Method for constructing a safety training scenario based on behavior trace

Abstract

The present application relates to the technical field of safety training and data analysis, and specifically discloses a safety training scenario construction method based on behavior tracing, which comprises the following steps: firstly, collecting the behavior time series data of the individual and the collaborative instruction time series data of the central control system in parallel; secondly, performing time stamp alignment and cross-modal fusion coding on the two types of data to generate joint representation data; thirdly, identifying and labeling the system-induced behavior nodes induced by the collaborative instruction logic through multi-scale sliding window analysis and bidirectional guidance intensity calculation; fourthly, synchronously reconstructing the behavior and instruction processes in a virtual simulation environment and constructing an interactive training scenario through high-correlation presentation of dynamic visual connectors; and finally, setting branch choices in the scenario, driving multi-path deduction and comparison, and generating a traceability analysis report containing individual decision evaluation and collaborative instruction quantitative optimization suggestions; the present application improves the accuracy and systematic improvement value of training.

Description

Technical Field

[0001] This invention relates to the field of safety training and data analysis technology, specifically to a method for constructing safety training scenarios based on behavior tracing. Background Technology

[0002] In complex collaborative operation scenarios such as airports, ports, and large industrial plants, the causes of safety accidents are often not simply individual operational errors, but rather deeply embedded in the complex dynamic interaction between "individual behavior" and "system instructions." Existing safety training methods, especially scenario-building technologies based on behavior tracing, mainly rely on the analysis of single-dimensional data such as individual behavioral trajectories and operation logs from historical accidents or violations. While these methods identify high-risk behavioral patterns through data mining and construct virtual training scenarios accordingly, and have a certain degree of targeting, their core flaw lies in the one-sidedness of causal attribution.

[0003] Existing safety training methods based on behavior tracing, by performing only one-way analysis of isolated individual behavioral data, fail to identify and separate specific behaviors induced by the deeper pressures exerted by the system's collaborative command logic. This leads to the erroneous attribution of what are essentially "system-induced behaviors" (such as a driver being forced to deviate from the prescribed route to complete a conflicting dispatch order) to spontaneous, deliberate violations by the workers, resulting in a severely distorted causal attribution of safety accidents. This attribution distortion causes the constructed training scenarios to remain superficially focused on punishing individual actions, failing to help trainees understand the transmission mechanism of system risks and their own real decision-making dilemmas, and failing to provide a basis for optimizing the system's command logic. Consequently, the training is unable to fundamentally prevent the recurrence of similar accidents. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing security training scenarios based on behavior tracing, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: The method for constructing security training scenarios based on behavior tracing includes the following steps: S1: Collect heterogeneous time-series data streams in the target operation scenario, wherein the heterogeneous time-series data streams include individual behavior time-series data and collaborative instruction time-series data; S2: Perform timestamp alignment and cross-modal joint encoding on individual behavior time series data and collaborative instruction time series data to generate joint representation data that can characterize the evolution of individual behavior patterns in the context of behavior corresponding to collaborative instruction time series data; S3: By analyzing the temporal dependencies and causal relationships between the temporal data of collaborative instructions and the temporal data of individual behaviors in the joint representation data, the system-induced behavior nodes induced by the collaborative instruction logic are identified and labeled. S4: Based on the information of system-induced behavior nodes, the behavior process and collaborative instruction process of historical events are synchronously reconstructed in a virtual simulation environment. The system-induced behavior nodes and the time sequence data of the collaborative instructions that trigger the corresponding nodes are presented with high correlation, forming an interactive training scenario that can dynamically reveal the causal chain between system instruction pressure and individual behavioral decisions. S5: Based on interactive training scenarios, different coping strategies are set at the system-induced behavior nodes. The system deduces along different causal paths according to the user operation-driven scenario, compares and displays the results of each path, and finally generates a comprehensive evaluation of individual decision quality and includes a source analysis report that includes suggestions for optimizing the collaborative instruction logic.

[0006] As a further aspect of the present invention: S2 specifically includes: Based on a unified time base, the time-series data of individual behaviors and the time-series data of collaborative instructions are timestamped to obtain time-aligned individual behavior time-series data streams and collaborative instruction time-series data streams; For time-aligned individual behavior time-series data streams and collaborative instruction time-series data streams, calculate the dynamic correlation between individual behavior time-series data and collaborative instruction time-series data at each time point; Based on the dynamic correlation, the individual behavior time-series data stream and the collaborative instruction time-series data stream are adaptively weighted and fused to generate joint representation data.

[0007] As a further aspect of the present invention: the dynamic correlation between the time-aligned individual behavior time-series data stream and the collaborative instruction time-series data stream is calculated point-in-time, specifically including: For each aligned time point, the cooperative instruction time series data is mapped to the standard state space in which the individual behavior time series data should be located, and the conditional probability density of the actual observations of the individual behavior time series data in the standard state space is calculated. By comparing the conditional probability density at the current time point with the conditional probability density at the previous time point, the probability density transition of individual behavior time-series data state caused by changes in cooperative instruction time-series data is calculated. By coupling the probability density transfer quantity with the state change quantity of the individual behavior time series data at the corresponding time point, a dynamic correlation degree value representing the intensity of the immediate influence of the collaborative instruction time series data on the individual behavior time series data is generated.

