Interactive life science display system and method

Abstract

The application discloses an interactive life science display system and method, and relates to the technical field of somatosensory interaction; the method comprises the following steps: collecting action time sequence characteristic data, and constructing a time sequence multidimensional array; extracting action characteristics from the time sequence multidimensional array in a set time window; calculating a speed stability score, a direction stability score and a trajectory continuity score based on the action characteristics, and setting contribution weights for the speed stability score, the direction stability score, the trajectory continuity score and a pressure characteristic, and obtaining an intention confidence by weighted summation; the application effectively reduces invalid occupation of resources by low-intention actions, significantly reduces core action feedback delay, improves the immersion and scientificity of popular science display, reduces misreading of scientific principles by users, simultaneously reduces equipment loss and maintenance cost, and enhances user participation interest.

Description

Interactive Life Science Demonstration Systems and Methods Technical Field

[0001] This invention relates to the field of motion-sensing interaction technology, and more specifically, to an interactive life science display system and method. Background Technology

[0002] In interactive life science demonstration scenarios, the core experience involves continuous interactions such as operating virtual gene scissors through motion-sensing devices and touching virtual cells with tactile devices. In these interactions, user actions are not isolated but consist of a series of steps, including subtle adjustments before raising a hand, precise touch operations, and corrective actions after positional deviations, forming a complete temporal sequence. Furthermore, the interactive intentions carried by different steps are significantly different.

[0003] The underlying technology supporting this type of interaction is based on the recognition and feedback triggering mechanism of single-frame motion features. The system collects motion coordinates and amplitude data in real time through motion sensing devices, captures touch pressure and duration information through haptic devices, and then determines whether the motion is valid based on preset fixed thresholds. For example, when the user's motion amplitude exceeds a certain value or the touch pressure reaches a set standard, it is determined to be a valid motion that requires a response, and then the corresponding haptic feedback is triggered, such as simulating cell stiffness or the resistance of gene scissors. The entire process relies on the logic of immediate response after the motion is completed.

[0004] However, this technical design neglects the correlation between action timing and intent, leading to a mismatch between interactive feedback and the user's true intent. On one hand, the system fails to distinguish the intent intensity of different steps within a continuous sequence of actions, misjudging low-intent actions such as subtle adjustments before lifting a hand or tentative touches before core operations as valid actions, prematurely occupying feedback resources. This results in a delay in the appropriate feedback when the user performs a core action. For example, when a user touches a virtual cell, they might first make a tentative 0.2-second touch with 0.3N pressure. This action is merely to confirm the location and has no clear interactive intent, but the system misjudges it as a valid action and triggers cell hardness feedback. When the user subsequently performs a core action with 1.2N pressure for 0.5 seconds, such as a core press, the hardness feedback corresponding to the true intent will be delayed by more than 0.3 seconds because feedback resources have already been occupied.

[0005] Meanwhile, when distinguishing the intensity of intent in different steps of a continuous action, the fixed threshold setting cannot adapt to the movement characteristics of different users. When children operate motion-sensing devices, their movements are large, but their intent is scattered. For example, when operating virtual gene scissors, there will be multiple swings with an amplitude of 5cm but no clear cutting target. The fixed threshold will misjudge these swings as core movements, causing tactile feedback to frequently jump between no resistance and strong resistance. When adults operate, their movements are small, but their intent is concentrated. For example, when performing precise cutting with an amplitude of 2cm, the fixed threshold may miss the judgment because it does not reach the preset 3cm standard, resulting in a complete lack of resistance feedback corresponding to the cutting action.

[0006] The aforementioned issues directly undermine the immersive and scientific rigor of science demonstrations. Users may mistake erroneous feedback triggered by tentative actions for genuine characteristics of biological tissues, leading to a misinterpretation of scientific principles. Children lose patience due to frequent changes in feedback, while adults lose interest due to a lack of feedback, thus hindering the effective transmission of science. Furthermore, ineffective feedback for low-intention actions continuously consumes system computing resources, increasing the processing delay for feedback in core operations. Long-term operation also accelerates the wear and tear on haptic devices, increasing maintenance costs.

[0007] In view of this, the present invention proposes an interactive life science demonstration system and method to solve the above problems. Summary of the Invention

[0008] To overcome the aforementioned deficiencies of the prior art and achieve the above objectives, the present invention provides the following technical solution: an interactive life science demonstration method, comprising:

[0009] Collect action time-series feature data and construct a multi-dimensional time series array;

[0010] Within a set time window, action features are extracted from a time series multidimensional array; the action features include pressure features, speed features, and amplitude features; based on the action features, speed stability scores, directional stability scores, and trajectory coherence scores are calculated, and contribution weights are assigned to the speed stability scores, directional stability scores, trajectory coherence scores, and pressure features, and the intention confidence is obtained by weighted summation;

[0011] Based on the intent confidence level and the set intent dynamic threshold, the user's actions are divided into different intent levels in real time, and corresponding feedback computing resources are allocated according to the divided intent levels.

[0012] Based on pre-collected user profile data, the dynamic threshold of intent and the weight of each contribution are adaptively adjusted.

[0013] Furthermore, the action timing feature data includes the continuous pressure value applied by the user when touching, the action amplitude and spatial coordinate sequence when the user waves the motion sensing device, and the velocity change rate of the spatial coordinate sequence of adjacent frames; the action amplitude is the Euclidean distance between the spatial coordinates of two consecutive frames in the spatial coordinate sequence.

[0014] Furthermore, within a defined time window, methods for extracting action features from a multidimensional time series array include:

[0015] Extract pressure features, which include the peak value of the pressure within a time window and the pressure duration;

[0016] Extract speed features, which are the average or maximum rate of change of speed within a time window; when the interactive scene needs to reflect the overall speed fluctuation of the action, the average rate of change of speed is selected; when the interactive scene needs to reflect the instantaneous speed fluctuation of the action, the maximum rate of change of speed is selected.

[0017] Extract amplitude features, which are the cumulative amplitudes of motion displacement within a time window;

[0018] The sliding window mechanism ensures that the action features are updated once within each time window. When a new data frame arrives, the old window slides out and the new window slides in, and the action features are updated accordingly.

[0019] Furthermore, methods for calculating velocity stability scores, directional stability scores, and trajectory coherence scores based on motion features include:

[0020] The velocity stability score is obtained by subtracting the velocity characteristics within the time window from 1;

[0021] The directional stability score is defined as: directional stability score = 1 - (average directional deviation / preset maximum deviation angle); the average directional deviation is obtained by calculating the angle deviation between the motion direction of each frame and the preset target direction, and taking the average of the directional deviations of several consecutive frames;

[0022] When the interactive scenario requires the user to perform an action according to a fixed standard trajectory point, the trajectory continuity score is calculated based on the trajectory deviation between the actual trajectory point formed by the spatial coordinate sequence of each frame and the standard trajectory point. The trajectory continuity score = 1 - (average trajectory deviation / preset tolerance); the trajectory deviation is the distance between the corresponding point of the actual trajectory point and the standard trajectory.

[0023] When the interaction scenario does not require the user to perform actions according to fixed standard trajectory points, the trajectory continuity score is calculated by evaluating the trajectory deviation rate, which is the proportion of the cumulative trajectory deviation to the cumulative displacement of the action within the time window.

[0024] Furthermore, methods for classifying user actions into different intent levels in real time based on intent confidence and a set dynamic intent threshold include:

[0025] The intent levels include low-intent actions, transitional actions, and core actions;

[0026] When the intent confidence level is lower than the set dynamic threshold for low intent ratio, the user's action is judged as a low intent action.

[0027] When the intent confidence level is between the intent dynamic threshold set at a low intent ratio and the intent dynamic threshold, the user's action is judged as a transitional action.

[0028] Otherwise, the user's action will be classified as a core action.

[0029] Furthermore, the method for allocating corresponding feedback computing resources according to the divided intent levels includes:

[0030] The feedback computing power resources are the computing capabilities of the system used to process motion data, calculate feedback logic, and drive the haptic feedback device.

[0031] Low-intention actions are marked as the lowest priority and allocated the lowest priority ratio of feedback computing resources.

[0032] Transitional actions are marked as medium priority and allocated a set medium priority ratio of feedback computing resources.

[0033] The core action is marked as the highest priority and can occupy all feedback computing resources.

[0034] Furthermore, methods for assigning contribution weights to velocity stability score, directional stability score, trajectory coherence score, and pressure feature include:

[0035] The contribution weight of the speed stability score is set based on the correlation between the speed change rate and the level of action intention; the smaller the speed change rate, the higher the contribution weight.

