Opencast coal mine VR safety training dynamic difficulty regulation and control method and system

Abstract

The invention provides an opencast coal mine VR safety training dynamic difficulty regulation and control method, which comprises the steps of collecting multi-dimensional interaction behavior data of a target student in a VR training scene in real time, and encoding the multi-dimensional interaction behavior data into a behavior vector sequence; inputting the behavior vector sequence into a pre-trained prediction model, and outputting a potential risk behavior and a risk occurrence probability; when the risk occurrence probability is higher than a preset threshold value, generating a risk event continuously caused by the potential risk behavior through a risk deduction model; dynamically generating a negative guide plot based on the risk event, and implanting the negative guide plot into the VR training scene; and dynamically adjusting a subsequent generation strategy of the negative guide plot according to a response result of the target student to the negative guide plot. The method has the technical effects of deeply understanding the behavior intention of the student, early predicting the potential risk and performing dynamic intervention.

Description

Technical Field

[0001] This invention relates to the field of safety training technology for open-pit coal mine operations, specifically to a method and system for dynamically adjusting the difficulty of VR safety training in open-pit coal mines. Background Technology

[0002] Open-pit coal mines operate in complex environments with numerous hazards. Utilizing VR technology to simulate scenarios such as slope inspection, mining truck transportation, and equipment operation allows trainees to familiarize themselves with procedures, identify potential hazards, and practice emergency response in a risk-free environment. Existing VR safety training systems already exist, and attempts are being made to introduce dynamic difficulty adjustment mechanisms to enhance training effectiveness.

[0003] Existing dynamic difficulty control technologies mainly follow two modes: The first is the preset script + fixed rule mode, where the system has built-in fixed training scripts for multiple difficulty levels, and trainees unlock the next difficulty level after passing the assessment; or static trigger conditions are set in a single scenario, such as entering a specific area to determine a violation. This method has extremely poor flexibility, makes it difficult to respond to the trainees' real-time status, the training process is rigid, and there is a lack of novelty when repeating training, which can easily make trainees feel bored.

[0004] The second type is a linear difficulty adjustment mode based on performance results. This mode collects macro-level performance data such as task completion time, number of errors, and point deductions, and adjusts the parameters of subsequent tasks according to preset linear rules based on these outcome indicators. However, this method only passively adjusts the difficulty after the student has made a clear mistake or even failed the task. This feedback is delayed, and the adjustment is based on macro-level data, making it difficult to reflect subtle behavioral signs before the student makes a mistake. For example, if a student is driving a mining truck and continuously drives close to the edge of the road (a critical safety behavior), the system will not intervene as long as the solid line is not crossed, until a line crossing or collision occurs. This is essentially a response to "already happened," rather than a prediction of "potential risks." It cannot predict the dangers the student may encounter in the next few steps or seconds based on their current operating habits, and therefore cannot implement proactive, preventative interventions.

[0005] In summary, existing technologies have significant shortcomings in terms of real-time and precision intervention, as well as risk predictability. Therefore, there is an urgent need for a dynamic difficulty control method and system for VR safety training in open-pit coal mines that can deeply understand trainees' behavioral intentions, predict potential risks early, and dynamically intervene. Summary of the Invention

[0006] In view of the aforementioned problems, this application is made to provide a method and system for dynamic difficulty control in VR safety training for open-pit coal mines that overcomes or at least partially solves the aforementioned problems, comprising: A method for dynamically adjusting the difficulty of VR safety training in open-pit coal mines includes the following steps: Real-time collection of multi-dimensional interactive behavior data of target learners in VR training scenarios, and encoding of the multi-dimensional interactive behavior data into a sequence of behavior vectors; The sequence of behavioral vectors is input into a pre-trained prediction model, which outputs potential risky behaviors and the probability of risk occurrence. When the probability of the risk occurring is higher than a preset threshold, a risk event that the potential risk behavior will continue to cause is generated through a risk inference model. Based on the aforementioned risk events, negative guidance scenarios are dynamically generated and associated with the current VR training scenario, and these negative guidance scenarios are then embedded into the VR training scenario. Based on the target learners' responses to the negative guidance scenarios, the generation strategy for subsequent negative guidance scenarios is dynamically adjusted.

[0007] Furthermore, the step of inputting the behavior vector sequence into a pre-trained prediction model and outputting the potential risk behavior and the probability of risk occurrence specifically includes: The behavior vector sequence is matched in real time with a baseline pattern in a preset non-optimal behavior pattern library. The baseline pattern defines a pattern template consisting of compliant but inefficient behaviors and high-risk behavior fragments. Calculate the dynamic time warped distance between the behavior vector sequence and the baseline pattern, and output the baseline pattern with the highest matching degree; The semantic label corresponding to the benchmark pattern with the highest matching degree is defined as a potential risky behavior; Calculate the probability that the potential risky behavior will evolve into a safety accident within a preset time window in the future.

[0008] Furthermore, the step of generating risk events that will result from the continued occurrence of the potential risky behavior through a risk extrapolation model when the probability of the risk occurring is higher than a preset threshold specifically includes: Using the potential risky behavior as the initial condition and the instantaneous state of the current VR training scenario as the initial environment for deduction, the risk deduction model is deduced forward along the time axis to output a first-level consequence event. Iteratively deduce the chain reaction triggered by the first-level consequence event until the risk event is generated; Output the event type, involved parties, severity level, and evolution path corresponding to the risk event.

[0009] Furthermore, the step of dynamically generating negative guidance scenarios related to the current VR training scenario based on the risk event, and embedding the negative guidance scenarios into the VR training scenario, specifically includes: Based on the event type, retrieve basic plot modules that conform to the physical rules and environmental constraints of the current VR training scenario from the preset plot element library; Using the objects involved as interactive entities, the intermediate links in the evolution path that can be externally intervened or occur naturally are instantiated as dynamic environmental interference events; Based on the basic plot module, the spatiotemporal parameters of the dynamic environmental interference event are dynamically adapted and calibrated to generate the negative guidance plot; Without interrupting the main task flow, the negative guidance plot is implanted into the current VR training scene in real time.

[0010] Furthermore, the step of dynamically adjusting the generation strategy of subsequent negative guidance scenarios based on the target learner's response to the negative guidance scenario specifically includes: Data on the target learner's response to the negative guidance scenario is obtained, and a response outcome evaluation vector is generated. The evaluation vector includes response delay time, operational accuracy, and scenario resolution efficiency. The response result evaluation vector is compared with the preset expected response standard, and the weakness weight of the target student is generated based on the comparison result. The matching similarity threshold of the corresponding benchmark pattern in the prediction model is adjusted based on the weight of the weak point; The triggering frequency of the negative guidance plot is adjusted based on the response result evaluation vector; The complexity of the negative guidance plot is dynamically adjusted based on the changing trend of the evaluation vector according to the historical response results.

[0011] Furthermore, the step of adjusting the triggering frequency of the negative guidance plot based on the response result evaluation vector specifically includes: Set the base trigger interval And the frequency modulation factor λ, the triggering wait time for the next negative guiding plot of the same type is ; When the accuracy of the operation is lower than the preset threshold, λ is set to a value greater than 1, and the trigger waiting time for the next negative guidance plot of the same type is output. When both the operation accuracy and the plot resolution efficiency are higher than the preset threshold, λ is set to a positive value less than 1, and the trigger waiting time for the next negative guidance plot of the same type is output.