[0008] As a further aspect of the present invention: S3 specifically includes: Multiple sliding analysis windows with different time lengths are set up on the joint characterization data; Within each sliding analysis window, the bidirectional guiding strength between the changing trends of collaborative instruction time-series data and individual behavior time-series data is calculated. Based on the bidirectional guidance intensity calculated under all sliding analysis windows, if and only if, within a set number of windows, the guidance intensity of instruction change preceding behavior change is consistently higher than the reverse guidance intensity and exceeds a set threshold, the individual behavior time series data of the corresponding time interval will be marked as a system-induced behavior node.

[0009] As a further aspect of the present invention: the bidirectional guidance strength between the changing trend of the collaborative instruction time series data and the changing trend of the individual behavior time series data specifically includes: The collaborative instruction time series data and individual behavior time series data within the window are decomposed into independent components that change continuously on the time axis; For each independent component of the collaborative instruction time series data, the collaborative instruction time series data is paired with all independent components of the individual behavior time series data, and the lead-lag correlation coefficients of the change sequence of the instruction component relative to the change sequence of the behavior component are calculated under all possible permutations and combinations. Statistical analysis was performed on all the lead-lag correlation coefficients obtained from the pairwise calculations, and the average correlation coefficients were extracted for the case where the change in instruction preceded the change in behavior, and the average correlation coefficients for the case where the change in behavior preceded the change in instruction. The average correlation coefficients of the two directions are multiplied by the normalized weight of the number of permutations and combinations of the corresponding directions to obtain the positive guidance strength value and the negative guidance strength value, which represent the causal possibility, and are used as the bidirectional guidance strength.

[0010] As a further aspect of the present invention: S4 specifically includes: Establish a unified spatiotemporal coordinate system and map the geometric trajectory of the reconstructed behavioral process and the time stamp of the cooperative instruction process to the spatiotemporal coordinate system; On the time axis of the spatiotemporal coordinate system, identify the start time of the system-induced behavior node, and extract the instruction point that is closest to the corresponding time and whose content has undergone key changes from the collaborative instruction process; Starting from the instruction point and ending at the corresponding system-induced behavior node, a visual connection body that dynamically changes over time is generated and rendered in the spatiotemporal coordinate system; By overlaying and fusing visual connectors with behavioral trajectories and instruction markers in a spatiotemporal coordinate system, the core of visualization for an immersive interactive environment is formed.

[0011] As a further aspect of the present invention: the generation process of the visual connective is as follows: Based on the temporal location of the instruction point and the spatial location of the corresponding behavior node, a set of path control points that evolve over time are calculated and defined. Using the cubic spline interpolation method, the path control points are smoothly fitted to generate a continuous spatiotemporal path curve, and the instantaneous radius of curvature at each moment is calculated as the dynamic curvature along the spatiotemporal path curve. Establish a nonlinear mapping relationship between dynamic curvature and pre-calculated guidance intensity, and dynamically determine the width and color brightness of the visual connector at each moment based on the nonlinear mapping relationship; In the spatiotemporal coordinate system, along the spatiotemporal path curve, based on the width and color brightness determined at each moment, a semi-transparent strip-shaped geometric body is rendered and drawn in real time as a visual connector.

[0012] As a further aspect of the present invention: S5 specifically includes: Capture the branch choices made by the user at the system-guided behavior nodes and convert the branch choices into corresponding logical control instructions; Based on logical control instructions, and combined with preset work rules and physical constraints, a continuous sequence of actions and state changes is generated in real time from the corresponding node until the task ends, based on the current state of the interactive training scenario. The virtual simulation environment is synchronously driven to extrapolate along the newly generated sequence in real time and the stored original historical sequence, and the differences in key state indicators on the two paths are recorded. The process records and results of the two path deductions are compared and displayed in a split-screen view on an integrated view. The differences in status indicators are quantified into specific optimization parameters for collaborative instruction logic and integrated into the source tracing analysis report.

[0013] As a further aspect of the present invention: the process of generating the continuous action sequence and the state change sequence specifically includes: The interactive training scenario is encoded at the system-induced behavior node, generating scenario state encoding parameters that include environmental parameters and individual pose. Based on the response strategy indicated by the logic control instructions, the corresponding target state vector and the set of allowed action operations are extracted from the preset operation rule base; Starting with the scenario state coding parameters and ending with the target state vector, a heuristic search is used in the state space, and the feasibility of the searched state transition sequence is verified and pruned in combination with physical constraints to generate a preliminary candidate state transition sequence. Perform time warping and smooth interpolation on the candidate state transition sequence to output a continuous sequence of actions and state changes that are continuous in time from the corresponding node until the end of the task.

[0014] The beneficial effects of this invention are: (1) By deeply analyzing the temporal dependencies and causal relationships between behavioral data and instruction data, this invention can accurately identify "system-induced behavioral nodes" induced by intrinsic risks such as system scheduling logic conflicts. This solves the problem of traditional safety training attributing systemic problems to individual operational errors, enabling the construction of training scenarios to trace from superficial behavioral violations to deep-seated system instruction pressures, thus improving the accuracy and root-cause relevance of training content and avoiding the training pitfall of "treating the symptoms but not the root cause".

