Stability evaluation system for air pistol based on time sequence feature extraction

Abstract

The present application relates to computer data processing and motion timing analysis technical field, specifically to the air pistol stability evaluation system based on timing feature extraction, system receives aiming point trajectory, gun body posture, trigger action and firing time mark, to establish unified time base with firing time mark, according to the division of continuous action stage according to the change of multiple source sequence, and segmented elastic registration is carried out under the condition of maintaining stage sequence. Based on the phase-ordered time delay influence diagram constructed by the stage characteristics, change direction, response sequence and cross-firing repetition, the closed-loop compensation chain between the gun body posture change, aiming correction, trigger sustained action and firing event is extracted, and the evaluation results are generated by combining the closed-loop integrity, time lag stability, stage connection continuity and registration uncertainty. The present application can reduce the stage dislocation caused by fixed window, distinguish the apparent stability formed by normal delay compensation and accidental offset, and make the results correspond to the continuous control relationship in the actual firing process.

Description

Technical Field

[0001] This invention relates to the field of computer data processing and motion timing analysis technology, specifically to an air pistol stability evaluation system based on timing feature extraction. Background Technology

[0002] Existing air pistol stability assessments typically use the aiming trajectory, gun attitude changes, trigger action changes, and firing time as data foundations during a single or multiple shots. The conventional approach involves extracting a fixed-length data window before and after the firing moment, filtering, normalizing, and statistically calculating various time-series data, then extracting features such as trajectory dispersion range, local fluctuation amplitude, angular motion change, trigger action duration, and the difference in changes before and after firing. A stability score is then generated through weighted summation, rule-based judgment, or a training-based classification model. Some schemes compress different firing samples to a uniform length for average trajectory comparison or similarity calculation. These schemes usually treat each data channel as synchronously changing objects, assuming that the positions of stable aiming, stable holding, progressive pull, firing disturbance, and post-firing hold phases are basically consistent across different shooters and firing processes within a fixed window, and determining the stability state based on the amplitude of a single channel or the combined relationship of statistical features from multiple channels.

[0003] Fixed windows and uniform length processing cannot maintain the true sequential relationship between action stages in different firing processes, nor can they handle the objective response time lag between aiming point changes, gun attitude corrections, and trigger action, leading to semantic misalignment of the time segments being compared. When the progressive pull phase of a firing occurs earlier or later, the data corresponding to the same time position may belong to the stable holding phase and the firing disturbance phase, respectively. When gun attitude deviations are offset by aiming corrections after a delay, simple synchronization statistics will interpret the normal compensation process as multi-channel fluctuations. When opposite changes in different channels are accidentally offset within a fixed window, a lower overall dispersion may result. The stability results obtained in this way are difficult to distinguish between true stability with continuous control relationships and apparent stability formed by stage mismatch, neglect of time lags, or accidental offsetting, and cannot objectively characterize the dynamic coordination state between various time-series features in a single firing process. Summary of the Invention

[0004] The purpose of this invention is to provide an air pistol stability evaluation system based on temporal feature extraction, which can solve the problems in the background art mentioned above.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An air pistol stability evaluation system based on temporal feature extraction includes a data interface, a memory, and a processor. The data interface receives aiming point trajectory sequences, gun body posture sequences, trigger action sequences, and firing timestamps. The memory stores programs that can be executed by the processor. When the processor executes the program, it establishes a unified time reference based on the firing timestamps. The continuous action phases in the firing process are determined based on the changing characteristics of the aiming point trajectory sequence, the gun body posture sequence, and the trigger action sequence. Segmented flexible registration is performed while maintaining the sequential order of the continuous action phases; A phase-order-preserving time delay influence diagram is constructed based on the stage characteristics, change direction, response sequence relationship and cross-firing repeatability after registration. The closed-loop compensation chain is extracted from the phase-sequence-preserving delay effect diagram, and the stability assessment result is generated.

[0006] Preferably, when the processor determines the continuous action phase, it calculates the local curvature change and convergence direction consistency for the aiming point trajectory sequence, calculates the angular motion energy change for the gun body attitude sequence, calculates the action start change and continuous change for the trigger action sequence, and forms a candidate boundary set for each change on a unified time reference. Establish a stage state chain in the order of entering stable aiming, stable holding, progressive pressing, firing disturbance and post-firing holding, map the candidate boundary set to the stage state chain, and select the boundary position of each continuous action stage based on the consistency of multi-source features between adjacent boundaries, boundary temporal monotonicity and firing time scale constraints.

[0007] Preferably, when the processor performs segmented elastic registration, it uses the start and end boundaries of each continuous action stage as fixed anchor points, establishes standardized stage time coordinates within each continuous action stage, and obtains the registration path within the stage based on the joint similarity of the changes in aiming point trajectory, gun body posture, and trigger action. The registration path within the stage satisfies the constraints of monotonically increasing time coordinates, continuous adjacent mapping points, and uncrossable stage boundaries. Multi-source sequences fired in the same instance share the stage-level scaling path, and each data channel only superimposes restricted residual offsets on the stage-level scaling path to obtain a registration timing sequence with consistent stage semantics.

[0008] Preferably, when the processor constructs the phase-preserving delay effect diagram, it sets the aiming point change characteristics, gun body attitude change characteristics, and trigger action change characteristics in each continuous action phase as nodes respectively; For any two nodes, calculate the directed candidate time delay within the stage sequence range defined by the relationship between adjacent action stages, and determine the directed connection based on the consistency of the change direction corresponding to the candidate time delay, the relationship between the previous and subsequent responses, the frequency of cross-shot occurrence, and the dispersion of the candidate time delay. Write the connection direction, time delay interval, change sign, and cross-fire repetition into the attributes of the directed connection, and delete directed connections that conflict with the phase order or appear only in one firing sample.

