A multi-case collision method based on a person involved in a case file

Abstract

The application relates to the technical field of case analysis, and discloses a multi-case collision method based on a case-involved personnel archive, which comprises the following steps: integrating multi-source heterogeneous data such as case-involved personnel, time tracks and geographical positions; constructing a case three-dimensional digital twin through hierarchical feature extraction and cross-modal alignment technology; discovering the cause-effect correlation between case elements based on a data driving and field knowledge double channel; automatically filtering illegal logical connections through a reality constraint library after constructing an initial cause-effect graph to generate a credible cause-effect network conforming to law enforcement specifications. Through the construction of a constraint-enhanced deduction framework, the application realizes physical logic compliance guarantee of counterfactual scenarios on the basis of maintaining the full-element digital reconstruction and dynamic correlation mining capacity of cases, generates an investigation strategy scheme conforming to reality law enforcement constraints and possessing tactical innovation, and significantly improves the executability and decision credibility of the deduction result in a complex actual combat environment.

Description

Technical Field

[0001] The present invention relates to the field of case analysis, and more particularly to a multi-case collision method based on files of persons involved in the case. Background Art

[0002] In response to the new challenges posed to investigative technology by the current trend of cross-regional, organized, and intelligent criminal cases, it is necessary to conduct a review and deduction of major criminal cases, analyze the motives of the cases and the causal reasoning of the case elements;

[0003] In retrospective learning, due to cross-departmental data barriers, information loss leads to information loss in calculations. This application uses constraint-enhanced causal reasoning and reinforcement learning strategies, the NOTEARS algorithm and PPO optimization to automatically filter illegal connections in computer systems, improving processing speed and compliance. Summary of the Invention

[0004] The present invention provides a multi-case collision method based on the archives of persons involved in the case, which solves the technical problems in related technologies such as information loss caused by causal intervention operations not taking into account cross-departmental data barriers, and excessive reliance on statistical utility by strategy evaluation indicators while ignoring compliance verification, resulting in a disconnect between deduction results and actual combat needs.

[0005] The present invention provides a multi-case collision method based on files of persons involved in the case, comprising the following steps:

[0006] S100, multimodal spatiotemporal fusion modeling: Integrates multi-source heterogeneous data such as individuals involved in the case, time trajectories, and geographic locations, and constructs a three-dimensional digital twin of the case through hierarchical feature extraction and cross-modal alignment technology;

[0007] S200, Constraint-Enhanced Causal Reasoning: This method uses a dual-channel approach of data-driven and domain knowledge to discover causal relationships between case elements. After constructing an initial causal graph, it automatically filters illegal logical connections through a real-world constraint library to generate a credible causal network that complies with law enforcement regulations.

[0008] S300, Robust Reinforcement Learning Strategy Deduction: Models the reconnaissance environment as a constrained Markov decision process, designs a dual-channel strategy network that integrates graph structure and temporal features, iteratively optimizes the reconnaissance strategy through an improved reinforcement learning algorithm, and simultaneously embeds a dynamic constraint penalty mechanism to ensure strategy feasibility;

[0009] S400, Constrained Counterfactual Analysis: Perform multi-dimensional intervention operations on the compliance causal graph to generate counterfactual scenarios, screen physically logically feasible alternative strategies through dual constraint detection, combine utility evaluation with spatiotemporal consistency testing to build a multi-dimensional credibility evaluation system, and output a visual deduction decision matrix.

[0010] Furthermore, in S100, the following steps are specifically included:

[0011] S110, multi-dimensional spatiotemporal feature extraction: construct a hierarchical spatiotemporal encoder to separate and extract features of different dimensions;

[0012] S120, cross-modal alignment correction: eliminating modality differences through spatiotemporal adversarial correction networks;

[0013] S130, Hierarchical Attention Fusion: Fusion of Multi-level Spatiotemporal Features.