[0036] The contribution weight of the directional stability score is set based on the correlation between the average directional deviation and the level of action intention; the smaller the average directional deviation, the higher the contribution weight.

[0037] The contribution weight of the trajectory coherence score is set based on the correlation between the average trajectory deviation / trajectory deviation rate and the level of action intent and conformity to the standard trajectory. The smaller the trajectory deviation or the lower the deviation rate, the higher the contribution weight.

[0038] The pressure peak and duration scores in the pressure characteristics have a higher pre-defined contribution weight than the velocity stability score, directional stability score, and trajectory coherence score.

[0039] The sum of all contribution weights is 1.

[0040] Furthermore, methods for obtaining user profile data include:

[0041] Collect all newly added user operation records daily, associate each record with the corresponding age group information, and extract key action features from each record; the key action features include the standard deviation of peak pressure, average action amplitude, and trajectory deviation rate, forming a feature dataset;

[0042] The feature dataset is divided into children's group and adult group according to age information, and clustering is performed separately. The clustering dimensions are the standard deviation of stress, the average amplitude of movement and the trajectory deviation rate. The neighborhood radius is set, and the minimum sample size is calibrated based on the similarity of the standard deviation of stress within the group, the average amplitude of movement and the trajectory deviation rate and the difference of the standard deviation of stress within the group and the trajectory deviation rate between the groups in the historical clustering.

[0043] For the children's group, records with peak stress standard deviation greater than the set children's standard deviation threshold, average movement amplitude greater than the children's movement amplitude threshold, or trajectory deviation rate greater than the children's deviation threshold are marked as children-fluctuating type; for the adults' group, records with peak stress standard deviation less than the adults' standard deviation threshold, average movement amplitude less than the adults' movement amplitude threshold, or trajectory deviation rate less than the adults' deviation threshold are marked as adults-stable type.

[0044] After clustering is completed, the final user profile data is obtained.

[0045] Furthermore, methods for adaptively adjusting the dynamic threshold of intent and the weights of each contribution based on pre-collected user profile data include:

[0046] A reinforcement learning model is used to adaptively adjust the contribution weights of speed stability score, direction stability score, trajectory coherence score, and pressure feature. The reinforcement learning model aims to maximize the accuracy of intent recognition. The state space includes the weight adjustment effects of user type, real-time action features, and historical interactions. The action space is the adjustment amount of each contribution weight, and the sum of each contribution weight is constrained to be 1, and the contribution weight of a single feature is within a preset range.

[0047] The reinforcement learning model is trained by using historical interaction records, which are labeled with action features, user type and actual intent level. The initial weight strategy is obtained through pre-training and then optimized.

[0048] During the optimization process, the reinforcement learning model loads initial weights according to user type, extracts real-time action features through a set time window, combines user type and historical recognition results to form a state, inputs it into the reinforcement learning model, and outputs contribution weights.

[0049] The intention confidence is calculated based on the contribution weight of the output. If the intention confidence matches the actual intention level, a positive reward is given; otherwise, a negative reward is given, driving the reinforcement learning model to update its weight strategy.

[0050] Interactive life science demonstration systems and methods, including:

[0051] The data acquisition module is used to collect action time-series feature data and construct a multi-dimensional time series array;

[0052] The intent calculation module is used to extract action features from a time series multidimensional array within a set time window; based on the action features, it calculates speed stability score, direction stability score, and trajectory coherence score, and sets contribution weights for speed stability score, direction stability score, trajectory coherence score, and pressure feature, and then calculates the intent confidence score by weighted summation.

[0053] The intent recognition module is used to classify user actions into different intent levels in real time based on intent confidence and set dynamic intent thresholds, and allocate corresponding feedback computing resources according to the classified intent levels.

[0054] The threshold adjustment module adaptively adjusts the dynamic threshold of intent and the weight of each contribution based on pre-collected user profile data.

[0055] Compared with existing technologies, the technical effects and advantages of the interactive life science demonstration system and method proposed in this invention are as follows:

[0056] This invention collects temporal feature data of actions using multiple sensors, constructs a multi-dimensional time-series array, extracts features such as pressure, speed, and amplitude within a sliding window, calculates speed stability, directional stability, and trajectory continuity scores, and weights these scores to obtain intent confidence. Based on confidence and dynamic thresholds, actions are categorized into low-intent, transitional, and core levels, with core actions receiving priority in resource allocation and more feedback computing power. Thresholds and feature weights are dynamically adjusted based on user profiles, identifying child-fluctuating and adult-stable user types to adapt to different user action characteristics. A rhythm matching mechanism is introduced, triggering feedback during actions according to the rhythms of children (slow start and slow finish) and adults (fast start and steady finish), and optimizing synchronization through cross-scene deviation correction.

[0057] This invention effectively reduces the ineffective use of resources by low-intention actions, significantly reduces the feedback delay of core actions, accurately adapts to the action characteristics of different users, improves the filtering rate of ineffective actions in children and the recognition rate of effective actions in adults, avoids feedback jumps and missed judgments, and synchronizes feedback with the rhythm of actions, enhancing the immersiveness and scientific nature of popular science presentations, reducing users' misinterpretation of scientific principles, while reducing equipment wear and maintenance costs and enhancing user participation interest. Attached Figure Description

[0058] Figure 1 is a schematic diagram of an interactive life science demonstration system according to an embodiment of the present invention;

[0059] Figure 2 is a flowchart of the interactive life science demonstration method according to an embodiment of the present invention;

[0060] Figure 3 is a flowchart of a method for classifying user actions into different intent levels in real time according to an embodiment of the present invention.

[0061] Figure 4 is a flowchart of a method for adaptively adjusting the dynamic threshold of intent and the weight of each contribution based on pre-collected user profile data according to an embodiment of the present invention.

[0062] Figure 5 is a schematic diagram of a virtual cell touch scene according to an embodiment of the present invention. Detailed Implementation

[0063] To further clarify the technical problem to be solved by this application and its background, the limitations of the prior art will be described in detail before proceeding with specific implementation methods.

[0064] Specifically, if low-intent actions are not identified in the different steps of a user's continuous actions, it will cause multi-dimensional waste of resources and directly affect system performance. For example, in terms of computing power, the feedback processing triggered by each low-intent action requires 15%-20% of the CPU computing power. In a child's operation scenario, there may be 10-15 such actions within 1 minute, accumulating to 150%-300% of the computing power, causing congestion in the core action processing queue and a significant increase in feedback latency from 50ms to over 200ms. In terms of memory, the action timing feature data generated by each action will be mistakenly stored as valid interaction records. After 8 hours of display, the memory usage will increase from 2GB to 6-10GB, triggering frequent memory reclamation and further slowing down the response speed. In terms of hardware, the drive motor of the haptic feedback device will generate 10-15 micro-mechanical vibrations each time it responds to a low-intent action, even if the intensity is low. Mis-triggered vibrations can cause the number of vibrations per day to surge from 1,000 to over 5,000, resulting in a 40% reduction in motor lifespan. These problems, when combined, will not only significantly reduce the smoothness of interaction, but also increase equipment wear and maintenance costs. They may even damage users' correct understanding of scientific principles due to feedback delays. Therefore, accurate recognition of low-intention actions is a key prerequisite for ensuring the efficient operation of the system and the effectiveness of popular science education.

[0065] The technical solutions of the embodiments of the present invention will be described in detail, clearly, and completely below with reference to the accompanying drawings. It should be particularly noted that the specific embodiments described below are only for better illustrating and explaining the technical solutions of the present invention, and are intended to enable those skilled in the art to better understand and implement the present invention, and should not be construed as limiting the scope of protection of the present invention. Without departing from the spirit and substance of the present invention, those skilled in the art can modify, adjust, or make equivalent substitutions based on the content disclosed in the present invention, and these should all be considered within the scope of protection of the present invention.

[0066] Example 1

[0067] Please refer to Figure 1. This embodiment discloses an interactive life science demonstration system, including a data acquisition module, an intent calculation module, an intent recognition module, and a threshold adjustment module. Each module is connected via wired and / or wireless connections to achieve data transmission.

[0068] The data acquisition module is used to collect action time-series feature data and construct a time-series multidimensional array.

[0069] The action timing feature data includes the continuous pressure value applied by the user when touching the device, the action amplitude and spatial coordinate sequence when the user waves the motion sensing device, and the velocity change rate of the spatial coordinate sequence of adjacent frames; the action amplitude is the Euclidean distance between the spatial coordinates of two consecutive frames in the spatial coordinate sequence.