[0009] Preferably, when the processor selects the boundary position of each continuous action stage, it establishes a boundary evidence vector for each candidate boundary, which includes the aiming point trajectory turning evidence, the gun body attitude energy turning evidence, the trigger action turning evidence, and the firing time marker distance, and constructs a directed boundary path on the stage state chain. The adjacent nodes of the directed boundary path correspond to the adjacent stage boundaries respectively. The path cost consists of the intra-stage feature stability, the adjacent stage feature difference, the cross-data channel boundary deviation, and the boundary order violation term. The stage boundary is determined by finding the complete boundary path with the minimum path cost, and boundary confidence information is generated based on the path cost difference between the complete boundary path and the adjacent candidate paths.

[0010] Preferably, when the processor obtains the registration path within the stage, it decomposes the registration displacement of the multi-source sequence into a shared displacement that characterizes the overall rhythm change of the same continuous action stage and a channel residual displacement that characterizes the local response difference of a single data channel. The shared displacement is determined by joint similarity, and the channel residual displacements are determined by the local differences between the corresponding data channels and the shared stage template, satisfying the constraints of weighted sum of residual displacements of each channel, continuous residual change rate, and no change in the order of original channel responses. The shared displacement and the channel residual displacement are combined to form the final mapping path, and the differences between multiple solutions of the mapping path are recorded as registration uncertainty information.

[0011] Preferably, when the processor determines the directed connection, it establishes a time delay sample set across firing cycles for each directed candidate time delay, and divides the time delay subset according to the continuous action stage, change sign and response direction; For each time-delay subset, calculate the conditional dependency after controlling the historical features of other nodes, and compare the conditional dependency with the direct dependency when the historical features of other nodes are not controlled. When the direct dependency is explained by the historical features of the common predecessor node, or when the candidate delay reverses direction in different firing samples, the corresponding directed connection is deleted. For the retained directed connections, connection confidence is generated based on the conditional dependency, the proportion of direction retention, and the degree of time delay concentration.

[0012] Preferably, when the processor extracts the closed-loop compensation chain, it takes the node corresponding to the gun body attitude change feature as the starting node, searches along the phase-sequence-preserving time delay influence map to find the directed path that sequentially passes through the node corresponding to the aiming point correction feature, the node corresponding to the trigger continuous action feature, and reaches the firing event node. For each directed path, the time delay intervals of each directed connection are checked for continuity according to the order of the continuous action phases. Paths that do not meet the compensation direction constraints are filtered out based on the correspondence between the direction of change of gun body attitude and the direction of correction of aiming point, and the order of change of aiming point deviation and continuous action of trigger. The preserved paths that share a common starting node and a common firing event node are merged into the closed-loop compensation chain.

[0013] Preferably, when the processor propagates the registration uncertainty information to the phase-preserving time delay influence map, for any directed connection, it re-determines the candidate time delay using the final mapping path and alternative mapping paths that have multiple solution differences with the final mapping path, forming a candidate time delay offset set. The timing order preservation ratio, connection direction preservation ratio, and candidate delay interval overlap ratio of the two nodes of the directed connection are determined based on the candidate delay offset set, and the timing order preservation ratio, connection direction preservation ratio, and candidate delay interval overlap ratio are combined into a connection robustness coefficient. The properties of the corresponding directed connection in the phase-order-preserving delay effect graph are corrected using the connection robustness coefficient.

[0014] Preferably, when the processor generates the stability evaluation result, it extracts the stage coverage ratio of the closed-loop compensation chain, the cross-fire repetition of each directed connection, the time delay connection deviation of adjacent directed connections, the compensation direction maintenance ratio, and the cross-fire occurrence ratio of the closed-loop compensation chain, and constructs a chain-level correlation matrix according to the relationship between shared nodes and shared directed connections between different closed-loop compensation chains. The feature weights corresponding to the shared nodes and the shared directed connections are merged according to the chain-level association matrix to form a stability feature vector composed of closed-loop integrity, time-delay stability, stage connection continuity and compensation redundancy. The stability evaluation result is then generated based on the stability feature vector.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. A unified time reference is established using the firing timescale. Based on the changing characteristics of the aiming point trajectory, rifle posture, and trigger action, continuous action stages are determined. While maintaining the sequence of stages, segmented flexible registration is performed, ensuring that data belonging to the same action semantics in different firing samples are placed in the corresponding stage positions. The registered stage characteristics are used to construct a phase-order-preserving time-delay influence diagram based on the direction of change, response sequence, repeatability across firing cycles, and time-delay dispersion. A closed-loop compensation chain consisting of rifle posture changes, aiming corrections, continuous trigger action, and the firing event is then extracted from the diagram. Therefore, stability evaluation is no longer based solely on the fluctuation amplitude of a single channel or the synchronous correlation within a fixed window, but rather on the consistency of action stages and the repeatable delay compensation relationship between multi-source time sequences. This distinguishes normal delay correction processes from disordered fluctuations and eliminates the apparent low-dispersion state formed by accidental cancellation between different channels, ensuring that the stability results correspond to the continuous control relationship in the actual firing process.

[0016] 2. The candidate stage boundaries are jointly determined by the aiming trajectory transition, attitude motion energy transition, trigger action transition, and firing timescale distance. The order and continuity of adjacent stages are constrained by directed boundary paths, reducing interference from single data channel anomalies on stage division. During segmented registration, the shared displacement formed by overall rhythm changes is separated from the residual displacement formed by local responses of single channels. Registration uncertainties generated by alternative mapping paths are propagated to time-delay connections, ensuring that connection direction, timing sequence, and time-delay intervals are retained only when consistent across multiple feasible registration results. False connections caused by common preceding features are further filtered out using conditional dependencies, and connection robustness, stage coverage, time-delay connection deviation, and cross-fire repetition are aggregated into a stability feature vector. These processes establish a hierarchical verification relationship between stage boundaries, registration paths, time-delay connections, and closed-loop compensation chains, reducing the impact of local noise, sample rhythm differences, and multiple registration solutions on the evaluation results. Attached Figure Description

[0017] Figure 1 The flowchart shows the overall process of air pistol stability assessment based on time-series feature extraction. Figure 2 Flowchart for determining the continuous action phase and segmented flexible registration; Figure 3 Flowchart for constructing the phase-order-preserving time-delay effect diagram and generating stability results; Figure 4 A comparison chart of direct dependencies, conditional dependencies, and explained variables; Figure 5 A comparison chart of the costs of complete boundary paths; Figure 6 This is a graph showing the composition of the robustness coefficients of the directed connections. Detailed Implementation

