An eye movement cognitive state inference method and system

Abstract

The application discloses an eye movement cognitive state inference method and system, relates to the field of advanced driving assistance systems and edge computing cross technology, and comprises the following steps: inputting a multi-modal joint feature generated by time stamp alignment of a real-time acquired eyeball micro movement sequence signal of a driver and synchronous cabin inertial measurement data into a trained lightweight neural network model for forward propagation calculation; based on stateless forward propagation output of the lightweight neural network model, low-delay parallel generation of an absolute state health quantitative index and a transient mental load quantitative index of the driver; and based on critical slowing down precursor features learned implicitly in the model, output of an advanced perceptual decoupling pre-warning signal; the method preposes complex nonlinear dynamics calculation to the cloud through lightweight forward propagation and physical knowledge distillation, and the vehicle-mounted end only needs to perform millisecond-level extremely fast inference, thereby meeting the hard real-time requirement of the ADAS system on a safety takeover request.

Description

Technical Field

[0001] This invention relates to the field of advanced driver assistance systems (ADAS) and edge computing, specifically to an eye-tracking cognitive state reasoning method and system. Background Technology

[0002] Advanced driver assistance systems are crucial for real-time monitoring of driver fatigue and cognitive decoupling.

[0003] To achieve absolute health status Ψ of the driver Tonic High-fidelity quantization decoupled from perception typically requires extracting complex nonlinear dynamic indicators in existing technologies. Common fatigue monitoring schemes often rely on cumulative integral models, such as Markov chain-based state evolution models, to track the accumulation of fatigue. These Markov chain-based state evolution models are essentially probabilistic statistical models that rely on "historical memory," assuming that the driver's current fatigue state evolves from the state at the previous moment. The system continuously collects data, viewing fatigue as a process of "accumulation" and "transition" over time (e.g., gradually transitioning from "awake" to "mild fatigue," and then to "severe fatigue"). The typical workflow can be divided into the following steps: (1) State division and initialization: First, several discrete cognitive / fatigue levels are preset (such as state A: awake; state B: fatigued; state C: micro-sleep). In the early stage of driving, the system defaults to the driver being in state A; (2) Continuous observation and feature extraction: The vehicle camera or sensor continuously collects the driver's eye movement, facial features or vehicle behavior data over time; (3) Probability transfer and integral accumulation (core mechanism): The system uses the Markov transition matrix, combined with the newly collected features and the fatigue state probability of the previous moment, to calculate the current state. This is equivalent to an "integrator". As long as the system continuously detects fatigue signals, the fatigue value will be accumulated layer by layer on the basis of historical data; (4) Threshold trigger warning: When the accumulated probability integral exceeds the set danger threshold, the system determines that the driver is fatigued and issues an alarm.

[0004] However, such models that rely on historical states have significant drawbacks in practical ADAS applications. In real driving environments, sensors may be temporarily obstructed due to driver head turns, glare from glasses, or other reasons, and data may be lost. Any such interruption disrupts the continuity of the historical state sequence, leading to catastrophic cumulative errors in the integral model. Secondly, high-precision topological determinism... D top The computation is complex, and its extremely high computational overhead makes it difficult to meet the millisecond-level safety response hard real-time requirements of ADAS systems on automotive edge computing chips with limited computing power. Summary of the Invention

[0005] To address the shortcomings of existing technologies, such as catastrophic cumulative errors caused by sensor occlusion or data loss, and the difficulty in meeting the millisecond-level hard real-time response requirements of ADAS systems on computationally limited automotive edge chips, this invention proposes an eye-tracking cognitive state reasoning method and system. By constructing an asymmetric architecture of "offline physical knowledge distillation + online stateless lightweight reasoning," the complex nonlinear dynamic index calculations are performed in the cloud, and the vehicle only needs to perform forward propagation of a lightweight neural network, thereby solving the problems existing in the prior art.

