Event clue dynamic studying and judging method based on large model agent

Abstract

The invention discloses an event clue dynamic study and judgment method based on a large model agent. The method comprises the steps that multi-modal event clues generated by the large model agent are acquired, and unified structured data input is formed; event clue key semantic feature vectors are extracted and coded into initial quantum state representation; executing quantum entanglement evolution by using a parameterized quantum gate to calculate an attention weight matrix; carrying out quantum characteristic field modeling and defining a dynamic clue potential energy function; determining a clue convergent point and an event causal path to construct an event association graph; the event causal correlation degree is evaluated in real time through quantum probabilistic reasoning, and an event evolution path is obtained; and adaptively adjusting the attention weight matrix based on the semantic feature similarity, carrying out iterative calculation until convergence, and outputting a final event clue evolution path. According to the method, the dynamic association study and judgment precision and real-time performance of heterogeneous event clues are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of multi-modal data analysis, and particularly relates to an event clue dynamic research and judgment method based on a large model agent. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, multi-modal information analysis technology relying on large-scale artificial intelligence models is increasingly mature and widely used in event monitoring and research and judgment fields. Currently, multi-modal event clues generated by large-scale artificial intelligence models usually include text, images, speech, and various sensor data. The existing technology usually first converts event clues of different modalities into structured data, realizes the reasoning of the causal relationship between events and the prediction of the evolution path through semantic feature extraction, fusion, and correlation analysis methods, so as to meet the needs of event monitoring and research and judgment in complex scenarios.

[0003] At present, the correlation analysis method for multi-modal event clues mainly includes feature fusion analysis methods based on attention mechanisms of deep learning networks and graph neural network structures. Among them, the attention mechanism calculates the semantic similarity weight between event clues to quantify the influence degree of different event clues on the target event; the graph neural network models the correlation relationship between event clues by using a graph structure to realize the fusion analysis between clues and the prediction of the event evolution path. However, the existing attention mechanism and graph neural network usually have problems such as low feature fusion efficiency, unstable attention weight calculation, and insufficient accuracy of dynamic correlation analysis of event clues when processing multi-modal event clues generated by large-scale artificial intelligence models, which makes it difficult to accurately identify the implicit nonlinear correlation characteristics and the potential event causal relationship evolving over time. Since the existing technology mainly uses classical neural networks and traditional probabilistic reasoning methods for feature fusion and event evolution path prediction, the computational efficiency is insufficient when processing a large amount of heterogeneous event clues in real time, and it is difficult to meet the needs of real-time and accuracy in high-dynamic and multi-change event research and judgment scenarios.

[0004] Therefore, how to provide an event clue dynamic research and judgment method based on a large model agent is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide an event clue dynamic research and judgment method based on a large model agent. In view of the low accuracy of the existing technology in real-time and accurate correlation analysis of heterogeneous multi-modal event clues, a dynamic research and judgment method based on quantum adaptive attention deep fusion quantum neural field network is proposed, which improves the real-time correlation analysis accuracy and research and judgment efficiency of heterogeneous event clues through quantum feature field modeling, quantum probability reasoning, and adaptive adjustment of semantic feature similarity.

[0006] The event clue dynamic research and judgment method based on a large model agent according to the embodiment of the application comprises:

[0007] A multi-modal event clue dynamically generated by a large model agent is acquired, and structured data input is constructed;

[0008] Corresponding feature extraction is respectively performed on each type of event clue in the structured data input, key semantic feature vectors are obtained, and the key semantic feature vectors are encoded into corresponding initial quantum state representations;

[0009] A quantum attention weight matrix is obtained by processing the initial quantum state representations using a pre-set parameterized quantum gate;

[0010] A continuous quantum feature field modeling is performed on all event clues, and a derivable dynamic clue potential function is defined;

[0011] Local extreme points and corresponding potential energy gradient paths of the dynamic clue potential function in the continuous quantum feature field are calculated, a potential clue convergence point and an event causal path of the event clue are determined, and an event correlation graph is constructed;

[0012] According to the event correlation graph, a historical clue path, and newly added event clues in a current time window, confidence evaluation is performed to obtain an event evolution path;

[0013] Event clues corresponding to each event node in the event evolution path are mapped back to key semantic feature vectors, and the quantum attention weight matrix is adjusted according to a semantic feature similarity calculation result, and continuous quantum feature field reconstruction and event causal path calculation are iteratively performed to obtain a final event clue evolution path.

[0014] Optionally, the key semantic feature vectors are encoded into corresponding initial quantum state representations, and the encoding specifically comprises:

[0015] A multi-dimensional feature normalization processing is performed on the key semantic feature vectors of each event clue to generate normalized key semantic feature vectors;

[0016] The number of quantum bits required in the corresponding initial quantum state representation is determined according to the number of feature dimensions contained in the normalized key semantic feature vectors, and a preset amplitude value range is set for each quantum bit, and the amplitude value range is divided into a plurality of non-overlapping sub-intervals;

[0017] An initial quantum bit amplitude distribution is formed according to a pre-set corresponding relationship between the feature dimensions and the sub-intervals;

[0018] A secondary normalization operation is performed on the initial quantum bit amplitude distribution to obtain a quantum state amplitude distribution;

[0019] setting an initial phase parameter for each quantum bit in the initial quantum state representation based on the quantum state amplitude distribution;

[0020] loading the quantum state amplitude distribution and the initial phase parameter into the initial quantum state representation one by one according to the preset quantum bit arrangement order, to obtain an initial quantum state representation corresponding to the key semantic feature vector.