[0018] In one embodiment, please refer to Figure 1 A stability evaluation system for an air pistol based on temporal feature extraction includes a data interface, a memory, and a processor. The data interface receives aiming point trajectory sequences, gun body posture sequences, trigger action sequences, and firing timestamps. The memory stores programs executable by the processor. When the processor executes the program, it establishes a unified time reference based on the firing timestamps, determines the continuous action stages in the firing process based on the changing characteristics of the aiming point trajectory sequences, gun body posture sequences, and trigger action sequences, and performs segmented elastic registration while maintaining the sequential order of the continuous action stages. Based on the registered stage characteristics, change directions, response sequence relationships, and cross-firing repeatability, a phase-preserving time-delay influence diagram is constructed. A closed-loop compensation chain is extracted from the phase-preserving time-delay influence diagram, and a stability evaluation result is generated. In this embodiment, the data interface is used to receive digital sequences already formed by the training acquisition device without changing the structure of the acquisition device. The processor associates each sequence and firing timestamp according to the same firing record identifier. The memory stores the original record, stage boundaries, registration mapping, graph connection attributes, closed-loop compensation chain, and final evaluation record, enabling each intermediate data to be traced along the same firing record.

[0019] Specifically, each firing record contains at least the position value of the aiming point in the two-dimensional reference plane as a function of time, the angle value of the gun body attitude as a function of time or the angular motion value derived from the angle value, the continuous quantity of the trigger action as a function of time, and the firing time stamp representing the moment of firing. The processor does not require each sequence to have the same start time and number of samples. Instead, it reads the time stamp carried by each sequence, converts the time stamp into a relative time value with the firing time stamp as the zero point, and obtains a unified time index for calculation only through piecewise interpolation. The original sequence is still retained in the order of reception, and an index mapping is maintained between the interpolated data and the original data so that the original observation position can be returned when the stage boundary or abnormal connection needs to be checked.

[0020] In this embodiment, the processor calculates adjacent position vectors, local turning angles, local curvatures, radial convergence, and tangential sway for the aiming point trajectory sequence; it calculates adjacent attitude differences, angular motion direction, angular motion energy, and energy transition for the gun body attitude sequence; and it calculates the action initiation changes, continuous changes, local stagnation, and change direction for the trigger action sequence. All features are used to form a time-aligned candidate feature table according to a unified time index. Each row in the candidate feature table corresponds to a relative time, and each column corresponds to a temporal feature, along with the original data source, interpolation status, and missing status. The processor extracts stage boundaries only within the same firing record to avoid directly splicing local changes from different firings into the same action process.

[0021] Preferably, the stage boundary evidence value is determined according to the following formula.

[0022] ; in, This represents the stage boundary evidence value at the unified time index position u. This represents the normalized change in local curvature at the aiming point. This represents the normalized change in the angular kinetic energy of the gun body. This represents the normalized change in trigger action. Indicates the position of the firing time stamp in the unified time index. Indicates the range of influence of the firing time marker. , , and These represent the non-negative weights of the four types of evidence, with the sum of the weights being 1. The symbols are... To represent exponentiation, the symbol is... To express absolute values, for example, when the aiming point transition and trigger action transition are both clear while the attitude energy transition is weak, it can be set to... , , , The two types of synchronous transitions at the corresponding positions will obtain higher boundary evidence, while positions that are only close to the firing time marker but lack multi-source changes will not be directly identified as stage boundaries.

[0023] Please refer to Figure 2 When the processor determines the continuous action stages in the firing process based on the changing characteristics of the aiming point trajectory sequence, the gun body attitude sequence, and the trigger action sequence, it calculates the local curvature change and convergence direction consistency for the aiming point trajectory sequence, the angular motion energy change for the gun body attitude sequence, and the action initiation change and continuous change for the trigger action sequence. It then forms a candidate boundary set for each change on a unified time reference and establishes a stage state chain in the order of entering stable aiming, stable holding, progressive pressing, firing disturbance, and post-firing holding. The candidate boundary set is mapped to the stage state chain, and the boundary position of each continuous action stage is selected based on the consistency of multi-source features between adjacent boundaries, the monotonicity of boundary timing, and firing time scale constraints.

[0024] Specifically, the processor selects candidate boundaries from the local extreme values ​​of the stage boundary evidence values, and merges multiple consecutive adjacent extreme values ​​caused by the same change event into a candidate interval. The position of the candidate interval is taken as the centroid of multi-source evidence. Entering the stable aiming stage requires that the radial convergence of the aiming point tends to become concentrated from discrete and that the angular motion energy of the gun body does not continue to rise. The stable holding stage requires that the convergence direction of the aiming point is in a continuous and stable state and that the trigger action has not formed a continuous change. The progressive pull stage requires that the trigger action shows a continuous unidirectional change and is adjacent to the local correction holding time of the aiming point. The firing disturbance stage takes the firing time scale as the core and covers the abrupt change interval before and after firing. The post-firing holding stage is composed of the continuous observation interval after the firing disturbance decays.

[0025] In a preferred embodiment, when a data channel has a short-term missing data, the processor does not determine the stage boundary separately through single-channel interpolation results. Instead, it marks the corresponding evidence as unavailable and renormalizes the weights of the remaining evidence. If the missing interval spans two candidate stages, the candidate boundaries on both sides of the interval are retained and the complete path is compared when solving the stage state chain. If the trigger action sequence is missing before the firing time marker, the starting boundary of the progressive pressing stage is jointly defined by the aiming convergence change and the attitude energy change. At the same time, the confidence information of the starting boundary is reduced to avoid interpreting the flat sequence caused by the missing data as a stable hold.

[0026] In this embodiment, the processor performs segmented elastic registration while maintaining the sequential order of continuous action stages. The start and end boundaries of each continuous action stage are used as fixed anchor points. Standardized stage time coordinates are established within each continuous action stage. The registration path within the stage is obtained based on the joint similarity of the changes in aiming point trajectory, gun attitude, and trigger action. The registration path within the stage satisfies the constraints of monotonically increasing time coordinates, continuous adjacent mapping points, and uncrossable stage boundaries. Multi-source sequences from the same firing share the stage-level scaling path. Each data channel only superimposes a restricted residual offset on the stage-level scaling path to obtain a registration timing sequence with consistent stage semantics.