[0012] Furthermore, the step of dynamically adjusting the complexity of the negative guidance plot based on the changing trend of the evaluation vector according to historical response results specifically includes: The statistical analysis of the assessment vectors of several responses to the same type of potential risk behavior within a preset period forms a historical sequence; Analyze the evolution trends of reaction delay time and operational accuracy in the historical sequence. When both show a continuous improvement trend, initiate a complexity upgrade. The complexity upgrade is achieved by combining and superimposing challenge increments of different dimensions, specifically including: superimposing secondary environmental interference while maintaining the core challenge unchanged; shortening the decision-making time window from the occurrence of the plot to the necessity of taking effective action; and reducing the explicitness of key hints in the plot. When the historical sequence shows that the accuracy of the operation drops significantly after the complexity level is increased, it automatically reverts to one complexity level.

[0013] A VR safety training system for open-pit coal mines with dynamic difficulty adjustment includes: The interactive behavior acquisition module is used to collect multi-dimensional interactive behavior data of the target trainees in the VR training scenario in real time, and encode the multi-dimensional interactive behavior data into a behavior vector sequence. The time-series risk prediction module is used to input the behavior vector sequence into a pre-trained prediction model and output the potential risk behaviors and the probability of risk occurrence. The counterfactual risk simulation module is used to generate risk events that will be caused by the continued occurrence of the potential risky behavior when the probability of the risk occurring is higher than a preset threshold, through a risk simulation model. The negative plot injection module is used to dynamically generate negative guidance plots associated with the current VR training scene based on the risk event, and to inject the negative guidance plots into the VR training scene; The dynamic difficulty adjustment module is used to dynamically adjust the generation strategy of subsequent negative guidance scenarios based on the target learners' response to the negative guidance scenarios.

[0014] A computer electronic device includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the computer program is executed by the processor, it implements the steps of the above-described method for dynamic difficulty adjustment of VR safety training in open-pit coal mines.

[0015] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a method for dynamic difficulty control in VR safety training for open-pit coal mines as described above.

[0016] This application has the following advantages: In the embodiments of this application, addressing the core deficiency of existing dynamic difficulty control technologies that only passively adjust the difficulty after a trainee has made a clear mistake or even failed the task, making it difficult to reflect the subtle behavioral signs before the trainee makes a mistake, this application provides a dynamic difficulty control method for VR safety training in open-pit coal mines. Specifically, it involves: real-time collection of multi-dimensional interactive behavior data of the target trainee in the VR training scene, and encoding the multi-dimensional interactive behavior data into a behavior vector sequence; inputting the behavior vector sequence into a pre-trained prediction model to output potential risk behaviors and the probability of risk occurrence; when the probability of risk occurrence is higher than a preset threshold, generating risk events that the potential risk behaviors will continue to cause through a risk inference model; dynamically generating negative guidance plots associated with the current VR training scene based on the risk events, and embedding the negative guidance plots into the VR training scene; and dynamically adjusting the generation strategy of subsequent negative guidance plots according to the target trainee's response to the negative guidance plots. By encoding multidimensional behavioral data into time-series vector sequences and performing model-based prediction, this approach overcomes the limitations of existing technologies, which can only respond to significant violations and cannot identify suboptimal behavioral sequences with cumulative risks. This achieves the technical effect of proactively identifying risk nascents and enabling preventative training. Furthermore, by dynamically generating natural environmental challenges strongly correlated with trainees' own behavioral logic, this approach forces trainees to face higher operational difficulties due to their suboptimal behaviors. This addresses the shortcomings of existing technologies, such as rigid feedback methods that rely heavily on interruptions and lecturing, severely damaging immersion and resulting in superficial teaching effects. It enables trainees to learn about the potential consequences of their own behaviors in a highly immersive and inspiring way without interrupting the main process. Finally, by dynamically adjusting personalized strategies based on response outcomes, this approach overcomes the limitations of existing technologies, which offer linear and uniform strategy adjustments and cannot precisely target individual trainees' weaknesses. This achieves the effect of targeted reinforcement of individual weaknesses and optimized training efficiency. Attached Figure Description

[0017] To more clearly illustrate the technical solution of this application, the drawings used in the description of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the steps of a method for dynamically adjusting the difficulty of VR safety training in open-pit coal mines, as provided in one embodiment of this application. Figure 2 This is a structural block diagram of a VR safety training dynamic difficulty control system for open-pit coal mines provided in one embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer electronic device provided in an embodiment of the present invention; 1. Computer electronic device; 2. External device; 3. Processing unit; 4. Bus; 5. Network adapter; 6. I / O interface; 7. Display; 8. Memory; 9. Random access memory; 10. Cache memory; 11. Storage system; 12. Program / utility; 13. Program module. Detailed Implementation

[0019] To make the objectives, features, and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0020] The inventors, through analysis of existing technologies, discovered that: 1. Fixed training scripts with multiple difficulty levels, where trainees unlock the next difficulty level after passing an assessment, are difficult to respond to based on the trainees' real-time status, resulting in a rigid training process; 2. Linear difficulty adjustment modes collect macro-level performance data such as trainees' task completion time, number of errors, and deductions, and adjust the parameters of subsequent tasks according to preset linear rules based on these outcome indicators. This method only passively adjusts the difficulty after trainees have made clear mistakes or even failed the task, making it difficult to reflect the subtle behavioral signs before trainees make mistakes.

[0021] Reference Figure 1 This application illustrates an embodiment of a method for dynamically adjusting the difficulty of VR safety training in open-pit coal mines, comprising the following steps: S110. Real-time collection of multi-dimensional interactive behavior data of target trainees in VR training scenarios, and encoding the multi-dimensional interactive behavior data into a behavior vector sequence; S120. Input the behavior vector sequence into the pre-trained prediction model and output the potential risk behavior and the probability of risk occurrence. S130. When the probability of the risk occurring is higher than a preset threshold, a risk event that the potential risk behavior will continue to cause is generated through a risk inference model. S140. Based on the risk event, dynamically generate a negative guidance plot associated with the current VR training scenario, and embed the negative guidance plot into the VR training scenario; S150. Based on the target learner's response to the negative guidance scenario, dynamically adjust the generation strategy for subsequent negative guidance scenarios.

[0022] In the embodiments of this application, addressing the core deficiency of existing dynamic difficulty control technologies that only passively adjust difficulty after a trainee has made a clear mistake or even failed a task, making it difficult to reflect the subtle behavioral signs before a trainee makes a mistake, this application provides a dynamic difficulty control method for VR safety training in open-pit coal mines. By encoding multidimensional behavioral data into time-series vector sequences and performing model-based prediction, this method solves the shortcomings of existing technologies that can only respond to significant violations and cannot identify non-optimal behavioral sequences with cumulative risks, achieving the technical effect of proactively identifying risk buds and realizing preventive training. By dynamically generating natural environmental challenges that are strongly correlated with the trainee's own behavioral logic, the method forces trainees to face higher operational difficulties due to their non-optimal behavior, solving the shortcomings of existing technologies that provide rigid feedback, rely mainly on interruptions and lecturing, severely damage the immersion, and have superficial teaching effects. This method enables trainees to learn about the possible consequences of their own behavior in a highly immersive and inspiring way without interrupting the main process. Through dynamic control of personalized strategies based on response results, this method solves the shortcomings of existing technologies that adjust strategies linearly and uniformly and cannot accurately strengthen individual trainees' weaknesses, achieving the effect of targeted strengthening of individual weaknesses and optimizing training efficiency.

[0023] The following will further explain a method for dynamically adjusting the difficulty of VR safety training in open-pit coal mines in this exemplary embodiment.

[0024] In one embodiment of the present invention, the "multi-dimensional interactive behavior data" mentioned in step S110 includes at least two of the following: spatial displacement trajectory, gaze focus sequence, operation sequence and interval duration with virtual device, and speed / acceleration control curve. The specific process of "real-time collection of multi-dimensional interactive behavior data of target trainees in VR training scenario and encoding the multi-dimensional interactive behavior data into behavior vector sequence" in step S110 can be further explained in conjunction with the following description.