[0021] Optionally, the initial quantum state representation is processed by using the preset parameterized quantum gate to obtain a quantum attention weight matrix, specifically:

[0022] According to the amplitude value and the phase value of each quantum bit in the initial quantum state representation, a sorting sequence of the amplitude values of all quantum bits in the initial quantum state of each event clue from high to low, and a sorting sequence of the phase values from small to large are determined;

[0023] The all quantum bits in the initial quantum state representation of each event clue are rearranged one by one according to the alternating order of amplitude and phase sorting to form an adaptive quantum bit recombination sequence;

[0024] The adaptive quantum bit recombination sequence is input one by one to the parameterized quantum gate for initial entanglement evolution;

[0025] According to the quantum state after the initial entanglement evolution, a phased probability measurement operation is performed, and the probability distribution of the current quantum state is determined through the phased probability measurement result, and the probability value of each quantum bit after measurement is recorded;

[0026] According to the phase difference value between the phased probability measurement result and each quantum bit of the current quantum state, a secondary sorting sequence of the quantum bits is determined, and secondary quantum bit recombination is performed;

[0027] The quantum state sequence input, real-time adaptive determination of the rotation gate rotation parameter, phased probability measurement and quantum bit adaptive recombination operation are repeatedly performed to complete multiple rounds of entanglement evolution;

[0028] When the whole multiple rounds of entanglement evolution process is completed, a joint probability distribution matrix of all quantum bits is obtained by implementing joint projection measurement on the global basis of the quantum state;

[0029] According to the joint probability distribution matrix, the quantum state correlation strength value between the initial quantum states of any two event clues is calculated, and is sequentially filled into the quantum attention weight matrix according to the order of the event clues, to obtain a complete quantum attention weight matrix.

[0030] Optionally, the preset parameterized quantum gate is specifically:

[0031] The quantum gate sequence is formed by a rotation gate and a controlled NOT gate in turn and alternately connected in series, wherein the initial values of the rotation axis direction and the rotation angle parameter of the rotation gate are determined according to the centroid of the current amplitude distribution of all quantum bits in the input quantum bit reordering sequence and the standard deviation of the amplitude value.

[0032] The pairing order of the control quantum bit and the target quantum bit of the controlled NOT gate is determined according to the amplitude difference between adjacent quantum bits in the input quantum bit reordering sequence from large to small;

[0033] In each round of entanglement evolution, the rotation axis direction and the rotation angle parameter of the rotation gate are updated in real time according to the stage probability measurement value obtained in the last round of evolution, and the pairing order of the control quantum bit and the target quantum bit of the controlled NOT gate is adjusted and updated according to the quantum bit amplitude change determined by the last round of stage probability measurement value.

[0034] Optionally, the continuous quantum feature field modeling of all event clues is performed, and a derivable dynamic clue potential function is defined, specifically:

[0035] The sum of the overall attention weights of each event clue is determined according to the quantum attention weight matrix as the spatial density initial value;

[0036] The amplitude value and the phase value of the key semantic feature vector are taken as the spatial coordinate axes of the continuous quantum feature field, and the spatial position coordinates of the event clues are determined according to the spatial density initial value;

[0037] The joint coordinate points of the event clues are constructed according to the spatial position coordinates and the time labels of the event clues;

[0038] The joint coordinate points of the event clues are continuously smoothed by using a cubic spline interpolation method, and the complete spatiotemporal continuous distribution of the event clues is obtained;

[0039] The local spatial density function is determined according to the complete spatiotemporal continuous distribution;

[0040] The dynamic clue potential function is constructed based on the local spatial density function, and the function value of the dynamic clue potential function at any spatial position is the negative logarithm of the local spatial density function value at the corresponding position, and the dynamic clue potential function satisfies the continuous derivable characteristics.

[0041] Optionally, the local extreme points and the corresponding potential energy gradient paths of the dynamic clue potential function in the continuous quantum feature field are determined, the potential clue convergence points and the event causal paths of the event clues are determined, and the event correlation graph is constructed, specifically:

[0042] The spatial coordinate positions of the local extreme points are solved by using a multidimensional Newton iteration method according to the dynamic clue potential function;

[0043] tracking the gradient path in the negative gradient direction of the dynamic clue potential function, determining a potential clue convergence point;

[0044] determining the start node and the end node of the event causal path according to the start and end positions of the gradient path;

[0045] taking the start node and the end node of the event causal path as the event nodes of the event correlation graph, and defining the directed correlation edges between the event nodes;

[0046] determining the time sequence of the correlation edges through the time label relationship between the event nodes, and constructing the event correlation graph with the time sequence relationship.

[0047] Optionally, the confidence assessment is performed according to the event correlation graph, the historical clue path and the newly added event clue in the current time window, and an event evolution path is obtained, specifically:

[0048] According to the attention weight values of all event clues in the historical clue path and the attention weight values of the newly added event clues, the cumulative values of the attention weights of all connected event clues are calculated to form an initial quantum probability amplitude distribution;

[0049] The amplitude values of all event nodes in the initial quantum probability amplitude distribution of each event node are mapped one by one to the quantum bits of the corresponding event nodes in the quantum probability reasoning space, and the amplitude values of the quantum bits are determined one by one to generate an initial quantum probability state.

[0050] According to the directed correlation edges between any two event nodes in the event correlation graph, a quantum operator matrix is constructed;

[0051] The constructed quantum operator matrix is applied to the initial quantum probability state in the quantum probability reasoning space to obtain a joint quantum probability amplitude value set;

[0052] The joint quantum probability amplitude value of each event causal path is subjected to an amplitude square operation to obtain a real-time probability value set;

[0053] The real-time probability values of all event causal paths in the real-time probability value set are compared and sorted one by one, and the event causal path with the largest real-time probability value is selected as the event evolution path in the current time window.