[0027] Preferably, the registration path within a stage is determined by the following formula.

[0028] ; in, Indicates candidate registration path The total cost, This represents the mapping point between the m-th time index in the registration stage and the n-th time index in the standard stage template. This indicates differences in the trajectory characteristics of the aiming point. This indicates differences in the gun's posture characteristics. This indicates differences in trigger action characteristics. This represents the amount of local scaling change along the path. This represents the quantity that violates path continuity. , , , and All weights are non-negative and sum to 1. For example, in the progressive pressure stage, we can set... , , , , The difference in trigger action has a high constraint strength at this stage, while local stretching and path continuity together prevent the registration path from turning back or jumping in order to reduce single-point differences.

[0029] Specifically, the standard stage template is not a pre-fixed single firing curve, but is iteratively formed from firing records in the same evaluation batch whose boundary confidence information meets the usage conditions. The processor first linearly maps each stage to the standardized stage time coordinate, calculates the median sequence of multi-source features according to the corresponding coordinates as the initial template, and then obtains the intra-stage registration path of each firing record based on the initial template. After registration is completed, the median sequence is recalculated and the template is updated. The update stops when the corresponding feature changes of two adjacent templates no longer change the path order. Since the template update uses median aggregation, local mutations in a single firing will not directly determine the common stage morphology.

[0030] Furthermore, when the processor obtains the registration path within a stage, it decomposes the registration displacement of the multi-source sequence into a shared displacement that characterizes the overall rhythm change of the same continuous action stage and a channel residual displacement that characterizes the local response difference of a single data channel. The shared displacement is determined by joint similarity, and the channel residual displacements are determined by the local differences between the corresponding data channels and the shared stage template, respectively. It also satisfies the constraints of weighted sum of channel residual displacements being limited, continuous residual change rate, and not changing the order of the original channel responses. The shared displacement and channel residual displacements are combined to form the final mapping path, and the differences between multiple solutions of the mapping path are recorded as registration uncertainty information.

[0031] Specifically, the shared displacement is used to absorb the rhythm difference of a single firing being too fast or too slow in a certain continuous action phase. The channel residual displacement is used to preserve the local response interval that originally existed between aiming correction, attitude correction and trigger action. The processor first obtains the shared registration path based on joint similarity, and then calculates the local offset relative to the shared path for the three types of data channels respectively. If a certain local offset will cause the aiming correction, which originally occurred later than the attitude change, to occur before the attitude change, then the local offset is truncated to the nearest feasible position that maintains the order of response. If multiple channel residual displacements increase in the same direction in the same time interval, then their weighted common component is deducted from each residual displacement and incorporated into the shared displacement to prevent the overall rhythm change from being repeatedly recorded in multiple channels.

[0032] In a preferred embodiment, the processor retains alternative mapping paths that have the same stage anchor point as the minimum total cost path in a local interval but use different feasible mapping points. The alternative mapping paths should satisfy the following conditions: monotonically increasing, continuous adjacent mappings, uncrossable stage boundaries, and no reversal of response order. The processor writes the time offset, local direction change, and channel sequence difference of each alternative mapping path relative to the final mapping path into the registration uncertainty information. Instead of directly selecting any alternative mapping to replace the final mapping path, the processor checks whether the candidate time delay remains stable in response to changes in feasible mappings during subsequent directed connection calculations.

[0033] In this embodiment, please refer to Figure 3 The processor constructs a phase-order-preserving time delay influence map based on the registered stage features, change direction, response sequence relationship, and cross-shot repeatability. The aiming point change features, gun body attitude change features, and trigger action change features in each continuous action stage are set as nodes. For any two nodes, directed candidate time delays are calculated within the stage sequence range defined by the relationship between adjacent action stages. Directed connections are determined based on the consistency of change direction, response sequence relationship, cross-shot frequency, and candidate time delay dispersion of the candidate time delays. The connection direction, time delay interval, change sign, and cross-shot repeatability are written into the attributes of the directed connection. Directed connections with conflicting stage sequences or those appearing only in one firing sample are deleted.

[0034] Preferably, the candidate connection confidence value of node i pointing to node j is calculated according to the following formula.

[0035] ; in, Indicates candidate time delay The connection trust value from node i to node j is given. Indicates the consistency of the direction of change. Indicates the degree of continuity of responses. Indicates the repetition rate across firing cycles. Represents the candidate time-delay discrete quantity. , , and They are respectively represented as , , and The corresponding non-negative weights and The minus sign indicates that the time-delay discrete quantity deducts from the connection reliability value, for example, taking... , , , Even if a connection matches in direction during a single firing, its reliability will still be limited if the repetition rate across firings is low or the time delay is dispersed.

[0036] Specifically, the processor performs baseline removal processing on the node feature sequence within each continuous action phase. The aiming point change feature adopts the position and direction changes relative to the phase start point. The gun body attitude change feature adopts the attitude difference and angular motion direction relative to the phase start point. The trigger action change feature adopts the action increment and persistence mark relative to the phase start point. Candidate time delays are obtained by moving the subsequent node sequence within the allowed phase sequence range and calculating the local dependencies with direction signs. If the maximum dependency position is outside the allowed range, no connection is established. If multiple adjacent candidate time delays have similar dependencies, they are merged into a time delay interval instead of being forcibly compressed into a single time delay point.

[0037] In a preferred embodiment, when the processor determines a directed connection, it establishes a time delay sample set across firings for each directed candidate time delay, and divides the time delay subsets according to the continuous action stage, change sign, and response direction. For each time delay subset, it calculates the conditional dependency after controlling the historical features of other nodes, and compares the conditional dependency with the direct dependency when the historical features of other nodes are not controlled. When the direct dependency is explained by the historical features of the common predecessor node, or when the candidate time delay reverses direction in different firing samples, the corresponding directed connection is deleted. For the retained directed connections, the connection confidence is generated based on the conditional dependency, the direction retention ratio, and the degree of time delay concentration.

