A satellite observation requirement generation method based on a triple capsule network

Abstract

The application discloses a kind of satellite observation demand generation methods based on triplets capsule network, it is related to artificial intelligence and remote sensing application technical field.The application first generates Top-N candidate triplets by network inference through graph neural network;Second, construct triplets feature matrix, and generate corresponding interaction matrix by multi-scale convolution of relationship modulation;Then, by primary capsule generation, relationship modulation projection and dynamic routing mechanism, interaction matrix is converted into interpretable semantic capsule vector;Finally, in combination with relationship semantic modulation function, semantic capsule vector is scored and sorted in decision, and the candidate triplet with the highest compatibility score is output as the completion decision result.The application realizes the unified modeling from multiple inferences to final decision, improves the accuracy, stability and interpretability of element completion decision when generating demand, and can be widely applied to satellite demand planning and task planning system in the field of earth observation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and remote sensing application technology, specifically to knowledge graph representation learning and element completion decision-making methods, and particularly to a satellite observation demand generation method based on triplet capsule networks, which can be applied to satellite element completion decision-making systems in fields such as meteorological monitoring, resource exploration, environmental protection, and disaster early warning. Background Technology

[0002] Satellite observation mission requirements typically consist of multiple elements, including satellite target, payload, imaging mode, spatial resolution, priority, and time window. In actual operational processes, due to diverse sources, inconsistent recording standards, or information gaps and delays, key elements are frequently missing or inconsistently expressed in mission records. Existing regression model-based completion methods often rely on single-step inference, making it difficult to guarantee stable and accurate predictions. Furthermore, results obtained from multiple inferences lack unified decision-making criteria and confidence assessments, hindering reliable support for subsequent mission planning. These shortcomings directly impact mission orchestration and resource scheduling, leading to plan conflicts, decreased execution efficiency, and difficulty in assessing observation quality.

[0003] In existing research, the decision-making methods for completing missing elements mainly fall into the following categories:

[0004] 1. Convolutional Neural Networks (CNNs): Utilizing the translation invariance of local receptive fields and convolutional kernels, certain local interaction patterns can be extracted. However, in triple (h, r, t) scenarios, CNNs often extract channel features independently and then separate and interface them with the discriminator, making it difficult to explicitly model the fine-grained triple couplings of h–r, r–t, h–t, and their conjunctions. At the same time, the lack of a unified modulation mechanism aligned with "relational semantics" leads to insufficient consistency and interpretability across relations.

[0005] 2. Recurrent Neural Networks (LSTM): These networks use gating mechanisms to characterize temporal dependencies and can simulate sequential interactions among multiple elements. However, triples are symmetric / permutation-sensitive structures, and LSTMs require manually setting the input order, often compressing triple coupling into sequential binary transitions, which is insufficient for expressing synchronous interactions within triples. Furthermore, relational semantics often remain in implicit states, lacking explicit constraints consistent with the scoring head, thus affecting decision stability.

[0006] 3. Transformer: It has the advantage of global modeling, but its basic operation is mainly based on pairwise attention (token-to-token). For ternary coupling patterns that require relational constraints, additional design is often still needed. When the sample size is limited or relational semantics are not explicitly injected, the interpretability of attention weights and cross-relational stability are also weak.

[0007] While the aforementioned methods have advanced the direction of feature completion, their limitations remain apparent: first, they lack specialized modeling for the fine-grained "trinity coupling" within triples; and second, they lack unified semantic constraints and calibrable confidence levels between feature extraction and decision scoring. Therefore, there is an urgent need for a completion decision-making method that can both guarantee prediction accuracy and provide a unified decision-making basis for multiple inference results, in order to meet the needs of actual satellite mission orchestration and scheduling. Summary of the Invention

[0008] In view of this, this invention proposes a satellite observation requirement generation method based on triplet capsule networks. Addressing the problems of existing satellite mission element completion methods, such as the lack of unified decision-making basis for multiple inference results, insufficient expression of local interactive features, and poor stability of the final completion result, this invention adopts capsule networks with the core idea of ​​"vector length representing existence and direction representing attitude," which is suitable for expressing structured semantic patterns. If capsule vectors are used to represent the existence and "attitude" attributes of triplet elements, and combined with reasonable feature combination strategies such as triplet feature matrices and a scoring mechanism consistent with decision-making, an end-to-end completion decision-making process can be constructed, which is expected to improve stability, interpretability, and cross-relational consistency. The aim is to achieve high-precision prediction and stable decision-making for missing elements by combining the reasoning capabilities of graph neural networks with the interpretable representation mechanism of capsule networks.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A method for generating satellite observation requirements based on triplet capsule networks includes the following steps:

[0011] Step 1: Collect satellite observation mission requirement data containing missing elements, and perform deduplication, missing value placeholders, unit conversion and vocabulary alignment on the heterogeneous data to form a unified mission sample library;

[0012] Step 2: For the task sample library, use graph neural networks for network inference to obtain the feature representation of entities / relationships, and then construct multiple sets of test samples containing head entities and relations;

[0013] Step 3: Perform TOP-N reasoning on each test sample to obtain N candidate triples containing head entity, relation and tail entity, which are denoted as the candidate set corresponding to the test sample.

[0014] Step 4: For each candidate triplet in the candidate set, generate the corresponding interaction matrix by constructing a triplet feature matrix and introducing multi-scale convolution with relation modulation.

[0015] Step 5: Through primary capsule generation, relation modulation projection, and dynamic routing mechanisms, the interaction matrix is ​​transformed into semantic capsule vectors;

[0016] Step 6: Guide the scoring mechanism through the relational semantic modulation function, and score the compatibility of candidate triples by combining semantic capsule vectors;

[0017] Step 7: Calculate the loss function based on the scoring results of the candidate triplets, and iteratively execute steps 4 to 7 for backpropagation training until the loss function value converges. Save the model parameters to obtain the final triplet capsule network.

[0018] Step 8: For the head entities and relationships input in actual applications, combine the final triplet capsule network to execute steps 3 to 6, thereby outputting the candidate triplet with the highest compatibility score as the completion decision result, and obtaining the complete satellite observation requirement generation result.

[0019] Furthermore, the fields of the satellite observation mission requirement data in step 1 cover satellite target, payload, imaging mode, spatial resolution, priority, and time window elements.

[0020] Furthermore, the specific method for step 2 is as follows:

[0021] Step 201, ID mapping: Extract the task sample library into entity type and relation type, assign a unique integer ID to all entities and relations, and establish a mapping between entity2id and relation2id;

[0022] Step 202, Embedding Initialization: Entity and relation vectors are uniformly initialized using Xavier, with a vector dimension of d;

[0023] Step 203, Graph Neural Network Application: Directly call the GNN model. The model input is the graph topology adjacency information and initial embedding composed of entities and relations. The output is the feature representation of the entity / relation: h, r, t; where h is the head entity, r is the relation type, and t is the tail entity. Then, construct multiple sets of test samples (h, r, ?) containing head entities and relations, where ? indicates that the tail entity corresponding to the test sample is unknown.

[0024] Furthermore, the specific method for step 3 is as follows:

[0025] For each test sample, with a fixed head entity h and relation r, iterate through all entities in the entire entity set and score them sequentially as candidate tail entities. Sort the entities by score from highest to lowest, and retain the top N candidate tail entities, denoted as Top-N. This yields N candidate triples (h, r, t) corresponding to the current test sample. The scoring principle for the candidate tail entities is as follows:

[0026]

[0027] It is the Sigmoid function; The corresponding h, r, and t are the first Repeat step 3 to obtain multiple candidate sets corresponding to test samples, based on the element values ​​of each dimension.

[0028] Furthermore, step 4 is specifically implemented as follows:

[0029] N candidate triples (h, r, t) are processed in parallel:

[0030] Step 401: Concatenate the candidate triples (h, r, t) by channel to form a triple feature matrix X=[h; r; t]; X is a two-dimensional tensor of shape (3, d), and X is normalized.