[0014] Furthermore, in S200, the following steps are specifically included:

[0015] S210, Data-driven Causal Discovery: Learning Initial Causal Structures Based on the NOTEARS Algorithm;

[0016] S220, Knowledge-guided Causal Discovery: Injecting domain knowledge rules to strengthen key relationships;

[0017] S230, Constraint Compliance Verification: Apply realistic constraints to filter out illegal associations.

[0018] Furthermore, in S300, the following steps are specifically included:

[0019] S310, Constrained MDP Modeling: Constructing an enhanced Markov decision process that complies with policing constraints;

[0020] S320, policy network initialization: building a dual-channel policy network architecture;

[0021] S330, Constrained Enhanced Strategy Optimization: uses the improved PPO algorithm for strategy training;

[0022] S340, Policy Compliance Verification: Ensure that the output policy meets real-world constraints.

[0023] Furthermore, in S310, the calculation formula for constrained MDP modeling is as follows:

[0024] State space definition:

[0025] ;

[0026] in represents graph tensor concatenation, is the time step, is the adjacency matrix of the compliant causal graph, is the digital twin feature at time t, represents the state vector at time t, Represents a tensor flattening operation;

[0027] Action space constraints:

[0028] ;

[0029] in is the i-th constraint, is the total number of constraints, is the kth action, is the indicator function, represents the action space;

[0030] Compound reward function:

[0031] ;

[0032] in To capture the reward weight, To capture the reward, is the clue reward weight, is the clue reward weight, is the penalty coefficient of the kth constraint, is the indicator function of action violation constraint, Indicates the Step reward.

[0033] Furthermore, in S320, the calculation formula for policy network initialization is as follows:

[0034] Graph feature extraction:

[0035] ;

[0036] in is a graph convolutional network, is the weight of the lth layer, is the l-th layer graph feature;

[0037] Time series feature extraction:

[0038] ;

[0039] in Represents the characteristics of the fused case digital twin, is the time step, is a bidirectional long short-term memory network, is the weight of the lth layer, is the temporal feature of the lth layer;

[0040] Strategic Network Convergence:

[0041] ;

[0042] in Indicates that the status Take action The probability of is the policy network weight, is the bias term, represents vector concatenation, For parameterized strategies, For action, For status, represents the 4th layer graph features, Indicates the timing characteristics of the 24th layer.

[0043] Furthermore, in S330, the calculation formula for constraint enhancement strategy optimization is as follows:

[0044] Constrained objective function:

[0045] ;

[0046] in is the reward discount factor, is the constraint weight, is the entropy coefficient, is the strategy parameter, is the number of time steps, is the number of constraints, is the reward at time t, is the policy entropy, is the kth constraint, Indicates the number of iterations.

[0047] Furthermore, in S340, the calculation formula for policy compliance verification is as follows:

[0048] Action filtering mechanism:

[0049] ;

[0050] in is the action probability ranking, is the constraint satisfaction judgment function, Indicates action Whether the constraints are satisfied , is the action-value function, is the value threshold, is the state at time t, is the kth action, represents the set of possible actions in step t;

[0051] Strategy distillation compression:

[0052] ;

[0053] in is the distillation temperature coefficient, is the distillation weight, is the strategy for the mth iteration, represents a set of multi-round iterative strategies, To be the best compliance strategy, , Represents knowledge distillation.

[0054] Furthermore, in S400, the following steps are specifically included:

[0055] S410, Causal Intervention Counterfactual Generation: Multi-dimensional intervention operations on the compliance graph;

[0056] S420, feasible strategy distillation: filtering counterfactual strategies that meet realistic constraints;

[0057] S430, Composite Credibility Assessment: Multidimensional Assessment of Integration Effectiveness and Compliance;

[0058] S440, Visual Decision Matrix Construction: Generate explainable deduction reports.

[0059] Furthermore, in S440, the calculation formula of the interpretable deduction report is as follows:

[0060] ;

[0061] in is the number of feasible scenarios, is the number of constraints, For strategic effects, For credibility, represents the decision matrix, represents the jth strategy, Represents the default rate.

