Operation behavior prediction and grading early warning method and system based on project reaction theory

Abstract

The invention discloses an operation behavior prediction and grading early warning method and system based on a project reaction theory, and belongs to the technical field of man-machine interaction and intelligent monitoring. The method comprises the following steps: collecting eye movement, limb and other multi-modal behavior data of an operator and extracting indexes; based on a project reaction theory, designing an evaluation project associated with a behavior index, and constructing and training a multi-dimensional project reaction theory model to evaluate the multi-dimensional operation capability of an operator by calculating a project index and generating a binary reaction matrix; in the real-time task, current comprehensive capability parameters of an operator are calculated based on the model, and the task success rate is predicted; and according to a preset threshold interval in which the success rate is located, triggering early warning intervention of different levels. According to the method, the problems of insufficient analysis of complex operation behaviors, limited prediction precision, individuation deficiency, early warning lag and the like in the prior art are solved, and more accurate and more individualized real-time evaluation and risk early warning of the operation behaviors are realized.

Description

Technical Field

[0001] This invention relates to the field of human-computer interaction and intelligent monitoring technology, specifically to a method and system for predicting and providing graded early warnings of operator behavior based on the fusion of multimodal behavior tracking and item response theory (IRT). Background Technology

[0002] With the rapid development of human-computer interaction technology, especially in demanding operational fields such as precision assembly and surgery, real-time monitoring and evaluation of operators' visual attention distribution, cognitive load, and operational quality during complex tasks have become crucial. Accurate prediction and timely early warning are key to ensuring operational safety and improving efficiency.

[0003] Currently, the analysis and prediction of operational behavior mainly face the following technical shortcomings: Insufficient depth of behavioral analysis: Existing methods mostly rely on single-dimensional behavioral analysis (such as simple rule-based matching or single-model classification), lacking a comprehensive and multi-dimensional quantitative evaluation of cognitive states (such as attention and search efficiency) during operation, resulting in limited prediction accuracy and difficulty in dealing with complex and ever-changing operational scenarios.

[0004] Lack of personalized modeling: Traditional solutions usually use uniform evaluation standards, which cannot adapt to individual differences in operators with different skill levels and operating habits. The early warning strategies are rigid and lack personalized adaptability.

[0005] Delayed early warning timing: Most early warning systems issue alerts after the fact based on operational results or obvious erroneous actions, lacking forward-looking prediction of potential risks, resulting in delayed early warning timing and failing to effectively prevent errors from occurring.

[0006] Insufficient Technological Integration: In existing technologies, Item Response Theory is mostly applied in fields such as educational assessment and psychological questionnaires to evaluate the potential abilities of test takers. However, research and practice on its deep integration with behavioral analysis technologies such as eye tracking and motion capture for real-time, dynamic prediction and early warning of operational behaviors are still lacking.

[0007] Therefore, there is an urgent need for a technical solution that can deeply integrate behavioral analysis and psychometric theories to achieve multi-dimensional, personalized, and real-time assessment of operators' abilities, and to provide forward-looking, graded early warnings based on the assessment results. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for predicting and classifying operational behavior based on Project Response Theory (IRT). This method aims to achieve accurate assessment of the operator's inherent operational capabilities and real-time prediction of task success rate through deep integration of multimodal behavioral data and IRT models, and to execute intelligent classified early warnings accordingly.

[0009] To achieve the above objectives, this invention provides a method for predicting and classifying operational behavior based on project response theory, comprising the following steps: S1. Collect operator behavior data and extract behavior indicators that include at least visual and physical behavior characteristics; S2. Based on project response theory, construct an operator competence assessment model, specifically including: S21. Define K evaluation items, each of which is associated with one or more of the behavioral indicators; S22. Based on the behavioral indicators, calculate the project index for each of the evaluation items; S23. Generate a binary response matrix based on the comparison results between the project index and the preset threshold; S24. Based on the binary response matrix, a multidimensional item response theory model is used to estimate the model parameters and the operator's ability parameters on each of the assessment items. S3. Real-time prediction and early warning based on the operational capability assessment model, specifically including: S31. Collect and process the current operator's behavior data in real time, and generate the current binary reaction vector according to steps S22 and S23; S32. Based on the model parameters estimated in step S24 and the current binary reaction vector, calculate the comprehensive capability parameters of the current operator; S33. Based on the comprehensive capability parameters and the model parameters, predict the success rate of the current operation task; S34. Compare the success rate with multiple preset warning thresholds, and trigger the corresponding level of warning intervention based on the comparison results.

