Shooting action mode recognition and prediction method and system

Abstract

The invention discloses a shooting action mode recognition and prediction method and system, and belongs to the technical field of mode recognition and prediction. The method specifically comprises the steps that a shooting action time sequence signal set and working condition information are collected, time alignment and amplitude normalization are carried out, and an aligned shooting action time sequence signal set and a working condition label set are obtained; based on the aligned shooting action time sequence signal set and the working condition label set, spatial and temporal distribution alignment is carried out through optimal transmission, a multi-layer multi-relation hypergraph is constructed, hyperbolic manifold embedding is carried out, and a hyperbolic hypergraph representation sequence and weight are obtained; constructing a conditional energy function based on the hyperbolic hypergraph representation sequence and the weight, performing optimization calculation through a micro-optimization layer to obtain a stage sequence, a mode label and a stage transfer parameter of the shooting action, constructing a multi-time-interval stage hazard rate matrix, and obtaining a mode recognition prediction result; the accuracy of shooting action mode recognition prediction in a new scene is improved, and the stability of shooting action mode recognition prediction is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of pattern recognition and prediction, and specifically relates to a shooting action pattern recognition and prediction method and system. Background Technology

[0002] In shooting training, assessment, and combat support, facing shooting maneuvers involving multiple stages, postures, personnel, and conditions, relying solely on instructor observation, video playback, and post-shooting statistics is insufficient to accurately and promptly identify the procedural characteristics of shooting actions, easily leading to delayed judgments and biased conclusions. However, acquiring multi-source time-series signals can record process information such as aiming trajectory, preload, and micro-jitter during shooting, providing data support for objective identification of shooting actions and result prediction. Therefore, in real firing ranges and mobile scenarios, how to accurately identify and predict shooting actions based on multi-source time-series signals is a crucial problem urgently needing to be solved in the field of shooting applications.

[0003] In existing technologies, shooting action pattern recognition and prediction methods are typically based on simple fusion of sensor signals, calculating statistical or deep features, and using classification and regression models with fixed time windows to determine the action stage and generate a firing sequence. These are then combined with empirical thresholds to generate pattern recognition and prediction results. However, existing shooting action pattern recognition and prediction methods suffer from the following problems: they ignore the data distribution shift of shooting actions under changing conditions, resulting in low accuracy in pattern recognition and prediction in new scenarios. This leads to poor stability and accuracy in pattern recognition and prediction of shooting actions in new scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a shooting action pattern recognition and prediction method and system to solve the problems of poor accuracy in the prior art.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] A shooting action pattern recognition and prediction method includes:

[0007] Collect the firing action timing signal set and working condition information, and perform time alignment and amplitude normalization based on the firing action timing signal set and working condition information to obtain the aligned firing action timing signal set and working condition label set;

[0008] Based on the aligned shooting action timing signal set and the working condition label set, spatiotemporal distribution alignment is performed through optimal transmission to construct a multi-layer multi-relationship hypergraph, and hyperbolic manifold embedding is performed to obtain the hyperbolic hypergraph representation sequence and weights;

[0009] Based on the hyperbolic hypergraph representation sequence and weights, a conditional energy function is constructed and optimized through a differentiable optimization layer to obtain the phase sequence, pattern label and phase transition parameters of the shooting action;

[0010] Based on stage sequences, pattern labels, and stage transition parameters, a multi-temporal stage hazard rate matrix is ​​constructed to obtain pattern recognition prediction results.

[0011] Specifically, based on the aligned shooting action timing signal set and the operational condition label set, spatiotemporal distribution alignment is performed through optimal transmission to construct a multi-layer, multi-relationship hypergraph, and hyperbolic manifold embedding is performed to obtain the hyperbolic hypergraph representation sequence and weights, including:

[0012] Based on the aligned firing action timing signal set and the working condition label set, a cost kernel is constructed and optimal transmission is performed to obtain the working condition aligned timing segment sequence and the same working condition baseline parameter set;

[0013] Based on the time-series segment sequence aligned with the working condition and the baseline parameter set of the same working condition, time period nodes, channel nodes and working condition nodes are generated through node mapping, and ternary hyperedges connecting the time period nodes, channel nodes and working condition nodes are generated to construct a multi-layer multi-relationship hypergraph.

[0014] Based on a multi-layer, multi-relation hypergraph, pre-selected relations that meet the preset selection rules are selected, and the pre-selected relations are used as seeds for cross-layer generation to obtain an augmented hypergraph and relation gating coefficients;

[0015] Based on the augmented hypergraph and relation gating coefficients, the curvature parameters of the hyperbolic manifold are determined. The time period nodes, channel nodes and working condition nodes in the augmented hypergraph are embedded and mapped in the hyperbolic manifold to obtain the hyperbolic hypergraph representation sequence.

[0016] Based on the hyperbolic hypergraph representation sequence, hyperedge aggregation is performed to calculate channel weights and operating condition coupling weights.

[0017] Specifically, based on the aligned firing action timing signal set and the condition label set, a cost kernel is constructed and optimal transmission is performed to obtain the condition-aligned timing segment sequence and the baseline parameter set of the same condition, including:

[0018] Based on the working condition tag set and the preset time length, the time window segment of the alignment shooting action timing signal set is divided, and the amplitude statistics and frequency domain statistics of each time window are calculated to obtain the initial time segment set with working condition tags.

[0019] Based on the initial set of time series segments, the time series segments are clustered within each working condition, and the statistical center and amplitude scale are calculated to obtain a prototype set of working condition-level segments.

[0020] Based on the prototype set of layered segments under different operating conditions, a layered transmission distance matrix is ​​constructed, and a cost kernel is generated by combining the operating condition labels to obtain the cost value.

[0021] Based on cost value, optimal transmission is performed, and the initial time series segments are rearranged and weighted to obtain the working condition aligned time series segment sequence and the same working condition baseline parameter set.

[0022] Specifically, based on a multi-layer, multi-relation hypergraph, pre-selected relations that meet preset filtering rules are selected. These pre-selected relations are then used as seeds for cross-layer generation to obtain an augmenting hypergraph and relation gating coefficients, including:

[0023] Based on a multi-layer, multi-relation hypergraph, the structure score and causal tendency score of each ternary hyperedge are calculated to generate a relation candidate set. The causal prior degree of each candidate relation in the relation candidate set is calculated to determine the directed connections and obtain an updated relation candidate set.

[0024] Based on the updated candidate relation set, pre-selected relations that meet the preset screening rules in terms of structure score and causal precedence are selected as seed relations. The starting time node of each sub-relation is taken as the root node, and the time node is expanded level by level according to the directed connection to obtain the causal chain structure set.

[0025] Based on the causal chain structure set, new ternary hyperedges are generated by combining the corresponding channel nodes and working condition nodes. The new ternary hyperedges are then added to the multi-layer, multi-relationship hypergraph with the causal chain as the skeleton to obtain the preliminary augmented hypergraph.

[0026] The contribution of each ternary hyperedge in the preliminary augmented hypergraph is statistically analyzed, the contribution is normalized, and hyperedges with a contribution less than a preset contribution threshold are deleted to obtain the augmented hypergraph and relation gating coefficients.

[0027] Specifically, based on the augmented hypergraph and relation gating coefficients, the curvature parameters of the hyperbolic manifold are determined. Embedding mapping is then performed on the time-period nodes, channel nodes, and operational condition nodes in the augmented hypergraph within the hyperbolic manifold to obtain the hyperbolic hypergraph representation sequence, including:

[0028] Based on the augmented hypergraph, the time period node subgraphs for each working condition are determined, the graph distance between each time period node node in each time period node subgraph is calculated, a graph distance value set is generated, and for each graph distance value in the graph distance value set, the number of time period nodes whose graph distance is less than or equal to the corresponding graph distance value is counted, and a growth curve is generated.

[0029] The growth curve is compared with the preset hyperbolic volume growth model, a curvature evaluation function is constructed, and the curvature evaluation function is optimized to obtain a set of curvature parameters for the working condition.

[0030] Based on the set of curvature parameters for the working conditions, a weighted graph Laplacian operator is constructed and spectral decomposition is performed to generate the initial embedding vectors of each node. The initial embedding vectors are then mapped to the hyperbolic manifold space of the corresponding working condition through hyperbolic exponential mapping to obtain the initial hyperbolic node representation set.

[0031] Based on the initial hyperbolic node representation set and relational gating coefficients, the nodes in the augmented hypergraph are sequentially taken as center nodes, and the adjacent nodes of each center node are determined. The hyperbolic node representations of the adjacent nodes are transformed into the tangent space of the hyperbolic manifold with the center node as the base point through hyperbolic logarithmic mapping, resulting in a convergent hyperbolic node representation set.

[0032] Based on the convergent set of hyperbolic node representations, geodesic skeleton curves passing through time period nodes are fitted in the hyperbolic manifold. The hyperbolic node representations of time period nodes are then subjected to geodesic orthogonal projection and splicing of projected coordinates to obtain a hyperbolic hypergraph representation sequence.

[0033] Specifically, the growth curve is compared with a pre-defined hyperbolic volume growth model to construct a curvature evaluation function. This function is then optimized to obtain a set of curvature parameters for the operating conditions, including:

[0034] For each graph distance value in the graph distance value set, calculate the difference between the number of predicted nodes of the preset hyperbolic volume growth model at the corresponding graph distance value and the number of nodes of the growth curve at the corresponding graph distance value, and obtain a multi-scale difference sequence;

[0035] The differences of the multi-scale difference sequence at each graph distance value are weighted and accumulated to generate a global difference value. The changes in the difference of the multi-scale difference sequence at the distance values ​​of adjacent graphs are determined to generate a shape offset. The global difference value and the shape offset are combined according to the preset combination weights to construct a curvature evaluation function.

[0036] Based on the curvature evaluation function and the preset step size, a set of discrete curvature parameter values ​​is generated within the preset curvature parameter search interval. The curvature evaluation function value corresponding to each curvature parameter value in the set of discrete curvature parameter values ​​is calculated to obtain the set of discrete curvature parameter values.

[0037] Based on the set of discrete values ​​of the curvature parameters, each discrete value of the curvature parameter is used as a graph node, adjacent pairs of discrete values ​​of the curvature parameters are used as graph edges, and the corresponding curvature evaluation function value is used as the node energy to construct a curvature candidate graph structure.

[0038] Extract the descent direction of node energy in the curvature candidate graph structure, and iteratively update the curvature candidate graph structure based on the descent direction until the node energy distribution meets the preset iteration termination condition. Stop the iteration, select the discrete curvature parameter values ​​corresponding to the node with the lowest node energy, and obtain the set of working condition curvature parameters.

