Behavior-driven twinborn prediction method

Abstract

The invention discloses a behavior-driven twinborn prediction method, and relates to the technical field of intelligent information processing and prediction modeling, and the method comprises the following steps: building a unified event time baseline, collecting nanosecond clock offset information of each data source, building a cross time sequence suspicion chart, recognizing time synchronization abnormal nodes, and forming a credible time anchor point set; and based on the trusted time anchor point set, executing anti-fact playback, reconstructing a historical evolution process of behavior data, generating a time offset vector set, and constructing a corresponding time sequence offset spectrum. According to the method, through construction of a unified time baseline, trusted anchor points, anti-fact replay, causal topology and time reversal control, time sequence dislocation identification, calibration and false trajectory elimination of multi-source behavior data are realized, a dynamic self-healing closed-loop prediction mechanism is established, and the twin system prediction accuracy and stability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent information processing and prediction modeling, in particular to a behavior-driven twin prediction method. BACKGROUND

[0002] "Behavior-driven twin prediction" refers to taking the behavior of an entity (person / device / group) as a key input to drive the state evolution and future inference of the digital twin: the system continuously collects behavior data such as operation instructions, interaction trajectories, strategy selection, and time-series events, extracts features and builds causal models, establishes a "behavior-mechanism-state-result" mapping, and makes the twin evolve in the simulation space according to the real decision logic, thereby predicting performance changes, demand fluctuations, abnormal risks, and resource occupation in advance and generating executable control recommendations. Unlike predictions based only on sensor states, it emphasizes the decisive impact of behavior triggers and strategy changes on the future of the system, combines multi-source data fusion (logs / sensors / locations / transactions), online calibration, and closed-loop feedback, so that the model can adaptively update as the behavior pattern migrates, and is suitable for scenarios such as life differences caused by operation strategies in device maintenance, congestion evolution caused by driving styles in traffic, and sales curve remodeling caused by user path changes in retail.

[0003] The prior art has the following disadvantages: In the prior art, the dynamic fusion of multi-source behavior data usually relies on a unified timestamp to achieve synchronization, but due to differences in sampling accuracy and transmission delay between different data sources, it is easy to cause cross-time sequence misalignment. Once misalignment occurs, different sources of data segments will produce unrealistic overlapping trajectories during the fusion process, thereby generating non-existent "virtual paths" in twin prediction. For example, in the traffic management scenario, misalignment between vehicle driving behavior data and road sensor detection data will form a false driving path, and the dispatching system may mistakenly determine that the road will be congested, thereby triggering an incorrect traffic diversion instruction, leading to actual road congestion or even traffic accidents. In the industrial device maintenance scenario, misalignment between operation log data and sensor monitoring data may generate non-existent downtime signs for the twin, and the dispatching system may trigger unnecessary downtime warnings or even forced shutdowns, severely affecting production efficiency and device safety.

[0004] The above information disclosed in the background section is only intended to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present application is to provide a behavior-driven twin prediction method to solve the problems in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a behavior-driven twin prediction method, comprising the following steps: Establish a unified event time baseline, collect nanosecond-level clock offset information from various data sources, construct a cross-time sequence suspicion diagram, identify time synchronization anomaly nodes, and form a set of reliable time anchor points; Based on a set of trusted time anchors, counterfactual replay is performed to reconstruct the historical evolution of behavioral data, generate a set of time offset vectors, and construct the corresponding time-series offset spectrum. Based on the time offset spectrum, cross-source phase traction is explored. Combined with residual consistency and phase-locked stability indicators, time misalignment boundary nodes are identified, and segment alignment coordinates are set. Based on segmented aligned coordinates, a multi-sampling-rate causal topology graph is constructed to reconstruct the mapping path from behavioral data to entity state, and non-consistent regions in the trajectory are marked to establish rejection regions. Based on the rejection region, adaptive time calibration is performed. The time offset vector set is used to elastically align multi-source behavioral data, and the correction results are written into the calibration matrix to form a cyclic correction mechanism. Based on the cyclic correction mechanism, time-reversal control commands are generated to drive the polarization metasurface to inject conjugate frame envelopes, construct energy suppression channels in the neighborhood of zero-phase anchor points to eliminate spurious trajectories, and update the weights of the causal topology graph to achieve self-healing and closed-loop regulation of the predicted path.

[0007] The preferred process for forming a set of reliable time anchors is as follows: A high-precision master clock source is selected as the synchronization reference. The time offset of each data source is collected by bidirectional round-trip measurement to construct a unified reference time mapping matrix and map the timestamps of all behavioral data to a unified event time baseline.

[0008] Under a unified time baseline, multi-source behavioral events are rearranged according to the mapped timestamp, the event density is statistically analyzed using a sliding time window, a two-dimensional time-source distribution map is plotted, a cross-time series suspect map is constructed, and areas with abnormal density are identified.

[0009] Based on the density anomaly regions in the suspect map, the event temporal density, spatial location, and behavioral semantics are extracted. Clustering is used to identify synchronous mutation behavior and determine the time synchronization anomaly nodes in each data source.

[0010] Events with high confidence, clear time location, and strong behavioral representativeness are selected from abnormal nodes to construct a set of credible time anchors, which are then combined to form a stable anchor group, providing a time series reference for behavioral paths.

[0011] Preferably, the time-off spectrum construction steps are as follows: Select time segments with large time spans and fluctuating event density from the anchor point set, and construct a behavior replay sequence that includes event time, behavior description, entity identifier, spatial location and source number; Under a unified time baseline, the behavior replay sequence is advanced at the millisecond level to determine whether there are spatial conflicts, logical breaks or causal anomalies between events within each time window, and to generate a set of counterfactual behavior trajectories. Identify anomalous event pairs from counterfactual behavior trajectories, extract event time, behavior sequence and offset direction, calculate time offset values ​​between various data sources, and construct a set of time offset vectors; The time offset vector is statistically analyzed according to the time window, and multi-scale time domain analysis is performed to generate a time-series offset spectrum, which is used as the guiding basis for subsequent time alignment and path correction.