[0010] Furthermore, the present invention also provides a system for implementing the above method, comprising: The data acquisition and processing module is used to collect operator behavior data and extract behavior indicators; The model building and training module is used to define assessment items, map behavioral indicators to item indices and binary responses, and train the model based on the multidimensional item response theory model to obtain model parameters. The real-time prediction and early warning module is used to receive real-time behavioral data, calculate the comprehensive ability parameters of the current operator and predict the task success rate based on the model parameters output by the model construction and training module, and trigger corresponding early warning intervention according to the early warning threshold range to which the success rate belongs. The early warning execution module is used to execute the early warning intervention.

[0011] Compared with the prior art, the present invention has the following significant advantages: By designing multidimensional assessment items (such as search efficiency and attention maintenance) and closely linking them with multimodal behavioral indicators (eye movement and action), the IRT model is used to quantitatively assess the operator's potential abilities, overcoming the limitations of single features or models and significantly improving the accuracy of behavior prediction.

[0012] The IRT model can estimate the individualized parameters (capability parameters θ) of each operator in different capability dimensions, so that the warning threshold and intervention strategy can be dynamically adjusted based on the individual capability baseline, thus realizing a truly personalized warning.

[0013] By using real-time calculations of task success rates to generate early warnings, risk trends can be identified before operational errors actually occur, achieving a shift from "post-event alerts" to "pre-event prevention," and significantly advancing the timing of warnings.

[0014] This innovative approach introduces item response theory, a well-established theory in psychometrics, into the real-time analysis of complex operational behaviors, providing a solid mathematical model foundation for behavioral prediction and enhancing the scientific rigor and interpretability of the solution. Attached Figure Description

[0015] Figure 1 This is a flowchart of the operational behavior prediction and graded early warning method provided in an embodiment of the present invention. Detailed Implementation

[0016] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that these descriptions are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0017] Example: A method for predicting and classifying operational behavior based on project response theory. This embodiment provides a complete and implementable method for predicting and classifying operational behavior based on Item Response Theory (IRT). This method is particularly suitable for scenarios with high requirements for operational precision and safety, such as minimally invasive surgery and precision instrument assembly. (See the overall flowchart of the reference method.) Figure 1The diagram illustrates the complete process from data acquisition and model training to real-time prediction and early warning. This method mainly includes the following steps: Step S1: Behavioral Data Collection and Processing This step aims to acquire high-quality, multimodal raw data on operator behavior and extract quantitative indicators from it that can be used for modeling.

[0018] First, deploy high-precision behavior tracking equipment. In a preferred embodiment, the operator's visual behavior data is collected using the iView X™ RED telemetry eye tracker from SMI, and the operator's limb (e.g., hand, tool) movement data is collected using the NOKOV metric optical 3D motion capture system. Before data acquisition begins, all equipment must undergo rigorous coordinate system calibration to ensure that the eye tracking data, 3D spatial coordinates, and motion capture data are in a unified reference system.

[0019] Subsequently, behavioral data from expert operators is collected to establish a standard dataset. For example, 30 experienced expert operators are invited to perform standard tasks, and the data collection frequency is set at 50Hz. This frequency ensures the accuracy of key behavioral indicators extraction while maintaining a moderate data volume that does not increase the real-time computing burden. 300 valid samples are collected for each operator, thus forming a rich and reliable benchmark database.

[0020] Next, key behavioral metrics are extracted from the raw data stream. These metrics include at least: Visual behavior metrics: fixation sequence (x, y coordinates), fixation duration (milliseconds), pupil diameter change (pixels, converted to millimeters based on the pixel density of the eye tracker used, such as 1 pixel ≈ 0.01 mm for SMI iView X™ RED), saccade path length (pixels), target detection time (seconds), number of regressions, and regression rate (Hz).

[0021] Limb behavior indicators: movement paths (x, y, z coordinate sequences) of limbs or tool ends.

[0022] Finally, data cleaning is performed to remove noise and outliers. Specifically, this includes: removing segments of eye-tracking data lost due to significant head movements by the operator; and filtering out physiologically unreasonable outliers with pupil diameters less than 20 pixels (approximately 0.2 mm) or greater than 80 pixels (approximately 0.8 mm). This threshold is set based on the typical human pupil diameter range under normal lighting conditions using the SMI iView X™RED eye tracker employed in this solution, conforming to the physiological characteristics of a healthy adult pupil diameter of 2-8 mm (converted based on device pixel density), ensuring the reliability of subsequent analysis.

