A computer-aided processing method and system for investigation based on big data analysis

Abstract

The application discloses a kind of computer-aided processing method and system based on big data analysis for investigation, comprising: collecting the personnel identity information of case to be investigated, space trajectory information and communication record information, and generating initial representation of investigation data by fusion;Initial representation is input into TabPFN network, to identify potential associated nodes and paths, and obtain initial association probability;Accordingly, build and dynamically update case association graph;Real-time focus matrix is generated using bionic visual attention mechanism;According to the focus matrix, the association probability is corrected, and the graph is updated in real time to generate suspect chain data;According to the feedback data of investigation, the focus matrix parameters are dynamically optimized.The application significantly improves the efficiency and accuracy of intelligent association analysis of investigation data.

Description

Technical Field

[0001] This invention relates to the field of computer-aided investigation and analysis technology, and in particular to a computer-aided processing method and system for criminal investigation based on big data analysis. Background Technology

[0002] With the rapid development of computer technology and big data analytics, computer-aided investigation analysis has become a crucial means to improve the efficiency and accuracy of criminal case investigations. Traditional computer-aided investigation analysis methods typically employ correlation analysis techniques based on static information comparison. This involves simply combining or statically matching structured information such as the identity information, spatial trajectory information, and communication record information of suspects to achieve preliminary correlation identification of case-related clues. However, in real-world investigation scenarios with large amounts of data or high case complexity, static methods struggle to effectively capture dynamically evolving relationships over time. This often necessitates extensive manual verification, resulting in low efficiency and insufficient accuracy of intelligent identification, significantly hindering the efficiency and quality of case investigations.

[0003] To address these issues, some existing technical solutions have begun to utilize dynamic graph networks and visual attention mechanisms to achieve dynamic correlation analysis of criminal case data. This includes building a correlation graph based on initial case data and then calculating the importance of nodes and paths using a single-scale attention mechanism to achieve preliminary dynamic correlation analysis. However, the attention mechanisms in these solutions typically lack effective feedback optimization capabilities and cannot simultaneously account for short-term dynamic fluctuations and long-term trend stability, making it difficult to capture key suspect nodes and potential correlation paths in real time and accurately. Furthermore, the node feature representation methods used in existing dynamic graph analysis solutions are relatively simplistic, failing to effectively integrate the complex cross-modal spatiotemporal relationships between personnel, trajectories, and communication information, severely impacting the accuracy and stability of the correlation analysis.

[0004] Therefore, how to provide a computer-aided processing method and system for criminal investigation based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a computer-aided processing method and system for criminal investigation based on big data analysis. Addressing the shortcomings of existing technologies in intelligent identification efficiency and accuracy during rapid correlation analysis of criminal investigation data, this invention proposes a method that integrates a TabPFN network and a biomimetic visual attention mechanism. By constructing a real-time dynamically updated case correlation graph and attention focus matrix, it achieves real-time dynamic correction of correlation probabilities and automatic push and feedback optimization of suspect chains. This invention offers the advantages of fast, high-precision, and real-time performance in criminal investigation data correlation analysis.

[0006] A computer-aided processing method for criminal investigation based on big data analysis according to an embodiment of the present invention includes:

[0007] Collect investigative data related to the cases to be investigated, and integrate them to generate an initial representation of the investigative data;

[0008] The TabPFN network is used to identify nodes and related paths with potential relationships in the initial representation of the case data in a fast reasoning manner, and the initial association probability score is obtained.

[0009] Based on the initial association probability scores, a case association graph is constructed, and the case association graph is dynamically updated based on new data.

[0010] A biomimetic visual attention mechanism is applied to the dynamically updated case association graph to generate a real-time updated attention focus matrix;

[0011] The initial association probability score output by the TabPFN network is weighted using the attention intensity values ​​in the attention focus matrix to obtain the corrected association probability score.

[0012] The weights of nodes and associated paths in the case association graph are updated in real time based on the corrected association probability scores, and dynamic suspect chain data is generated based on the updated weights.

[0013] It receives feedback confirmation data from investigators after conducting on-site investigations based on dynamic suspect chain data, and dynamically optimizes and updates the focus matrix.

[0014] Optionally, the process of collecting and fusing investigative data related to the case under investigation to generate an initial representation of the investigative data specifically includes:

[0015] Spatial trajectory information of relevant personnel in the case to be investigated is selected as the temporal reference for data fusion, and a multi-dimensional spatiotemporal consistency mapping relationship between spatial trajectory information and communication record information is established.

[0016] Based on the multidimensional spatiotemporal consistency mapping relationship, personnel location information, facial feature information and communication subject identification information are extracted, and iterative enumeration is performed to obtain a preliminary matching association set;

[0017] For each matching combination in the preliminary matching association set, the coordinate difference between the geographical coordinates of the communication base station corresponding to the communication record and the location coordinates corresponding to the spatial trajectory information is extracted and calculated to obtain a high-confidence effective association set.

[0018] Based on each high-confidence matching combination in the effective association set, a multidimensional structured association matrix is ​​automatically generated, and the temporal and spatial continuity values ​​of each association unit in the multidimensional structured association matrix are calculated one by one to obtain the spatiotemporal structural coupling relationship.

[0019] Based on the temporal and spatial continuity values ​​in the multidimensional structured correlation matrix, a multimodal information fusion map is generated, and a structural coupling strength level label is assigned to each connection edge.

[0020] Based on the structural coupling strength level markers, the trend characteristics of relevant personnel in each case to be investigated are evaluated one by one. Based on the evaluation results, the node relationship subgraphs with high structural coupling strength levels and stable trends are automatically selected as the initial representation of the case investigation data.

[0021] Optionally, the TabPFN network includes a multi-timescale dynamic attention computation layer, an attention focus adaptive feedback layer, a spatiotemporally coupled attention fusion layer, a multi-granularity association graph dynamic correction layer, and a real-time reinforcement feedback self-optimization layer.

[0022] The multi-timescale dynamic attention calculation layer is used to perform sliding window attention calculations on both short-timescale and long-timescale values ​​on the time stamp sequence in the initial representation of the case data, and to obtain multi-scale attention feature vectors.

[0023] The attention focus adaptive feedback layer is used to receive multi-scale attention feature vectors and feedback confirmation data, and to adjust the attention feature vectors of the attention mechanism.

[0024] The spatiotemporal coupled attention fusion layer is used to receive the updated attention feature vector, and to perform cross-modal cross-attention calculation on the attention feature vectors calculated by the spatial trajectory information channel and the communication record information channel respectively, so as to obtain the coupled attention feature vector with dual constraints.

[0025] The multi-granularity association graph dynamic correction layer is used to construct an initial case association graph based on the coupled attention feature vector, and output the case association graph through online adaptive adjustment;

[0026] The real-time reinforcement feedback self-optimization layer is used to take the feedback confirmation data as the reinforcement feedback signal input, and drive the reinforcement learning loop to dynamically adjust the attention weight allocation strategy in the multi-timescale dynamic attention calculation layer and the attention focus adaptive feedback layer.