[0062] The beneficial effects of the present invention are:

[0063] By constructing a constraint-enhanced deduction framework, the present invention achieves physical and logical compliance assurance for counterfactual scenarios while maintaining the ability to digitally reconstruct all elements of a case and conduct dynamic correlation mining. This generates investigative strategy plans that are both compliant with real-world law enforcement constraints and tactically innovative, significantly improving the feasibility and decision-making credibility of deduction results in complex actual combat environments, and promoting the paradigm shift of investigative command from experience-based to data-intelligent driven. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a flow chart of a multi-case collision method based on files of persons involved in the case proposed by the present invention;

[0065] Figure 2 The present invention Figure 1 Flow chart of sub-steps in S100;

[0066] Figure 3 The present invention Figure 1 Flow chart of sub-steps in S200;

[0067] Figure 4 The present invention Figure 1 Flow chart of sub-steps in S300;

[0068] Figure 5 The present invention Figure 1 Flowchart of sub-steps in S400. DETAILED DESCRIPTION

[0069] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described for some examples may be combined in other examples.

[0070] like Figure 1-Figure 5 As shown, a multi-case collision method based on the files of persons involved in the case includes the following steps:

[0071] S100, multimodal spatiotemporal fusion modeling: Integrates multi-source heterogeneous data such as case personnel, time trajectories, and geographic locations, and constructs a digital twin of the case through hierarchical feature extraction and cross-modal alignment technology;

[0072] In one embodiment of the present invention, the following steps are specifically included:

[0073] S110, multi-dimensional spatiotemporal feature extraction: constructing a hierarchical spatiotemporal encoder (MGSTE) to separate and extract features of different dimensions;

[0074] enter: ;

[0075] in Indicates the The spatial distribution characteristics of individuals, , The number of people involved in the case, Indicates the The temporal characteristics of a time slice, , is the number of time slices, Indicates the Spatial information of a geographic location, , is the geographical location number;

[0076] Personnel relationship code:

[0077] ;

[0078] Where GAT represents the graph attention network, represents the spatial features of people output by GAT, represents the number of attention heads, represents the weight matrix of the first layer of GAT, Represents the spatial distribution input of personnel;

[0079] Temporal Mode Encoding:

[0080] ;

[0081] in represents the time window, represents the position encoding dimension, Represents the model dimension, represents the number of attention heads, represents the time series feature input, represents the temporal Transformer model, Represents the temporal feature encoding output;

[0082] Geospatial coding:

[0083] ;

[0084] in represents the GeoHash-based spatial convolutional encoder, Indicates the spatial aggregation radius, with a value of 500 meters. Indicates the GeoHash encoding length, the value is 32. Indicates the input of geographic location information. Represents geospatial feature encoding output;

[0085] Fusion output:

[0086] ;

[0087] in is the number of samples, is the feature dimension after splicing, is layer normalization, is the feature splicing operation, Represents the fused multimodal features;

[0088] S120, Cross-modal Alignment Correction: Eliminating modality differences through spatiotemporal adversarial correction networks (STACN);

[0089] Generator Network:

[0090] ;

[0091] ;

[0092] in is the weight matrix of the generator, is the generator function, is the input feature, For the generated alignment features, Represents a set of generator parameters;

[0093] Discriminator network:

[0094] ;

[0095] in is the sigmoid function, represents vector concatenation, is the discriminator function, is the discriminator weight matrix, , For the generated alignment features, Represents the fused multimodal features;

[0096] Spatiotemporal consistency constraints:

[0097] ;

[0098] in is the geographic consistency weight, is the time consistency weight, is the spherical distance, (Dynamic Time Warping) is the dynamic time warping distance, are the generated and real geographic locations, are the generated and real time series features respectively;

[0099] Adversarial training objectives:

[0100] ;

[0101] in For the generator, is the discriminator, Represents the fused multimodal features, is the spatiotemporal consistency loss, represents the expectation operation;

[0102] S130, Hierarchical Attention Fusion: Fusion of multi-level spatiotemporal features;

[0103] Spatial Attention:

[0104] ;