[0039] Specifically, based on the hyperbolic hypergraph representation sequence and weights, a conditional energy function is constructed and optimized through a differentiable optimization layer to obtain the phase sequence, pattern label, and phase transition parameters of the shooting action, including:

[0040] Based on the hyperbolic hypergraph representation sequence and weights, the time period nodes and adjacent hyperedges are aggregated to obtain the stage observation feature sequence and the stage transition feature sequence;

[0041] Based on the phase observation feature sequence, phase transition feature sequence, preset phase set and shooting action pattern set, time chain, phase chain and pattern node are generated, a three-part graph structure is constructed, and observation factor, time transition factor and pattern consistency factor are determined.

[0042] Based on the observation factor, time transition factor, model consistency factor and preset energy construction rules, the factor energy of each observation factor, time transition factor and model consistency factor is calculated, and the factor energy is weighted and summed according to the preset energy weight to construct the conditional energy function.

[0043] Based on the conditional energy function, block gradient updates are performed in the differentiable optimization layer, and constraint optimization is performed through simplex projection and near-end projection to obtain the current update stage indicator vector, the current update mode indicator vector, and the current update stage transition parameters.

[0044] The current update phase indicator vector and the current update mode indicator vector are decoded by maximum component to obtain the phase sequence, mode label and phase transition parameters of the shooting action.

[0045] Specifically, based on the conditional energy function, block gradient updates are performed in the differentiable optimization layer, and constrained optimization is performed through simplex projection and near-end projection to obtain the current update stage indicator vector, the current update mode indicator vector, and the current update stage transition parameters, including:

[0046] Based on the conditional energy function, hot-start initialization is performed in the differentiable optimization layer to generate initial values ​​for the stage indicator vector, the mode indicator vector, and the stage transition parameters. Random perturbations are introduced to break the symmetry and obtain the initial optimization state.

[0047] Based on the initial optimization state, the gradient of the conditional energy function with respect to the initial values ​​of the stage indicator vector, the initial values ​​of the mode indicator vector, and the initial values ​​of the stage transition parameters is calculated. The gradient is then updated in blocks alternately according to the opposite direction to obtain the unprojected intermediate optimization state, which includes the intermediate values ​​of the stage indicator vector, the intermediate values ​​of the mode indicator vector, and the intermediate values ​​of the stage transition parameters.

[0048] Based on the unprojected intermediate optimization state, simplex projection is performed on the intermediate values ​​of the stage indicator vector and the mode indicator vector, and near-end projection is performed on the intermediate values ​​of the stage transition parameters to satisfy preset order constraints and sparsity constraints, so as to obtain updated stage indicator vector, mode indicator vector and stage transition parameters.

[0049] Based on the updated stage indicator vector, mode indicator vector, and stage transition parameters, the change magnitude of the conditional energy function between two adjacent updates is calculated. When the change magnitude is less than a preset stopping threshold, the update is stopped, and the current update stage indicator vector, the current update mode indicator vector, and the current update stage transition parameters are obtained.

[0050] Specifically, based on stage sequences, pattern labels, and stage transition parameters, a multi-temporal stage hazard rate matrix is ​​constructed to obtain pattern recognition prediction results, including:

[0051] Based on the phase sequence and phase transition parameters, the dwell time and remaining time to firing in each phase of the phase sequence are statistically analyzed to obtain a set of phase time interval statistical features.

[0052] Based on the statistical feature set of time intervals in each stage, a multi-time interval hazard rate grid map is constructed. The stage condition hazard rate of each grid point is calculated by forward recursion to obtain the multi-time interval stage hazard rate matrix.

[0053] Based on the multi-time interval stage hazard rate matrix, the preset set of predicted time intervals and the survival analysis relationship, the survival probability and firing probability of each predicted time interval are calculated, and the firing probability of each predicted time interval is normalized and calibrated to obtain the multi-time interval firing probability distribution.

[0054] Based on the multi-time interval firing probability distribution and pattern labels, the firing probability of each predicted time interval is compared, and the predicted time interval with the highest firing probability is selected to obtain the pattern recognition prediction result.

[0055] A shooting action pattern recognition and prediction system, comprising:

[0056] The signal acquisition module is used to acquire the firing action timing signal set and working condition information. Based on the firing action timing signal set and working condition information, time alignment and amplitude normalization are performed to obtain the aligned firing action timing signal set and working condition label set.

[0057] The optimal transmission module is used to perform spatiotemporal distribution alignment based on the aligned shooting action timing signal set and the working condition label set, construct a multi-layer multi-relationship hypergraph, and perform hyperbolic manifold embedding to obtain the hyperbolic hypergraph representation sequence and weights;

[0058] The function construction module is used to construct the conditional energy function based on the hyperbolic hypergraph representation sequence and weights, and perform optimization calculations through a differentiable optimization layer to obtain the phase sequence, mode label and phase transition parameters of the shooting action;

[0059] The results output module is used to construct a multi-temporal stage hazard rate matrix based on stage sequence, pattern label and stage transition parameters, and obtain pattern recognition prediction results.

[0060] Compared with existing technologies, the beneficial effects of this invention include: collecting a set of timing signals of shooting actions and operational information, performing time alignment and amplitude normalization, and constructing aligned timing features of shooting actions; based on the aligned timing signal set of shooting actions and operational label set, constructing a multi-layer, multi-relationship hypergraph through optimal transmission, and embedding it in a hyperbolic manifold to construct hierarchical relationships and causal structures between time periods, sensing channels, and operational conditions; based on the hyperbolic hypergraph representation sequence and weights, constructing a conditional energy function, and performing constraint optimization in a differentiable optimization layer to obtain the stage sequence, pattern label, and stage transition parameters of shooting actions; based on the stage sequence, pattern label, and stage transition parameters, constructing a multi-time-distance stage hazard rate matrix to obtain a multi-time-distance firing probability distribution, and obtaining prediction results for firing time distance and action patterns; solving the problem in existing technologies where data distribution shifts due to changes in operational conditions, resulting in low accuracy of pattern recognition prediction for shooting actions in new scenarios, improving the accuracy of pattern recognition prediction results for shooting actions in new scenarios, enhancing the stability of pattern recognition prediction for shooting actions, and improving the accuracy of pattern recognition prediction for shooting actions in new scenarios. Attached Figure Description

[0061] Figure 1 A flowchart of a shooting action pattern recognition and prediction method provided by the present invention;

[0062] Figure 2 A schematic diagram of hyperbolic embedding and multi-temporal prediction of shooting action provided by the present invention;

[0063] Figure 3 A flowchart for generating pattern recognition prediction results provided by this invention;

[0064] Figure 4 The present invention provides a structural diagram of a shooting action pattern recognition and prediction system. Detailed Implementation

[0065] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0066] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0067] Example 1:

[0068] Please see Figures 1-3 The present invention provides an embodiment of a shooting action pattern recognition and prediction method, comprising the following specific steps:

[0069] Step S1: Collect the firing action timing signal set and working condition information. Based on the firing action timing signal set and working condition information, perform time alignment and amplitude normalization to obtain the aligned firing action timing signal set and working condition label set.

[0070] In this embodiment, the shooting action timing signal set includes triaxial acceleration signals, triaxial angular velocity signals, key point trajectory signals, grip force signals, electromyography signals, muzzle velocity signals, and trigger signals, etc. The working condition information includes site type, target distance, indoor or outdoor markings, gun type, ammunition type, and shooter posture category, etc. It should be noted that after collecting the shooting action timing signal set and working condition information, the data needs to be preprocessed. The preprocessing specifically includes: outlier handling, data standardization, missing value imputation, time alignment, and feature smoothing.

[0071] Based on the shooting action timing signal set and working condition information, a shooting action time axis is constructed with the sampling time of various signals as the horizontal axis. According to the trigger time in the trigger signal, the signal sequence of each shooting is shifted on the shooting action time axis. According to the preset time range before and after the trigger time, effective signal segments including the preparation, aiming, pre-pressing and firing stages are extracted to obtain a time-aligned shooting action timing signal set. The time range is set by those skilled in the art according to the actual situation.

[0072] The time-aligned firing action timing signal set is obtained by normalizing the amplitude of various signals according to their amplitude range, so that the signals under different firing times and different working conditions are at a uniform amplitude scale.

[0073] Based on the working condition information, the field type, target distance, indoor or outdoor markings, gun type, ammunition type and shooter posture category are encoded according to the preset encoding rules to generate working condition tags corresponding to each shooting session. The working condition tags of each shooting session are then summarized to obtain a working condition tag set. The encoding rules are set by those skilled in the art according to the actual situation.

[0074] like Figure 2 As shown, step S2: Based on the aligned shooting action timing signal set and the working condition label set, the spatiotemporal distribution is aligned through optimal transmission to construct a multi-layer multi-relationship hypergraph, and hyperbolic manifold embedding is performed to obtain the hyperbolic hypergraph representation sequence and weights.

[0075] The specific steps of step S2 are as follows:

[0076] Step S201: Based on the aligned shooting action timing signal set and the working condition label set, construct a cost kernel and perform optimal transmission to obtain the working condition aligned timing segment sequence and the same working condition baseline parameter set.

[0077] The specific steps of step S201 are as follows:

[0078] Step S2011: Based on the working condition tag set and the preset time length, the time window segment of the alignment shooting action timing signal set is divided, and the amplitude statistics and frequency domain statistics of each time window are calculated to obtain the initial time segment set with working condition tags.

[0079] In this embodiment, based on the alignment firing action timing signal set, with the sampling time as the horizontal axis, the time window length and window step size are set according to the preset time length. The alignment firing action timing signal of each firing is slidably divided along the time axis, and n1 consecutive sampling points are divided into a time window. According to the working condition label corresponding to each firing in the working condition label set, each time window and its corresponding working condition label are associated to obtain a time window sequence. The time length is set by those skilled in the art according to the actual situation.

[0080] Based on the time window sequence, the amplitude statistics and frequency domain statistics of the triaxial acceleration signal, triaxial angular velocity signal, key point trajectory signal, grip force signal, electromyography signal, and muzzle velocity signal within each time window are calculated. The amplitude statistics, frequency domain statistics, and corresponding working condition labels of each time window are combined to form a feature vector. The feature vectors of all time windows are summarized to obtain an initial time series segment set with working condition labels. The amplitude statistics include mean, variance, and energy, etc., and the frequency domain statistics include dominant frequency and frequency band energy distribution, etc.

[0081] Step S2012: Based on the initial set of time series segments, cluster the time series segments within each working condition, and calculate the statistical center and amplitude scale to obtain a prototype set of working condition-level segments.

[0082] In this embodiment, the initial time series segment set is divided based on the working condition label, and the initial time series segments belonging to the same working condition label are divided into the same working condition segment subset to obtain each working condition segment subset.