[0070] Methods for obtaining action time sequence feature data include:

[0071] Electronic skin pressure sensors are used to acquire continuous pressure values ​​applied by a user during touch. For example, when a user touches a virtual cell model with their fingertip, the pressure sensor on the surface of the electronic skin records the real-time pressure changes applied by the fingertip. Based on flexible electronic materials such as piezoresistive or capacitive principles, electronic skin pressure sensors can sample pressure at a frequency of 50 to 100 times per second or more and output corresponding electrical signals, while being synchronized with other sensors.

[0072] The motion-sensing interactive device incorporates a 6-axis inertial measurement unit (IMU) and an infrared positioning device to record motion amplitude and spatial coordinate sequences. The IMU includes a three-axis accelerometer and a three-axis gyroscope, detecting changes in acceleration and rotational attitude of the device. The infrared positioning device uses an external infrared camera to capture infrared markers on the device, achieving spatial coordinate positioning. The IMU and infrared positioning device work together at a sampling frequency of approximately 80Hz, recording spatial coordinates every 12.5 milliseconds to ensure continuous capture even at high speeds. Thus, when a user performs actions such as holding a virtual gene scissors or waving the device, the system can record the spatial coordinates and motion amplitude of the three-dimensional trajectory of the device's core components, such as the fingertips of the scissors or gloves, in real time.

[0073] The velocity change rate is obtained by differential calculation of adjacent frames in the spatial coordinate sequence. Specifically, the system compares the difference in motion amplitude between two consecutive frames at unit time intervals Δt, and calculates the velocity change rate, which is the ratio of the difference in motion amplitude between two consecutive frames to Δt. For example, in this embodiment, the velocity change rate is calculated every approximately 10 to 12.5 milliseconds to capture subtle acceleration or deceleration processes. The velocity change rate is aligned with pressure and spatial coordinates through a unified timestamp to achieve synchronous acquisition of multi-sensor data.

[0074] The aforementioned sensors work together to record every subtle movement of the user during interaction as a digital signal. For example, when a child swings a virtual gene scissor, the position sensor records the amplitude of the scissor glove's movement and the spatial coordinate sequence, while the electronic skin records the continuous pressure generated when the scissor blade contacts the virtual gene chain, and simultaneously calculates the rate of change of the scissor glove's velocity every 10 milliseconds. The temporal feature data of the movements are synchronously fused through timestamps, fully reflecting the real-time process of the user's operation and laying the foundation for subsequent intent analysis.

[0075] The acquired motion timing data is encapsulated in real-time into a time-series multidimensional array. Each data frame in the time-series multidimensional array includes a timestamp, pressure, spatial coordinates, motion amplitude, and velocity change rate. Specifically, the columns of the array are as follows: Timestamp (ms): Records the time of capture of this data frame in milliseconds, aligning data from different sensors to the same time axis. Pressure Value (N): The pressure value detected by the electronic skin in this time frame. If no pressure action occurs in the frame, the pressure is 0 or a background value. Coordinate Position (x, y, z): Three-dimensional coordinates or displacement increments given by the position sensor, reflecting the spatial position change of the user holding the device. Motion Amplitude: Records the amount of motion displacement between adjacent frames. Velocity Change Rate: The difference in motion amplitude between adjacent frames divided by the time interval, i.e., the ratio of velocity change, used to measure the magnitude of acceleration.

[0076] The system performs real-time sliding processing on a multidimensional time series array using a fixed time window. In this embodiment, a time window length of 50ms is selected. That is, whenever 50 milliseconds of continuous data are accumulated, the system performs feature extraction and analysis on the data within the fixed time window. The window continuously covers the entire interaction process using a sliding method: after each short time step, for example, adding a new frame of data every 12.5ms, the window slides forward by the corresponding step, thereby achieving real-time analysis of the action flow.

[0077] The time window is set to 50ms to effectively extract and analyze action features in the time series multidimensional array by accumulating continuous data for this duration. At the same time, with a sliding step size of 12.5ms, the window can continuously cover the entire interaction process, thereby achieving real-time analysis of the action flow and ensuring that the system can respond to dynamic changes in actions in a timely manner.

[0078] Within a 50ms time window, the system preprocesses and extracts motion features from the multidimensional time series array. The motion features include pressure features, velocity features, and amplitude features, as detailed below:

[0079] Extract pressure features, including the peak pressure value and pressure duration within a time window. For example, if the maximum pressure detected within the time window is 0.3 N and the duration is 0.2 s, then record the pressure peak value and the corresponding duration.

[0080] Speed ​​features are extracted, which are statistical values ​​of the rate of change of speed within a time window. For example, the average or maximum rate of change can be used to measure the stability of the action. When the interaction scenario needs to reflect the overall speed fluctuation of the action, the average rate of change of speed is selected; when the interaction scenario needs to reflect the instantaneous speed fluctuation of the action, the maximum rate of change of speed is selected. Large changes in speed features within the time window indicate unstable user actions, such as actions with a maximum rate of change of speed exceeding 50%. Small changes in speed features indicate relatively stable user actions, such as actions with a maximum rate of change of speed less than 20%.

[0081] Amplitude features are extracted, which are the cumulative amplitudes of motion displacement within a time window, used to measure the magnitude of the motion. The cumulative amplitude of the core motion within the same time window should be greater than that of the tentative small-amplitude motions.

[0082] The aforementioned sliding window mechanism ensures that action features are updated every 50ms, effectively capturing subtle changes in user intent. When a new data frame arrives, the old window slides out and the new window slides in, updating the action features accordingly, thus providing continuous and up-to-date input data for subsequent intent recognition models.

[0083] The intent calculation module is used to extract action features from a time series multidimensional array within a set time window; calculate speed stability score, direction stability score and trajectory coherence score based on action features, and set contribution weights for speed stability score, direction stability score, trajectory coherence score and pressure feature, and obtain the intent confidence score by weighted summation.

[0084] Intent confidence is used to quantify the intensity of a user's interaction intent at this moment, comprehensively considering the contribution weights of pressure, time, speed, and amplitude features in the action characteristics. First, the dimensions of each action feature are removed. Specifically, this is done based on the feature value range of core actions in historical interaction data to eliminate the influence of differences in the units of different features: For pressure peak (unit N), the pressure peak range of core actions in historical data is statistically analyzed, for example, the minimum value is 0.1N and the maximum value is 2N. The pressure peak collected in real time is converted into a dimensionless value in the range of 0-1 using the formula "normalized pressure value = (real-time pressure value - historical minimum value) ÷ (historical maximum value - historical minimum value)". For example, when the real-time pressure is 1N, the normalized value = (1-0.1) ÷ (2-0.1) ≈ 0.47. For duration (in seconds), the duration range of historical core actions is statistically analyzed, for example, a minimum of 0.1 seconds and a maximum of 1 second. The real-time duration is converted to a dimensionless value in the 0-1 range using the formula "Normalized Duration = (Real-time Duration - Historical Minimum) ÷ (Historical Maximum - Historical Minimum)". For example, when the real-time duration is 0.5 seconds, the normalized value = (0.5 - 0.1) ÷ (1 - 0.1) ≈ 0.44. The speed stability score, direction stability score, and trajectory continuity score have already been converted to dimensionless values ​​in the 0-1 range using formulas and require no additional processing. After dimensionless processing, all action features are on the same 0-1 scale. When substituted into the intent confidence calculation formula, the weighted calculation of pressure, time, and other features is not affected by unit differences, ensuring the effectiveness of weight allocation and improving the accuracy of intent quantification. In this embodiment, the intent confidence (CI) is calculated as follows:

[0085] CI = (P 峰值 ×W 峰值 +T 持续 ×W 持续 +S 速度 ×W 速度 +S 方向 ×W 方向 +S 轨迹 ×W 轨迹 );

[0086] In the formula, P 峰值 T represents the peak pressure within the time window. 持续 S is the duration of the pressure peak sustained within the time window. 速度 S is used to score speed stability. 方向 S is the score for directional stability. 轨迹 For trajectory coherence score; W 峰值 W is the weighted value for the contribution of the pressure peak within the time window. 持续 W is the weighted value for the duration of the pressure peak within the time window. 速度 W is the weight that contributes to the speed stability score.方向 W is the contribution weight to the directional stability score. 轨迹 The contribution weight to the trajectory coherence score.