[0025] As described in the following steps The system collects multi-dimensional interactive behavior data of target trainees in VR training scenarios in real time at a preset sampling frequency, and performs noise reduction and normalization preprocessing on the collected raw data. A non-uniform sampling strategy combining event triggering and fixed-period sampling is adopted, and recording is only performed when the interaction state changes significantly or when the period point is reached. The processed data is then encapsulated into structured data units. Based on preset feature extraction rules, key features are extracted from each structured data unit. The key features include at least behavior type, quantization parameters, and context labels. The key features corresponding to each data unit are arranged in chronological order and combined into the behavior vector sequence, where each vector is associated with a precise timestamp, and the interval between vectors in the sequence reflects the dynamic sparsity of the interaction.

[0026] It should be noted that multi-dimensional interactive behavior data is a dataset acquired from different perception and interaction channels, comprehensively depicting the trainee's state and intentions in the virtual environment. Spatial displacement trajectory refers to the continuous change in the trainee's position in the three-dimensional virtual space, typically calculated from the pose data of a head-mounted display (HMD) and / or hand controllers, used to analyze movement paths, trends towards hazards, etc. Eye focus sequence refers to the trainee's visual attention points in the virtual scene and the order in which they change over time, usually acquired through eye-tracking technology, used to determine whether they are paying attention to key instruments, warning signs, or potential risk areas. "Sequence and interval of operation with virtual equipment" refers to the order in which the trainee interacts with virtual simulation equipment (such as console buttons, valves, and cab controls) and the time interval between adjacent operations, used to assess the standardization, proficiency, and presence of hesitation or haste in their operational procedures. Speed / acceleration control curves specifically refer to the continuous description of the rate of motion and its rate of change during equipment movement or manipulation, such as the travel speed of a mining truck or the lifting acceleration of a bucket, used to identify aggressive, unstable, or slow-responding operating patterns. Non-uniform sampling is an adaptive data recording method that differs from fixed-interval sampling. Its core is driven by interactive events (such as button presses, gaze timeouts, and sudden speed changes), supplemented by a minimum maximum sampling period. This ensures no redundant data is generated during low-activity periods and captures key details during high-activity periods, thus efficiently generating a sparse but information-dense data stream. A "structured data unit" is a preprocessed, packaged, and formatted data packet containing multi-dimensional raw data collected at a specific moment or short time interval. Each unit can be considered a complete "behavioral snapshot." A behavioral vector sequence is a sequence of feature vectors ordered by timestamps. Each feature vector represents a set of key features at a specific moment. The sparsity of the sequence is reflected in the non-uniform distribution of vectors along the time axis, corresponding to the non-uniform sampling strategy and accurately reflecting the real rhythm of alternating activity and stillness in human-computer interaction.

[0027] As an example, spatial displacement trajectories are acquired using an inertial measurement unit (IMU) and an infrared positioning system integrated into the VR headset. Eye-tracking modules integrated into the VR headset acquire gaze focus sequences. Operation sequence and interval durations are obtained through button / touch event logs and haptic feedback response records from the VR controllers. Velocity / acceleration control curves are generated by numerically differentiating the displacement trajectory or by directly reading physical parameters provided by the virtual simulation engine.

[0028] In one specific implementation, the VR system reads raw sensor data from the HMD and controllers at a frequency of 90Hz. A Kalman filter is used to denoise the positional data, and a median filter is used to smooth jitter in the eye-tracking data. Subsequently, all numerical data (such as coordinates and velocity) are normalized to the [0,1] interval to eliminate the influence of dimensions. A base period T (e.g., 500ms) is set, and a set of event triggers are defined, such as displacement changes exceeding a threshold, the gaze first focusing on a key object, and any button being pressed. Whenever any trigger event occurs, or the period T is reached, all currently preprocessed multi-dimensional data, the gaze target ID, the most recent operation event, and the current velocity are packaged into a structured data unit with a millisecond-level timestamp. For each structured data unit, a feature mapping table is created for parsing. For example, if the gaze target ID in the data unit maps to a pressure gauge in the virtual scene, the extracted behavior type is GAUGE_INSPECTION; the quantization parameter is the duration of this gaze, calculated by the difference in timestamps between this duration and the previous data unit containing the same target ID; the context label is CONTEXT_HAUL_ROAD_DRIVING (based on location information, the learner is in a road driving scenario). The extracted features are then transformed into a fixed-dimensional numerical vector. Arranging all the feature vectors generated in chronological order forms a sparsely distributed (because sampling is only done at event or periodic points) behavior vector sequence, which serves as the direct input to the subsequent risk prediction model.

[0029] In one embodiment of the present invention, the specific process of step S120, "inputting the behavior vector sequence into a pre-trained prediction model and outputting potential risk behaviors and the probability of risk occurrence," can be further explained in conjunction with the following description.

[0030] As described in the following steps The behavior vector sequence is matched in real time with a baseline pattern in a preset non-optimal behavior pattern library. The baseline pattern defines a pattern template consisting of compliant but inefficient behaviors and high-risk behavior fragments. Calculate the dynamic time warped distance between the behavior vector sequence and the baseline pattern, and output the baseline pattern with the highest matching degree; The semantic label corresponding to the benchmark pattern with the highest matching degree is defined as a potential risky behavior; Calculate the probability that the potential risky behavior will evolve into a safety accident within a preset time window in the future.

[0031] It's important to note that the pre-trained prediction model is a machine learning model capable of identifying the mapping relationship between behavioral sequences and risks. Its core is learning the mapping function from "non-optimal behavioral patterns" to "accident probabilities," rather than simply matching fixed rules. The non-optimal behavioral pattern library is a structured knowledge base that stores a series of compliant but inefficient or potentially high-risk behavioral sequence templates—i.e., baseline patterns—extracted from expert experience and accident analysis. Examples include "a sequence where the line of sight stays on critical equipment for less than 50% of the normal value during inspection" and "a sequence where a mining truck continuously travels at 85%-95% of the speed limit on a curve." Dynamic time warping distance is an algorithm used to measure the similarity between two time-series sequences that may be generated at different speeds. It allows sequences to be stretched or compressed on the time axis to find the optimal alignment, effectively matching the duration variations caused by differences in proficiency when different trainees perform the same type of behavior. Semantic labels are textual descriptions with clear safety meanings assigned to each baseline pattern, such as "distracted inspection" and "aggressive edge driving," transforming machine-recognized patterns into human-understandable qualitative risk descriptions. The preset time window defines the predicted future time range. The probability of risk occurrence represents the expected mathematical likelihood of a definable safety incident (such as a collision or fall) occurring within the specified time window, assuming the current behavior pattern continues. The probability calculation integrates pattern matching degree, current environmental state, and historical statistical patterns.

