CN-121980193-A - Building rubbish illegal dumping behavior chain reconstruction method based on time sequence map reasoning

CN121980193ACN 121980193 ACN121980193 ACN 121980193ACN-121980193-A

Abstract

The invention provides a construction waste illegal dumping behavior chain reconstruction method based on time sequence spectrum reasoning, which relates to the technical field of knowledge spectrums and comprises the steps of constructing a time sequence knowledge spectrum through multi-source data fusion, and dynamically representing and reasoning potential association by utilizing a time sequence diagram neural network learning node, and further generating and evaluating candidate behavior paths to screen out a reconstruction behavior chain with highest causal strength. The invention realizes accurate and dynamic reconstruction and reasoning of the complex illegal action chain.

Inventors

SUN SHUMENG
Jia Songrui
YAN MIN
ZHAO LEIZHEN
ZHANG GUOQIANG

Assignees

北京亦庄智能城市研究院集团有限公司

Dates

Publication Date: 20260505
Application Date: 20260331

Claims (9)

1. The construction waste illegal dumping behavior chain reconstruction method based on time sequence map reasoning is characterized by comprising the following steps of: performing cross-space alignment on target entities in multi-source monitoring video data containing time stamps and performing space-time matching verification on the target entities and vehicle positioning track data to obtain fusion track data, and constructing a time sequence knowledge graph containing entity nodes, relationship edges and time attributes according to the fusion track data and entity basic information data; carrying out dynamic representation learning on nodes in the time sequence knowledge graph by adopting a time sequence diagram neural network, and predicting potential association relations and evolution trends among nodes by aggregating historical state information of neighbor nodes in a time dimension to obtain an inference relation set; generating a plurality of candidate behavior paths by using a suspected dumping event as a query source through a time constraint graph traversal algorithm based on the time sequence knowledge graph and the reasoning relation set; And performing causal strength evaluation on the plurality of candidate behavior paths, screening out a behavior path with highest causal strength as a reconstruction behavior chain by calculating causal dependency degree of adjacent events in the paths and time sequence consistency constraint satisfaction degree of the whole path, and performing self-adaptive adjustment on parameters of the time sequence diagram neural network based on evidence strength feedback of the reconstruction behavior chain.
2. The method of claim 1, wherein dynamically representing and learning nodes in the time sequence knowledge graph by using a time sequence graph neural network, predicting potential association relations and evolution trends thereof between nodes by aggregating historical state information of neighbor nodes in a time dimension, and obtaining an inference relation set, comprises: Extracting a historical state sequence from each node in the time sequence knowledge graph according to a multi-scale time window; Calculating a dynamic importance score based on historical interaction frequency between the neighbor node and the current node, type semantic similarity of a relation edge and state change amplitude of the neighbor node in a preamble time period, selecting an effective neighbor set according to the dynamic importance score, and carrying out weighted aggregation on a historical state sequence of the effective neighbor set to generate dynamic representation of the current node; Calculating the path reliability of each candidate time sequence path based on the dynamic representation similarity of each intermediate node in the path, the time sequence consistency of the connection relation and the attenuation factor of the path length, and carrying out aggregation on the path reliability of the plurality of candidate time sequence paths to calculate the association strength score between node pairs; and screening out node pairs with the association strength score exceeding a preset association threshold and the evolution trend feature presenting an enhancement mode as potential association relations to generate an inference relation set.
3. The method of claim 2, wherein calculating path confidence for each candidate timing path comprises: calculating a time interval rationality score according to the time stamp deviation of the relationship edges between adjacent nodes in the candidate time sequence path and the historical time interval distribution of the same relationship type in the time sequence knowledge graph; matching event sequence modes in candidate time sequence paths with a predefined causal constraint rule set, wherein the causal constraint rule set comprises a hard constraint rule and a soft constraint rule, applying punishment weights to paths violating the hard constraint rule, and applying rewarding weights to paths conforming to the soft constraint rule to obtain causal sequence consistency scores; performing weighted fusion on the time interval rationality score and the causal sequence consistency score to generate a time sequence consistency score; Calculating the similarity between the dynamic representation vectors of each intermediate node in the candidate time sequence path to obtain a dynamic representation similarity score, and calculating an attenuation factor according to the path length; And carrying out weighted summation on the dynamic representation similarity score and the time sequence consistency score to obtain a path basic score, and multiplying the path basic score with the attenuation factor to generate path credibility.