[0087] The contribution weight of the speed stability score is set based on the correlation between the speed change rate and the level of action intent. The smaller the speed change rate, the stronger the correlation with the explicit intent, and the higher the contribution weight. The contribution weight of the directional stability score is set based on the correlation between the average directional deviation and the level of action intent. The smaller the average directional deviation, the stronger the correlation with the explicit intent, and the higher the contribution weight. The contribution weight of the trajectory coherence score is set based on the correlation between the average trajectory deviation / trajectory deviation rate and the level of action intent and the conformity to the standard trajectory. The smaller the trajectory deviation or the lower the deviation rate, the stronger the correlation with the explicit intent, and the higher the contribution weight. The pressure peak and duration in the pressure feature have preset contribution weights greater than those of the speed stability score, directional stability score, and trajectory coherence score. The sum of all contribution weights is 1. For example, in the virtual gene scissors cutting scenario, the system presets the contribution weights of each action feature as follows: the contribution weight of the pressure peak is set to 0.4, the contribution weight of the duration of the pressure peak is set to 0.3, the contribution weight of the speed stability score is set to 0.15, the contribution weight of the directional stability score is set to 0.1, and the contribution weight of the trajectory coherence score is set to 0.05. This is because, during the cutting operation, the magnitude of the pressure peak directly reflects the force with which the user grips the scissors; the greater the force, the clearer the cutting intention. The duration reflects the continuity of the user's cutting action; the longer the duration, the higher the focus on the cutting target. Both of these have the most significant impact on the intensity of the action intention, so they are preset with a high weight. On the other hand, the speed stability score reflects whether the cutting action is smooth, the directional stability score reflects whether the cutting direction is aligned with the gene chain, and the trajectory coherence score measures whether the cutting trajectory conforms to the standard arc. Although these scores can help determine the intention, their influence is relatively weak. Therefore, they are preset with a lower weight to adapt to the differences in the contribution of each feature to intention recognition in actual operation.

[0088] The speed stability score is obtained by subtracting the speed characteristic from 1. The speed characteristic reflects the instability of the motion speed; the larger the value, the more drastic the speed fluctuation. By subtracting the speed characteristic from 1, the speed characteristic is converted into a speed stability score. That is, the more stable the speed, the closer the speed stability score is to 1; conversely, the more drastic the fluctuation, the closer the speed stability score is to 0. For example, if the speed characteristic measured within the time window is 40%, then the speed stability score = 1 - 0.4 = 0.6.

[0089] The directional stability score reflects the consistency of the user's movement direction when waving the motion-sensing device. The system uses gyroscope data to calculate the angular deviation between the motion direction in each frame and the preset target direction, and further calculates the average directional deviation over several consecutive frames. The preset target direction is the expected motion direction for a specific interaction scenario, determined based on the target of the operation in the scenario, such as the vertical downward direction when touching a virtual cell or the cutting direction when operating virtual gene scissors, serving as a reference benchmark for measuring the consistency of the motion direction. The directional stability score can be defined as: S 方向 = 1 - (Average Directional Deviation / 90°). Setting 90° as the normalization upper limit angle, that is, if the continuous motion direction has almost no deviation, and the deviation is close to 0°, then the directional stability score is approximately 1; if the directional deviation is large, for example, deviating from the target direction by 30°, then S... 方向 =1-30 / 90≈0.67. The directional stability score decreases significantly when the direction of the action is variable, thus reducing the weight of actions with no clear directional intention in the calculation of intention confidence.

[0090] The trajectory coherence score measures the smoothness and consistency of the actual motion trajectory. The system reconstructs the actual motion path from the continuous frame spatial coordinate sequence of the user's operation of the motion-sensing interaction device in chronological order and compares it with a predefined standard trajectory. The standard trajectory is an ideal motion path preset for a specific interaction scenario; for example, touching a cell should be a straight line with a vertical downward press, and a scissor cutting operation should be a smooth arc. The trajectory deviation is the distance between the actual trajectory point and the corresponding point on the standard trajectory. When the interaction scenario requires the user to perform an action according to a fixed standard trajectory point, the trajectory coherence score, i.e., S, is calculated based on the distance deviation of each frame spatial coordinate sequence. 轨迹= 1 - (Average trajectory deviation / Preset tolerance). The smaller the deviation, the higher the trajectory continuity score. In this type of scenario, the absolute distance deviation between the actual trajectory and the standard trajectory directly reflects the accuracy of the operation. For example, the smaller the deviation, the higher the stability of the user's execution according to the preset path, and the score is more in line with the judgment of smoothness and consistency. When the interaction scenario does not require the user to perform actions according to fixed standard trajectory points, the trajectory continuity score is calculated by evaluating the trajectory deviation rate, that is, the proportion of the cumulative deviation to the amplitude feature. That is, the trajectory continuity score is high when the trajectory deviation rate is low. The trajectory deviation rate is measured by the proportion of the cumulative deviation to the total action amplitude. It can avoid misjudging the trajectory continuity due to the high deviation value caused by the large absolute amplitude. It can more objectively reflect the degree of relative deviation. Even if the absolute deviation is slightly large, but the proportion of the total amplitude is low, such as a deviation of 1cm accounting for 10% of a 10cm amplitude, it can still be judged as trajectory continuity, which is suitable for the needs of scenarios with variable action amplitude. The system will use the trajectory continuity score to dynamically adjust the intention dynamic threshold. For example, if the trajectory deviation of an adult user's action is consistently very small, the intention dynamic threshold requirement can be temporarily reduced. The dynamic threshold of intent is used as a standard to judge the level of action intent. It is dynamically adjusted in combination with user type. For example, the dynamic threshold of intent is lowered for children and raised for adults. This distinguishes between low-intent actions, transitional actions and core actions. The initial value is only used as an initial reference.

[0091] Through the user intent recognition model, the system can output an intent confidence value at each time window, assessing the strength of the user's current action intent in real time. As the user gradually transitions from preparatory / exploratory actions to explicit core actions, the characteristics of each action will show significant differences: for example, increased pressure peak, extended duration, and improved speed stability score, thereby causing a corresponding increase in intent confidence, providing a basis for the next step of action classification and feedback scheduling.

[0092] The intent recognition module is used to classify user actions into different intent levels in real time based on intent confidence and set dynamic intent thresholds, and allocate corresponding feedback computing resources according to the classified intent levels.

[0093] Please refer to Figure 3. The intent levels include low-intent actions, transitional actions, and core actions, and the specific classification criteria are as follows:

[0094] When the intent confidence level is lower than the set intent dynamic threshold for a low intent ratio, the user's action is classified as a low intent action (P3 level). The low intent ratio is set based on the characteristic differences between low intent actions and core actions. For example, the pressure and duration of low intent actions are usually less than 70% of those of core actions. Therefore, the low intent ratio can be set to 0.7 to clearly distinguish low intent actions from actions of other intent levels. For example, in this embodiment, the low intent ratio is set to 0.7 and the intent dynamic threshold is 0.6. Then, when CI < 0.42, the user's action is considered a low intent action. Low intent actions include minor adjustments or tentative touches by the user before performing the main operation, where the intent is unclear. For P3 level actions, the system marks them as the lowest priority and allocates only a small amount of feedback computing resources. The feedback computing resources are the system's computing power used to process action data, calculate feedback logic, and drive the haptic feedback device, supporting the operation of the feedback mechanism and ensuring that feedback can respond to actions in a timely manner. A small amount of feedback computing power is allocated based on the processing needs of low-intention actions, specifically approximately 5% of the total system feedback computing power. This is because low-intention actions do not require complex feedback and only need a small amount of computing power to process basic data, avoiding the consumption of large resources required by core actions. This value is determined based on the total system computing power and the processing load of low-intention actions to balance resource utilization and feedback efficiency. The system sets a "preemptible" flag for the feedback computing power resources of low-intention actions, meaning that the feedback computing power resources can be immediately interrupted and reclaimed if a higher priority demand arises.

[0095] When the intent confidence level falls between the set low intent ratio threshold and the set intent dynamic threshold, the user's action is classified as a transitional action (P2 level). Transitional actions are in an intermediate state between the preparation phase and the core action, such as positioning calibration or minor corrections before the core action. Transitional actions are characterized by: peak pressure approximately 50%–70% of a typical core action, duration approximately 0.3–0.5 seconds, and speed change rate of 20%–50%. The system marks these actions as P2 level medium priority and allocates a relatively small amount of feedback computing resources. Specifically, the allocated feedback computing resources can be set to approximately 30% of the total system feedback computing resources, depending on the processing requirements of the transitional action. This is because transitional actions, as an intermediate state between the preparation phase and the core actions, require handling feedback logic of moderate complexity, such as positioning calibration and minor corrections. They need sufficient computing power to generate moderately strong or delayed small feedback to alert the user, while avoiding excessive resource consumption that could affect the processing of the core actions. This value is determined based on the processing load and system resource balance corresponding to the peak pressure and duration of the transitional action, achieving a balance between alerting the user and ensuring resources are available for the core actions. Corresponding vibration or resistance feedback can be moderately strong or delayed, used to alert the user without prematurely consuming excessive resources. Similarly, the system sets a "preemptible" flag for the feedback computing resources of transitional actions. When the core action arrives, P2-level tasks can also be preempted by the core action, but with higher priority than P3 to ensure the necessity of normal transitional feedback.