[0027] Optionally, obtaining the initial association probability score through the TabPFN network specifically involves:

[0028] The initial representation of the case investigation data is divided into a sequence of personnel identity information, a sequence of spatial trajectory information, and a sequence of communication record information, with each personnel node in the personnel identity information sequence serving as the index basis.

[0029] Based on the position coordinates and corresponding timestamp information in the spatial trajectory information sequence, the spatial displacement vector of each personnel node at different timestamps is calculated one by one, and matched one by one to form a spatiotemporal sequence matching set.

[0030] A spatial mapping set is formed based on the communication subject identification information in the communication record information sequence and the geographical location information of the communication base station;

[0031] Based on the spatiotemporal sequence matching set and the spatial mapping set, the spatial location change features and communication activity features of the nodes in each set at each timestamp are extracted, and the feature vectors are encoded one by one in a preset spatiotemporal feature vector format to form a standardized node feature representation sequence.

[0032] The standardized node feature representation sequence is used as the input of the TabPFN network. The similarity value of the feature vector between each node in the sequence and other nodes is calculated in the first forward inference method of the TabPFN network, and then transformed into the potential association relationship between the nodes.

[0033] Construct a topology of associated paths based on the potential relationships between nodes, and assign an initial association probability score to each path.

[0034] Optionally, the step of constructing a case association graph based on the initial association probability score and dynamically updating the case association graph based on new data specifically involves:

[0035] Using the initial association probability score as the initial edge weight, the node identifier corresponding to each person's identity information in the personnel identity information sequence is selected one by one to construct the initial case association graph.

[0036] Real-time acquisition and processing of newly generated spatial trajectory information and communication record information during the reconnaissance process to form a standardized node feature representation sequence of the newly added data;

[0037] Based on the spatial displacement vector features and communication activity features corresponding to each new node in the standardized node feature characterization sequence of newly added data, cross-comparison of temporal position and spatial coordinates is performed with the features of existing nodes to determine the spatiotemporal position similarity.

[0038] Based on the determined spatiotemporal similarity, the standardized node feature representation sequence of the newly added data is input into the TabPFN network. The similarity value of the feature vector between each newly added node and the existing nodes is calculated one by one by the forward inference of the TabPFN network, and the new potential association between the newly added node and the existing node is determined.

[0039] Based on the newly added potential relationships, insert the new nodes and the newly added potential relationship paths between the new nodes and existing nodes into the case relationship graph;

[0040] The edge weights between nodes are updated based on the similarity values ​​of the feature vectors corresponding to the newly added potential relationships, resulting in a dynamically updated case association graph.

[0041] Optionally, the step of performing a biomimetic visual attention mechanism on the dynamically updated case association graph to generate a real-time updated attention focus matrix specifically involves:

[0042] Based on the changes in the association probability value of each node in the dynamically updated case association graph within multiple consecutive preset time windows, the initial focus of attention node is identified.

[0043] Taking the initial focus node as the central node, calculate the correlation probability gradient value of the path between the central node and adjacent nodes in the case association map one by one, and determine the attention radiation range of each central node.

[0044] Based on the scope of attention, an initial attention focus matrix is ​​constructed with each central node as the origin of the matrix coordinates, and the initial attention intensity of the central node to each node within the scope of attention is determined.

[0045] For newly added nodes and newly added associated paths in the case association graph after real-time dynamic updates, calculate the association probability value between the newly added node and the existing central node, and determine whether it is within the attention and radiation range of the existing central node.

[0046] For newly added nodes that are determined to be within the attention radiation range of existing central nodes, the correlation probability gradient value of the newly added node is weighted and fused with the normalized value of the matrix element corresponding to the existing initial attention focus matrix, and the attention intensity value of each node in the initial attention focus matrix is ​​updated.

[0047] For newly added nodes that are determined not to be within the attention radiation range of any existing central node, the newly added nodes are automatically marked as new initial attention focus nodes, a new initial attention focus matrix is ​​re-established, and they are merged into the existing initial attention focus matrix in real time to form a merged attention focus matrix.

[0048] Based on the fused attention focus matrix, attention intensity thresholds are filtered one by one, and the filtered attention focus matrix is ​​output.

[0049] Optionally, the step of using the attention intensity values ​​in the attention focus matrix to weight the initial association probability score output by the TabPFN network to obtain the corrected association probability score specifically involves:

[0050] Extract the attention intensity value of each node to other nodes in the attention focus matrix, and match them one by one with the initial association probability score output by the TabPFN network according to the node identifier;

[0051] For each pair of nodes, the initial association probability score and the corresponding attention intensity value are used to calculate the ratio of the attention intensity value to the initial association probability score, thus obtaining the initial association correction factor.

[0052] Based on the initial association correction factor corresponding to each pair of nodes, calculate the statistical mean of the initial association correction factors among all nodes one by one, and determine a unified association correction coefficient.

[0053] Based on a unified correlation correction coefficient and the attention intensity value between each pair of nodes in the attention focus matrix, the correlation probability correction weight of each pair of nodes is calculated one by one.

[0054] The weights are adjusted based on the association probability between nodes, and the adjusted association probability score between nodes is calculated one by one.

[0055] Optionally, the step of updating the weights of nodes and associated paths in the case association graph in real time based on the corrected association probability scores, and generating dynamic suspect chain data based on the updated weights, specifically involves:

[0056] Based on the corrected correlation probability score, the original edge weights of the paths between every two nodes in the case correlation graph are replaced one by one;

[0057] Using the updated graph edge weights as input, a depth-first search method is used to traverse all associated paths in the case association graph starting from each node, and the cumulative probability value of each path is calculated.

[0058] Based on the cumulative probability values ​​obtained from all paths, calculate the standardized probability value for each path one by one;

[0059] Based on the calculated path standardization probability values, each path is determined and marked as belonging to a suspected chain;

[0060] For each path marked as a suspected chain, extract the personnel identity information identifiers corresponding to each node in the path, the corrected correlation probability scores of the paths between nodes after real-time updates, and the topological structure of the path, and extract the suspected chain data one by one.

[0061] Based on the generated suspect chain data, a graphical representation of the chain topology is established;

[0062] At preset real-time push intervals, the data of each suspected chain and its corresponding graphical representation of the chain topology are pushed to the investigator's terminal in the form of a data stream.

[0063] Optionally, the receiving of feedback confirmation data after on-site investigation by investigators based on dynamic suspect chain data, and the dynamic optimization and updating of the attention focus matrix, specifically involves:

[0064] It receives feedback confirmation data sent by investigators' terminals in real time. The feedback confirmation data includes on-site investigation confirmation marker information for each dynamic chain of suspicion. The on-site investigation confirmation marker information is divided into valid confirmation marker information and invalid confirmation marker information.

