Urban information reasoning enhancement method and system based on recessive time-space relationship

Abstract

The invention belongs to the technical field of artificial intelligence and smart city crossing, and particularly relates to a city information reasoning enhancement method and system based on a recessive time-space relationship, and the method comprises the steps: carrying out the initial vector representation of a learning event; modeling an event sequence and predicting subsequent events; constructing and verifying a causal rule based on multi-source evidence; performing an anti-factual intervention to validate the causal link; constructing a multi-hop reasoning path and fusing multi-modal information; integrating data driving and causal reasoning results, and outputting an enhanced city information reasoning result; the system correspondingly comprises an event representation learning module, a sequence modeling and prediction module, a causal rule construction and verification module, a multi-hop reasoning and multi-modal fusion module, an information fusion and collaborative reasoning module and a result generation and output module. And the cognition and decision support capability of the city information model in a complex scene is improved.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and smart cities, specifically to a method and system for enhancing urban information reasoning based on implicit spatiotemporal relationships. Background Technology

[0002] In the field of City Information Modeling (CIM) and related geospatial analysis, existing technologies primarily rely on the integration of static or quasi-static 3D models with Geographic Information Systems (GIS) to achieve digital mapping of urban space. When it comes to event reasoning and situation prediction, existing technologies typically employ the following two mainstream methods:

[0003] (1) Data-driven statistical learning models. These methods convert entities and events in the urban environment into low-dimensional vector representations using deep neural networks, and perform similarity calculations or correlation predictions in this vector space. For example, by training the model to learn historical event sequences, the most likely subsequent events under specific initial conditions can be predicted.

[0004] (2) Knowledge base model based on predefined rules. This type of method relies on a rule base built from expert experience and performs reasoning through logical matching. These rules are often macroscopic and deterministic and are directly encoded into the business logic of the city information model.

[0005] Although the aforementioned existing technologies have achieved some application in urban visualization and management, they have significant shortcomings in realizing deep cognitive reasoning. The specific shortcomings and their technical reasons are as follows: a. The superficiality and "black box" problem of causal relationship mining: Most existing data-driven models learn in an end-to-end manner. Although they can discover statistical associations from data, the models themselves are "black boxes," and their decision-making process lacks transparency. They may incorrectly use accompanying non-causal associations as the basis for reasoning, leading to unreliable conclusions. This is because the model's learning objective is to minimize prediction error, rather than to reveal the inherent causal mechanisms between variables. The model structure itself lacks modules designed to distinguish between correlation and causation, and it also lacks the ability to incorporate domain common sense and physical constraints.

[0006] b. Insufficient ability to model fine-grained, multi-hop spatiotemporal relationships: Existing methods mostly focus on single-hop relationships at the entity level (continuous heavy rainfall → traffic disruption), making it difficult to capture and infer multi-hop, chain-like evolutionary relationships such as "continuous heavy rainfall → surge in underground pipeline load → causing local road collapse → leading to regional traffic disruption." These relationships are typical "implicit spatiotemporal relationships" because they involve cross-domain entities, spatiotemporal delay effects, and indirect causal relationships, exceeding the processing scope of traditional GIS spatial analysis or statistical models. This results in the inference of complex urban events remaining at a superficial level, unable to predict indirect and secondary events. The technical bottleneck lies in the fact that traditional CIM and GIS models excel at expressing spatial topological relationships but lack mechanisms for explicitly modeling complex logical relationships and temporal evolution paths between events; while neural network models also have inherent difficulties in handling long-range dependencies and multi-step inference.

[0007] c. Insufficient utilization of multimodal evidence and fragile inference links: Current inference processes are often disconnected from the rich multi-source, multimodal data (such as text reports, sensor time-series data, oblique photogrammetry models, and BIM component information) in urban information models. Inference relies on only a small number of input features, failing to effectively align and integrate spatiotemporal evidence, resulting in a lack of solid evidence support in the inference link and poor generalization ability and reliability. Technically, it is difficult to embed unstructured multimodal data (text, images, geographic coordinates) into a unified, location-centric semantic space and to effectively interact with causal inference models. Data from different sources exhibit heterogeneity in spatiotemporal scale and semantics, making effective alignment and conflict resolution difficult.