[0012] Preferably, the steps for setting segment alignment coordinates are as follows: Based on the time offset spectrum, a misalignment concentration section is selected, and a cross-source phase traction probe is performed. By shifting the time axis of the behavioral event, a cross-source time response curve is formed. Based on the traction, the residual consistency index is calculated, the trend of event time difference changes is analyzed, the synchronization accuracy of different behavior types is evaluated, and the effectiveness of phase traction is determined. Perform phase-locked stability analysis, statistically analyze the direction and magnitude of the change in the phase difference of events within consecutive sub-windows, and identify time periods with continuous synchronization capabilities; Based on the residual consistency index and phase-locked stability analysis results, time jump points and behavioral change points are identified, and timestamps, behavioral differences and residual jump values ​​are extracted to form a set of misalignment boundary nodes. An alignment transition zone is constructed with the misalignment boundary node as the center. The behavior event time translation operation is performed according to the offset difference to establish a segmented alignment coordinate structure between the time axis and the behavior sequence.

[0013] Preferably, the steps for establishing the rejection region are as follows: Based on the segmented alignment coordinates, a unified sampling mapping is performed on the multi-source behavioral data to construct a set of behavioral events with multiple sampling rates and rearrange them onto a unified timeline; A causal topology graph is constructed based on a set of behavioral events. Causal connections between behavioral events are established through temporal sequence, spatial continuity, and changes in entity state, generating a behavioral path structure with directionality and weight. Identify trajectory abrupt changes, semantic conflicts, and areas of reduced confidence in the behavioral path to form a set of trajectory inconsistency segments; Based on the set of trajectory inconsistency segments, a trajectory rejection region is set, conflicting behavioral events and causal edges are removed, and a high-confidence virtual transition path is constructed to maintain the integrity of the causal structure.

[0014] Preferably, the cyclic correction mechanism is formed as follows: Based on the marked behavioral conflict segments in the trajectory rejection area, and combined with the time offset vector set, the timestamps of behavioral events are flexibly aligned to generate a corrected behavioral data stream. Write the corrected time results into the time series calibration matrix, record the behavior event number, original timestamp, calibrated timestamp, offset adjustment value and adjustment source, and mark the data source with continuous offset as the drift channel; During the behavior prediction process, the time-series calibration matrix is ​​invoked to perform dynamic matching on new events, triggering real-time calibration operations and constructing a cyclic correction mechanism to achieve closed-loop updates and continuous stability of behavior time calibration.

[0015] Preferably, the operation of writing the correction result into the time-series calibration matrix further includes performing drift channel marking on data sources that have a high proportion of the same offset direction within multiple consecutive calibration windows, and forcibly calling the corresponding correction record for all new events in the drift channel in the subsequent behavior prediction to complete the pre-calibration.

[0016] Preferably, based on the cyclic correction mechanism, time-reversal control commands are generated to drive the polarization metasurface to inject conjugate frame envelopes, construct energy suppression channels in the neighborhood of zero-phase anchor points to eliminate spurious trajectories, and simultaneously update the causal topology graph weights. The steps are as follows: Based on the persistent drift path and the residual erroneous behavior segments in the trajectory rejection region, a time-reversal phase control instruction is constructed, and the activation time, behavior entity, semantic duality and execution intensity in the reverse behavior path are defined. The programmable polarization metasurface is driven to inject a modified frame envelope conjugate with the original trajectory, and an energy suppression channel with the opposite direction is generated in the neighborhood of the zero-phase time anchor point to perform pulse-level trajectory elimination operation; Freeze the topological edge weights corresponding to the original behavior path, add a new corrected behavior path and update the edge weights to complete the self-healing of the causal topological graph structure, and ensure that the closed-loop control of the subsequent prediction path is consistent with the logic.

[0017] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention addresses the timing drift problem caused by differences in sampling accuracy and transmission delays between different data sources by establishing a unified event time baseline and a set of reliable time anchors. Through counterfactual replay and phase traction mechanisms, it captures and quantifies potential misalignment regions in behavioral segments. Furthermore, by constructing a causal topology graph and rejection regions, it effectively identifies and masks false trajectories, achieving the discrimination and isolation of abnormal behavioral paths. Building upon this, it introduces elastic time calibration and correction write-back to construct an online cyclic correction mechanism, enabling the system to have adaptive calibration capabilities. Finally, through time inversion control and conjugate frame injection, it eliminates interfering behavioral trajectories at the physical level and dynamically updates the weights of the causal structure graph, achieving real-time self-healing and closed-loop feedback of the system's predicted paths. Overall, this method breaks through the traditional time synchronization methods centered on static alignment, realizing a dynamic fusion prediction mechanism driven by behavioral causality and aiming for full-time closed-loop operation. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a flowchart of the behavior-driven twin prediction method of the present invention. Detailed Implementation

[0020] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0021] This invention provides, for example Figure 1 The behavior-driven twin prediction method shown includes the following steps: Establish a unified event time baseline, and collect nanosecond-level clock offset information from multiple data sources based on the unified event time baseline. Construct a cross-time series suspicion map, extract time synchronization anomaly nodes based on the cross-time series suspicion map, determine a set of credible time anchor points, and form a stable anchor point group for anomaly evolution reference. To address the temporal misalignment issue caused by inconsistent sampling accuracy and transmission delays during the fusion of multi-source behavioral data, it is necessary to construct a high-precision unified event time baseline and establish a stable set of time anchor points from this baseline for reference in anomaly evolution. The specific implementation steps are as follows: A unified event time baseline is established. Specifically, a reference clock source with extremely high stability and time-keeping accuracy is selected as the benchmark for time synchronization of all behavioral data. This reference clock can be the master clock integrating a temperature-compensated crystal oscillator and a GPS timing device, with a frequency accuracy of less than 50 nanoseconds per day and a timing resolution within 10 nanoseconds. To collect actual time offset information from external data sources, a two-way round-trip measurement method is used to actively perform request-response interactions for time synchronization signals with all data sources. The transmission time under the local reference clock is recorded when the request signal is sent, and the return time is recorded after receiving the response signal, and compared with the timestamp of the source data device carried in the response. For each measurement, the signal propagation delay, response generation delay, and intermediate communication link delay are recorded. Through multiple rounds of sampling, the average offset value and offset stability of each data source are statistically analyzed. Finally, based on the offset measurement results of all data sources, a time mapping matrix based on the unified reference clock is formed, mapping the timestamps of all original behavioral data to the unified event time baseline.