[0038] Specifically, when the gun attitude change node precedes both the aiming point change node and the trigger action node, and the aiming point change node and the trigger action node exhibit a direct dependency, the processor uses the historical characteristics of the gun attitude change node as a control variable to recalculate the residual dependency between the aiming point change node and the trigger action node. If the residual dependency is close to the background fluctuation within the same time-delay subset, it is determined that the direct dependency between the two is mainly explained by the common preceding node, and the corresponding connection is deleted. If the residual dependency remains stable and does not reverse in direction during multiple firings, the corresponding connection is retained, thereby avoiding the formation of multiple repeated propagation paths for the same preceding change in the graph.

[0039] Preferably, the conditional dependency explanatory power is determined according to the following formula.

[0040] ; in, This represents the dependency between node i and node j explained by the set of control variables H. This represents the direct dependency when the historical characteristics of other nodes are not controlled. This represents the conditional dependency retained after the set of control variables H participates in the conditional calculation, for example, when... and hour, This indicates that most direct dependencies can be explained by common historical features, and the processor removes candidate connections accordingly. and At that time, candidate connections are retained and participate in the connection confidence calculation as conditional dependencies.

[0041] Please refer to Figure 4 , Figure 4 The horizontal axis represents the directed connections to be tested, numbered E1 to E4 in order; the vertical axis represents the dependency, and the bar charts show the direct dependency, conditional dependency, and dependency explained by the historical features of other nodes. E1 has a direct dependency of approximately 0.68, a conditional dependency of approximately 0.54, and an explained value of approximately 0.14, indicating that this connection retains a strong residual dependency after controlling for common predecessor nodes and can continue to be considered an effective connection in the phase-order-preserving lag influence graph. E2 and E3 have direct dependencies of approximately 0.60 and 0.52, and conditional dependencies of approximately 0.50 and 0.46, respectively, with relatively small explained values, indicating that their direct association is not primarily caused by common historical features. E4 has a direct dependency of approximately 0.44, a conditional dependency of approximately 0.24, and an explained value of approximately 0.20, the highest explained value among the four connections, indicating that this connection is more significantly influenced by common predecessor nodes, but still retains some conditional dependency after control, therefore it was not deleted in the current sample. Figure 4 The screening function of the corresponding conditional dependency interpretation formula: By comparing the dependency changes before and after control, the processor distinguishes between the true response relationship and the pseudo-connection caused by the common preceding changes, thus avoiding the repeated propagation of the same gun body attitude change in the graph structure.

[0042] In this embodiment, when the processor extracts the closed-loop compensation chain from the phase-preserving time delay influence map, it takes the node corresponding to the gun body attitude change feature as the starting node, searches along the phase-preserving time delay influence map for directed paths that sequentially pass through the node corresponding to the aiming point correction feature, the node corresponding to the trigger continuous action feature, and reach the firing event node. For each directed path, it checks whether the time delay intervals of each directed connection are continuously connected according to the order of the continuous action stages. Based on the correspondence between the gun body attitude change direction and the aiming point correction direction, and the order of the aiming point deviation change and the trigger continuous action, it filters out paths that do not meet the compensation direction constraints. The retained paths with a common starting node and a common firing event node are merged into a closed-loop compensation chain.

[0043] Specifically, the change in gun attitude in the closed-loop compensation chain is not based solely on the magnitude of the change in attitude, but is paired with the direction of subsequent aiming point correction. When the change in gun attitude causes the aiming point to deviate from the stage reference center, the subsequent aiming point correction should be later in phase than the attitude change and oriented towards the reference center. The continuous trigger action should occur after the aiming correction enters the convergence interval and should not interrupt the correction process. The firing event node should be within the subsequent time delay range of the continuous trigger action. The processor considers directed paths that meet the above timing and direction conditions as reducible paths and excludes paths that only rely on the numerical cancellation at the same time point but lack a sequential response relationship.

[0044] Furthermore, the processor merges multiple retained paths between the same starting node and the same firing event node according to shared nodes and shared directed connections. When two paths use different aiming correction nodes in the stable holding phase but merge into the same trigger continuous action node in the progressive pressing phase, the processor retains the two branches and their connection attributes and establishes a common chain identifier. When two paths form a duplicate connection only because adjacent candidate time delay intervals are split, the processor merges the connected time delay intervals and recalculates the cross-firing repeatability, so that the closed-loop compensation chain can express the alternative local corrections in the same compensation process without repeatedly accumulating the common part.

[0045] In a preferred embodiment, when the processor selects the boundary position of each continuous action stage, it establishes a boundary evidence vector for each candidate boundary, which includes the aiming point trajectory turning evidence, the gun body attitude energy turning evidence, the trigger action turning evidence, and the firing time scale distance. It also constructs a directed boundary path on the stage state chain. The adjacent nodes of the directed boundary path correspond to the adjacent stage boundaries respectively. The path cost consists of the feature stability within the stage, the feature difference between adjacent stages, the boundary deviation across data channels, and the boundary order violation term. The stage boundary is determined by finding the complete boundary path with the minimum path cost, and boundary confidence information is generated based on the path cost difference between the complete boundary path and the adjacent candidate path.

[0046] Preferably, the path cost of the complete boundary path is calculated according to the following formula.

[0047] ; in, Represents the complete boundary path Path cost, This represents the cumulative amount by which the internal characteristics of each stage deviate from the state of the corresponding stage. This represents the cumulative amount of insufficient characteristic differences between adjacent stages. This represents the cumulative amount of positioning deviation at the same boundary from different data channels. This indicates a violation of the phase boundary order. , , and They are respectively represented as , , and The corresponding non-negative weights sum to 1, for example, taking , , , At this time, any stage inversion will increase the path cost significantly, while when there is a small range of deviation in the multi-source boundary, the feasible boundary combination can still be determined by comparing the entire path as a whole.