[0054] Optionally, the event clues corresponding to each event node in the event evolution path are mapped back to the key semantic feature vector, and the quantum attention weight matrix is adjusted according to the semantic feature similarity calculation result, and the continuous quantum feature field reconstruction and event causal path calculation are iteratively executed to obtain the final event clue evolution path, specifically:

[0055] Map each event node in the event evolution path obtained in the current time window back to the event thread corresponding to the event node in the structured data input respectively, to obtain the original key semantic feature vector of each event node corresponding to the event thread;

[0056] According to the key semantic feature vector of the event node, the cosine similarity of the corresponding key semantic feature vectors of any two event nodes connected by a directed association edge is calculated respectively;

[0057] The cosine similarity between each pair of associated event nodes is calculated by the numerical difference value of the attention weight value between the corresponding event nodes in the quantum attention weight matrix, to obtain a numerical difference value set;

[0058] The numerical difference value set is summed to obtain the numerical difference accumulation value of the current quantum attention weight matrix as a whole, and the numerical difference accumulation value is compared with the preset convergence threshold;

[0059] When the numerical difference accumulation value is greater than the preset convergence threshold, the attention weight value between the corresponding event nodes in the quantum attention weight matrix is proportionally scaled and adjusted to update the attention weight value of the quantum attention weight matrix;

[0060] Based on the updated quantum attention weight matrix, the continuous quantum feature field modeling, dynamic clue potential function, potential gradient path calculation, event association graph reconstruction and quantum probability reasoning of the event evolution path are re-executed to obtain the updated event evolution path.

[0061] The beneficial effects of the present application are:

[0062] (1) The present application realizes efficient and accurate association analysis between heterogeneous multi-modal event threads by fusing quantum feature field modeling and quantum probability reasoning method based on quantum adaptive attention mechanism, effectively improves the accuracy and real-time performance of event thread dynamic association, and enhances the reliability of event evolution path recognition and prediction in complex scenarios.

[0063] (2) The present application realizes effective identification and reasoning of the potential causal relationship between event threads by gradient path analysis and event association graph construction method of dynamic clue potential function, significantly improves the reasoning accuracy of event evolution path, and shows better adaptability and stability in multi-modal heterogeneous data scenarios.

[0064] (3) The application effectively solves the problems of low feature fusion efficiency and unstable attention weight calculation in the prior art by adaptively adjusting the quantum attention weight matrix through semantic feature similarity in real-time analysis of multi-modal event clues, breaks through the recognition bottleneck of nonlinear event correlation features in the prior art, significantly improves the event clue correlation accuracy, and effectively improves the practical application level of multi-modal event clue dynamic research and judgment in the field of artificial intelligence. BRIEF DESCRIPTION OF DRAWINGS

[0065] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:

[0066] Figure 1 A flowchart of the event clue dynamic research and judgment method based on a large model agent proposed by the application. DETAILED DESCRIPTION

[0067] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams that only schematically show the basic structure of the application and thus only show the components relevant to the application.

[0068] Reference Figure 1 The event clue dynamic research and judgment method based on a large model agent includes:

[0069] Obtaining multi-modal event clues including text, images, speech and sensor data dynamically generated by a large model agent, and respectively analyzing the original data content and corresponding time labels of each event clue to form a unified structured data input;

[0070] Respectively performing corresponding feature extraction on each type of event clue in the structured data input to obtain key semantic feature vectors of each event clue in the original feature space, and encoding the key semantic feature vectors into corresponding initial quantum state representations;

[0071] Using a pre-set parameterized quantum gate to perform multi-round iterative entanglement evolution and quantum state measurement operation on the initial quantum state representation, calculating the attention weight value between any two event clues, and obtaining a complete quantum attention weight matrix between event clues;

[0072] According to the attention weight of each event clue in the quantum attention weight matrix and the time label of the corresponding event clue, modeling the continuous quantum feature field of all event clues, and defining a derivable dynamic clue potential function according to the spatial distribution of the event clues in the quantum feature field;

[0073] The local extreme point of the dynamic clue potential function in the continuous quantum characteristic field and the corresponding potential energy gradient path are calculated, the potential clue convergence point of the event clue and the event causal path are determined, and the event correlation graph is constructed based on the potential clue convergence point and the event causal path;

[0074] According to the event correlation graph, the historical clue path and the newly added event clue in the current time window, the real-time confidence evaluation of the correlation degree and the causal relationship between events is performed through the quantum probability reasoning method, and the event evolution path with the highest causal correlation degree is obtained;

[0075] The event clues corresponding to each event node in the event evolution path are mapped and compared with the key semantic feature vectors in the original feature space, the semantic feature similarity between the key semantic feature vectors is calculated, and the quantum attention weight matrix is adaptively adjusted according to the semantic feature similarity. The continuous quantum characteristic field reconstruction and event causal path calculation are iteratively performed until the quantum attention weight matrix and the semantic feature similarity in the original feature space simultaneously satisfy the preset convergence condition, and the final event clue evolution path is output.

[0076] In the embodiment, the multi-modal event clues generated by the large model agent in real time are obtained, including text, image, voice and sensor data, and the original data content and corresponding time label of each event clue are analyzed respectively to form a unified structured data input, specifically:

[0077] Collecting multi-modal event clue data generated by the large model agent in real time, the multi-modal event clue data includes text data, image data, voice data and sensor data;

[0078] Performing semantic analysis operation on the collected text data, extracting semantic entities, keywords and semantic structures reflecting event attributes in the text data, and recording the time stamp of the text data generation;

[0079] Performing feature segmentation and target recognition operation on the collected image data, determining the object category, appearance feature and position coordinate associated with the event in the image data, and recording the time stamp of the image data generation;

[0080] Performing speech feature extraction and speech recognition operation on the collected voice data, determining the text transcription content, speech emotion feature and speaker identity feature corresponding to the voice data, and recording the time stamp of the voice data generation;

[0081] Performing signal preprocessing and feature extraction operation on the collected sensor data, determining the corresponding physical measurement parameter and state feature in the sensor data, and recording the time stamp of the sensor data generation;

[0082] The parsed text data, image data, voice data and sensor data are structured and coded with respective time stamps to form a unified structured data input.