[0032] As an example, the training process of the predictive model is a data-driven supervised learning process, the core of which lies in constructing a high-quality dataset of risk-labeled behavioral sequences. Specifically, in the offline phase, a large number of trainees (including novices, experienced personnel, and experts) are organized to operate in an open-pit coal mine VR training scenario covering various typical risks. Multi-dimensional interactive behavioral data, completely consistent with the online phase, is recorded throughout and encoded into historical behavioral vector sequences. These behavioral sequences are then replayed by coal mine safety experts and training instructors. Experts not only mark sequences that have led to virtual accidents (such as collisions and falls), but more importantly, they identify and label behavioral segments that, while not directly causing accidents, exhibit "compliant but inefficient" or "potentially risky" tendencies. For example, labeling a sequence where "when inspecting equipment in a noisy environment, the gaze lingers on key instruments for less than half the standard time." Each labeled sequence is assigned two types of labels: a semantic label, corresponding to the template name in the non-optimal behavior pattern library, such as "attention-distracted inspection"; and a risk probability label, a tuple (P, T), where P is the confidence level (between 0 and 1) of the behavior pattern, as assessed by experts based on experience, that it will trigger a definable incident within a specific time window T (e.g., 30 seconds). A neural network model suitable for processing time-series data is employed, such as a Long Short-Term Memory (LSTM) network, a Gated Recurrent Unit (GRU), or a Temporal Convolutional Network (TCN). The model input is a variable-length sequence of behavior vectors, and the output consists of two parallel branches: a pattern classification branch, which outputs the matching degree between the sequence and each baseline pattern in the non-optimal behavior pattern library; and a risk regression branch, which outputs a continuous value representing the probability of risk occurrence. The loss function consists of two weighted parts: a classification loss, using the cross-entropy loss function to ensure the model accurately identifies the non-optimal pattern category to which the behavior sequence belongs; and a regression loss, using the mean squared error (MSE) or smoothed L1 loss function to ensure that the risk probability value output by the model is as close as possible to the probability label P assigned by the expert. The model is trained using a prepared labeled dataset, and the model parameters are iteratively adjusted through backpropagation and optimization algorithms (such as Adam) to minimize the overall loss function. Early stopping is employed during training to prevent overfitting, and a validation set is used to monitor the model's generalization ability on unseen data. The construction of a non-optimal behavior pattern library complements model training; initial templates can be derived from descriptions of unsafe behaviors in safety procedures and incident reports. During model training, clustering algorithms (such as temporal clustering of a large number of labeled sequences) can discover new and frequently occurring non-optimal patterns. After expert confirmation, this pattern library is dynamically expanded and optimized, while the prediction model is updated.

[0033] In one specific implementation, the prediction model first performs dynamic time warping (DTW) calculations on the sparse behavior vector sequence and each "baseline pattern" in the non-optimal behavior pattern library to obtain a set of similarity scores. Assuming the highest match (smallest DTW distance) is found with the baseline pattern named "edge-hesitating driving," the semantic label "edge-hesitating driving" corresponding to this pattern is defined as the currently identified potential risk behavior. The prediction model further analyzes the tightness of the match, the stability of the current sequence (e.g., speed variance), and the virtual environment context (e.g., the current road surface is slippery). The prediction model's built-in probability calculation unit integrates these factors and outputs a quantified probability of risk occurrence, such as "the probability of a skidding or lane-crossing accident occurring within the next 20 seconds is 0.65 (65%)." This probability is compared with a preset threshold (e.g., 0.6). Since 0.65 > 0.6, the trigger condition is met, and the process proceeds to step S130.

[0034] In one embodiment of the present invention, the specific process of step S130, "when the probability of the risk occurring is higher than a preset threshold, a risk event that the potential risk behavior will continue to cause is generated through a risk inference model," can be further explained in conjunction with the following description.

[0035] As described in the following steps Using the potential risky behavior as the initial condition and the instantaneous state of the current VR training scenario as the initial environment for deduction, the risk deduction model is deduced forward along the time axis to output a first-level consequence event. Iteratively deduce the chain reaction triggered by the first-level consequence event until the risk event is generated; Output the event type, involved parties, severity level, and evolution path corresponding to the risk event.

[0036] It's important to note that a risk simulation model is a rule-based, physics engine-based, or causal graph-based simulation system used for rapid, low-cost hypothetical simulations in a virtual environment. The model's inputs are the current state and the hypothetical behavior, and its output is a possible chain of events. An instantaneous state refers to a complete snapshot of all relevant entities (students, devices, environmental parameters) in the VR scene at the moment the simulation begins, including position, speed, physical properties, and environmental conditions, serving as the deterministic initial conditions for the simulation. Forward simulation refers to simulating the development of events chronologically from the current moment, corresponding to reverse attribution analysis. A chain reaction refers to an initial consequence acting as a new "cause," triggering a cascading process of subsequent "effects." For example, "a wheel running over a loose stone" (a first-level consequence event) may lead to "a slight skid on the vehicle" (a second-level consequence event), which may then trigger "the driver overcorrecting the steering wheel" (a third-level consequence event), etc. An evolutionary path is a step-by-step description of the causal chain from potential risky behavior to the final risky event, and is key to understanding how risks are gradually realized.

[0037] As an example, the construction of a risk extrapolation model is a knowledge engineering and simulation modeling process that integrates domain knowledge, physical laws, and case data. Specifically, the first step is to establish a knowledge base and rule base: integrating or configuring the physical engine parameters used in the VR environment (such as rigid body dynamics, collision detection, and friction coefficient) to ensure that the extrapolation conforms to basic physical laws. For example, defining the impact of different road surface materials (dry, wet, loose) on tire adhesion; drawing causal chain diagrams of typical open-pit coal mine accidents (such as vehicle rollovers, slope collapses, and object strikes), where nodes represent states or events (such as speeding, wet road surfaces, and sharp turns), and edges represent causal relationships and probabilistic influences, forming the logical framework of the extrapolation; and performing structured analysis on historical real-world accident cases and "quasi-accident" cases recorded in high-fidelity VR simulations to extract key causal chains, which are then transformed into scripted scene fragments or state transition rules that can be understood by the extrapolation engine. Risk extrapolation models typically employ a hybrid architecture combining rule-based reasoning systems and lightweight physical simulations. The rule engine is responsible for activating relevant rule chains from the causal graph based on the input potential risk behaviors and the current scene state. For deductions involving complex kinematics (such as vehicle loss of control), the rule engine invokes a lightweight physics simulation thread stripped of high-fidelity rendering. Starting with the current scene state, this thread runs a physics simulation in the background for several seconds, predicting the future trajectory and state of objects (e.g., "the vehicle will contact the guardrail in 1.5 seconds"). During validation, the risk deduction model is run using numerous test scenarios with known processes and outcomes to verify whether the deduced causal chains and final events match known facts. For example, given a known sequence of driving behaviors leading to a rollover, the model is checked to see if it can deduce the rollover result. Finally, the probability parameters in the causal graph and some parameters in the physics engine (such as material properties) are adjusted to ensure that the frequency and severity distribution of risk events deduced by the model aligns with historical statistical data or expert expectations. This is an iterative optimization process.

[0038] In one specific implementation, the counterfactual risk simulation module captures the instantaneous state of the current VR scene: the mining truck's position (X,Y), speed V, steering wheel angle θ, distance to the right shoulder D, and road surface friction coefficient μ. Simultaneously, it uses the identified potential risk behavior "edge-hesitating driving" and its key parameters (such as a continuous rightward tendency) as behavioral constraints for the simulation. Assuming the student's behavior pattern remains unchanged (continuing to move closer to the right), the physics engine simulates the next 3 seconds. The simulation results show that due to the continued approach to the edge, the right rear wheel is about to run over a virtual pile of loose gravel pre-distributed randomly on the shoulder, which is a "Level 1 consequence event." The simulation continues with the physical effects of the wheel running over the gravel: the right rear wheel's adhesion drops instantly, causing the vehicle to generate a slight counter-clockwise yaw moment. Based on the current vehicle speed and the student's average reaction time, this yaw is determined to cause the rear of the vehicle to skid towards the center line of the road; this skidding event is judged as a sufficiently serious terminal risk event.