[0105] ;

[0106] in represents the spatial attention matrix, represents the normalization function, represents the spatial query vector, represents the spatial key vector, 、 represents the spatial attention weight matrix, represents the spatial feature dimension, Represents fusion feature input;

[0107] Time Attention:

[0108] ;

[0109] ;

[0110] in represents the temporal attention matrix, represents a causal mask operation, represents the time query vector, represents the time key vector, 、 represents the temporal attention weight matrix, Represents the time feature dimension;

[0111] Cross-modal fusion:

[0112] ;

[0113] ;

[0114] in Represents the characteristics of the fused case digital twin, ,in represents the hidden layer dimension, represents a time slice, represents the number of fusion layers, Indicates the Layer fusion weights, Indicates the Layers can learn weight parameters, Indicates the Layers can learn weight parameters, Indicates the Layer spatial attention, Indicates the Layer-time attention, represents element-wise product, Indicates the Layer feature vector, , Indicates the number of samples;

[0115] S200, Constraint-Enhanced Causal Reasoning: Discovers causal relationships between case elements based on data-driven and domain knowledge-based approaches. After constructing an initial causal graph, it automatically filters illegal logical connections through a real-world constraint library to generate a compliant causal graph adjacency matrix.

[0116] In one embodiment of the present invention, the following steps are specifically included:

[0117] S210, Data-driven Causal Discovery: Learning Initial Causal Structures Based on the NOTEARS Algorithm;

[0118] Structural Equation Modeling:

[0119] ;

[0120] in represents the structural equation output, represents a parameterized structural equation model, represents the set of model parameters, Indicates the layer weight matrix, , represents the Gaussian error linear unit activation function, represents the Gaussian noise term;

[0121] Acyclic Constraint Optimization:

[0122] ;

[0123] ;

[0124] in is the DAG constraint weight, is the target dimension, is the predicted value, is the Frobenius norm, is the Hadamard product, is the trace of the matrix, represents the matrix exponential, is the weight matrix, represents the acyclic constraint regularization term;

[0125] Intermediate output:

[0126] ;

[0127] in represents the data-driven causal adjacency matrix, , Indicates the number of nodes, represents the third layer weight matrix;

[0128] S220, Knowledge-guided Causal Discovery: Injecting domain knowledge rules to strengthen key relationships;

[0129] Rule code:

[0130] ;

[0131] in is the rule encoding weight matrix, For the knowledge base, is a multi-layer perceptron, represents the rule encoding matrix;

[0132] Knowledge Enhanced Fusion:

[0133] ;

[0134] in represents element-wise product, is the sigmoid activation function, is the data-driven adjacency matrix, is the rule encoding matrix, represents the knowledge enhancement fusion matrix, ;

[0135] Importance reweighting:

[0136] ;

[0137] in is the temperature coefficient, is the (i,j) association rule in the rule set, Enhance the adjacency matrix elements for knowledge, represents the reweighted knowledge-enhanced adjacency matrix element, Indicates the rule importance score;

[0138] S230, constraint compliance verification: applying realistic constraints to filter out illegal associations;

[0139] Constraint matching detection:

[0140] ;

[0141] in Represents an edge The compliance flag (1 for compliance, 0 for non-compliance), Is the constraint violation indicator function, which is 1 if the condition is met, otherwise it is 0. is the edge (i,j), is the total number of constraints, is the kth constraint;

[0142] Compliance edge weight calculation:

[0143] ;

[0144] ;

[0145] in is the mean, is the standard deviation, for The historical characteristics of the node, Enhance the adjacency matrix elements for knowledge, To constrain the matching results, Indicates compliance edge rights, represents the normalization function;

[0146] Sparse processing:

[0147] ;

[0148] ;

[0149] in is the thinning threshold, is the number of nodes, To comply with the boundary rights, are diagonal elements (self-loop weights), express The standard deviation of Indicates the final compliance edge right;

[0150] Final output: compliant causal graph adjacency matrix ;

