hours on Mars after the start of forecasting; wherein, Figure 4 The left is a schematic diagram of the MSE error change curve. Figure 4 The middle is a schematic diagram of the MAE error change curve. Figure 4Right is R 2 Schematic diagram of error change curve;

[0034] Figure 5 is the spatial error distribution of different models at the 2nd, 6th and 12th Mars local time after the start of the forecast in one embodiment, and the average value 12 hours after the start of the forecast; wherein, Figure 5 (a) is a schematic diagram of MSE error distribution. Figure 5 (b) is a schematic diagram of MAE error distribution. Figure 5 (c) is a schematic diagram of BIAS error distribution;

[0035] Figure 6 This is a schematic diagram of the results of predicting the global sandstorm on Mars using the sand and dust activity prediction model provided by the present invention in one embodiment; wherein: Figure 6 (a) The actual and predicted spatial distribution of rapidly changing local dust storms during the global dust storm period on Mars at 2, 6, and 12 Mars local time after the reporting time. Figure 6 (b) is the curve showing the MAPE of Dust A and Dust B in a rapidly changing local dust storm as the prediction step length increases;

[0036] Figure 7 A schematic diagram of a structure of a device for predicting dust activity on Mars in one embodiment;

[0037] Figure 8 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0039] In the process of implementing the technology of this solution, the inventor found that it is difficult for the existing technology to accurately and quickly predict the dust activity on Mars. Based on this, the inventor found that deep learning has the ability to respond quickly and improve itself, and it is also expected to solve the communication limitations and human risk problems in interstellar exploration. Since space missions are usually carried out remotely, long-distance communication will bring two obstacles: the communication time between the spacecraft and the earth is too long, resulting in a waste of time and productivity; the information transmission between the spacecraft and the earth will also discard data due to long-distance bandwidth limitations. Therefore, fast-response robots based on artificial intelligence have advantages in space exploration. In addition, the development of Mars exploration is very fast, and the total amount of data sets and the growth rate of data volume are also increasing. It is necessary to make full use of the self-improvement ability of deep learning and make full use of this data. In addition, analyzing Mars data through deep learning has the potential to provide deeper insights in many fields such as climate, atmosphere, and possible future living environment on Mars. Deep learning is a field of artificial intelligence that uses collected data to learn or improve its performance. Deep learning algorithms are able to capture, analyze and describe complex data patterns, which makes them a valuable tool for studying various phenomena in many industrial and scientific fields. Deep learning technology has also been successfully applied to the analysis of spacecraft and space acquisition data.

[0040] Therefore, the inventors applied deep learning to the atmospheric field of the planet to achieve more autonomous and comprehensive use of meteorological data collected and analyzed by rovers such as orbiters and landers.

[0041] Since dust storms usually follow specific trajectories and show coherent sequence changes, the inventors constructed a neural network model based on an encoder-decoder architecture, in which a ConvGRU layer is set, so that the constructed neural network model has excellent sequence modeling capabilities, and the local characteristics of the evolution of dust storms in the Martian atmosphere can be better captured through the convolution layer.

[0042] The following will describe the implementation of the present invention in detail with reference to the accompanying drawings in the embodiment diagram of the present invention.

[0043] In one embodiment, Figure 1 As shown, a method for predicting dust activity on Mars is provided, comprising the following steps:

[0044] Step 102, obtaining a training set, where the training set is two-dimensional time series data.

[0045] Specifically, a data set including a training set, a test set, and a validation set is obtained. This embodiment uses the column dust optical depths (DOD) in the EnsembleMars Atmosphere Reanalysis System (EMARS) dataset version 1.0 to conduct a study on the global dust prediction on Mars. DOD is the same as the optical thickness of the air column, which is equivalent to the direct imaging method of visible light. The typical global average visible wavelength columnar dust optical depth is about 0.3-0.5, which can often be observed in the spring and summer of the Martian year, when there is less atmospheric dust. When a global dust storm occurs, the global average visible wavelength columnar dust optical depth reaches more than 1.5-2, and can even reach more than 7. The DOD in EMARS covers the world, with a latitude of -90° to 90° and a longitude of 0° to 360°. It is divided into 36 and 60 grids in the longitude and latitude directions, respectively, so its spatial resolution is 5°×6°; while the time span spans 24 to 33 years in the Martian year, and its time resolution is 1 hour in the local time of Mars. The use of DOD in EMARS for global dust analysis and prediction research on Mars can be extended to further research on visible light imaging images and has a certain degree of versatility.