[0083] Within each subset of operating condition segments, clustering is performed based on the amplitude statistics and frequency domain statistics of each initial time segment using a preset clustering method. Initial time segments with a similarity not lower than a preset clustering threshold are divided into the same segment cluster, resulting in a set of segment clusters for each operating condition. The clustering method and clustering threshold are set by those skilled in the art according to the actual situation.

[0084] Based on the fragment cluster set for each operating condition, the mean values ​​of the amplitude statistics and frequency domain statistics of the initial time-series fragments within each fragment cluster are calculated to obtain the statistical center of the fragment cluster.

[0085] Based on the amplitude statistics of the initial time series segments within each segment cluster, the corresponding amplitude scale is calculated. The statistical center and amplitude scale of the segment clusters under each working condition are combined to obtain the working condition-layered segment prototype set.

[0086] Step S2013: Based on the prototype set of working condition layered segments, construct a layered transmission distance matrix, and combine it with the working condition label to generate a cost kernel to obtain the cost value.

[0087] In this embodiment, based on the set of segment prototypes under different working conditions, the segment prototypes under each working condition are numbered according to a preset hierarchical order. A weighted distance metric is constructed based on the difference in statistical center and the difference in amplitude scale of each segment prototype. The weighted distance between any two segment prototypes under each working condition is calculated. The weighted distance is then arranged according to the working condition number and the hierarchical number to obtain the hierarchical transmission distance matrix.

[0088] Based on the hierarchical transmission distance matrix and the working condition label set, it is determined whether the working conditions corresponding to the segment prototypes are the same, and working condition weights are assigned to each weighted distance according to the determination results.

[0089] Determine whether the prototype level numbers of the fragments are adjacent, and assign level weights to each weighted distance based on the determination result.

[0090] The weighted distance, working condition weight, and hierarchical weight are combined to obtain a cost kernel that represents the transmission cost between fragment prototype pairs. The values ​​of each element in the cost kernel are used as the cost value.

[0091] Step S2014: Based on cost value, perform optimal transmission, rearrange and weighted summarize the initial time series segments to obtain the working condition aligned time series segment sequence and the same working condition baseline parameter set.

[0092] In this embodiment, a cost matrix is ​​constructed based on the cost value. The initial time series segment set with working condition labels is used as the source distribution, and the working condition layered segment prototype set is used as the target distribution. Optimal transmission optimization is performed according to the preset quality conservation constraints to obtain the optimal transmission coefficient matrix between each initial time series segment and each segment prototype. The quality conservation constraints are set by those skilled in the art according to the actual situation.

[0093] Based on the optimal transmission coefficient matrix, the initial time segments in each working condition are weighted and sorted. According to the order of the corresponding segment prototypes, the initial time segments with the optimal transmission coefficient not less than the preset transmission threshold are arranged in descending order of transmission coefficient to obtain the working condition aligned time segment sequence. The transmission threshold is set by those skilled in the art according to the actual situation.

[0094] Based on the time-series segments aligned to the operating conditions, the mean and variance of the amplitude statistics and frequency domain statistics of the time-series segments corresponding to each operating condition are calculated by weighting, so as to obtain the baseline parameters of the same operating condition under each operating condition. The baseline parameters of the same operating condition under each operating condition are summarized to obtain the baseline parameter set of the same operating condition.

[0095] Step S202: Based on the time-series segment sequence aligned with the working condition and the baseline parameter set of the same working condition, generate time period nodes, channel nodes and working condition nodes through node mapping, and generate ternary hyperedges connecting the time period nodes, channel nodes and working condition nodes to construct a multi-layer multi-relationship hypergraph.

[0096] In this embodiment, based on the working condition aligned time sequence, the start and end times of each working condition aligned time sequence segment on the time axis and the amplitude statistics and frequency domain statistics corresponding to the working condition aligned time sequence segment are taken as a time period unit, and each time period unit is numbered to generate a time period node set.

[0097] Based on signals such as triaxial acceleration, triaxial angular velocity, key point trajectory, grip force, electromyography, and muzzle velocity in the time-series segments aligned with the working conditions, a set of channel nodes is generated.

[0098] Based on the operating condition information in the baseline parameter set of the same operating condition, different operating conditions are encoded to generate a set of operating condition nodes.

[0099] Based on the set of time period nodes, the set of channel nodes, and the set of working condition nodes, each time period node is associated with its corresponding channel node and working condition node, generating a ternary superedge connecting the time period nodes, channel nodes, and working condition nodes.

[0100] Based on the chronological order of nodes on the time axis, time relationship labels are added to the ternary hyperedges. Based on the similarity between working conditions in the same working condition baseline parameter set, working condition relationship labels are added to the ternary hyperedges. The ternary hyperedges and nodes with different relationship labels are combined to construct a multi-layer, multi-relationship hypergraph.

[0101] Step S203: Based on the multi-layer, multi-relation hypergraph, select the pre-selected relations that meet the preset selection rules, and use the pre-selected relations as seeds to generate across layers to obtain the augmented hypergraph and relation gating coefficients.

[0102] The specific steps of step S203 are as follows:

[0103] Step S2031: Based on the multi-layer, multi-relation hypergraph, calculate the structure score and causal tendency score of each ternary hyperedge, generate a candidate relation set, calculate the causal precedence of each candidate relation in the candidate relation set, determine the directed connections, and obtain the updated candidate relation set.

[0104] In this embodiment, based on a multi-layer, multi-relationship hypergraph, the number of ternary hyperedges sharing the same time period node, the number of ternary hyperedges sharing the same channel node, and the number of ternary hyperedges sharing the same working condition node are counted. The number of the three types of hyperedges is weighted and summed according to a preset sharing weighting rule to obtain the structural score of each ternary hyperedge. The sharing weighting rule is set by those skilled in the art according to the actual situation.

[0105] Obtain the chronological order and time interval of the time period nodes connected by each ternary hyperedge on the time axis. Based on the chronological order and time interval, construct a causal tendency index to reflect the strength of the tendency of the previous time period node to point to the next time period node, and calculate the causal tendency score of each ternary hyperedge according to the causal tendency index.

[0106] Based on the structure score and causal tendency score of each ternary hyperedge, and according to the preset structure score threshold and causal tendency score threshold, ternary hyperedges that simultaneously satisfy the structure score threshold and causal tendency score threshold are selected to obtain a candidate set of relations. The structure score threshold and causal tendency score threshold are set by those skilled in the art according to the actual situation.

[0107] Candidate relations that overlap at the same time period node are selected from the candidate relation candidate set. The causal precedence degree of each candidate relation is calculated based on the causal tendency score of the corresponding ternary hyperedge and the time sequence of the time period node.

[0108] Based on the causal precedence, determine the temporal order of any two candidate relations. Take the time period node corresponding to the earlier candidate relation as the starting point and the time period node corresponding to the later candidate relation as the ending point, establish a directed connection and mark it. Summarize all candidate relations with directed connection marks to obtain the updated relation candidate set.

[0109] Step S2032: Based on the updated candidate relation set, select pre-selected relations that meet the preset screening rules in terms of structure score and causal precedence as seed relations, take the starting time node of each sub-relation as the root node, and expand the time node level by level according to the directed connection to obtain the causal chain structure set.

[0110] In this embodiment, based on the updated candidate relation set, candidate relations with a structure score greater than or equal to the structure score threshold and a causal precedence degree greater than or equal to the causal precedence degree threshold are selected to obtain a pre-selected relation set.

[0111] Based on the pre-selected relation set, the time period node that serves as the starting point in each pre-selected relation is taken as the root node, and the time period node corresponding to the root node is taken as the starting node. Starting from the starting node, according to the directed connections and chronological order in the pre-selected relation, the direct successor time period node is searched sequentially. When a successor time period node is found, it is taken as the new starting node and the search continues. When a time period node no longer serves as the starting time period node of any pre-selected relation in the pre-selected relation set, the search along that path is terminated. The time period nodes that are sequentially passed from the root node to the current time period node are arranged in chronological order to form a causal chain. The entire causal chain sequence is summarized to obtain the causal chain structure set.

[0112] Step S2033: Based on the causal chain structure set, generate new ternary hyperedges by combining the corresponding channel nodes and working condition nodes, and add the new ternary hyperedges to the multi-layer multi-relationship hypergraph with the causal chain as the skeleton to obtain the preliminary augmented hypergraph.

[0113] In this embodiment, based on the causal chain structure set, the time period node sequence on each causal chain in the causal chain structure set is read sequentially according to the time order. According to the existing connection relationship of each time period node in the multi-layer multi-relationship hypergraph, the channel node and working condition node connected to each time period node are retrieved. The two adjacent time period nodes on the same causal chain and the channel node and working condition node that appear at the same time period node are combined to obtain n2 relation combination sets composed of adjacent time period node pairs, channel nodes and working condition nodes.

[0114] Based on the set of relationship combinations and the amplitude and frequency domain statistics of adjacent time period nodes on each channel node, relationship combinations with a similarity degree not lower than a preset similarity threshold are selected. For each relationship combination, according to the time order on the causal chain, the time period node that ranks earlier in the time order among the two adjacent time period nodes corresponding to the relationship combination is determined as the starting time period node. The starting time period node, the associated channel node, and the operating condition node are combined into a triplet to obtain a new triplet hyperedge. The similarity is set by those skilled in the art according to the actual situation.

[0115] Based on the order of arrangement on the causal chain, new ternary hyperedges are added to the multi-layered multi-relational hypergraph, forming a connection structure that grows step by step along the direction of the causal chain on the basis of the original ternary hyperedges. The new ternary hyperedges and the multi-layered multi-relational hypergraph are then combined to obtain a preliminary augmented hypergraph.

[0116] Step S2034: Calculate the contribution of each ternary hyperedge in the preliminary augmented hypergraph, normalize the contribution, and delete hyperedges with a contribution less than a preset contribution threshold to obtain the augmented hypergraph and relation gating coefficients.

[0117] In this embodiment, based on the preliminary augmented hypergraph and the causal chain structure set, the number of times each ternary hyperedge in the preliminary augmented hypergraph is included in the causal chain structure set, and the number of different channel nodes connected by each ternary hyperedge are counted. The counts and quantities are weighted and summed according to a preset ternary weighting rule to obtain the original contribution of each ternary hyperedge. The ternary weighting rule is set by those skilled in the art according to the actual situation.

[0118] Based on the original contribution of each ternary hyperedge, the original contribution is normalized, and the normalized contribution is limited to the range of zero to one. Ternary hyperedges with normalized contribution less than a preset contribution threshold are deleted from the initial augmented hypergraph to obtain the augmented hypergraph. The contribution threshold is set by those skilled in the art according to the actual situation.