[0015] (2) This invention constructs an interactive scenario in virtual simulation that integrates dynamic visual connectors and supports users in making decision-making deductions and comparing multiple paths at key nodes. This not only visualizes and concretizes implicit system pressure, enhancing the immersion and cognitive depth of training, but also outputs an assessment of individual decision-making ability and specific optimization parameters for collaborative instruction logic through the generated quantitative traceability analysis report. This achieves both training personnel's emergency response skills under system pressure and providing data-driven decision-making basis for optimizing the system itself, forming a closed loop from training to system improvement. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention is a method for constructing security training scenarios based on behavior tracing, comprising the following steps: S1: Collect heterogeneous time-series data streams in the target operation scenario, wherein the heterogeneous time-series data streams include individual behavior time-series data and collaborative instruction time-series data; S2: Perform timestamp alignment and cross-modal joint encoding on individual behavior time series data and collaborative instruction time series data to generate joint representation data that can characterize the evolution of individual behavior patterns in the context of behavior corresponding to collaborative instruction time series data; S3: By analyzing the temporal dependencies and causal relationships between the temporal data of collaborative instructions and the temporal data of individual behaviors in the joint representation data, the system-induced behavior nodes induced by the collaborative instruction logic are identified and labeled. S4: Based on the information of system-induced behavior nodes, the behavior process and collaborative instruction process of historical events are synchronously reconstructed in a virtual simulation environment. The system-induced behavior nodes and the time sequence data of the collaborative instructions that trigger the corresponding nodes are presented with high correlation, forming an interactive training scenario that can dynamically reveal the causal chain between system instruction pressure and individual behavioral decisions. S5: Based on interactive training scenarios, different coping strategies are set at the system-induced behavior nodes. The system deduces along different causal paths according to the user operation-driven scenario, compares and displays the results of each path, and finally generates a comprehensive evaluation of individual decision quality and includes a source analysis report that includes suggestions for optimizing the collaborative instruction logic.

[0020] In S1, time-series data reflecting the physical state and operational actions of individual workers in the target work scenario are collected. This data is obtained through an onboard data acquisition terminal deployed on the individual worker (e.g., a ground support vehicle). This terminal integrates a Global Navigation Satellite System (GNSS) receiver, an inertial measurement unit (IMU), and a vehicle bus interface. The GNSS receiver records the vehicle's three-dimensional spatial coordinates and precise UTC timestamps at a preset frequency (e.g., 10 times per second); the IMU synchronously collects the vehicle's real-time speed, acceleration, and heading angle data; and the vehicle bus interface continuously reads the vehicle's operational status signals, such as turn signals, brakes, gear positions, and the take-off and landing status of work devices (e.g., luggage lifts). All data is initially encapsulated at the acquisition terminal and appended with a unified high-precision timestamp, forming a continuous time-series data stream of individual behaviors.

[0021] Secondly, time-series data reflecting the task requirements and logic issued by the central control system to the individual operations are collected. This data is directly extracted from the background database and communication logs of the airport's central dispatch system. Specifically, the collected data includes: structured task instructions issued by the dispatch system to specific vehicle terminals (such as instruction number, target aircraft position, operation type, and priority), the precise timestamp of the instruction issuance, the recipient identifier of the instruction, and the status update record of the instruction (such as received, confirmed, and executed). This data is exported in the form of database records or log files to ensure its temporal order and the integrity of the instruction logic. Finally, the above two types of data streams are aggregated and initially organized to form the heterogeneous time-series data stream required for subsequent processing.

[0022] In S2, firstly, a timestamp alignment operation is performed. Individual behavior time-series data and collaborative instruction time-series data are aligned using a common, high-precision time server signal as a unified time reference. For each type of data, all data points are converted to this unified time reference based on their own timestamps. Since the two types of data may have different acquisition frequencies, a linear interpolation method is used to interpolate the lower-frequency data stream. This ensures that at each identical time interval (e.g., one point per second), corresponding individual behavior data sampling points and collaborative instruction data sampling points are obtained, forming strictly time-aligned individual behavior time-series data streams and collaborative instruction time-series data streams.

[0023] Secondly, the dynamic correlation between the two types of data is calculated point-in-time. The core of this process is constructing a "standard state space" as a reference benchmark. This standard state space is obtained by analyzing a large amount of historical normal operational data: collecting historical collaborative instruction data and its corresponding verified compliant historical individual behavior data; for each typical instruction type (e.g., "proceed to gate 207"), calculating the mean and standard deviation distribution of all behavioral parameters (such as speed and heading) of the operational individual over a subsequent period under that instruction. This distribution is defined as the standard state space corresponding to that instruction type. During real-time calculations, for each aligned time point, the corresponding instruction type and standard state space are first determined based on the collaborative instruction data at the current time point. Then, the probability that the actually observed individual behavior data value at the current time point falls within this standard state space is calculated. This probability value is the conditional probability density, specifically calculated using a Gaussian kernel density estimation algorithm based on the mean and standard deviation of the behavioral parameters under that instruction in historical data. Next, the difference between the conditional probability density at the current time point and the conditional probability density at the previous time point is calculated. This difference is the probability density transfer, which quantifies the degree to which behavior deviates from the standard state due to changes in instruction content or environment. Finally, this probability density transfer is multiplied by the change in individual behavior data at the same time point relative to the previous time point (e.g., the rate of change of velocity), resulting in an intermediate product. This intermediate product is then normalized so that its value falls between 0 and 1. The final value generated is the dynamic correlation degree at that time point. The closer this value is to 1, the stronger the influence of the collaborative instruction data on the individual behavior data at that moment.

