CN-121998088-A - Urban information reasoning enhancement method and system based on implicit space-time relationship

CN121998088ACN 121998088 ACN121998088 ACN 121998088ACN-121998088-A

Abstract

The invention belongs to the technical field of artificial intelligence and smart city intersection, and particularly relates to a city information reasoning enhancement method and system based on implicit space-time relationship, wherein the method comprises the steps of learning an initial vector representation of an event; modeling an event sequence and predicting a subsequent event; constructing and verifying causal rules based on multi-source evidence; the system correspondingly comprises event representation learning, sequence modeling and prediction, causal rule construction and verification, multi-hop reasoning and multi-mode fusion, information fusion and collaborative reasoning, and result generation and output modules.

Inventors

XIE HOULI
Ye Ruoyu
CHEN XUAN
FENG JIANGFAN
WANG CHUANDONG
Yan Lupeng
Huang Longhang
WEN XIN
LIU CHUAN
WANG CHUNLE

Assignees

重庆市建设信息中心

Dates

Publication Date: 20260508
Application Date: 20260121

Claims (10)

1. A city information reasoning enhancement method based on implicit space-time relationship is characterized by comprising the following key steps: Learning event representations based on a historical event sequence and predicting at least one candidate follow-up event as a result event using a time sequence model Simultaneously outputting the result event Is a predictive confidence level of (2); step two, aiming at the result event Determining at least one potentially causative event from a sequence of historical events or contextual events To form causal event pairs Extracting space-time evidence from multi-source heterogeneous space-time data, and based on a preset space domain omega and a key monitoring index threshold value And a facility load capacity threshold Quantifying the spatiotemporal evidence to construct an interpretable causal rule in the form of "IF [ quantized conditional evidence ] THEN [ result event ]; step three, for the causal event pairs Performing a counterfacts intervention operation Calculating the result event before and after intervention Probability difference of (2) If the |DeltaP| is greater than a predetermined reliability threshold If not, generating a verification failure signal and triggering at least one operation according to the signal, namely returning to the predicting step to trigger the prediction of new candidate follow-up events or returning to the rule constructing step to adjust the determination strategy of the cause event or the quantization parameter of evidence; Step four, based on a pre-constructed urban knowledge graph, the causal event pairs are generated Setting up a multi-hop reasoning path for connecting reasons and results by starting from an event entity in the system; And fifthly, dynamically distributing the fusion weight of the data-driven reasoning result and the causal reasoning result based on the prediction confidence and the causal rule confidence, and fusing the semantic information of the multi-hop reasoning path to generate final urban information reasoning enhancement output.
2. The method according to claim 1, wherein in step one, the "learning event representation" is specifically: aggregating vector representations of event elements including time, place, entity, and behavior by an attention mechanism Expressed as: Wherein, the Is an element Is learned by a multi-layer perceptron MLP and Softmax function: for the vector representation of the event elements, Is the number of elements.
3. The method according to claim 1, wherein in the second step, "quantifying time-space evidence" is specifically: based on preset spatial domain Key monitoring index threshold And a facility load capacity threshold Calculating the associated intensity probability of the causal event pair by an activation function and a normalization function 。
4. The method according to claim 1, wherein in step three, the "condition for determining that the causal rule is authentic" is specifically: The absolute value |Δp| of the probability difference Δp is greater than a predetermined reliability threshold And the positive and negative of delta P accord with the cause event For result events Producing an expectation of positive or negative effects.
5. The urban information reasoning enhancement method based on implicit space-time relationship according to claim 1, wherein in the fourth step, the "constructing a multi-hop reasoning path" is specifically: capturing deep semantic associations between event elements by iteratively mining spatiotemporal entities in a geographic medium Principal of behavior And its interaction relationship, forming an interpretable reasoning path, expressed as follows: Wherein, the In order to provide a number of path hops, Is the first The attention weight of the jump relation is learned through a neural network; Is a relational predicate function that represents interactions between entities.
6. The method according to claim 1, wherein in step five, the step of dynamically assigning the fusion weights and generating the final output is specifically: Based on the prediction confidence And the credibility By normalizing a function Calculating data driven inference weights And causal reasoning weights Wherein , The final fusion output is: Wherein For the data-driven prediction result, Is a causal reasoning result.
7. The method according to claim 1, wherein in step five, "fusing semantic information of a multi-hop inference path" specifically includes: A cross-modal alignment mechanism is introduced, embedding text description, visual features and geographic coordinates into unified semantic space and performing weighted fusion, and fusing the representation : Wherein, the For the attention weight of each modality, For a projection layer that is mode-specific, Is an original characteristic representation of each modality.
8. A system for urban information inference enhancement based on implicit spatiotemporal relationships for implementing the method of any of claims 1-7, comprising: (1) The event sequence modeling and predicting module is used for processing the historical event sequence by using the timing model and outputting candidate follow-up events and the prediction confidence thereof; (2) The causal rule constructing and verifying module is in communication connection with the predicting module and is used for receiving the candidate follow-up event and constructing interpretable causal rules from multi-source heterogeneous space-time data, and the causal rule constructing and verifying module comprises a counterfactual intervention verifying unit which is used for executing counterfactual intervention operation and calculating a probability difference; (3) The multi-hop reasoning and multi-mode fusion module performs multi-hop path exploration from the current event entity based on the pre-constructed urban knowledge graph, performs weighted summation on the relation vectors on the paths by using an attention mechanism, generates semantic reasoning path interpretation, and simultaneously introduces a cross-mode alignment mechanism to embed text description, visual features and geographic coordinates into a unified semantic space; (4) The information fusion and collaborative reasoning module is in communication connection with the event sequence modeling and prediction module and the causal rule construction and verification module, and is used for receiving candidate follow-up events from the prediction module and the prediction confidence thereof, and causal rule reasoning results and credibility thereof passing verification from the causal verification module, dynamically adjusting fusion weights based on the results, and integrating data-driven reasoning results and causal reasoning results; The causal rule construction and verification module is configured to feed back signals to the event sequence modeling and prediction module to trigger the prediction of new candidate follow-up events when the anti-facts are not verified, so as to form a closed-loop iterative reasoning process.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the method according to any of claims 1-7 when executing the program.
10. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method according to any of claims 1-7.