[0022] A cross-temporal sequence suspicion map is constructed to identify potential time misalignment locations. Under a unified time baseline, behavioral events from all data sources are rearranged along the timeline according to their mapped timestamps, forming a cross-source behavioral event time series. A two-dimensional distribution map is plotted with time as the horizontal axis and different data sources as the vertical axis, where each behavioral event corresponds to a discrete point on both the horizontal and vertical axes. Subsequently, a window scan is performed on the multi-source events within each time period, with the sliding time window width set to 5 milliseconds and moving 2 milliseconds at a time. For each window, the number of events appearing within that time range and their sources are counted. If the number of events in a window is significantly higher than the statistical expectation of the source data for that time period, and there is no obvious causal relationship between events from different sources in terms of spatial location, content attributes, and operation objects, then that window is marked as a suspected misalignment area. Time periods with consecutive abnormal densities are extracted from all scanned windows as cross-temporal sequence abnormal areas. Finally, all abnormal areas are summarized to form a cross-temporal sequence suspicion map for subsequent identification of synchronization abnormal nodes.

[0023] Extracting time synchronization anomaly nodes from a cross-temporal sequence suspect graph. First, event feature clustering is performed on each anomaly region. The temporal density distribution, event description keywords, object identifiers, and physical spatial location markers of events within the anomaly region are extracted to construct a five-dimensional feature vector to measure the differences between events. High-density event groups are identified as preliminary aggregation regions using a density peak clustering algorithm. Subsequently, the temporal span and source heterogeneity of each aggregation region are calculated. If multiple data sources simultaneously exhibit abrupt behavioral signals within an aggregation region, and the signal time interval is less than 1 millisecond, it is considered a synchronous abrupt change. The data point with the largest temporal location change within this region is identified as the representative node of this abrupt change, and its event index, source number, mapping timestamp, and spatial location description are extracted. Through this method, the boundary nodes where misalignment occurs between multiple sources are obtained; these nodes are the time synchronization anomaly nodes, possessing high representativeness and value in indicating temporal structure breaks.

[0024] Based on the identified time synchronization anomalies, a set of reliable time anchors is constructed, and further filtered to form a stable anchor group. First, nodes with mixed behavioral content, low event confidence, or excessive source device time fluctuations are removed from all anomalies. Then, based on factors such as the number of repeated observations, whether a node appears in multiple anomaly segments, and whether it appears at the starting or turning point of a key behavioral chain, its event stability, temporal independence, and behavioral driving strength scores are calculated. Nodes with higher scores are included in the anchor candidate set. Within the anchor candidate set, nodes are sorted according to the time interval between nodes, the continuity of behavioral semantics, and the causal coherence of events. Nodes with global representativeness and behavioral chain structural support are selected to form a stable anchor group. Each anchor in this group has a clear temporal location, event meaning, participating entity identifier, and spatial location information, and can be traced back to the original behavioral record through data tracing. After the anchor group is determined, it serves as the temporal reference architecture for all behavioral path evolution, state changes, and anomaly inference processes in behavioral-driven twin prediction, ensuring high temporal consistency and physical rationality in subsequent data fusion, path prediction, and behavioral simulation.

[0025] Based on a set of reliable time anchors, a counterfactual replay operation is performed to reconstruct the historical evolution of extreme segments of behavioral data. The set of time offset vectors between data sources is calculated from the counterfactual replay sequence, and the corresponding time offset spectrum is generated based on the set of vectors, which is used as the basis for subsequent time series correction. After establishing a unified event time baseline and extracting a set of reliable time anchors consisting of multiple behavioral temporal intersection nodes, counterfactual replay needs to be performed based on this anchor set to reconstruct the behavioral evolution path and extract temporal misalignment features to generate a temporal offset spectrum, supporting subsequent data alignment and path correction. The specific steps are as follows: The target behavioral segments for counterfactual replay are identified, and a behavioral replay sequence oriented towards temporal causal chain reconstruction is constructed. Specifically, within a time interval composed of multiple credible time anchors, the time span and event density fluctuation between adjacent anchors are calculated one by one. Key segments with a time span exceeding fifty milliseconds, drastic changes in event density, and spanning multiple data sources are selected as the target replay intervals. Within this interval, all original behavioral data are extracted, retaining fields such as event occurrence time, behavior name, associated entity identifier, execution duration, behavior spatial location, behavior purpose description, and source device number. These data are then rearranged according to a unified time baseline to construct a behavioral replay sequence with temporal integrity and semantic continuity. This sequence not only includes the physical order of events but also encompasses the dependencies between behaviors, spatial displacement logic, and entity participation trajectories, used to reconstruct the ideal evolutionary process that each data source should present under the assumption of no misalignment.

[0026] A counterfactual replay environment with time constraints and behavioral consistency is constructed, and the evolution process is progressively advanced according to the event sequence. During actual execution, the behavioral replay sequence is divided into continuous windows with millisecond-level granularity. Each time window allows only one or more behavioral events to occur in parallel, provided that these events do not conflict in terms of entity, location, or function. For example, if two events act on the same mechanical arm in the same time window with opposite directions, the replay within that window will be marked as a conflict scenario; if two events act on different entities, are more than two meters apart, and have no overlapping behavioral purposes, parallel replay is allowed. Within each window, events are loaded sequentially, and the event status is recorded to ensure logical consistency with the behavioral outcome in the previous window, such as whether there is spatial overlap with inconsistent directions, or whether the behavioral consequences occur prematurely or are delayed. By progressively replaying events to a unified timeline and performing spatial consistency, behavioral logical integrity, and event causal compliance checks at each step, a set of counterfactual trajectories of the actual behavioral evolution is obtained. By comparing the behavioral evolution with that shown by real data, we can clearly observe the fictitious trajectories caused by misalignment, the overlap of non-physical behaviors, and the phenomenon of sudden jumps.