[0024] Finally, adaptive weighted fusion is performed to generate joint representation data. The dynamic correlation degree value calculated in the previous step is used as the basic weight. For each time point, the collaborative instruction time series data (after numerical encoding) at that point is multiplied by the dynamic correlation degree value at that point to obtain a weighted instruction feature vector; simultaneously, the individual behavior time series data at that point is multiplied by one and subtracted from the dynamic correlation degree value to obtain a weighted behavior feature vector. Subsequently, these two weighted feature vectors are concatenated along the feature dimension to form an extended joint feature vector. This joint feature vector not only contains the original behavior and instruction information, but also, through the dynamic correlation degree weight, highlights the feature contribution of instruction data at times when the instruction influence is strong, while weakening the random fluctuation features that may exist in behavior data at times when the instruction influence is weak. Arranging the joint feature vectors of all time points in chronological order constitutes the final joint representation data, which fully depicts the evolution of individual behavior patterns driven by collaborative instructions.

[0025] In S3, multiple sliding analysis windows with different time lengths are set on the joint representation data. Specifically, three different time length windows are set, such as 5 seconds, 10 seconds, and 15 seconds, to capture interaction patterns at different time scales. The sliding step size of the windows is uniformly set to 1 second. For a given joint representation data sequence, starting from the beginning time point of the sequence, each time point is used as the starting point of the window, and the three window lengths are applied to extract data, thereby generating multiple sets of data segments covering the entire time series but with different time spans. These data segments are the sliding analysis windows. Each window contains the time series data components of cooperative instructions and individual behavior corresponding to that time period.

[0026] Within each sliding analysis window, the bidirectional guiding strength between the trends in collaborative command time-series data and individual behavior time-series data is calculated. Before the calculation begins, the collaborative command time-series data and individual behavior time-series data within the window are preprocessed. By extracting the main changing components, each data sequence is decomposed into 3 to 5 independent components that change smoothly and continuously on the time axis, with each component representing a specific change pattern in the data. Subsequently, for each independent component of the collaborative command time-series data, it is paired with each independent component of the individual behavior time-series data to form multiple component pairs. For each component pair, the following operations are performed: Fix the instruction component sequence, and shift the behavior component sequence forward and backward along the time axis by a certain number of time units (e.g., from -10 units to +10 units). After each shift, calculate the Pearson correlation coefficient between the shifted behavior component sequence and the fixed instruction component sequence. The Pearson correlation coefficient is equal to the sum of the products of the differences between corresponding data points in each sequence and their respective mean values, divided by the square root of the product of the sum of the squares of the differences between all data points in the first sequence and their mean values ​​in the second sequence. Among the correlation coefficients calculated at all shift positions, the one with the largest absolute value is taken as the lead-lag correlation coefficient for that component pair in the current shift direction (instruction-led or behavior-led). After traversing all component pairs and shift directions, two sets of correlation coefficients are obtained: one set shows the correlation coefficients when all instruction component changes lead the behavior component changes; the other set shows the correlation coefficients when all behavior component changes lead the instruction component changes. Then, the arithmetic mean of these two sets of correlation coefficients is calculated. Finally, the average value of the instruction leading group is multiplied by the proportion of the number of correlation coefficients in that group to the total number of calculations to obtain the positive guidance strength value; the average value of the behavior leading group is multiplied by the proportion of the number of correlation coefficients in that group to the total number of calculations to obtain the negative guidance strength value. This positive guidance strength value and the negative guidance strength value together constitute the two-way guidance strength of this window.

[0027] Finally, by synthesizing the calculation results of all sliding analysis windows, system-induced behavior nodes are identified and labeled. After calculating the bidirectional guidance intensity of all windows, statistical analysis is performed. A quantity threshold is set, defined as 2 / 3 of the total number of windows. Simultaneously, an intensity threshold is set, for example, 0.7, which is determined by taking the 95th percentile of the distribution of bidirectional guidance intensity in a large amount of historical normal operation data. Each time point is examined along the time axis. For each examined time point, all sliding analysis windows containing that time point are identified. The number of windows whose positive guidance intensity value simultaneously meets the following two conditions is counted: first, the positive guidance intensity value of the window is greater than its own negative guidance intensity value; second, the positive guidance intensity value of the window is greater than the intensity threshold of 0.7. If the number of windows meeting this condition exceeds the set quantity threshold (i.e., exceeds 2 / 3 of the total number of relevant windows), then the time point is determined to be within a system-induced behavior interval. All these consecutive time points forming the interval are marked on the original individual behavior time series data; the data within this interval is labeled as a system-induced behavior node.

[0028] In S4, this step specifically involves constructing an interactive training scenario within a virtual simulation environment that can intuitively and dynamically reveal the causal chain between system command pressure and individual behavioral decisions. The scenario is constructed strictly based on the identified system-induced behavioral nodes, using highly correlated visualization techniques to concretize the systemic pressures hidden in historical events. The specific implementation process follows a four-stage sequence.

