SSC-KA: A Framework of Space Situational Knowledge Acquisition for Status-Cognition of Satellites

With the rapid development of space technology, the environment of the space domain has become more and more complex and changeable, which brought great difficulties in cognition of space domain activity. As space domain awareness (SDA) required, any relevant information and knowledge from various sources are needed as much as possible, while all of those can be sorted and integrated for effective cognition of space objects, including the cognition of their status. This paper proposes a knowledge integration framework (SSC-KA) designed for the cognition of space target and its status. In the framework (SSC-KA), open-source information and data acquired from multi-kinds of sensors are sent into parallel channels, and then processed by the algorithm this paper designed into sequence data as the channels’ output. Furthermore, the rules and four statuses defined in this paper can be judged for the anomaly detection of satellites. Based on the space domain knowledge acquisition framework of SSC-KA, this paper describes a complete abnormal state detection method for satellites step by step through multi-level feature engineering. Therefore, the method is used to analyze four different statuses of satellites in this paper, to verify the validity and feasibility of the application of the method in the cognition of spatial events, thus laying the foundation for the cognition of Space Domain Awareness.

turized electronics, commercialization and standardization of 28 The associate editor coordinating the review of this manuscript and approving it for publication was Rosario Pecora . traditionally bespoke satellite subsystem components, and 29 access to launch vehicles as secondary payloads have also 30 contributed to the improved accessibility of space and a 31 large increase in satellites and satellite operators. Nowadays, 32 commercial entities (Space X, One Web, Planet Labs) have 33 already sent their giant constellation into the earth's orbit for 34 space-based internet access. Reference [3], [4], [5], [6], [7] 35 Since then space situational awareness has recently become 36 an important research topic due to the enormous amount of 37 space objects. 38 Space situational awareness (SSA) [8] is the perception 39 of the elements in the environment within a volume of time 40 and space, the (organizational) comprehension of their mean-41 ing, and the projection of their status shortly (in the space 42 domain). SSA is the first step of space domain knowledge, 43 • No standard method of calibrating sensors and informa-66 tion sources has been developed; 67 • Tasking is addressed to individual sensors for specific 68 data rather than to a comprehensive system for informa-69 tion required to address needs and requirements; 70 • No rigorous understanding of space environment effects 71 and impacts on space objects; 72 • No framework that encourages and enables big data 73 analysis, and supports an investigative 'from data to 74 discovery' paradigm. 75 In general, we lack a consistent method to understand all 76 of the causes and effects relating space objects and events. 77 Enabled by the most recent advancement in sensor tech-78 nology, researchers and operational engineers rely on a large 79 amount of tracking data that can be processed to iden-80 tify, characterize, and understand the intention of space 81 objects [12]. Improving methods applied to the SDA domain 82 task will require up-to-date approaches and algorithms that 83 will be able to predict and prevent satellite anomalies. 84 In this paper, we aim at designing an effective framework 85 for multi-source information on relevant space assets to pro-86 vide a more accurate result for satellite Anomaly detection. 87 Multi-source information fusion is a sophisticated estima-88 tion process that allows users to assess complex situations 89 more accurately by effectively combining core evidence in 90 the massive, diverse, and sometimes conflicting information 91 received from multiple sources. It involves integrating infor-92 mation from these multiple sources to produce specific and 93 comprehensive unified estimates about an entity, activity, 94 or event. Multisource fusion systems seek to combine infor-95 mation from multiple sources and sensors in a wide variety 96 of applications to achieve analysis and decision-supporting 97 inferences that cannot be achieved with a single sensor or 98 source. Designing and implementing data and information 99 fusion systems requires a multidisciplinary approach, as seen 100 in the diagram below that shows the disciplines and methods 101 needed to achieve holistic system designs.

103
As for the SDA tasks, the information fusion method is still 104 under research. With the rapid development of the theory 105 and technology of AI, more and more difficult and complex 106 SDA tasks can be solved now. Firstly, the machine learning 107 technology offers emergent solutions to many industrial sys-108 tems, such as transportation [13], manufacturing [14], video 109 surveillance [15], climate change [16], and net-working [17]. 110 Machine learning technology [18], [19] plays an important 111 role in solving practical problems. The article [20] present 112 an adaptive strategy of active control information updates for 113 use in dynamic collaborative activity, which shows applicable 114 VOLUME 10, 2022 ranking methods of semantic data mining for selecting rele-115 vant information in collaborative activity.

116
Especially for the situation awareness problem where the 117 original analytical model is complex or it is difficult to 118 obtain the real model, the machine learning algorithm helps 119 to provide a solution to some extent. As to administrate the 120 space domain assets, anomaly detection is the basic layer  The main purpose of this investigation is to design a 130 framework for several sources of information fusion to diag-131 nose whether the status of a satellite is abnormal. The 132 framework is composed of multi-channel information pro-133 cess and made the fusion algorithm based on the atten-134 tion mechanism for the structure data from every channel.

135
The following section formalizes the framework and our 136 approach.

139
In this paper, we propose a framework for Space Situation 140 knowledge acquisition in order to integrate space situational 141 resources from multiple sources more quickly, comprehen-142 sively and systematically for Space Situation Awareness 143 activities and Space Domain Awareness activities. The frame-144 work of space situation knowledge acquisition (SSC-KA) is 145 showed in figure 2. The framework consists of three main 146 phases, which are multi-channel source information, pre-147 processing to generate data, and multi-channel to generate 148 knowledge conclusions for decision supporting.