4. The method of claim 1, wherein generating a plurality of candidate behavior paths based on the time-series knowledge-graph and the set of inference relationships using a graph traversal algorithm with time constraints with suspected dumping events as query sources, comprises: taking a suspected dumping event node as a starting node, extracting a time stamp attribute of the node as a time constraint datum point, and constructing a time constraint filter based on a causal transfer time window range corresponding to each relationship type in the reasoning relationship set to carry out time constraint and time interval constraint judgment on the time stamp of the candidate relationship side; Expanding adjacent relation edges from the time sequence knowledge graph in a breadth-first mode, filtering the expanded relation edges by applying the time constraint filter, extracting target nodes meeting time constraint, and storing the target nodes and path information into a path record structure, wherein the path information comprises a node identification sequence, a relation edge type sequence and a time stamp sequence; And judging whether the target node meets the path termination condition, if so, outputting the current path, and if not, continuing to expand until traversing is completed, and generating a plurality of candidate behavior paths.
5. The method of claim 1, wherein performing causal strength evaluation on the plurality of candidate behavior paths, and screening the behavior path with the highest causal strength as a reconstruction behavior chain by calculating causal dependency of adjacent events in the paths and time sequence consistency constraint satisfaction of the whole path, comprises: Calculating local causal dependencies based on historical co-occurrence frequencies and conditional probabilities between event types for adjacent event pairs in each candidate behavior path; calculating a time rationality weight according to the time interval between adjacent event pairs and the causal transfer time window range of the corresponding relation type in the reasoning relation set; carrying out aggregation operation on causal association intensity scores of all adjacent event pairs in the candidate behavior path to obtain path-level causal dependency; Extracting a time stamp sequence of each relation edge in the candidate behavior path, calculating time interval distribution characteristics between adjacent time stamps, performing matching degree calculation on the time interval distribution characteristics and a preset time sequence consistency constraint rule, and generating the time sequence consistency constraint satisfaction degree of the whole path; and carrying out weighted fusion on the path-level causal dependency degree and the time sequence consistency constraint satisfaction degree to obtain causal strength comprehensive scores of the candidate behavior paths, and selecting the candidate behavior path with the highest causal strength comprehensive score as a reconstruction behavior chain.
6. The method of claim 5, wherein weighting and fusing the local causal dependencies with the temporal rationality weights to obtain a causal correlation strength score for each adjacent event pair, comprises: Counting the accuracy and the data integrity of each event type in the verified behavior chain from historical behavior data, and carrying out normalized fusion calculation on the accuracy and the data integrity to obtain the credibility score of the event type; extracting reliability scores of two event types in adjacent event pairs, converting the two reliability scores into fusion weight coefficients of local causal dependency and time rationality weights through a nonlinear mapping function, and distributing the fusion weights into preset intervals according to the numerical range of the reliability scores by the nonlinear mapping function; and carrying out weighted summation on the local causal dependency degree and the time rationality weight according to the fusion weight coefficient to obtain a causal correlation strength score.
7. The method of claim 1, wherein adaptively adjusting parameters of the timing diagram neural network based on evidence strength feedback of the reconstructed behavioral chain comprises: Extracting evidence nodes from the reconstruction behavior chain, calculating a data source reliability score, a time integrity score and a space precision score for each evidence node, and carrying out weighted fusion on the data source reliability score, the time integrity score and the space precision score to obtain an evidence node quality score; calculating the space-time consistency verification intensity among the evidence nodes, and carrying out cross correction on the quality scores of all the evidence nodes based on the space-time consistency verification intensity to generate corrected evidence intensity; comparing the prediction result of the reconstruction behavior chain with the artificial labeling result to calculate a prediction deviation, and carrying out weighted distribution on the prediction deviation according to the corrected evidence intensity of each evidence node to generate a differential feedback signal; And carrying out gradient update on parameters related to corresponding evidence nodes in the time sequence diagram neural network by utilizing the differential feedback signals.
8. An electronic device, comprising: A processor; A memory for storing processor-executable instructions; Wherein the processor is configured to invoke the instructions stored in the memory to perform the method of any of claims 1 to 7.
9. A computer readable storage medium having stored thereon computer program instructions, which when executed by a processor, implement the method of any of claims 1 to 7.