[0031] Step 402: Introduce a one-dimensional convolution kernel with multiple scales s, and perform low-rank modulation on the convolution kernel using the relation vector r. Its parameterization form is as follows:

[0032]

[0033] Where s is the length of the one-dimensional convolution kernel; Modulate the convolution kernel for the relation at scale s; It is a static convolution kernel; represents the m-th basic kernel at scale s; M represents the number of basic kernels at scale s. The scalar modulation coefficients are generated from the relation vector r, and their calculation method is as follows:

[0034] ;

[0035] in, For the weight vector, For bias, It is a relation vector;

[0036] Step 403: Convolve X with the modulated convolution kernel, then pass it through a gated linear unit (GLU) and layer normalization to suppress noise and highlight key segments, thus obtaining the interaction matrix. :

[0037]

[0038]

[0039]

[0040] in, For the convolution output at scale s, set stride=1; The mapping matrix is ​​used to project the relation vector r into a space consistent with the convolution output channels; GLU is the gate function. The relation-related gated bias vector is added to the convolution output in GLU channel by channel to achieve dynamic modulation based on relational semantics; the convolution output is concatenated in the channel dimension to form a unified interaction matrix F.

[0041] Furthermore, step 5 is specifically implemented as follows:

[0042] Step 501, Primary Capsule Generation: The interaction matrix F is decomposed by row to obtain M sets of primary capsules. , i=1,2,3,……M;

[0043] Step 502, relation modulation projection:

[0044]

[0045] in, It is a static projection matrix. For relation modulation tensors, Here, j is the relation vector; j is the index of the higher-level capsule, j=1,2,3,…… ; This refers to the number of high-level capsules;

[0046] Step 503, Dynamic Routing Mechanism: For the j-th high-level capsule, the predicted vectors from each primary capsule are converged to form... And then non-linearly compressed by squash to Routing logits are updated iteratively based on consistency, and the entire routing process is represented by the following formula:

[0047]

[0048]

[0049]

[0050]

[0051]

[0052] in The projection matrix modulated by the relation vector r; Let be the prediction vector of the i-th primary capsule for the j-th higher-level capsule; For the route logit, it represents the matching prior between the i-th primary capsule and the j-th higher-level capsule, with an initial value of 0; Let be the coupling coefficient, satisfying ; The weighted sum of all predicted vectors; For the j-th high-level capsule, the length is compressed to the (0,1) interval; The capsule compression function is defined as follows: , Represents the vector norm; Represents the vector dot product. Used for updating ;

[0053] Iteratively execute the routing process until... and Both converge;

[0054] Step 504, based on Each high-level capsule output constructs the corresponding semantic capsule vector. ; .

[0055] Furthermore, step 6 is performed as follows:

[0056] Step 601, Relational Semantic Modulation Function: Generates a modulation vector of consistent dimension from the relation vector r, emphasizing the semantic subspace related to r:

[0057]

[0058] in, The weight matrix for relation modulation is used to project r into a space of the same dimension as the score vector; It is the bias vector; This is a relational semantic modulation function used to assign differentiated weights to different dimensions in the scoring function; For non-linear activation functions, defined as follows: ;

[0059] Step 602, Scoring Function and Ranking: Perform compatibility scoring with diagonal modulation on candidate triples:

[0060]

[0061] in, This indicates element-wise multiplication, i.e., multiplying by elements. Multiply by each corresponding dimension of g(r); Indicates vector transpose; This is the compatibility score of the candidate triple (h,r,t).

[0062] Furthermore, the specific method for calculating the loss function in step 7 is as follows:

[0063] Loss calculation: For a given test sample (h, r, ?), let the set of candidate tail entities be T={t1, t2,...,t...} N}, where the only correct tail entity is t + The loss function is then defined as:

[0064]

[0065] in, This is a scaling factor used to adjust the sensitivity of score differences between candidates in softmax.

[0066] Due to the adoption of the above technical solution, the beneficial effects of this invention compared with the prior art are as follows:

[0067] Compared to existing technologies, the core advantage of this invention lies in achieving unified modeling from multiple inferences to the final decision. Specifically, this is reflected in two aspects: First, this invention explicitly captures the fine-grained interaction relationships within triples through a feature matrix and convolutional interaction mechanism, effectively compensating for the shortcomings of traditional feature modeling methods. Second, and more importantly, this invention introduces a dynamic routing and relationship modulation scoring mechanism from capsule networks based on the candidate ranking results, ensuring that the final decision result has both higher accuracy and interpretable confidence information. This mechanism makes the completion process no longer dependent on the results of a single inference, but achieves stable, reliable, and traceable completion decisions. Attached Figure Description