[0065] For suspected chains marked as validly confirmed in the on-site investigation and confirmation information, the attention intensity value between the corresponding nodes of the path is increased successively according to the number of times the chain is validly confirmed;

[0066] For suspected chains marked as invalid confirmations in the on-site investigation and confirmation information, the attention intensity values ​​between the corresponding nodes in the path are reduced successively according to the number of times the chain is invalidally confirmed;

[0067] Based on the changes in the attention intensity values ​​of all suspected chains, the overall attention intensity change of the attention focus matrix is ​​statistically analyzed and calculated one by one, and the attention intensity change is compared with the preset dynamic optimization threshold.

[0068] When the magnitude of the change in attention intensity exceeds the preset dynamic optimization threshold, the parameters of the attention focus matrix are automatically optimized globally.

[0069] When the change in attention intensity does not exceed the preset dynamic optimization threshold, the adjustment results of attention intensity values ​​between local nodes that have already been executed are maintained, and no global dynamic optimization update is performed.

[0070] Based on the optimized and updated state of the attention focus matrix parameters, output the optimized and updated attention focus matrix.

[0071] Optionally, a computer-aided processing system for criminal investigation based on big data analysis includes:

[0072] The case investigation data collection and initial characterization module is used to collect the identity information, spatial trajectory information and communication record information of the persons involved in the case to be investigated, and integrate them to generate the initial characterization of the case investigation data;

[0073] The TabPFN network inference module is used to receive the initial representation of the investigation data, perform fast forward inference, and output the initial association probability scores between nodes.

[0074] The case association graph construction and dynamic update module is used to construct an initial case association graph based on the initial association probability score output by the TabPFN network inference module, and dynamically update the nodes and association paths in the case association graph based on the newly collected case data in real time.

[0075] The attention focus matrix generation and update module is used to perform biomimetic visual attention mechanism processing on the dynamically updated case association map, generate the attention focus matrix, and dynamically update the attention focus matrix based on the information of newly added nodes and associated paths in real time.

[0076] The association probability correction calculation module is used to perform weighted calculations on the initial association probability scores output by the TabPFN network inference module based on the attention intensity values, so as to obtain the corrected association probability scores between each node.

[0077] The dynamic suspect chain generation and visualization push module is used to update the edge weights of nodes and related paths in the case association graph in real time based on the corrected association probability scores, and generate dynamic suspect chain data based on the updated edge weights, and push it to the investigator's terminal in real time in the form of a graphical representation of a chain topology.

[0078] The feedback confirmation data receiving and attention focus matrix dynamic optimization module is used to receive feedback confirmation data in real time after the investigator's terminal confirms the case based on the dynamic suspect chain data. The confirmation data is used to dynamically optimize and update the parameters of the attention focus matrix in the attention focus matrix generation and update module.

[0079] The beneficial effects of this invention are:

[0080] (1) By integrating the TabPFN network and the bionic visual attention mechanism, this invention constructs a dynamic case association map and attention focus matrix in real time, realizing rapid intelligent identification and accurate dynamic correction of the association nodes and paths of the case data. This effectively improves the accuracy and real-time performance of the intelligent association analysis of the case data, and enhances the rapid response and investigation efficiency of investigators in complex cases.

[0081] (2) By adopting dynamic suspect chain automatic generation and real-time visualization push technology, the present invention realizes the intuitive presentation of suspect targets and paths, significantly improves the data interactivity and interpretation convenience of the investigator's terminal, and shows better adaptability and application effect in real-time investigation scenarios of complex cases.

[0082] (3) In terms of the accuracy of case investigation data analysis, this invention effectively solves the technical problems of insufficient dynamic adjustment capability and low data correlation reliability in the prior art by using a dynamic optimization mechanism of attention focus matrix driven by feedback confirmation data. It breaks through the bottleneck of low efficiency in traditional investigation data processing, realizes closed-loop linkage between data processing and feedback optimization, and effectively improves the practical performance of intelligent case investigation. Attached Figure Description

[0083] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0084] Figure 1 This is a flowchart of a computer-aided processing method for criminal investigation based on big data analysis proposed in this invention. Detailed Implementation

[0085] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0086] refer to Figure 1 A computer-aided processing method for criminal investigation based on big data analysis, comprising:

[0087] Collect structured personal identity information, time and space trajectory information, and communication record information involved in the case to be investigated, and integrate various types of information to generate an initial representation of the case investigation data;

[0088] The initial representation of the case investigation data is input into the TabPFN network, which uses a fast reasoning method to identify nodes and related paths with potential relationships in the initial representation of the case investigation data, and obtains the initial association probability score between each node.

[0089] Based on the initial association probability score, a case association graph is constructed with the identity information of the persons involved in the case as nodes and the initial association probability score between nodes as edge weights. The case association graph is then dynamically updated based on the newly collected data in real time.

[0090] A biomimetic visual attention mechanism is applied to the dynamically updated case association graph to generate a real-time updated attention focus matrix, wherein the attention focus matrix contains the attention intensity values ​​of nodes with high association probability values ​​and association paths that change over time.

[0091] The initial association probability scores output by the TabPFN network are weighted using the attention intensity values ​​in the attention focus matrix to obtain the corrected association probability scores between each node.

[0092] The weights of nodes and associated paths in the case association graph are updated in real time based on the corrected association probability score, and dynamic suspect chain data is generated according to the updated weights and pushed to the investigators' terminals in a visual manner in real time.

[0093] The system receives feedback confirmation data from investigators after conducting on-site investigations based on dynamic suspect chain data, and dynamically optimizes and updates the parameters of the attention focus matrix in the bionic visual attention mechanism based on the feedback confirmation data.

[0094] In this embodiment, the collection of investigation data related to the case to be investigated and the fusion to generate an initial representation of the investigation data specifically includes:

[0095] Spatial trajectory information of relevant personnel in the case to be investigated is selected as the temporal reference for data fusion. Based on the timestamp corresponding to each spatial location coordinate in the spatial trajectory information, the difference between the occurrence time of each communication record and the corresponding time of the spatial trajectory location is retrieved one by one in the communication record information. A time difference threshold is set, and a multi-dimensional spatiotemporal consistency mapping relationship between spatial trajectory information and communication record information is established between communication records with time differences within the threshold and spatial location coordinates.

[0096] Based on the multidimensional spatiotemporal consistency mapping relationship between spatial trajectory information and communication record information, the personnel location information recorded in the spatial trajectory information, the facial feature information contained in the personnel identity information, and the communication subject identification information contained in the communication record information are extracted one by one. The three information combination patterns are automatically iteratively enumerated to obtain a preliminary matching association set.

[0097] For each matching combination in the preliminary matching association set, the geographical coordinates of the communication base station corresponding to the communication record and the location coordinates corresponding to the spatial trajectory information are extracted respectively. The spatial location coordinate difference between the two is calculated one by one. The spatial location similarity is judged according to the preset spatial coordinate difference threshold. The matching combination with spatial location similarity higher than the preset threshold is automatically associated with confidence assessment. Weak association combinations with association confidence lower than the preset threshold are eliminated to obtain a high-confidence effective association set.