[0151] S300, Robust Reinforcement Learning Strategy Deduction: Modeling the investigative environment as a constrained Markov decision process, designing a dual-channel strategy network that integrates the compliance causal graph adjacency matrix and case digital twin features, iteratively optimizing the investigative strategy through an improved reinforcement learning algorithm, and simultaneously embedding a dynamic constraint penalty mechanism to ensure strategy feasibility;

[0152] In one embodiment of the present invention, the specific steps are as follows:

[0153] S310, Constrained MDP Modeling: Constructing an enhanced Markov decision process that complies with policing constraints;

[0154] State space definition:

[0155] ;

[0156] in represents graph tensor concatenation, is the time step, is the adjacency matrix of the compliant causal graph, is the digital twin feature at time t, represents the state vector at time t, Represents a tensor flattening operation;

[0157] Action space constraints:

[0158] ;

[0159] Retention satisfaction The 9 core actions of restraint;

[0160] in is the i-th constraint, is the total number of constraints, is the kth action, is the indicator function, represents the action space;

[0161] Compound reward function:

[0162] ;

[0163] in To capture the reward weight, To capture the reward, is the clue reward weight, is the clue reward weight, is the penalty coefficient of the kth constraint, is the indicator function of action violation constraint, Indicates the Step reward;

[0164] S320, policy network initialization: building a dual-channel policy network architecture;

[0165] Graph feature extraction:

[0166] ;

[0167] in is a graph convolutional network, is the weight of the lth layer, is the l-th layer graph feature;

[0168] Time series feature extraction:

[0169] ;

[0170] in Represents the characteristics of the fused case digital twin, is the time step, is a bidirectional long short-term memory network, is the weight of the lth layer, is the temporal feature of the lth layer;

[0171] Strategic Network Convergence:

[0172] ;

[0173] in Indicates that the status Take action The probability of is the policy network weight, is the bias term, represents vector concatenation, For parameterized strategies, For action, For status, represents the 4th layer graph features, Indicates the 24th layer timing characteristics;

[0174] S330, Constrained Enhanced Strategy Optimization: uses the improved PPO algorithm for strategy training;

[0175] Constrained objective function:

[0176] ;

[0177] in is the reward discount factor, is the constraint weight, is the entropy coefficient, is the strategy parameter, is the number of time steps, is the number of constraints, is the reward at time t, is the policy entropy, is the kth constraint, Indicates the number of iterations;

[0178] Adaptive penalty mechanism:

[0179] ;

[0180] in is the adjustment rate, is the initial value, As the basic increment, is the number of iterations, For the The constraint violation rate of iterations, represents the constraint weight of the n+1th iteration, represents the constraint weight of the nth iteration;

[0181] Importance Sampling Updates:

[0182]

[0183] in is the clipping threshold, is the number of iterations, is the KL divergence coefficient, is the advantage function, For the old strategy, For the current strategy, is the KL divergence, is the strategy parameter, Indicates the Strategy parameters for iterations;

[0184] S340, Policy Compliance Verification: Ensures that the output policy meets real-world constraints;

[0185] Action filtering mechanism:

[0186] ;

[0187] in is the action probability ranking, is the constraint satisfaction judgment function, Indicates action Whether the constraints are satisfied , is the action-value function, is the value threshold, is the state at time t, is the kth action, represents the set of possible actions in step t;

[0188] Strategy distillation compression:

[0189] ;

[0190] in is the distillation temperature coefficient, is the distillation weight, is the strategy for the mth iteration, represents a set of multi-round iterative strategies, To be the best compliance strategy, , Represents knowledge distillation;

[0191] S400, Constrained Counterfactual Analysis: This system generates counterfactual scenarios by performing multi-dimensional intervention operations on the compliance causal graph. It then uses dual constraint detection to filter counterfactual strategies that meet real-world constraints. It then combines utility evaluation with spatiotemporal consistency testing to construct a multi-dimensional credibility evaluation system, outputting a visual decision matrix.