[0068] Figure 1 This is an overall flowchart of a satellite observation requirement generation method based on a triplet capsule network in an embodiment of the present invention. Detailed Implementation

[0069] The invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0070] A method for generating satellite observation requirements based on triplet capsule networks, such as Figure 1 As shown, it includes the following steps:

[0071] Step 1: Collect satellite observation mission requirement data containing missing elements, and perform deduplication, missing value placeholders, unit conversion and vocabulary alignment on the heterogeneous data to form a unified mission sample library;

[0072] Step 2: For the task sample library, use graph neural networks for network inference to obtain the feature representation of entities / relationships, and then construct multiple sets of test samples containing head entities and relations;

[0073] Step 3: Perform TOP-N reasoning on each test sample to obtain N candidate triples containing head entity, relation and tail entity, which are denoted as the candidate set corresponding to the test sample.

[0074] The data preparation and multiple inference to generate candidate results aim to generate Top-N candidate tail entities from satellite observation elements through multiple inferences and form a parallel processing list, thereby providing standardized input and a constrained candidate space for subsequent triple feature matrix construction, convolutional interaction, capsule routing and relational semantic decision scoring.

[0075] Step 4: For each candidate triplet in the candidate set, generate the corresponding interaction matrix by constructing a triplet feature matrix and introducing multi-scale convolution with relation modulation.

[0076] This step aims to extract and enhance fine-grained interaction patterns between head entities, relations, and candidate tail entities by constructing a triplet feature matrix and introducing multi-scale convolution with relation modulation, thereby generating a unified interaction feature sequence F, which provides structured input for subsequent capsule generation and dynamic routing.

[0077] Step 5: Through primary capsule generation, relation modulation projection, and dynamic routing mechanisms, the interaction matrix is ​​transformed into semantic capsule vectors;

[0078] This step aims to transform the feature sequence obtained from convolutional interactions into interpretable semantic capsule vectors through primary capsule generation, relational modulation projection, and dynamic routing mechanisms, thereby achieving stable high-level representations at the candidate triplet level and providing a unified input for subsequent scoring and decision ranking.

[0079] Step 6: Guide the scoring mechanism through the relational semantic modulation function, and score the compatibility of candidate triples by combining semantic capsule vectors;

[0080] Step 7: Calculate the loss function based on the scoring results of the candidate triplets, and iteratively execute steps 4 to 7 for backpropagation training until the loss function value converges. Save the model parameters to obtain the final triplet capsule network.

[0081] Step 8: For the head entities and relationships input in actual applications, combine the final triplet capsule network to execute steps 3 to 6, thereby outputting the candidate triplet with the highest compatibility score as the completion decision result, and obtaining the complete satellite observation requirement generation result.

[0082] Furthermore, the fields of the satellite observation mission requirement data in step 1 cover satellite target, payload, imaging mode, spatial resolution, priority, and time window elements.

[0083] Furthermore, the specific method for step 2 is as follows:

[0084] Step 201, ID mapping: Extract the task sample library into entity type and relation type, assign a unique integer ID to all entities and relations, and establish a mapping between entity2id and relation2id;

[0085] Step 202, Embedding Initialization: The entity and relation vectors are uniformly initialized using Xavier, with a vector dimension of d; in this embodiment, d is set to 64. This initialization is used to provide a stable starting point for subsequent training;

[0086] Step 203, Graph Neural Network Application: Directly call the GNN model. The model input is the graph topology adjacency information and initial embedding composed of entities and relations. The output is the feature representation of the entity / relation: h, r, t; where h is the head entity, r is the relation type, and t is the tail entity. Then, construct multiple sets of test samples (h, r, ?) containing head entities and relations, where ? indicates that the tail entity corresponding to the test sample is unknown.

[0087] Furthermore, the specific method for step 3 is as follows:

[0088] For each test sample, with a fixed head entity h and relation r, iterate through all entities in the entire entity set and score them sequentially as candidate tail entities. Sort the entities by score from highest to lowest, and retain the top N candidate tail entities, denoted as Top-N. This yields N candidate triples (h, r, t) corresponding to the current test sample. The scoring principle for the candidate tail entities is as follows:

[0089]

[0090] It is the Sigmoid function; The corresponding h, r, and t are the first Repeat step 3 to obtain multiple candidate sets corresponding to test samples. In this embodiment, N is 20.