[0098] Based on each high-confidence matching combination in the effective association set, a multidimensional structured association matrix is ​​automatically generated with personnel identity information as the vertical index and spatial location information and communication subject identification information as the horizontal index. The temporal continuity and spatial continuity values ​​of each association unit in the multidimensional structured association matrix are calculated one by one to quantitatively represent the spatiotemporal structural coupling relationship between personnel identity information, spatial trajectory information and communication record information.

[0099] Based on the temporal and spatial continuity values ​​in the multidimensional structured correlation matrix, a multimodal information fusion graph is generated one by one, with personnel identity information as the core node and spatiotemporal structural coupling relationship as the connecting edge. Each connecting edge in the multimodal information fusion graph is assigned a structural coupling strength level label.

[0100] Based on the structural coupling strength level markers in the multimodal information fusion graph, the trend characteristics of the structural coupling strength between each person's identity information node and its adjacent nodes in the fusion graph are calculated over time. Based on the quantitative evaluation results of the trend characteristics, node relationship subgraphs with high structural coupling strength levels and stable trends are automatically selected as the initial representation of the investigation data.

[0101] In this embodiment, the TabPFN network includes a multi-timescale dynamic attention calculation layer, an attention focus adaptive feedback layer, a spatiotemporally coupled attention fusion layer, a multi-granularity association graph dynamic correction layer, and a real-time reinforcement feedback self-optimization layer.

[0102] The multi-timescale dynamic attention calculation layer is used to perform sliding window attention calculations on the correlation between features in the initial representation of the case data based on the timestamp sequence in the initial representation of the case data. The attention calculation results of different time scales are fused through the attention weight adaptive mechanism to output a multi-scale attention feature vector that can simultaneously reflect the short-term sudden correlation trend and the long-term stable correlation trend.

[0103] The attention focus adaptive feedback layer is used to receive the multi-scale attention feature vector output by the multi-timescale dynamic attention calculation layer, as well as the feedback confirmation data after on-site reconnaissance. The weight parameters of the attention mechanism are dynamically and in real time adjusted through the feedback confirmation data to automatically adapt to the feature association change trend reflected by the on-site reconnaissance confirmation data, thereby realizing the online dynamic adaptive update of the attention feature vector.

[0104] The spatiotemporal coupled attention fusion layer is used to receive the adaptively updated attention feature vector output by the attention focus adaptive feedback layer, and to perform cross-modal cross-attention calculation on the attention feature vectors calculated by the spatial trajectory information channel and the communication record information channel respectively, and output coupled attention feature vectors that fuse the dual constraints of spatial position and communication behavior one by one, so as to improve the accurate coupling and association ability between cross-modal features in the investigative data association reasoning.

[0105] The multi-granularity association graph dynamic correction layer is used to construct an initial case association graph based on the coupled attention feature vector output by the spatiotemporal coupled attention fusion layer. It automatically triggers online adaptive adjustment of the association graph topology based on the changing trend of attention weights of associated nodes and associated paths, updates the structural relationship between associated nodes and paths in real time, and outputs a real-time dynamically corrected case association graph.

[0106] The real-time reinforcement feedback self-optimization layer is used to take the feedback confirmation data after the investigator's terminal confirms the case on-site as the reinforcement feedback signal input. The feedback confirmation data drives the reinforcement learning loop to dynamically adjust the attention weight allocation strategy in the multi-timescale dynamic attention calculation layer and the attention focus adaptive feedback layer, and outputs the network parameters after reinforcement feedback adjustment to form a real-time dynamic optimization loop for the accuracy of case data correlation analysis.

[0107] In this embodiment, the initial representation of the case investigation data is input into the TabPFN network to identify nodes and related paths with potential relationships in the initial representation of the case investigation data through fast reasoning, and to obtain the initial association probability score between each node. Specifically:

[0108] The initial representation of the case investigation data is divided into a sequence of personnel identity information, a sequence of spatial trajectory information, and a sequence of communication record information, with each personnel node in the personnel identity information sequence serving as the index basis for candidate nodes;

[0109] Based on the position coordinates and corresponding timestamp information in the spatial trajectory information sequence, the spatial displacement vector of each personnel node at different timestamps is calculated one by one, and the spatial displacement vector is matched one by one with the corresponding personnel node in the personnel identity information sequence to form a spatiotemporal sequence matching set of nodes and spatial trajectory information.

[0110] Based on the communication subject identification information and the geographical location information of the communication base station in the communication record information sequence, the spatial association mapping relationship between the communication activity location of each personnel node and the personnel spatial displacement vector is established one by one, forming a spatial mapping set between nodes and communication record information.

[0111] Based on the spatiotemporal sequence matching set of spatial trajectory information and the spatial mapping set of node and communication record information, the spatial location change features and communication activity features of nodes in each set at each timestamp are extracted, and feature vectors are encoded one by one in a preset spatiotemporal feature vector format to form a standardized node feature representation sequence suitable for TabPFN network input.

[0112] The standardized node feature representation sequence is used as the input of the TabPFN network. The similarity value of the feature vector between each node in the sequence and other nodes is calculated one by one in the forward inference method of the TabPFN network. The feature vector similarity value is automatically converted into the potential association relationship between nodes according to the preset threshold.

[0113] The system automatically constructs a topology of associated paths with nodes as vertices based on the potential relationships between nodes, and calculates the initial association probability score for each path in the topology as the initial probability input for subsequent updates to the case association graph.

[0114] In this embodiment, the step of constructing a case association graph based on the initial association probability score, with the identity information of the persons involved in the case as nodes and the initial association probability score between nodes as edge weights, and dynamically updating the case association graph based on the newly collected data in real time, specifically involves:

[0115] Using the initial association probability score as the initial edge weight, the node identifier corresponding to each person's identity information in the initial representation of the case data is selected one by one to construct an initial case association graph with unique nodes and determined edge weights.

[0116] Real-time acquisition of newly generated spatial trajectory information and communication record information during the reconnaissance process, and processing of the newly added data to form a standardized node feature representation sequence of the new data, including both new and existing nodes;

[0117] Based on the spatial displacement vector features and communication activity features corresponding to each new node in the standardized node feature representation sequence of the newly added data, cross-compare the temporal position and spatial coordinates with the features of existing nodes one by one. The spatiotemporal similarity between the new nodes and existing nodes is determined by the difference in temporal features and the difference in spatial features of the cross-compare.

[0118] Based on the determined spatiotemporal similarity, the standardized node feature representation sequence of the newly added data is input into the TabPFN network. The similarity value of the feature vector between each newly added node and the existing nodes is calculated one by one by the forward inference of the TabPFN network, and the potential new association between the newly added node and the existing node is automatically determined according to the preset threshold.

[0119] Based on the automatically identified new potential relationships, each new node and the new potential relationship path between the new node and existing nodes is inserted into the case association graph to dynamically adjust the topology of the case association graph.