[0083] In the embodiment, the key semantic feature vector is encoded into the corresponding initial quantum state representation, specifically:

[0084] The key semantic feature vector of each event clue is subjected to multi-dimensional feature normalization processing to constrain the feature component values of each dimension in the key semantic feature vector to between 0 and 1, thereby generating a normalized key semantic feature vector.

[0085] The number of quantum bits required in the corresponding initial quantum state representation is determined according to the number of feature dimensions contained in the normalized key semantic feature vector, and a preset amplitude value range is set for each quantum bit, and the amplitude value range is divided into a plurality of non-overlapping subintervals, so that each subinterval uniquely corresponds to a feature dimension.

[0086] According to the preset correspondence between the feature dimensions and the subintervals, the values of the feature dimensions in the normalized key semantic feature vector are mapped one by one into the subintervals of the corresponding quantum bit amplitudes, thereby forming an initial quantum bit amplitude distribution.

[0087] A quantum state amplitude normalization method is used to perform a secondary normalization operation on the initial quantum bit amplitude distribution, so that the sum of squares of all quantum bit amplitude components is strictly equal to 1, thereby obtaining a quantum state amplitude distribution that satisfies the definition condition of the quantum state. The quantum state amplitude normalization method is to accumulate the square values of the amplitude components in the initial quantum bit amplitude distribution to obtain the modulus square sum of the amplitude distribution, and use the square root of the modulus square sum of the amplitude distribution as a normalization factor to divide each amplitude component in the initial quantum bit amplitude distribution by the normalization factor, thereby obtaining a quantum state amplitude distribution that satisfies the definition condition of the quantum state.

[0088] Based on the quantum state amplitude distribution, an initial phase parameter is set for each quantum bit in the initial quantum state representation, the value range of the initial phase parameter is between 0 and 2π, and the setting of the initial phase parameter and the amplitude value of the corresponding feature dimension have a one-to-one mapping relationship.

[0089] According to the preset quantum bit arrangement order, the quantum state amplitude distribution and the initial phase parameter are loaded one by one into the initial quantum state representation, so that each quantum bit in the initial quantum state representation has both amplitude and phase values, thereby obtaining an initial quantum state representation corresponding to the key semantic feature vector.

[0090] In the embodiment, the initial quantum state representation is iteratively evolved and quantum state measurement is operated by using the pre-set parameterized quantum gate for multiple rounds, the attention weight values between any two event clues are calculated, and a complete quantum attention weight matrix between event clues is obtained, specifically as follows:

[0091] Before performing the multiple rounds of iterative entanglement evolution, the amplitude values of all quantum bits in the initial quantum state of each event clue are sorted in descending order, and the phase values of all quantum bits in the initial quantum state of each event clue are sorted in ascending order according to the amplitude values and the phase values of each quantum bit in the initial quantum state representation;

[0092] All quantum bits in the initial quantum state representation of each event clue are rearranged one by one in an alternating order of amplitude and phase sorting to form an adaptive quantum bit reordering sequence which is different from the initial quantum state representation and has a determined sequence;

[0093] The adaptive quantum bit reordering sequence is input one by one into the pre-set parameterized quantum gate for initial entanglement evolution, so that the quantum state after the initial entanglement evolution reflects the amplitude space distribution characteristics of the current input sequence;

[0094] After the initial entanglement evolution, a phased probability measurement operation is performed according to the quantum state after the initial entanglement evolution, the probability distribution of the current quantum state is determined through the phased probability measurement result, and the probability value of each quantum bit after measurement is recorded at the same time;

[0095] According to the phase difference value between the phased probability measurement result and each quantum bit of the current quantum state, a secondary sorting sequence of quantum bits is determined, and secondary quantum bit reordering is strictly performed according to the secondary sorting sequence to generate a second quantum state sequence for the next round of entanglement evolution;

[0096] The quantum state sequence input, real-time adaptive determination of the rotation parameter of the rotation gate, phased probability measurement and quantum bit adaptive reordering operation are repeatedly performed to complete multiple rounds of entanglement evolution;

[0097] When the entire multiple rounds of entanglement evolution process is completed, a global quantum state cooperative measurement is performed based on the final entangled state by implementing a joint projection measurement on the global basis of the quantum state, and a joint probability distribution matrix of all quantum bits is obtained, each matrix element of the joint probability distribution matrix representing the cooperative probability value of each quantum bit relative to all other quantum bits;

[0098] The quantum state correlation strength values between the initial quantum states of any two event clues are calculated according to the joint probability distribution matrix, and all the calculated quantum state correlation strength values are sequentially filled into the quantum attention weight matrix according to the order of the event clues, so as to obtain the quantum attention weight matrix in which each matrix element corresponds to the quantum state correlation strength between the event clues.

[0099] In the embodiment, the pre-configured parameterized quantum gate is specifically:

[0100] The quantum gate sequence is formed by alternately connecting rotation gates and controlled non-gate in sequence, wherein the initial values of the rotation axis direction and the rotation angle parameter of the rotation gate are determined according to the centroid of the current amplitude distribution of all quantum bits in the input quantum bit reordering sequence and the standard deviation of the amplitude value; the pairing order of the control quantum bit and the target quantum bit of the controlled non-gate is determined according to the amplitude difference between adjacent quantum bits in the input quantum bit reordering sequence in descending order;

[0101] In each subsequent round of entanglement evolution, the rotation axis direction and the rotation angle parameter of the rotation gate are updated in real time according to the stage probability measurement value obtained in the last round of evolution, and the pairing order of the control quantum bit and the target quantum bit of the controlled non-gate is updated in real time according to the quantum bit amplitude change determined by the last round of stage probability measurement value.