[0008] More specifically, existing data-driven models suffer from causal confounding problems. For example, a model might incorrectly associate "holidays" (a common confounding factor) with "traffic congestion," when the true causal factor might be "large-scale events" occurring during holidays. Furthermore, existing rule-based models cannot dynamically adapt and validate; their predefined rules (such as "rainfall → flooding") frequently generate false alarms when quantification thresholds are not met or when multi-source evidence is lacking. Summary of the Invention

[0009] In view of this, the present invention aims to overcome the shortcomings of existing urban information modeling technologies in causal relationship mining, fine-grained multi-hop relationship reasoning, and deep fusion of multimodal evidence. By introducing counterfactual intervention and multi-source evidence quantification, it proposes an urban information reasoning enhancement method and system based on implicit spatiotemporal relationships to systematically solve the above problems.

[0010] First, the core concepts involved in this invention will be explained: The "implicit spatiotemporal relationship" described in this invention refers to an inter-event relationship that satisfies at least one of the following characteristics: (a) there is a spatiotemporal delay between cause and effect, and they do not co-occur immediately; (b) the intermediate nodes of the causal relationship chain involve entities from different fields (such as meteorology, transportation, and municipal affairs); and (c) it cannot be directly discovered in traditional spatial topological analysis (such as adjacency and inclusion) or statistical correlation analysis. The "interpretable causal rule" described in this invention refers to an "IF-THEN" type computer-executable rule whose conditional part is composed of quantified spatiotemporal evidence, and whose causal validity is verified through counterfactual intervention.

[0011] On the one hand, this invention provides a method for enhancing urban information reasoning based on implicit spatiotemporal relationships. Logically, this method can be summarized as the following key steps, executed sequentially and iteratively based on feedback: Step 1: Learn event representations based on historical event sequences, and use a time series model to predict at least one candidate subsequent event as the result event. Simultaneously output the result event. The prediction confidence level; Step 2: Regarding the resulting event Identify at least one potential causal event from the historical event sequence or contextual events. To constitute a causal event Extracting spatiotemporal evidence from multi-source heterogeneous spatiotemporal data, and based on a pre-defined spatial domain. Key monitoring indicator thresholds and facility carrying capacity threshold The spatiotemporal evidence is quantified to construct an interpretable causal rule in the form of "IF [quantified conditional evidence] THEN [outcome event]". Step 3: For the causal event pair Perform counterfactual intervention operations Calculate the outcome events before and after the intervention. probability difference If the absolute value of the probability difference ΔP, |ΔP|, is greater than a preset reliability threshold. If the causal rule is found to be credible, its credibility is output; otherwise, a verification failure signal is generated, and at least one of the following operations is triggered based on the signal: return to the prediction step to trigger the prediction of new candidate subsequent events, or return to the rule construction step to adjust the determination strategy of the causal event or the quantification parameters of the evidence. Step 4: Based on the pre-built city knowledge graph, analyze the causal event pairs... Starting from the event entity in the process, multi-hop path exploration is carried out to generate multi-hop reasoning paths that connect cause and effect; Step 5: Based on the prediction confidence and the causal rule credibility, dynamically allocate the fusion weights of the data-driven reasoning results and the causal reasoning results, and fuse the semantic information of the multi-hop reasoning path to generate the final enhanced urban information reasoning output.

[0012] The following sections, S1 to S8, further explain the detailed implementation of the above key steps: S1: Learn the initial event representation to implement the "learn event representation and predict" step.

[0013] For input events The event's key elements (such as time, location, entity, and behavior) are extracted and converted into vector representations using a pre-trained word vector model (such as Word2Vec or BERT). Then, these element vectors are aggregated using an attention mechanism to obtain the overall representation of the event. Let the event... The set of elements is The vector of each element The event representation is calculated as follows:

[0014] in, It is an element The attention weights are learned through a multilayer perceptron (MLP) and a softmax function:

[0015] in, It is a multilayer perceptron used to calculate the importance score of each element, thereby capturing the differentiated contribution of different elements to the representation of events.

[0016] S2: Update the event representation and predict the result, thus implementing the "learning event representation and prediction" step.