[0119] Based on the normalized contribution of each ternary hyperedge in the augmented hypergraph, the ternary hyperedges are divided into time-type ternary hyperedges, channel-type ternary hyperedges, and operation-type ternary hyperedges according to whether the time period nodes connected by the ternary hyperedges are adjacent on the time axis, whether they are connected to the same channel node, and whether they are connected to the same operating condition node. The mean of the normalized contribution of time-type ternary hyperedges, channel-type ternary hyperedges, and operation-type ternary hyperedges is calculated by weighting, and the corresponding relation gating coefficients are obtained.

[0120] Step S204: Based on the augmented hypergraph and relation gating coefficients, determine the curvature parameters of the hyperbolic manifold, and perform embedding mapping on the time period nodes, channel nodes and working condition nodes in the augmented hypergraph in the hyperbolic manifold to obtain the hyperbolic hypergraph representation sequence.

[0121] The specific steps of step S204 are as follows:

[0122] Step S2041: Based on the augmented hypergraph, determine the time period node subgraphs for each working condition, calculate the graph distance between each time period node node in each time period node subgraph, generate a graph distance value set, take a graph distance value for each graph distance value in the graph distance value set, count the number of time period nodes whose graph distance is less than or equal to the corresponding graph distance value, and generate a growth curve.

[0123] In this embodiment, based on the augmented hypergraph and the work condition label set, the time period nodes connected to each work condition node and the connection relationships formed by the ternary hyperedges are extracted from the augmented hypergraph to obtain the time period node subgraph set of each work condition.

[0124] Based on the subgraph of nodes in each time period, a hop count rule is set within the same subgraph of nodes in the same time period. According to the hop count rule, the minimum hop count between any two nodes in each time period is calculated. The non-repeating values ​​of the obtained minimum hop count are arranged in ascending order to obtain the graph distance value set. The hop count rule is that if there is a connection path between any two nodes in the time period formed by n3 ternary hyperedges connected end to end, then the number of ternary hyperedges on the connection path is taken as the hop count between the two nodes in the time period.

[0125] Based on the graph distance value set, for each graph distance value in the set, each time period node in the subgraph of each time period node is used as a reference node. The number of different time period nodes that can be connected by a ternary hyperedge path when the graph distance is less than or equal to the graph distance value is counted, and the average of the statistical results of each reference node is calculated to obtain the number of time period nodes corresponding to the graph distance value.

[0126] Based on the graph distance in ascending order, each graph distance value is paired with the corresponding number of nodes in the time period to generate a growth curve that varies with the graph distance.

[0127] Step S2042: Compare the growth curve with the preset hyperbolic volume growth model, construct the curvature evaluation function, and optimize the curvature evaluation function to obtain the set of working condition curvature parameters.

[0128] The specific steps of step S2042 are as follows:

[0129] Step S20421: For each graph distance value in the graph distance value set, calculate the difference between the number of predicted nodes of the preset hyperbolic volume growth model at that graph distance value and the number of nodes of the growth curve at that graph distance value, and obtain a multi-scale difference sequence.

[0130] In this embodiment, based on the graph distance value set, each graph distance value in the graph distance value set is used as input in sequence and substituted into a preset hyperbolic volume growth model to obtain the number of predicted nodes under each graph distance value. The hyperbolic volume growth model is set by those skilled in the art according to the actual situation.

[0131] Read the actual number of nodes corresponding to each graph distance value from the growth curve, and take the difference between the predicted number of nodes and the actual number of nodes as the difference component corresponding to each graph distance value. Arrange the difference components corresponding to each graph distance value in ascending order of graph distance value to form a multi-scale difference sequence.

[0132] Step S20422: Weighted summation of the differences of the multi-scale difference sequence at the distance values ​​of each graph to generate a global difference quantity, and determination of the difference changes of the multi-scale difference sequence at the distance values ​​of adjacent graphs to generate a shape offset. The global difference quantity and the shape offset are combined according to the preset combination weights to construct a curvature evaluation function.

[0133] In this embodiment, based on the multi-scale difference sequence and the set of graph distance values, the differences of the multi-scale difference sequence at each graph distance value are weighted and summed according to the preset global weight coefficient to obtain a global difference quantity used to represent the overall degree of fitting deviation. The global weight coefficient is set by those skilled in the art according to the actual situation.

[0134] Based on the multi-scale difference sequence, the absolute difference of the difference components corresponding to the distance values ​​of adjacent graphs is calculated, and each absolute difference is weighted and summed according to the preset shape weight coefficient to obtain the shape offset used to represent the degree of change in the shape of the growth curve. The shape weight coefficient is set by those skilled in the art according to the actual situation.

[0135] Based on the global difference, shape offset, and preset combined weights, a curvature evaluation function is constructed by weighted combination. The combined weights are set by those skilled in the art according to the actual situation.

[0136] Step S20423: Based on the curvature evaluation function and the preset step size, generate a set of discrete curvature parameter values ​​within the preset curvature parameter search interval, and calculate the curvature evaluation function value corresponding to each curvature parameter value in the set of discrete curvature parameter values ​​to obtain the set of discrete curvature parameter values.

[0137] In this embodiment, based on the preset lower boundary value, upper boundary value, preset step size, and step size increment of the curvature parameter search interval, n4 curvature parameter values ​​are generated sequentially starting from the lower boundary value in an incremental step size manner. When the generated curvature parameter value is not greater than the upper boundary value, the generated curvature parameter value is added sequentially to the curvature parameter discrete value set to obtain a curvature parameter discrete value set covering the entire curvature parameter search interval.

[0138] Based on the set of discrete values ​​of curvature parameters, the discrete values ​​of each curvature parameter are input into the curvature evaluation function for calculation, and the curvature evaluation function value corresponding to each discrete value of curvature parameter is obtained. The discrete values ​​of curvature parameters and the corresponding curvature evaluation function values ​​are then paired to obtain the set of discrete values ​​of evaluated curvature parameters.

[0139] Step S20424: Based on the set of discrete values ​​of the curvature parameters, each discrete value of the curvature parameter is used as a graph node, adjacent pairs of discrete values ​​of the curvature parameters are used as graph edges, and the corresponding curvature evaluation function value is used as the node energy to construct a curvature candidate graph structure.

[0140] In this embodiment, based on the set of discrete values ​​of the curvature parameter, the discrete values ​​of each curvature parameter in the set are sorted from largest to smallest according to their numerical values, and each sorted discrete value of the curvature parameter is mapped to a graph node to obtain a node set.

[0141] Based on the sorting results, each pair of adjacent discrete values ​​of curvature parameters is determined as a pair of adjacent nodes, and graph edges are constructed for each pair of adjacent nodes to obtain the edge set.

[0142] Based on the set of discrete values ​​of the curvature parameters, the curvature evaluation function value corresponding to each discrete value of the curvature parameter is introduced into the corresponding graph node as node energy. The node set, edge set and node energy are combined to construct the curvature candidate graph structure.

[0143] Step S20425: Extract the descent direction of node energy in the curvature candidate graph structure, iteratively update the curvature candidate graph structure based on the descent direction until the node energy distribution meets the preset iteration termination condition, stop the iteration, and select the discrete value of curvature parameter corresponding to the node with the lowest node energy to obtain the set of working condition curvature parameters.

[0144] In this embodiment, based on the curvature candidate graph structure, the node energy of each graph node is used as the initial energy distribution. The weighted average value of the node energy of each graph node and its adjacent nodes is calculated. The weighted average value is subtracted from the node energy to obtain the local energy difference. Nodes with a local energy difference less than zero are identified as candidate nodes for energy decrease.

[0145] Based on the energy decrease candidate node, the neighboring node whose energy is less than that of the energy decrease candidate node is selected as the target node among the neighboring nodes connected to the energy decrease candidate node. An energy transfer path is determined along the corresponding graph edge between the energy decrease candidate node and the target node, and the node energy is updated along the energy transfer path according to the preset energy transfer coefficient to obtain the updated node energy distribution. The energy transfer coefficient is set by those skilled in the art according to the actual situation.

[0146] Based on the above method, the node energy of all nodes in the curvature candidate graph structure is updated, and the curvature candidate graph structure is iteratively updated until the node energy change amplitude of all nodes after two adjacent updates is less than the preset energy change threshold. Then the iteration stops, and the discrete value of the curvature parameter corresponding to the node with the smallest node energy at this time is taken as the curvature parameter of the working condition. The curvature parameters of each working condition are summarized to obtain the set of working condition curvature parameters. The energy change threshold is set by those skilled in the art according to the actual situation.

[0147] Step S2043: Based on the set of curvature parameters of the working conditions, construct a weighted graph Laplacian operator and perform spectral decomposition to generate the initial embedding vector of each node. Map the initial embedding vector to the hyperbolic manifold space of the corresponding working condition through hyperbolic exponential mapping to obtain the initial hyperbolic node representation set.

[0148] In this embodiment, based on the time period nodes, channel nodes, and working condition nodes connected by each ternary hyperedge in the augmented hypergraph, for any working condition, the ternary hyperedges associated with the working condition node are traversed in the augmented hypergraph. The connection weight between any two nodes connected by each ternary hyperedge is set as the relation gating coefficient corresponding to the ternary hyperedge. The connection weight between two nodes that do not have a ternary hyperedge connection relationship under the working condition is set to zero, thus obtaining the weighted adjacency matrix for each working condition.

[0149] Based on the weighted adjacency matrix under each working condition, the connection weights of the corresponding rows or columns of each node in the weighted adjacency matrix are summed to obtain the degree value of each node under each working condition. The degree matrix composed of the degree values ​​and the weighted adjacency matrix are then combined to construct the weighted graph Laplacian operator corresponding to each working condition.

[0150] Based on the set of curvature parameters for each working condition, spectral decomposition is performed on the weighted graph Laplacian operator corresponding to each working condition. From the feature vectors, n5 feature vectors with feature values ​​less than a preset feature value threshold are selected as the initial embedding vectors for each node corresponding to each working condition. The feature value threshold is set by those skilled in the art according to the actual situation.

[0151] Based on the set of curvature parameters for each working condition, hyperbolic exponential mapping is performed on the initial embedding vector of each node under each working condition according to the curvature parameters corresponding to each working condition. This yields the initial hyperbolic node representation of each node in the hyperbolic manifold space corresponding to each working condition. The initial hyperbolic node representations corresponding to each working condition are then summarized to obtain the set of initial hyperbolic node representations.

[0152] Step S2044: Based on the initial hyperbolic node representation set and relational gating coefficients, the nodes in the augmented hypergraph are sequentially taken as center nodes, the adjacent nodes of each center node are determined, and the hyperbolic node representations of the adjacent nodes are transformed into the tangent space of the hyperbolic manifold with the center node as the base point through hyperbolic logarithmic mapping to obtain a convergent hyperbolic node representation set.