[0006] An eye-tracking cognitive state reasoning method includes the following steps: Collect historical eye-tracking time-series signals of the driver during the driving process and synchronized inertial measurement data of the carriage; Phase space reconstruction is performed on historical eye movement time series signals to extract absolute state health to characterize current cognitive reserve capacity and topological deterministic indicators to characterize the degree of degradation of the eye movement phase space manifold structure. Historical eye-tracking time-series signals and inertial measurement data from the train carriage are timestamped to generate multimodal joint features. These multimodal joint features are then input into a lightweight neural network model for training. Absolute state health is used as the primary task label, and topological determinism is used as the auxiliary supervision label to obtain the primary task prediction output and auxiliary output of the network model. Based on the primary task prediction output and auxiliary output, a joint loss function is constructed, which includes the primary task prediction loss and the physical auxiliary loss. The parameters of the lightweight neural network model are updated by calculating the joint loss function and through backpropagation until the convergence condition is met, thus completing the training of the lightweight neural network model. The multimodal joint features generated by aligning the real-time driver eye movement time series signals with the synchronous inertial measurement data of the carriage with timestamps are input into a lightweight neural network model trained for forward propagation, and the driver's cognitive state inference results are output. The inference results include absolute state health, transient mental load and perceptual decoupling early warning signals.

[0007] Furthermore, the step of aligning the historical eye movement time series signal and the carriage inertial measurement data with timestamps to generate multimodal joint features specifically includes: after aligning the eye movement sequence signal with the carriage inertial measurement data with timestamps, using early feature splicing or cross-modal cross-attention mechanism to perform feedforward fusion of multi-source data during the forward propagation of the lightweight neural network model to obtain multimodal joint features.

[0008] Furthermore, after reconstructing the phase space of the historical eye-tracking time-series signal, the method further includes extracting perceptual decoupling early warning tags, specifically including the following steps: Determine the physical anchor time of perceptual decoupling in historical eye-tracking time series signals; Within a preset time window prior to the physical anchor point time, calculate the time-shifted variance and the first-order autocorrelation coefficient of the state transition energy index. The critical slowdown precursor interval is defined based on the abrupt increase characteristics of time-shifted variance and lag first-order autocorrelation coefficient. This critical slowdown precursor interval is then incorporated as a supervisory soft label into the training process of the lightweight neural network model, enabling the lightweight neural network model to implicitly internalize the nonlinear fluctuation characteristics of the time series before the collapse.

[0009] Furthermore, the absolute state health is calculated using an electrodynamic isomorphic decoupling model based on stateless assessment, specifically including the following steps: Topological deterministic indices characterizing the degree of degradation of the manifold structure in eye movement phases are extracted and global calibration values ​​are used in the driver's awake baseline state. Construct a capacity ceiling Φ(t) to represent the absolute cognitive reserve; Extract the state transition energy index, which represents the actual kinetic energy consumed by the driver's sympathetic nervous system activation, and the sample entropy, which represents the complexity of the representational signal, within the current sliding time window. SampEn ; Introducing a preset physical penalty coefficient, based on the state transition energy index STEI With topological determinism indicators D top The ratio of is used to construct a damping operator that characterizes the dynamic workload discounting; Multiply the capacity ceiling by the sample entropy and divide by the damping operator Γt to obtain the absolute state health within the current time window, which is free from compensatory artifacts and independent of historical states.

[0010] Furthermore, the absolute state health Ψ Tonic(t) Represented as: ; Where γ is the physical penalty coefficient; Capacity ceiling; τ is the damping operator; τ is the absolute critical threshold for triggering decoupling warning.

[0011] Furthermore, a nonlinear dynamics physics engine is used to reconstruct the phase space of historical eye-tracking time series. During the phase space reconstruction process, topological dynamic features characterizing the capacity and compensatory kinetic energy evolution of the underlying structure of the eye-tracking manifold are extracted in the multidimensional phase space as another hidden layer supervision signal to decouple transient mental load and absolute fatigue state from external inducements. This hidden layer supervision signal is added to the joint loss function for supervised training, so that the lightweight neural network model can adaptively identify transient high cognitive input triggered by any factor and has the ability to block system false alarms caused by transient compensatory illusions from the bottom layer.