[0192] In one embodiment of the present invention, the following steps are specifically included:

[0193] S410, Causal Intervention Counterfactual Generation: Multi-dimensional intervention operations on the compliance graph;

[0194] Definition of intervention budget:

[0195] ;

[0196] in For the intervention mask, is the noise distribution, is the noise variance, is the counterfactual map, is the tensor product, is the element-wise product, Indicates causal interference equation;

[0197] Constraint Enhancement Generation:

[0198] ;

[0199] ;

[0200] in For the The adaptive weights of the constraints, For constraints The number of violations, is the Frobenius norm, is the rectified linear unit, represents the counterfactual generation loss, Represents the counterfactual graph pair constraint the extent of the violation;

[0201] S420, feasible strategy distillation: filtering counterfactual strategies that meet realistic constraints;

[0202] Double Binding Detection:

[0203] ;

[0204] in is the strict threshold, the minimum number that satisfies the constraint, is the total number of constraints, is the actual threshold, is the indicator function, is the kth constraint, is the counterfactual map, represents the set of feasible counterfactual scenarios;

[0205] Strategy effectiveness evaluation:

[0206] ;

[0207] ;

[0208] in is the default rate, is the reward of the j-th strategy at time t, is the i-th sample, is the discount factor, is the proportion of the j-th strategy violating the k-th constraint, represents the effectiveness score of the j-th counterfactual strategy, represents the number of time steps;

[0209] S430, Composite Credibility Assessment: Multidimensional Assessment of Integration Effectiveness and Compliance;

[0210] Shapley value correction:

[0211] ;

[0212] in is the original Shapley value, For the The proportion of strategies violating constraints, is the total number of constraints, represents the modified Shapley value of the j-th strategy;

[0213] Spatiotemporal consistency test:

[0214] ;

[0215] in is the structural similarity, For the Time-to-time counterfactual predictions, is the digital twin at time t, represents the spatiotemporal consistency score of the j-th strategy;

[0216] Final credibility calculation:

[0217] ;

[0218] ;

[0219] in is a stability indicator, is the number of time steps, is the action of the j-th strategy at time t, is the modified Shapley value, is the consistency score, represents the final credibility score of the j-th strategy;

[0220] S440, Visual Decision Matrix Construction: Generate explainable deduction reports;

[0221] 3D heat map encoding:

[0222] ;

[0223] in is the number of feasible scenarios, is the number of constraints, For strategic effects, For credibility, represents the decision matrix, represents the jth strategy, represents the default rate;

[0224] Pareto front analysis:

[0225] ;

[0226] in is the utility of the i-th strategy, is the credibility of the i-th strategy, is the jth counterfactual strategy, represents the Pareto frontier strategy set, represents the i-th counterfactual strategy, represents the utility of the j-th strategy;

[0227] Final output: set of feasible counterfactual scenarios (retain an average of 12 scenarios), decision matrix ;

[0228] In one embodiment of the present invention, after obtaining the final output, the technical process of the execution phase is as follows:

[0229] Step 1: Strategy recommendation and situation adaptation:

[0230] Multi-dimensional strategy ranking:

[0231] Decision matrix based on the above output , generate recommendation sequences according to confidence and utility:

[0232] ;

[0233] in is the credibility of the j-th strategy, is the utility score of the j-th strategy, Indicates the The priority score of each strategy;

[0234] Output: Top 3 recommended strategy sets ;

[0235] Context suitability calculation:

[0236] Combined with real-time police deployment Fine-tune your strategy:

[0237] ;

[0238] in is the police deployment action set at time t, is the consistency of the j-th strategy, is the jth counterfactual strategy, Indicates the The contextual adaptability of a strategy, Indicates the cardinality of a set;

[0239] Step 2: Human-machine collaborative decision verification:

[0240] Visual decision sandbox: Loading on the digital twin platform ;

[0241] Overlay display: police force heat map, traffic network, communication base stations and other real-time data layers;

[0242] Hybrid Enhancement Decision:

[0243] ;

[0244] in Recommended strategy for Top, For artificial strategies, is the mixed scoring function, represents the final mixed decision strategy;