[0046] In one embodiment, the data set includes periods with global sandstorms and periods without global sandstorms, and the data ratio of the periods with global sandstorms and the periods without global sandstorms is 1:1.

[0047] It can be understood that in Mars years 25 (MY25) and MY28, global dust storm events occurred in the second half of the northern hemisphere (180°-360°Ls). To ensure the integrity of the dataset, the dataset was selected based on the years with the strongest temporal continuity in the EMARS data, and included data when global dust storm events occurred and when global dust storm events did not occur. The selected datasets without global dust storm events include MY29, 30, and 31. MY25 and MY29 were used as training sets. The second half of MY28 was randomly divided into two parts, forming the test set and validation set with MY30 and MY31 respectively. The dataset was row-normalized and divided into training set, validation set, and test set in a ratio of 8:1:1, with 16,000 / 2,000 / 2,000 pairs of data randomly selected from each set. The proportional division of the dataset is conducive to improving the generalization ability of the model and ensuring that the model has good generalization ability on unseen data to avoid information leakage. The training set and test set are used to adjust the hyperparameters and model structure, and the validation set is used to finally evaluate the performance of the model. Dust activity on Mars has obvious seasonal changes and different characteristics, especially when there is a global dust storm season. These changes and characteristics will affect the learning of our deep neural network. Therefore, it is necessary to fully consider the division of the data set. By including the data of the period with global dust storms and the period without global dust storms in each set of data at a ratio of 1:1 as shown in Table 1, ensure that the model fully learns the dust activity pattern. The design of the data set aims to improve the model's ability to predict global dust storms and avoid overfitting the training set data.

[0048] Table 1

[0049]

[0050]

[0051] To further verify the reliability of the dataset, the seasons when the Martian dust events occur are also considered, and different seasonal data are used in the training and validation sets. The dataset is divided into spring and summer, which represent the first half of the year in the northern hemisphere, and autumn and winter, which represent the second half of the year and contain global dust storm events. Compared with datasets containing active dust activities, the error is larger when only datasets containing calmer dust activities are used for training. In order to reduce the error value, more dust activity data needs to be added to the dataset to make it more representative and reasonable.

[0052] Step 104, construct a pre-trained model for predicting sand and dust activity, which includes an encoder and a decoder; the encoder includes a number of ConvGRU layers with increasing depth, and the output of the previous ConvGRU layer is the input of the next ConvGRU layer, and finally the encoder outputs an intermediate hidden layer; the decoder includes a number of ConvGRU layers with decreasing depth, and the intermediate hidden layer serves as the input of each ConvGRU layer in the decoder; the number of ConvGRU layers of the decoder is less than the number of ConvGRU layers of the encoder.

[0053] It can be understood that the encoder-decoder architecture and ConvGRU layer used in this embodiment can be obtained by setting an input sequence (x 1 ,x 2 ,...x T ) and recurrent hidden layer (h t ), calculate an output sequence by iterating the following equation. It can be mapped to the sequence (y 1 ,y 2 ,...y T ), which is the output sequence.

[0054] h t =tanh(W hh h t-1 +W xh x t );

[0055] y t =W hy h t ;

[0056] Where W hh , W xh and W hy are all weight matrices.

[0057] Specifically, the adopted ConvGRU layer is based on the combination of the recursive neural network expansion algorithm and the convolutional neural network. The core idea of ​​ConvGRU is to combine matrix operations and convolution operations by applying GRU units at each position of the convolution calculation, so that the model can extract spatial features in continuous data. In this way, ConvGRU can both utilize the memory function of the GRU model and extract spatial features through convolution calculation. In addition, ConvGRU is very compact because it can use fewer parameters to complete complex calculation tasks. The main calculation formula of ConvGRU is as follows:

[0058] Z t =σ(W z *x t +U z *Ht-1 );

[0059] R t =σ(W r *x t +U r *H t-1 );

[0060]

[0061]

[0062] In the formula, Z t is the update gate, R t To reset the gate, is the candidate gate (memory state), * is the convolution operation (used to extract spatial features), ⊙ is the Hadamard product, x t Input time series; f is the ReLU activation function.