[0017] Given the sequence of events (the events mentioned above) We use temporal models (such as LSTM, Transformer) to update event representations to capture context dependencies. Let the context representation at time t be... Then, a set of candidate subsequent events is generated based on the updated representation. Its preliminary forecasts provide context for subsequent refined selection.

[0018] S3: Calculate the event relevance score and select subsequent events to implement the "learn event representation and predict" step.

[0019] Regarding the above event and candidate events Calculate each of the preceding events and The relevance score. Scoring function. Using bilinear transformation or cosine similarity, the formula is as follows:

[0020] in, This represents the correlation calculation function, which selects the candidate event with the highest score as the output. It is the first The vector representation of the events mentioned above at time t. It is the vector representation of the j-th candidate event. It is a learnable parameter matrix.

[0021] Then, for each candidate event Based on the combined scores of all the events mentioned above (e.g., taking the average), select the candidate event with the highest score as the output:

[0022] S4: Construct causal rules based on multi-source evidence to achieve the step of "determining the cause and constructing causal rules".

[0023] Given a pair of causal events Based on the reasoning rules defined by domain experts (e.g., "continuous heavy rainfall may lead to urban flooding"), spatiotemporal correlation evidence is extracted from multi-source data (such as sensor time series data, text reports, and BIM models). Through spatiotemporal alignment and conflict resolution of multi-source evidence, causal rules with geographical interpretability are constructed.

[0024]

[0025] in, Represents the normalization function. Represents the ReLU function. For spatial domain, for Time and location Key monitoring indicators (such as sensor readings). For position The facility carrying capacity coefficient (such as the road carrying capacity coefficient). For position Environmental constraints This serves as the activation threshold for the corresponding indicator. This evidence is transformed into causal logic rules in the form of "IF [conditional evidence] THEN [outcome event]", providing a transparent and reliable reasoning chain for emergency decision-making.

[0026] It should be noted that the probabilities calculated by the causal rules constructed in this step... As a conditional probability based on observational data, it quantifies the probability of observing a causal event. Under the conditions that it occurs, the resulting event The likelihood of the event occurring. This probability value synthesizes the quantitative results of multi-source spatiotemporal evidence and reflects the likelihood of the event occurring. and The statistical association strength between them. However, this association may be affected by confounding factors, so it cannot be determined solely by... Insufficient to confirm and To investigate the true causal relationship between the two, this invention subsequently introduces a counterfactual intervention verification step (S5) to further examine and quantify the causal effect.

[0027] S5: Implement counterfactual intervention to verify the causal link, thus achieving the "implement counterfactual intervention verification" step.

[0028] Based on the constructed causal rules, we simulated scenarios after changes in key factors to verify the physical interpretability of the causal chain.

[0029] Counterfactual intervention Regarding the cause of the change Parameters, observed events The probability changes are compared with theoretical calculations:

[0030] in, Indicates causal intervention operation, if If the result is consistent with expectations (e.g., the error is less than a threshold), the causal rule is considered reliable; otherwise, a verification failure signal is generated, triggering at least one of the following closed-loop feedback mechanisms: ① Feedback to steps S2 / S3: Based on the currently updated context event representation, re-execute event prediction and selection to generate new candidate subsequent events; ② Feedback to step S4: Based on the current verification status, adjust the evidence extraction strategy or rule construction parameters, and re-execute the construction and verification of causal rules.

[0031] This iterative process continues until the counterfactual intervention verification in step S5 passes, or the number of iterations reaches the preset upper limit.

[0032] The counterfactual intervention operation By representing causal events in simulation models or structural causal models The variable values ​​are fixed to a specific state, while the distribution of other irrelevant variables remains unchanged, and then the probability after intervention is calculated through simulation. .

[0033] S6: Construct a multi-hop relationship reasoning link to realize the "multi-hop path exploration" step.

[0034] Capture deep semantic relationships between event elements and iteratively mine spatiotemporal entities in geographic media. Subject of the act Their interactions form an interpretable reasoning path, as shown below:

[0035] in, For the first The attention weights for skip relationships are learned through a neural network; It is a relational predicate function that represents the interaction between entities (such as "cause" or "influence").