[0102] In the embodiment, the continuous quantum feature field modeling is performed on all event clues according to the attention weight of each event clue in the quantum attention weight matrix and the time label of the corresponding event clue, and a derivable dynamic clue potential function is defined according to the spatial distribution of the event clues in the quantum feature field, which is specifically:

[0103] According to the attention weight values between each event clue and all other event clues in the quantum attention weight matrix, the overall attention weight sum of each event clue relative to all event clues is determined, and the overall attention weight sum is taken as the initial spatial density value of each event clue in the continuous quantum feature field;

[0104] Taking the amplitude values and phase values of all feature dimensions in the key semantic feature vector corresponding to each event clue as the multi-dimensional space coordinate axes of the continuous quantum feature field, the spatial position coordinates of each event clue in the continuous quantum feature field space coordinate system are determined according to the initial spatial density value of the event clue;

[0105] According to the time label value of the event clue, the spatial position coordinates of each event clue are mapped and associated with the corresponding time label in the spatial coordinate system of the continuous quantum feature field, so as to construct an event clue joint coordinate point with space-time positioning in the continuous spatial coordinate system;

[0106] According to the spatial position coordinates and time label values between the joint coordinate points of adjacent event clues, high-order interpolation operations of spatial coordinates and time coordinates are respectively performed by using a cubic spline interpolation method, so as to realize continuous and smooth mapping of spatial coordinates and time coordinates between any adjacent event clues, and obtain complete spatiotemporal continuous distribution of all event clues in the continuous quantum characteristic field;

[0107] Based on the complete spatiotemporal continuous distribution, a local spatial density function is defined at any position in the continuous quantum characteristic field space, and a value of the local spatial density function is determined by calculating a Gaussian weighted cumulative value of all event clue spatial density initial values within a preset radius range around the corresponding position, and represents a local spatial density value of the event clues at each position in the quantum characteristic field space.

[0108] A dynamic clue potential function is constructed by using the local spatial density function, a function value of the dynamic clue potential function at any spatial coordinate position in the continuous quantum characteristic field is a negative logarithm of a local spatial density function value of the corresponding spatial position, and the dynamic clue potential function satisfies continuous derivable characteristics at any position in the quantum characteristic field space.

[0109] In the embodiment, the local extreme points and the corresponding potential energy gradient paths of the dynamic clue potential function in the continuous quantum characteristic field are calculated, the potential clue convergence points and the event causal paths of the event clues are determined, and an event correlation graph is constructed based on the potential clue convergence points and the event causal paths, specifically:

[0110] Within the spatial coordinate range of the continuous quantum characteristic field, a local extreme point solving is performed on a spatial neighborhood of all event clue joint coordinate points according to the dynamic clue potential function by using a multidimensional Newton iteration method, and a spatial coordinate position of the local extreme point is obtained.

[0111] Based on a derivative calculation method of the dynamic clue potential function, each event clue joint coordinate point is taken as a starting position, a gradient path tracking is performed along a negative gradient direction of the dynamic clue potential function in the quantum characteristic field space, and a local extreme point at which the gradient path terminates at a spatial coordinate position is taken as a potential clue convergence point of the event clue.

[0112] For the gradient paths calculated for all event clue joint coordinate points, a spatial coordinate distance from the starting position to the termination position is judged, the gradient paths with a spatial coordinate distance less than a set threshold are reserved, and the event clue joint coordinate points corresponding to the starting and termination positions of the paths are respectively determined as a starting node and a termination node of the event causal path.

[0113] The event causal path starting node and the event causal path ending node are taken as event nodes in the event association graph, and a directed association edge between the event nodes is defined according to a gradient path direction relationship between the event causal path starting node and the event causal path ending node;

[0114] The event nodes and the association edges are gradually increased in the event association graph until the mapping of all event causal paths satisfying the spatial coordinate distance threshold condition in the event association graph is completed;

[0115] The time sequence order of the association edges between the event nodes is determined through a time label numerical relationship between the event clues joint coordinate points corresponding to the event causal path starting node and the event causal path ending node, and the time sequence order is marked on the association edges to construct the event association graph with the time sequence relationship.

[0116] In the embodiment, the real-time confidence evaluation of the correlation degree and the causal relationship between events is performed through a quantum probability reasoning method according to the event association graph, the historical clue path and the newly added event clue in the current time window, and an event evolution path with the highest causal correlation degree is obtained, specifically as follows:

[0117] The accumulated numerical value of the attention weight of all connected event clues is calculated for each event node according to the attention weight numerical value of all event clues in the historical clue path connected to each event node in the event association graph and the attention weight numerical value of the newly added event clue in the current time window in the quantum attention weight matrix, and the accumulated numerical value is determined as the initial quantum probability amplitude value of each event node to form the initial quantum probability amplitude distribution of the event node.

[0118] The amplitude numerical value of all event nodes in the initial quantum probability amplitude distribution of each event node is mapped to the quantum bit of the corresponding event node in the quantum probability reasoning space one by one, and the quantum bit amplitude value is determined one by one, so as to generate an initial quantum probability state capable of completely expressing the initial state of the event node in the quantum probability reasoning space.

[0119] According to the directed association edge between any two event nodes in the event association graph, the attention weight numerical value corresponding to each directed association edge is determined as the initial quantum association strength value of the association edge, and a quantum operator matrix for expressing the causal correlation relationship between the event nodes is constructed according to all initial quantum association strength values, each element of the quantum operator matrix is one-to-one mapped with the attention weight numerical value corresponding to the directed association edge between the event nodes in the event association graph.

