Earth observation method based on temporal knowledge graph interpretable reasoning technology

Abstract

A method for Earth observation based on interpretable reasoning technology for time-series knowledge graphs uses closed-loop path rules to collect potential multi-hop closed-loop path rules from a time-series knowledge graph in an offline phase. A network model consisting of a rule filter and a rule reasoner is constructed and trained based on an Earth observation mission process dataset. In an online phase, the trained network model is used to predict real-time data to be measured. Based on the closed-loop path rules, missing processes in the Earth observation mission are completed. This method can meet the interpretability requirements of Earth observation mission processes and has broad application prospects.

Description

Technical Field

[0001] The present invention relates to a technology in the field of knowledge reasoning in knowledge graphs, specifically a ground observation method based on explainable reasoning technology of temporal knowledge graphs. Background Art

[0002] The Earth observation mission process consists of various tasks that must be executed sequentially to complete the Earth observation, maintaining a strict temporal relationship. With the rapid development of aerospace technology, Earth observation mission processes are becoming increasingly complex. Temporal knowledge graphs excel at storing knowledge with temporal relationships. They use the quadruple (head entity, relationship, tail entity, and timestamp) of the temporal knowledge graph to store Earth observation mission processes. Inference methods are then used to further analyze and mine the knowledge implicit in these processes. For example, given a mission requirement to observe the South China Sea at a specific moment, the rules implicit in the temporal knowledge in the knowledge graph can be used to provide a decision-making solution, select the most appropriate satellite to observe that area, and complete the mission process. This will undoubtedly provide unprecedented knowledge support for Earth observation.

[0003] The Earth observation mission process places high demands on the reliability of the reasoning process. Reasoning methods must explain how the inference results were obtained. Only on this basis can potential errors be identified and higher-quality inference results be provided. Current mainstream knowledge graph reasoning methods use embedding learning algorithms to mine implicit features of entities and relationships. However, due to the lack of intuitive understanding of implicit features, the interpretability of reasoning cannot be guaranteed. Traditional graph traversal mining algorithms based on explicit logical rules have good interpretability, but due to the need to traverse the entire knowledge graph, i.e., the search scope is large, resulting in significant efficiency deficiencies. Summary of the Invention

[0004] In response to the above-mentioned deficiencies in the prior art, the present invention proposes an earth observation method based on interpretable reasoning technology of time series knowledge graph. The obtained earth observation task process data is preprocessed to obtain rules in the time series knowledge graph, and the obtained rules are screened to mine implicit knowledge in the time series knowledge graph. Reasoning is performed based on the rules with higher scores after screening. This can ensure the interpretability of the reasoning method and meet the requirements of the earth observation task process for interpretable reasoning, and has broad application prospects.

[0005] The present invention is achieved through the following technical solutions:

[0006] The present invention relates to an earth observation method based on interpretable reasoning technology of temporal knowledge graph. In the offline stage, closed-loop path rules are used to collect potential multi-hop closed-loop path rules in the temporal knowledge graph, and a network model including a rule filter and a rule reasoner is constructed. The network model is trained based on an earth observation task process data set. In the online stage, the trained network model is used to predict real-time test data, and the missing processes in the earth observation task are completed according to the closed-loop path rules.

[0007] When there are n pieces of information (r1, r2, ..., r n ), then the closed-loop path rule is used to infer that there is a relationship r between x and y at a certain moment in the future; the closed-loop path rule [(r, t, d), (r1, t1, d1), (r2, t2, d2), ..., (r n ,t n ,d n )], including a head consisting of Horn clauses and a body of multiple quadruples of length n, specifically: r(x,y,t,d)←r1(x,z1,t1,d1)∧r2(z1,z2,t2,d2)∧···∧r n (z n-1 ,y,t n ,d n ), where n is the length of the closed-loop path rule, x, y, z i Represents an entity in the knowledge graph, r i Represents the relationship in the knowledge graph, t i Indicates the start timestamp of the relationship, and t>t1>t2>...>t n , d i Indicates the duration of the relationship.