[0039] In one embodiment of the present invention, the specific process of step S140, "dynamically generating a negative guidance plot associated with the current VR training scene based on the risk event, and embedding the negative guidance plot into the VR training scene," can be further explained in conjunction with the following description.

[0040] As described in the following steps Based on the event type, retrieve basic plot modules that conform to the physical rules and environmental constraints of the current VR training scenario from the preset plot element library; Using the objects involved as interactive entities, the intermediate links in the evolution path that can be externally intervened or occur naturally are instantiated as dynamic environmental interference events; Based on the basic plot module, the spatiotemporal parameters of the dynamic environmental interference event are dynamically adapted and calibrated to generate the negative guidance plot; Without interrupting the main task flow, the negative guidance plot is implanted into the current VR training scene in real time.

[0041] It's important to note that the plot element library is a reusable collection of virtual assets and logical rules, containing various pre-made environmental objects (such as rocks, puddles, and animals), equipment states (such as flashing instruments and unusual sounds), weather effects (such as sudden fog), and their triggering and performance logic. Instantiation refers to the process of selecting specific virtual assets from the element library based on the abstract description in the deduction results, and determining their specific attributes (size, position, and timing of appearance) in the scene to create a realistic and interactive object. Dynamic adaptation and calibration are used to ensure that the generated plot seamlessly integrates with the current scene. For example, the size of the generated rocks must match the proportion of the road shoulder, the position must be a reasonable distance ahead beyond the trainee's current field of vision, and the timing of appearance must match the vehicle's speed, ensuring that the plot appears naturally rather than out of thin air. Not interrupting the main task flow is a key principle of negative guidance plot design. The insertion and development of the plot should not be done through pop-ups, forced pauses, or task resets, but rather occur naturally as part of the environment. Trainees still need to complete the original driving task, but they will need to cope with this additional challenge during the process.

[0042] As an example, a scripted event system is adopted, in which plot modules are written as dynamically invoked scripts that are injected and executed at runtime. Plot generation instructions are published through the event bus of the virtual environment engine, which are listened to and responded to by various pre-registered environmental objects in the scene (such as roads, weather systems, and virtual devices), collaboratively generating complex plots and achieving distributed, loosely coupled plot construction.

[0043] In one specific implementation, based on the "event type: vehicle skidding and evolution path," a basic plot module related to "road debris causing imbalance" is retrieved from the plot element library. This module defines the basic logic such as the type of debris (e.g., small stones) and its appearance effect (dust rising). The "running over loose road debris" step in the evolution path is instantiated. Using the predicted "position where the right rear wheel is about to run over" as a reference, a visually plausible small stone model is dynamically generated on the right side of the road about 15 meters in front of the trainee's vehicle, and a simple physical collision object is added to it. The timing of the stone's appearance is calibrated according to the current vehicle speed to ensure that the trainee has a reasonable reaction time (e.g., 2-3 seconds). At the same time, it is checked whether the generated position conflicts with other existing scene objects and fine-tuned, ultimately generating a complete set of negative guidance plot instructions. The VR rendering engine receives the instructions and loads the stone model at the specified position in the next frame rendering cycle. Since the position is outside the edge of the trainee's current field of vision, the model will naturally enter the field of vision as the vehicle moves forward, as if it were already on the road, achieving seamless integration. The trainee's primary driving duties were not interrupted, but he now faces a new challenge: he needs to react promptly because he is driving on the right.

[0044] In one embodiment of the present invention, the specific process of step S150, "dynamically adjusting the generation strategy of subsequent negative guidance scenarios based on the target learner's response to the negative guidance scenario," can be further explained in conjunction with the following description.

[0045] As described in the following steps Data on the target learner's response to the negative guidance scenario is obtained, and a response outcome evaluation vector is generated. The evaluation vector includes response delay time, operational accuracy, and scenario resolution efficiency. The response result evaluation vector is compared with the preset expected response standard, and the weakness weight of the target student is generated based on the comparison result. The matching similarity threshold of the corresponding benchmark pattern in the prediction model is adjusted based on the weight of the weak point; The triggering frequency of the negative guidance plot is adjusted based on the response result evaluation vector; The complexity of the negative guidance plot is dynamically adjusted based on the changing trend of the evaluation vector according to the historical response results.

[0046] It's important to note that the response outcome evaluation vector is a multi-dimensional, quantitative set of performance indicators designed to comprehensively and objectively assess the quality of trainees' responses to unexpected challenges, rather than a binary judgment of right or wrong. The weakness weight is a dynamic parameter in the target trainee's individual ability profile, used to quantify their vulnerability to specific types of risks (such as "insufficient spatial awareness" or "slow emergency response"). A higher weight indicates a greater need for intensive training in that area. The matching similarity threshold is an adjustable parameter in the prediction model, used to control the detection sensitivity to specific suboptimal behaviors. Lowering the threshold makes detection more sensitive (easier to trigger intervention), while raising the threshold is more lenient. The generation strategy includes a series of decision rules such as when to generate (trigger frequency), what level of difficulty (complexity) to generate, and what risks to target (achieved by adjusting the detection threshold).

[0047] As an example, a multi-objective decision-making framework is used to balance different evaluation metrics (such as safety and efficiency) and comprehensively derive adjustment strategies. Reinforcement learning algorithms are applied, treating the system as an intelligent agent, using the long-term improvement of learners' abilities as rewards to automatically optimize and generate policy parameters. A case-based personalized recommendation approach is adopted to select the most effective strategy adjustment scheme for learners with similar weaknesses.

[0048] In one specific implementation, when faced with an approaching rock, the trainee swerved sharply to the left to avoid it. The vehicle swayed slightly but did not veer off the road, and then corrected its course. The dynamic difficulty control module recorded: reaction delay time: 1.2 seconds (from the rock entering the main field of view to the start of the turn); operation accuracy: 0.7 (successful avoidance but clumsy action, points deducted); scenario resolution efficiency: 0.9 (quickly resolved the crisis without causing task failure); combined into an evaluation vector [1.2, 0.7, 0.9]. Comparing the evaluation vector with the expected standard [<1.0s, >0.8, >0.85] showed that the operation accuracy was below standard, thus increasing the weight of the weak item "emergency operation precision" in the trainee's personal profile. Since emergency operation precision is related to edge driving risk, the matching similarity threshold for the baseline pattern of "edge-driving" in the prediction model was slightly reduced from 0.6 to 0.58, which means that it may also provide an early warning when the trainee exhibits a slight edge driving tendency in the future. Based on the operational accuracy score of 0.7 (below the threshold of 0.8), it was determined that the response needed improvement. Therefore, a frequency penalty strategy was applied to shorten the triggering wait time for the next similar negative guidance scenario related to edge driving. Reviewing the trainee's historical evaluation vectors from the last three spatial judgment challenges revealed that while reaction delays were decreasing, operational accuracy fluctuated significantly. Therefore, it was decided not to increase the complexity of the next scenario, but rather maintain the current difficulty level to reinforce operational standardization.

[0049] In one embodiment of the present invention, the specific process of "adjusting the triggering frequency of the negative guidance plot according to the response result evaluation vector" can be further explained in conjunction with the following description.

[0050] As described in the following steps Set the base trigger interval And the frequency modulation factor λ, the triggering wait time for the next negative guiding plot of the same type is ; When the accuracy of the operation is lower than the preset threshold, λ is set to a value greater than 1, and the trigger waiting time for the next negative guidance plot of the same type is output. When both the operation accuracy and the plot resolution efficiency are higher than the preset threshold, λ is set to a positive value less than 1, and the trigger waiting time for the next negative guidance plot of the same type is output.