[0153] In this embodiment, based on the relation gating coefficient and the category of each ternary hyperedge in the augmented hypergraph, the relation gating coefficient corresponding to each ternary hyperedge is used as the connection weight of the node pairs connected by each ternary hyperedge, and the connection weight of the node pairs that do not have a ternary hyperedge connection relationship is set to zero, thus obtaining the node connection weight matrix.

[0154] Based on the node connection weight matrix and a preset order, each node in the augmented hypergraph is taken as a central node in turn. Nodes with connection weights greater than zero to each central node are selected to form a set of neighboring nodes of each central node. The connection weights between each central node and its corresponding neighboring nodes are normalized to obtain the aggregate weight of each neighboring node to each central node. The order is set by those skilled in the art according to the actual situation.

[0155] Based on the initial hyperbolic node representation set, the hyperbolic node representations of each center node and each adjacent node corresponding to each center node are used as hyperbolic coordinates in the hyperbolic manifold space corresponding to each working condition. According to the preset hyperbolic logarithmic mapping rule, the hyperbolic coordinates of each center node and each adjacent node corresponding to each center node are mapped to the tangent space of each hyperbolic manifold with each center node as the base point, thus obtaining the tangent space representation of each center node and each adjacent node corresponding to each center node in the corresponding hyperbolic manifold tangent space. The hyperbolic logarithmic mapping rule is set by those skilled in the art according to the actual situation.

[0156] Based on the aggregation weights, the tangent space representations of each central node and its corresponding neighboring nodes in the tangent space of the hyperbolic manifold, the aggregation tangent space representations of each central node are weighted and summed according to the aggregation weights within the tangent space of the hyperbolic manifold corresponding to each central node, thus obtaining the aggregation tangent space representation of each central node.

[0157] Based on the aggregated tangent space representation of each central node, and combined with the set of curvature parameters for the working conditions, the radial correction term corresponding to each central node is determined, and the aggregated tangent space representation of each central node is radially corrected to obtain the updated tangent space representation of each central node.

[0158] Based on the updated tangent space representation of each central node, according to the preset hyperbolic exponential mapping rule, the updated tangent space representation of each central node is mapped from the hyperbolic manifold tangent space corresponding to each central node under each working condition back to the hyperbolic manifold space of each working condition, and the original hyperbolic node representation of each central node is replaced. The hyperbolic exponential mapping rule is set by those skilled in the art according to the actual situation.

[0159] The above update operation is performed sequentially on each node in the augmented hypergraph, and the hyperbolic node representation set is iteratively updated until the change amplitude of the hyperbolic node representation of each node between two adjacent updates is less than the preset change amplitude threshold. The update is then stopped, and the hyperbolic node representation set at this time is taken as the converged hyperbolic node representation set. The change amplitude threshold is set by those skilled in the art according to the actual situation.

[0160] Step S2045: Based on the convergent hyperbolic node representation set, fit the geodesic skeleton curve passing through the time period nodes in the hyperbolic manifold, perform geodesic orthogonal projection and projection coordinate stitching on the hyperbolic node representation of the time period nodes to obtain the hyperbolic hypergraph representation sequence.

[0161] In this embodiment, based on the convergent hyperbolic node representation set and the temporal order of the time period nodes, the hyperbolic node representations of each time period node are arranged to obtain the time period node trajectory.

[0162] According to the preset hyperbolic space geodesic fitting rules, a fitting curve passing through the trajectory of time period nodes is constructed in the hyperbolic manifold space corresponding to each working condition. According to the preset hyperbolic distance metric, the distance between the fitting curve and the hyperbolic node representation of each time period node is calculated, and the distance is taken as the hyperbolic distance. The hyperbolic distance metric and the hyperbolic space geodesic fitting rules are set by those skilled in the art according to the actual situation.

[0163] The fitted curve is optimized to minimize the sum of squares of the hyperbolic distances, resulting in a geodesic skeleton curve of the trajectories of the nodes in the specified time period. The parameter positions of each node in the specified time period on the geodesic skeleton curve are then determined.

[0164] Based on the geodesic skeleton curves under each working condition, the normal geodesic direction pointing to the geodesic skeleton curves under each working condition is determined in the hyperbolic manifold space corresponding to each working condition, based on the hyperbolic node representation of each time period node.

[0165] Based on the hyperbolic distance metric, the point with the minimum hyperbolic distance to the geodesic skeleton curve along the normal geodesic direction is searched, and this point is used as the projection point of the corresponding time period node. The arc length coordinates of each projection point on the geodesic skeleton curve and the hyperbolic distance from each time period node to each projection point are extracted. The arc length coordinates and hyperbolic distances are concatenated to obtain the projection coordinate vector of each time period node. The projection coordinate vectors are then arranged in the time order of the time period nodes to obtain the hyperbolic hypergraph representation sequence.

[0166] Step S205: Based on the hyperbolic hypergraph representation sequence, perform hyperedge aggregation and calculate channel weights and operating condition coupling weights.

[0167] In this embodiment, based on the hyperbolic hypergraph representation sequence and the augmented hypergraph, the projected coordinate vectors of each time period node in the hyperbolic hypergraph representation sequence are aggregated. According to the time period node, channel node, and working condition node connected by each ternary hyperedge in the augmented hypergraph, the ternary hyperedges related to the same time period node are selected. The projected coordinate vectors of the time period node, the hyperbolic node representations of the channel node and the working condition node are weighted and combined using the corresponding relational gating coefficients to obtain the hyperbolic hyperedge representations corresponding to each ternary hyperedge. The hyperbolic hyperedge representations are then arranged according to the time order of the time period nodes to obtain the hyperbolic hyperedge representation sequence.

[0168] Based on the hyperbolic hyperedge representation sequence and channel nodes, the hyperbolic hyperedge representations connected to the same channel node are summarized, and the hyperbolic hyperedge representations corresponding to each channel are weighted and summed according to the preset channel aggregation rules to obtain the aggregated representation of each channel.

[0169] Based on the aggregated representation of each channel, the channel response intensity of each channel is calculated, and the channel response intensity of each channel is normalized to obtain a set of channel weights. The aggregation rules are set by those skilled in the art according to the actual situation.

[0170] Based on the hyperbolic hyperedge representation sequence, channel weight set, and working condition nodes, the hyperbolic hyperedge representations connected to each working condition node are divided into hyperedge subsets corresponding to each working condition. The aggregate representation of each working condition is obtained by weighted summation according to the channel weights of the hyperbolic hyperedge representations in each hyperedge subset.

[0171] Based on the preset coupling metric rules and the aggregated representation of each working condition, the working condition coupling strength is calculated, and the working condition coupling strength is normalized to obtain the working condition coupling weight set. The coupling metric rules are set by those skilled in the art according to the actual situation.

[0172] Step S3: Based on the hyperbolic hypergraph representation sequence and weights, construct the conditional energy function and perform optimization calculations through a differentiable optimization layer to obtain the phase sequence, pattern label, and phase transition parameters of the shooting action.

[0173] The specific steps of step S3 are as follows:

[0174] Step S301: Based on the hyperbolic hypergraph representation sequence and weights, aggregate the time period nodes and adjacent hyperedges to obtain the stage observation feature sequence and the stage transition feature sequence.

[0175] In this embodiment, based on the augmented hypergraph, hyperbolic hypergraph representation sequence, channel weight set, working condition coupling weight set, and time order of time period nodes, the projected coordinate vectors of each time period node in the hyperbolic hypergraph representation sequence are traversed to determine the hyperbolic hyperedge representation connected to each time period node.

[0176] Based on the channel weight and operating condition coupling weight corresponding to each hyperbolic hyperedge, the hyperbolic hyperedge representations connected to each time period node are weighted and summed. The result of the weighted sum is used as the time period observation feature vector of each time period node. The time period observation feature vectors are arranged according to the time order of the time period nodes to obtain the time period observation feature sequence.

[0177] Based on the time period observation feature sequence and the preset stage division rules, the time period nodes are divided on the shooting action time axis. A continuous n6 time period nodes are divided into a stage node. According to the preset aggregation weight within the stage, the time period observation feature vectors within each stage node are weighted and summed to obtain the stage observation feature vector of each stage node.

[0178] The stage observation feature vectors are arranged according to the time order of the stage nodes to obtain the stage observation feature sequence. The stage division rules and the aggregation weights within the stages are set by those skilled in the art according to the actual situation.

[0179] Based on the stage observation feature sequence, the stage observation feature vectors of adjacent stage nodes are combined according to the time order of stage nodes. According to the preset transfer feature construction rules, the stage observation feature vectors of two adjacent stages are concatenated and differentially operated to generate the stage transfer feature vectors corresponding to each adjacent stage. The stage transfer feature vectors are arranged according to the time order of stage nodes to obtain the stage transfer feature sequence. The transfer feature construction rules are set by those skilled in the art according to the actual situation.

[0180] Step S302: Based on the stage observation feature sequence, stage transition feature sequence, preset stage set and shooting action pattern set, generate time chain, stage chain and pattern node, construct a three-part graph structure, and determine the observation factor, time transition factor and pattern consistency factor.

[0181] In this embodiment, based on the stage observation feature sequence, the stage transition feature sequence, the preset stage set and the shooting action mode set, time nodes are generated according to the arrangement order of the stage observation feature vectors in the stage observation feature sequence, and the time nodes are connected according to the order of the stage observation feature vectors corresponding to the time nodes on the shooting action time axis to obtain a time chain. The stage set and the shooting action mode set are set by those skilled in the art according to the actual situation.

[0182] Based on the stage set, each stage is mapped to a graph stage node, and the graph stage nodes are connected according to a preset stage arrangement rule to obtain a stage chain. The stage arrangement rule is set by those skilled in the art according to the actual situation.

[0183] Based on the set of shooting action patterns, each shooting action pattern is mapped to a pattern node, and time nodes, stage nodes, and pattern nodes are regarded as three types of nodes, forming a node set of a ternary graph structure.

[0184] Based on the time chain, stage chain, stage observation feature sequence and preset observation metric rules, the stage observation feature vector corresponding to each time node and the compatibility metric corresponding to each stage node are calculated. The compatibility metric is used as the observation factor to obtain the observation factor set. The observation metric rules are set by those skilled in the art according to the actual situation.

[0185] Based on the stage transition feature sequence, adjacent time nodes in the time chain and adjacent stage nodes in the stage chain are combined respectively, and the transition compatibility metric corresponding to each combination is calculated. Each transition compatibility metric is used as a time transition factor to obtain a set of time transition factors.

[0186] Based on the set of shooting action patterns and the stage chain, the stage sequence corresponding to each shooting action pattern is extracted, the matching compatibility metric of the pattern node and the stage node on the stage sequence is calculated, and each matching compatibility metric is used as a pattern consistency factor to obtain a set of pattern consistency factors.