[0091] Furthermore, step 4 is specifically implemented as follows:

[0092] N candidate triples (h, r, t) are processed in parallel:

[0093] Step 401: Concatenate the candidate triples (h, r, t) by channel to form a triple feature matrix X=[h; r; t]; X is a two-dimensional tensor of shape (3, d), and X is normalized.

[0094] Step 402: Introduce a one-dimensional convolution kernel with multiple scales s, and perform low-rank modulation on the convolution kernel using the relation vector r. Its parameterization form is as follows:

[0095]

[0096] Where s is the length of the one-dimensional convolution kernel; Modulate the convolution kernel for the relation at scale s; It is a static convolution kernel; represents the m-th basic kernel at scale s; M represents the number of basic kernels at scale s. The scalar modulation coefficients are generated from the relation vector r, and their calculation method is as follows:

[0097] ;

[0098] in, For the weight vector, For bias, It is a relation vector;

[0099] In this embodiment, s=3 and M=5; by combining a finite number of base kernels with scalar linear combinations, the convolution kernel can continuously deform with relational semantics, thereby exhibiting differentiated responses on different candidates;

[0100] Step 403: Convolve X with the modulated convolution kernel, then pass it through a gated linear unit (GLU) and layer normalization to suppress noise and highlight key segments, thus obtaining the interaction matrix. :

[0101]

[0102]

[0103]

[0104] in, For the convolution output at scale s, set stride=1; The mapping matrix is ​​used to project the relation vector r into a space consistent with the convolution output channels; GLU is the gate function. The relation-related gated bias vector is added channel-wise to the convolutional output in the GLU to achieve dynamic modulation based on relational semantics. The convolutional output is concatenated channel-wise to form a unified interaction matrix F. The 2M setting is for "half-gating" in the GLU. The feature matrix convolution uses shared weights and mini-batch parallelism. The convolutional output is concatenated channel-wise to form a unified F.

[0105] Furthermore, step 5 is specifically implemented as follows:

[0106] Step 501, Primary Capsule Generation: The interaction matrix F is decomposed by row to obtain M sets of primary capsules. , i=1,2,3,……M;

[0107] Step 502, relation modulation projection:

[0108]

[0109] in, It is a static projection matrix. For relation modulation tensors, Here, j is the relation vector; j is the index of the higher-level capsule, j=1,2,3,…… ; This refers to the number of high-level capsules;

[0110] Step 503, Dynamic Routing Mechanism: For the j-th high-level capsule, the predicted vectors from each primary capsule are converged to form... And then non-linearly compressed by squash to Routing logits are updated iteratively based on consistency, and the entire routing process is represented by the following formula:

[0111]

[0112]

[0113]

[0114]

[0115]

[0116] in The projection matrix modulated by the relation vector r; Let be the prediction vector of the i-th primary capsule for the j-th higher-level capsule; For the route logit, it represents the matching prior between the i-th primary capsule and the j-th higher-level capsule, with an initial value of 0; Let be the coupling coefficient, satisfying ; The weighted sum of all predicted vectors; For the j-th high-level capsule, the length is compressed to the (0,1) interval; The capsule compression function is defined as follows: , Represents the vector norm; Represents the vector dot product. Used for updating ;

[0117] Iteratively execute the routing process until... and All convergences are achieved; this routing process typically converges after 2–3 iterations.

[0118] Step 504, based on Each high-level capsule output constructs the corresponding semantic capsule vector. ; .

[0119] Furthermore, step 6 is performed as follows:

[0120] Step 601, Relational Semantic Modulation Function: Generates a modulation vector of consistent dimension from the relation vector r, emphasizing the semantic subspace related to r:

[0121]

[0122] in, The weight matrix for relation modulation is used to project r into a space of the same dimension as the score vector; It is the bias vector; This is a relational semantic modulation function used to assign differentiated weights to different dimensions in the scoring function; For non-linear activation functions, defined as follows: ;

[0123] Step 602, Scoring Function and Ranking: Perform compatibility scoring with diagonal modulation on candidate triples:

[0124]

[0125] in, This indicates element-wise multiplication, i.e., multiplying by elements. Multiply by each corresponding dimension of g(r); Indicates vector transpose; This is the compatibility score of the candidate triple (h,r,t);

[0126] Furthermore, the specific method for calculating the loss function in step 7 is as follows:

[0127] Loss calculation: For a given test sample (h, r, ?), let the set of candidate tail entities be T={t1, t2,...,t...} N}, where the only correct tail entity is t + The loss function is then defined as:

[0128]

[0129] in, This is a scaling factor used to adjust the sensitivity of score differences between candidates in softmax, and the correct tail entity t + The correct tail entity is pre-selected by humans.