[0051] It should be noted that the base trigger interval It is a preset and standardized training interval for specific types of potential risk events, representing the default time interval or task progress interval between two similar reinforcement trainings when the trainee's performance meets the standards. The rate modulation factor λ is a coefficient dynamically calculated based on real-time evaluation results, used to scale the base interval to achieve flexible frequency control. λ > 1 means accelerated triggering (higher frequency), and λ < 1 means delayed triggering (lower frequency).

[0052] As an example, apply different frequency adjustment strategies according to different situations: A. When the operation accuracy is lower than the preset qualified threshold and the scenario resolution efficiency is also lower than the preset high-efficiency threshold, it is determined that the response is extremely poor, and apply the reinforcement punishment strategy: set λ to ( > 1, such as = 1.8), so that the waiting time for the next negative guidance scenario of the same type to trigger is shortened to / ; B. When the operation accuracy is lower than the preset qualified threshold but the scenario resolution efficiency is higher than the preset high-efficiency threshold, it is determined that the response mode has high risk, and apply the targeted high-frequency strategy: set λ to ( > 1, and > ), such as = 2.0), so that the waiting time for the next negative guidance scenario of the same type to trigger is shortened to / , and mark that the trainee has a high-risk tendency of prioritizing efficiency at the expense of accuracy; C. When the operation accuracy is higher than the preset qualified threshold but the scenario resolution efficiency is lower than the preset high-efficiency threshold, it is determined that the response is safe but slow, and apply the gentle reminder strategy: set λ to (1 < < ), such as = 1.2), so that the waiting time for the next negative guidance scenario of the same type to trigger is shortened to / ; D. When the operation accuracy is higher than the preset qualified threshold and the scenario resolution efficiency is higher than the preset high-efficiency threshold, it is determined that the response is good, and apply the incentive strategy: set λ to (0 < reward < 1, such as reward = 0.7), so that the trigger waiting time is extended to / reward. Establish a short-term memory window to perform weighted average processing on consecutive response results to avoid drastic fluctuations in trigger frequency caused by random fluctuations in single performance.

[0053] In a specific implementation, set the base trigger interval = 300 seconds, meaning that after completing a normal 5-minute task flow by default, the student may encounter the same edge-driving challenge again. The evaluation vector shows an operational accuracy of 0.7 (<0.8) and a scenario resolution efficiency of 0.9 (>0.85). According to the preset rule: when the accuracy is below the threshold, regardless of efficiency, it is considered that more training is needed. Taking λ = 1.5, the next trigger waiting time is calculated as 300 / 1.5 = 200 seconds. This means that the student's driving behavior will be re-evaluated after approximately 200 seconds (or at the corresponding task progress point). If the student still exhibits a tendency towards "edge-driving," since the detection threshold has been lowered previously (see step S150), it will be detected more sensitively, and new negative guidance scenarios will be triggered at a higher frequency (shorter intervals) (for example, this time it might be a virtual prompt of oncoming traffic at a curve), thus providing intensive corrective training.

[0054] In one embodiment of the present invention, the specific process of "dynamically adjusting the complexity of the negative guidance plot based on the changing trend of the evaluation vector according to the historical response results" can be further explained in conjunction with the following description.

[0055] As described in the following steps The statistical analysis of the assessment vectors of several responses to the same type of potential risk behavior within a preset period forms a historical sequence; Analyze the evolution trends of reaction delay time and operational accuracy in the historical sequence. When both show a continuous improvement trend, initiate a complexity upgrade. The complexity upgrade is achieved by combining and superimposing challenge increments of different dimensions, specifically including: superimposing secondary environmental interference while maintaining the core challenge unchanged; shortening the decision-making time window from the occurrence of the plot to the necessity of taking effective action; and reducing the explicitness of key hints in the plot. When the historical sequence shows that the accuracy of the operation drops significantly after the complexity level is increased, it automatically reverts to one complexity level.

[0056] It should be noted that the preset period is a time window or task phase window used to assess trainees' ability trends, such as "within the same training module," "within the most recent hour of training," or "the most recent 10 similar challenges." Challenge increments are pluggable elements that increase training difficulty from multiple dimensions while keeping the core teaching points (such as avoiding road debris) unchanged. These increments aim to simulate more complex and stressful real-world work scenarios. Explicitness refers to the visibility of risk warnings or key plot information; reducing explicitness requires trainees to have more acute observation and risk prediction abilities.

[0057] As an example, the dimensions of complexity are not limited to environmental interference, time pressure, and prompts, but can also include: increasing the number of multitasking tasks that need to be handled simultaneously (such as needing to pay attention to dashboard alarms while avoiding obstacles), introducing ethical dilemmas in decision-making (such as avoiding foreign objects potentially endangering virtual pedestrians), etc.

[0058] In one specific implementation, the historical sequence of evaluation vectors for trainees dealing with four negative guidance scenarios involving road debris in a previous transportation module was analyzed. The analysis revealed that the average reaction delay time consistently decreased from 2.0 seconds to 1.3 seconds, while the average operational accuracy consistently increased from 0.6 to 0.82. Since both key indicators showed a continuous improvement trend, it was determined that the trainees' basic abilities in this type of challenge had been solidified, and the complexity upgrade procedure was initiated. Two options were selected from the challenge increment library: "overlaying minor environmental interference" and "reducing the visibility of prompts." When the "road debris" scenario is generated again, a light dust effect will be dynamically activated while generating small stones, slightly reducing the visibility of the road ahead. Simultaneously, the stone textures, which were originally clearly distinguishable from the road surface, will be replaced with textures with lower contrast to the road surface color, making them less likely to be immediately noticed. After the upgrade, if the trainee's operational accuracy drops sharply to 0.5 under the new complex scenario, a rollback mechanism will be immediately triggered. In the trainee's next training session, the scenario complexity will be restored to the level before the upgrade, ensuring that the challenge matches the trainee's abilities.

[0059] Reference Figure 2 The diagram illustrates a structural block diagram of a VR safety training dynamic difficulty control system for open-pit coal mines according to an embodiment of this application, comprising: The interactive behavior acquisition module 210 is used to collect multi-dimensional interactive behavior data of the target trainee in the VR training scenario in real time, and encode the multi-dimensional interactive behavior data into a behavior vector sequence. The time-series risk prediction module 220 is used to input the behavior vector sequence into a pre-trained prediction model and output the potential risk behavior and the probability of risk occurrence. The counterfactual risk deduction module 230 is used to generate risk events that will be caused by the continued occurrence of the potential risky behavior through a risk deduction model when the probability of the risk occurrence is higher than a preset threshold. The negative plot injection module 240 is used to dynamically generate negative guidance plots associated with the current VR training scene based on the risk event, and to implant the negative guidance plots into the VR training scene; The dynamic difficulty adjustment module 250 is used to dynamically adjust the generation strategy of subsequent negative guidance scenarios based on the target learner's response to the negative guidance scenario.

[0060] In one embodiment of the present invention, the interactive behavior acquisition module 210 includes: The first interactive behavior acquisition submodule is used to collect multi-dimensional interactive behavior data of the target trainees in the VR training scenario in real time at a preset sampling frequency, and to perform noise reduction and normalization preprocessing on the collected raw data. The second interactive behavior acquisition submodule is used to adopt a non-uniform sampling strategy that combines event triggering and fixed period sampling. It records only when the interactive state changes significantly or when the period point is reached, and encapsulates the processed data into structured data units. The third interactive behavior acquisition submodule is used to extract key features from each of the structured data units based on preset feature extraction rules. The key features include at least behavior type, quantization parameters and context labels. The fourth interactive behavior acquisition submodule is used to arrange the key features corresponding to each data unit in chronological order and combine them into the behavior vector sequence, wherein each vector is associated with a precise timestamp, and the interval between vectors in the sequence reflects the dynamic sparsity of the interaction.