[0008] For example, in chronological order, user A uploads disaster monitoring requirements to demand planning department B, which then uploads a requirements list to control system department C. Control system department C then uploads a task plan to information system department D. Information system department D will then transmit the data product to user A at some point in the future, thus forming a three-hop closed-loop path rule:

[0009] The network model includes: a rule filter unit and a rule reasoner unit, wherein: the rule filter unit scores the closed-loop path rules corresponding to each rule head according to the feature information of the closed-loop path rules contained in the temporal knowledge graph, and the top q closed-loop path rules in each rule head are used as the result obtained by the rule filter unit; the rule reasoner unit processes the quadruple in the temporal knowledge graph according to the closed-loop path rule information filtered out by the rule filter unit, obtains the loss function value corresponding to the closed-loop path rule, and trains the network model according to the loss function value.

[0010] The rule filter unit calculates the scores of all closed-loop path rules and sorts them, obtains the rule with the highest score as the screening result, and completes the screening of the rules. It specifically includes: a rule encoder, a transformer encoder for capturing rule features, and a sub-network for converting features into rule scores, wherein: the rule encoder encodes and fuses the attributes contained in the closed-loop path rule according to the closed-loop path rule to obtain encoding information; the transformer encoder performs feature extraction processing based on the encoding information to obtain a feature matrix; the sub-network converts the feature matrix of size (n+1)×m into a score to obtain the score of the closed-loop path rule.

[0011] The coding information includes: the coding of the relationship r, the coding of the start timestamp t, the coding of the duration d and the coding of the relationship sequence position, wherein: the coding of the relationship r, the coding of the start timestamp t and the coding of the duration d are all encoded by using an embedding function to encode different values ​​to generate a one-dimensional space vector of size 1×m, where m is the dimension size of the coding; the coding of the relationship sequence position is based on the relationship r in the closed-loop path rule. i The order of appearance is encoded, which is consistent with the position encoding of the transformer, that is, Where: p i (t) is the i-th dimension value of the position encoding vector at position t. Similarly, this encoding also generates a one-dimensional space vector of size 1×m.

[0012] The transformer encoder has 8 attention heads. By adding the corresponding positions of the four encodings, each tuple (r i ,t i ,d i ), then an n-hop closed-loop path rule (including the head and body of the rule) will generate a feature matrix of (n+1)×m size.

[0013] The sub-network includes: a fully connected layer, a ReLU activation layer and a dropout regularization layer.

[0014] The rule reasoner unit improves the existing neural network model based on translation invariance, and includes: an entity encoder, a relationship encoder, a negative sample generator and a transfer model-based reasoner, wherein: the entity encoder and the relationship encoder generate entity encoding information and relationship encoding information respectively according to the quadruple data in the temporal knowledge graph; the negative sample generator replaces the tail entity in the quadruple by uniform random negative sampling according to the quadruple information in the temporal knowledge graph to obtain the negative sample corresponding to each quadruple; the reasoner performs network training processing according to the n-hop closed-loop path rule with a higher score, and continuously optimizes through the rule filter unit and the rule reasoner unit to obtain the n-hop closed-loop path rule with higher quality in the temporal knowledge graph.

[0015] The rule reasoner unit is trained in the following manner: the transferred head entity encoding information and tail entity encoding information are compared to obtain a loss function, and the parameters in the rule filter unit and the rule reasoner unit are updated through reverse propagation.

[0016] The entity coding is to adopt the idea of ​​dynamic entity and regard the entity coding information as a continuous function that changes with time, that is, Among them: the sin function is used to retain the information of the entity's periodic changes over time, and the tanh function is used to retain the information of the entity's monotonically changing over time. Used to hold constant information about an entity.

[0017] The relationship encoding information includes: the encoding of the relationship r, the encoding of the start timestamp t, and the encoding of the duration d. These three encodings are similar to the rule encoder in the rule filter unit. The relationship encoder fuses the three encodings according to the set weights to obtain the relationship encoding information, that is, V(r, t, d) = ω r +θ1·ω t +θ2·ω d .

[0018] The replacement means that if the number of entities is N and the number of negative samples is k, through negative sampling, each iteration only needs to train k+1 of them to improve the computational efficiency of the model.

[0019] The reasoner based on the transfer model is designed according to the principle of translation invariance, that is, the head entity should be approximately equal to the tail entity after the relationship transformation; according to the n-hop closed loop path rule, the head entity should be approximately equal to the tail entity after the n relationships in the rule body are transferred in sequence, that is, V(e h ,t)×V(r1,t1,d1)×V(r2,t2,d2)×...×V(r n ,t n ,d n )≈V(et ,t).