[0120] The weights of the edges between nodes are updated one by one based on the similarity values ​​of the feature vectors corresponding to the new potential associations between the new nodes and the existing nodes. The updated edge weights are represented as association probability values, and the real-time association strength between nodes is expressed by the updated association probability values, thus obtaining the dynamically updated case association graph.

[0121] In this embodiment, the biomimetic visual attention mechanism is applied to the dynamically updated case association graph to generate a real-time updated attention focus matrix. This attention focus matrix includes the attention intensity values ​​of nodes and association paths with high association probabilities that change over time. Specifically:

[0122] Based on the change sequence of the association probability value of each node in the dynamically updated case association map within multiple consecutive preset time windows, the rate of change of the association probability value of each node between adjacent time windows is calculated one by one, and the initial focus of attention nodes whose association probability change rate exceeds the preset threshold are automatically identified according to the preset change rate threshold.

[0123] Taking the initial focus node as the central node, the association probability gradient value of the path between the central node and other adjacent nodes in the case association map is calculated one by one. The association probability gradient value represents the attenuation strength of the association probability when the central node propagates to the adjacent nodes, and the attention radiation range of each central node is automatically determined according to the magnitude of the association probability gradient value.

[0124] Based on the scope of attention, an initial attention focus matrix is ​​constructed one by one with each central node as the origin of the matrix coordinates. The value of each matrix element in the initial attention focus matrix is ​​the normalized value of the correlation probability gradient between the central node and the corresponding node within the scope of attention, thus determining the initial attention intensity of the central node to each node within the scope of attention.

[0125] For newly added nodes and newly added associated paths in the case association graph after real-time dynamic updates, calculate the association probability value between the newly added node and the existing central node, and automatically determine whether the newly added node is within the attention and radiation range of the existing central node based on the association probability value between the newly added node and the existing central node.

[0126] For newly added nodes that are determined to be within the attention radiation range of existing central nodes, the correlation probability gradient value of the newly added node is weighted and fused with the normalized value of the matrix element corresponding to the existing initial attention focus matrix, and the attention intensity value of each node in the initial attention focus matrix is ​​updated in real time.

[0127] For any new node that is determined not to be within the attention radiation range of any existing central node, the new node is automatically marked as a new initial attention focus node. A new initial attention focus matrix centered on the new node is re-established based on the correlation probability gradient value of the new node, and then merged into the existing initial attention focus matrix in real time to form a fused attention focus matrix.

[0128] Based on the fused attention focus matrix, attention intensity thresholds are filtered one by one, retaining nodes and paths whose attention intensity values ​​exceed the preset attention intensity threshold, and outputting the filtered attention focus matrix.

[0129] In this embodiment, the step of using the attention intensity values ​​in the attention focus matrix to weight the initial association probability scores output by the TabPFN network to obtain the corrected association probability scores between each node is specifically as follows:

[0130] Extract the attention intensity value of each node to other nodes in the attention focus matrix, and match them one by one with the initial association probability score output by the TabPFN network according to the node identifier;

[0131] For each pair of nodes, the ratio of the initial association probability score to the corresponding attention intensity value is calculated one by one to obtain the initial association correction factor.

[0132] Based on the initial association correction factor corresponding to each pair of nodes, calculate the statistical mean of the initial association correction factors among all nodes one by one, and determine a unified association correction coefficient based on the statistical mean.

[0133] Based on a unified correlation correction coefficient and the attention intensity values ​​between each pair of nodes in the attention focus matrix, the correlation probability correction weight for each pair of nodes is calculated one by one:

[0134] W ij =1+β·S ij ;

[0135] Among them, W ij S represents the adjusted weight for the association probability between node i and node j. ij β represents the attention intensity of node i to node j, and β represents the unified association correction coefficient;

[0136] Based on the adjusted weights of the association probabilities between nodes, calculate the adjusted association probability score between nodes one by one:

[0137] P′ ij =P ij ×W ij ;

[0138] Among them, P′ ij P represents the modified association probability score between node i and node j. ij This represents the initial association probability score between node i and node j output by the TabPFN network.

[0139] In this embodiment, the weights of nodes and associated paths in the case association graph are updated in real time based on the corrected association probability scores, and dynamic suspect chain data is generated according to the updated weights, which is then pushed to the investigators' terminals in a visualized manner in real time. Specifically:

[0140] Based on the corrected association probability score, the original edge weights of the paths between every two nodes in the case association graph are replaced one by one. The updated graph edge weights are the corrected association probability scores between the corresponding nodes.

[0141] Using the updated graph edge weights as input, a depth-first search method is used to traverse all associated paths in the case association graph starting from each node, and the cumulative probability value of each path is calculated. The cumulative probability value of the path is defined as the product of the edge weight values ​​between all adjacent nodes in the traversed path.

[0142] Based on the cumulative probability values ​​calculated for all paths, the standardized probability value for each path is calculated one by one. The standardized probability value is defined as the ratio of the cumulative probability value of the current path to the sum of the cumulative probability values ​​of all paths.

[0143] Based on the calculated path standardization probability value, each path is determined to be a suspected chain. When the path standardization probability value is higher than the preset standardization probability threshold, the corresponding path is automatically marked as a suspected chain.

[0144] For each path marked as a suspected chain, extract the personnel identity information identifier corresponding to each node in the path, the real-time updated corrected correlation probability score between nodes, and the topology of the path, and generate suspected chain data with personnel identity information identifier as nodes and corrected correlation probability score as edge weights.

[0145] Based on the generated suspected chain data, a graphical representation of the chain topology is established one by one. In the graphical representation, the visual features of the nodes are clearly represented by personnel identification information, and the edge weights between nodes are clearly marked by real-time values ​​of the corrected association probability scores.

[0146] At preset real-time push intervals, the data of each suspected chain and its corresponding graphical representation of the chain topology are pushed to the investigator's terminal in real time as a data stream.

[0147] In this embodiment, the step of receiving feedback confirmation data after the investigators have conducted on-site investigation and confirmation based on dynamic suspect chain data, and then dynamically optimizing and updating the parameters of the attention focus matrix in the bionic visual attention mechanism based on the feedback confirmation data, specifically involves:

[0148] It receives feedback confirmation data sent by investigators' terminals in real time. The feedback confirmation data includes on-site investigation confirmation marker information for each dynamic chain of suspicion. The on-site investigation confirmation marker information is divided into valid confirmation marker information and invalid confirmation marker information.

[0149] For the suspected chains marked as validly confirmed in the on-site investigation and confirmation information, the attention intensity values ​​in the attention focus matrix corresponding to the path between each node in the chain are extracted one by one, and the attention intensity values ​​between the corresponding nodes of the path are increased successively according to the number of times the chain is validly confirmed.

[0150] For the suspected chains marked as invalid confirmation in the on-site investigation and confirmation information, the attention intensity values ​​in the attention focus matrix corresponding to the path between each node in the chain are extracted one by one, and the attention intensity values ​​between the corresponding nodes of the path are reduced successively according to the number of times the chain is invalidally confirmed.