[0027] Based on the conflicting event points and logical breakpoints identified during behavior replay, a set of time offset vectors across data sources is constructed. In practice, all event pairs exhibiting logical anomalies in the counterfactual trajectory are traversed, extracting their event occurrence time, participating entities, event sequence number, and behavior deviation type, and comparing these with the collection time of their respective original sources in the real data. Events in the event pair whose time is earlier than the replay trajectory are marked as "early offset," and events whose time is later than the trajectory are marked as "lagging offset," while their offset time values ​​are calculated. By statistically analyzing the offsets of all conflicting event pairs, a set of time offset vectors is generated, with the data source number as the primary index and behavior type, offset direction, offset value, confidence level, and anomaly scenario number as additional dimensions. During this process, noise fragments caused by the uncertainty of the behavior content itself are removed, retaining only event pairs where the same entity's behavior occurs repeatedly at different times or in reverse order within spatially overlapping areas, ensuring the representativeness and interpretability of the offset vectors. This set of offset vectors truly reflects the drift patterns of each data source's behavior relative to the actual evolution logic at the event level under the time architecture constructed with trusted anchors.

[0028] A continuous temporal migration spectrum is constructed based on a set of time migration vectors to characterize the migration patterns and temporal distortion between different data sources, serving as a guiding principle for subsequent corrections. Operationally, all migration vectors are divided into continuous non-overlapping windows along a unified time axis, with each window width set to five milliseconds. The average migration value, maximum migration value, and migration direction percentage of all migration events within each window are statistically analyzed, along with the distribution of event behavior types. Subsequently, the statistical values ​​from all windows are used to construct a time migration intensity sequence, and multi-scale time-domain analysis is performed on this sequence. Migration patterns of different frequency components are extracted using continuous wavelet transform, generating a two-dimensional temporal migration spectrum. In this spectrum, the horizontal axis represents the window position on the unified time baseline, the vertical axis represents the migration frequency distribution, and image brightness represents migration intensity. The spectrum allows for visual observation of which time periods exhibit periodic misalignment (fixed frequency, stable intensity), which time periods exhibit sudden jumps (unfixed frequency, drastic intensity), and whether resonant misalignment patterns exist between multiple sources (multi-band overlap enhancement). Ultimately, this spectrogram is used as the core input parameter for subsequent temporal elastic alignment, behavioral path reconstruction, and causal topology repair, ensuring that the correction process is based on real behavioral evolution data and has temporal coherence and physical consistency.

[0029] Based on the time offset spectrum, the cross-source phase traction test results between different data sources are obtained. By jointly comparing the residual consistency index and the phase-locked stability index, the boundary nodes with time misalignment in each data source are identified. Precise segment alignment coordinates are set at the misalignment boundary node positions to ensure the time continuity between different time periods. After constructing the time-off spectrum, to identify misaligned segments and time boundaries between multi-source behavioral data, it is necessary to progressively complete cross-source phase pulling, stability analysis, misalignment boundary identification, and precise alignment coordinate setting based on the spectrum, ensuring the temporal continuity and behavioral consistency of the data fusion process. The specific steps are as follows: Based on the constructed time offset spectrum, a cross-source phase traction probe operation is performed to clarify the potential synchronization relationship between data sources. In the specific implementation, two data sources with overlapping behavioral regions are selected, and their misalignment concentration areas are located in the spectrum. A 20-millisecond time interval is defined using this as the traction window. Within this interval, all behavioral events from both data sources are extracted, including behavioral content, occurrence time, target, spatial location, and source number. Using the more time-stable data source as a reference and the other data source as the traction target, a linear time axis offset probe is implemented. Specifically, the behavioral events of the traction target data source are progressively shifted forward or backward, with each shift step being 0.5 milliseconds. After each adjustment, the sum of the behavioral time difference and the sum of the behavioral spatial distance difference between the two data sources within the time window are recalculated, and the trend is recorded. If the sum of the time differences generated by multiple consecutive shift operations tends to decrease, it indicates a potential synchronization trend, termed the traction response. The entire process does not rely on a global synchronization mechanism but instead constructs cross-source time response curves in the micro-scale time domain for subsequent stability analysis.

[0030] Based on the successful phase traction trial, a residual consistency index is calculated, and statistical analysis is performed on the residual event time difference after traction to assess the degree of behavioral synchronization. In this step, all event pairs within the traction window are traversed, and the change in time difference for each pair before and after traction adjustment is recorded to construct a residual sequence. Mathematical analysis is performed on the residual sequence, including the mean, variance, maximum difference, and standard deviation. The residual change rate is also normalized as an additional indicator. If the average residual value approaches zero and the variance decreases significantly after adjustment, the traction operation is considered effective. Furthermore, the behavioral types of each event pair are categorized, and the residual change trends are compared by type to confirm whether there is a significant impact of behavioral category on synchronization accuracy. For example, separate statistics are performed on mechanical control events, visual recognition events, and position change events. If the residual changes of all types are highly consistent, the residual consistency is rated as highly reliable. Residual consistency, as a quantitative indicator for judging whether phase traction has formed a true alignment basis, has strong temporal logical guidance in this implementation process.

[0031] After the residual consistency reaches the threshold, phase-locked stability analysis is performed on the traction window to determine whether the adjusted time series has continuous synchronization capability. The procedure involves extending the time segment by 20 milliseconds before and after the segment to cover the entire behavioral evolution. Within each 1-millisecond sub-window, the phase difference (i.e., time difference) data between event pairs is extracted, and the direction and magnitude of phase difference changes between consecutive sub-windows are statistically analyzed. If the phase difference direction remains consistent across five or more sub-windows, and the change magnitude is less than 2 milliseconds, the segment is considered to have phase-locked stability. Higher stability indicates that the time relationship between the two data sources is more likely to maintain a stable synchronized state, making it valuable as a segment alignment reference. Ultimately, a continuous phase-locked stable segment is obtained, corresponding to the low-frequency, slowly varying region in the spectrogram, proving that the synchronization between behavioral rhythms exists not only in local events but also in the overall evolution of the behavioral rhythm.