[0023] Step S2: Based on project response theory, construct an operator's operational ability assessment model. This step is the model training phase, the core of which is to convert the behavioral data into a format that the IRT model can process and solve for the corresponding model parameters.

[0024] S21. Define K assessment items. This embodiment defines three key operational cognitive ability assessment items, each associated with a set of behavioral indicators: Search efficiency is a metric related to target discovery time, scan path length, and replay rate, used to assess an operator's ability to quickly and accurately locate target areas. The primary measurement dimension is θ1 (search efficiency), and its difficulty is influenced by factors such as target size, target-background contrast, and search area size.

[0025] Attention maintenance exercise: This exercise is related to fixation duration, pupil diameter change, and number of regressions, and is used to assess an operator's ability to maintain focus and stability during key operational phases. The primary measure of this exercise is the ability dimension θ2 (attention maintenance ability), and its difficulty is mainly influenced by task duration, the number of distracting factors, and environmental complexity.

[0026] Anticipatory observation project: This project is linked to target discovery time and replay rate, and is used to assess the operator's ability to anticipate task progress and shift attention in advance. The primary measure of this project is the ability dimension θ3 (anticipatory observation ability), and its difficulty is mainly affected by incomplete information, time pressure, and task complexity.

[0027] S22. Calculate the project index for each evaluation item. For each item, combine its associated multiple behavioral indicators into a composite index. Preferably, a weighted summation method is used for calculation: For the Attention Maintenance Index (AMI): AMI = w1*F1 + w2*F2 + w3*F3 + w4*F4. Where F1 is fixation duration stability (calculated as 1 - |fixation duration / average fixation duration - 1|), F2 is pupil change stability (calculated as 1 / (1 + pupil diameter change rate)), F3 is the return rate reasonableness (obtained by comparing with the percentile of the expert baseline distribution; for example, if the current return rate is within the top 30% percentile of the expert distribution, it is assigned a full score of 1; otherwise, it is assigned a linear score), and F4 is attention fluctuation (calculated as 1 - (maximum fixation duration - minimum fixation duration) / average fixation duration). The weight vector [w1, w2, w3, w4] can be set to [0.35, 0.25, 0.25, 0.15]. The AMI index reflects the operator's ability to maintain attention, and the discrimination parameter a2 mainly reflects the AMI index's ability to distinguish the attention maintenance abilities of different operators.

[0028] The search efficiency index (SI) is defined as follows: SI = α × (1 − target discovery time) + β × (1 − scan path length) + γ × (1 − replay rate). Here, target discovery time and scan path length are normalized to the interval [0,1], and the weight vector [α,β, γ] is set to [0.5, 0.3, 0.2]. This SI index reflects the operator's visual search efficiency, and the discrimination parameter a1 mainly reflects the SI index's ability to distinguish the visual search efficiency of different operators.

[0029] For the Expected Observation Index (PT): PT = max(0, 1 − Target Discovery Time / Average Expert Discovery Time) × (1 − Response Rate). Here, the average expert discovery time is the statistical mean of the target discovery times of 30 expert operators. A non-negative truncation is applied during the calculation phase using the max(0,・) function to avoid negative values ​​when the target discovery time > the average expert discovery time, ensuring the PT index is always positive (higher values ​​indicate stronger ability). If the average expert discovery time is lower than a preset minimum effective threshold (e.g., 0.05 seconds), then PT = (1 − Response Rate) to avoid division by zero errors or numerical instability. This PT index reflects the operator's expected observation ability, and the discrimination parameter a3 mainly reflects the PT index's ability to distinguish the expected observation abilities of different operators.

[0030] S23. Generate a binary response matrix. A threshold is set for each item, determined by the expert performance distribution. For example, the threshold T1 for the attention maintenance item can be the 30th percentile of the expert group's AMI index; the threshold T2 for the search efficiency item can be the 40th percentile of the expert group's SI index; and the threshold T3 for the anticipatory observation item can be the 35th percentile of the expert group's PT index. For each operator's sample, if their item index is higher than the item's threshold, the response to that item is recorded as "1" (indicating competence meets the standard); otherwise, it is recorded as "0". Finally, the responses of all samples from all operators are organized into an N (operator) × K (item) binary response matrix. In this embodiment, a 30 × 3 matrix is ​​obtained.