[0063] Each time a new input arrives, the gate R is reset. t Will control whether to clear the previous state and update gate Z t will control the amount of new information written to this state, At time step t is the intermediate memory unit, H t is the hidden layer.

[0064] The encoder-decoder architecture used is a neural network structure based on a recurrent neural network (RNN). The encoder is used to transform the input historical two-dimensional time series into a fixed-size hidden layer vector, and this fixed-size hidden layer vector is used as the input of the decoder to transform it into a future two-dimensional time series. From a probabilistic point of view, this model is to learn a historical time series (x 1 ,x 2 ,...x T ) under the condition that the future time series (y 1 ,y 2 ,...y T ) is expressed as:

[0065]

[0066] In one embodiment, the encoder includes 4 ConvGRU layers and the decoder includes 3 ConvGRU layers.

[0067] Each ConvGRU layer of the encoder reads each input two-dimensional time series data in sequence, and the H in each cell t According to the above calculation H tAfter receiving the sequence marked with t=0, that is, the end of the sequence, the hidden state of the entire input sequence encoder, that is, the middle hidden layer C, is obtained.

[0068] The ConvGRU layers of the decoder predict the historical time series by inputting the intermediate hidden layer C generated by the encoder. However, unlike the network structure in the encoder, it needs to input H output by the previous network layer. t-1 , and the y in the previous layer t-1 Conditioned on the intermediate hidden layer C generated by the encoder.

[0069] In one embodiment, the hidden state of the decoder at the prediction time is expressed as:

[0070] H t =φ(H t-1 ,y t-1 ,C);

[0071] In the formula, φ represents the nonlinear activation function; H t-1 represents the previous hidden state; y t-1 represents the previous time series; C represents the middle hidden layer.

[0072] The structure of the dust activity prediction model constructed by the present invention is divided into an encoder and a decoder. A two-dimensional time series is input into the encoder to update the transmitted hidden layer, and the output hidden layer is retained as the intermediate hidden layer C of the dust activity prediction model, and is input into the decoder to be used as the input of the next time step to generate a predicted target sequence.

[0073] from Figure 2It can be seen that by using different number of ConvGRU layers in the neural network construction of the encoder (ENCODER) and the decoder (DECODER), the encoder consists of 4 layers of ConvGRU layers, in which the connections of the third and fourth layers of neural networks are stacked together to form a unit, and input is made at the moment t=0, which helps to improve the expressiveness of the model and enable it to handle more data patterns and complexity, which is very useful for many complex machine learning tasks, especially for tasks with long-term temporal dependencies and dependencies on previous and next temporal information in this dataset. It is used to extract key features in time series, and the depth increases layer by layer to 64, 96, 128, and 256 respectively. The decoder consists of 3 layers of ConvGRU, and the depth decreases layer by layer to 256, 128, and 64 respectively to generate predictions of future dust. The decoder usually needs to generate an output smaller than the input matrix (for example, to generate a reduced matrix or a specific label), so it may not need as many layers as the encoder. This can reduce the computational complexity of the model and increase the speed of training and reasoning. At the same time, since the task of the decoder is usually to generate relatively simple outputs, it can have fewer layers, thereby reducing the risk of overfitting. A deeper encoder can help capture the complex features of the input image, while a shallower decoder can map these features to the desired output. The design details for the model parameters are shown in Table 2.

[0074] Table 2

[0075]

[0076] Step 106, input the two-dimensional time series data into each ConvGRU layer in the encoder in sequence, train the dust activity prediction pre-training model, and obtain a trained dust activity prediction model.

[0077] It is understandable that after training the prediction model with the training set, the prediction performance of the model needs to be tested with the test set. When testing the model performance, the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2 ) and mean absolute percentage error (MAPE) were used as evaluation indicators.

[0078] Step 108, predicting the future Martian dust using the trained dust activity prediction model.

[0079] The above-mentioned Mars dust activity prediction method obtains a training set, which is two-dimensional time series data; constructs a dust activity prediction pre-training model, which includes an encoder and a decoder; the encoder includes a plurality of ConvGRU layers with increasing depth, and the output of the previous ConvGRU layer is the input of the next ConvGRU layer, and finally the encoder outputs an intermediate hidden layer; the decoder includes a plurality of ConvGRU layers with decreasing depth, and the intermediate hidden layer is used as the input of each ConvGRU layer in the decoder; the number of ConvGRU layers of the decoder is less than the number of ConvGRU layers of the encoder; the two-dimensional time series data is sequentially input into each ConvGRU layer in the encoder, and the dust activity prediction pre-training model is trained to obtain a trained dust activity prediction model; and the future Mars dust is predicted by the trained dust activity prediction model.