[0096] When the intent confidence level reaches or exceeds the intent dynamic threshold, such as CI ≥ 0.6, the user's action is classified as a core action (P1 level). Core actions carry the main interaction purpose, such as deterministic pressing or cutting, and should be processed with the highest priority. The system marks them as P1 level, allowing them to occupy all feedback computing resources. Once a P1 level action is detected, the system immediately triggers a resource preemption mechanism: within 50ms, feedback computing resources marked as "preemptible" are suspended, for example, terminating ongoing P3 level haptic feedback calculations, and the freed-up feedback computing resources are quickly reallocated to core action processing. At the same time, the driving rights of key feedback hardware such as haptic motors are exclusively reserved and locked to the core action until the core action is completed. By estimating the duration of the pressure feature in the action characteristics, the system can predict when the core action will end, and release feedback computing resources or reduce the feedback intensity at the appropriate time. This scheduling strategy ensures that the feedback required by the core action can be responded to in a timely and sufficient manner, without delay due to previous irrelevant actions, such as the hardness of a virtual cell when pressed or the resistance of gene scissors cutting, providing the user with timely feedback.

[0097] Resource preemption and priority allocation mechanisms greatly improve the utilization efficiency of system feedback computing resources and the accuracy of feedback responses. The system guarantees that the higher the confidence level of the intent, the more resources are allocated to the corresponding user's action, thus optimizing the matching degree between interactive feedback and the user's true intent.

[0098] The threshold adjustment module adaptively adjusts the dynamic threshold of intent and the weight of each contribution based on pre-collected user profile data.

[0099] The user profile data includes the user's age group, the user's operation history trajectory data, and corresponding action feature data.

[0100] Considering the differences in interactive actions between children and adults, this system introduces user profile data to dynamically adjust the dynamic threshold of intent and the weight of action features, so that the algorithm can adapt well to different groups of people.

[0101] Before a user begins interacting with the system, the system obtains the user's age group information through interface options or user registration, initially classifying the user into children or adults. In addition, the system continuously records the user's operation history trajectory data, saving a complete action sequence every 5 interactions, including action characteristics such as stress peak, average action amplitude, and trajectory stability. Every day, the system performs cluster analysis on newly added data, updating user profiles based on age group information and operation trajectory data. Users are subdivided into two main categories: children-fluctuating and adults-stable. Specifically: First, all newly added user operation records for the day are collected, each record is associated with corresponding age group information, and initially classified as either children or adults. Key action features are extracted from each record, including peak stress standard deviation, average action amplitude, and trajectory deviation rate, forming a feature dataset. Subsequently, the feature dataset is divided into children's and adults according to age group information, and clustering is performed separately. For example, the density-based clustering algorithm DBSCAN is used, with peak stress standard deviation, average action amplitude, and trajectory deviation rate as clustering dimensions. A reasonable neighborhood radius is set, and the minimum sample size is calibrated based on the similarity of peak stress standard deviation, average action amplitude, and trajectory deviation rate within groups and the differences in peak stress standard deviation, average action amplitude, and trajectory deviation rate between groups in historical clustering. Records with similar characteristics are clustered together. For the children's group, records with a peak stress standard deviation greater than a set children's standard deviation threshold, an average movement amplitude greater than a children's movement amplitude threshold, or a trajectory deviation rate greater than a children's deviation threshold are labeled as "child-fluctuating type." For the adults' group, records with a peak stress standard deviation less than an adults' standard deviation threshold, an average movement amplitude less than an adults' movement amplitude threshold, or a trajectory deviation rate less than an adults' deviation threshold are labeled as "adult-stable type." After clustering, the mean and standard deviation of key movement features for children and adults are calculated, and the standard deviation threshold, movement amplitude threshold, and deviation threshold are fine-tuned accordingly. For example, the peak stress standard deviation threshold for the children's group can fluctuate by 5% based on the mean of newly added data. Each user's operation records are matched with the clustering results, and their profile labels are updated. Finally, the updated user profiles and threshold parameters are stored in the system for the next day's intent recognition and feedback scheduling, ensuring that the classification always aligns with the user's latest operating habits. Through daily clustering and adaptive updates of threshold parameters, the system ensures accurate and timely characterization of user operating habits.

[0102] The setting of various thresholds, such as the standard deviation threshold for children, the amplitude threshold for children's movements, and the standard deviation threshold for adults, is determined in stages based on historical interaction data and combined with the statistical patterns of user operation characteristics, as detailed below:

[0103] In the initial threshold setting phase, for the children's group, the distribution range of key movement features is calculated, and the upper limit of the occurrence of key movement features is taken as the initial threshold. For example, the 80th percentile of the standard deviation of peak stress in children is 0.3N, so the initial standard deviation threshold for children is set to 0.3; the 80th percentile of the average movement amplitude is 5cm, so the initial movement amplitude threshold for children is set to 5; and the 80th percentile of the trajectory deviation rate is 15%, so the initial deviation threshold for children is set to 15. For the adults' group, the lower limit of the occurrence of feature values ​​is taken as the initial threshold. For example, the 20th percentile of the standard deviation of peak stress in adults is 0.2N, so the initial standard deviation threshold for adults is set to 0.2; the 20th percentile of the average movement amplitude is 3cm, so the initial movement amplitude threshold for adults is set to 3; and the 20th percentile of the trajectory deviation rate is 8%, so the initial deviation threshold for adults is set to 8. This ensures that the initial thresholds can distinguish the typical operating characteristics of the two types of users.

[0104] During the daily update phase, the system calculates the mean and standard deviation of key features for both the children's and adult groups based on the clustering results of the newly added data for that day, and then fine-tunes the initial thresholds. For example, if the mean standard deviation of the peak pressure for the children's group increases by 3% compared to the historical average, the children's standard deviation threshold will be increased by 3%, such as from 0.3N to 0.309N; if the mean average movement amplitude for the adult group decreases by 2% compared to the historical average, the adult movement amplitude threshold will be decreased by 2%, such as from 3cm to 2.94cm. The fine-tuning range is controlled within ±5% to avoid drastic fluctuations in the thresholds, ensuring that they always align with the current user group's operating habits and providing a stable benchmark for user profile segmentation.

[0105] Please refer to Figure 4. Based on user profile data, the system configures the dynamic threshold of intent confidence and the weights of each contribution differently, as follows:

[0106] A reinforcement learning model is employed to adaptively adjust the contribution weights of speed stability score, direction stability score, trajectory coherence score, and pressure feature. The reinforcement learning model aims to maximize intent recognition accuracy as its reward objective. The state space includes the user type, real-time action features, and the weight adjustment effects of historical interactions. The action space represents the adjustment amount of each contribution weight, with the sum of all contribution weights constrained to 1 and the contribution weight of each individual feature within a preset range. The reinforcement learning model is trained using historical interaction records labeled with action features, user type, and actual intent level. After pre-training, an initial weight strategy is obtained and then optimized. During optimization, the reinforcement learning model loads the initial weights according to the user type, extracts real-time action features through a set time window, combines the user type and historical recognition results to form a state, inputs it into the reinforcement learning model, and outputs contribution weights. Intent confidence is calculated based on the output contribution weights. If the intent confidence matches the actual intent level, a positive reward is given; otherwise, a negative reward is given, driving the reinforcement learning model to update its weight strategy.

[0107] For example, the state space includes user type, real-time action features, and the weight adjustment effects of the past 5 interactions; the action space is the adjustment amount of the contribution weights of each feature, with the total weight constrained to 1 and the individual weight value ranging from 0.05 to 0.6. The training data uses over 100,000 historical interaction records, each labeled with action features, user type, and actual intent level. The initial weight strategy is obtained through supervised pre-training and then optimized online through reinforcement learning. In the real-time process, the reinforcement learning model first loads the initial weights according to the user type; every 50ms, a sliding window extracts real-time action features, which, together with the user type and historical recognition results, constitute the state input model; the reinforcement learning model outputs the contribution weights of each feature based on the current state. For example, when a child user detects a sharp directional fluctuation, the weight of directional stability is increased and the weight of trajectory coherence is decreased; when an adult user performs a fine operation, the weights of pressure peak and trajectory coherence are increased; if the intent confidence based on the output weights accurately matches the actual action level, the reinforcement learning model receives a +1 reward; otherwise, it receives a -0.5 reward to drive policy updates. The reinforcement learning model updates the Q-value table with new data after every 100 interactions. The activation function forces the total weight to be 1 and the weight of each individual weight to not exceed the upper limit, ensuring that the adjustment adapts to changes in user operation habits.