[0120] The constructed quantum operator matrix is respectively applied to the initial quantum probability state in the quantum probability reasoning space, and the joint quantum probability amplitude distribution of each event causal path is calculated through the product operation of the matrix and the quantum state, so as to obtain the joint quantum probability amplitude value set of all event causal paths in the event correlation graph;

[0121] The joint quantum probability amplitude value of each event causal path is subjected to amplitude square operation, and a real-time probability value set is obtained, and the real-time probability value of each event causal path is the calculation result after the joint quantum probability amplitude value of the path is squared;

[0122] The real-time probability values of all event causal paths in the real-time probability value set are compared and sorted one by one, and the event causal path with the largest real-time probability value is selected as the event evolution path in the current time window, and the event evolution path is represented in the form of event nodes and directed correlation edges between connected event nodes;

[0123] After the event evolution path in the current time window is determined, the determined event evolution path is taken as the basis of the historical clue path in the next time window, and the initial quantum probability amplitude distribution is updated according to the new event clue in the next time window, and the quantum probability reasoning step is executed again to continuously obtain the event evolution path with the largest real-time probability value in the next time window.

[0124] In the embodiment, the event clues corresponding to each event node in the event evolution path are mapped and compared with the key semantic feature vectors in the original feature space, the semantic feature similarity between the key semantic feature vectors is calculated, and the quantum attention weight matrix is adaptively adjusted according to the semantic feature similarity, and the continuous quantum feature field reconstruction and event causal path calculation are iteratively executed until the quantum attention weight matrix and the semantic feature similarity in the original feature space satisfy the preset convergence condition at the same time, and the final event clue evolution path is output, specifically:

[0125] Each event node in the event evolution path obtained in the current time window is respectively mapped back to the event clue corresponding to the event node in the structured data input, and the original key semantic feature vector of each event node corresponding to the event clue is obtained;

[0126] According to the key semantic feature vector of the event node, the cosine similarity of the corresponding key semantic feature vectors of any two event nodes connected by a directed correlation edge is calculated, and the semantic feature similarity value between each pair of correlated event nodes is obtained;

[0127] performing a square operation on the obtained difference value, to obtain a numerical difference value set between the attention weight matrix and the semantic feature similarity;

[0128] performing a summation calculation on the numerical difference values in the numerical difference value set to obtain a numerical difference accumulation value of the current quantum attention weight matrix as a whole, and comparing the numerical difference accumulation value with a preset convergence threshold to determine whether a convergence condition is reached;

[0129] When the numerical difference accumulation value is greater than the preset convergence threshold, taking the semantic feature similarity value between each pair of associated event nodes as a reference, the attention weight values between the corresponding event nodes of the quantum attention weight matrix are proportionally scaled and adjusted to update the attention weight values of the quantum attention weight matrix;

[0130] Based on the updated quantum attention weight matrix, the steps of continuous quantum feature field modeling, dynamic clue potential energy function, potential energy gradient path calculation, event association graph reconstruction, and quantum probability inference of event evolution path are re-executed to obtain an updated event evolution path;

[0131] The steps of semantic feature similarity calculation, quantum attention weight matrix adjustment, and continuous quantum feature field reconstruction corresponding to the event evolution path are repeatedly executed until the numerical difference accumulation value of the quantum attention weight matrix is less than or equal to the convergence threshold, and the event clue evolution path that finally satisfies the convergence condition is output.

[0132] Embodiment 1:

[0133] In order to verify the feasibility of the application in implementation, the application is applied to the real-time dynamic research and judgment task of a sudden event of an intelligent emergency command system. This task needs to analyze multi-modal heterogeneous clues such as text, image, voice and sensor data in real time to quickly identify the event evolution path and assist decision-making. In this scenario, the existing technology generally uses a classical attention mechanism combined with a traditional graph neural network method to analyze event clues. Due to the lack of effective nonlinear association expression and real-time dynamic iteration mechanism, the multi-modal data fusion efficiency is low, and the real-time analysis accuracy is difficult to meet the actual demand. Especially when the event clue evolves rapidly, the research and judgment result error is significant, which affects the timeliness of decision-making.

[0134] In the implementation process, first, the large model agent generates event clue data in real time, covering text description, on-site image, voice call record, and various types of sensor real-time collection information, and parses it into a unified structured data format, recording the corresponding time label. Subsequently, feature extraction operations are performed on event clue data of each modality, wherein event keywords and entities are extracted from text data, appearance features and position coordinates of targets are extracted from image data, speaker emotion features and identity information are extracted from voice data, and physical measurement parameters and state features are extracted from sensor data. After completing feature extraction, the system converts the obtained key semantic feature vector into an initial quantum state representation through a quantum state encoding method. In specific implementation, the key semantic feature vector is normalized to the 0 to 1 interval, and the number of quantum bits is determined according to the number of feature dimensions, and then the feature vector dimension values are mapped to the corresponding quantum bit amplitudes, and after amplitude normalization, initial phase information is loaded, and finally an initial quantum state representation satisfying the physical constraint condition is obtained.

[0135] In the quantum state processing stage, the system takes the adaptively ordered quantum bit reorganization sequence as input through the preset parameterized quantum gate sequence, and obtains the quantum attention weight matrix between the clues through multiple rounds of iterative entanglement evolution and quantum measurement. In the specific experiment process, after the quantum bit amplitude value and phase value sequence are adaptively arranged, the parameterized quantum gate is constructed by alternately connecting the rotation gate and the controlled non gate, and the rotation gate parameters are updated in real time based on the current amplitude centroid and standard deviation in each round of evolution. After 5 rounds of entanglement evolution, a stable quantum attention weight matrix is obtained.

[0136] Subsequently, the system models the continuous quantum feature field of the event clues according to the quantum attention weight matrix, realizes the continuous smooth distribution of the spatial and temporal dimensions by using the cubic spline interpolation method, defines the dynamic clue potential function based on this, and calculates the potential gradient path to determine the potential clue convergence point and the event causal path, and then constructs the event association graph reflecting the event time sequence relationship.

[0137] After completing the construction of the event association graph, the system constructs the quantum operator matrix to perform probability reasoning by taking the attention weight of the historical clue path corresponding to the event node and the current new clue as the initial value of the quantum probability amplitude, and obtains the real-time probability value of the event causal path. In actual application, after 20 rounds of dynamic time window iteration, the system always completes each reasoning within 0.5 seconds, which is much lower than the average 3-second calculation delay required by traditional methods, effectively meeting the efficiency requirements of real-time dynamic research and judgment of emergencies.