[0036] To effectively integrate multimodal information (such as text descriptions, visual features, and geographic coordinates) with the aforementioned inference path, this invention introduces a cross-modal alignment mechanism. This mechanism first maps the features of each modality to a unified semantic space through a modality-specific projection layer, then assigns fusion weights to each modality using an attention mechanism, and finally performs weighted fusion. The fusion representation is as follows:

[0037] in, For the attention weights of each modality, A modality-specific projection layer is formed, thereby creating a knowledge-enhanced reasoning chain that improves the ability to identify implicit causal relationships.

[0038] S7: Integrate the results of common sense reasoning (S1-S3) and causal reasoning (S4-S6) into a deep fusion to achieve the step of "dynamically allocating weights and fusing information".

[0039] By designing a collaborative reasoning mechanism, high-probability events obtained from statistical learning are weighted and combined with logical events supported by causal rules:

[0040] in, and These are the confidence weights for data-driven and causal inference, respectively, and their magnitudes are related to the corresponding prediction confidence. And the credibility of causal rules Positive correlation, and can be dynamically adjusted according to the application scenario; for example, it can be achieved through a normalization function. (e.g., using the Softmax function) to perform calculations: , ; and These are data-driven prediction results and causal logic reasoning results, respectively.

[0041] S8: Output the reasoning result to achieve the "generate final output" step.

[0042] The results of collaborative reasoning are integrated to generate a final enhanced output of urban information reasoning. The output includes event prediction, causal rules, multi-hop reasoning paths, and counterfactual analysis to support urban management decisions (such as emergency response and planning optimization).

[0043] Corresponding to the above method, the present invention also provides an urban information reasoning enhancement system, comprising: (1) Event representation learning module, used to execute step S1 and learn the initial event representation.

[0044] (2) Event sequence modeling and prediction module, used to execute steps S2 and S3, to process historical event sequences using time series models, and output candidate subsequent events and their prediction confidence.

[0045] (3) A causal rule construction and verification module, used to execute steps S4 and S5, establishes a bidirectional communication connection with the event sequence modeling and prediction module: on the one hand, it receives the candidate subsequent events from the prediction module and constructs interpretable causal rules from multi-source heterogeneous spatiotemporal data; on the other hand, when the counterfactual intervention verification fails, the module sends a feedback control signal to the event sequence modeling and prediction module, which at least includes the rule verification failure status and an instruction to request re-prediction, thereby triggering the prediction of new candidate events and forming a closed-loop iterative reasoning process. It includes a counterfactual intervention verification unit, used to perform counterfactual intervention operations and calculate the conditional probability difference.

[0046] (4) Multi-hop reasoning and multimodal fusion module, used to execute step S6, based on the pre-built city knowledge graph, to explore multi-hop paths starting from the current event entity, and use attention mechanism to perform weighted summation of relation vectors on the path to generate semantic reasoning path explanation. At the same time, a cross-modal alignment mechanism is introduced to embed text description, visual features and geographic coordinates into a unified semantic space.

[0047] (5) Information fusion and collaborative reasoning module, used to execute step S7, is connected to the event sequence modeling and prediction module and the causal rule construction and verification module. It is used to receive candidate subsequent events and their prediction confidence from the prediction module, as well as the verified causal rule reasoning results and their credibility from the causal verification module, and dynamically adjust the fusion weight based on this to integrate the data-driven reasoning results and the causal reasoning results.

[0048] (6) Reasoning result generation and output module, used to execute step S8, integrate collaborative reasoning results, and output city information.

[0049] Corresponding to the above method, the present invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method when executing the program.

[0050] Corresponding to the above method, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method.

[0051] The beneficial effects of this invention are as follows: By introducing a counterfactual intervention verification mechanism, the problem of the "black box" of causal relationships in data-driven models is solved, and interpretable causal reasoning is realized. Through multi-hop chain reasoning supported by knowledge graphs, implicit spatiotemporal relationships that are difficult to capture by traditional methods are explicitly mined. Through cross-modal alignment and fusion mechanisms, the collaborative utilization of multi-source heterogeneous evidence is realized, and the robustness of the reasoning chain is enhanced.