[0245] Scoring function:

[0246] ;

[0247] in Scoring machine learning, Expert confidence level;

[0248] Step 3: Dynamic execution and monitoring:

[0249] The policy instructions are translated as follows:

[0250]

[0251] Real-time effect monitoring:

[0252] Build execution effectiveness index:

[0253] ;

[0254] in is the number of clues found at time t, is the total number of clues, is the number of hours of delay at time t, is the execution effect index at time t;

[0255] Anomaly Detection: When When the warning is triggered;

[0256] Step 4: Feedback-driven model evolution:

[0257] Incremental learning mechanism:

[0258] ;

[0259] in represents the updated model parameters, represents the original model parameters, is the learning rate, are model parameters, is the number of strategies, For the A counterfactual strategy, To actually execute the strategy, represents the reality gap measurement function;

[0260] Reality Gap Metrics:

[0261] ;

[0262] in For the counterfactual reward, For actual rewards, is the number of time steps, for constraint satisfaction judgment;

[0263] Knowledge base dynamic updates:

[0264] New rules: ;

[0265] in Indicates a new rule set. Indicates new rules, Indicates the rule support;

[0266] Constraint library extensions: ;

[0267] in Indicates a new constraint set. represents the new constraint, Indicates the number of violations;

[0268] In one embodiment of the present invention, the following actual combat verification case is given:

[0269] In a cross-border drug case:

[0270] Strategy Execution:

[0271] Deployed the top two recommended strategies (communication tracing + logistics monitoring) within 72 hours, generating 37 specific directives involving four law enforcement agencies;

[0272] Execution effect:

[0273] Amount of drugs seized: The predicted value was 83.5kg, while the actual amount seized was 81.2kg (error <3%);

[0274] Suspect location speed: 2.7 times faster than traditional methods;

[0275] Feedback optimization: added 2 new transport route constraint rules and updated the regional feature library of the spatiotemporal encoder.

[0276] The above describes the embodiments of the present invention, but the present invention is not limited to the above specific implementation methods. The above specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make many forms under the guidance of the present invention, all of which are protected by the present invention.

Claims

1. A multi-case collision method based on the files of persons involved in the case, characterized in that: Executed by a computer system, including the following steps: S100, multimodal spatiotemporal fusion modeling: Integrates heterogeneous data from multiple sources, including case personnel, time trajectories, and geographic locations, and constructs a digital twin of the case through hierarchical feature extraction and cross-modal alignment technology. Hierarchical feature extraction uses a graph neural network and a temporal Transformer model. S200, Constraint-Enhanced Causal Reasoning: Discovers causal relationships between case elements based on data-driven and domain knowledge-based approaches. After constructing an initial causal graph, it automatically filters illegal logical connections through a real-world constraint library to generate a compliant causal graph adjacency matrix. S300, Robust Reinforcement Learning Strategy Deduction: Modeling the investigative environment as a constrained Markov decision process, designing a dual-channel strategy network that integrates the compliance causal graph adjacency matrix and case digital twin features, iteratively optimizing the investigative strategy through an improved reinforcement learning algorithm, and simultaneously embedding a dynamic constraint penalty mechanism to ensure strategy feasibility; In S300, the following steps are specifically included: S310, Constrained MDP Modeling: Constructing an enhanced Markov decision process that complies with policing constraints; S320, policy network initialization: building a dual-channel policy network architecture; In S320, the calculation formula for policy network initialization is as follows: Graph feature extraction: ； in is a graph convolutional network, is the weight of the lth layer, is the l-th layer graph feature, is the adjacency matrix of the compliance causal graph; Time series feature extraction: ； in Represents the characteristics of the fused case digital twin, is the time step, is a bidirectional long short-term memory network, is the weight of the lth layer, is the temporal feature of the lth layer; Strategic Network Convergence: ； in Indicates that the status Take action The probability of is the policy network weight, is the bias term, represents vector concatenation, For parameterized strategies, For action, For status, represents the 4th layer graph features, Indicates the 24th layer timing characteristics; S330, Constrained Enhanced Strategy Optimization: uses the improved PPO algorithm for strategy training; S340, Policy Compliance Verification: Ensures that the output policy meets real-world constraints; S400, Constrained Counterfactual Analysis: Perform multi-dimensional intervention operations on the compliance causal graph to generate counterfactual scenarios, filter counterfactual strategies that meet real-world constraints through dual constraint detection, combine utility evaluation with spatiotemporal consistency testing to build a multi-dimensional credibility evaluation system, and output a visual strategy decision matrix.