[0031] S24. Estimate IRT model parameters and capability parameters. A multidimensional three-parameter logistic (3PL) model is used as the IRT model in this embodiment. The model expresses the probability P(θ) of an operator's success in a task as: P(θ) = c + (1-c) / [1 + exp(-1.7*(a1θ1 + a2θ2 + a3θ3 - b))] where θ = [θ1, θ2, θ3] is a three-dimensional ability vector, corresponding to the potential abilities of search efficiency, attention maintenance, and anticipatory observation, respectively; each evaluation item corresponds to a set of discrimination parameters (a1, a2, a3), representing the sensitivity of the item to the three different ability dimensions of search efficiency, attention maintenance, and anticipatory observation, respectively; b is a difficulty parameter, which characterizes the overall difficulty level of the task and is affected by the indicators of the three items of search efficiency, attention maintenance, and anticipatory observation; c is a guessing parameter, and the range of the guessing parameter c is set based on the random guessing success probability interval of historical operation data statistics. For example, in the preliminary experiment, the average value of this probability statistics is 7%, so the default value can be set to 0.07. The Expectation-Maximization (EM) algorithm is used to iteratively estimate the model parameters (a1, a2, a3, b, c). During initialization, the prior distribution of the capability vector θ is assumed to be a multivariate normal distribution. The maximum number of iterations is set to 100, and the convergence accuracy is 0.001. The algorithm iterates based on the binary response matrix generated in step S23. During model training, if convergence is not achieved after 100 iterations, the iterations are automatically extended to 200. If convergence is still not achieved, an anomaly is recorded and a manual verification mechanism is triggered. During real-time prediction: if convergence is not achieved after 100 iterations, the previous valid parameters are used directly, and an anomaly is recorded in the background for verification during idle periods.

[0032] S25. Establish a baseline level for operational capability. After the model training is completed, input the behavioral data of expert operators into the trained model to calculate their average capability parameters [θ1_expert, θ2_expert, θ3_expert] in three dimensions. This vector serves as the baseline level for operational capability and is used for reference and comparison in subsequent real-time evaluation.

[0033] Step S3: Real-time prediction and early warning based on the operational capability assessment model. This step is the application phase of the model, which involves real-time monitoring and early warning of new tasks for new operators or new tasks for the same operator.

[0034] S31. Real-time data processing and feature generation. While the operator is performing the task, their eye movement and motion data are collected in real time at a frequency of not less than 50Hz. Following the methods in steps S1 and S22, the current operator's indices on the three evaluation items are calculated in real time, and then a 1 × 3 current binary response vector is generated based on preset thresholds.

[0035] S32. Calculate the current integrated capability parameters. Input the current binary response vector obtained in step S31, along with the model parameters (a1, a2, a3, b, c) trained in step S24, into the IRT model. Using methods such as Bayesian estimation, deduce the current operator's real-time, personalized three-dimensional integrated capability parameters [θ1_current, θ2_current, θ3_current]. This process can be interpreted or calibrated with reference to the expert baseline established in step S25.

[0036] S33. Predict the current task success rate and determine the trend. Substitute the current operator's comprehensive ability parameters [θ1_current, θ2_current, θ3_current] into the IRT model formula in step S24 to calculate the predicted current task success rate P(θ_current). Simultaneously, maintain a sliding time window of length 3, with each window corresponding to 3 seconds of data collection (150 samples), and a sliding step of 0.5 seconds, meaning adjacent windows overlap by 0.5 seconds to ensure the continuity and sensitivity of the trend determination. Store the success rate data of the three most recent windows [P1, P2, P3], and calculate the trend change value ΔP, i.e., ΔP = P1 - P3, representing the difference in success rate between the previous window and the current window. The length and step of the sliding time window can be configured according to the specific task's operation cycle and real-time requirements.

