CN-121980192-A - Satellite observation demand generation method based on triplet capsule network

CN121980192ACN 121980192 ACN121980192 ACN 121980192ACN-121980192-A

Abstract

The invention discloses a satellite observation demand generation method based on a triplet capsule network, and relates to the technical field of artificial intelligence and remote sensing application. The method comprises the steps of firstly carrying out network reasoning through a graph neural network to generate Top-N candidate triples, secondly constructing a triplet feature matrix, generating a corresponding interaction matrix through multi-scale convolution of relation modulation, then converting the interaction matrix into interpretable semantic capsule vectors through a primary capsule generation, relation modulation projection and dynamic routing mechanism, and finally carrying out decision scoring and sequencing on the semantic capsule vectors by combining a relation semantic modulation function to output candidate triples with highest compatibility scores as a complement decision result. The invention realizes unified modeling from multiple reasoning to final decision, improves the accuracy, stability and interpretability of element completion decision in the process of demand generation, and can be widely applied to satellite demand planning and task planning systems in the field of earth observation.

Inventors

YU XIAOGANG
XING YING
CHEN JINYONG
WANG QIANG
ZHI JUN
GAO CHAOHUI
SHEN PENG
GUO XUTONG
Wu Kaizheng

Assignees

中国电子科技集团公司第五十四研究所
北京市遥感信息研究所

Dates

Publication Date: 20260505
Application Date: 20260330

Claims (8)