2. A multi-case collision method based on files of persons involved in a case according to claim 1, characterized in that: In S100, the following steps are specifically included: S110, multi-dimensional spatiotemporal feature extraction: construct a hierarchical spatiotemporal encoder to separate and extract features of different dimensions; S120, cross-modal alignment correction: eliminating modality differences through spatiotemporal adversarial correction networks; S130, Hierarchical Attention Fusion: Fusion of Multi-level Spatiotemporal Features.

3. A multi-case collision method based on files of persons involved in a case according to claim 2, characterized in that: In S200, the following steps are specifically included: S210, Data-driven Causal Discovery: Learning Initial Causal Structures Based on the NOTEARS Algorithm; S220, Knowledge-guided Causal Discovery: Injecting domain knowledge rules to strengthen key relationships; S230, Constraint Compliance Verification: Apply realistic constraints to filter out illegal associations.

4. The multi-case collision method based on the files of persons involved in the case according to claim 1 is characterized in that: In S310, the calculation formula for constrained MDP modeling is as follows: State space definition: ； in represents graph tensor concatenation, is the time step, is the adjacency matrix of the compliant causal graph, is the digital twin feature at time t, represents the state vector at time t, Represents a tensor flattening operation; Action space constraints: ； in is the i-th constraint, is the total number of constraints, is the kth action, is the indicator function, represents the action space, For action; Compound reward function: ； in To capture the reward weight, To capture the reward, is the clue reward weight, is the clue reward weight, is the penalty coefficient of the kth constraint, is the indicator function of action violation constraint, Indicates the Step reward.

5. The multi-case collision method based on the files of persons involved in the case according to claim 1 is characterized in that: In S330, the calculation formula for constraint enhancement strategy optimization is as follows: Constrained objective function: ； in is the reward discount factor, is the constraint weight, is the entropy coefficient, is the strategy parameter, is the number of time steps, is the number of constraints, is the reward at time t, is the policy entropy, is the kth constraint, Indicates the number of iterations.

6. The multi-case collision method based on the files of persons involved in the case according to claim 1 is characterized in that: In S340, the calculation formula for policy compliance verification is as follows: Action filtering mechanism: ； in is the action probability ranking, is the constraint satisfaction judgment function, Indicates action Whether the constraints are satisfied , is the action-value function, is the value threshold, is the state at time t, is the kth action, represents the set of possible actions in step t; Strategy distillation compression: ； in is the distillation temperature coefficient, is the distillation weight, is the strategy for the mth iteration, represents a set of multi-round iterative strategies, To be the best compliance strategy, , Represents knowledge distillation.

7. The multi-case collision method based on the files of persons involved in the case according to claim 1 is characterized in that: In S400, the following steps are specifically included: S410, Causal Intervention Counterfactual Generation: Multi-dimensional intervention operations on the compliance graph; S420, feasible strategy distillation: filtering counterfactual strategies that meet realistic constraints; S430, Composite Credibility Assessment: Multidimensional Assessment of Integration Effectiveness and Compliance; S440, Visual Decision Matrix Construction: Generate explainable deduction reports.

8. The multi-case collision method based on the files of persons involved in the case according to claim 7 is characterized in that: In S440, the calculation formula for the interpretable deduction report is as follows: ； in is the number of feasible scenarios, is the number of constraints, For strategic effects, For credibility, represents the decision matrix, represents the jth strategy, Represents the default rate.

Images