[0037] S34. Tiered Early Warning Intervention. The early warning threshold is jointly set based on historical operation data statistical analysis and domain expert experience, and can be dynamically adjusted as the system continuously learns operator behavior patterns. A dual mechanism of "fixed period + sliding window" is used to balance adaptability and stability: the system triggers an update every 30 days (configurable as needed) or after accumulating 100 valid new operation samples (whichever comes first). These two thresholds can be flexibly configured according to the actual task execution frequency, data update needs, and early warning stability requirements. During the update, only the quantiles (e.g., 10%, 30%, 50%) of the early warning threshold are recalculated based on the most recent 3000 valid samples (the upper limit of the sliding window to avoid interference from old data). It should be noted that this update window is much larger than the 3-second sliding window for trend judgment. The core logic is: the trend window focuses on real-time sensitive early warning, capturing short-term fluctuations in the success rate; the update window focuses on data distribution stability, avoiding frequent threshold fluctuations affecting early warning consistency, comparing the predicted success rate P(θ) with multiple preset early warning thresholds, and triggering the corresponding level of early warning. Primary warning (prompt): When 0.6 < P(θ) ≤ 0.7, or 0.7 < P(θ) ≤ 0.8 and ΔP > 0.1 (the success rate decreases by more than 10% for 3 consecutive windows), it is judged as low risk or potential low risk. The system gives a gentle visual reminder by popping up a semi-transparent prompt box (such as "Please concentrate") at the edge of the display screen on the operation interface.

[0038] Intermediate warning (warning): When 0.4 < P(θ) ≤ 0.6, or 0.6 < P(θ) ≤ 0.7 and ΔP > 0.1 (the success rate decreases by more than 10% for 3 consecutive windows), it is judged as medium risk or escalating risk. The system activates the voice prompt system to clearly point out the specific area that needs attention with voice instructions (such as "Please note the stability of the instrument in the lower left corner") to guide the operator to make adjustments.

[0039] Advanced warning (intervention): When P(θ) ≤ 0.4, or 0.4 < P(θ) ≤ 0.6 and ΔP > 0.15 (the success rate decreases by more than 15% for 3 consecutive windows), it is judged as high risk or rapid risk escalation. The system first initiates transitional intervention: immediately control the operating device to enter the deceleration operation mode [such as the robot movement speed drops to 30% of the normal speed, this ratio is set based on the safety standards in the field of precision operation and multi-scenario experimental data, and different deceleration ratios (range 10% - 50%) can be preset according to the task type (such as suturing, cutting, precision assembly)]. At the same time, through the dual emergency prompts of voice + full-screen highlighted graphics and text, clearly inform the operator of the risk level and the interruption countdown (default 5 seconds, adjustable from 3 - 10 seconds); during the deceleration operation stage, if the operator does not confirm "Continue operation" through the preset physical button (anti-misoperation design, long press for 2 seconds to take effect) within the countdown, the system locks the key operation permissions through the software interface (such as pausing the robot movement), forcibly interrupts the operation that may cause errors, and at the same time displays detailed graphics, text or animation guidance information on the main display screen to guide the operator to return to a safe state; to avoid secondary risks in emergency situations, the system adds a safety fallback mechanism, allowing the operator to manually override the locked state through the preset physical button or authorization password (needs to be bound to the operator's identity, the update period can be set to a fixed duration (such as every 90 days) or based on the operation frequency (such as after every 500 tasks)) during the countdown or the equipment deceleration stage to ensure the flexibility and safety of manual intervention in emergency scenarios.

[0040] Corresponding system embodiment A system for implementing the above method, characterized by including: Data acquisition and processing module: Integrates the hardware units of the specific model eye tracker and motion capture system as described above, and the software units for data cleaning and index extraction.

[0041] Model building and training module: This module contains a configuration library that stores project definitions, exponent calculation formulas, thresholds and weights, as well as a computing engine that runs the EM algorithm to estimate IRT model parameters.

[0042] Real-time prediction and early warning module: includes real-time data pipeline, capability parameter estimation algorithm, success rate prediction model and early warning decision logic.

[0043] Warning Execution Module: This module integrates a graphical interface generator, a speech synthesis engine, and an interface that can communicate with an external operating system or device controller to execute warning actions at different levels.

[0044] Other embodiments This invention is not limited to the specific embodiments described above. For example, the number of evaluation items, the selection of specific behavioral indicators, the synthesis formula of the item index, the specific form of the IRT model (such as the 2PL model), the specific value of the warning threshold, and the form of the warning (such as haptic feedback) can all be adjusted and replaced according to different application scenarios, and these variations and replacements all fall within the protection scope of this invention.