[0080] The present invention constructs an encoder and decoder framework, and sets a number of ConvGRU layers with increasing depth layer by layer in the encoder to extract key features in the time series; sets a number of ConvGRU layers with decreasing depth layer by layer in the decoder to generate predictions of future dust, and the number of ConvGRU layers of the decoder is less than the number of ConvGRU layers of the encoder, because the decoder usually needs to generate an output smaller than the input matrix, so it may not need as many layers as the encoder. Since the task of the decoder is usually to generate relatively simple outputs, it can have fewer levels, thereby reducing the risk of overfitting. A deeper encoder can help capture the complex features of the input image, while a shallower decoder can map these features to the desired output. Through such a setting, the computational complexity of the model can be reduced and the speed of training and reasoning can be improved.

[0081] In one of the embodiments, the training set is preprocessed; the data of the periods with global sandstorms and the periods without global sandstorms are divided into physical elements; the physical elements include: air pressure, temperature, wind speed, wind direction and sky optical thickness; the air pressure, temperature, wind speed, wind direction and sky optical thickness are divided into input dimensions.

[0082] It is understandable that Martian dust activity is a complex phenomenon affected by many factors. Thermal tides play a vital role in the excitation and propagation of Martian dust. In addition, the enhancement of low-altitude convergence flow and Hadley circulation is also closely related to Martian dust activity. The enhanced atmospheric circulation system will cause turbulence on the surface of Mars, promoting dust suspension and migration. The formation and movement of low-pressure cyclones will also cause changes in airflow, leading to increased Martian dust activity. In addition, seasonal changes also have a significant impact on Martian dust activity. Due to the different positions of Mars orbiting the sun, solar radiation and temperature distribution will also change, which in turn affects the generation and migration of dust. Therefore, dust activity will show different patterns and intensities in different seasons.

[0083] When approaching the surface of Mars, due to the unstable atmospheric state (strong convection) and the rapid increase in dust content, it is impossible to detect the dust content close to the ground and other atmospheric parameters, such as temperature, density, etc., and this part of the altitude is the main activity area of ​​the sandstorm, and the dust content there accounts for the vast majority of the dust content at the entire altitude, so other observation methods are needed to determine it. Although the data of the optical thickness of the air column cannot obtain dust information in the vertical direction, it is most sensitive to the radiation information of a large amount of dust content near the surface, and can intuitively distinguish the outline of the sandstorm, and is used to study the origin, evolution, and development process of the sandstorm.

[0084] According to the dynamic mechanism of Martian dust activity, relevant physical factors are considered as inputs of the model for prediction. These physical factors include atmospheric pressure, temperature, wind speed and wind direction. Specifically, there are four variables in the EMARS dataset: air pressure P, temperature T, wind speed U and wind direction V. By experimenting with different combinations of these physical quantities, dust events on Mars can be predicted more accurately, and the impact of each variable on dust activity can be deeply understood.

[0085] As in common image prediction, at the input layer of the model, the data pairs are divided into five dimensions: prediction time series length (seqlen), batch size (batch size), color channel (channel), length (length) and width (width). We also set five input dimensions, in which a channel is assigned to each element. That is, in the channel dimension, we put five physical elements: air pressure P, temperature T, wind speed U, wind direction V and sky optical thickness DOD, so that we can consider the relationship between different elements at the same time during the convolution operation. Figure 3 In this paper, we tested the combination of DOD with surface pressure, near-surface temperature at 767Pa, DOD with latitude and radial wind field at the same altitude from the EMARS dataset. In addition, we compared the input of DOD alone with the input of the above variables together into the model.

[0086] The prediction error accuracy of different combinations was investigated. The results showed that the prediction error accuracy improved when combining DOD with any of the variables. Furthermore, the prediction accuracy of the model was the best when all variables were considered, indicating that each of the variables we selected played a role in explaining dust activity.

[0087] It is particularly noteworthy that when the variables are paired, it can be observed that the combination of DOD and longitudinal wind significantly reduces the prediction error, followed by the combination of DOD and radial wind. This combination effectively captures the impact of wind on dust activity, thereby improving the prediction accuracy. The model provided in this embodiment shows that the wind field is the main cause of dust activity.