[0187] Based on time nodes, stage nodes, model nodes, sets of observation factors, sets of time transition factors, and sets of model consistency factors, factor connections are established between corresponding node pairs in the ternary graph structure to construct a ternary graph structure containing observation factors, time transition factors, and model consistency factors.

[0188] Step S303: Based on the observation factor, time transition factor, model consistency factor and preset energy construction rules, calculate the factor energy of each observation factor, time transition factor and model consistency factor, and perform weighted summation of the factor energies according to preset energy weights to construct the conditional energy function.

[0189] In this embodiment, based on the set of observation factors, the set of time transition factors, the set of model consistency factors, and the preset energy construction rules, the compatibility metrics corresponding to each observation factor, each time transition factor, and each model consistency factor are mapped to factor energies, resulting in the set of observation factor energy, the set of time transition factor energy, and the set of model consistency factor energy. The energy construction rules are set by those skilled in the art according to the actual situation.

[0190] Based on the energy sets of observation factors, time transition factors, and pattern consistency factors, a conditional energy function is constructed by weighting and summing the energies of each factor according to preset energy weights. The energy weights are set by those skilled in the art based on actual conditions.

[0191] Step S304: Based on the conditional energy function, perform block gradient updates in the differentiable optimization layer, and perform constraint optimization through simplex projection and near-end projection to obtain the current update stage indicator vector, the current update mode indicator vector, and the current update stage transition parameters.

[0192] The specific steps of step S304 are as follows:

[0193] Step S3041: Based on the conditional energy function, perform hot start initialization in the differentiable optimization layer, generate initial values ​​for the stage indicator vector, mode indicator vector, and stage transition parameters, and introduce a small random perturbation to break the symmetry to obtain the initial optimization state.

[0194] In this embodiment, based on the conditional energy function, a differentiable optimization layer is constructed according to the stage nodes, firing action patterns, and adjacent stage combinations. In the differentiable optimization layer, a stage indicator vector is constructed to characterize stage selection based on the number of stage nodes, a mode indicator vector is constructed to characterize firing action pattern selection based on the number of firing action patterns, and a stage transition parameter vector is constructed to characterize stage transition intensity based on the number of adjacent stage combinations. Each component of the stage indicator vector is mapped one-to-one with the stage node, each component of the mode indicator vector is mapped one-to-one with the firing action pattern, and each component of the stage transition parameter vector is mapped one-to-one with the adjacent stage combination.

[0195] Based on the differentiable optimization layer, according to the preset hot start initialization rules, each component of the stage indicator vector, mode indicator vector and stage transition parameter vector is initialized to obtain the initial value of the stage indicator vector, the initial value of the mode indicator vector and the initial value of the stage transition parameter. The hot start initialization rules are set by those skilled in the art according to the actual situation.

[0196] Based on the preset random perturbation amplitude, small random perturbations are made on the initial values ​​of the stage indicator vector, the initial value of the mode indicator vector, and the initial value of the stage transition parameter to break the symmetry of the variable values ​​and obtain the initial optimization state of the differentiable optimization layer. The random perturbation amplitude is set by those skilled in the art according to the actual situation.

[0197] Step S3042: Based on the initial optimization state, calculate the gradient of the conditional energy function with respect to the initial values ​​of the stage indicator vector, the initial values ​​of the mode indicator vector, and the initial values ​​of the stage transition parameters. Update the gradient in blocks alternately according to the opposite direction of the gradient to obtain the unprojected intermediate optimization state. The unprojected intermediate optimization state includes the intermediate values ​​of the stage indicator vector, the intermediate values ​​of the mode indicator vector, and the intermediate values ​​of the stage transition parameters.

[0198] In this embodiment, based on the initial optimized state and the conditional energy function, according to the initial values ​​of the stage indicator vector, the initial values ​​of the mode indicator vector and the initial values ​​of the stage transition parameters, the partial derivatives of the conditional energy function with respect to the initial values ​​of the stage indicator vector, the initial values ​​of the mode indicator vector and the initial values ​​of the stage transition parameters are calculated. The partial derivatives are then combined to form the corresponding gradient vectors, thus obtaining the gradients of the stage indicator vector, the mode indicator vector and the stage transition parameters.

[0199] Based on the gradient of the stage indicator vector, the gradient of the mode indicator vector, and the gradient of the stage transition parameter, and according to the preset step size parameter, the differentiable optimization layer performs block-by-block alternating updates in the opposite direction of the gradient, keeping the initial values ​​of the mode indicator vector and the stage transition parameter unchanged, and updating the initial value of the stage indicator vector in the opposite direction of the gradient of the stage indicator vector to obtain the intermediate value of the stage indicator vector. The step size parameter is set by those skilled in the art according to the actual situation.

[0200] Based on the intermediate value of the phase indicator vector, keeping the intermediate value of the phase indicator vector and the initial value of the phase transition parameter unchanged, the initial value of the mode indicator vector is updated in the opposite direction of the gradient of the mode indicator vector to obtain the intermediate value of the mode indicator vector.

[0201] Based on the intermediate values ​​of the stage indicator vector and the mode indicator vector, keeping the intermediate values ​​of the stage indicator vector and the mode indicator vector unchanged, the initial values ​​of the stage transition parameters are updated in the opposite direction of the gradient of the stage transition parameters to obtain the intermediate values ​​of the stage transition parameters.

[0202] The intermediate values ​​of the stage indicator vector, the mode indicator vector, and the stage transition parameters are combined to obtain the unprojected intermediate optimized state.

[0203] Step S3043: Based on the unprojected intermediate optimization state, perform simplex projection on the intermediate values ​​of the stage indicator vector and the mode indicator vector, and perform near-end projection on the intermediate values ​​of the stage transition parameters that satisfy preset order constraints and sparsity constraints to obtain updated stage indicator vectors, mode indicator vectors and stage transition parameters.

[0204] In this embodiment, based on the unprojected intermediate optimization state, the intermediate values ​​of the stage indicator vector and the mode indicator vector are extracted. According to the preset simplex projection rule, the intermediate values ​​of the stage indicator vector are subjected to simplex projection, the components of each stage indicator vector that are less than zero are set to zero, and the components of each stage indicator vector are normalized to obtain the updated values ​​of the stage indicator vector that satisfy the non-negativity constraint and the normalization constraint. The simplex projection rule is set by those skilled in the art according to the actual situation.

[0205] Based on the simplex projection rule, the intermediate value of the mode indicator vector is subjected to simplex projection so that the components of the mode indicator vector satisfy the preset non-negativity constraints and normalization constraints, thereby obtaining the updated value of the mode indicator vector. The non-negativity constraints and normalization constraints are set by those skilled in the art according to the actual situation.

[0206] Based on the intermediate values ​​of stage transition parameters in the unprojected intermediate optimization state, a near-end projection rule is constructed according to preset order constraints and sparsity constraints. The intermediate values ​​of stage transition parameters are then projected near-end to satisfy the order constraints on the stage sequence, thereby obtaining updated values ​​of stage transition parameters that satisfy the order constraints and sparsity constraints. The order constraints and sparsity constraints are set by those skilled in the art according to the actual situation.

[0207] Based on the updated values ​​of the phase indicator vector, the pattern indicator vector, and the phase transition parameters, the updated values ​​of the phase indicator vector, the pattern indicator vector, and the phase transition parameters are combined to obtain the updated phase indicator vector, the pattern indicator vector, and the phase transition parameters.

[0208] Step S3044: Based on the updated stage indicator vector, mode indicator vector, and stage transition parameters, calculate the change amplitude of the conditional energy function between two adjacent updates. When the change amplitude is less than the preset stop threshold, stop updating and obtain the current update stage indicator vector, the current update mode indicator vector, and the current update stage transition parameters.

[0209] In this embodiment, based on the updated stage indicator vector, mode indicator vector, and stage transition parameters, the updated stage indicator vector, mode indicator vector, and stage transition parameters are substituted into the conditional energy function for calculation to obtain the conditional energy function value corresponding to this update.

[0210] Based on the stage indicator vector, mode indicator vector, and stage transition parameters obtained from the previous update at the end of steps S3042 and S3043, the conditional energy function value corresponding to the previous update is calculated.

[0211] Calculate the absolute value of the difference between the conditional energy function value of this update and the conditional energy function value of the previous update to determine the magnitude of change of the conditional energy function between two adjacent updates.

[0212] Based on the energy change amplitude, a preset stop threshold is used to determine the stage indicator vector, mode indicator vector, and stage transition parameters. The iteration is stopped when the energy change amplitude is less than the stop threshold. The stage indicator vector, mode indicator vector, and stage transition parameters at this point are then used as the current updated stage indicator vector, the current updated mode indicator vector, and the current updated stage transition parameters, respectively. The stop threshold is set by those skilled in the art based on the actual situation.

[0213] Step S305: Perform maximum component decoding on the current update phase indicator vector and the current update mode indicator vector to obtain the phase sequence, mode label and phase transition parameters of the shooting action.

[0214] In this embodiment, based on the current update stage indicator vector, the current update mode indicator vector, and the current update stage transition parameters, the components of the current update stage indicator vector are compared according to the preset maximum component decoding rule. The component with the largest value is determined, and the corresponding stage identifier is extracted. The stage identifiers are arranged according to the arrangement order of the stage indicator vectors on the shooting action time axis to obtain the stage sequence of the shooting action. The maximum component decoding rule is set by those skilled in the art according to the actual situation.

[0215] Based on the maximum component decoding rule, each component of the current update mode indicator vector is compared to determine the shooting action mode corresponding to the component with the largest value, and this shooting action mode is used as the mode label of the shooting action.

[0216] Based on the current update phase transition parameters, the values ​​of each parameter in the current update phase transition parameters are used as the phase transition parameters of the shooting action to obtain the phase transition parameters of the shooting action.

[0217] Step S4: Based on the stage sequence, pattern label, and stage transition parameters, construct a multi-time-distance stage hazard rate matrix to obtain the pattern recognition prediction results.

[0218] The specific steps of step S4 are as follows:

[0219] Step S401: Based on the phase sequence and phase transition parameters, calculate the dwell time and remaining time to firing for each phase in the phase sequence to obtain a set of phase time distance statistical features.

[0220] In this embodiment, based on the stage sequence and stage transition parameters, and according to the arrangement order of the stage sequence on the shooting action time axis, the adjacent stages in the stage sequence are traversed, and the start and end positions of each stage on the shooting action time axis are determined by combining the non-zero components in the stage transition parameters.

[0221] Based on the preset shooting action timeline time resolution, the start and end positions of each stage, the dwell time of each stage is calculated, and the dwell times of each stage are arranged according to the order of the stage sequence to obtain the stage dwell time sequence. The time resolution is set by those skilled in the art according to the actual situation.