[0012] Furthermore, the joint loss function is expressed as: ; in, Predicting losses for the main task. D top For the physical truth value of topological determinism, The structural capacity prediction value is the output of the network-assisted output. STEI The physical true value of the state transition energy index is... λ1 and λ2 are the predicted compensatory kinetic energy output by the network, and are the preset loss balance weight coefficients.

[0013] Furthermore, the topological dynamic characteristics characterizing the evolution of the capacity and compensatory kinetic energy of the eye-moving manifold's underlying structure include at least one of topological determinism, state transition energy index, or transient spatial extensibility.

[0014] Furthermore, the perceptual decoupling pre-warning signal is specifically determined by extracting baseline features representing the pure physiological background of the system, topological geometric features representing the local manifold distortion law of the eye movement phase space, and dynamic features representing the evolution law of the system's compensatory kinetic energy. Among them, the dynamic features include at least one of the following: the recent mean, time-shift variance, or first-order evolutionary gradient of the transient mental load quantification index within a preset time window; the baseline features include the health descent gradient calculated from the first difference of the absolute state health quantification index; the topological geometric features include at least one of the following: the recent mean of the topological deterministic index, the time-shift variance of the spatial extension, the second-order evolutionary acceleration of the spatial extension, or the nonlinear product cross term of the transient mental load quantification index and the topological deterministic index.

[0015] The present invention also includes an eye-tracking cognitive state reasoning system, comprising: The acquisition module is used to collect the driver's historical eye movement time series signals and synchronous inertial measurement data of the carriage during the driving process; The extraction module is used to reconstruct the phase space of historical eye movement time series signals to extract absolute state health to characterize current cognitive reserve capacity and topological deterministic indicators to characterize the degree of degradation of the eye movement phase space manifold structure. The model training module is used to align historical eye-tracking time-series signals and inertial measurement data of the carriage with timestamps to generate multimodal joint features. These multimodal joint features are then input into a lightweight neural network model for training. Absolute state health is used as the primary task label, and topological determinism is used as the auxiliary supervision label to obtain the primary task prediction output and auxiliary output of the network model. Based on the primary task prediction output and auxiliary output, a joint loss function is constructed, comprising primary task prediction loss and physical auxiliary loss. The parameters of the lightweight neural network model are updated by calculating the joint loss function and through backpropagation until the convergence condition is met, thus completing the training of the lightweight neural network model. The output module is used to input the multimodal joint features generated by aligning the real-time acquired driver eye movement time series signal with the synchronous inertial measurement data of the carriage with the timestamps into the trained lightweight neural network model for forward propagation, and output the driver's cognitive state inference results; the inference results include absolute state health, transient mental load and perceptual decoupling early warning signal.

[0016] This invention provides a method for reasoning about eye-tracking cognitive states, which has the following beneficial effects: This invention calculates absolute state health using a stateless evaluation method that relies solely on the current time window's cross-sectional features. It then uses topological determinism indicators as auxiliary supervisory labels for physical knowledge distillation. This enables a lightweight neural network to penetrate appearance compensation artifacts during real-time inference and directly output absolute state health, transient mental load, and early warning signals reflecting true cognitive reserves. The method constructs an asymmetric architecture of "offline physical knowledge distillation + online stateless lightweight inference," pre-completing the calculation of complex nonlinear dynamic indicators in the cloud. The vehicle-mounted terminal only needs to execute the forward propagation of the lightweight neural network. This fundamentally solves the technical challenges of existing fatigue monitoring solutions, which rely on historical state accumulation models such as Markov models, resulting in catastrophic cumulative errors when sensors are obstructed or data is lost. Furthermore, it addresses the difficulty of meeting the millisecond-level hard real-time response requirements of ADAS systems on computationally limited vehicle edge chips. This invention achieves complete immunity to state forgetting and integral drift problems and compresses vehicle-mounted inference latency to the millisecond level, providing reliable technical support for the active safety takeover of ADAS systems. Attached Figure Description