[0151] Based on the changes in the attention intensity values ​​of all suspected chains, the overall attention intensity change of the attention focus matrix is ​​statistically analyzed and calculated one by one, and the attention intensity change is compared with the preset dynamic optimization threshold.

[0152] When the change in attention intensity exceeds a preset dynamic optimization threshold, global dynamic optimization is automatically performed on the parameters of the attention focus matrix. This global dynamic optimization specifically includes:

[0153] The attention radiation range corresponding to each initial attention focus node is recalculated and adjusted one by one. Based on the correlation probability gradient value, the normalized attention intensity value between each node in the attention focus matrix is ​​optimized and updated one by one to obtain the globally dynamically optimized and updated attention focus matrix.

[0154] When the change in attention intensity does not exceed the preset dynamic optimization threshold, the adjustment results of attention intensity values ​​between local nodes that have already been executed are maintained, and no global dynamic optimization update is performed.

[0155] Based on the optimized and updated state of the attention focus matrix parameters, the attention focus matrix is ​​output in real time, either globally dynamically optimized or locally adjusted.

[0156] In this embodiment, a computer-aided processing system for criminal investigation based on big data analysis includes:

[0157] The case investigation data collection and initial characterization module is used to collect the identity information, spatial trajectory information and communication record information of the persons involved in the case to be investigated, and integrate them to generate the initial characterization of the case investigation data;

[0158] The TabPFN network inference module is used to receive the initial representation of the case data, perform fast forward inference, identify nodes and related paths with potential relationships in the initial representation of the case data, and output the initial association probability score between nodes.

[0159] The case association graph construction and dynamic update module is used to construct an initial case association graph based on the initial association probability score output by the TabPFN network inference module, with personnel identity information as nodes and the initial association probability score between nodes as edge weights, and dynamically update the nodes and association paths in the case association graph based on the newly collected case data in real time.

[0160] The attention focus matrix generation and update module is used to perform biomimetic visual attention mechanism processing on the dynamically updated case association map to generate an attention focus matrix. The attention focus matrix contains the attention intensity values ​​of nodes and association paths with high association probability values ​​that change over time, and the attention focus matrix is ​​dynamically updated based on the information of newly added nodes and association paths in real time.

[0161] The association probability correction calculation module is used to perform weighted calculation on the initial association probability score output by the TabPFN network inference module based on the attention intensity value in the attention focus matrix, so as to obtain the corrected association probability score between each node.

[0162] The dynamic suspect chain generation and visualization push module is used to update the edge weights of nodes and related paths in the case association graph in real time based on the corrected association probability score output by the association probability correction calculation module, and generate dynamic suspect chain data based on the updated edge weights, and push it to the investigator's terminal in real time in a graphical representation of a chain topology.

[0163] The feedback confirmation data receiving and attention focus matrix dynamic optimization module is used to receive feedback confirmation data in real time after the investigator's terminal confirms the case based on the dynamic suspect chain data. Based on the feedback confirmation data, the module dynamically optimizes and updates the parameters of the attention focus matrix in the attention focus matrix generation and update module, and outputs the attention focus matrix after global dynamic optimization or local optimization adjustment.

[0164] Example 1:

[0165] To verify the feasibility of this invention in practice, the system of this invention was applied to various complex cases handled daily by the investigation department of a public security agency. For multiple cases involving complex personal identity information, large amounts of spatial trajectory data, and numerous and intertwined communication records, dynamic data analysis and processing were implemented in order to achieve efficient and rapid correlation analysis of case information and real-time and accurate identification of key suspect chains.

[0166] In this application scenario, traditional investigative analysis methods generally employ a step-by-step verification and manual screening approach based on a single data source. Specifically, investigators acquire identity information, communication records, and spatial trajectory data of various suspects, manually comparing these data to confirm any suspicious connections between individuals, and then gradually deducing possible modus operandi and chains of suspicion. This traditional method is limited by the experience and data processing capabilities of investigators. When processing large amounts of data, it is manpower-intensive, has a long analysis cycle, and is prone to overlooking crucial data details. Especially in cases where information is frequently updated, it struggles to achieve real-time and effective dynamic correlation, severely impacting the efficiency of case investigation.

[0167] In actual operation, the case investigation data acquisition and initial characterization module of this invention first automatically collects the identity information of the personnel involved in the case, spatial trajectory data, and communication record information, and integrates the above-mentioned multi-source heterogeneous data into the initial characterization of the case investigation data. Subsequently, the TabPFN network inference module receives and performs fast forward inference, identifies nodes and paths with potential relationships based on the spatiotemporal similarity of the node feature vectors in the data, and outputs the initial association probability scores between nodes. Then, the case association graph construction and dynamic update module constructs an initial association graph based on the above-mentioned initial association probability scores, with personnel identities as nodes, and updates it dynamically in real time.

[0168] After the initial atlas is generated, the attention focus matrix generation and update module uses a biomimetic visual attention mechanism to capture the changes in the correlation probability of nodes and paths in the atlas over time through multi-scale attention calculation. This automatically generates a real-time attention focus matrix, effectively highlighting nodes and paths that investigators should focus on. Subsequently, the correlation probability correction calculation module integrates the attention intensity of the attention focus matrix with the initial correlation probability score, and obtains a corrected correlation probability score through weighted calculation, achieving dynamic correction and precise optimization of the case correlation analysis results.

[0169] The dynamic suspect chain generation and visualization push module updates the edge weights of nodes and paths in the case association graph in real time based on the corrected association probability score. Then, it automatically generates suspect chain data with the updated weights and pushes it to the investigator's terminal in real time at fixed intervals in an intuitive chain topology graphical interface, helping the investigator to quickly and accurately grasp the key doubts of the case.

[0170] In subsequent on-site investigations, investigators confirmed the suspected chain data received in real time and fed back relevant data to the system's feedback confirmation data receiving and attention focus matrix dynamic optimization module. This module receives feedback confirmation data in real time and adjusts the attention intensity values ​​in the attention focus matrix accordingly based on whether the investigators' confirmation marks are valid or invalid. It automatically calculates the adjustment range and compares it with a preset optimization threshold. If the threshold is exceeded, a global optimization program is initiated; otherwise, only local optimization is performed, thereby achieving continuous self-iterative optimization of the accuracy of case data analysis.

[0171] To verify the practical application effect of the solution of the present invention, this embodiment selects a typical case scenario, processes the same data using the system of the present invention and the traditional analysis method respectively, and compares the system performance. The detailed results are shown in Table 1:

[0172] Table 1. Performance Comparison of Case Data Correlation Analysis between the Invention Method and Traditional Methods

[0173]

[0174]

[0175] As shown in Table 1, compared with traditional manual data analysis methods, the system of this invention demonstrates a significant advantage when handling large-scale cases with numerous related paths. Regarding analysis time, taking case 004 as an example, traditional manual analysis took as much as 96 hours, while the system of this invention only took 120 minutes, significantly shortening the data processing cycle. In terms of analysis accuracy, the system of this invention has a significantly higher accuracy rate than traditional methods. Especially in the highly complex case 003, the accuracy rate of the traditional method was 74.6%, while the accuracy rate of the system of this invention increased to 91.7%, demonstrating a remarkable ability to accurately identify errors.