[0222] Based on the phase sequence and phase dwell time sequence, according to the preset phase label configuration, the phase identifier corresponding to the firing action is determined, and the position of the phase identifier is determined in the phase sequence. The phase end time corresponding to the position is taken as the firing time. The phase label configuration is set by those skilled in the art according to the actual situation.

[0223] Based on the stage dwell time series, the time difference from the start time to the firing time of each stage is calculated. Each time difference is used as the remaining time from the start time to the firing time of each stage. The dwell time of each stage and the corresponding remaining time of each stage are combined to obtain the stage time distance statistical feature vector. Based on the stage identifier, the stage time distance statistical feature vector is sorted to obtain the stage time distance statistical feature set.

[0224] Step S402: Based on the statistical feature set of time intervals in each stage, construct a multi-time interval hazard rate grid map, and calculate the stage condition hazard rate of each grid point through forward recursion to obtain the multi-time interval stage hazard rate matrix.

[0225] In this embodiment, n7 target prediction time intervals are determined based on the stage time interval statistical feature set and the preset multi-time interval configuration set. The multi-time interval configuration set is set by those skilled in the art according to the actual situation.

[0226] Based on the target prediction time interval, the remaining time axis to firing is divided into n8 time intervals. Based on the stage identifier and time interval, a grid set is constructed. The combination of each stage and each time interval is regarded as a grid point, and a multi-time interval hazard rate grid map is constructed.

[0227] Based on the multi-time interval hazard rate grid map, according to the stage time interval statistical feature set, the number of firing events and non-firing events in each stage within each time interval are counted. According to the preset hazard rate construction rules, the empirical hazard rate of each grid point is calculated, and the empirical hazard rate is used as the initial stage conditional hazard rate of each grid point. The hazard rate construction rules are set by those skilled in the art according to the actual situation.

[0228] According to the preset forward recursion rule, the grid point corresponding to the time interval with the shortest remaining time is determined. Starting from the grid point corresponding to the time interval with the shortest remaining time, the stage condition hazard rate of each stage at adjacent time interval grid points is forward recursively recursively and gradually corrected along the direction of increasing remaining time to obtain the updated stage condition hazard rate of each grid point. The forward recursion rule is set by those skilled in the art according to the actual situation.

[0229] Based on the updated stage condition hazard rate of each grid point, the stage condition hazard rate of each stage in each time interval is filled into a preset two-dimensional array structure to construct a multi-time interval stage hazard rate matrix. The two-dimensional array structure is set by those skilled in the art according to the actual situation.

[0230] Step S403: Based on the multi-time interval stage hazard rate matrix, the preset set of predicted time intervals and the survival analysis relationship, calculate the survival probability and firing probability of each predicted time interval, and normalize and calibrate the firing probability of each predicted time interval to obtain the multi-time interval firing probability distribution.

[0231] In this embodiment, based on the multi-time-distance stage hazard rate matrix and the preset set of prediction time intervals, the stage condition hazard rate of each stage in the multi-time-distance stage hazard rate matrix in each prediction time interval is sorted out. The set of prediction time intervals is set by those skilled in the art according to the actual situation.

[0232] According to the preset stage aggregation rules, the stage condition hazard rates of each stage within the same prediction time interval are weighted and summarized to obtain an equivalent condition hazard rate sequence arranged according to the prediction time interval. The stage aggregation rules are set by those skilled in the art based on the actual situation.

[0233] Based on the equivalent conditional hazard rate sequence, and according to the preset survival analysis relationship, the equivalent conditional hazard rate in each prediction time interval is converted into the corresponding survival probability sequence and firing probability sequence, so as to obtain the survival probability and firing probability of each prediction time interval. The survival analysis relationship is set by those skilled in the art according to the actual situation.

[0234] The survival probability and firing probability of each predicted time interval are normalized and calibrated to obtain the multi-time interval firing probability distribution.

[0235] Step S404: Based on the multi-time interval firing probability distribution and pattern label, compare the firing probability of each predicted time interval, select the predicted time interval with the highest firing probability, and obtain the pattern recognition prediction result.

[0236] In this embodiment, based on the multi-time interval firing probability distribution and pattern label, the firing probability corresponding to each predicted time interval is arranged according to the order of the predicted time intervals. The firing probability of each predicted time interval is compared, and the predicted time interval with the largest firing probability value is determined. This predicted time interval is used as the firing predicted time interval of the current shooting action.

[0237] Based on the pattern label and the firing prediction time interval, the pattern label is used as the shooting action category, and the firing prediction time interval is used as the corresponding time prediction result. The combination yields the pattern recognition prediction result.

[0238] exist Figure 2In the diagram, the elliptical region on the left represents the firing action timing signal set before alignment, and the colored dashed lines within the elliptical region represent m timing segments formed by connecting n sampling points when collecting the firing action timing signal set and operational information. These timing segments represent local actions such as aiming, pre-pressing, and micro-jittering. Different colors represent timing segments under different operational conditions. The directions of the colored dashed lines are random and scattered throughout the elliptical region, indicating that the action segments under different operational conditions are not aligned. The positions of the same type of action segments vary greatly under different operational conditions. The entire left ellipse represents the chaotic and positionally offset unaligned firing action segments under multiple operational conditions.

[0239] The large circle on the right represents the hyperbolic manifold space. The orange dashed curve in the large circle represents the geodesic skeleton curve fitted in the hyperbolic manifold space, that is, the skeleton trajectory of the firing action stage. The curve starts from the center of the circle and gradually bends outward, representing the firing action from the initial stage of preparation and preload to the later stage of near-fire and follow-up actions. The pentagonal nodes, star nodes, triangular nodes, square nodes and circular nodes on the orange dashed line each represent a typical stage. There are not only five typical stages, but only five types of nodes are drawn here for the sake of clarity. The nodes are arranged along the path of the orange dashed line, representing that each typical stage has a clear order of evolution over time. The typical stages include preparation, preload, stabilization, firing and follow-up.

[0240] The blue and green dotted arcs in the large circle represent the multi-layered clustering of shooting action segments under different conditions on both sides of the skeleton trajectory of the shooting action stage after being projected onto the hyperbolic manifold space. This represents the convergence of the original scattered distribution of the same type of action segments in the hyperbolic manifold space after alignment into a hierarchical structure arranged along the skeleton trajectory of the shooting action stage.

[0241] The short black arrow pointing from the left ellipse to the right large circle indicates that the misaligned shooting action sequence segments under multiple conditions are mapped to the hyperbolic manifold space after optimal transmission alignment and multi-layer, multi-relationship hypergraph modeling.

[0242] The thick black horizontal line below the large circle on the right represents the prediction time range, i.e. the prediction time interval axis. The small thick vertical lines on the thick line divide the prediction time interval axis into m1 prediction time intervals. Each interval is a preset set of prediction time intervals. The diagonal rectangle on the prediction time interval axis falls on the m2th interval of the prediction time interval axis, which means that the m2th interval is the time period with the highest selected firing probability.

[0243] The long black arrow in the large circle on the right represents inputting the phase sequence, pattern label, and phase transition parameters corresponding to the typical phase nodes on the skeletal trajectory of the shooting action phase in the hyperbolic manifold space into the multi-time-range phase hazard rate matrix, and mapping the predicted multi-time-range firing probability distribution onto the prediction time-range axis below the large circle on the right.

[0244] Example 2:

[0245] Please see Figure 4 The present invention provides an embodiment of a shooting action pattern recognition and prediction system, comprising: a signal acquisition module, an optimal transmission module, a function construction module, and a result output module.

[0246] The signal acquisition module is used to acquire the firing action timing signal set and working condition information, and based on the firing action timing signal set and working condition information, to perform time alignment and amplitude normalization to obtain the aligned firing action timing signal set and working condition label set.

[0247] The optimal transmission module is used to perform spatiotemporal distribution alignment based on the aligned shooting action timing signal set and the working condition label set, construct a multi-layer multi-relationship hypergraph, and perform hyperbolic manifold embedding to obtain the hyperbolic hypergraph representation sequence and weights.

[0248] The function construction module is used to construct a conditional energy function based on the hyperbolic hypergraph representation sequence and weights, and to perform optimization calculations through a differentiable optimization layer to obtain the phase sequence, pattern label and phase transition parameters of the shooting action.

[0249] The result output module is used to construct a multi-temporal stage hazard rate matrix based on stage sequence, pattern label and stage transition parameters, and obtain pattern recognition prediction results.

[0250] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0251] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A shooting action pattern recognition and prediction method, characterized in that, include: Collect the firing action timing signal set and working condition information, and perform time alignment and amplitude normalization based on the firing action timing signal set and working condition information to obtain the aligned firing action timing signal set and working condition label set; Based on the aligned shooting action timing signal set and the working condition label set, spatiotemporal distribution alignment is performed through optimal transmission to construct a multi-layer multi-relationship hypergraph, and hyperbolic manifold embedding is performed to obtain the hyperbolic hypergraph representation sequence and weights; Based on the hyperbolic hypergraph representation sequence and weights, a conditional energy function is constructed and optimized through a differentiable optimization layer to obtain the phase sequence, pattern label and phase transition parameters of the shooting action; Based on stage sequences, pattern labels, and stage transition parameters, a multi-temporal stage hazard rate matrix is ​​constructed to obtain pattern recognition prediction results.

2. The shooting action pattern recognition and prediction method according to claim 1, characterized in that, The method, based on the aligned shooting action timing signal set and the working condition label set, performs spatiotemporal distribution alignment through optimal transmission, constructs a multi-layer, multi-relationship hypergraph, and performs hyperbolic manifold embedding to obtain a hyperbolic hypergraph representation sequence and weights, including: Based on the aligned firing action timing signal set and the working condition label set, a cost kernel is constructed and optimal transmission is performed to obtain the working condition aligned timing segment sequence and the same working condition baseline parameter set; Based on the time-series segment sequence aligned with the working condition and the baseline parameter set of the same working condition, time period nodes, channel nodes and working condition nodes are generated through node mapping, and ternary hyperedges connecting the time period nodes, channel nodes and working condition nodes are generated to construct a multi-layer multi-relationship hypergraph. Based on a multi-layer, multi-relation hypergraph, pre-selected relations that meet the preset selection rules are selected, and the pre-selected relations are used as seeds for cross-layer generation to obtain an augmented hypergraph and relation gating coefficients; Based on the augmented hypergraph and relation gating coefficients, the curvature parameters of the hyperbolic manifold are determined. The time period nodes, channel nodes and working condition nodes in the augmented hypergraph are embedded and mapped in the hyperbolic manifold to obtain the hyperbolic hypergraph representation sequence. Based on the hyperbolic hypergraph representation sequence, hyperedge aggregation is performed to calculate channel weights and operating condition coupling weights.