[0017] Figure 1 This is a logical architecture diagram of the absolute state health engine in an embodiment of the present invention; Figure 2This is an anatomical diagram of the physical mechanism of pseudo-wake-up in an embodiment of the present invention; Figure 3 This is a heatmap showing parameter optimization and penalty coefficients in an embodiment of the present invention; Figure 4 This is a diagram showing the advanced prediction ROC curve and optimized confusion matrix in an embodiment of the present invention; Figure 5 This is a time profile of the critical slowdown (CSD) predicted by advance decoupling in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] This invention proposes a low-latency reasoning method for eye-tracking cognitive states based on physical knowledge distillation, applied to in-vehicle edge computing devices; this method is used for real-time calculation of decoupling warnings and absolute state health (Ψ). Tonic The stateless evaluation method eliminates the vulnerability of traditional relative entropy metrics to "compensatory false wake-ups" and constructs a physics engine based on the isomorphism between dynamic terminal voltage and battery discharge rate (C-rate). The offline distillation architecture introduces topological determinism. D top As a capacity constraint of the underlying structure, it forces the lightweight vehicle model to internalize the prior of topological distortion, thereby achieving orthogonal decoupling between instantaneous mental load and absolute state health at the physical level.

[0020] 1. Decoupling computation mechanism for stateless absolute health based on battery electrodynamic isomorphism: When performing offline physical knowledge distillation training, it is necessary to pre-extract the absolute state health (Ψ). Tonic This serves as the hard physics truth label. To completely resolve the cumulative integration error caused by occlusion from onboard sensors, this invention constructs a stateless evaluation architecture that relies solely on the current time window. The offline physics engine strictly executes the following computation mechanism in the multi-dimensional phase space:

[0021] (1) Extracting the structural ceiling representing cognitive capacity: Extracting the topological determinism representing the eye-movement manifold control structure ( D top Its physical meaning is equivalent to the "cognitive working memory pool" available in the nervous system. It utilizes global calibration values ​​from the subject's baseline state of wakefulness (…). ), to construct an absolute capacity ceiling (Φt):

[0022] ; This mechanism ensures that when the system faces deep sleep deprivation leading to topological manifold collapse ( D top When the value reaches →0, even if there is a violent compensatory struggle on the surface, the absolute health limit of the system is forcibly reset to zero.

[0023] (2) Extracting Real Mental Load and Damping Operator: The State Transition Energy Index (STEI) within the current sliding time window is extracted to characterize the "actual kinetic energy consumed" by sympathetic nerve activation. In neuro-ergonomics, real mental load is not a static scalar, but rather the ratio of actual kinetic energy consumed to available cognitive capacity. Therefore, a physical penalty coefficient γ is introduced to construct a damping operator (Γt) representing the dynamic workload discounting:

[0024] ; This mechanism is perfectly isomorphic to the discharge rate (C-rate) of a physical battery: STEI is equivalent to the discharge current. D top This is equivalent to the rated capacity. When the structural capacity is exhausted, any tiny compensatory kinetic energy will trigger a mathematical singularity (extremely high discharge rate), thereby triggering a severe voltage sag similar to the Peukert effect. After cross-validation using a two-dimensional grid search optimization space, the optimal physical penalty coefficient is γ≈2.30, at which point the system can achieve a compensatory artifact discount rate of up to 90%.

[0025] (3) Stateless final equation solidification: Extract the apparent signal complexity (sample entropy SampEn) and combine it with the capacity ceiling and damping operator to calculate the final absolute state health: ; This equation completely resolves the "Oculomotor Freezing" paradox; under this stateless mechanism, no matter how much micro-saccade noise increases the sample entropy, the collapse... D top This will directly reduce the capacity ceiling to zero and dampen the denominator to infinity, instantly locking the health status at the bankruptcy threshold.