[0176] The actual verification of the above embodiments fully demonstrates that the present invention, by employing the TabPFN network to deeply integrate biomimetic visual attention mechanism technology, can capture the potential correlations of case data in real time and accurately. Combined with dynamic suspect chain generation, real-time visualization push, and feedback-driven adaptive optimization of attention focus matrix parameters, it successfully solves the problems of long time consumption, low efficiency, and insufficient accuracy of traditional criminal investigation data analysis technology. It significantly improves the efficiency and accuracy of actual case investigation and provides a powerful intelligent data analysis support for complex case investigation work, with broad application prospects.

[0177] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A computer-aided processing method for criminal investigation based on big data analysis, characterized in that, include: Collect investigative data related to the cases to be investigated, and integrate them to generate an initial representation of the investigative data; The TabPFN network is used to identify nodes and related paths with potential relationships in the initial representation of the case data in a fast reasoning manner, and the initial association probability score is obtained. The TabPFN network includes a multi-timescale dynamic attention computation layer, an attention focus adaptive feedback layer, a spatiotemporally coupled attention fusion layer, a multi-granularity association graph dynamic correction layer, and a real-time reinforcement feedback self-optimization layer. The multi-timescale dynamic attention calculation layer is used to perform sliding window attention calculations on both short-timescale and long-timescale values ​​on the time stamp sequence in the initial representation of the case data, and to obtain multi-scale attention feature vectors. The attention focus adaptive feedback layer is used to receive multi-scale attention feature vectors and feedback confirmation data, and to adjust the attention feature vectors of the attention mechanism. The spatiotemporal coupled attention fusion layer is used to receive the updated attention feature vector, and to perform cross-modal cross-attention calculation on the attention feature vectors calculated by the spatial trajectory information channel and the communication record information channel respectively, so as to obtain the coupled attention feature vector with dual constraints. The multi-granularity association graph dynamic correction layer is used to construct an initial case association graph based on the coupled attention feature vector, and output the case association graph through online adaptive adjustment; The real-time reinforcement feedback self-optimization layer is used to take the feedback confirmation data as the reinforcement feedback signal input to drive the reinforcement learning loop to dynamically adjust the attention weight allocation strategy in the multi-timescale dynamic attention calculation layer and the attention focus adaptive feedback layer. Based on the initial association probability scores, a case association graph is constructed, and the case association graph is dynamically updated based on new data. A biomimetic visual attention mechanism is applied to the dynamically updated case association graph to generate a real-time updated attention focus matrix; Based on the changes in the association probability value of each node in the dynamically updated case association graph within multiple consecutive preset time windows, the initial focus of attention node is identified. Taking the initial focus node as the central node, calculate the correlation probability gradient value of the path between the central node and adjacent nodes in the case association map one by one, and determine the attention radiation range of each central node. Based on the scope of attention, an initial attention focus matrix is ​​constructed with each central node as the origin of the matrix coordinates, and the initial attention intensity of the central node to each node within the scope of attention is determined. For newly added nodes and newly added associated paths in the case association graph after real-time dynamic updates, calculate the association probability value between the newly added node and the existing central node, and determine whether it is within the attention and radiation range of the existing central node. For newly added nodes that are determined to be within the attention radiation range of existing central nodes, the correlation probability gradient value of the newly added node is weighted and fused with the normalized value of the matrix element corresponding to the existing initial attention focus matrix, and the attention intensity value of each node in the initial attention focus matrix is ​​updated. For newly added nodes that are determined not to be within the attention radiation range of any existing central node, the newly added nodes are automatically marked as new initial attention focus nodes, a new initial attention focus matrix is ​​re-established, and they are merged into the existing initial attention focus matrix in real time to form a merged attention focus matrix. Based on the fused attention focus matrix, attention intensity thresholds are filtered one by one, and the filtered attention focus matrix is ​​output. The initial association probability score output by the TabPFN network is weighted using the attention intensity values ​​in the attention focus matrix to obtain the corrected association probability score. The weights of nodes and associated paths in the case association graph are updated in real time based on the corrected association probability scores, and dynamic suspect chain data is generated based on the updated weights. It receives feedback confirmation data from investigators after conducting on-site investigations based on dynamic suspect chain data, and dynamically optimizes and updates the focus matrix.

2. The computer-aided processing method for criminal investigation based on big data analysis according to claim 1, characterized in that, The process of collecting and fusing investigative data related to the case to be investigated to generate an initial representation of the investigative data specifically includes: Spatial trajectory information of relevant personnel in the case to be investigated is selected as the temporal reference for data fusion, and a multi-dimensional spatiotemporal consistency mapping relationship between spatial trajectory information and communication record information is established. Based on the multidimensional spatiotemporal consistency mapping relationship, personnel location information, facial feature information and communication subject identification information are extracted, and iterative enumeration is performed to obtain a preliminary matching association set; For each matching combination in the preliminary matching association set, the coordinate difference between the geographical coordinates of the communication base station corresponding to the communication record and the location coordinates corresponding to the spatial trajectory information is extracted and calculated to obtain a high-confidence effective association set. Based on each high-confidence matching combination in the effective association set, a multidimensional structured association matrix is ​​automatically generated, and the temporal and spatial continuity values ​​of each association unit in the multidimensional structured association matrix are calculated one by one to obtain the spatiotemporal structural coupling relationship. Based on the temporal and spatial continuity values ​​in the multidimensional structured correlation matrix, a multimodal information fusion map is generated, and a structural coupling strength level label is assigned to each connection edge. Based on the structural coupling strength level markers, the trend characteristics of relevant personnel in each case to be investigated are evaluated one by one. Based on the evaluation results, the node relationship subgraphs with high structural coupling strength levels and stable trends are automatically selected as the initial representation of the case investigation data.

3. The computer-aided processing method for criminal investigation based on big data analysis according to claim 1, characterized in that, The initial association probability score obtained through the TabPFN network is specifically as follows: The initial representation of the case investigation data is divided into a sequence of personnel identity information, a sequence of spatial trajectory information, and a sequence of communication record information, with each personnel node in the personnel identity information sequence serving as the index basis. Based on the position coordinates and corresponding timestamp information in the spatial trajectory information sequence, the spatial displacement vector of each personnel node at different timestamps is calculated one by one, and matched one by one to form a spatiotemporal sequence matching set. A spatial mapping set is formed based on the communication subject identification information in the communication record information sequence and the geographical location information of the communication base station; Based on the spatiotemporal sequence matching set and the spatial mapping set, the spatial location change features and communication activity features of the nodes in each set at each timestamp are extracted, and the feature vectors are encoded one by one in a preset spatiotemporal feature vector format to form a standardized node feature representation sequence. The standardized node feature representation sequence is used as the input of the TabPFN network. The similarity value of the feature vector between each node in the sequence and other nodes is calculated in the first forward inference method of the TabPFN network, and then transformed into the potential association relationship between the nodes. Construct a topology of associated paths based on the potential relationships between nodes, and assign an initial association probability score to each path.