[0130] This loss function directly maximizes the correct tail entity t by normalizing the entire candidate set. +The normalized probabilities are used to ensure that the training objective aligns with the final Top-N ranking decision. Unlike pointwise or pairwise losses, this design avoids the problem of inconsistency between the training objective and the inference objective, thus improving the reliability of the completion decision.

[0131] Those skilled in the art will recognize that the described embodiments are intended to help readers understand the principles of the invention and should be understood as not limiting the scope of protection of the invention to the described embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A method for generating satellite observation requirements based on triplet capsule networks, characterized in that, Includes the following steps: Step 1: Collect satellite observation mission requirement data containing missing elements, and perform deduplication, missing value placeholders, unit conversion and vocabulary alignment on the heterogeneous data to form a unified mission sample library; Step 2: For the task sample library, use graph neural network for network inference to obtain feature representations of entity / relation types, and then construct multiple sets of test samples containing head entities and relation types; Step 3: Perform TOP-N reasoning on each test sample to obtain N candidate triples containing head entity, relation type and tail entity, which are denoted as the candidate set corresponding to the test sample. Step 4: For each candidate triplet in the candidate set, generate the corresponding interaction matrix by constructing a triplet feature matrix and introducing multi-scale convolution with relation modulation. Step 5: Through primary capsule generation, relation modulation projection, and dynamic routing mechanisms, the interaction matrix is ​​transformed into semantic capsule vectors; Step 6: Guide the scoring mechanism through the relational semantic modulation function, and score the compatibility of candidate triples by combining semantic capsule vectors; Step 7: Calculate the loss function based on the scoring results of the candidate triplets, and iteratively execute steps 4 to 7 for backpropagation training until the loss function value converges. Save the model parameters to obtain the final triplet capsule network. Step 8: For the head entity and relation type input in actual application, combine the final triplet capsule network to execute steps 3 to 6, so as to output the candidate triplet with the highest compatibility score as the completion decision result, and obtain the complete satellite observation requirement generation result.

2. The satellite observation requirement generation method based on triplet capsule networks according to claim 1, characterized in that, The fields in the satellite observation mission requirement data in Step 1 cover satellite target, payload, imaging mode, spatial resolution, priority, and time window elements.

3. The satellite observation demand generation method based on triplet capsule networks according to claim 1, characterized in that, The specific method for step 2 is as follows: Step 201, ID mapping: Extract the task sample library into entity types and relation types, assign a unique integer ID to all entities and relation types, and establish a mapping between entity2id and relation2id; Step 202, Embedding Initialization: Entity and relation types are uniformly initialized using Xavier, with a vector dimension of d; Step 203, Graph Neural Network Application: Directly call the GNN model. The model input is the graph topology adjacency information and initial embedding composed of entities and relation types. The output is the feature representation of the entity / relation type: h, r, t; where h is the head entity, r is the relation type, and t is the tail entity. Then, construct multiple sets of test samples (h, r, ?) containing head entities and relation types, where ? indicates that the tail entity corresponding to the test sample is unknown.

4. The satellite observation requirement generation method based on triplet capsule networks according to claim 1, characterized in that, The specific method for step 3 is as follows: For each test sample, with a fixed head entity h and relation type r, iterate through all entities in the entire entity set and score them sequentially as candidate tail entities. Sort the entities by score from highest to lowest, and retain the top N candidate tail entities, denoted as Top-N. This yields N candidate triples (h, r, t) corresponding to the current test sample, where t is the tail entity. The scoring principle for the candidate tail entities is as follows: It is the Sigmoid function; d represents the dimensions of h, r, and t. The corresponding h, r, and t are the first Repeat step 3 to obtain multiple candidate sets corresponding to test samples, based on the element values ​​of each dimension.