[0020] The network training process specifically includes:

[0021] Step 1: Divide the Earth observation mission process time series knowledge graph data sample set into a training set, a validation set, and a test set.

[0022] Step 2: Set the maximum value of n in the n-hop closed-loop path rule, and recursively obtain the n-hop closed-loop path rule corresponding to each rule r based on the test set.

[0023] Step 3: Use the rule filter unit to filter out the n-hop closed-loop path rules with higher scores corresponding to each rule r.

[0024] Step 4: In the training set, use the rule reasoner unit for each data according to the n-hop closed-loop path rules with higher scores that are screened out to obtain the loss function, and finally pass the update parameters in the rule filter unit and the rule reasoner unit in reverse.

[0025] Step 5: Continuously iterate steps 3 to 4, select the network with the highest evaluation index in the validation set as the trained interpretable reasoning method, and finally obtain high-quality n-hop closed-loop path rules in the temporal knowledge graph.

[0026] The present invention relates to a system for implementing the above-mentioned method, comprising: a rule filter unit and a rule reasoner unit, wherein: the rule filter unit scores the closed-loop path rules corresponding to each rule head according to the characteristic information of the closed-loop path rules contained in the temporal knowledge graph, and filters out the top q closed-loop path rules in each rule head; the rule reasoner unit processes the quadruple in the temporal knowledge graph according to the closed-loop path rules obtained by filtering, obtains the loss function value corresponding to the closed-loop path rule, and trains the network model according to the loss function value.

[0027] Technical Effects

[0028] This invention aggregates multi-scale features in a temporal knowledge graph through a neural network built with a rule-based reasoner, reducing the search space and computational complexity. Compared to embedded learning algorithms, this invention uses a rule filter to select high-scoring closed-loop path rules in the current temporal knowledge graph. Simultaneously, the rule-based reasoner performs reasoning based on these selected rules, resulting in a more reliable and interpretable inference result. This invention incorporates start timestamps and duration information into the encoding of entities and relationships, facilitating the capture of frequent periodic events in Earth observation missions.

[0029] This method combines the advantages of logical rule algorithms and embedded learning algorithms to ensure the interpretability of the reasoning method and effectively mine implicit knowledge from the fine-grained, frequently occurring data generated by Earth observation mission processes. A neural network approach is used to aggregate multi-scale features in the temporal knowledge graph and evaluate the rules selected in the previous step. Furthermore, a transformer encoder is used to select existing rules, ensuring the interpretability of the reasoning method. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Flowchart for realizing the present invention;

[0031] Figure 2 A network structure diagram for interpretable reasoning methods;

[0032] Figure 3 An example diagram of dynamic entity encoding at time t;

[0033] Figure 4 Example graph encoding a relationship at time t with duration d. DETAILED DESCRIPTION

[0034] like Figure 1 As shown, this embodiment relates to an explainable reasoning method for a time series knowledge graph of an Earth observation mission process, specifically including:

[0035] Step A: Number the entities, relationships, and time information contained in the Earth Observation Mission Process dataset. The inverse relationship of each relationship is stored as a new relationship in the dataset. The inverse relationship of a four-tuple (head entity, relationship, tail entity, time) is (tail entity, inverse of relationship, head entity, time). Reasoning about the second entity simultaneously completes two different reasoning tasks: reasoning about the head entity or the tail entity. The dataset is then divided into training, validation, and verification sets in a ratio of 0.6:0.2:0.2.

[0036] Step B: Use a recursive algorithm to search for closed-loop path rules from the training set. To improve search efficiency, set the maximum value of n in the closed-loop path rule to 5. To limit the number of rules found, set the maximum difference between the start time of the rule header and the rule body in the closed-loop path rule to 24 hours, and set the minimum unit of the relationship duration to 2 hours. Each relationship in the knowledge graph serves as a rule header, and the searched rule body is stored in the corresponding rule header to generate a closed-loop path rule set.

[0037] Step C: Set the learning rate to 0.001, the number of iterations to 200, the batch size for each training round to 512, the optimizer to the Adam optimizer, and the dropout value to 0.5. By using the rule filter, for each relationship in the knowledge graph, select the closed-loop path rules with the top 10 scores from the closed-loop path rule set. Among them: the input closed-loop path rule length is 6, including 1 rule head and 5 rule bodies, the number of attention mechanism heads is 8, and the encoding generated by the rule encoder needs to be a multiple of 8, and the size is set to 128. After encoding, the input matrix size is 6×128. After passing through the transformer encoder, the output feature matrix size is 6×128. Finally, it passes through the sub-network and two fully connected layers to obtain a 1×1 value as the score of the current closed-loop path rule.