[0138] In addition, to further improve the research and judgment accuracy, the system calculates the semantic feature similarity based on the original key semantic feature vectors of each event node, and adaptively adjusts the quantum attention weight matrix, continuously iterates the continuous quantum feature field reconstruction and event evolution path reasoning process until the difference between the attention weight matrix and the semantic similarity is less than the set threshold 0.001, and finally outputs a highly reliable event clue evolution path. In actual testing, usually after 8 to 12 iterations, it can converge.

[0139] In order to reflect the performance advantages of the present application, 5 real emergency scene event samples are selected for research and judgment result testing, and 5 key research and judgment indicators are selected for each sample for actual measurement and prediction comparison. The detailed data results are shown in Table 1:

[0140] Table 1 Comparison of actual measurement and prediction of event clue dynamic research performance indicators

[0141]

[0142] As can be seen from Table 1, the method proposed in the present application performs outstandingly in core indicators such as event evolution node number, average path length, research and judgment accuracy, attention weight error and average calculation delay, especially in real-time calculation delay, the time consumption of each reasoning is significantly better than that of existing classical methods, and the actual measured delay is always kept below 0.5 seconds. Specifically, taking the S-003 sample as an example, the predicted node number is consistent with the actual measured value, the path length error is only 0.1 node, the research and judgment accuracy is 94.5%, the attention weight error is controlled at about 0.01, and the average calculation delay is only 0.44 seconds. Compared with the traditional graph neural network and classical attention mechanism combined method, the accuracy is improved by about 8%, and the delay is reduced by nearly 85%, significantly improving the real-time and reliability of event dynamic research.

[0143] In summary of the above implementation effects, the event clue dynamic research method based on large model agent proposed in the present application can efficiently and accurately process heterogeneous multi-modal event clues, and quickly and stably identify the dynamic evolution path of the event, which has strong application value and promotion potential in emergency decision-making and risk control scenarios.

[0144] The above is only the preferred specific implementation method of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical solution and inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for dynamic analysis of event clues based on large-scale intelligent agents, characterized in that, include: Acquire multimodal event cues dynamically generated by large model agents and construct structured data inputs; For each type of event clue in the structured data input, corresponding feature extraction is performed to obtain key semantic feature vectors, which are then encoded into the corresponding initial quantum state representation; The initial quantum state representation is processed using pre-set parameterized quantum gates to obtain the quantum attention weight matrix; Continuous quantum characteristic field modeling is performed for all event clues, and a differentiable dynamic clue potential energy function is defined; Calculate the local extrema of the dynamic cue potential energy function in the continuous quantum characteristic field and the corresponding potential energy gradient path, determine the potential cue convergence point and event causal path of event cues, and construct the event correlation graph; Based on the event association diagram, historical clue paths, and newly added event clues within the current time window, confidence assessment is performed to obtain the event evolution path; The event clues corresponding to each event node in the event evolution path are mapped back to the key semantic feature vectors. The quantum attention weight matrix is ​​adjusted according to the semantic feature similarity calculation results. The continuous quantum feature field reconstruction and event causal path calculation are iteratively executed to obtain the final event clue evolution path.

2. The method for dynamic analysis of event clues based on large-scale intelligent agents according to claim 1, characterized in that, The process of encoding the key semantic feature vector into the corresponding initial quantum state representation specifically involves: Perform multi-dimensional feature normalization on the key semantic feature vectors of each event clue to generate normalized key semantic feature vectors; The number of qubits required in the corresponding initial quantum state representation is determined based on the number of feature dimensions contained in the normalized key semantic feature vector, and an amplitude range is preset for each qubit, and the amplitude range is divided into multiple non-overlapping sub-intervals. Based on the pre-defined correspondence between the feature dimensions and the sub-intervals, an initial quantum bit amplitude distribution is formed; A second normalization operation is performed on the initial qubit amplitude distribution to obtain the quantum state amplitude distribution; Based on the quantum state amplitude distribution, an initial phase parameter is set for each qubit in the initial quantum state representation; According to the preset quantum bit arrangement order, the quantum state amplitude distribution and initial phase parameters are loaded one by one into the initial quantum state representation to obtain the initial quantum state representation corresponding to the key semantic feature vector.

3. The method for dynamic analysis of event clues based on large-scale intelligent agents according to claim 1, characterized in that, The initial quantum state representation is processed using pre-set parameterized quantum gates to obtain the quantum attention weight matrix, specifically as follows: Based on the amplitude and phase values ​​of each qubit in the initial quantum state representation, determine the sorting sequence of amplitude values ​​from high to low and the sorting sequence of phase values ​​from small to large for all qubits in the initial quantum state of each event clue; All qubits in the initial quantum state representation of each event clue are rearranged one by one in an alternating order of amplitude and phase to form an adaptive qubit recombination sequence; The adaptive qubit recombination sequence is input one by one into a parameterized quantum gate to perform the initial entanglement evolution; Perform a phased probability measurement operation based on the quantum state after the initial entanglement evolution, determine the probability distribution of the current quantum state through the phased probability measurement results, and record the probability value of each quantum bit after measurement at the same time. Based on the phase difference between the phase measurement results and the phase difference between each quantum bit in the current quantum state, the secondary sorting sequence of the quantum bits is determined, and secondary quantum bit recombination is performed. Repeatedly execute quantum state sequence input, real-time adaptive determination of rotation gate rotation parameters, staged probability measurement and adaptive recombination of qubits to complete multiple rounds of entanglement evolution; After completing all rounds of entanglement evolution, the joint probability distribution matrix of all qubits is obtained by performing joint projection measurements on the global basis of quantum states. The quantum state correlation strength between the initial quantum states of any two event cues is calculated based on the joint probability distribution matrix, and then filled into the quantum attention weight matrix in the order of the event cues to obtain the complete quantum attention weight matrix.