4. The computer-aided processing method for criminal investigation based on big data analysis according to claim 1, characterized in that, The process of constructing a case association graph based on the initial association probability scores and dynamically updating the case association graph based on new data is as follows: Using the initial association probability score as the initial edge weight, the node identifier corresponding to each person's identity information in the personnel identity information sequence is selected one by one to construct the initial case association graph. Real-time acquisition and processing of newly generated spatial trajectory information and communication record information during the reconnaissance process to form a standardized node feature representation sequence of the newly added data; Based on the spatial displacement vector features and communication activity features corresponding to each new node in the standardized node feature characterization sequence of newly added data, cross-comparison of temporal position and spatial coordinates is performed with the features of existing nodes to determine the spatiotemporal position similarity. Based on the determined spatiotemporal similarity, the standardized node feature representation sequence of the newly added data is input into the TabPFN network. The similarity value of the feature vector between each newly added node and the existing nodes is calculated one by one by the forward inference of the TabPFN network, and the new potential association between the newly added node and the existing node is determined. Based on the newly added potential relationships, insert the new nodes and the newly added potential relationship paths between the new nodes and existing nodes into the case relationship graph; The edge weights between nodes are updated based on the similarity values ​​of the feature vectors corresponding to the newly added potential relationships, resulting in a dynamically updated case association graph.

5. The computer-aided processing method for criminal investigation based on big data analysis according to claim 1, characterized in that, The modified association probability score is obtained by weighting the initial association probability score output by the TabPFN network using the attention intensity values ​​in the attention focus matrix. Specifically: Extract the attention intensity value of each node to other nodes in the attention focus matrix, and match them one by one with the initial association probability score output by the TabPFN network according to the node identifier; For each pair of nodes, the initial association probability score and the corresponding attention intensity value are used to calculate the ratio of the attention intensity value to the initial association probability score, thus obtaining the initial association correction factor. Based on the initial association correction factor corresponding to each pair of nodes, calculate the statistical mean of the initial association correction factors among all nodes one by one, and determine a unified association correction coefficient. Based on a unified correlation correction coefficient and the attention intensity value between each pair of nodes in the attention focus matrix, the correlation probability correction weight of each pair of nodes is calculated one by one. The weights are adjusted based on the association probability between nodes, and the adjusted association probability score between nodes is calculated one by one.

6. The computer-aided processing method for criminal investigation based on big data analysis according to claim 1, characterized in that, The method involves updating the weights of nodes and associated paths in the case association graph in real time based on the corrected association probability score, and generating dynamic suspect chain data based on the updated weights. Specifically: Based on the corrected correlation probability score, the original edge weights of the paths between every two nodes in the case correlation graph are replaced one by one; Using the updated graph edge weights as input, a depth-first search method is used to traverse all associated paths in the case association graph starting from each node, and the cumulative probability value of each path is calculated. Based on the cumulative probability values ​​obtained from all paths, calculate the standardized probability value for each path one by one; Based on the calculated path standardization probability values, each path is determined and marked as belonging to a suspected chain; For each path marked as a suspected chain, extract the personnel identity information identifiers corresponding to each node in the path, the corrected correlation probability scores of the paths between nodes after real-time updates, and the topological structure of the path, and extract the suspected chain data one by one. Based on the generated suspect chain data, a graphical representation of the chain topology is established; At preset real-time push intervals, the data of each suspected chain and its corresponding graphical representation of the chain topology are pushed to the investigator's terminal in the form of a data stream.

7. The computer-aided processing method for criminal investigation based on big data analysis according to claim 1, characterized in that, The receiving of feedback confirmation data after on-site investigation by investigators based on dynamic suspect chain data, and the dynamic optimization and updating of the attention focus matrix, specifically: It receives feedback confirmation data sent by investigators' terminals in real time. The feedback confirmation data includes on-site investigation confirmation marker information for each dynamic chain of suspicion. The on-site investigation confirmation marker information is divided into valid confirmation marker information and invalid confirmation marker information. For suspected chains marked as validly confirmed in the on-site investigation and confirmation information, the attention intensity value between the corresponding nodes of the path is increased successively according to the number of times the chain is validly confirmed; For suspected chains marked as invalid confirmations in the on-site investigation and confirmation information, the attention intensity values ​​between the corresponding nodes in the path are reduced successively according to the number of times the chain is invalidally confirmed; Based on the changes in the attention intensity values ​​of all suspected chains, the overall attention intensity change of the attention focus matrix is ​​statistically analyzed and calculated one by one, and the attention intensity change is compared with the preset dynamic optimization threshold. When the magnitude of the change in attention intensity exceeds the preset dynamic optimization threshold, the parameters of the attention focus matrix are automatically optimized globally. When the change in attention intensity does not exceed the preset dynamic optimization threshold, the adjustment results of attention intensity values ​​between local nodes that have already been executed are maintained, and no global dynamic optimization update is performed. Based on the optimized and updated state of the attention focus matrix parameters, output the optimized and updated attention focus matrix.

8. A computer-aided processing system for criminal investigation based on big data analysis, comprising executing the computer-aided processing method for criminal investigation based on big data analysis as described in any one of claims 1 to 7, characterized in that, include: The case investigation data collection and initial characterization module is used to collect the identity information, spatial trajectory information and communication record information of the persons involved in the case to be investigated, and integrate them to generate the initial characterization of the case investigation data; The TabPFN network inference module is used to receive the initial representation of the investigation data, perform fast forward inference, and output the initial association probability scores between nodes. The case association graph construction and dynamic update module is used to construct an initial case association graph based on the initial association probability score output by the TabPFN network inference module, and dynamically update the nodes and association paths in the case association graph based on the newly collected case data in real time. The attention focus matrix generation and update module is used to perform biomimetic visual attention mechanism processing on the dynamically updated case association map, generate the attention focus matrix, and dynamically update the attention focus matrix based on the information of newly added nodes and associated paths in real time. The association probability correction calculation module is used to perform weighted calculations on the initial association probability scores output by the TabPFN network inference module based on the attention intensity values, so as to obtain the corrected association probability scores between each node. The dynamic suspect chain generation and visualization push module is used to update the edge weights of nodes and related paths in the case association graph in real time based on the corrected association probability scores, and generate dynamic suspect chain data based on the updated edge weights, and push it to the investigator's terminal in real time in the form of a graphical representation of a chain topology. The feedback confirmation data receiving and attention focus matrix dynamic optimization module is used to receive feedback confirmation data in real time after the investigator's terminal confirms the case based on the dynamic suspect chain data. The confirmation data is used to dynamically optimize and update the parameters of the attention focus matrix in the attention focus matrix generation and update module.

Images