[0032] Fourth, after obtaining the joint index of residual consistency and phase-locked loop stability, misalignment boundary node identification is performed. Specifically, all traction windows are mapped as continuous segments on the time axis, and the synchronous jump points of residual value changes and phase-locked loop indices are observed. At these jump points, symmetrical windows are constructed, and the event sequences before and after are extracted and compared to see if there are abrupt changes in the evolution of the behavioral trajectory. For example, if the operation objects in the previous segment are concentrated in one type of entity, and the next segment suddenly changes to another type, or if the previous segment's behavior is initialization and the next segment is an abnormal stop, then the node is marked as a potential misalignment boundary node. Furthermore, the differences in the number of events, the number of active entities, and the spatial distribution within 20 milliseconds before and after the node are calculated. If significant differences occur and the causal chain of behavior is interrupted, then the node is confirmed as a misalignment boundary node. Finally, a set of boundary nodes is generated, with each node recording a timestamp, data source combination, residual jump value, phase-locked loop signal instability point, and behavioral difference score for subsequent alignment.

[0033] Based on the identified misaligned boundary nodes, segmented alignment coordinates are set to achieve local reconstruction of the timeline. The operation involves defining a 10-millisecond alignment transition zone before and after each boundary node, and redistributing behavioral events within this zone. All behavioral events from lagging data sources are shifted forward, and those from earlier data sources are shifted backward. The shift magnitude is obtained by linear interpolation of the offset difference between the preceding and following segments to avoid abrupt jumps. During the adjustment process, the original order of events is maintained, and the priority of behavioral dependencies is strictly followed to ensure no logical errors. After alignment, the correspondence between the boundary of each alignment segment and the original timeline is recorded as an alignment mapping record, forming a precise segmented alignment coordinate structure. This structure serves as a basic reference in subsequent behavioral prediction, ensuring that multi-source data can be traced back to a unified behavioral timeline position at any time, preventing fictitious paths from reappearing due to time misalignment.

[0034] Based on the segmented aligned coordinates, a causal topology graph model with multiple sampling rates is constructed. The mapping path from multi-source behavioral data to entity state is reconstructed according to the causal topology graph model. At the same time, non-consistent trajectory regions in the mapping path are identified, trajectory rejection regions are established, and false behavioral trajectory fragments caused by time misalignment are shielded. After completing the segmentation and alignment of multi-source data and obtaining a unified time reference coordinate, it is necessary to establish a behavioral causal structure under multiple sampling rates, identify inconsistencies in the misaligned reconstructed behavioral paths, and set rejection zones to mask non-realistic trajectories, thereby improving the reliability and accuracy of behavioral prediction. The specific steps are as follows: A multi-sampling rate behavioral event set is constructed on a unified timeline based on segmented aligned coordinates, providing a foundation for subsequent causal topology modeling. This operation begins by remapping all data from various sources in their respective aligned time coordinates. Each original behavioral event needs to record its behavior type, executing entity, operation object, operation duration, behavior spatial location, event triggering conditions, and behavior completion status. Due to differences in sampling periods among different data sources, unified alignment processing of multiple sampling rates needs to be achieved through time reconstruction. Based on the least common multiple of the aligned time series, a unified sampling granularity is determined, for example, every 10 milliseconds as a time unit, and various types of behavioral data are interpolated and mapped to this time unit. For behavioral data with high sampling frequency, the event closest to the target sampling point is selected as the master mapping point, while retaining the start and end times and the duration of the behavior. For data with low sampling frequency, when no direct event exists within a time point, the state of the previous event is maintained to the current sampling point through the event extension principle. The final constructed behavioral event set is arranged on a unified timeline with a 10-millisecond granularity. Each time point contains the corresponding behavioral content from all data sources, serving as the data foundation for subsequent causal graph construction.

[0035] A causal topology graph is constructed based on the reconstructed timeline of behavioral events to clarify the impact path of behaviors on entity state changes. In this graph construction process, the entire behavioral sequence is first divided into continuous time nodes with time as the horizontal axis. Each node represents a set of behaviors at a specific point in time. Each individual event in each behavior set includes information such as event time, target entity, action performed (e.g., opening, closing, moving, issuing commands), start and end points, and trigger signal source identifier. Then, causal connections between nodes are identified, based on three dimensions: temporal sequence, spatial continuity, and the propagation of entity state changes. If a behavioral event in time node A acts on entity P, and entity P undergoes a state change in the next time node B, and the two behaviors are spatially continuous, a causal edge is considered to exist between node A and node B. To enhance structural rationality, it is also necessary to verify whether the preceding and following behaviors have the same triggering conditions or behavioral command succession relationships, such as whether they are triggered by the same source, whether the target object is consistent, and whether the operation sequence order is logically consistent. This information is encoded as edge attributes in the causal graph, forming a directed graph structure with weights, directions, entity labels, and spatial ranges. Ultimately, this topology diagram fully expresses the path from behavior event-driven changes to entity state changes, and characterizes the behavior synthesis logic between multi-source data in a structured way.

[0036] A topological graph structure is used to perform consistency checks on behavioral trajectory paths, identifying and marking inconsistent path regions caused by temporal misalignment or source data interference. This process is handled in two dimensions: First, path jump behavior is analyzed in the time dimension. If there are abrupt state transitions between consecutive nodes, instantaneous changes in the behavioral object, or spatial displacements exceeding a preset range (e.g., more than 2 meters), these are marked as path abrupt change regions. Second, content consistency is checked in the behavioral semantic dimension. If two or more behavioral paths act on the same entity simultaneously within a certain time period, and the behavioral intentions are contradictory (e.g., simultaneous "on" and "off" behaviors), these are judged as behavioral conflict paths. Simultaneously, the edge weight changes in the behavioral paths are statistically analyzed. If the weights of multiple consecutive causal edges are lower than a set confidence threshold (e.g., less than 0.3), it indicates that the behavioral data source is unstable or the behavioral logic chain is incomplete, and this path segment is also marked as a potential abnormal path. All path abrupt change segments, semantic conflict segments, and low-confidence path segments are uniformly included as trajectory inconsistency fragments, recording their start and end times, the numbers of the involved behavioral events, the entity objects, the spatial location segments, and the information on changes in inference weights, forming a set of trajectory suspicious segments.