[0108] The adjusted formula for calculating the confidence level of a child's intention is shown below:

[0109] CI = (P 峰值 ×0.3+T 持续 ×0.2+S 速度 ×0.1+S 方向 ×0.3+S 轨迹 ×0.1);

[0110] The adjusted formula for calculating adult intent confidence is shown below:

[0111] CI = (P 峰值 ×0.5+T 持续 ×0.2+S 速度 ×0.1+S 方向 ×0.1+S 轨迹 ×0.1);

[0112] Increasing the contribution weight of pressure features allows actions that apply pressure continuously but with small force to obtain higher intention confidence, thereby overcoming the problem of missing core actions due to insufficient action amplitude.

[0113] Personalized intent recognition is achieved through dynamic adjustment of the aforementioned intent threshold. Daily updates to cluster labels and dynamic intent thresholds ensure the system evolves along with the behavioral patterns of the user group. This prevents children from losing patience due to frequent changes in feedback, while also ensuring that adults receive timely feedback for accurate operations, enhancing the immersive experience of the science popularization interaction.

[0114] The following specific scenario illustrates how the method of the present invention can complete intent recognition, resource scheduling, and feedback response within 50ms.

[0115] In the virtual cell touch scenario, the user uses a haptic stylus to touch virtual cells on the electronic skin surface of the interactive screen. Initially, the user might lightly probe the cell surface with the stylus tip, applying pressure of approximately 0.3N for about 0.2 seconds, with a speed change rate as high as 60%, resulting in a slightly unstable movement. The system detects these characteristics within a 50ms window, calculating an intent confidence score of approximately 0.28, which is below the dynamic threshold of 70%. Therefore, this action is classified as a low-intent tentative touch, and the system allocates only the minimum 5% of the P3 level feedback computing power resources, providing extremely weak tactile feedback, such as a slight touch sensation, and marking this feedback computing power resource as immediately available for preemption. Immediately afterward, the user confirms the location and begins to actually press the cell, with the pressure suddenly increasing to 1.2N for about 0.5 seconds, the speed change rate decreasing to 15%, and the movement becoming stable. This new feature causes the intent confidence score to jump to 0.92, exceeding the threshold of 0.6. The system immediately identified the core pressing action, upgraded it to P1 priority, and triggered resource preemption: within 50ms, it terminated the previous P3-level feedback process, releasing and concentrating 80% of the feedback computing resources for cellular pressing feedback. The haptic module immediately output simulated cellular hardness resistance. The entire process ensured that the user felt the firm cellular feedback within approximately 100ms after the pressing action began, synchronized with the application of pressure. By promptly stopping ineffective feedback from tentative touches and prioritizing the response to core pressing, the consumption of feedback computing resources for low-intention actions was reduced by 90% in this scenario, and the user's subjective perception of feedback delay almost disappeared.

[0116] Please refer to Figure 5. In the virtual gene scissors cutting scenario, the user holds a haptic scissor device to perform a cutting operation on the virtual gene chain. Before starting the actual cutting, the user may make several empty cuts or swinging motions aimed at the target. For example, children often make large swings of about 5cm without intending to close the scissors to cut. These movements have very low pressure, such as less than 0.3N, and the direction changes drastically. Older systems might frequently trigger scissor resistance feedback due to the large amplitude, causing meaningless resistance abrupt changes. In this invention, the system monitors the directional stability score and pressure characteristics. Since the intention confidence is much lower than the dynamic threshold of 0.45 for children's intentions, such swings are judged as low-intention transitional movements. Therefore, these swings are only assigned extremely low resources at the P3 level, and the haptic device remains basically idle or only provides a slightly relaxed resistance sensation, without frequently triggering strong resistance feedback. Subsequently, when the user performs a precise cutting motion on the gene chain, characterized by the scissors rapidly closing along a pre-defined arc path, a sharp increase in pressure (e.g., 0.8N when gripping the scissors), a significant reduction in directional deviation (the scissors' direction of movement deviates from the target by only 15°), lasting approximately 0.6 seconds, the system detects a confidence jump exceeding the intention dynamic threshold. At this point, the action is identified as a core cut and elevated to P1 level. The system immediately applies simulated shearing resistance; that is, at the instant the scissors close, the user's hand experiences the realistic resistance and vibration feedback of the virtual gene chain being cut. Through resource preemption, any previous low-level feedback is canceled, ensuring that the shearing resistance is output synchronously within tens of milliseconds, maintaining consistency with the rhythm of the scissors closing. For similar precise cuts by adult users, the system, by dynamically reducing the intention dynamic threshold and increasing the contribution weight of pressure features, can also correctly identify the core action and provide timely resistance feedback, avoiding the missed triggering of important feedback due to insufficient action amplitude in traditional methods.

[0117] In summary, this embodiment acquires subtle temporal features of user actions in a timely manner through high-speed multi-sensor acquisition and sliding window feature extraction. It employs an intent confidence model to fuse action features such as pressure, time, and speed, and quantitatively assesses the intensity of user intent by combining speed stability score, directional stability score, and trajectory coherence score. Based on intent confidence, it performs three-class classification of actions and uses a pre-defined resource priority table for on-demand allocation and preemptive scheduling of computing power and feedback devices. Simultaneously, it dynamically adjusts the dynamic threshold of intent and the contribution weight of action features using user profile data, optimizing recognition sensitivity for different user groups. This method effectively avoids the waste of system resources and erroneous feedback caused by low-intent actions, ensuring timely and accurate feedback for core interactions, and significantly enhancing the immersive experience and scientific accuracy of interactive life science displays.

[0118] Example 2

[0119] Existing feedback mechanisms follow the logic of responding immediately after an action is completed. This disconnects from the rhythm at which users expect feedback during the core operation, further exacerbating the sense of incongruity in the interaction. For example, when a user operates a virtual syringe to simulate drug injection, they expect to feel the resistance of the drug being pushed halfway through the injection. However, the system waits until the entire injection is completed before triggering feedback, causing a mismatch between the user's actual experience and their expected rhythm, creating a sense of disconnect where the action has been performed but the feedback is not synchronized.

[0120] Based on this, this embodiment provides an interactive life science demonstration method, which also includes:

[0121] The rhythm matching module estimates the degree of action completion during the user's actions and dynamically adjusts the timing of feedback triggering based on different user profile data, and introduces a cross-scene rhythm deviation adaptive mechanism.

[0122] To further align with users' desired interaction rhythm, this embodiment provides an action completion prediction and feedback trigger rhythm matching mechanism. The system estimates the action completion rate during the action's execution and dynamically adjusts the timing of feedback triggering based on different user profile data. It also introduces a cross-scenario rhythm deviation adaptive mechanism to ensure that feedback is synchronized with the user's action progress.

[0123] The system does not rely on pre-set fixed duration parameters, but dynamically predicts the total duration of the current action based on real-time motion characteristics. The specific method is as follows:

[0124] The total expected amplitude is a fixed parameter preset for different interaction scenarios, set based on the typical amplitude characteristics of the core action in that scenario: for example, the total expected amplitude of the virtual cell touch scenario is preset to 3cm, which can be obtained by referring to the average amplitude of the core pressing action in the historical data of that scenario; the total expected amplitude of the virtual gene scissors cutting scenario is preset to 5cm, which can be obtained based on the spatial trajectory length of the typical cutting action.

[0125] Based on the preset total expected amplitude, the criterion for determining the first 30% of the process is as follows: when the cumulative real-time action amplitude reaches 30% of the total expected amplitude, it is defined as entering the first 30% stage. For example, when the total expected amplitude is 3cm, a cumulative real-time amplitude of 0.9cm is considered to be in the first 30% of the process; when the total expected amplitude is 5cm, a cumulative real-time amplitude of 1.5cm triggers the first 30% monitoring. During this stage, the system calculates the average speed and elapsed time, and then uses the formula: Total duration of the current action = Elapsed time ÷ 30%, to estimate the total duration of the current action. This ensures that the process division is based on a clear preset benchmark, avoids deviations caused by ad-hoc definitions, and improves the stability of total duration estimation and feedback triggering.