[0048] Please refer to Figure 5 , Figure 5 The horizontal axis represents the candidate boundary paths arranged in ascending order of total cost, from path 1 to path 5; the vertical axis represents the cost of the boundary path. The bar chart uses different filled squares to distinguish the total cost of the path and the difference between the path and the best value. The total cost of path 1 is approximately 0.361, making it the lowest-cost path among the candidate boundary combinations in this batch. Its boundary time combinations are -1.270 seconds, -0.640 seconds, -0.160 seconds, and 0.310 seconds, corresponding to the complete boundary path ultimately selected by the processor. The total costs of paths 2 to 4 are approximately 0.363, 0.365, and 0.368, respectively, and the differences from the optimal path are all less than 0.007, indicating that these candidate boundary combinations are relatively close to the optimal path in local stage positions and can be used as neighboring paths in boundary confidence calculation. The total cost of path 5 is approximately 0.403, and the difference between the path and the best value is approximately 0.042, which is significantly higher than the previous four paths. This is mainly because the combined constraints of the candidate boundary order and the stability of the internal characteristics of the stage are further violated. Figure 5 This demonstrates the purpose of the complete boundary path formula: instead of selecting each local extremum independently, the processor compares the cumulative cost of the entire boundary path in the stage state chain and transforms the cost interval between the optimal path and neighboring paths into boundary confidence information.

[0049] Specifically, the boundary confidence information is composed of the difference between the minimum path cost and the costs of several adjacent candidate paths, the multi-source consistency of the candidate boundary evidence vector, and the stability of the boundary position in the candidate paths. The processor does not directly equate the boundary confidence information with the existence of a stage, but rather attaches the boundary confidence information to the corresponding stage features and subsequent directed connections. When the boundary of a certain stage moves significantly in multiple approximate paths, the candidate time-delay connections generated within the stage enter the uncertainty test process to avoid multiple solutions of the boundary position being treated as a definite time sequence in subsequent graph calculations.

[0050] Furthermore, when the processor propagates the registration uncertainty information to the phase-order-preserving delay influence map, for any directed connection, it re-determines the candidate delays using the final mapping path and alternative mapping paths that have multiple solutions to the final mapping path, forming a candidate delay offset set. Based on the candidate delay offset set, it determines the timing order preservation ratio, connection direction preservation ratio, and candidate delay interval overlap ratio of the nodes at both ends of the directed connection. The timing order preservation ratio, connection direction preservation ratio, and candidate delay interval overlap ratio are combined into a connection robustness coefficient, which is used to correct the properties of the corresponding directed connection in the phase-order-preserving delay influence map.

[0051] Preferably, the connection robustness coefficient is calculated according to the following formula.

[0052] ; in, This represents the robustness coefficient of the connection from node i to node j. This indicates the proportion of node timing order maintained when using the final mapping path and alternative mapping paths. This indicates that the connection direction remains proportional. Indicates the overlap ratio of candidate time delay intervals. , and They are respectively represented as , and The corresponding non-negative weights sum to 1, for example, taking , , At that time, the stability of the timing sequence constitutes the main constraint. If the timing sequence of a connection is reversed under different feasible registrations, even if the direction is maintained and the intervals overlap locally, the connection robustness coefficient will still decrease and limit its entry into the closed-loop compensation chain.

[0053] Please refer to Figure 6 , Figure 6 The horizontal axis represents directed connections E1 to E4; the vertical axis represents proportions or coefficients. The bar charts show order preservation, orientation preservation, time delay overlap, and robustness coefficients, respectively. The order preservation and orientation preservation proportions for E1 to E3 are both approximately 0.94, indicating that the response order and direction of change for these three connections are relatively stable under the final mapping path and feasible alternative mapping paths. E3 has the highest time delay overlap proportion, approximately 0.506, indicating a large common coverage of candidate time delay intervals under different registration paths. E4 has the lowest time delay overlap proportion, approximately 0.056, while the order preservation and orientation preservation proportions drop to approximately 0.86, resulting in a robustness coefficient of approximately 0.725, the lowest among the four connections. E1 has a robustness coefficient of approximately 0.843, the highest in the current batch, reflecting its ability to maintain relatively reliable connection properties even under the influence of multiple registration solutions. Figure 6The connection robustness coefficient is used to propagate registration uncertainty information to the phase-order-preserving time-delay effect diagram: only when the node order, connection direction, and time-delay interval overlap remain stable can the corresponding connection be suitable for entering the main path of the closed-loop compensation chain.

[0054] Specifically, when the processor uses connection robustness coefficients to correct directed connection attributes, it retains the original connection direction, original change sign, original time delay interval, and original cross-shot repetition, and adds a robustness mark and correction confidence. The correction confidence is obtained by combining the original connection confidence and the connection robustness coefficient. If there are candidate connections under the final mapping path but the connections disappear under most alternative mapping paths, the connection is marked as a registration sensitive connection and does not participate in the main path merging of the closed-loop compensation chain. If the connection maintains the direction and timing order under each feasible mapping but the time delay interval is shifted, the connection is retained and the common coverage of each time delay interval is used as the main time delay interval.

[0055] In this embodiment, when the processor generates the stability assessment result, it extracts the stage coverage ratio of the closed-loop compensation chain, the cross-fire repetition of each directed connection, the time delay connection deviation of adjacent directed connections, the compensation direction maintenance ratio, and the cross-fire occurrence ratio of the closed-loop compensation chain. It then constructs a chain-level correlation matrix according to the relationship between shared nodes and shared directed connections between different closed-loop compensation chains. Based on the chain-level correlation matrix, it merges the feature weights corresponding to shared nodes and shared directed connections to form a stability feature vector composed of closed-loop integrity, time delay stability, stage connection continuity, and compensation repetition. Finally, it generates the stability assessment result based on the stability feature vector.

[0056] Preferably, the overall stability value is determined according to the following formula.

[0057] ; in, This represents the overall stability value. Indicates the completeness of the closed loop. Indicates time-delay stability. Indicates the continuity of stage transitions. Indicates compensation for repetition. This represents the uncertainty formed by the combination of boundary confidence information, registration uncertainty information, and registration sensitive connectivity. , , , and They are respectively represented as , , , and The corresponding non-negative weights sum to 1, for example, taking , , , , At that time, whether the closed loop is complete and whether it is repeated across times has a high constraint strength, while samples with uncertain evidence will not obtain the same comprehensive value as stable closed loop samples even if the local fluctuations are small.