[0088] In one embodiment, the input dimensions include predicted time series length, batch size, color channels, length and width.

[0089] In one embodiment, each ConvGRU layer of the encoder has an input channel corresponding to the input dimension.

[0090] It is understandable that in the "Spirit" and "Opportunity" missions, the occurrence of dust storms will affect the absorption of solar energy by the battery array. This causes the battery to not work properly, so the solar panels need to be cleaned or recycled in time during the dust storm. In addition, dust storms will reduce visibility on the surface of Mars, increasing the difficulty of human navigation on the surface of Mars in the future. Therefore, we need to select appropriate input and output time dimensions to predict the occurrence of dust storms in advance. The front of a dust storm can propagate eastward at a speed of 20-28 m / s and can spread about 800 km in 12 hours. Therefore, in order for the Mars rover to be able to respond before the dust storm arrives, we choose a forecast duration of 12 hours and a forecast time step of 1 hour to meet the requirements of solar panel storage and detection instrument operation. This choice of forecast time dimension can provide the necessary protection and data support for the Mars exploration mission.

[0091] In one embodiment, in order to verify the effectiveness of the prediction model provided by the present invention, the prediction comparison is performed by inputting the test set into the model provided by the present invention and other five models. For ease of description, the prediction model provided by the present invention is referred to as ConvGRU-S2S.

[0092] We used the autumn and winter of MY28 and the whole year of MY31 as test sets, and used the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2 ) and mean absolute percentage error (MAPE) as evaluation indicators, the test results show that our model ConvGRU-S2S performs best on the test set with a RMSE of 8.73×10-4 , MAE is 13.48×10 -3 , R 2 98.12×10 -2 The RMSE of the ConvLSTM-S2S model, which has a similar structure to ours, is 12.86×10 -4 , MAE is 23.37×10 -3 , R2 is 92.07×10 -2 Overall, our model achieves better prediction results on the test set. The prediction results are shown in Table 3.

[0093] Table 3

[0094]

[0095] On the other hand, there are global dust storms with interannual variability on Mars, which can affect local temperatures by up to 60K and affect the activity of gravity waves, which has been confirmed by the research results of the Mars Monitoring System team. Due to the intensity of weather activities, the average prediction results over a period of time are difficult to reflect the true prediction performance of the model. Therefore, in the process of model training, we refer more to the loss function change curves of the training set and the validation set to train the model. However, in order to further verify the prediction effect of the model and consider the stability of the model prediction, we need to test it on the test set. In addition, it is key to test on the spatial distribution of errors to ensure the reliability of the model prediction results.

[0096] like Figure 4 As shown in the figure, it can be seen that the prediction error of each model is positively correlated with the prediction length. In the ConvGRU-S2S model represented by the bold solid line and dots in the figure, the prediction error of our model does not decrease significantly over time, and has good stability over time. Although the prediction error is similar to that of the ConvLSTM-S2S model, as the forecast time increases, ConvGRU-S2S outperforms ConvLSTM-S2S in model fitting performance evaluation. In the 12-hour forecast, the R2 of ConvLSTM-S2S is lower than 0.9, while ConvGRU-S2S is still higher than 0.95.

[0097] In addition to considering the temporal performance stability of the model, we should also pay attention to the spatial distribution of the error to evaluate the prediction ability of the model in different regions. By improving the prediction ability of the model in certain regions, we can more accurately predict the global dust distribution in practical applications. In order to evaluate the performance of the model in different regions, we selected three indicators from the test set and analyzed the prediction error of the global dust distribution. Figure 5The error distribution of different models predicting global dust distribution is shown. It can be seen that the error distribution of different models in different regions is relatively small. The mean square error and mean absolute error near the Tharsis Mountains are relatively small after 12 Mars local times, and the errors are mainly concentrated in the Hellas Basin. This shows that the model is very suitable for predicting the area near the Tharsis Mountains. In addition, from Figure 5 In (c), we can see that the ConvLSTM-S2S model predicts low values ​​for global dust distribution overall. In contrast, the error distribution of the ConvGRU-S2S model in different regions is more concentrated than other models and closer to 0. This shows that the model's prediction ability in different regions is relatively stable and accurate.