5. The satellite observation demand generation method based on triplet capsule networks according to claim 1, characterized in that, The specific method for step 4 is as follows: N candidate triples (h, r, t) are processed in parallel, where h is the head entity, r is the relation type, and t is the tail entity. Step 401: Concatenate the candidate triples (h, r, t) by channel to form a triple feature matrix X=[h; r; t]; X is a two-dimensional tensor of shape (3, d), where d is the dimension of h, r, and t, and X is normalized. Step 402: Introduce a one-dimensional convolution kernel with multiple scales s, and perform low-rank modulation on the convolution kernel using relation type r. Its parameterization form is as follows: Where s is the length of the one-dimensional convolution kernel; Modulate the convolution kernel for the relation at scale s; It is a static convolution kernel; represents the m-th basic kernel at scale s; M represents the number of basic kernels at scale s. The scalar modulation coefficients are generated from the relation type r, and their calculation method is as follows: ； in, For the weight vector, For bias, It is a relation type; Step 403: Convolve X with the modulated convolution kernel, then pass it through a gated linear unit (GLU) and layer normalization to suppress noise and highlight key segments, thus obtaining the interaction matrix. : in, For the convolution output at scale s, set stride=1; The mapping matrix is ​​used to project the relation type r into a space consistent with the convolution output channels; GLU is the gate function. The gated bias vector, which is related to the relation type, is added to the convolution output in GLU channel by channel to achieve dynamic modulation based on relation semantics; the convolution output is concatenated in the channel dimension to form a unified interaction matrix F.

6. The satellite observation demand generation method based on triplet capsule networks according to claim 5, characterized in that, The specific method for step 5 is as follows: Step 501, Primary Capsule Generation: The interaction matrix F is decomposed by row to obtain M sets of primary capsules. , i=1,2,3,……M; Step 502, relation modulation projection: in, It is a static projection matrix. For relation modulation tensors, The relation type is j; j is the high-level capsule index, j=1,2,3,…… ; This refers to the number of high-level capsules; Step 503, Dynamic Routing Mechanism: For the j-th high-level capsule, the predicted vectors from each primary capsule are converged to form... And then non-linearly compressed by squash to Routing logits are updated iteratively based on consistency, and the entire routing process is represented by the following formula: in This is the projection matrix modulated by relation type r; Let be the prediction vector of the i-th primary capsule for the j-th higher-level capsule; For the route logit, it represents the matching prior between the i-th primary capsule and the j-th higher-level capsule, with an initial value of 0; Let be the coupling coefficient, satisfying ; The weighted sum of all predicted vectors; For the j-th high-level capsule, the length is compressed to the (0,1) interval; The capsule compression function is defined as follows: , Represents the vector norm; Represents the vector dot product. Used for updating ; Iteratively execute the routing process until... and Both converge; Step 504, based on Each high-level capsule output constructs the corresponding semantic capsule vector. ; .

7. The satellite observation demand generation method based on triplet capsule networks according to claim 6, characterized in that, The specific method for step 6 is as follows: Step 601, Relational Semantic Modulation Function: Generates a modulation vector of consistent dimension from relation type r, emphasizing the semantic subspace related to r: in, The weight matrix for relation modulation is used to project r into a space of the same dimension as the score vector; It is the bias vector; This is a relational semantic modulation function used to assign differentiated weights to different dimensions in the scoring function; For non-linear activation functions, defined as follows: ; Step 602, Scoring Function and Ranking: Perform compatibility scoring with diagonal modulation on candidate triples: in, This indicates element-wise multiplication, i.e., multiplying by elements. Multiply by each corresponding dimension of g(r); Indicates vector transpose; This is the compatibility score of the candidate triple (h,r,t).

8. The satellite observation demand generation method based on triplet capsule networks according to claim 7, characterized in that, The specific method for calculating the loss function in step 7 is as follows: Loss Calculation: For a given test sample (h, r, ?), where h is the head entity, r is the relation type, and ? indicates that the tail entity corresponding to this test sample is unknown, let the candidate tail entity set be T={t1, t2,...,t...} N }, where the only correct tail entity is t + The loss function is then defined as: in, This is a scaling factor used to adjust the sensitivity of score differences between candidates in softmax.

Images