[0029] The first stage involves establishing a unified spatiotemporal coordinate system and mapping the data. First, a four-dimensional virtual spatiotemporal coordinate system is defined, where three dimensions (denoted as the X-axis, Y-axis, and Z-axis) represent three-dimensional physical space, and the fourth dimension (T-axis) represents time. The reconstructed historical behavioral process, i.e., the geometric trajectory of the individual worker, is mapped to this coordinate system: the three-dimensional spatial coordinates of each recorded point in the trajectory (e.g., three-dimensional rectangular coordinates converted from longitude, latitude, and altitude) are assigned corresponding (X, Y, Z) values, and their precise timestamps are linearly converted into coordinate values ​​on the T-axis, thus transforming the continuous motion trajectory into a parametric curve in this coordinate system. Simultaneously, the collaborative instruction process is mapped to the same coordinate system: each instruction is marked as a discrete point on the T-axis with its precise timestamp (called an instruction marker), and represented in space as an icon or text label associated with the instruction content at a fixed height (e.g., a certain distance above the plane of the individual's movement trajectory). Its position in (X, Y, Z) space is usually determined based on the target location involved in the instruction (such as the aircraft position coordinates) or the individual's real-time position when receiving the instruction. Through this mapping, the behavioral process and the instruction process are integrated and expressed within a unified spatiotemporal framework.

[0030] The second stage involves identifying key command points and associating them with system-induced behavior nodes. This is done within the established spatiotemporal coordinate system. On the axis, the start time point of each system-induced behavior node marked in step S3 (denoted as ) Next, the time series of the coordinated instruction process is traced back. On the axis, find the location The most recent instruction marker in time. A necessary condition for this instruction point to be considered a "cooperative instruction point triggering the corresponding node" is that the instruction content it represents has undergone a "critical change" compared to its preceding instruction. The determination of a critical change is based on a quantified threshold: calculating the degree of difference between the two instructions in key parameters. For example, for instructions involving target location, the degree of difference... Defined as the Euclidean distance between the target locations of two commands. When D is greater than a preset distance threshold... When the distance reaches 50 meters (for example), it is considered a critical change. If the nearest instruction point does not meet the critical change condition, the search continues backward until an instruction point that meets the condition is found. The time position of that point is recorded as follows. Spatial location is denoted as The spatial location of the corresponding system-induced behavior node is taken as the spatial location of the individual worker at the start time, denoted as... .

[0031] The third stage involves generating and rendering dynamically changing visual connectivities. This stage aims to create a connectivity from the command point. Extending to system-induced behavior nodes A link that visually and dynamically reflects the transmission process of command pressure. First, a spatiotemporal path curve is generated. From... arrive Within a time interval, a series of time points are sampled at fixed intervals (e.g., 0.1 seconds). For each Its corresponding spatial path control point Calculations are performed by combining spatial location and guidance intensity. Specifically, spatial linear interpolation is first performed: the calculation starts from... arrive On the straight path, relative to time The corresponding points, among which, Indicates the first At a certain point in time, Indicates the first A straight path Indicates the first A straight path. Then, based on that moment... Corresponding guiding strength value (Calculated in step S2), apply an offset perpendicular to the straight path direction to the above linear interpolation points. Offset amount and Proportional to the scene scale, with the scaling factor set according to the scene scale. This method yields a set of path control points. Subsequently, a cubic spline interpolation algorithm was used to smoothly fit this set of control points, generating a continuous and smooth spatiotemporal path curve. Along this curve, the time-series data was calculated. instantaneous radius of curvature This calculation uses numerical methods to solve for the first and second derivatives of the curve at that point, and applies the geometric definition of the radius of curvature (textual description: the radius of curvature is equal to the cube of the magnitude of the tangential velocity vector at that point, divided by the magnitude of the cross product of the tangential velocity vector and the normal acceleration vector at that point). It is called the dynamic radius of curvature.

[0032] Subsequently, a nonlinear mapping relationship is established between visual attributes and guidance intensity and dynamic curvature. The visual connective dynamically changes on two core attributes: width and color brightness. These two attributes are determined by the following two mathematical formulas: Width calculation formula: ;in, Indicates time The width of the visual connector, typically in meters, is mapped to the virtual environment. This is a width baseline constant, a positive value preset based on the visualization effect (e.g., set to 0.1). Indicates time The guiding strength value ranges from 0 to 1. Indicates time The dynamic radius of curvature. To prevent Too small a radius will cause abnormal width; set a minimum radius threshold. (e.g., 0.01), when calculated Less than season The physical meaning of this formula is that the greater the guidance intensity, the wider the visual connective, highlighting the intensity of the instruction pressure; at the same time, the greater the path curvature (the smaller the radius), the wider the visual connective, indicating a drastic change in the direction of movement (which may represent an avoidance action or a decision-making contradiction).

[0033] Color brightness calculation formula: ;in, Indicates time The color brightness value of the visual connector is normalized to the range of 0 to 1, with 1 representing the brightest. Use a base brightness constant (e.g., set to 0.3) to ensure that the connector remains visible even when the guide intensity is 0. This is the brightness adjustment factor (e.g., set to 0.7), used to control the extent to which the guiding intensity affects brightness. For a moment The guiding intensity is determined by the squared term used here to create a non-linear mapping, making the high guiding intensity regions more prominent in brightness.

[0034] Finally, based on the above calculations, real-time rendering is performed in a spatiotemporal coordinate system. Along the spatiotemporal path curve, at each sampling time point... Calculated As the width of the cross-section of the connecting body at that point, Using a semi-transparent material, a strip-shaped geometric shape with dynamically changing width and brightness over time is drawn to represent the color brightness of the connecting body at that point; this is the visual connecting body.