[0098] Rapidly developing dust storms have diurnal variations, which have important implications for the rapid transport of heat and other substances

[33] Local and regional dust storms will continue to develop and condense, occasionally becoming global-scale dust storms.

[34] Rapidly developing dust storms on Mars, like extreme precipitation on Earth, could have serious consequences for future missions to Mars, and predicting the distribution of dust during rapidly developing dust storms could improve early warning capabilities for dust storms and help scientists and rovers prepare. At the same time, by predicting rapidly developing dust storms on Mars, we can test the predictive ability of models to handle rapidly changing dust activity. This is important for improving the predictive power of models as well as the robustness of models.

[0099] During the global dust storm of MY28268°Ls (approximately July 2, 2007), two rapidly changing regional-level dust storms appeared near the Martian equator: Dust A and Dust B, the former of which weakened rapidly and the latter of which strengthened rapidly. The initial growth stage of the dust storm was a complex process, with the DOD of Dust A decreasing from 0.99 to approximately 0.69 within 12 hours, while the DOD of Dust B increased from 0.70 to approximately 0.86. In model training, we pay more attention to the global error of the model on Mars, but local errors are equally important, so in this experiment we pay more attention to the error changes in the two dust regions. In order to study the predictive effect of the ConvGRU-S2S model on the development and attenuation of such rapidly changing dust storms, we used the global dust distribution data of the previous 12 hours to input the model hour by hour and obtained Figure 6 (a) shows the prediction results.

[0100] According to the figure Figure 6(a) shows the experimental results. We observed that the dust activity of dust A predicted by the model was weakening, which is consistent with the actual dust changes. However, at 12 hours, the predicted dust activity range was larger, indicating that the impact of local dust storms still existed. In contrast, the prediction results for dust B showed that the model predicted range and intensity were closer to the actual situation, and strong local dust storms broke out in the dust area within a few hours. In addition, Figure 6 The experimental results in (b) show that the overall prediction result of dust B is slightly better than that of dust A, but the DOD prediction error of the model for these two rapidly changing local dust storms does not exceed 10%, indicating that the model has a high prediction accuracy. This also provides more reliable support for our forecast research and proves the application potential of the ConvGRU-S2S model in predicting the spatiotemporal distribution of rapidly changing local dust storms.

[0101] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0102] In one embodiment, Figure 7 As shown, a device for predicting dust activity on Mars is provided, comprising: a data acquisition module 202, a model building module 204, a pre-training module 206 and a prediction module 208, wherein:

[0103] The data acquisition module 202 is used to acquire a training set, which is two-dimensional time series data.

[0104] The model construction module 204 is used to construct a pre-trained model for predicting dust activity. The pre-trained model for predicting dust activity includes an encoder and a decoder; the encoder includes several ConvGRU layers, and the output of the previous ConvGRU layer is the input of the next ConvGRU layer, and finally the encoder outputs an intermediate hidden layer; the decoder includes several ConvGRU layers, and the intermediate hidden layer serves as the input of each ConvGRU layer in the decoder.

[0105] The pre-training module 206 is used to input the two-dimensional time series data into each ConvGRU layer in the encoder in sequence, train the dust activity prediction pre-training model, and obtain a trained dust activity prediction model.

[0106] The prediction module 208 is used to predict the future dust on Mars by using the trained dust activity prediction model.

[0107] For the specific definition of the Mars dust activity prediction device, please refer to the definition of the Mars dust activity prediction method above, which will not be repeated here. Each module in the above-mentioned Mars dust activity prediction device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0108] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 8 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, 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 operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store Martian dust activity prediction data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for predicting Martian dust activity is implemented.

[0109] Those skilled in the art will understand that Figure 8 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0110] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0111] Step 102, obtaining a training set, where the training set is two-dimensional time series data.

[0112] Step 104, construct a pre-trained model for predicting sand and dust activity, which includes an encoder and a decoder; the encoder includes a number of ConvGRU layers with increasing depth, and the output of the previous ConvGRU layer is the input of the next ConvGRU layer, and finally the encoder outputs an intermediate hidden layer; the decoder includes a number of ConvGRU layers with decreasing depth, and the intermediate hidden layer serves as the input of each ConvGRU layer in the decoder; the number of ConvGRU layers of the decoder is less than the number of ConvGRU layers of the encoder.

[0113] Step 106, input the two-dimensional time series data into each ConvGRU layer in the encoder in sequence, train the dust activity prediction pre-training model, and obtain a trained dust activity prediction model.