[0052] Meanwhile, this invention is not a simple superposition of various technologies, but rather an organic synergy between data-driven models and causal logic models through the design of a closed-loop feedback verification mechanism; through multi-hop reasoning and multimodal evidence alignment, it provides a computable path for the explicitization of implicit spatiotemporal relationships; and it achieves a leap in capabilities from correlation prediction to causal explanation and from single-point judgment to chain deduction, significantly enhancing the credibility and decision support value of urban information reasoning results. Attached Figure Description

[0053] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 The overall flowchart provided for embodiments of the present invention; Figure 3 This is a schematic diagram comparing the effects of the present invention with those of the prior art. Detailed Implementation

[0054] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0055] This invention provides a method and system for enhancing urban information reasoning based on implicit spatiotemporal relationships, as detailed below. Figure 1-3 The present invention will be described in further detail below.

[0056] Figure 1 This is a flowchart of the key steps in a method for enhancing urban information reasoning based on implicit spatiotemporal relationships. Figure 2 This illustrates an architecture diagram of an urban information reasoning enhancement system based on implicit spatiotemporal relationships according to an embodiment of the present invention.

[0057] The system adopts a modular design, mainly including an event representation learning module, an event sequence modeling and prediction module, a causal rule construction and verification module, a multi-hop reasoning and multimodal fusion module, an information fusion and collaborative reasoning module, and a reasoning result generation and output module. These modules are sequentially connected through standard interfaces, forming a complete processing chain from data input to reasoning output, enabling deep reasoning and enhanced analysis of urban information.

[0058] Example 1: Urban Information Reasoning in Rainfall Events 1.1 Learning the initial event representation.

[0059] First, the input event is vectorized. Specifically, the system receives event description text from the city information model platform, such as "Continuous heavy rainfall in XX district for more than 3 hours". The pre-trained natural language processing model BERT parses the text, identifies and extracts core event elements, including time "more than 3 hours", location "XX district", entity "rainfall", and behavior "continuous heavy rainfall". Each element is converted into a high-dimensional vector. .

[0060] Subsequently, these feature vectors are fed into an attention aggregation layer, which computes the attention weight for each feature using a multilayer perceptron (MLP). :

[0061] Finally, the overall distributed representation of the event is obtained by weighted summation. :

[0062] This process dynamically highlights elements that are more critical to the current reasoning task, such as "continuous heavy rainfall" and "XX district," laying the foundation for subsequent in-depth reasoning.

[0063] 1.2 Update event representations and predict event outcomes.

[0064] The system creates a sequence of current events and their historical context events. This sequence of events represents… It is fed into a time series model to capture sequence dependencies and outputs an updated representation containing contextual information. .

[0065] 1.3 Calculate the event relevance score and select subsequent events.

[0066] The system selects several candidate events from a predefined pool of candidate events. For each candidate event... The system calculates its relationship with each preceding event in the sequence. Relevance score, scoring function Using bilinear transformation or cosine similarity:

[0067] in, This is the learnable parameter matrix.

[0068] The system employs a time-weighted strategy to calculate the comprehensive score for each candidate event and selects the candidate event with the highest score as the preliminary prediction result. :

[0069] 1.4 Constructing causal rules based on multi-source evidence.

[0070] The system provides interpretable causal logic to support the above predictions. For a causal event pair to be verified... , cause event The event was caused by "continuous heavy rainfall". For the term "urban flooding," the system extracts spatiotemporal evidence from multi-source data. Specifically, the system focuses on areas covered by rainfall. Internally, it integrates key monitoring indicators such as real-time rainfall. Pipeline drainage capacity and terrain elevation The conditional probability of this causal rule being true can be calculated using the following formula:

[0071] in, The function ensures that only values ​​exceeding the threshold are considered valid. Only then was the evidence activated. This is the Sigmoid function. This process transforms abstract causal hypotheses into quantifiable chains of evidence supported by data, converting numerical evidence into interpretable causal rules in the form of "IF [condition] THEN [outcome]".

[0072] 1.5 Implement counterfactual intervention to verify causal links.