[0126] Based on user profile data, the system predefines typical action rhythm categories for different user groups. Based on historical operation records, the speed-time distribution of user actions is clustered into two types: a slow-start, slow-fall type for children and a fast-start, steady-fall type for adults. Specifically: First, a sufficient amount of historical user operation records is accumulated, such as 100,000+ interaction records. Each record contains complete action time sequence feature data, and the speed-time distribution curves, i.e., the speed value sequence at different time points, are extracted. Then, a clustering algorithm is used to group these speed-time curves. By analyzing the morphological characteristics of the curves, such as the time position of the speed peak and the ratio of the acceleration phase to the deceleration phase, the curves are clustered into two categories. Among them, the curves of children are characterized by a long acceleration process from the start to the peak, with the peak appearing in the middle of the action, and a smooth deceleration process from the peak to the end, i.e., the slow-start, slow-fall type. The curves of adults are characterized by a rapid rise in speed to the peak, i.e., the peak appears in the early stage of the action, and then remains relatively stable or decelerates slowly to the end, i.e., the fast-start, steady-fall type. Finally, the two rhythm types are associated with the child-fluctuating and adult-stable labels in the user profile and stored as the system's default rhythm classification template for subsequent matching of feedback trigger intervals based on user type.

[0127] Among them, children's slow-start, slow-finish type is characterized by a slow start to the action, reaching peak speed only in the middle and later stages, with relatively smooth acceleration and deceleration throughout the process. Most children exhibit an overall slow rhythm in various interactions such as touching, swiping, and pressing. Adult users' fast-start, steady-finish type is characterized by a rapid acceleration to peak speed in the early stages of the action, followed by a smooth finish. Adults often approach the target with force at the beginning of the action, and then maintain a constant speed or make small adjustments to complete the operation.

[0128] Based on different user groups, typical action rhythm categories are defined, and the system sets general feedback trigger interval parameters so that the timing of feedback activation matches the user's expected rhythm point.

[0129] For children's slow-starting, slow-ending movements, the basic interval for feedback triggering is set at 50%–60% of the total movement duration. That is, feedback is provided at an appropriate point in the middle when the prediction is more than halfway complete. For example, if a child's movement is estimated to last about 2 seconds, the feedback signal will be triggered at approximately 1.0–1.2 seconds.

[0130] For adult users' fast-paced, steady-motion actions, the feedback trigger interval is advanced to 20%–30% of the total duration. This means providing feedback shortly after the action begins, satisfying adult users' expectation of immediate feedback in the early stages. If an adult action is expected to be completed in 1 second, the system will trigger the corresponding feedback in approximately 0.2–0.3 seconds.

[0131] Furthermore, the system dynamically adjusts the trigger interval based on the stability of the action type. This is achieved by calculating the variance σ of the rate of change of velocity during an action segment. 2 Determine whether the action is stationary or fluctuating. If σ 2 A very small value, such as less than 5%, indicates smooth and stable movement; if σ 2 A large value, such as greater than 15%, indicates significant jitter or unevenness in the movement rhythm. For fluctuating movements, the system appropriately widens the trigger interval by approximately 20% to tolerate rhythm instability; for steady movements, the default interval is maintained. For example, the default 50%–60% trigger interval for children will be widened to 45%–65% when the movement is judged to be fluctuating; the default 20%–30% interval for adults will be widened to 15%–35% when the movement is fluctuating. In this way, when the user's movements are inconsistent and not smooth, the feedback trigger timing will be dynamically adjusted within a wider range, avoiding missing the optimal feedback point or causing misjudgment due to rhythm fluctuations. When the movement is highly stable, the feedback trigger timing is relatively concentrated, thus precisely meeting the user's requirement for feedback synchronization under a stable rhythm.

[0132] To adapt to different interaction scenarios and continuously optimize feedback synchronization, the system also introduces a cross-scenario rhythm deviation adaptive mechanism, dynamically adjusting compensation values ​​based on differences in scene action characteristics. Different interaction scenarios exhibit inherent differences in action characteristics: for example, the virtual cell touch scenario primarily involves vertical pressing with a small amplitude, typically between 2cm and 5cm, and significant pressure characteristics, with the core action pressure exceeding 1N; the virtual gene scissors cutting scenario primarily involves arc-shaped swinging with a large amplitude, typically between 5 and 10cm, and a more prominent speed change rate characteristic, with the core action speed change rate less than 30%. Based on historical data, the system pre-classifies scenarios into types such as "pressing," "swinging," and "drag" according to "action amplitude range + dominant characteristic (pressure / speed)," assigning scene feature weights to each type. For example, the pressure feature weight for the pressing scenario is 0.6, and the speed feature weight is 0.4; the opposite applies to the swinging scenario.

[0133] Each time feedback is triggered, the system calculates the rhythm matching deviation: the rhythm matching deviation is the difference between the actual feedback trigger time and the theoretically planned trigger time, where the theoretically planned trigger time = action start time + estimated total duration × trigger completion rate. If a rhythm matching deviation exceeds a set threshold, for example, greater than 100ms, a global dynamic compensation value is invoked and adjusted based on the feature weights of the current scene type. For example, in a press-type scene, which is more sensitive to pressure feedback delay, the compensation value is increased by 10%; in a swing-type scene, which is more sensitive to speed feedback delay, the compensation value is decreased by 5%, before being applied to subsequent triggers.

[0134] Meanwhile, if a certain type of action is detected to have a large rhythm matching deviation in multiple consecutive interactions, such as a steady action having a rhythm matching deviation greater than 80ms for 5 consecutive times, the trigger interval coefficient of the corresponding action type will be updated, for example, from 1.0 to 1.1, and further fine-tuned according to the scene type: the coefficient adjustment range for the same action type is 1.0 to 1.1 in the pressing scene and 1.05 to 1.15 in the swinging scene, to ensure that the coefficient is adapted to the scene characteristics.

[0135] This mechanism's cross-scenario transfer capability is based on a dual matching of scenario type and user type: the compensation parameters calculated in one interaction scenario, such as the need for a child user to respond 10ms in advance in a "virtual touch" (press type) scenario, can be directly reused when transferred to other similar scenarios such as "virtual injection push"; when transferred to different types of scenarios, such as "virtual gene scissors cutting" (swing type), it needs to be converted through scenario feature weights, such as 10ms in advance for press type corresponding to 8ms in advance for swing type, and then applied to the feedback triggering of the same type of user.

[0136] Through this closed-loop correction that incorporates scene characteristics, the system's rhythm matching strategy adapts to scene differences while maintaining cross-scene universality. This ensures that whether users are touching cells, dragging gene chains, or performing other interactive operations, they can enjoy accurate feedback rhythms that match the scene's characteristics. By matching action completion with rhythm, the system's feedback triggering has shifted from the traditional "immediate response after action completion" to "rhythmic response during action." For example, in simulated injection, where users expect to feel resistance or other process-related feedback during the injection process, the system can output some resistance tactile feedback halfway through the injection, synchronizing with the action and avoiding the disjointed feeling caused by waiting for feedback to finish. This mechanism significantly enhances the user's immersive experience, ensuring that the rhythm of interactive feedback matches the user's expectations.

[0137] Example 3

[0138] Please refer to Figure 2. This embodiment provides an interactive life science demonstration method, including:

[0139] Collect action time-series feature data and construct a multi-dimensional time series array;

[0140] Within a set time window, action features are extracted from a time series multidimensional array; the action features include pressure features, speed features, and amplitude features; based on the action features, speed stability scores, directional stability scores, and trajectory coherence scores are calculated, and contribution weights are assigned to the speed stability scores, directional stability scores, trajectory coherence scores, and pressure features, and the intention confidence is obtained by weighted summation;

[0141] Based on the intent confidence level and the set intent dynamic threshold, the user's actions are divided into different intent levels in real time, and corresponding feedback computing resources are allocated according to the divided intent levels.

[0142] Based on pre-collected user profile data, the dynamic threshold of intent and the weight of each contribution are adaptively adjusted.