1. The satellite observation demand generation method based on the triplet capsule network is characterized by comprising the following steps of: step 1, collecting satellite observation task demand data containing missing elements, and performing deduplication, missing value occupation, unit conversion and word list alignment on heterologous data in the satellite observation task demand data to form a unified task sample library; step2, aiming at a task sample library, adopting a graph neural network to perform network reasoning to obtain the characteristic representation of the entity/relationship, and further constructing a plurality of groups of test samples containing head entities and relationships; step 3, TOP-N reasoning is carried out on each group of test samples, N groups of candidate triples comprising a head entity, a relation and a tail entity are obtained, and the candidate triples are recorded as candidate sets corresponding to the group of test samples; Step 4, aiming at each candidate triplet in the candidate set, generating a corresponding interaction matrix by constructing a triplet feature matrix and introducing multi-scale convolution of relation modulation; Step 5, converting the interaction matrix into a semantic capsule vector through primary capsule generation, relation modulation projection and a dynamic routing mechanism; step 6, a scoring mechanism is guided through a relation semantic modulation function, and compatibility scoring is carried out on the candidate triples by combining semantic capsule vectors; Step 7, calculating a loss function based on the scoring result of the candidate triplet, and iteratively executing the steps 4-7 to perform back propagation training until the loss function value converges, and storing model parameters to obtain a final triplet capsule network; And 8, aiming at the head entity and the relation input during practical application, executing the steps 3-6 by combining the final triplet capsule network, thereby outputting the candidate triplet with the highest compatibility score as a completion decision result and obtaining a complete satellite observation demand generation result.
2. The method of claim 1, wherein the fields of the satellite observation task demand data in step 1 cover satellite target, payload, imaging mode, spatial resolution, priority, and time window elements.
3. The method for generating satellite observation requirements based on a triplet capsule network according to claim 1, wherein the specific manner of step 2 is as follows: step 201, ID mapping, namely extracting a task sample library as entity type and relation type, distributing unique integer ID for the whole entity and relation, and establishing entity2ID and relation2ID mapping; Step 202, embedding and initializing, namely uniformly initializing an entity and a relation vector by using an Xavier, wherein the vector dimension is d; in step 203, the graph neural network application directly calls a GNN model, inputs graph topology adjacency information and initial embedding formed by entities and relations, and outputs and obtains feature representations of the entities/relations, namely h, r and t, wherein h is a head entity, r is a relation type, and t is a tail entity, so as to construct a plurality of groups of test samples (h, r,.
4. The method for generating satellite observation requirements based on a triplet capsule network according to claim 1, wherein the specific manner of step 3 is as follows: For each group of test samples, fixing a head entity h and a relation r, traversing each entity in the whole entity set, sequentially marking the entities as candidate tail entities, reserving the first N candidate tail entities after sorting from high to low according to scores, marking the first N candidate tail entities as Top-N to obtain N groups of candidate triples (h, r, t) corresponding to the current test sample, wherein the principle of marking the candidate tail entities is as follows: is a Sigmoid function; corresponding to h, r and t respectively And (3) repeatedly executing the step (3) to obtain candidate sets corresponding to the plurality of groups of test samples.
5. The method for generating satellite observation requirements based on a triplet capsule network according to claim 1, wherein the specific manner of step 4 is as follows: the N candidate triples (h, r, t) are processed in parallel: Step 401, splicing candidate triples (h, r, t) according to channels to form a triplet feature matrix X= [ h; r; t ]; X is a two-dimensional tensor with the shape of (3, d), and carrying out normalization processing on X; Step 402, introducing a one-dimensional convolution kernel of a multi-scale s, and performing low-rank modulation on the convolution kernel by a relation vector r, wherein the parameterization form is as follows: Wherein s is the length of a one-dimensional convolution kernel; Modulating a convolution kernel for the relationship at scale s; is a static convolution kernel; m is the number of the base cores under the scale s; is a scalar modulation coefficient, is generated by a relation vector r, and is calculated by the following way: ; Wherein, the As a weight vector of the weight vector, In order for the offset to be a function of, Is a relationship vector; step 403, convolving with the modulated convolution check X, and obtaining an interaction matrix by gating the linear unit GLU and layer normalization to suppress noise and highlight key segments : Wherein, the For the convolution output of the scale s, the step size is set to stride=1; for mapping matrices, for projecting the relation vector r into a space consistent with the convolved output channels, the GLU is a gating function, And the convolution output is spliced according to the channel in the channel dimension to form a uniform interaction matrix F.
6. The method for generating satellite observation requirements based on a triplet capsule network according to claim 1, wherein the specific manner of step 5 is as follows: step 501, primary capsule generation, namely decomposing an interaction matrix F according to rows to obtain M primary capsule sets ,i=1,2,3,......M; Step 502, relation modulation projection: Wherein, the In the form of a static projection matrix, The tensor is modulated for the relationship, J is a high-level capsule index, j=1, 2, 3. ; Is the number of high-rise capsules; step 503, dynamic routing mechanism, for the j-th high-level capsule, converging the prediction vectors from each primary capsule to form And non-linearly compressed by squash to form The route logits is iteratively updated according to consistency, and the whole routing process is represented by the following formula: Wherein the method comprises the steps of Is a projection matrix modulated by a relation vector r; a predictive vector for the ith primary capsule to the jth higher-level capsule; for routing logic, the matching priori between the ith primary capsule and the jth high-level capsule is represented, and the initial value is 0; For the coupling coefficient, satisfy ; A weighted sum of all the prediction vectors; Outputting for the j-th high-level capsule, and compressing the length to the (0, 1) interval; Is defined as a capsule compression function , Representing a vector norm; the vector inner product is represented by the vector, For updating ; Iteratively performing the routing process to And (3) with Converging; Step 504, based on Corresponding semantic capsule vector is constructed by outputting and constructing each high-level capsule ; 。
7. The method for generating satellite observation requirements based on a triplet capsule network according to claim 6, wherein the specific manner of step 6 is as follows: In step 601, a relation semantic modulation function, namely generating a modulation vector with consistent dimensions by a relation vector r, and emphasizing a semantic subspace related to r: Wherein, the A weight matrix modulated for the relation, for projecting r into a space in the same dimension as the scoring vector; is a bias vector; A relation semantic modulation function is used for giving different dimensionalities differential weights in a scoring function; is a nonlinear activation function defined as ; Step 602, scoring function and sorting, namely scoring compatibility of candidate triples with diagonal modulation: Wherein, the Representing element-by-element multiplication, i.e. pairing Correspondingly multiplying each dimension of g (r); Representing the vector transpose; i.e., the compatibility score of the candidate triplet (h, r, t).
8. The method for generating satellite observation requirements based on a triplet capsule network according to claim 7, wherein the specific way of calculating the loss function in step 7 is as follows: loss calculation: for a given test sample (h, r,? let the candidate tail entity set be t= { T 1 , t 2 ,...,t N }, where the only correct tail entity is T + , the loss function is defined as: Wherein, the Is a scaling factor for adjusting the sensitivity of the inter-candidate score differences in softmax.