149
In the first stage, knowledge from various sources should 150 all be obtained, which including open-source (news, report, 151 websites, etc.) and sensors (optical sensors, ground-based 152 radar sensors, and space-based sensors). In the second stage, 153 the multi-source knowledge will be pre-processed separately 154 for generating information that can be understood by the sys-155 tem. In the third stage, the information will be processed into 156 structured data by several different channels with targeted 157 models (encoder-decoder and other types of models) so that 158 the data from each channel will be calculated based on rules. 159 Finally, the data can be used for the space situation knowledge 160 representation by judging the status of space targets such as 161 satellites.

162
The characteristic of the framework SSC-KA is that    radar, is closely related to the target's attitude motion, shape 202 size, surface material, and other attributes. It is one of the key 203 information sources for tracking and recognizing the space 204 target. To fully exploit the attitude motion information in the 205 RCS sequence of space target, the dynamic RCS observation 206 sequence is studied. The major method of obtaining RCS data 207 is shown in table 2.

208
From the perspectives of the attitude motion period inver-209 sion and the attitude motion pattern recognition, the inversion 210 of the rotation period and precession period is studied, and 211 the RCS feature extraction method for distinguishing the 212 three-axis stable target from the unstable rotation target is 213 explored.

215
As the SSC-KA framework showed, the multi-sources 216 of information put into the system would be processed 217 in parallel channels individually. Through the pretraining 218 VOLUME 10, 2022 process, the information has transformed into sequence-219 structure data set. Then for each channel, the sequence 220 data can be processed independently and specifically. In the 221 SSC-KA framework, the major three-channel is described as  According to the variation rule of the OCS, the inversion and 245 identification of the attitude, geometry, and working state of 246 space objects can be performed. In this paper, we take the 247 OCS curves as experimental data.

248
In the OCS-Data Process, the OCS data will be transformed where, x is the input sequence data, n is the length of sequence 259 data, W and S respectively represent convolution kernel and 260 sliding step number, and f conv are the output vector after 261 convolution.

262
In the Decoder, two GRU layers are used to reconstruct 263 OCS curve. The Classifier consists of three full connection 264 layers using ReLU activation functions and one output layer 265 using sigmoid activation functions. The eigenvectors gener-266 ated by the encoder are the inputs to the classifier. The output 267 layer uses the sigmoid function to map features to categories.  Here the model (as Figure 5 showed) for the RCS-Data 284 processing channel constructed in this paper consists of an 285 input layer, four hidden layers and an output layer. The hidden 286 layer is divided into GRU hidden layer and full connection 287 layer. The number of nodes in the input layer is equal to the 288 sample input dimension. Moreover, the Bidirectional Gated 289 Recurrent Unit (bi-GRU) can be adopted by the model in 290 GRU hidden layer to learn the complete before and after 291 information of time series for higher accuracy. And the ReLU 292 function between the feature output layer and the feature 293 output layer makes the feature extracted in network training 294 more effective. In this channel, the TLE data will be processed into the 301 orbital elements' tensor data.

303
The attention mechanism is often used for training neural net-304 works, which allows models to learn alignments between dif-305 ferent modalities. In this paper, the self-attention mechanism 306 is adopted to further capture the sensor-source dependence 307 between sequences in sensor data samples.

308
The different sources of satellites' information will form

319
Here the multi-attention mechanism is adopted with the

335
where each raw satellite motion characteristic S motion is a 336 sequence of the orbit features of the satellite and S attitude is 337 a sequence of the attitude status of the satellite.

338
Since the data through the process above, the final task of We describe the satellite status comprehensively by its 344 attitude and orbital motion as table 3 shown.

366
In this section, we demonstrate the validity of the framework It can be seen from the distribution attributes of different 384 features of different GEO satellites in Figure 9 that some fea-385 tures are basically general. Therefore, we can initially believe 386 that not every feature attribute of a satellite has important 387 feature significance. However, due to the system's need for 388 computing power and speed, we consider feature dimension 389 reduction. The input satellite attribute information is quanti-390 fied and reduced to low-dimensional data for satellite feature 391 representation.

392
From the importance rank of the features (Fig 10) from 393 open-source orbital information acquired, the satellites' 394 motion information can be calculated by the attention mech-395 anism above the original features, which can make the 396 motion information of the targets in GEO belt embedding 397 into two-dimensional space for characterization, as shown 398 in Figure 11.      It can be seen that the loss function of the training set and 431 verification set basically remains unchanged and converges 432 to 0 after more than 40 training rounds, and the accuracy 433 curve converges to 1, proving that the GRU deep neural net-434 work model constructed has no under-fitting or over-fitting 435 phenomenon, achieving good training effects.

437
In this section, we collected the multi-source of information 438 from 29 Mar 2022 04:00:00.000 UTCG to 30 Mar 2022 439 04:00:00.000 UTCG, among which is a station (40.0386 N, 440 105.597 E). In order to analyze the effect of the method we 441 proposed, we take the measurements as input to decide the 442 status of space objects.

443
The experiment set as the table 6 shown as the satellites' 444 motion mode discerption, and the attitude mode set as the 445 fig16 showed. In this section, the normal attitude mode we 446 set is the normal attitude to ground triaxial stability.    The follow-up study will continue to explore the effective 518 approach for further deepen situational understanding and 519 cognition.  He is currently a Lecturer with Space Engi-651 neering University. His research interests include 652 machine learning, computer vision, and data 653 mining. 654 655 VOLUME 10, 2022