[0026] 2. Four-dimensional fatigue degradation phenotype mapping based on electrodynamic clustering: To address the industry challenge of vastly different fatigue coping strategies among drivers, the system internalizes four electrodynamic degradation phenotypes through unsupervised K-Means clustering, enabling it to cope with various complex failure modes. (1) Forced maintenance and intermittent microsleep (Cluster 0): The driver attempts to maintain topology capacity (high Dtop), but compensates by generating pulsed high-frequency kinetic energy (STEI spikes). The formula instantly identifies these high C-rate discharges as "transient voltage sags", accurately capturing microsleep events.

[0027] (2) Violent Restart Type (Cluster 1): When the driver's structural capacity is on the verge of collapse, an extremely large compensatory current is drawn to attempt to restart the system. Due to the extremely low underlying capacity, the formula locks the assessed power in the dangerous 20% range through an extreme penalty term.

[0028] (3) Completely decoupled type (Cluster 2): The driver abandons compensation (extremely low STEI), but the structural capacity completely collapses (D top →0). Even if random eye saccharides increase the apparent entropy value, the Φ(t) multiplier in the formula will still force its output to zero.

[0029] (4) Progressive collapse type (Cluster 3): It presents linear capacity leakage and inefficient compensation, and the model can output a smooth health degradation curve.

[0030] Once the hard physical truth labels constructed using the above mechanism complete the knowledge distillation of the lightweight model in the cloud, the network deployed on the vehicle's NPU can implicitly trigger the internally learned phenotypic responses, achieving millisecond-level ultra-fast edge inference without historical dependencies.

[0031] The eye-tracking cognitive state low-latency inference method proposed in this invention, such as Figure 1 As shown, the specific steps include: S1. Real-time acquisition of the driver's eye movement sequence signals and synchronous inertial measurement data of the carriage.

[0032] S2. The eye micro-movement sequence signal is time-stamped and the inertial measurement data of the carriage is aligned to generate multimodal joint features.

[0033] Specifically, the step of using eye micro-motion sequence signals and carriage inertial measurement data as feature inputs includes: aligning the eye micro-motion sequence signals and carriage inertial measurement data with timestamps; employing early feature splicing or cross-modal cross-attention mechanisms to perform feedforward fusion of multi-source data during the forward propagation of the lightweight neural network model; enabling the lightweight neural network model to adaptively learn the transmission relationship from the vehicle's three-axis rigid body excitation to the passive displacement of the eyeball, so as to adaptively cancel common-mode mechanical noise at the feature extraction layer.

[0034] S3. Training of eye-tracking cognitive state model based on physical knowledge distillation.

[0035] S3.1. Obtain historical eye-tracking time series datasets and synchronized carriage inertial measurement data.

[0036] S3.2. Use a nonlinear dynamics physics engine to perform phase space reconstruction on historical eye movement time series, and extract physical indicators that reflect the evolution of eye movement manifold dynamics. The physical indicators include at least the absolute state health, which characterizes the current cognitive reserve capacity, and the topological deterministic indicators that reflect topological distortion (characterizing the degree of degradation of the eye movement phase space manifold structure). The absolute state health is used as the hard physics truth value label, and the parameters that reflect topological distortion, such as the topological deterministic indicators, are used as hidden layer supervision signals.

[0037] Furthermore, in the process of extracting physical indicators using a nonlinear dynamics physics engine, the steps also include extracting pre-warning labels for perceptual decoupling: determining the physical anchor time in the historical eye-tracking time series where perceptual decoupling occurs; calculating the time-shift variance and the first-order autocorrelation coefficient of the state transition energy index within a preset time window before the physical anchor time; defining the critical slowdown precursor interval based on the jump characteristics of the time-shift variance and the first-order autocorrelation coefficient, and adding the critical slowdown precursor interval as a supervised soft label to the physical knowledge distillation training, so that the lightweight neural network model implicitly internalizes the nonlinear fluctuation characteristics of the time series before collapse.