[0045] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A method for predicting and classifying operational behavior based on project response theory, characterized in that, Includes the following steps: S1. Collect operator behavior data and extract behavior indicators that include at least visual and physical behavior characteristics; S2. Based on project response theory, construct an operator competence assessment model, specifically including: S21. Define K evaluation items, each of which is associated with one or more of the behavioral indicators; S22. Based on the behavioral indicators, calculate the project index for each of the evaluation items; S23. Generate a binary response matrix based on the comparison result between the project index and the preset threshold; S24. Based on the binary response matrix, a multidimensional item response theory model is used to estimate the model parameters and the operator's ability parameters on each of the assessment items. S3. Real-time prediction and early warning based on the operational capability assessment model, specifically including: S31. Collect and process the current operator's behavior data in real time, and generate the current binary reaction vector according to steps S22 and S23; S32. Based on the model parameters estimated in step S24 and the current binary reaction vector, calculate the comprehensive capability parameters of the current operator; S33. Based on the comprehensive capability parameters and the model parameters, predict the success rate of the current operation task; S34. Compare the success rate with multiple preset warning thresholds, and trigger the corresponding level of warning intervention based on the comparison results.

2. The method according to claim 1, characterized in that, In step S1, visual behavior data is collected through an eye-tracking device, and limb behavior data is collected through a motion capture device; the frequency of the behavior data collection is not less than 50Hz; the behavior indicators include at least the fixation point sequence, fixation duration, pupil diameter change, saccade path length, target discovery time, number of retrospectives, retrospective rate, and limb movement path.

3. The method according to claim 1, characterized in that, In step S21, the K evaluation items include search efficiency, attention maintenance, and predictability observation items; The search efficiency item is related to the target discovery time, scan path length, and return rate. The attention maintenance items are related to fixation duration, pupil diameter change, and number of retrospectives. The anticipated observation items are associated with the target discovery time and the replay rate.

4. The method according to claim 1, characterized in that, In step S22, the calculation of the item index of the attention maintenance item includes: a weighted sum based on four sub-indicators: fixation duration stability, pupil change stability, return rate reasonableness, and attention fluctuation. Fixation duration stability is calculated based on the deviation of fixation duration from its average value. Pupil change stability is calculated based on the pupil diameter change rate. Return rate reasonableness is obtained by comparing the current return rate with the expert benchmark distribution. Attention fluctuation is calculated based on the ratio of the difference between the maximum and minimum fixation duration to the average fixation duration.

5. The method according to claim 1, characterized in that, In step S24, the multidimensional item response theory model is a multidimensional three-parameter logistic model, which expresses the probability of task success as a function of the ability vector; the model parameters include a discrimination parameter corresponding to each evaluation item, a difficulty parameter, and a guessing parameter; the model parameters are solved iteratively using the expectation-maximization algorithm; the ability vector is a three-dimensional vector, which respectively characterizes the operator's potential ability level in search efficiency, attention maintenance, and predictive observation.

6. The method according to claim 1, characterized in that, Step S2 further includes step S25: based on behavioral data samples from multiple expert operators, calculate expert capability parameters using the operational capability assessment model as a baseline level for operational capability.

7. The method according to claim 1, characterized in that, In step S34, the warning levels include at least a primary warning, an intermediate warning, and a high-level warning; When the success rate is within a first preset range, the primary warning is triggered, and the primary warning includes a visual cue. When the success rate is below the first preset range in the second preset range, the intermediate warning is triggered, and the intermediate warning includes a voice prompt. When the success rate is lower than the second preset range, the advanced warning is triggered, which includes operation blocking and detailed guidance.

8. The method according to claim 1, characterized in that, Step S1 further includes cleaning the collected raw behavioral data, which includes removing data lost due to head movement and removing data whose pupil diameter values ​​exceed a preset reasonable range.

9. A predictive and graded early warning system for operational behavior based on item response theory, characterized in that, The system for implementing the method of any one of claims 1 to 8 comprises: The data acquisition and processing module is used to collect operator behavior data and extract behavior indicators; The model building and training module is used to define assessment items, map behavioral indicators to item indices and binary responses, and train the model based on the multidimensional item response theory model to obtain model parameters. The real-time prediction and early warning module is used to receive real-time behavioral data, calculate the comprehensive ability parameters of the current operator and predict the task success rate based on the model parameters output by the model construction and training module, and trigger corresponding early warning intervention according to the early warning threshold range to which the success rate belongs. The early warning execution module is used to execute the early warning intervention.

10. The system according to claim 9, characterized in that, The data acquisition and processing module includes an eye-tracking unit and a 3D motion capture unit.

Images