[0037] Based on the set of suspicious segments, a trajectory rejection region is defined and injected into the behavioral data processing chain to shield and eliminate false trajectories. Specifically, a set of temporal and spatial behavioral elimination rules is constructed for each inconsistent path segment. These rules include: within the rejection time window, all behavioral events originating from a specific data source, acting on a specific entity, and belonging to the category of high-conflict actions are removed from the original behavioral dataset; behavioral events with uncertain sources or excessively low edge weights are treated as missing values ​​in subsequent data fusion and do not participate in state calculations. In the causal topology graph structure, nodes within the rejection region and their connected causal edges are deleted, disconnecting them from upstream and downstream behavioral chains to prevent path offset due to a single erroneous behavior. To ensure the continuity of the behavioral chain, completion can be performed through behavioral events on high-confidence paths, constructing virtual transition paths in the rejection segment to reconnect broken state evolutions and maintain the integrity of the model's reasoning capabilities. The entire rejection region determination process corresponds one-to-one with the original causal structure, behavioral time-series data, and entity state feedback logic, forming a transparent and verifiable path purification mechanism.

[0038] Based on the trajectory rejection region, an adaptive time calibration operation is performed. The time offset vector set is used to perform elastic time alignment on multi-source behavioral data to generate a corrected behavioral data stream. The correction result is written into the time-series calibration matrix to freeze the identified drift channels and establish a cyclic correction mechanism for online prediction correction. After constructing the trajectory rejection region, to ensure the stability and fusion consistency of multi-source behavioral data in the time dimension, adaptive time calibration needs to be performed on data channels with offsets. This involves elastically aligning the behavioral event timeline using a vectorization strategy and constructing a traceable calibration matrix structure to achieve a closed-loop correction capability for historical drift freezing and future behavior prediction. The specific steps are as follows: Based on the behavioral conflict segments recorded in the trajectory rejection region, and combined with the time offset vector set, a flexible time alignment operation is performed on multi-source behavioral data. At the start of the operation, all behavioral segments identified by the trajectory rejection mechanism as potentially time-distorted need to be re-extracted, especially those behavioral events that exhibit semantic conflicts, time jumps, or entity path breaks across data sources. These behavioral events all carry timestamps, entity identifiers, behavioral types, spatial locations, behavioral execution durations, and source data stream numbers. For each rejected behavioral event, based on the historical drift characteristics recorded in the existing time offset vector set, its average offset value, maximum offset amplitude, fluctuation trend, and corrected confidence interval within a similar time window are extracted, and a dynamic calibration window with the original event as the reference point is constructed. Within this window, a flexible time alignment operation is performed, specifically by progressively fine-tuning the event's timestamp forward or backward, with each adjustment not exceeding 1 millisecond. After each adjustment, the temporal consistency indicators exhibited by the same entity in other data sources are evaluated, mainly including: relative time difference, event overlap rate, behavioral response consistency, and spatial trajectory intersection density. If the adjustment significantly improves consistency metrics (e.g., time difference decreases to within 5 milliseconds, spatial trajectory overlap exceeds 80%), then the corrected time of the event is confirmed as the result of the current step adjustment and marked as "effective elastic alignment behavior". The entire process performs a complete time correction loop for each type of data source, forming a corrected behavioral data stream that is structurally balanced, behaviorally continuous, and temporally consistent.

[0039] After completing the time-location correction of all events, these results need to be structured and recorded in a unified time-series calibration matrix to construct a traceable, freezeable, and updatable calibration structure. The calibration matrix is ​​organized using a five-tuple structure, including: event number, original timestamp, post-calibration timestamp, offset adjustment value, and time adjustment reason. The offset adjustment value is used for subsequent prediction of behavioral offset estimation, and the adjustment reason field records which trajectory rejection segment triggered the adjustment, which time offset vector provided the reference, and its correction confidence level. In terms of matrix organization, each data source independently maintains a behavior calibration table sorted by time, and the calibration results from all data sources are cross-referenced through a unified calibration index table. This design ensures that during behavior prediction or state judgment, the calibrated time values ​​in the matrix can be directly used to replace the original timestamps, and guarantees a consistent behavioral time-series structure for the same event across different data sources. During the matrix construction process, a freeze-marking operation is also performed on data channels where time offsets are repeatedly corrected. Specifically, if a data source is found to have more than 90% of its behavioral events adjusted across three consecutive calibration windows, with consistent offset directions and similar magnitudes, it is considered to have a persistent drift trend. In this case, the data source is marked as a "drift channel," and all subsequent new behavioral events from this channel must first undergo calibration matrix matching and mandatory correction before being allowed into the prediction process. The freeze operation is implemented in the matrix using a status identifier field and can be queried and verified in real time through the frozen channel index.

[0040] After generating and writing the corrected behavior data stream and matrix, a cyclical correction mechanism is established for the behavior prediction process, enabling continuous closed-loop data correction capabilities across the temporal dimension. Under this mechanism, whenever a new behavior event is input, it is first compared with existing calibration records in the temporal calibration matrix to find the historical correction record closest to the current event's source, behavior type, entity identifier, and event time period. If a match is successful, the system will call the corresponding offset correction parameters to perform an initial adjustment to the current event's timestamp. This adjustment does not take effect immediately but enters an online behavior consistency verification process, primarily checking whether the event and the current behavior chain context have issues such as disordered behavior order, entity conflicts, or broken path logic caused by temporal changes. If consistency verification passes, the calibration takes effect and is written to the matrix; if verification fails, the current event is added to the "event pool to be calibrated," and delayed calibration is performed again after collecting more context data. Simultaneously, for new behavior events that continuously appear and significantly differ from historical offset vectors, their offset trends are clustered to identify whether new time drift patterns have emerged. If a new drift feature is confirmed, a new drift correction path is automatically created, and an independent correction index structure is added to the calibration matrix. This enables the time calibration mechanism to achieve self-expansion, self-updating, and self-iterative capabilities, ensuring that the entire behavior-driven prediction architecture always maintains a time-series closed loop and structural stability.