[0038] Furthermore, in the process of reconstructing the phase space of historical eye-tracking time series using a nonlinear dynamics physics engine, the process also includes extracting topological dynamic features in the multidimensional phase space that characterize the capacity and compensatory kinetic energy evolution of the underlying structure of the eye-tracking manifold as another hidden layer supervision signal that decouples transient mental load and absolute fatigue state from external inducements. This hidden layer supervision signal is added to the joint loss function for supervised training, so that the lightweight neural network model after physical knowledge distillation can adaptively identify instantaneous high cognitive input triggered by any factor and has the ability to block system false alarms caused by instantaneous compensatory illusions from the bottom layer.

[0039] This also includes: extracting topological dynamic features characterizing the evolution of the capacity and compensatory kinetic energy of the underlying structure of the eye-movement manifold in a multidimensional phase space, serving as hidden layer supervision signals to decouple transient mental load from absolute fatigue state, independent of external inducements; incorporating these hidden layer supervision signals into a joint loss function for supervised training, enabling a lightweight neural network model distilled from physical knowledge to adaptively identify transient high cognitive input triggered by any factor, and possessing the ability to block system false alarms caused by transient compensatory illusions from the ground up. Among these, the topological dynamic features characterizing the evolution of the capacity and compensatory kinetic energy of the underlying structure of the eye-movement manifold include at least topological determinism characterizing the available cognitive capacity (…). D top), and the State Transition Energy Index (STEI) or transient spatial extension, which characterizes the intensity of dynamic compensatory struggle. S spread ).

[0040] Furthermore, the step of extracting physical indicators and using them as hard physics truth labels includes employing an electrodynamic isomorphic decoupling model based on stateless evaluation to calculate absolute state health. The specific calculation mechanism includes the following steps: extracting topological deterministic indicators characterizing the eye-tracking manifold control structure. D top And using global calibration values ​​of subjects in a conscious baseline state. Construct a capacity ceiling Φt to represent the absolute cognitive reserve; extract the state transition energy index within the current sliding time window. STEI To characterize the instantaneous kinetic energy consumed during compensatory struggle; to extract the sample entropy within the current time window. SampEn To characterize the cognitive complexity of representations; a pre-defined physical penalty coefficient γ is introduced to calculate the state transition energy index STEI and the topological determinism index D. top The ratio is used to characterize the actual mental workload. A damping operator Γt is constructed with a preset constant as the base and the product of the actual mental workload and the physical penalty coefficient as the exponent. The capacity ceiling Φt is multiplied by the sample entropy SampEn and divided by the damping operator Γt to obtain the absolute state health degree that is stripped of compensatory artifacts and independent of historical states within the current time window.

[0041] Ψ is calculated using the following stateless exponential discount equation. Tonic(t) : ; Where γ is the physical penalty coefficient; Capacity ceiling; τ is the damping operator; τ is the absolute critical threshold for triggering decoupling warning, the preferred calibration value of the physical penalty coefficient γ is 2.30, and the absolute critical threshold τ for triggering decoupling warning is set to 50.0%.

[0042] like Figure 2 The diagram illustrates the physical mechanism of pseudo-awakening. This comparison highlights the discrepancy between apparent indicators and the underlying physical reality. Figure 2 (a) represents surface arousal: subjects exhibit dramatic fluctuations in physiological signals; (b) represents topological determinism: acting as a "metabolic auditor," it reveals that the system completely lacks structural stability at this point. D top (c) represents the true health level of the energy: in the absence of structural support, high-energy kinetic energy struggles ( This can lead to the rapid depletion of system reserves, ultimately triggering the system's brittle cliff.

[0043] like Figure 3 The figure shows the absolute state health ( Parameter optimization heatmap. This heatmap shows the system utility score as... and critical wake-up threshold The trend of change. The global optimal solution. As the main calibration anchor point for the decoupling engine, the darker areas represent combinations of failure parameters with excessively high false alarm rates or insufficient state separation.

[0044] S3.3 Construct a lightweight neural network model. After aligning the historical eye-tracking time series dataset and the carriage inertial measurement data with timestamps, use them as inputs to the lightweight neural network model as multimodal joint features to obtain the network's main task prediction output and network auxiliary output.