Description

Building rubbish illegal dumping behavior chain reconstruction method based on time sequence map reasoning Technical Field The invention relates to a knowledge graph technology, in particular to a building rubbish illegal dumping behavior chain reconstruction method based on time sequence graph reasoning. Background In the field of supervision and traceability of illegal dumping behavior of construction waste, the prior art generally relies on independent analysis of multi-source surveillance videos and vehicle track data. It is common practice to process video data and positioning data separately, and associate a particular vehicle with a suspected dumping event by simple spatiotemporal threshold matching. Video analysis focuses on identifying entities such as transportation vehicles using object detection algorithms, while trajectory data is used to outline the path of movement of the vehicle. Such processed information is often stored and presented in isolated event records or static relationship charts, which are intended to provide clues to law enforcement. However, the conventional methods described above have significant drawbacks. On one hand, due to heterogeneous data sources and split processing processes, accurate alignment and coherent tracking of the inter-camera and inter-period entity behaviors are difficult to achieve, and behavior chain interruption or false correlation is easy to occur due to information fragmentation. On the other hand, the static analysis model cannot effectively describe the dynamic evolution process of the relationship among entities in the illegal behaviors, such as the change of the interactive mode among vehicles, sites and personnel along with time, so that the inference of a concealed complete illegal chain with causal logic from discrete events becomes extremely difficult, and the predictability and the traceability of supervision are restricted. Disclosure of Invention The embodiment of the invention provides a building rubbish illegal dumping behavior chain reconstruction method based on time sequence map reasoning, which can solve the problems in the prior art. In a first aspect of the embodiment of the invention, a method for reconstructing a chain of illegal dumping behavior of construction waste based on time sequence map reasoning is provided, which comprises the following steps: performing cross-space alignment on target entities in multi-source monitoring video data containing time stamps and performing space-time matching verification on the target entities and vehicle positioning track data to obtain fusion track data, and constructing a time sequence knowledge graph containing entity nodes, relationship edges and time attributes according to the fusion track data and entity basic information data; carrying out dynamic representation learning on nodes in the time sequence knowledge graph by adopting a time sequence diagram neural network, and predicting potential association relations and evolution trends among nodes by aggregating historical state information of neighbor nodes in a time dimension to obtain an inference relation set; generating a plurality of candidate behavior paths by using a suspected dumping event as a query source through a time constraint graph traversal algorithm based on the time sequence knowledge graph and the reasoning relation set; And performing causal strength evaluation on the plurality of candidate behavior paths, screening out a behavior path with highest causal strength as a reconstruction behavior chain by calculating causal dependency degree of adjacent events in the paths and time sequence consistency constraint satisfaction degree of the whole path, and performing self-adaptive adjustment on parameters of the time sequence diagram neural network based on evidence strength feedback of the reconstruction behavior chain. The method comprises the steps of carrying out dynamic representation learning on nodes in the time sequence knowledge graph by adopting a time sequence diagram neural network, predicting potential association relations and evolution trends among nodes by aggregating historical state information of neighbor nodes in a time dimension, and obtaining an inference relation set, wherein the inference relation set comprises the following steps: Extracting a historical state sequence from each node in the time sequence knowledge graph according to a multi-scale time window; Calculating a dynamic importance score based on historical interaction frequency between the neighbor node and the current node, type semantic similarity of a relation edge and state change amplitude of the neighbor node in a preamble time period, selecting an effective neighbor set according to the dynamic importance score, and carrying out weighted aggregation on a historical state sequence of the effective neighbor set to generate dynamic representation of the current node; Calculating the path reliability of each candidate time seque