4. The method for dynamic analysis of event clues based on large-scale intelligent agents according to claim 3, characterized in that, The pre-set parameterized quantum gate is specifically as follows: A quantum gate sequence consisting of alternating series of rotating gates and controlled NOT gates, wherein the initial values ​​of the rotation axis direction and rotation angle parameters of the rotating gate are determined based on the centroid and standard deviation of the amplitude distribution of all qubits in the input qubit recombination sequence; The pairing order of the control qubits and the target qubits of the controlled NOT gate is determined by sorting the amplitude differences between adjacent qubits in the input qubit recombination sequence from largest to smallest. In each round of entanglement evolution, the rotation axis direction and rotation angle parameters of the rotating gate are updated in real time based on the stage probability measurement values ​​obtained from the previous round of evolution. The pairing order of the control qubit and the target qubit of the controlled NOT gate is adjusted and updated based on the qubit amplitude change determined by the stage probability measurement values ​​of the previous round.

5. The method for dynamic analysis of event clues based on large-scale intelligent agents according to claim 1, characterized in that, The continuous quantum characteristic field modeling of all event clues and the definition of a differentiable dynamic clue potential energy function are as follows: The sum of the overall attention weights for each event cue is determined based on the quantum attention weight matrix, and used as the initial value for the spatial density; Using the amplitude and phase values ​​of key semantic feature vectors as the spatial coordinate axes of the continuous quantum feature field, the spatial location coordinates of event clues are determined based on the initial value of spatial density. Construct joint coordinate points for event clues based on their spatial location coordinates and time tags; A cubic spline interpolation method is used to perform continuous smooth mapping of the joint coordinate points of the event clues, thereby obtaining the complete spatiotemporal continuous distribution of the event clues; Determine the local spatial density function based on the complete spatiotemporal continuous distribution; A dynamic cue potential energy function is constructed based on the local spatial density function. The function value of the dynamic cue potential energy function at any spatial location is the negative logarithm of the local spatial density function value at the corresponding location, and it satisfies the continuous differentiability property.

6. The method for dynamic analysis of event clues based on large-scale intelligent agents according to claim 1, characterized in that, The calculation of the local extrema of the dynamic clue potential energy function in the continuous quantum characteristic field and the corresponding potential energy gradient path determines the potential clue convergence point and event causal path of the event clues, and constructs an event correlation graph, specifically as follows: Based on the dynamic clue potential energy function, the spatial coordinates of the local extreme points are solved using the multidimensional Newton iteration method; Tracing the gradient path along the negative gradient direction of the dynamic cue potential energy function to determine potential cue convergence points; The starting and ending nodes of the event causal path are determined based on the starting and ending positions of the gradient path. The starting and ending nodes of the event causal path are used as event nodes in the event association graph, and directed association edges are defined between event nodes. The temporal sequence of associated edges is determined by the time tag relationship between event nodes, and an event association graph with temporal relationship is constructed.

7. The method for dynamic analysis of event clues based on large-scale intelligent agents according to claim 1, characterized in that, The process of assessing confidence based on the event association graph, historical clue paths, and newly added event clues within the current time window to obtain the event evolution path is as follows: Based on the attention weight values ​​of all event clues in the historical clue path and the attention weight values ​​of newly added event clues, the cumulative value of the attention weights of all connected event clues is calculated to form the initial quantum probability amplitude distribution. The amplitude values ​​of all event nodes in the initial quantum probability amplitude distribution of each event node are mapped one by one to the qubits of the corresponding event nodes in the quantum probability inference space, and the amplitude value of each qubit is determined to generate the initial quantum probability state. Construct a quantum operator matrix based on the directed association edges between any two event nodes in the event association graph; The constructed quantum operator matrix is ​​applied to the initial quantum probability state in the quantum probability inference space to obtain a set of joint quantum probability amplitude values. Perform an amplitude squaring operation on the joint quantum probability amplitude value of each event causal path to obtain a set of real-time probability values; The real-time probability values ​​of all event causal paths in the real-time probability value set are compared and sorted one by one, and the event causal path with the largest real-time probability value is selected as the event evolution path within the current time window.

8. The method for dynamic analysis of event clues based on large-scale intelligent agents according to claim 1, characterized in that, The process involves mapping event clues corresponding to each event node in the event evolution path back to key semantic feature vectors, adjusting the quantum attention weight matrix based on the semantic feature similarity calculation results, iteratively performing continuous quantum feature field reconstruction and event causal path calculation to obtain the final event clue evolution path, specifically as follows: Map each event node in the event evolution path obtained in the current time window back to the event clue corresponding to the event node in the structured data input, and obtain the original key semantic feature vector of the event clue corresponding to each event node; Based on the key semantic feature vectors of the event nodes, calculate the cosine similarity of the corresponding key semantic feature vectors for any two event nodes that are connected by a directed edge. The numerical difference between the cosine similarity between each pair of related event nodes and the attention weight values ​​between the corresponding event nodes in the quantum attention weight matrix is ​​calculated to obtain a set of numerical differences. The numerical differences in the set of numerical differences are summed to obtain the total numerical difference of the current quantum attention weight matrix, and the total numerical difference is compared with a preset convergence threshold. When the cumulative value of the numerical difference exceeds the preset convergence threshold, the attention weight values ​​of the quantum attention weight matrix are updated by scaling the attention weight values ​​between the corresponding event nodes in the quantum attention weight matrix. Based on the updated quantum attention weight matrix, the continuous quantum feature field modeling, dynamic cue potential function, potential gradient path calculation, event association graph reconstruction, and quantum probabilistic inference of event evolution path are re-executed to obtain the updated event evolution path.

Images