[0038] In step D, the selected rules are trained using a rule-based reasoner. During each round of training, 10 negative samples are randomly generated for each data point based on the training set samples. This reduces the algorithm's computational complexity and enriches the training set samples to improve the generalization of the inference method. By iterating steps C and D, the accuracy of the inference algorithm is verified using a validation dataset, screening for higher-quality rules and achieving the highest interpretable inference method with the highest evaluation metrics. Finally, the inference results are obtained using the selected rules and the trained rule-based reasoner network, using the quadruple to be inferred from the test set as input.

[0039] like Figure 2 As shown, the network structure of the reasoning method includes two parts: a rule filter and a rule reasoner.

[0040] The rule filter utilizes a transformer encoder, including a rule encoder, a transformer encoder for capturing rule features, and a subnetwork for converting features into rule scores. The searched closed-loop path rule is input into the rule encoder to obtain encoded information for the closed-loop path rule. This encoded information incorporates the relational order and temporal information contained in the closed-loop path rule and serves as the input to the transformer encoder. The output of the transformer encoder serves as the input to the subnetwork, ultimately obtaining a score for the closed-loop path rule. The rule filter calculates and ranks the scores of all closed-loop path rules, obtaining the rules with the highest scores to complete the rule screening.

[0041] The rule encoder contains four parts of encoding: encoding of the relation r, encoding of the relation order position, encoding of the start timestamp t, and encoding of the duration d. These four parts of encoding can retain the relation order and time information implicit in the rule to the greatest extent. Among them, the encoding of the relation r, the encoding of the start timestamp t, and the encoding of the duration d all use the embedding function in the pytorch deep learning framework to encode different values ​​to generate a one-dimensional space vector of size 1×128. The encoding process of the relation order position is: Where: p i (t) is the i-th dimension of the positional encoding vector at position t. This encoding also generates a 1×128 one-dimensional spatial vector. Adding the corresponding positions of the four encoded parts, a 5-hop closed-loop path rule generates a 6×128 matrix. After feature extraction by the transformer encoder, the output is a 6×128 feature matrix, where the number of attention heads is 8.

[0042] The subnetwork used to convert features into rule scores consists of a fully connected layer, a Relu activation layer, and a dropout regularization layer. The score of the current closed-loop path rule is calculated through the subnetwork. The subnetwork sequence is: 6×128-dimensional to 128-dimensional fully connected layer -> Relu activation function -> dropout regularization layer -> 128-dimensional to 1-dimensional fully connected layer -> Relu activation function. For each relationship in the knowledge graph, the top 10 closed-loop path rules with the highest scores are selected from the closed-loop path rule set to serve as the reasoning basis for the rule reasoner.

[0043] The rule reasoner includes: an encoder for obtaining entity encoding, an encoder for obtaining relationship encoding, a negative sample generator and a reasoner based on a transfer model, wherein: the quadruple data in the temporal knowledge graph is used as the input of the entity and relationship encoder to obtain entity and relationship encoding vectors of size 1×128, and a 5-hop closed-loop path rule with a higher score is obtained from the rule filter, and then the reasoner is used, according to the rules of the transfer model, the encoding of the head entity is transferred by using the encoding of the rule in the closed-loop path rule, the transferred head entity encoding information and the tail entity encoding information are compared to obtain the loss function, and the parameters in the rule filter and the rule reasoner are updated through reverse transfer, and the rule filter and the rule reasoner are continuously optimized to obtain a high-quality n-hop closed-loop path rule in the temporal knowledge graph.

[0044] like Figure 3 As shown, the entity encoder adopts the idea of ​​dynamic entity, regards the encoding information of the entity as a continuous function that changes with time, and uses the formula The encoding is divided into three parts: information that changes periodically over time represented by the sin function, information that changes monotonically over time represented by the tanh function, and constant information that does not change over time. Among them: time information is calculated by splicing "year, month, day, hour, minute, second" into a 1×128 vector, and ultimately each entity will be mapped to a 1×128 space vector. Figure 4 As shown. The relational encoder is based on V(r, t, d) = ω r +θ1·ω t +θ2·ω d , incorporating the start time and duration information, each relationship will eventually be mapped into a 1×128 space vector.