[0058] Specifically, the stage coverage ratio is calculated by corresponding the number of continuous action stages actually traversed by the closed-loop compensation chain to the number of available stages in the preset stage state chain. The time delay connection deviation is calculated by the distance between the previous response endpoint and the next response start point of adjacent directed connections. The compensation direction maintenance ratio is calculated by the correspondence between the attitude offset direction and the aiming correction direction, the aiming deviation change direction and the trigger continuous action sequence. The proportion of the closed-loop compensation chain appearing across times is aggregated according to firing records with the same node type sequence and compatible time delay interval. The processor identifies shared nodes and shared connections in the chain-level association matrix to avoid the same attitude change or the same trigger continuous action being repeatedly included in the stability feature vector due to participation in multiple local branches.

[0059] In a preferred embodiment, the stability assessment results include a comprehensive stability value, a stability feature vector, the identifier of the main closed-loop compensation chain, the identifier of the registration-sensitive connection, and the source of uncertainty. The processor converts the comprehensive stability value into a discrete stability level according to a pre-saved hierarchical mapping relationship. The hierarchical mapping relationship only encodes the range of comprehensive values ​​without changing the stability feature vector. If two firing batches have the same stability level but different sources of uncertainty, the system retains records of boundary uncertainty, registration uncertainty, or connection uncertainty respectively, so that subsequent verification can locate the stage boundary, mapping path, or graph connection level.

[0060] Furthermore, during the evaluation of multiple consecutive firings, the processor uses the single firing record as the basic calculation object to form a phase-order-preserving time-delay influence graph. Then, based on the node type, stage position, connection direction, and compatible time-delay interval, the graph structure in multiple firings is matched. The processor does not directly calculate the average of the original time series. Only when the corresponding connection meets the cross-firing repetition condition is it written into the batch-level closed-loop compensation chain. Occasionally occurring single low-fluctuation trajectories will not independently constitute the basis for batch-level stability. When there is a local disturbance in a single firing but the closed-loop compensation chain still maintains the stage order and direction relationship, its connection attributes participate in the batch-level calculation according to the actual time-delay dispersion.

[0061] Specifically, when the aiming point trajectory exhibits relatively small dispersion within a fixed time window and the trigger action begins earlier than the aiming correction convergence time, the processor classifies the trigger action into the progressive pull stage according to the stage state chain. It also checks the response sequence from the aiming correction node to the trigger continuous action node in the phase-sequence-preserving delay influence diagram. If the connection direction is incompatible with the stage sequence, the data does not form a complete closed-loop compensation chain. When the aiming correction pointing to the reference center occurs after a stable time delay following a change in the gun's attitude and the trigger continuous action is after the correction convergence, even if the changes in each channel do not occur synchronously, the processor still forms a closed-loop compensation chain according to the delay compensation relationship.

[0062] In a preferred embodiment, when multiple data channels exhibit opposite directions and cancel each other out within the same fixed window, the processor does not use the composite amplitude after cancellation as the primary judgment value. Instead, it retains the aiming point change node, the gun body attitude change node, and the trigger action node respectively. Directed connections are established based on the change sign, stage position, and candidate time delay of each node. If the opposite changes lack a stable sequential relationship or a cross-repetition relationship, the corresponding directed connection is deleted in the direction maintenance test or repeatability test. If the opposite changes constitute a continuous relationship of aiming back to center after attitude shift, the compensation direction attribute is written into the closed-loop compensation chain.

[0063] Preferably, the processor associates and stores the original data, stage features, registration timing, phase-preserving delay effect diagram, and stability assessment results using the same firing record identifier and the same stage identifier. When adjusting the stage boundary, only the registration path and directed connection of the adjusted stage and its adjacent stages are recalculated. When a connection is deleted or merged, the corresponding closed-loop compensation chain and chain-level correlation matrix are updated synchronously. The original data remains read-only, and all recalculated results retain the source mapping, so that the stability results can be traced back from the closed-loop compensation chain to the directed connection, from the directed connection to the registration feature, and from the registration feature to the original time index.

[0064] In this embodiment, the system maintains firing references for multi-source records through a unified time base, eliminates semantic misalignment of actions within a fixed window through continuous action stage constraints, preserves overall rhythm changes and local response delays through shared displacement and channel residual displacement within stages, expresses the direction of change, response sequence, and cross-repetition relationship through a phase-preserving delay influence diagram, eliminates pseudo-connections caused by common preceding features through conditional dependencies, restricts the transmission of connection attributes to multiple solutions through registration uncertainty information, and forms a stability feature vector by a closed-loop compensation chain and a chain-level correlation matrix. The entire data processing chain is completed based on the received time-series data and does not involve changes to the air pistol structure.

Claims

1. A stability evaluation system for air pistols based on temporal feature extraction, characterized in that, It includes a data interface, a memory, and a processor. The data interface receives aiming point trajectory sequence, gun body posture sequence, trigger action sequence, and firing time marker. The memory stores programs that can be executed by the processor. When the processor executes the program, it establishes a unified time reference based on the firing time marker. The continuous action phases in the firing process are determined based on the changing characteristics of the aiming point trajectory sequence, the gun body posture sequence, and the trigger action sequence. Segmented flexible registration is performed while maintaining the sequential order of the continuous action phases; A phase-order-preserving time delay influence diagram is constructed based on the stage characteristics, change direction, response sequence relationship and cross-firing repeatability after registration. The closed-loop compensation chain is extracted from the phase-sequence-preserving delay effect diagram, and the stability assessment result is generated.

2. The air pistol stability evaluation system based on temporal feature extraction according to claim 1, characterized in that, When the processor determines the continuous action phase, it calculates the local curvature change and convergence direction consistency of the aiming point trajectory sequence, calculates the angular motion energy change of the gun body posture sequence, calculates the action start change and continuous change of the trigger action sequence, and forms a candidate boundary set of each change on a unified time reference. Establish a stage state chain in the order of entering stable aiming, stable holding, progressive pressing, firing disturbance and post-firing holding, map the candidate boundary set to the stage state chain, and select the boundary position of each continuous action stage based on the consistency of multi-source features between adjacent boundaries, boundary temporal monotonicity and firing time scale constraints.