Description

Urban information reasoning enhancement method and system based on implicit space-time relationship Technical Field The invention relates to the technical field of intersection of artificial intelligence and smart cities, in particular to a method and a system for enhancing urban information reasoning based on implicit space-time relations. Background In the field of urban information model (CIM) and related geospatial analysis, the prior art solutions mainly rely on integration of static or quasi-static three-dimensional models with a Geographic Information System (GIS) to implement digital mapping of urban space. When event reasoning and situation prediction are involved, the prior art generally adopts the following two main stream methods: (1) Based on a data-driven statistical learning model. Such methods convert entities and events in urban environments into low-dimensional vector representations through deep neural networks, and perform similarity calculation or association prediction in the vector space. For example, a sequence of historical events is learned by training a model to predict the most likely subsequent events that occur under certain initial conditions. (2) Knowledge base models based on predefined rules. Such methods rely on logic rule bases built from expert experience to infer by logical matching. These rules tend to be macroscopic and deterministic, coded directly into the business logic of the city information model. Although the prior art has certain application in city visualization and management, the prior art has obvious defects in realizing deep cognitive reasoning, and specific defects and technical reasons thereof are as follows: a. The problem of the superficial nature of causal relation mining and the 'black box' is that the existing data-driven models are mostly learned in an end-to-end manner, and the models are 'black boxes' and the decision process lacks transparency although statistical correlations can be found from the data. It may erroneously take the concomitant non-causal association as a basis for reasoning, resulting in an unreliable conclusion. This is because the learning goal of the model is to minimize the prediction error, rather than reveal the causal mechanisms inherent between variables. The model structure itself is not designed with a module for distinguishing between relevance and causality, nor is it capable of integrating into the general knowledge of the field and physical constraints. B. The modeling capability of the fine-granularity multi-hop space-time relationship is insufficient, namely the existing method focuses on the single-hop relationship (continuous heavy rainfall→traffic interruption) of the entity level, and is difficult to capture and infer the multi-hop chain evolution relationship of continuous heavy rainfall→underground pipe network load surge→local road surface collapse→regional traffic interruption. The relationship is a typical 'implicit space-time relationship', and the relationship is related to cross-domain entity, space-time delay effect and indirect causal relationship, so that the processing range of the traditional GIS space analysis or statistical model is exceeded. This results in deduction of urban complex events remaining in shallow layers, indirect and secondary events being unforeseeable. The technical bottleneck is that the traditional CIM and GIS models are good at expressing the space topological relation, but lack of a mechanism for explicitly modeling complex logic relation and time sequence evolution paths among events, and the neural network model also has inherent difficulties in processing long-range dependence and multi-step reasoning. C. The multi-mode evidence information is underutilized, and an inference link is fragile, so that the current inference process is often disjointed with rich multi-source and multi-mode data (such as text report, sensor time sequence data, oblique photography model and BIM component information) in the urban information model. Inference relies on only a small number of input features, and the spatiotemporal evidence cannot be aligned and fused effectively, so that the inference links lack solid evidence support, and the generalization capability and reliability are poor. Technically, unstructured multimodal data (text, images, geographic coordinates) are difficult to embed into a unified, location-centric semantic space and interact efficiently with causal inference models. The data from different sources are heterogeneous on a space-time scale and semantically, and effective alignment and conflict resolution are difficult to achieve. More specifically, existing data driven models suffer from causal confusion problems. For example, the model may erroneously relate "holidays" (a common confounding factor) to "traffic congestion", whereas in reality, the actual cause and effect may be "large activities" occurring during holidays. Meanwhile, the existing rule model cannot