[0045] The negative sample generator replaces the tail entity in the quadruple by uniform random negative sampling. The negative sampling size is set to 10, which can greatly improve the computational efficiency of the model. The transfer model-based reasoner is designed according to the principle of translation invariance, that is, V(e h ,t)×V(r1,t1,d1)×V(r2,t2,d2)×...×V(r n ,t n ,d n )≈V(e t ,t), calculate the loss function by comparing the distance between the left and right sides of the formula in: is the entity inferred by applying the closed-loop path rule, The negative samples are obtained by negative sampling. The network parameters of the inference method are optimized by back propagation using the loss function.

[0046] This example prioritizes the use of a time series dataset of the Earth observation mission process. This dataset simulates the Earth observation mission process from October 15, 2022, to November 30, 2022. It contains 877 entities, 9 time series relationships, and a time granularity of 1 second. The training set contains 68,361 time series data items, the validation set contains 22,787 time series data items, and the test set contains 22,787 time series data items, for a total of 113,935 time series data items. The inference results are evaluated using two criteria: the mean reciprocal ranking (MRR) indicator, defined as: Where: r i is the ranking of the correct answer in the inference results; the proportion of the correct answer ranked less than or equal to k in the inference, Hits@k, is defined as: Where: δ(r i ≤k) means that the result is 1 when the condition in the brackets is met, otherwise it is 0.

[0047] We used the inference method to test the data in the test set and compared our method with five embedding representation learning algorithms, including TransE, DistMult, TTransE, HyTE, and De-DistMult, based on two evaluation criteria. The results of different inference methods are shown in Table 1:

[0048] Table 1 Comparison between this method and existing methods

[0049] method MRR (%) Hits@3(%) Hits@10(%) TransE 11.54 7.35 14.55 DistMult 11.98 7.28 14.73 TTransE 9.22 6.33 13.69 HyTE 9.92 6.17 13.62 De-DistMult 13.24 7.81 15.56 This method 16.48 13.67 18.38

[0050] The higher the MRR and Hits@k values, the better the inference results. As can be seen from the table, the MRR, Hits@3, and Hits@10 results of this method are 16.48%, 13.67%, and 18.38%, respectively, all outperforming existing methods. This method infers closed-loop path rules mined from existing knowledge graphs. Due to the cyclical nature of Earth observation mission processes, the mined rules are consistent with historical facts, resulting in more accurate results. It also has better reliability and interpretability than existing mainstream embedding representation learning algorithms. Some of the mined closed-loop path rules are shown in Table 2:

[0051] Table 2 Some closed-loop path rules in the Earth observation mission process dataset

[0052]

[0053]

[0054] As can be seen from the table, this method can effectively mine implicit closed-loop path rules from the fine-grained and frequently occurring data of the Earth observation mission process, for example:

[0055] Rule 1: Since an area may be continuously observed after an observation request is issued, once a satellite payload has observed an area, it is very likely that the payload will observe the area again within 24 hours.

[0056] Rule 2: After the user submits a terrain detection request to the demand planning department, the information system department will eventually return the data product results to the user;

[0057] Rule 3: Due to the limited number of satellite ground system resources, when a large number of mission plans are transmitted, the same control system department may transmit multiple mission plans to the same information system department.

[0058] This method infers closed-loop path rules mined from existing knowledge graphs. Due to the cyclical nature of Earth observation mission processes, the mined rules are consistent with historical facts. Compared with existing technologies, this method predicts higher entity rankings, with the evaluation metrics MRR and Hits@k outperforming other methods. Compared to mainstream embedding representation learning algorithms, this method, based on the superior MRR and Hits@k evaluation metrics, selects existing rules through the transformer encoder, resulting in mined rules with greater reliability and interpretability.

[0059] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principles and purpose of the present invention. The scope of protection of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. All implementation schemes within its scope shall be subject to the constraints of the present invention.