[0061] In one embodiment of the present invention, the time-series risk prediction module 220 includes: The first time-series risk prediction submodule is used to match the behavior vector sequence with the baseline pattern in the preset non-optimal behavior pattern library in real time. The baseline pattern defines a pattern template composed of compliant but inefficient behaviors and high-risk behavior fragments. The second time-series risk prediction submodule is used to calculate the dynamic time warping distance between the behavior vector sequence and the benchmark pattern, and output the benchmark pattern with the highest matching degree. The third time-series risk prediction submodule is used to define the semantic label corresponding to the benchmark pattern with the highest matching degree as a potential risk behavior. The fourth time-series risk prediction submodule is used to calculate the probability of the potential risk behavior evolving into a safety accident within a preset time window in the future.

[0062] In one embodiment of the present invention, the counterfactual risk deduction module 230 includes: The first counterfactual risk deduction submodule is used to take the potential risk behavior as the initial condition, the instantaneous state of the current VR training scene as the initial environment for deduction, and to perform forward deduction along the time axis in the risk deduction model to output a first-level consequence event. The second counterfactual risk deduction submodule is used to iteratively deduce the chain reaction triggered by the first-level consequence event until the risk event is generated; The third counterfactual risk deduction submodule is used to output the event type, involved objects, severity level and evolution path corresponding to the risk event.

[0063] In one embodiment of the present invention, the negative plot injection module 240 includes: The first negative plot injection submodule is used to retrieve a basic plot module that conforms to the physical rules and environmental constraints of the current VR training scenario from a preset plot element library according to the event type. The second negative plot injection submodule is used to instantiate the intermediate links in the evolution path that can be externally intervened or occur naturally into dynamic environmental interference events, using the involved objects as interactive entities. The third negative plot injection submodule is used to dynamically adapt and calibrate the spatiotemporal parameters of the dynamic environmental interference event based on the basic plot module, and generate the negative guiding plot. The fourth negative plot injection submodule is used to inject the negative guiding plot into the current VR training scene in real time without interrupting the main task flow.

[0064] In one embodiment of the present invention, the dynamic difficulty control module 250 includes: The first dynamic difficulty control submodule is used to acquire the target learner's response behavior data to the negative guidance scenario and generate a response result evaluation vector, which includes reaction delay time, operation accuracy and scenario resolution efficiency. The second dynamic difficulty adjustment submodule is used to compare the response result evaluation vector with the preset expected response standard, and generate the weak item weight of the target student based on the comparison result. The third dynamic difficulty adjustment submodule is used to adjust the matching similarity threshold of the corresponding benchmark pattern in the prediction model based on the weight of the weak point; The fourth dynamic difficulty control submodule is used to adjust the trigger frequency of the negative guidance plot based on the response result evaluation vector; The fifth dynamic difficulty adjustment submodule is used to dynamically adjust the complexity of the negative guidance plot based on the changing trend of the evaluation vector according to the historical response results.

[0065] In one embodiment of the present invention, the fourth dynamic difficulty adjustment submodule includes: The frequency parameter initialization unit is used to set the basic trigger interval. And the frequency modulation factor λ, the triggering wait time for the next negative guiding plot of the same type is ; The first trigger condition judgment unit is used to set λ to a value greater than 1 when the operation accuracy is lower than a preset threshold, and output the trigger waiting time for the next negative guidance plot of the same type. The second trigger condition judgment unit is used to set λ to a positive value less than 1 and output the trigger waiting time for the next negative guidance plot of the same type when both the operation accuracy and the plot resolution efficiency are higher than a preset threshold.

[0066] In one embodiment of the present invention, the fifth dynamic difficulty adjustment submodule includes: The historical data aggregation unit is used to statistically analyze the evaluation vectors of several responses to the same type of potential risk behavior within a preset period, forming a historical sequence. The capability trend analysis unit is used to analyze the evolution trend of reaction delay time and operational accuracy in the historical sequence. When both show a continuous improvement trend, the complexity upgrade is initiated. The complexity upgrade is achieved by combining and superimposing challenge increments of different dimensions, specifically including: superimposing secondary environmental interference while keeping the core challenge unchanged; shortening the decision-making time window from the occurrence of the plot to the necessity of taking effective action; and reducing the explicitness of key prompts in the plot. The difficulty balancing rollback unit is used to automatically roll back one level of complexity when the accuracy of the operation displayed in the historical sequence decreases significantly after the complexity level is upgraded.

[0067] Reference Figure 3 The diagram illustrates a computer electronic device for implementing a method for dynamic difficulty adjustment in VR safety training for open-pit coal mines according to the present invention, which may specifically include the following: The aforementioned computer electronic device 1 is manifested in the form of a general-purpose computing device. The components of the computer electronic device 1 may include, but are not limited to: one or more processors or processing units 3, memory 8, and a bus 4 connecting different system components (including memory 8 and processing unit 3).

[0068] Bus 4 represents one or more of several bus architectures, including memory buses or memory controllers, peripheral buses, graphics acceleration ports, processors, or local buses using any of the various bus architectures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Audio / Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0069] Computer electronic device 1 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by computer electronic device 1, including volatile and non-volatile media, removable and non-removable media.

[0070] Memory 8 may include computer system readable media in the form of volatile memory, such as random access memory 9 and / or cache memory 10. Computer electronic device 1 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 11 may be used to read and write non-removable, non-volatile magnetic media (commonly referred to as a "hard disk drive"). Although Figure 3 As not shown, a disk drive for reading and writing to a removable non-volatile disk (such as a "floppy disk") and an optical disk drive for reading and writing to a removable non-volatile optical disk (such as a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 4 via one or more data media interfaces. The memory may include at least one program product having a set (e.g., at least one) of program modules 13 configured to perform the functions of the embodiments of this application.

[0071] A program / utility 12 having a set (at least one) of program modules 13 may be stored, for example, in memory. Such program modules 13 include—but are not limited to—an operating system, one or more application programs, other program modules 13, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 13 typically perform the functions and / or methods described in the embodiments of this application.

[0072] The computer electronic device 1 can also communicate with one or more external devices 2 (e.g., keyboard, pointing device, display 7, camera, etc.), and with one or more devices that enable an operator to interact with the computer electronic device 1, and / or with any device that enables the computer electronic device 1 to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed through the I / O interface 6. Furthermore, the computer electronic device 1 can also communicate with one or more networks (e.g., local area network (LAN)), wide area network (WAN), and / or public networks (e.g., the Internet) through the network adapter 5. Figure 3 As shown, network adapter 5 communicates with other modules of computer electronic device 1 via bus 4. It should be understood that, although... Figure 3 Not shown, it may be combined with other hardware and / or software modules, including but not limited to: microcode, device drivers, redundant processing unit 3, external disk drive array, RAID system, tape drive and data backup storage system 11, etc.

[0073] The processing unit 3 executes various functional applications and data processing by running programs stored in memory 8, such as implementing a method for dynamic difficulty control of VR safety training in open-pit coal mines provided in the embodiments of this application.

[0074] That is, when the processing unit 3 executes the above procedure, it performs the following: real-time collection of multi-dimensional interactive behavior data of the target trainee in the VR training scene, and encoding the multi-dimensional interactive behavior data into a behavior vector sequence; inputting the behavior vector sequence into a pre-trained prediction model, and outputting potential risk behaviors and the probability of risk occurrence; when the probability of risk occurrence is higher than a preset threshold, generating risk events that the potential risk behaviors will continue to cause through the risk inference model; dynamically generating negative guidance plots associated with the current VR training scene based on the risk events, and embedding the negative guidance plots into the VR training scene; dynamically adjusting the generation strategy of the subsequent negative guidance plots according to the target trainee's response to the negative guidance plots.