[0143] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0144] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An interactive life science demonstration method, characterized in that, include: Action temporal feature data is collected to construct a time series multidimensional array. The action temporal feature data includes the continuous pressure value applied by the user when touching the device, the action amplitude and spatial coordinate sequence when the user waves the motion sensing device, and the velocity change rate of the spatial coordinate sequence of adjacent frames. The action amplitude is the Euclidean distance between the spatial coordinates of two consecutive frames in the spatial coordinate sequence. Within a set time window, action features are extracted from the time series multidimensional array. The action features include pressure features, velocity features, and amplitude features. Extract pressure features, including the peak pressure value and pressure duration within a time window; extract velocity features, which are the average or maximum rate of change of velocity within a time window; extract amplitude features, which are the cumulative amplitude of motion displacement within a time window. Based on action features, speed stability score, direction stability score, and trajectory coherence score are calculated. Specifically, the speed stability score is obtained by subtracting the speed feature within the time window from 1; the direction stability score is defined as: direction stability score = 1 - (average direction deviation / preset maximum deviation angle); the average direction deviation is obtained by calculating the angle deviation between the motion direction of each frame and the preset target direction, and taking the average of the direction deviations of several consecutive frames; when the interaction scenario requires the user to perform actions according to fixed standard trajectory points, the trajectory coherence score is calculated based on the trajectory deviation between the actual trajectory points formed by the spatial coordinate sequence of each frame and the standard trajectory points, trajectory coherence score = 1 - (average trajectory deviation / preset tolerance); the trajectory deviation is the distance between the actual trajectory point and the corresponding point of the standard trajectory; when the interaction scenario does not require the user to perform actions according to fixed standard trajectory points, the trajectory coherence score is calculated by evaluating the trajectory deviation rate, that is, the proportion of the cumulative trajectory deviation to the cumulative amplitude of the action displacement within the time window; these are the speed stability score, direction stability score, and trajectory coherence score. The intention confidence score is obtained by weighting the contribution weights of stress features and summing them. Based on the intention confidence score and a set dynamic threshold, user actions are divided into different intention levels in real time, and corresponding feedback computing resources are allocated according to the divided intention levels. The dynamic threshold and contribution weights are adaptively adjusted based on pre-collected user profile data. The user profile data acquisition method includes: collecting all newly added user operation records daily, associating each record with corresponding age group information, and extracting key action features from each record. These key action features include the standard deviation of stress peak, average action amplitude, and trajectory deviation rate, forming a feature dataset. The feature dataset is divided into children's and adults according to age group information, and clustering is performed separately. The clustering dimensions are the standard deviation of stress peak, average action amplitude, and trajectory deviation rate. A neighborhood radius is set, and the minimum sample size is calibrated based on the similarity of stress peak standard deviation, average action amplitude, and trajectory deviation rate within groups and the differences between groups in these parameters during historical clustering. After clustering, the final user profile data is obtained.

2. The interactive life science demonstration method according to claim 1, characterized in that, Within a set time window, methods for extracting motion features from a multidimensional array of time series include: when the interaction scene needs to reflect the overall speed fluctuation of the motion, the average speed change rate is selected; when the interaction scene needs to reflect the instantaneous speed fluctuation of the motion, the maximum speed change rate is selected; the sliding window mechanism ensures that the motion features are updated once within each time window. When a new data frame arrives, the old window slides out and the new window slides in, and the motion features are updated accordingly.

3. The interactive life science demonstration method according to claim 1, characterized in that, The method for classifying user actions into different intent levels in real time based on intent confidence and a set intent dynamic threshold includes: the intent levels include low-intent actions, transitional actions, and core actions; when the intent confidence is lower than a first intent dynamic threshold for the proportion of low intent, the user's action is determined to be a low-intent action; when the intent confidence is between the first intent dynamic threshold and the second intent dynamic threshold for the proportion of low intent, the user's action is determined to be a transitional action; otherwise, the user's action is determined to be a core action.

4. The interactive life science demonstration method according to claim 3, characterized in that, The method for allocating corresponding feedback computing resources according to the divided intent levels includes: low intent actions, which are marked as the lowest priority and allocated a set minimum priority ratio of feedback computing resources; transitional actions, which are marked as medium priority and allocated a set medium priority ratio of feedback computing resources; and core actions, which are marked as the highest priority and can occupy all feedback computing resources. The feedback computing resources are the computing capabilities of the system used to process action data, calculate feedback logic, and drive the haptic feedback device.

5. The interactive life science demonstration method according to claim 1, characterized in that, The method for setting contribution weights for speed stability score, directional stability score, trajectory coherence score, and stress feature includes: the contribution weight of speed stability score is set based on the correlation between speed change rate and action intention level; the smaller the speed change rate, the higher the contribution weight. The contribution weight of directional stability score is set based on the correlation between average directional deviation and action intention level; the smaller the average directional deviation, the higher the contribution weight. The contribution weight of trajectory coherence score is set based on the correlation between average trajectory deviation / trajectory deviation rate and action intention level and conformity to standard trajectory; the smaller the trajectory deviation or the lower the deviation rate, the higher the contribution weight. The preset contribution weights for stress peak and duration in stress feature are greater than the contribution weights of speed stability score, directional stability score, and trajectory coherence score. The sum of all contribution weights is 1.

6. The interactive life science demonstration method according to claim 1, characterized in that, The methods for obtaining user profile data include: for the children group, records with a peak stress standard deviation greater than a set children's standard deviation threshold, an average movement amplitude greater than a children's movement amplitude threshold, or a trajectory deviation rate greater than a children's deviation threshold are marked as children-fluctuating type; for the adults group, records with a peak stress standard deviation less than an adults' standard deviation threshold, an average movement amplitude less than an adults' movement amplitude threshold, or a trajectory deviation rate less than an adults' deviation threshold are marked as adults-stable type.

7. The interactive life science demonstration method according to claim 1, characterized in that, The method for adaptively adjusting the dynamic threshold of intent and the weights of each contribution based on pre-collected user profile data includes: using a reinforcement learning model to adaptively adjust the contribution weights of speed stability score, direction stability score, trajectory coherence score, and pressure feature, wherein the reinforcement learning model aims to maximize the accuracy of intent recognition; the state space includes the user type, real-time action features, and the weight adjustment effect of historical interactions; the action space is the adjustment amount of each contribution weight, and the sum of each contribution weight is constrained to be 1, and the contribution weight of a single feature is within a preset range; the reinforcement learning model is trained through historical interaction records, which are labeled with action features, user type, and actual intent level, and the initial weight strategy is obtained after pre-training and then optimized; during the optimization process, the reinforcement learning model loads the initial weights according to the user type, extracts real-time action features through a set time window, combines the user type and historical recognition results to form a state, inputs it into the reinforcement learning model, and outputs the contribution weights; the intent confidence is calculated based on the output contribution weights, and if the intent confidence matches the actual intent level, a positive reward is given, otherwise a negative reward is given, driving the reinforcement learning model to update the weight strategy.

8. An interactive life science demonstration system, used to implement the interactive life science demonstration method according to any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect action temporal feature data and construct a time series multidimensional array. The action temporal feature data includes the continuous pressure value applied by the user when touching, the action amplitude and spatial coordinate sequence when the user waves the motion sensing device, and the velocity change rate of the spatial coordinate sequence of adjacent frames. The action amplitude is the Euclidean distance between the spatial coordinates of two consecutive frames in the spatial coordinate sequence. The intent calculation module is used to extract action features from the time series multidimensional array within a set time window. Based on motion features, speed stability score, directional stability score, and trajectory coherence score are calculated. Specifically, the speed stability score is obtained by subtracting the speed feature within the time window from 1; the directional stability score is defined as: directional stability score = 1 - (average directional deviation / preset maximum deviation angle); the average directional deviation is obtained by calculating the angular deviation between the motion direction of each frame and the preset target direction, and taking the average of the directional deviations of several consecutive frames; when the interactive scenario requires the user to perform actions according to fixed standard trajectory points, the trajectory coherence score is calculated based on the trajectory deviation between the actual trajectory points formed by the spatial coordinate sequence of each frame and the standard trajectory points, trajectory coherence score = 1 - (average trajectory deviation / preset tolerance); the trajectory deviation is the distance between the actual trajectory point and the corresponding point of the standard trajectory; when the interactive scenario does not require the user to perform actions according to fixed standard trajectory points, the trajectory coherence score is calculated by evaluating the trajectory deviation rate, that is, the proportion of the cumulative trajectory deviation to the cumulative amplitude of the motion displacement within the time window; contribution points are set for speed stability score, directional stability score, trajectory coherence score, and pressure feature. The system calculates the intention confidence score by weighting and summing the contributions. An intention recognition module categorizes user actions into different intention levels in real-time based on the intention confidence score and a set dynamic threshold, and allocates corresponding feedback computing resources according to these levels. A threshold adjustment module adaptively adjusts the dynamic threshold and contribution weights based on pre-collected user profile data. The method for acquiring user profile data includes: collecting all newly added user operation records daily, associating each record with corresponding age group information, and extracting key action features from each record. These key action features include the standard deviation of peak pressure, average action amplitude, and trajectory deviation rate, forming a feature dataset. The feature dataset is divided into children's and adults based on age group information, and clustering is performed separately. The clustering dimensions are the standard deviation of peak pressure, average action amplitude, and trajectory deviation rate. A neighborhood radius is set, and the minimum sample size is calibrated based on the similarity of the standard deviation of peak pressure, average action amplitude, and trajectory deviation rate within groups and the differences between groups in these parameters during historical clustering. After clustering, the final user profile data is obtained.

Images