[0045] S3.4 Construct a joint loss function that includes task prediction loss and physical auxiliary loss. The task prediction loss is calculated from the network's main task prediction output and the hard physical truth label, while the physical auxiliary loss is calculated from the network's auxiliary output and the hidden layer supervision signal. By calculating the joint loss function and updating the parameters of the lightweight neural network model via backpropagation until the convergence condition is met, a fixed model suitable for low-latency deployment of edge devices is obtained.

[0046] Joint loss function Constructed using the following formula: ; in, Predicting losses for the main task. D top For the physical truth value of topological determinism, The structural capacity prediction value is the output of the network-assisted output. STEI The physical truth value of the state transition energy index. λ1 and λ2 are the predicted compensatory kinetic energy output by the network, and are the preset loss balance weight coefficients.

[0047] S4. The multimodal joint features generated by aligning the real-time acquisition of the driver's eye micro-movement sequence signals and the synchronous inertial measurement data of the carriage with timestamps are directly input into a pre-deployed lightweight neural network model for forward propagation calculation. Based on the stateless forward propagation output of the lightweight neural network model, the absolute state health quantification index and transient mental load quantification index of the driver are generated in parallel with low latency. Based on the critical slowdown precursor features implicitly learned within the model, an early perceptual decoupling warning signal is output.

[0048] The steps for outputting an early warning signal for perceptual decoupling include: (1) extracting baseline features representing the pure physiological background of the system for early warning determination; the baseline features include at least: the health descent gradient (Psi_Drain_Velocity) calculated by the first difference of the absolute state health quantification index. (2) extracting dynamic features representing the evolution law of compensatory kinetic energy of the system for early warning determination; the dynamic features include at least any one of the following: the recent mean, time-shift variance, or first-order evolution gradient of the transient mental load quantification index within a preset time window. (3) extracting topological geometric features representing the law of local manifold distortion in eye movement phase space for early warning determination; the topological geometric features include at least any one of the following: the recent mean of the topological deterministic index, the time-shift variance of spatial extension, the second-order evolution acceleration of spatial extension, or the nonlinear product cross term of the transient mental load quantification index and the topological deterministic index.

[0049] like Figure 5 The figure shows the forward prediction performance verification of the topology architecture. Figure 5 The evaluation results are presented under a rigorous leave-one-out-of-subject (LOSO) cross-validation protocol. Subfigure A is the receiver operating characteristic (ROC) plot: When predicting impending perceptual decoupling within the golden time window, the model exhibits extremely high discrimination (AUC > 0.95); Subgraph B is the confusion matrix: by utilizing a threshold optimized based on F2-Score to prioritize life safety interventions (maximizing recall) while maintaining strong noise suppression capabilities (precision), this architecture effectively alleviates the frequent false alarms of the "boy who cried wolf" inherent in traditional ADAS monitoring.

[0050] Figure 5 This is a time profile of the advance prediction for spontaneous perceptual decoupling. It shows the dynamic evolution of the prediction index within a 9-second advance transition window. The monotonous increase in discrimination (AUC, solid blue line) directly reflects the physical phenomenon of critical slowing (CSD), confirming that autonomic regenerative capacity exhibits a gradual, deterministic decline, rather than instantaneous failure. Within the optimal ADAS engineering intervention window ( to Within the shaded area, the engine achieved a golden cross between high recall (red dotted line) and strong precision (green dashed line), thus mitigating the impact of absolute system structural collapse. Prior to this, a mathematically validated time buffer was provided for proactive and flexible takeover.

[0051] Compared with the prior art, the present invention has achieved the following extremely significant engineering and technical effects: Stateless evaluation eliminates historical accumulated errors: Unlike traditional Markov or Kalman integral tracking, the absolute health formula of this invention relies solely on the cross-sectional evaluation of the current time window. This perfectly eliminates the state forgetting and integral drift problems commonly encountered in in-vehicle environments, such as sensor occlusion, driver head turning, or system restarts.