[0035] The fourth stage involves the overlay, fusion, and interactive construction of the core visualization. The dynamic visual connectors generated in the third stage are overlaid with the behavioral trajectory curves mapped in the first stage and the instruction markers extracted in the second stage, all displayed within the same 3D rendering viewport of the virtual simulation scene. The visual connectors act like a "causal force line," clearly connecting the instruction points that trigger pressure with the behavioral nodes affected by that pressure in space and time. Users (trainees) can freely change their perspective within this environment to observe the entire causal chain. When the user hovers their visual focus over or clicks on key points of the visual connectors, instruction markers, or behavioral trajectories, the system triggers the display of more detailed data, such as the precise guidance intensity value at that moment, the full text of the instruction, and the vehicle speed. This scene, integrating multi-dimensional data and a dynamic visualized causal chain, constitutes the core visualization of the immersive interactive training environment, providing an intuitive and information-rich operational background for subsequent interactive decision-making simulations.

[0036] In S5, the user's branch selection is first captured and converted into a logical control instruction. When the virtual simulation scenario reaches a system-induced behavior node, the scenario pauses and presents the user with multiple branch options designed based on the node's historical background and operational rules. For example, option one is "Follow the historical behavior," option two is "Adhere to standard procedures and wait," and option three is "Try an alternative safe path." After the user confirms their selection via an interactive device (such as a click or controller selection), the selection is captured. Subsequently, a predefined interpretation rule maps this selection into a structured logical control instruction. This instruction contains at least the following fields: instruction type (e.g., "stop and wait," "path replanning"), instruction strength parameters (e.g., waiting duration, expected destination of the new path), and the identifier of the system-induced behavior node that triggered the instruction. This logical control instruction serves as the input signal driving subsequent dynamic scenario deduction.

[0037] Secondly, continuous action sequences and state change sequences are generated in real time based on logical control commands. This generation begins with encoding the complete state of the interactive training scenario at the system-induced behavior node. This encoding process extracts and quantifies all relevant parameters in the current virtual environment, including the precise pose of the individual operator (3D coordinates, heading angle, velocity vector), the state of the operating device (e.g., the height of the lifting platform), the positions of surrounding static and dynamic obstacles, and environmental conditions (e.g., simulated weather visibility). These parameters, after normalization, collectively constitute a multi-dimensional scenario state encoding parameter vector. Next, based on the command type of the logical control command, the corresponding entry is retrieved from a pre-set operation rule knowledge base. This knowledge base defines the ideal target state under different scenarios and the set of basic operations allowed to achieve the target using "if-then" rules. For example, for the "stop and wait" command, the target state is that the vehicle speed is zero and it is in a safe area; the allowed operation set only includes "braking" and "staying still." For the "path replanning" command, the target state is arriving at the designated machine position; the allowed operation set includes "accelerating," "decelerating," and "turning." The retrieved target state is quantized into a target state vector, and the allowed operations form a discrete list of actions.

[0038] Subsequently, a heuristic search is performed in an abstract state space, starting with the current scenario state encoding parameter vector and ending with the target state vector. Each state point in the state space represents a possible configuration of the task individual and the environment. The search process uses a strategy called "best-first search." At each expansion step, starting from the current state, every action in the allowed action set is attempted, simulating the state transition caused by pre-set physical constraints (such as maximum vehicle acceleration and turning radius limits) within a short time (e.g., 0.5 seconds), generating new candidate states. Simultaneously, an evaluation function (or heuristic function) calculates the "distance" from each candidate state to the target state vector. This distance is a weighted sum of differences across multiple dimensions, such as positional deviation, heading deviation, and time consumption. The search algorithm prioritizes candidate states with smaller evaluation function values ​​(i.e., closer to the target). While generating state transition paths, rigorous feasibility verification is performed, such as checking whether the new state collides with obstacles or exceeds the task boundary. Any action branch leading to an infeasible state is immediately pruned. Finally, the search algorithm outputs a state transition path from the starting point to the ending point, consisting of a series of discrete state points, which is the preliminary candidate state transition sequence.

[0039] However, the candidate sequence is discrete in time and may not be smooth. Therefore, time warping and smooth interpolation are required. Time warping first estimates a reasonable duration for each state transition in the sequence, calculated based on vehicle dynamics constraints, such as the shortest time required to accelerate from speed A to speed B. Then, the time axis of the entire path is uniformly resampled at a fixed high frequency (e.g., 100 times per second). For each new sampling time point, a new, smoothly transitioning state value (including position, speed, heading, etc.) is calculated using cubic Hermitian interpolation based on its neighboring original discrete state points and their corresponding speed and acceleration information. After this processing, the final output is a temporally continuous and kinematically smooth sequence of continuous actions (describing "what to do," such as steering wheel angle and throttle depth at each moment) and a sequence of state changes (describing "what state," such as coordinates and speed at each moment).

[0040] Secondly, the system synchronously drives the virtual environment to perform dual-path simulations and records the differences. While generating a new sequence, the system initiates a parallel simulation process. One simulation thread strictly follows the newly generated continuous action sequence and state change sequence to drive the work individuals and related environmental elements in the virtual environment. The other simulation thread synchronously replays the stored historical original sequence of the actual event. The two simulation processes are strictly time-synchronized. During the simulation, a series of predefined key state indicators are continuously monitored and recorded, such as: the minimum distance between the work individual and the nearest obstacle, the total task time, whether rule violations occurred (such as entering a restricted area), average speed, and the number of sudden braking and turning events. At the end of each path simulation, the same indicators recorded for the two paths are compared item by item, and their differences are calculated. For example, "the total task time of the new path increased by 120 seconds compared to the historical path," and "the minimum distance to the bridge in the new path is consistently maintained above 3 meters, while the minimum distance in the historical path is 0.5 meters (a collision occurred)."