Claims

1. An interpretable reasoning method for the time series knowledge graph of Earth observation mission processes, characterized by: In the offline phase, closed-loop path rules are used to collect potential multi-hop closed-loop path rules in the time series knowledge graph. A network model containing rule filters and rule reasoners is constructed and trained based on the Earth observation mission process dataset. In the online phase, the trained network model is used to predict real-time test data and, based on the closed-loop path rules, the missing processes in the Earth observation mission are completed. The network model includes: a rule filter unit and a rule reasoner unit, wherein: the rule filter unit scores the closed-loop path rules corresponding to each rule head according to the feature information of the closed-loop path rules contained in the time series knowledge graph, and the top q closed-loop path rules in each rule head are used as the result obtained by the rule filter unit; the rule reasoner unit processes the quadruple in the time series knowledge graph according to the closed-loop path rule information filtered by the rule filter unit, obtains the loss function value corresponding to the closed-loop path rule, and trains the network model according to the loss function value; The rule filter unit calculates the scores of all closed-loop path rules and sorts them, obtains the rule with the highest score as the screening result, and completes the screening of the rules. Specifically, it includes: a rule encoder, a transformer encoder for capturing rule features, and a sub-network for converting features into rule scores, wherein: the rule encoder encodes and fuses the attributes contained in the closed-loop path rules according to the closed-loop path rules to obtain encoding information; the transformer encoder performs feature extraction processing based on the encoding information to obtain a feature matrix; the sub-network converts (n + 1) The feature matrix of size m is converted into a score to obtain the score of the closed-loop path rule; The rule reasoner unit improves the existing neural network model based on translation invariance, and includes: an entity encoder, a relationship encoder, a negative sample generator and a transfer model-based reasoner, wherein: the entity encoder and the relationship encoder generate entity encoding information and relationship encoding information respectively according to the quadruple data in the temporal knowledge graph; the negative sample generator replaces the tail entity in the quadruple by uniform random negative sampling according to the quadruple information in the temporal knowledge graph to obtain the negative sample corresponding to each quadruple; the reasoner performs network training processing according to the n-hop closed-loop path rule with a higher score, and continuously optimizes through the rule filter unit and the rule reasoner unit to obtain the n-hop closed-loop path rule with higher quality in the temporal knowledge graph.

2. The interpretable reasoning method for the time series knowledge graph of the Earth observation mission process according to claim 1 is characterized in that: When there are n pieces of information ( , ,..., ), then the closed-loop path rule is used to infer that there is a relationship r between x and y at a certain moment in the future; the closed-loop path rule , including a header consisting of Horn clauses and a body of multiple quadruple entries of length n, specifically: , where: n is the length of the closed loop path rule, x, y, Represents an entity in the knowledge graph, Represents the relationship in the knowledge graph, Indicates the start timestamp of the relationship, and t > > > ... > , Indicates the duration of the relationship.

3. The interpretable reasoning method for the time series knowledge graph of the Earth observation mission process according to claim 1 is characterized in that: The coding information includes: the coding of the relationship r, the coding of the start timestamp t, the coding of the duration d and the coding of the relationship sequence position, wherein: the coding of the relationship r, the coding of the start timestamp t and the coding of the duration d are all generated by encoding different values ​​using an embedding function. A one-dimensional space vector of size m, where m is the dimension of the encoding; the encoding of the relationship order position is based on the relationship in the closed loop path rule The order of appearance is encoded, which is consistent with the position encoding of the transformer, that is, ,in: is the value of the i-th dimension of the position encoding vector at position t. Similarly, the encoding also generates 1 One-dimensional space vector of size m; The transformer encoder has 8 attention heads. By adding the corresponding positions of the four encodings, each tuple in an n-hop closed-loop path rule is obtained. Corresponding 1 If a one-dimensional space vector of size m is used, then an n-hop closed-loop path rule (including the head and body of the rule) will generate (n + 1) m-sized feature matrix; The sub-network includes: a fully connected layer, a ReLU activation layer and a dropout regularization layer.

4. The interpretable reasoning method for the time series knowledge graph of the Earth observation mission process according to claim 1 is characterized in that: The entity coding is to adopt the idea of ​​dynamic entity and regard the entity coding information as a continuous function that changes with time, that is, , where: the sin function is used to retain the information of the entity's periodic changes over time, and the tanh function is used to retain the information of the entity's monotonically changing over time. Used to retain constant information of the entity; The relationship encoding information includes: the encoding of the relationship r, the encoding of the start timestamp t, and the encoding of the duration d. These three encodings are similar to the rule encoder in the rule filter unit. The relationship encoder fuses the three encodings according to the set weights to obtain the encoding information of the relationship, that is, ; The replacement means that if the number of entities is N and the number of negative samples is k, through negative sampling, each iteration only needs to train k+1 of them to improve the computational efficiency of the model; The transfer model-based reasoner is designed based on the principle of translation invariance, that is, the head entity should be approximately equal to the tail entity after the relationship transformation. According to the n-hop closed-loop path rule, the head entity should be approximately equal to the tail entity after being transferred through n relationships in the rule body, that is: .