[0073] The system performs robustness verification on the constructed causal rules. This is achieved by executing causal intervention operations. In the simulation environment, the "continuous heavy rainfall" was altered. The state of rainfall Set the value to a level below the urban flooding threshold and observe the resulting event "Urban Flooding". The probability change. This is determined by calculating the probability difference before and after the intervention. and with preset threshold A comparison is made to assess the reliability of the causal relationship. If If the result is not significant, the process will proceed to step four, prompting you to adjust the evidence extraction strategy or rule parameters.

[0074] 1.6 Constructing multi-hop relationship reasoning links.

[0075] In order to uncover the implicit spatiotemporal relationship between "continuous heavy rainfall" and "regional traffic disruption", the multi-hop reasoning and multimodal fusion module was put into operation. The implementation of this function relies on a pre-built urban knowledge graph.

[0076] Starting with the entity "continuous heavy rainfall," the system explored multiple paths within the graph and successfully discovered a path: "continuous heavy rainfall" → (leading to) → "surge in underground pipeline load" → (triggered) → "local road collapse" → (further causing) → "regional traffic disruption." This path fully reveals the implicit causal chain that traditional single-hop reasoning could not capture.

[0077] For each path, the system uses an attention mechanism to process the relation vectors along the path. By performing a weighted summation, we obtain the semantic representation of the path:

[0078] in, For the first The attention weights for skip relationships are learned through a neural network; It is a relational predicate function that represents the interaction between entities, such as "cause" or "influence".

[0079] Simultaneously, a cross-modal alignment mechanism is introduced to embed text descriptions, visual features, and geographic coordinates into a unified semantic space and fuse them into a joint representation:

[0080] in, The attention weights for each module, A modality-specific projection layer is used to form a knowledge-enhanced reasoning chain.

[0081] 1.7 Output of reasoning results.

[0082] The results of the aforementioned steps are ultimately integrated in the information fusion and collaborative reasoning module, which uses data to drive predictions. Results of causal logical reasoning To integrate and form a final decision:

[0083] Among them, the weighting coefficient and The system can be dynamically adjusted according to the application scenario. Ultimately, it outputs a comprehensive reasoning enhancement report, including: the final prediction conclusion ("High probability of traffic disruption in the area"), supporting causal rules, multi-hop reasoning paths revealing implicit relationships, and counterfactual analysis ("If rainfall stops, the risk of traffic disruption decreases by XX%"). This report is pushed to the city's emergency command platform via API, completing a full urban information reasoning enhancement process.

[0084] Example 2: Urban Information Reasoning in Traffic Accidents To further illustrate the reasoning ability of this invention in other urban events, a typical traffic accident chain is used as a second embodiment for description.

[0085] 2.1 Initial Event Representation Learning: The system receives event sequences from the traffic management platform, starting with "road construction on XX section" as the initial event. Based on historical patterns, these events may evolve into a chain of events such as "traffic convergence" leading to "rear-end collisions." The event representation learning module identifies and extracts core event elements, including "construction," "traffic flow," and "rear-end collisions," and uses an attention mechanism to aggregate these elements to obtain vector representations of each event.

[0086] 2.2 Update event representation and predict event outcome: In the event sequence modeling and prediction module, the system predicts the candidate subsequent events that may occur after "road construction" based on historical traffic event sequences, forming a candidate event library, including "vehicle flow convergence", "traffic congestion", "vehicle rear-end collision", etc.

[0087] 2.3 Calculating Event Relevance Scores and Selecting Subsequent Events: The system selects several candidate events from the aforementioned candidate event library. For each candidate event, its relevance score to each preceding event in the sequence is calculated. A time-series weighted average is used to calculate the comprehensive score for each candidate event, and the candidate event with the highest score is selected as the preliminary prediction result.

[0088] 2.4 Constructing Causal Rules Based on Multi-Source Evidence: In the causal rule construction and verification module, causal rules are constructed for causal events. Evidence is extracted from multi-source data, such as the road occupancy area and duration of construction, real-time traffic flow monitoring data, road design capacity, and weather visibility. The strength of the evidence is quantified using the following formula:

[0089] in, for Time and location Construction impact coefficient, For traffic flow, Visibility factor , The activation thresholds are for construction impact and traffic flow.