3. The shooting action pattern recognition and prediction method according to claim 2, characterized in that, The process involves constructing a cost kernel and performing optimal transmission based on the aligned firing action timing signal set and the condition label set, resulting in a condition-aligned timing segment sequence and a baseline parameter set for the same condition, including: Based on the working condition tag set and the preset time length, the time window segment of the alignment shooting action timing signal set is divided, and the amplitude statistics and frequency domain statistics of each time window are calculated to obtain the initial time segment set with working condition tags. Based on the initial set of time series segments, the time series segments are clustered within each working condition, and the statistical center and amplitude scale are calculated to obtain a prototype set of working condition-level segments. Based on the prototype set of layered segments under different operating conditions, a layered transmission distance matrix is ​​constructed, and a cost kernel is generated by combining the operating condition labels to obtain the cost value. Based on cost value, optimal transmission is performed, and the initial time series segments are rearranged and weighted to obtain the working condition aligned time series segment sequence and the same working condition baseline parameter set.

4. The shooting action pattern recognition and prediction method according to claim 3, characterized in that, The process involves using a multi-layer, multi-relation hypergraph to filter out pre-selected relations that meet preset filtering rules. These pre-selected relations are then used as seeds for cross-layer generation to obtain an augmenting hypergraph and relation gating coefficients, including: Based on a multi-layer, multi-relation hypergraph, the structure score and causal tendency score of each ternary hyperedge are calculated to generate a relation candidate set. The causal prior degree of each candidate relation in the relation candidate set is calculated to determine the directed connections and obtain an updated relation candidate set. Based on the updated candidate relation set, pre-selected relations that meet the preset screening rules in terms of structure score and causal precedence are selected as seed relations. The starting time node of each sub-relation is taken as the root node, and the time node is expanded level by level according to the directed connection to obtain the causal chain structure set. Based on the causal chain structure set, new ternary hyperedges are generated by combining the corresponding channel nodes and working condition nodes. The new ternary hyperedges are then added to the multi-layer, multi-relationship hypergraph with the causal chain as the skeleton to obtain the preliminary augmented hypergraph. The contribution of each ternary hyperedge in the preliminary augmented hypergraph is statistically analyzed, the contribution is normalized, and hyperedges with a contribution less than a preset contribution threshold are deleted to obtain the augmented hypergraph and relation gating coefficients.

5. The shooting action pattern recognition and prediction method according to claim 4, characterized in that, The hyperbolic manifold curvature parameters are determined based on the augmented hypergraph and relation gating coefficients. Then, time-period nodes, channel nodes, and operational condition nodes in the augmented hypergraph are embedded and mapped within the hyperbolic manifold to obtain a hyperbolic hypergraph representation sequence, including: Based on the augmented hypergraph, the time period node subgraphs for each working condition are determined, the graph distance between each time period node node in each time period node subgraph is calculated, a graph distance value set is generated, and for each graph distance value in the graph distance value set, the number of time period nodes whose graph distance is less than or equal to the corresponding graph distance value is counted, and a growth curve is generated. The growth curve is compared with the preset hyperbolic volume growth model, a curvature evaluation function is constructed, and the curvature evaluation function is optimized to obtain a set of curvature parameters for the working condition. Based on the set of curvature parameters for the working conditions, a weighted graph Laplacian operator is constructed and spectral decomposition is performed to generate the initial embedding vectors of each node. The initial embedding vectors are then mapped to the hyperbolic manifold space of the corresponding working condition through hyperbolic exponential mapping to obtain the initial hyperbolic node representation set. Based on the initial hyperbolic node representation set and relational gating coefficients, the nodes in the augmented hypergraph are sequentially taken as center nodes, and the adjacent nodes of each center node are determined. The hyperbolic node representations of the adjacent nodes are transformed into the tangent space of the hyperbolic manifold with the center node as the base point through hyperbolic logarithmic mapping, resulting in a convergent hyperbolic node representation set. Based on the convergent set of hyperbolic node representations, geodesic skeleton curves passing through time period nodes are fitted in the hyperbolic manifold. The hyperbolic node representations of time period nodes are then subjected to geodesic orthogonal projection and splicing of projected coordinates to obtain a hyperbolic hypergraph representation sequence.

6. The shooting action pattern recognition and prediction method according to claim 5, characterized in that, The process involves comparing the growth curve with a preset hyperbolic volume growth model, constructing a curvature evaluation function, and optimizing the curvature evaluation function to obtain a set of working condition curvature parameters, including: For each graph distance value in the graph distance value set, calculate the difference between the number of predicted nodes of the preset hyperbolic volume growth model at the corresponding graph distance value and the number of nodes of the growth curve at the corresponding graph distance value, and obtain a multi-scale difference sequence; The differences of the multi-scale difference sequence at each graph distance value are weighted and accumulated to generate a global difference value. The changes in the difference of the multi-scale difference sequence at the distance values ​​of adjacent graphs are determined to generate a shape offset. The global difference value and the shape offset are combined according to the preset combination weights to construct a curvature evaluation function. Based on the curvature evaluation function and the preset step size, a set of discrete curvature parameter values ​​is generated within the preset curvature parameter search interval. The curvature evaluation function value corresponding to each curvature parameter value in the set of discrete curvature parameter values ​​is calculated to obtain the set of discrete curvature parameter values. Based on the set of discrete values ​​of the curvature parameters, each discrete value of the curvature parameter is used as a graph node, adjacent pairs of discrete values ​​of the curvature parameters are used as graph edges, and the corresponding curvature evaluation function value is used as the node energy to construct a curvature candidate graph structure. Extract the descent direction of node energy in the curvature candidate graph structure, and iteratively update the curvature candidate graph structure based on the descent direction until the node energy distribution meets the preset iteration termination condition. Stop the iteration, select the discrete curvature parameter values ​​corresponding to the node with the lowest node energy, and obtain the set of working condition curvature parameters.

7. The shooting action pattern recognition and prediction method according to claim 6, characterized in that, The conditional energy function is constructed based on the hyperbolic hypergraph representation sequence and weights, and optimized through a differentiable optimization layer to obtain the phase sequence, mode label, and phase transition parameters of the shooting action, including: Based on the hyperbolic hypergraph representation sequence and weights, the time period nodes and adjacent hyperedges are aggregated to obtain the stage observation feature sequence and the stage transition feature sequence; Based on the phase observation feature sequence, phase transition feature sequence, preset phase set and shooting action pattern set, time chain, phase chain and pattern node are generated, a three-part graph structure is constructed, and observation factor, time transition factor and pattern consistency factor are determined. Based on the observation factor, time transition factor, model consistency factor and preset energy construction rules, the factor energy of each observation factor, time transition factor and model consistency factor is calculated, and the factor energy is weighted and summed according to the preset energy weight to construct the conditional energy function. Based on the conditional energy function, block gradient updates are performed in the differentiable optimization layer, and constraint optimization is performed through simplex projection and near-end projection to obtain the current update stage indicator vector, the current update mode indicator vector, and the current update stage transition parameters. The current update phase indicator vector and the current update mode indicator vector are decoded by maximum component analysis to obtain the phase sequence, mode label and phase transition parameters of the shooting action, wherein the phase transition parameters are the current update phase transition parameters.

8. The shooting action pattern recognition and prediction method according to claim 7, characterized in that, The conditional energy function-based block gradient update is performed in the differentiable optimization layer, and constrained optimization is performed through simplex projection and near-end projection to obtain the current update stage indicator vector, the current update mode indicator vector, and the current update stage transition parameters, including: Based on the conditional energy function, hot-start initialization is performed in the differentiable optimization layer to generate initial values ​​for the stage indicator vector, the mode indicator vector, and the stage transition parameters. Random perturbations are introduced to break the symmetry and obtain the initial optimization state. Based on the initial optimization state, the gradient of the conditional energy function with respect to the initial values ​​of the stage indicator vector, the initial values ​​of the mode indicator vector, and the initial values ​​of the stage transition parameters is calculated. The gradient is then updated in blocks alternately according to the opposite direction to obtain the unprojected intermediate optimization state, which includes the intermediate values ​​of the stage indicator vector, the intermediate values ​​of the mode indicator vector, and the intermediate values ​​of the stage transition parameters. Based on the unprojected intermediate optimization state, simplex projection is performed on the intermediate values ​​of the stage indicator vector and the mode indicator vector, and near-end projection is performed on the intermediate values ​​of the stage transition parameters to satisfy preset order constraints and sparsity constraints, so as to obtain updated stage indicator vector, mode indicator vector and stage transition parameters. Based on the updated stage indicator vector, mode indicator vector, and stage transition parameters, the change magnitude of the conditional energy function between two adjacent updates is calculated. When the change magnitude is less than a preset stopping threshold, the update is stopped, and the current update stage indicator vector, the current update mode indicator vector, and the current update stage transition parameters are obtained.

9. The shooting action pattern recognition and prediction method according to claim 8, characterized in that, The process involves constructing a multi-temporal stage hazard rate matrix based on stage sequences, pattern labels, and stage transition parameters to obtain pattern recognition prediction results, including: Based on the phase sequence and phase transition parameters, the dwell time and remaining time to firing in each phase of the phase sequence are statistically analyzed to obtain a set of phase time interval statistical features. Based on the statistical feature set of time intervals in each stage, a multi-time interval hazard rate grid map is constructed. The stage condition hazard rate of each grid point is calculated by forward recursion to obtain the multi-time interval stage hazard rate matrix. Based on the multi-time interval stage hazard rate matrix, the preset set of predicted time intervals and the survival analysis relationship, the survival probability and firing probability of each predicted time interval are calculated, and the firing probability of each predicted time interval is normalized and calibrated to obtain the multi-time interval firing probability distribution. Based on the multi-time interval firing probability distribution and pattern labels, the firing probability of each predicted time interval is compared, and the predicted time interval with the highest firing probability is selected to obtain the pattern recognition prediction result.

10. A shooting action pattern recognition and prediction system, used to implement the shooting action pattern recognition and prediction method according to any one of claims 1-9, characterized in that, include: The signal acquisition module is used to acquire the firing action timing signal set and working condition information. Based on the firing action timing signal set and working condition information, time alignment and amplitude normalization are performed to obtain the aligned firing action timing signal set and working condition label set. The optimal transmission module is used to perform spatiotemporal distribution alignment based on the aligned shooting action timing signal set and the working condition label set, construct a multi-layer multi-relationship hypergraph, and perform hyperbolic manifold embedding to obtain the hyperbolic hypergraph representation sequence and weights. The function construction module is used to construct the conditional energy function based on the hyperbolic hypergraph representation sequence and weights, and perform optimization calculations through a differentiable optimization layer to obtain the phase sequence, mode label and phase transition parameters of the shooting action; The results output module is used to construct a multi-temporal stage hazard rate matrix based on stage sequence, pattern label and stage transition parameters, and obtain pattern recognition prediction results.

Images