3. The air pistol stability evaluation system based on temporal feature extraction according to claim 1, characterized in that, When the processor performs segmented elastic registration, it uses the start and end boundaries of each continuous action stage as fixed anchor points, establishes standardized stage time coordinates within each continuous action stage, and obtains the registration path within the stage based on the joint similarity of the changes in aiming point trajectory, gun body posture, and trigger action. The registration path within the stage satisfies the constraints of monotonically increasing time coordinates, continuous adjacent mapping points, and uncrossable stage boundaries. Multi-source sequences fired in the same instance share the stage-level scaling path, and each data channel only superimposes restricted residual offsets on the stage-level scaling path to obtain a registration timing sequence with consistent stage semantics.

4. The air pistol stability evaluation system based on temporal feature extraction according to claim 1, characterized in that, When the processor constructs the phase-preserving time delay influence map, it sets the aiming point change characteristics, gun body attitude change characteristics, and trigger action change characteristics in each continuous action phase as nodes respectively. For any two nodes, calculate the directed candidate time delay within the stage sequence range defined by the relationship between adjacent action stages, and determine the directed connection based on the consistency of the change direction corresponding to the candidate time delay, the relationship between the previous and subsequent responses, the frequency of cross-shot occurrence, and the dispersion of the candidate time delay. Write the connection direction, time delay interval, change sign, and cross-fire repetition into the attributes of the directed connection, and delete directed connections that conflict with the phase order or appear only in one firing sample.

5. The air pistol stability evaluation system based on temporal feature extraction according to claim 2, characterized in that, When the processor selects the boundary position of each continuous action stage, it establishes a boundary evidence vector for each candidate boundary, which includes the aiming point trajectory turning evidence, the gun body attitude energy turning evidence, the trigger action turning evidence, and the firing time marker distance, and constructs a directed boundary path on the stage state chain. The adjacent nodes of the directed boundary path correspond to the adjacent stage boundaries respectively. The path cost consists of the intra-stage feature stability, the adjacent stage feature difference, the cross-data channel boundary deviation, and the boundary order violation term. The stage boundary is determined by finding the complete boundary path with the minimum path cost, and boundary confidence information is generated based on the path cost difference between the complete boundary path and the adjacent candidate paths.

6. The air pistol stability evaluation system based on temporal feature extraction according to claim 3, characterized in that, When the processor calculates the registration path within the stage, it decomposes the registration displacement of the multi-source sequence into a shared displacement that characterizes the overall rhythm change of the same continuous action stage and a channel residual displacement that characterizes the local response difference of a single data channel. The shared displacement is determined by joint similarity, and the channel residual displacements are determined by the local differences between the corresponding data channels and the shared stage template, satisfying the constraints of weighted sum of residual displacements of each channel, continuous residual change rate, and no change in the order of original channel responses. The shared displacement and the channel residual displacement are combined to form the final mapping path, and the differences between multiple solutions of the mapping path are recorded as registration uncertainty information.

7. The air pistol stability evaluation system based on temporal feature extraction according to claim 4, characterized in that, When the processor determines the directed connection, it establishes a time delay sample set across firing cycles for each directed candidate time delay, and divides the time delay subset according to the continuous action stage, change sign and response direction; For each time-delay subset, calculate the conditional dependency after controlling the historical features of other nodes, and compare the conditional dependency with the direct dependency when the historical features of other nodes are not controlled. When the direct dependency is explained by the historical features of the common predecessor node, or when the candidate delay reverses direction in different firing samples, the corresponding directed connection is deleted. For the retained directed connections, connection confidence is generated based on the conditional dependency, the proportion of direction retention, and the degree of time delay concentration.

8. The air pistol stability evaluation system based on temporal feature extraction according to claim 4, characterized in that, When the processor extracts the closed-loop compensation chain, it takes the node corresponding to the gun body attitude change feature as the starting node, searches along the phase-sequence-preserving time delay influence map, sequentially passes through the node corresponding to the aiming point correction feature, the node corresponding to the trigger continuous action feature, and reaches the firing event node in a directed path. For each directed path, the time delay intervals of each directed connection are checked for continuity according to the order of the continuous action phases. Paths that do not meet the compensation direction constraints are filtered out based on the correspondence between the direction of change of gun body attitude and the direction of correction of aiming point, and the order of change of aiming point deviation and continuous action of trigger. The preserved paths that share a common starting node and a common firing event node are merged into the closed-loop compensation chain.

9. The air pistol stability evaluation system based on temporal feature extraction according to claim 6, characterized in that, When the processor propagates the registration uncertainty information to the phase-order-preserving time delay influence map, for any directed connection, it re-determines the candidate time delay using the final mapping path and the alternative mapping path that has multiple solution differences with the final mapping path, forming a candidate time delay offset set. The timing order preservation ratio, connection direction preservation ratio, and candidate delay interval overlap ratio of the two nodes of the directed connection are determined based on the candidate delay offset set, and the timing order preservation ratio, connection direction preservation ratio, and candidate delay interval overlap ratio are combined into a connection robustness coefficient. The properties of the corresponding directed connection in the phase-order-preserving delay effect graph are corrected using the connection robustness coefficient.

10. The air pistol stability evaluation system based on temporal feature extraction according to claim 1, characterized in that, When the processor generates the stability evaluation result, it extracts the phase coverage ratio of the closed-loop compensation chain, the cross-fire repetition of each directed connection, the time delay connection deviation of adjacent directed connections, the compensation direction maintenance ratio, and the cross-occurrence ratio of the closed-loop compensation chain, and constructs a chain-level association matrix according to the relationship between shared nodes and shared directed connections between different closed-loop compensation chains. The feature weights corresponding to the shared nodes and the shared directed connections are merged according to the chain-level association matrix to form a stability feature vector composed of closed-loop integrity, time-delay stability, stage connection continuity and compensation redundancy. The stability evaluation result is then generated based on the stability feature vector.

Images