[0090] 2.5 Implement counterfactual intervention to verify causal links: The system constructs causal rules, such as "IF [construction occupancy > 2 hours & traffic flow > threshold & visibility < standard] THEN [risk of rear-end collision increases]", and performs causal intervention operations. Changing the cause event in a simulation environment “ The state of "" and observe the resulting event. “ The probability change before and after the intervention. and with preset threshold By comparing and simulating the probability change of vehicle rear-end collisions under the condition of "no construction", the following calculations were performed. Verify the reliability of the causal relationship. If If the result is not significant, the feedback will be sent to step 2.4, indicating that the evidence extraction strategy or rule parameters need to be adjusted.

[0091] 2.6 Constructing a Multi-hop Relationship Reasoning Path: In the multi-hop reasoning and multimodal fusion module, based on the urban traffic knowledge graph, a complete reasoning path is mined: "Construction Occupies Road" → (leading to) → "Lane Reduction" → (causing) → "Traffic Convergence" → (resulting in) → "Shortened Following Distance" → (ultimately leading to) → "Rear-end Collision". This path reveals implicit causal relationships that are difficult to discover with traditional single-hop reasoning. Simultaneously, the system integrates multimodal information such as construction permit texts, traffic monitoring videos, and GPS trajectory data, enhancing the evidentiary support for the reasoning through a cross-modal alignment mechanism.

[0092] 2.7 Output Reasoning Results: The information fusion and collaborative reasoning module weights and fuses the data-driven prediction results with the causal reasoning results, and outputs a comprehensive report that includes multi-hop reasoning paths, counterfactual analysis and risk levels. This report is then pushed to the traffic command center to provide decision support for traffic accident early warning and road network scheduling.

[0093] Figure 3The diagram illustrates a comparison between the technical solution of this invention and existing technologies. The left side shows three main shortcomings of existing City Information Modeling (CIM) technologies: the black-boxing of causal relationships leads to a lack of interpretability; the absence of multi-hop relationships results in broken reasoning chains; and insufficient multimodal fusion leads to inadequate utilization of evidence. The right side shows three technical effects of the method of this invention: achieving interpretable causal reasoning through counterfactual verification; supporting multi-hop chain reasoning paths through explicit modeling; and achieving deep multimodal fusion through cross-modal alignment. The arrows indicate the specific solutions this invention provides to each technical shortcoming.

[0094] In summary, this invention provides a method and system for enhancing urban information reasoning based on implicit spatiotemporal relationships. This method uses a time-series model for data-driven event prediction and constructs interpretable causal rules based on multi-source evidence, then uses counterfactual intervention to verify the credibility of the rules. Furthermore, it performs multi-hop reasoning through a knowledge graph and integrates multimodal information. Finally, it dynamically integrates the results of data-driven and causal reasoning based on credibility, overcoming the problems of black boxes in causal relationships, insufficient multi-hop reasoning capabilities, and inadequate utilization of multimodal information in existing technologies. This significantly improves the interpretability and reliability of urban information reasoning and is suitable for complex decision-making scenarios such as urban emergency management.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for enhancing urban information reasoning based on implicit spatiotemporal relationships, characterized in that, This includes the following key steps: Step 1: Learn event representations based on historical event sequences, and use a time series model to predict at least one candidate subsequent event as the result event. Simultaneously output the result event. The prediction confidence level; Step 2: Regarding the resulting event Identify at least one potential causal event from a sequence of historical events or contextual events. To constitute a causal event Extracting spatiotemporal evidence from multi-source heterogeneous spatiotemporal data, and based on a preset spatial domain Ω and key monitoring indicator thresholds. and facility carrying capacity threshold The spatiotemporal evidence is quantified to construct an interpretable causal rule in the form of "IF [quantified conditional evidence] THEN [outcome event]". Step 3: For the causal event pair Perform counterfactual intervention operations Calculate the outcome events before and after the intervention. probability difference ; If |ΔP| is greater than a preset reliability threshold If the causal rule is found to be credible, its credibility is output. Otherwise, a verification failure signal is generated, and at least one of the following operations is triggered based on the signal: return to the prediction step to trigger the prediction of new candidate subsequent events, or return to the rule building step to adjust the determination strategy of the causal event or the quantification parameters of the evidence. Step 4: Based on the pre-built city knowledge graph, analyze the causal event pairs... Construct multi-hop reasoning paths connecting causes and effects by starting from the event entities in the text; Step 5: Based on the prediction confidence and causal rule credibility, dynamically allocate the fusion weights of data-driven reasoning results and causal reasoning results, and fuse the semantic information of multi-hop reasoning paths to generate the final enhanced output of urban information reasoning.