[0041] Based on the cyclic correction mechanism, a time-reversal phase control command is generated to drive the programmable polarization metasurface to inject a correction frame envelope conjugate with the original trajectory. An energy suppression channel with the opposite direction is generated in the neighborhood of the zero-phase time anchor point, and a pulse-level trajectory elimination operation is performed. At the same time, the weight coefficients in the causal topology graph model are updated to realize the dynamic self-healing and closed-loop control of the data prediction path. After completing the elastic temporal alignment of behavioral data and establishing a temporal calibration matrix, it is necessary to further proactively eliminate potentially misleading behavioral trajectories in the time dimension. This is achieved by constructing a time-reversal signal, injecting a corrected frame structure, and linking it with causal topology updates, thereby realizing the precise elimination of false paths and the closed-loop self-healing of predicted links. The specific steps are as follows: After the behavioral data has undergone time elastic calibration and persistent offset paths have been identified, time-reversal phase control commands need to be constructed based on the evolution pattern of the drift channel. This operation begins by identifying behavioral segments in the trajectory rejection region that have not yet been completely extinguished. Even after time calibration, these behavioral segments still exhibit behavioral rhythms inconsistent with the multi-source data fusion logic, or form logical islands in the causal topology. To address such residual erroneous trajectories, a reverse signal channel needs to be constructed in the neighborhood of the zero-phase time anchor point. First, the time index, drift direction, offset amplitude, and behavioral event chain corresponding to the behavioral segment are extracted from the calibration matrix. Then, a behavioral path that is spatially consistent with the behavioral trajectory but reversed in the time dimension is generated. Each behavioral event in the path must meet two conditions: first, its behavioral type must be semantically dual to the original event, such as "ascending" corresponding to "descending," and "starting" corresponding to "closing"; second, its behavioral entity must be completely consistent with the original event, ensuring that the behavioral energy of the target entity is directionally eliminated when the interference is performed. Based on this, a set of control instructions is constructed to describe the activation time, entity number, semantic inversion method, time anchor point position, control strength, and execution duration of each event in the reverse behavior path. This set of control instructions is the time inversion phase control instruction, used to drive physical behavior interference operations.

[0042] The execution of time-reversal phase control commands activates a programmable polarization metasurface, injecting a modified frame envelope signal conjugate to the original trajectory into the target behavior trajectory path region, and generating an energy suppression channel in the neighborhood of the zero-phase anchor point. During execution, the programmable polarization metasurface is first driven to release energy units matching the original behavior path based on the behavior entity number and modified frame activation time range defined in the aforementioned control command set. Each modified frame unit contains four key structures: first, a behavior event semantic label reversal structure to cancel the original behavior during execution; second, a spatial path reconstruction structure to precisely align the modified behavior with the original trajectory; third, a time conjugate structure to ensure it occurs within the reverse time window of the original event; and fourth, a behavior intensity adjustment structure to align the behavior energy level for effective interference. These modified frames constitute a complete frame envelope sequence and are injected sequentially under the control of the programmable polarization metasurface. In the neighborhood of the zero-phase anchor point, by adjusting the injection angle and polarization direction, a behavior energy convergence point is formed, and an energy suppression channel is constructed in reverse. This channel expands in a cone shape in front of the behavior path, forming a behavior energy weakening region with strong directionality and high absorption efficiency. When the original false behavioral trajectories in the path pass through the channel, they will gradually disappear until they are completely extinguished due to phase reversal, energy overlap and semantic offset, forming a pulse-level trajectory suppression effect, thus achieving the purpose of clearing false paths without deleting behavioral data.

[0043] After completing the trajectory elimination operation, the weight values ​​of the affected path segments in the causal topology graph structure must be updated promptly to achieve dynamic self-healing and closed-loop control of the prediction network. This update unfolds in the following steps: First, identify all original behavioral event nodes covered by the correction frame envelope and freeze their edge structures in the topology graph. The freezing operation means setting the weight coefficients of the original behavioral paths to the minimum value in the topology graph and marking them as "untriggerable". Next, add the reverse behavioral event sequence constructed in the correction frame as a new topology path segment, calculate the behavioral credibility, contextual logic consistency, and system feedback response strength for each edge, and assign new weight values ​​to the edge structure based on these. Subsequently, trigger the topology graph structure consistency verification process: scan all downstream paths originating from the correction node layer by layer to ensure there are no isolated subgraphs, no behavioral conflict paths, and no causal loop links. If graph structure errors are found, backtrack to the correction frame injection stage to re-evaluate the frame path construction; if the verification passes, update the topology version and write it into the prediction path control table. Finally, based on the new causal topology, in the next behavior prediction task, the updated path will be preferentially included in the candidate path set, and the original false path will be shielded to ensure that the future prediction results will no longer be affected by it, thus achieving structural closed-loop self-healing of the path.

[0044] This invention addresses the timing drift problem caused by differences in sampling accuracy and transmission delays between different data sources by establishing a unified event time baseline and a set of reliable time anchors. Through counterfactual replay and phase traction mechanisms, it captures and quantifies potential misalignment regions in behavioral segments. Furthermore, by constructing a causal topology graph and rejection regions, it effectively identifies and masks false trajectories, achieving the discrimination and isolation of abnormal behavioral paths. Building upon this, it introduces elastic time calibration and correction write-back to construct an online cyclic correction mechanism, enabling the system to have adaptive calibration capabilities. Finally, through time inversion control and conjugate frame injection, it eliminates interfering behavioral trajectories at the physical level and dynamically updates the weights of the causal structure graph, achieving real-time self-healing and closed-loop feedback of the system's predicted paths. Overall, this method breaks through the traditional time synchronization methods centered on static alignment, realizing a dynamic fusion prediction mechanism driven by behavioral causality and aiming for full-time closed-loop operation.

[0045] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A behavior-driven twin prediction method, characterized in that, Includes the following steps: Establish a unified event time baseline, collect nanosecond-level clock offset information from various data sources, construct a cross-time sequence suspicion diagram, identify time synchronization anomaly nodes, and form a set of reliable time anchor points; Based on a set of trusted time anchors, counterfactual replay is performed to reconstruct the historical evolution of behavioral data, generate a set of time offset vectors, and construct the corresponding time-series offset spectrum. Based on the time offset spectrum, cross-source phase traction is explored. Combined with residual consistency and phase-locked stability indicators, time misalignment boundary nodes are identified, and segment alignment coordinates are set. Based on segmented aligned coordinates, a multi-sampling-rate causal topology graph is constructed to reconstruct the mapping path from behavioral data to entity state, and non-consistent regions in the trajectory are marked to establish rejection regions. Based on the rejection region, adaptive time calibration is performed. The time offset vector set is used to elastically align multi-source behavioral data, and the correction results are written into the calibration matrix to form a cyclic correction mechanism. Based on the cyclic correction mechanism, time-reversal control commands are generated to drive the polarization metasurface to inject conjugate frame envelopes, construct energy suppression channels in the neighborhood of zero-phase anchor points, eliminate spurious trajectories, and update the weights of the causal topology graph.