[0075] In this application embodiment, the application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for dynamic difficulty adjustment of VR safety training in open-pit coal mines as provided in all embodiments of the application.

[0076] That is, when the program is executed by the processor, it performs the following: real-time collection of multi-dimensional interactive behavior data of the target learner in the VR training scenario, and encoding the multi-dimensional interactive behavior data into a behavior vector sequence; inputting the behavior vector sequence into a pre-trained prediction model, and outputting potential risk behaviors and the probability of risk occurrence; when the probability of risk occurrence is higher than a preset threshold, generating risk events that the potential risk behaviors will continue to cause through a risk inference model; dynamically generating negative guidance plots associated with the current VR training scenario based on the risk events, and embedding the negative guidance plots into the VR training scenario; dynamically adjusting the generation strategy of subsequent negative guidance plots according to the target learner's response to the negative guidance plots.

[0077] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0078] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including—but not limited to—electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0079] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the operator's computer, partially on the operator's computer, as a standalone software package, partially on the operator's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the operator's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider). The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably.

[0080] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0081] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0082] The above provides a detailed description of the dynamic difficulty control method and system for VR safety training in open-pit coal mines provided by this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for dynamically adjusting the difficulty of VR safety training in open-pit coal mines, characterized in that, Including the following steps: Real-time collection of multi-dimensional interactive behavior data of target learners in VR training scenarios, and encoding of the multi-dimensional interactive behavior data into a sequence of behavior vectors; The sequence of behavioral vectors is input into a pre-trained prediction model, which outputs potential risky behaviors and the probability of risk occurrence. When the probability of the risk occurring is higher than a preset threshold, a risk event that the potential risk behavior will continue to cause is generated through a risk inference model. Based on the aforementioned risk events, negative guidance scenarios are dynamically generated and associated with the current VR training scenario, and these negative guidance scenarios are then embedded into the VR training scenario. Based on the target learners' responses to the negative guidance scenarios, the generation strategy for subsequent negative guidance scenarios is dynamically adjusted.

2. The method for dynamic difficulty control of VR safety training in open-pit coal mines according to claim 1, characterized in that, The step of inputting the behavior vector sequence into a pre-trained prediction model and outputting the potential risk behavior and the probability of risk occurrence specifically includes: The behavior vector sequence is matched in real time with a baseline pattern in a preset non-optimal behavior pattern library. The baseline pattern defines a pattern template consisting of compliant but inefficient behaviors and high-risk behavior fragments. Calculate the dynamic time warped distance between the behavior vector sequence and the baseline pattern, and output the baseline pattern with the highest matching degree; The semantic label corresponding to the benchmark pattern with the highest matching degree is defined as a potential risky behavior; Calculate the probability that the potential risky behavior will evolve into a safety accident within a preset time window in the future.

3. The method for dynamic difficulty control of VR safety training in open-pit coal mines according to claim 1, characterized in that, The step of generating risk events that will be caused by the continued occurrence of the potential risk behavior through a risk inference model when the probability of the risk occurrence is higher than a preset threshold specifically includes: Using the potential risky behavior as the initial condition and the instantaneous state of the current VR training scenario as the initial environment for deduction, the risk deduction model is deduced forward along the time axis to output a first-level consequence event. Iteratively deduce the chain reaction triggered by the first-level consequence event until the risk event is generated; Output the event type, involved parties, severity level, and evolution path corresponding to the risk event.

4. The method for dynamic difficulty control of VR safety training in open-pit coal mines according to claim 3, characterized in that, The step of dynamically generating negative guidance scenarios associated with the current VR training scenario based on the risk event, and embedding the negative guidance scenarios into the VR training scenario, specifically includes: Based on the event type, retrieve basic plot modules that conform to the physical rules and environmental constraints of the current VR training scenario from the preset plot element library; Using the objects involved as interactive entities, the intermediate links in the evolution path that can be externally intervened or occur naturally are instantiated as dynamic environmental interference events; Based on the basic plot module, the spatiotemporal parameters of the dynamic environmental interference event are dynamically adapted and calibrated to generate the negative guidance plot; Without interrupting the main task flow, the negative guidance plot is implanted into the current VR training scene in real time.

5. The method for dynamic difficulty control of VR safety training in open-pit coal mines according to claim 2, characterized in that, The step of dynamically adjusting the generation strategy for subsequent negative guidance scenarios based on the target learner's response to the negative guidance scenario specifically includes: Data on the target learner's response to the negative guidance scenario is obtained, and a response outcome evaluation vector is generated. The evaluation vector includes response delay time, operational accuracy, and scenario resolution efficiency. The response result evaluation vector is compared with the preset expected response standard, and the weakness weight of the target student is generated based on the comparison result. The matching similarity threshold of the corresponding benchmark pattern in the prediction model is adjusted based on the weight of the weak point; The triggering frequency of the negative guidance plot is adjusted based on the response result evaluation vector; The complexity of the negative guidance plot is dynamically adjusted based on the changing trend of the evaluation vector according to the historical response results.

6. The method for dynamic difficulty control of VR safety training in open-pit coal mines according to claim 5, characterized in that, The step of adjusting the trigger frequency of the negative guidance plot based on the response result evaluation vector specifically includes: Set the base trigger interval And the frequency modulation factor λ, the triggering wait time for the next negative guiding plot of the same type is ; When the accuracy of the operation is lower than the preset threshold, λ is set to a value greater than 1, and the trigger waiting time for the next negative guidance plot of the same type is output. When both the operation accuracy and the plot resolution efficiency are higher than the preset threshold, λ is set to a positive value less than 1, and the trigger waiting time for the next negative guidance plot of the same type is output.

7. The method for dynamic difficulty control of VR safety training in open-pit coal mines according to claim 5, characterized in that, The step of dynamically adjusting the complexity of the negative guidance plot based on the changing trend of the evaluation vector according to historical response results specifically includes: The statistical analysis of the assessment vectors of several responses to the same type of potential risk behavior within a preset period forms a historical sequence; Analyze the evolution trends of reaction delay time and operational accuracy in the historical sequence. When both show a continuous improvement trend, initiate a complexity upgrade. The complexity upgrade is achieved by combining and superimposing challenge increments of different dimensions, specifically including: superimposing secondary environmental interference while maintaining the core challenge unchanged; shortening the decision-making time window from the occurrence of the plot to the necessity of taking effective action; and reducing the explicitness of key hints in the plot. When the historical sequence shows that the accuracy of the operation drops significantly after the complexity level is increased, it automatically reverts to one complexity level.

8. A VR safety training dynamic difficulty control system for open-pit coal mines, characterized in that, include: The interactive behavior acquisition module is used to collect multi-dimensional interactive behavior data of the target trainees in the VR training scenario in real time, and encode the multi-dimensional interactive behavior data into a behavior vector sequence. The time-series risk prediction module is used to input the behavior vector sequence into a pre-trained prediction model and output the potential risk behaviors and the probability of risk occurrence. The counterfactual risk simulation module is used to generate risk events that will be caused by the continued occurrence of the potential risky behavior when the probability of the risk occurring is higher than a preset threshold, through a risk simulation model. The negative plot injection module is used to dynamically generate negative guidance plots associated with the current VR training scene based on the risk event, and to inject the negative guidance plots into the VR training scene; The dynamic difficulty adjustment module is used to dynamically adjust the generation strategy of subsequent negative guidance scenarios based on the target learners' response to the negative guidance scenarios.

9. A computer electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images