[0041] Finally, a split-screen comparison display and source tracing analysis report are generated. After the simulation, a split-screen display is presented in an integrated view: one side plays a recording of the complete simulation process of the user's selected path, while the other side simultaneously plays a recording of the original historical path. Key time points and differences in status indicators are highlighted with floating labels. Based on the recorded differences in status indicators, quantitative analysis is performed, and a final source tracing analysis report is generated. The first part of the report assesses the quality of individual decisions: it analyzes the performance of the user's selected operations in multiple dimensions such as safety (e.g., the degree of collision risk reduction), compliance (number of rule violations), and efficiency (increased time consumption), and provides a comprehensive score. The second part of the report generates specific optimization suggestions for the collaborative instruction logic: for example, if the analysis finds that the "stop waiting" strategy can completely avoid accidents but significantly increase time consumption, the report may suggest optimizing the instruction logic to "during peak periods when flight delays cause overlap between the X and Y gate paths, the scheduling instruction should be forcibly supplemented with a minimum time interval of Z seconds"; if the "path replanning" strategy can significantly improve the safety distance with an acceptable efficiency loss, it suggests optimizing the instruction logic to "when the system detects a path conflict, it should automatically provide the relevant vehicles with at least one safe alternative path suggestion verified by conflict detection." These recommendations are presented in the form of quantified parameters (such as minimum time interval Z seconds and safe distance threshold), directly pointing to the optimizable points of the cooperative instruction system.

[0042] The working principle of this invention is as follows: First, time-series data of individual behaviors reflecting individual operations and time-series data of collaborative instructions reflecting central control logic are collected in parallel in the target operation scenario, forming a heterogeneous time-series data stream. Next, the two types of data are timestamped and then combined by calculating dynamic correlation and adaptive weighted fusion to generate joint representation data that can characterize the evolution of behavior under the instruction background. Then, by setting a multi-scale sliding window on the joint representation data and calculating the bidirectional guidance strength, the system-induced behavior nodes induced by the collaborative instruction logic are identified and labeled. Furthermore, historical behaviors and instruction processes are synchronously reconstructed in a virtual simulation environment, and the highly correlated induced nodes and source instructions are presented by generating a dynamic visual connector, constructing an interactive training scenario. Finally, branch selection is set at the induced nodes of this scenario, different causal paths are deduced in real time according to user operations, the decision quality is evaluated by comparing the path results, and a source analysis report containing quantitative optimization suggestions for collaborative instruction logic is generated.

[0043] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for constructing safety training scenarios based on behavior tracing, characterized in that, Includes the following steps: S1: Collect heterogeneous time-series data streams in the target operation scenario, wherein the heterogeneous time-series data streams include individual behavior time-series data and collaborative instruction time-series data; S2: Perform timestamp alignment and cross-modal joint encoding on individual behavior time series data and collaborative instruction time series data to generate joint representation data that can characterize the evolution of individual behavior patterns in the context of behavior corresponding to collaborative instruction time series data; S3: By analyzing the temporal dependencies and causal relationships between the temporal data of collaborative instructions and the temporal data of individual behaviors in the joint representation data, the system-induced behavior nodes induced by the collaborative instruction logic are identified and labeled. S4: Based on the information of system-induced behavior nodes, the behavior process and collaborative instruction process of historical events are synchronously reconstructed in a virtual simulation environment. The system-induced behavior nodes and the time sequence data of the collaborative instructions that trigger the corresponding nodes are presented with high correlation, forming an interactive training scenario that can dynamically reveal the causal chain between system instruction pressure and individual behavioral decisions. S5: Based on interactive training scenarios, different coping strategies are set at the system-induced behavior nodes. The system deduces along different causal paths according to the user operation-driven scenario, compares and displays the results of each path, and finally generates a comprehensive evaluation of individual decision quality and includes a source analysis report that includes suggestions for optimizing the collaborative instruction logic.

2. The method for constructing a safety training scenario based on behavior tracing according to claim 1, characterized in that, S2 specifically includes: Based on a unified time base, the time-series data of individual behaviors and the time-series data of collaborative instructions are timestamped to obtain time-aligned individual behavior time-series data streams and collaborative instruction time-series data streams; For time-aligned individual behavior time-series data streams and collaborative instruction time-series data streams, calculate the dynamic correlation between individual behavior time-series data and collaborative instruction time-series data at each time point; Based on the dynamic correlation, the individual behavior time-series data stream and the collaborative instruction time-series data stream are adaptively weighted and fused to generate joint representation data.

3. The method for constructing a safety training scenario based on behavior tracing according to claim 2, characterized in that, The time-aligned individual behavior time-series data stream and collaborative instruction time-series data stream are used to calculate the dynamic correlation between the individual behavior time-series data and the collaborative instruction time-series data point by point, specifically including: For each aligned time point, the cooperative instruction time series data is mapped to the standard state space in which the individual behavior time series data should be located, and the conditional probability density of the actual observations of the individual behavior time series data in the standard state space is calculated. By comparing the conditional probability density at the current time point with the conditional probability density at the previous time point, the probability density transition of individual behavior time-series data state caused by changes in cooperative instruction time-series data is calculated. By coupling the probability density transfer quantity with the state change quantity of the individual behavior time series data at the corresponding time point, a dynamic correlation degree value representing the intensity of the immediate influence of the collaborative instruction time series data on the individual behavior time series data is generated.