2. The behavior-driven twin prediction method according to claim 1, characterized in that, The process of forming a set of reliable time anchors is as follows: A high-precision master clock source is selected as the synchronization reference. The time offset of each data source is collected by bidirectional round-trip measurement. A unified reference time mapping matrix is ​​constructed, and the timestamps of all behavioral data are mapped to a unified event time baseline. Under a unified time baseline, multi-source behavioral events are rearranged according to the mapped timestamp, the event density is statistically analyzed by sliding time windows, a two-dimensional time-source distribution map is drawn, a cross-time series suspicion map is constructed, and areas with abnormal density are identified. Based on the density anomaly regions in the suspect map, the event temporal density, spatial location and behavioral semantics are extracted, and the synchronous mutation behavior is identified by clustering to determine the time synchronization anomaly nodes in each data source. Events with high confidence, clear time location, and strong behavioral representativeness are selected from abnormal nodes to construct a set of credible time anchors, which are then combined to form a stable anchor group, providing a time series reference for behavioral paths.

3. The behavior-driven twin prediction method according to claim 2, characterized in that, The steps for constructing the temporal offset spectrum are as follows: Select time segments with large time spans and fluctuating event density from the anchor point set to construct behavior replay sequences; Under a unified time baseline, the behavior replay sequence is advanced at the millisecond level to determine whether there are spatial conflicts, logical breaks or causal anomalies between events within each time window, and to generate a set of counterfactual behavior trajectories. Identify anomalous event pairs from counterfactual behavior trajectories, extract event time, behavior sequence and offset direction, calculate time offset values ​​between various data sources, and construct a set of time offset vectors; The time offset vector is statistically analyzed according to the time window, and multi-scale time domain analysis is performed to generate a time-series offset spectrum.

4. The behavior-driven twin prediction method according to claim 3, characterized in that, The steps for setting segment alignment coordinates are as follows: Based on the time offset spectrum, a misalignment concentration section is selected, and a cross-source phase traction probe is performed. By shifting the time axis of the behavioral event, a cross-source time response curve is formed. Based on the traction, the residual consistency index is calculated, the trend of event time difference changes is analyzed, the synchronization accuracy of different behavior types is evaluated, and the effectiveness of phase traction is determined. Perform phase-locked stability analysis, statistically analyze the direction and magnitude of the change in the phase difference of events within consecutive sub-windows, and identify time periods with continuous synchronization capabilities; Based on the residual consistency index and phase-locked stability analysis results, time jump points and behavioral change points are identified, and timestamps, behavioral differences and residual jump values ​​are extracted to form a set of misalignment boundary nodes. An alignment transition zone is constructed with the misalignment boundary node as the center. The behavior event time translation operation is performed according to the offset difference to establish a segmented alignment coordinate structure between the time axis and the behavior sequence.

5. The behavior-driven twin prediction method according to claim 4, characterized in that, The steps for establishing a denial-of-identity zone are as follows: Based on the segmented alignment coordinates, a unified sampling mapping is performed on the multi-source behavioral data to construct a set of behavioral events with multiple sampling rates and rearrange them onto a unified timeline; A causal topology graph is constructed based on a set of behavioral events. Causal connections between behavioral events are established through temporal sequence, spatial continuity, and changes in entity state, generating a behavioral path structure with directionality and weight. Identify trajectory abrupt changes, semantic conflicts, and areas of reduced confidence in the behavioral path to form a set of trajectory inconsistency segments; Based on the set of trajectory inconsistency segments, a trajectory rejection region is set, conflicting behavioral events and causal edges are removed, and a high-confidence virtual transition path is constructed to maintain the integrity of the causal structure.

6. The behavior-driven twin prediction method according to claim 5, characterized in that, The formation process of the cyclical correction mechanism is as follows: Based on the marked behavioral conflict segments in the trajectory rejection area, and combined with the time offset vector set, the timestamps of behavioral events are flexibly aligned to generate a corrected behavioral data stream. Write the corrected time results into the time series calibration matrix, record the behavior event number, original timestamp, calibrated timestamp, offset adjustment value and adjustment source, and mark the data source with continuous offset as the drift channel; During the behavior prediction process, the time-series calibration matrix is ​​invoked to perform dynamic matching on new events, triggering real-time calibration operations and constructing a cyclic correction mechanism to achieve closed-loop updates and continuous stability of behavior time calibration.

7. The behavior-driven twin prediction method according to claim 6, characterized in that, The operation of writing the correction results into the time-series calibration matrix further includes performing drift channel marking on data sources that have a high proportion of the same offset direction within multiple consecutive calibration windows, and forcibly calling the corresponding correction record for all new events in the drift channel in subsequent behavior prediction to complete the pre-calibration.

8. The behavior-driven twin prediction method according to claim 6, characterized in that, Based on the cyclic correction mechanism, time-reversal control commands are generated to drive the polarization metasurface to inject conjugate frame envelopes. Energy suppression channels are constructed in the neighborhood of zero-phase anchor points to eliminate spurious trajectories, and the causal topology graph weights are updated synchronously. The steps are as follows: Based on the persistent drift path and the residual erroneous behavior segments in the trajectory rejection region, a time-reversal phase control instruction is constructed, and the activation time, behavior entity, semantic duality and execution intensity in the reverse behavior path are defined. The programmable polarization metasurface is driven to inject a modified frame envelope conjugate with the original trajectory, and an energy suppression channel with the opposite direction is generated in the neighborhood of the zero-phase time anchor point to perform pulse-level trajectory elimination operation. Freeze the topological edge weights corresponding to the original behavior path, add a new corrected behavior path and update the edge weights to complete the self-healing of the causal topological graph structure, and ensure that the closed-loop control of the subsequent prediction path is consistent with the logic.

Images