[0114] Step 108, predicting the future Martian dust using the trained dust activity prediction model.

[0115] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0116] Step 102, obtaining a training set, where the training set is two-dimensional time series data.

[0117] Step 104, construct a pre-trained model for predicting sand and dust activity, which includes an encoder and a decoder; the encoder includes a number of ConvGRU layers with increasing depth, and the output of the previous ConvGRU layer is the input of the next ConvGRU layer, and finally the encoder outputs an intermediate hidden layer; the decoder includes a number of ConvGRU layers with decreasing depth, and the intermediate hidden layer serves as the input of each ConvGRU layer in the decoder; the number of ConvGRU layers of the decoder is less than the number of ConvGRU layers of the encoder.

[0118] Step 106, input the two-dimensional time series data into each ConvGRU layer in the encoder in sequence, train the dust activity prediction pre-training model, and obtain a trained dust activity prediction model.

[0119] Step 108, predicting the future Martian dust using the trained dust activity prediction model.

[0120] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and 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 embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0121] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0122] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the present invention. It should be noted that, for a person of ordinary skill in the art, several modifications and improvements may be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for predicting dust activity on Mars, characterized in that: The method comprises: Obtaining a training set, wherein the training set is two-dimensional time series data; Constructing a dust activity prediction pre-training model, the dust activity prediction pre-training model includes an encoder and a decoder; the encoder includes a plurality of ConvGRU layers with increasing depth, and the output of the previous ConvGRU layer is the input of the next ConvGRU layer, and finally the encoder outputs an intermediate hidden layer; the decoder includes a plurality of ConvGRU layers with decreasing depth, and the intermediate hidden layer is used as the input of each ConvGRU layer in the decoder; the number of ConvGRU layers of the decoder is less than the number of ConvGRU layers of the encoder; Inputting the two-dimensional time series data into each ConvGRU layer in the encoder in sequence, training the dust activity prediction pre-training model, and obtaining a trained dust activity prediction model; The future dust on Mars is predicted using the trained dust activity prediction model.

2. The method for predicting dust activity on Mars according to claim 1, characterized in that: The training set includes a period with global sandstorms and a period without global sandstorms; the data ratio of the period with global sandstorms and the period without global sandstorms is 1:

1.

3. The method for predicting dust activity on Mars according to claim 2, characterized in that: It also includes preprocessing the training set; Dividing the data of the period with global sandstorm and the period without global sandstorm into physical elements; The physical factors include: air pressure, temperature, wind speed, wind direction and sky optical thickness; The air pressure, the temperature, the wind speed, the wind direction and the sky optical thickness are divided into input dimensions.

4. The method for predicting dust activity on Mars according to claim 3, characterized in that: The input dimensions include the predicted time series length, batch size, color channel, length and width.

5. The method for predicting dust activity on Mars according to claim 3 or 4, characterized in that: Each ConvGRU layer of the encoder has an input channel corresponding to the input dimension.

6. The method for predicting dust activity on Mars according to any one of claims 1 to 4, characterized in that: The encoder includes 4 ConvGRU layers, and the decoder includes 3 ConvGRU layers.

7. The method for predicting dust activity on Mars according to claim 8, characterized in that: The hidden state of the decoder at the prediction time is expressed as: H t =φ(H t-1 ,y t-1 ,C); In the formula, φ represents the nonlinear activation function; H t-1 represents the previous hidden state; y t-1 represents the previous time series; C represents the middle hidden layer.

8. A device for predicting dust activity on Mars, characterized in that: The device comprises: A data acquisition module is used to acquire a training set, wherein the training set is two-dimensional time series data; A model building module is used to build a pre-trained model for predicting dust activity, wherein the pre-trained model for predicting dust activity includes an encoder and a decoder; the encoder includes a plurality of ConvGRU layers, and the output of the previous ConvGRU layer is the input of the next ConvGRU layer, and finally the encoder outputs an intermediate hidden layer; the decoder includes a plurality of ConvGRU layers, and the intermediate hidden layer is used as the input of each ConvGRU layer in the decoder; A pre-training module, used for sequentially inputting the two-dimensional time series data into each ConvGRU layer in the encoder, training the dust activity prediction pre-training model, and obtaining a trained dust activity prediction model; The prediction module is used to predict the future Martian dust by using the trained dust activity prediction model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images