5. The interpretable reasoning method for the time series knowledge graph of the Earth observation mission process according to claim 1 is characterized in that: The network training process specifically includes: Step 1: Divide the Earth observation mission process time series knowledge graph data sample set into a training set, a validation set, and a test set; Step 2: Set the maximum value of n in the n-hop closed-loop path rule, and recursively obtain the n-hop closed-loop path rule corresponding to each rule r based on the test set; Step 3: Use the rule filter unit to filter out the n-hop closed-loop path rules with higher scores corresponding to each rule r; Step 4: In the training set, use the rule reasoner unit to apply the n-hop closed-loop path rules with higher scores to each data point, obtain the loss function, and then backpropagate to update the parameters in the rule filter unit and the rule reasoner unit. Step 5: Continuously iterate steps 3 and 4, select the network with the highest evaluation index in the validation set as the trained interpretable reasoning method, and finally obtain high-quality n-hop closed-loop path rules in the temporal knowledge graph.

6. The interpretable reasoning method for the time series knowledge graph of the Earth observation mission process according to any one of claims 1 to 5 is characterized in that: include: Step A: Number the entities, relationships, and time information contained in the Earth Observation Mission Process dataset, and store the inverse relationship of each relationship as a new relationship in the dataset. The inverse relationship of the quadruple (head entity, relationship, tail entity, time) is (tail entity, inverse of relationship, head entity, time). Reasoning on the second entity can simultaneously complete two different reasoning tasks: reasoning about the head entity or reasoning about the tail entity. Then, the dataset is divided into a training set, a validation set, and a validation set in a ratio of 0.6:0.2:0.

2. Step B: Use a recursive algorithm to search for closed-loop path rules from the training set. To improve search efficiency, set the maximum value of n in the closed-loop path rule to 5. To limit the number of rules found, set the maximum start time difference between the rule header and the rule body in the closed-loop path rule to 24 hours, and set the minimum unit of the relationship duration to 2 hours. Each relationship in the knowledge graph is used as a rule header, and the searched rule body is stored in the corresponding rule header to generate a closed-loop path rule set. Step C: Set the learning rate to 0.001, the number of iterations to 200, the batch size for each training round to 512, the optimizer to Adam optimizer, the dropout value to 0.5, and use the rule filter to select the top 10 closed-loop path rules from the closed-loop path rule set for each relationship in the knowledge graph. The input closed-loop path rule length is 6, including 1 rule head and 5 rule bodies, the number of attention mechanism heads is 8, the code generated by the rule encoder needs to be a multiple of 8, and the size is set to 128. After encoding, the input matrix size is 6 128, after the transformer encoder, the output feature matrix size is 6 128, and finally through the sub-network, through 2 fully connected layers, we get 1 The value of 1 is the score of the current closed-loop path rule; Step D: Use the rule reasoner to train the screened rules. During each round of training, 10 negative samples are randomly generated for each data based on the training set samples to reduce the computational complexity of the algorithm and enrich the training set samples to improve the generalization ability of the reasoning method. By iterating steps C and D, the accuracy of the reasoning algorithm is verified using the validation data set, and higher-quality rules are screened out to obtain the explainable reasoning method with the highest evaluation index. Finally, the quadruple to be inferred in the test set is used as input, and the screened rules and the trained rule reasoner network are used to obtain the reasoning result.

7. A system for implementing the explainable reasoning method for the time series knowledge graph of Earth observation mission processes as described in any one of claims 1 to 6, characterized in that: include: A rule filter unit and a rule reasoner unit, wherein: the rule filter unit scores the closed-loop path rules corresponding to each rule head according to the feature information of the closed-loop path rules contained in the temporal knowledge graph, and filters out the top q closed-loop path rules in each rule head; the rule reasoner unit processes the quadruple in the temporal knowledge graph according to the filtered closed-loop path rules, obtains the loss function value corresponding to the closed-loop path rule, and trains the network model according to the loss function value.

Images