2. The method according to claim 1, characterized in that, In step one, the "learning event representation" specifically refers to: An attention mechanism is used to aggregate vector representations of event elements, including time, location, entity, and behavior; event vector. Represented as: in, It is an element The attention weights are learned through a multilayer perceptron (MLP) and a softmax function: A vector representation of event elements. The quantity of elements.

3. The method according to claim 1, characterized in that, In step two, the "quantification of spatiotemporal evidence" specifically refers to: Based on the preset spatial domain Thresholds of key monitoring indicators and facility carrying capacity threshold The correlation strength probability of the causal event pairs is calculated using activation and normalization functions. .

4. The method according to claim 1, characterized in that, In step three, the "conditions for determining the credibility of a causal rule" specifically refer to: The absolute value of the probability difference ΔP, |ΔP|, is greater than a preset reliability threshold. And the sign of ΔP corresponds to the cause event. For the result event The expectation of producing a positive or negative impact.

5. The urban information reasoning enhancement method based on implicit spatiotemporal relationships according to claim 1, characterized in that, In step four, "constructing a multi-hop inference path" specifically refers to: Capture deep semantic relationships between event elements and iteratively mine spatiotemporal entities in geographic media. Subject of the act Their interactions form an interpretable reasoning path, as shown below: in, This represents the number of hops in the path. For the first The attention weights for skip relationships are learned through a neural network; It is a relational predicate function that represents the interaction between entities.

6. The method according to claim 1, characterized in that, In step five, the "dynamically allocating fusion weights and generating the final output" specifically refers to: Based on the predicted confidence level and the aforementioned credibility Through normalization function Computational data-driven inference weights Weight of causal reasoning ;in , The final fusion output is: ,in For data-driven prediction results, This is the result of causal reasoning.

7. The method according to claim 1, characterized in that, In step five, "integrating semantic information from multi-hop reasoning paths" specifically includes: A cross-modal alignment mechanism is introduced, embedding text descriptions, visual features, and geographic coordinates into a unified semantic space and performing weighted fusion. This fusion representation... : in, For the attention weights of each modality, For modality-specific projection layers, This represents the original features of each modality.

8. A city information reasoning enhancement system based on implicit spatiotemporal relationships, used to implement the method of any one of claims 1-7, characterized in that, include: (1) Event sequence modeling and prediction module, used to process historical event sequences using time series models and output candidate subsequent events and their prediction confidence; (2) Causal rule construction and verification module, which is connected to the prediction module, is used to receive the candidate subsequent events and construct interpretable causal rules from multi-source heterogeneous spatiotemporal data. It includes a counterfactual intervention verification unit, which is used to perform counterfactual intervention operations and calculate probability differences. (3) Multi-hop reasoning and multimodal fusion module: Based on the pre-built city knowledge graph, it explores multi-hop paths starting from the current event entity, and uses the attention mechanism to perform weighted summation of the relation vectors on the path to generate semantic reasoning path explanation. At the same time, it introduces a cross-modal alignment mechanism to embed text description, visual features and geographic coordinates into a unified semantic space. (4) Information fusion and collaborative reasoning module, which is communicatively connected to the event sequence modeling and prediction module and the causal rule construction and verification module, is used to receive candidate subsequent events and their prediction confidence from the prediction module, as well as the verified causal rule reasoning results and their credibility from the causal verification module, and dynamically adjust the fusion weight based on this to integrate the data-driven reasoning results and the causal reasoning results. The causal rule construction and verification module is configured to: when the counterfactual verification fails, send a signal to the event sequence modeling and prediction module to trigger the prediction of new candidate subsequent events, thus forming a closed-loop iterative reasoning process.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the program, implements the method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

Images