Description

Satellite observation demand generation method based on triplet capsule network Technical Field The invention relates to the technical field of artificial intelligence and remote sensing application, in particular to a satellite observation demand generation method based on a triplet capsule network, which belongs to a knowledge graph representation learning and element completion decision method, and can be applied to a satellite element completion decision system in the fields of meteorological monitoring, resource exploration, environmental protection, disaster early warning and the like. Background Satellite observation task requirements are typically made up of a number of elements such as "satellite objectives, loading, imaging modes, spatial resolution, priority, time windows" and the like. In an actual business process, due to various sources, different recording calibers or information deletion, delay and other reasons, the problem of key element deletion or inconsistent expression often occurs in task recording. The existing completion mode based on the regression model often depends on single reasoning, the prediction result is difficult to ensure stability and accuracy, and the results obtained based on multiple reasoning lack of uniform decision basis and confidence level evaluation, so that reliable support is difficult to provide for subsequent task planning. These deficiencies directly affect task orchestration and resource scheduling, resulting in scheduling conflicts, reduced execution efficiency, or difficult quality of observation. In the existing research, the complement decision method aiming at the missing elements mainly comprises the following ideas: 1. convolutional Neural Network (CNN) a certain local interaction pattern can be extracted by using the translational invariance of local receptive fields and convolutional kernels. However, in the scene of the triplet (h, r, t), the CNN often extracts channel characteristics independently, and then separates and interfaces with the discrimination head, so that fine-granularity ternary coupling of h-r, r-t, h-t and conjunctions thereof is difficult to model explicitly, and meanwhile, a unified modulation mechanism aligned with 'relation semantics' is absent, so that consistency and interpretation of cross relation are insufficient. 2. The cyclic neural network (LSTM) is characterized by a gating mechanism to describe time sequence dependence, so that sequential interaction of multiple elements can be simulated. However, the triple belongs to a symmetrical/displacement sensitive structure, the LSTM needs to manually set an input sequence, ternary coupling is often compressed into sequential binary transition, synchronous interaction expression in the triple is insufficient, meanwhile, the relation semantics are remained in an implicit state, and the explicit constraint consistent with a scoring head is lacked, so that decision stability is affected. 3. Transformer has global modeling advantages, but basic operation is mainly focus-to-focus (focus-to-focus), and additional design is still needed for a ternary coupling mode requiring relation condition constraint, and when the sample size is limited or relation semantics are not explicitly injected, the interpretation of the focus weight and the cross relation stability are weak. Although the method promotes the development of element complement, the limitation is still obvious, namely, the method lacks special modeling for fine-granularity 'ternary coupling' in the triplet, and lacks unified relation semantic constraint and calibratable confidence between feature extraction and decision scoring. Therefore, a complementary decision method capable of guaranteeing prediction accuracy and providing a unified decision basis for multiple reasoning results is needed to meet the requirements of actual satellite task arrangement and scheduling. Disclosure of Invention In view of the above, the invention provides a satellite observation demand generation method based on a triplet capsule network. Aiming at the problems that the existing satellite task element completion method lacks uniform decision basis in multiple reasoning results, the local interaction feature expression is insufficient, the stability of the final completion result is poor, and the like, a capsule network is adopted to take the 'vector length representation existence and direction representation gesture' as core ideas, so that the method is suitable for expressing the structural semantic mode. If the capsule vector is used for representing the existence and the 'gesture' attribute of the triplet element, and the reasonable feature combination strategy such as the triplet feature matrix and the like are combined with the scoring mechanism consistent with the decision to construct an end-to-end complement decision flow, the stability, the interpretability and the cross-relation consistency are expected to be improved. The method