4. The method for constructing a safety training scenario based on behavior tracing according to claim 1, characterized in that, S3 specifically includes: Multiple sliding analysis windows with different time lengths are set up on the joint characterization data; Within each sliding analysis window, the bidirectional guiding strength between the changing trends of collaborative instruction time-series data and individual behavior time-series data is calculated. Based on the bidirectional guidance intensity calculated under all sliding analysis windows, if and only if, within a set number of windows, the guidance intensity of instruction change preceding behavior change is consistently higher than the reverse guidance intensity and exceeds a set threshold, the individual behavior time series data of the corresponding time interval will be marked as a system-induced behavior node.

5. The method for constructing a safety training scenario based on behavior tracing according to claim 4, characterized in that, The bidirectional guiding strength between the changing trends of the computational collaborative instruction time-series data and the changing trends of the individual behavior time-series data specifically includes: The collaborative instruction time series data and individual behavior time series data within the window are decomposed into independent components that change continuously on the time axis; For each independent component of the collaborative instruction time series data, the collaborative instruction time series data is paired with all independent components of the individual behavior time series data, and the lead-lag correlation coefficients of the change sequence of the instruction component relative to the change sequence of the behavior component are calculated under all possible permutations and combinations. Statistical analysis was performed on all the lead-lag correlation coefficients obtained from the pairwise calculations, and the average correlation coefficients were extracted for the case where the change in instruction preceded the change in behavior, and the average correlation coefficients for the case where the change in behavior preceded the change in instruction. The average correlation coefficients of the two directions are multiplied by the normalized weight of the number of permutations and combinations of the corresponding directions to obtain the positive guidance strength value and the negative guidance strength value, which represent the causal possibility, and are used as the bidirectional guidance strength.

6. The method for constructing a safety training scenario based on behavior tracing according to claim 1, characterized in that, S4 specifically includes: Establish a unified spatiotemporal coordinate system and map the geometric trajectory of the reconstructed behavioral process and the time stamp of the cooperative instruction process to the spatiotemporal coordinate system; On the time axis of the spatiotemporal coordinate system, identify the start time of the system-induced behavior node, and extract the instruction point that is closest to the corresponding time and whose content has undergone key changes from the collaborative instruction process; Starting from the instruction point and ending at the corresponding system-induced behavior node, a visual connection body that dynamically changes over time is generated and rendered in the spatiotemporal coordinate system; By overlaying and fusing visual connectors with behavioral trajectories and instruction markers in a spatiotemporal coordinate system, the core of visualization for an immersive interactive environment is formed.

7. The method for constructing a safety training scenario based on behavior tracing according to claim 6, characterized in that, The generation process of the visual connective is as follows: Based on the temporal location of the instruction point and the spatial location of the corresponding behavior node, a set of path control points that evolve over time are calculated and defined. Using the cubic spline interpolation method, the path control points are smoothly fitted to generate a continuous spatiotemporal path curve, and the instantaneous radius of curvature at each moment is calculated as the dynamic curvature along the spatiotemporal path curve. Establish a nonlinear mapping relationship between dynamic curvature and pre-calculated guidance intensity, and dynamically determine the width and color brightness of the visual connector at each moment based on the nonlinear mapping relationship; In the spatiotemporal coordinate system, along the spatiotemporal path curve, based on the width and color brightness determined at each moment, a semi-transparent strip-shaped geometric body is rendered and drawn in real time as a visual connector.

8. The method for constructing a safety training scenario based on behavior tracing according to claim 1, characterized in that, S5 specifically includes: Capture the branch choices made by the user at the system-guided behavior nodes and convert the branch choices into corresponding logical control instructions; Based on logical control instructions, and combined with preset work rules and physical constraints, a continuous sequence of actions and state changes is generated in real time from the corresponding node until the task ends, based on the current state of the interactive training scenario. The virtual simulation environment is synchronously driven to extrapolate along the newly generated sequence in real time and the stored original historical sequence, and the differences in key state indicators on the two paths are recorded. The process records and results of the two path deductions are compared and displayed in a split-screen view on an integrated view. The differences in status indicators are quantified into specific optimization parameters for collaborative instruction logic and integrated into the source tracing analysis report.

9. The method for constructing a safety training scenario based on behavior tracing according to claim 8, characterized in that, The generation process of the continuous action sequence and state change sequence specifically includes: The interactive training scenario is encoded at the system-induced behavior node, generating scenario state encoding parameters that include environmental parameters and individual pose. Based on the response strategy indicated by the logic control instructions, the corresponding target state vector and the set of allowed action operations are extracted from the preset operation rule base; Starting with the scenario state coding parameters and ending with the target state vector, a heuristic search is used in the state space, and the feasibility of the searched state transition sequence is verified and pruned in combination with physical constraints to generate a preliminary candidate state transition sequence. Perform time warping and smooth interpolation on the candidate state transition sequence to output a continuous sequence of actions and state changes that are continuous in time from the corresponding node until the end of the task.

Images