[0052] Unified electrodynamic isomorphism completely suppresses all compensatory artifacts: Cognitive load is mapped to the battery's discharge rate (C-rate). Whether it is the high current compensation caused by "violent restart" or the capacity collapse caused by "complete decoupling", it will be adaptively identified and forcibly damped by a single formula, completely eliminating the false negatives caused by "high entropy artifacts".

[0053] Bridging the computing power mismatch gap and achieving millisecond-level hard real-time response: The complex topological determinism and exponential discounting, which are extremely CPU-intensive, are pre-positioned on the cloud server; the vehicle edge device only needs to deploy a lightweight network, and the forward inference time is compressed to the 5ms level.

[0054] Establish a universally applicable key early warning line: The structural "half-life" threshold of τ=50% was established, which not only ensures a detection rate of 50.4% in L3 fatigue, but also suppresses the false alarm rate in normal state to 10.8%, providing a standardized trigger interface with great engineering application value for downstream ADAS chassis control.

[0055] Based on the same inventive concept, this invention also proposes an in-vehicle eye-tracking monitoring system based on physical knowledge distillation, comprising: The acquisition module is used to collect historical eye movement time series signals of the driver during the driving process and synchronous inertial measurement data of the carriage.

[0056] The extraction module is used to reconstruct the phase space of historical eye movement time series signals to extract absolute state health to characterize current cognitive reserve capacity and topological deterministic indicators to characterize the degree of degradation of the eye movement phase space manifold structure.

[0057] The model training module is used to align historical eye-tracking time-series signals and inertial measurement data from the train carriage with timestamps to generate multimodal joint features. These multimodal joint features are then input into a lightweight neural network model for training. Absolute state health is used as the primary task label, and topological determinism is used as the auxiliary supervision label to obtain the primary task prediction output and auxiliary output of the network model. Based on the primary task prediction output and auxiliary output, a joint loss function is constructed, which includes the primary task prediction loss and the physical auxiliary loss. The parameters of the lightweight neural network model are updated by calculating the joint loss function and through backpropagation until the convergence condition is met, thus completing the training of the lightweight neural network model.

[0058] The output module is used to input the multimodal joint features generated by aligning the real-time acquired driver eye movement time series signal with the synchronous inertial measurement data of the carriage with the timestamps into the trained lightweight neural network model for forward propagation, and output the driver's cognitive state inference results; the inference results include absolute state health, transient mental load and perceptual decoupling early warning signal.

[0059] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for reasoning about eye-tracking cognitive states, characterized in that, Includes the following steps: Collect historical eye-tracking time-series signals of the driver during the driving process and synchronized inertial measurement data of the carriage; Phase space reconstruction is performed on historical eye movement time series signals to extract absolute state health to characterize current cognitive reserve capacity and topological deterministic indicators to characterize the degree of degradation of the eye movement phase space manifold structure. Historical eye-tracking time-series signals and inertial measurement data from the train carriage are timestamped to generate multimodal joint features. These multimodal joint features are then input into a lightweight neural network model for training. Absolute state health is used as the primary task label, and topological determinism is used as the auxiliary supervision label to obtain the primary task prediction output and auxiliary output of the network model. Based on the primary task prediction output and auxiliary output, a joint loss function is constructed, which includes the primary task prediction loss and the physical auxiliary loss. The parameters of the lightweight neural network model are updated by calculating the joint loss function and through backpropagation until the convergence condition is met, thus completing the training of the lightweight neural network model. The multimodal joint features generated by aligning the real-time driver eye movement time series signals with the synchronous inertial measurement data of the carriage with timestamps are input into a lightweight neural network model trained for forward propagation, and the driver's cognitive state inference results are output. The inference results include absolute state health, transient mental load and perceptual decoupling early warning signals.

2. The eye-tracking cognitive state reasoning method according to claim 1, characterized in that, The step of aligning historical eye movement time series signals and carriage inertial measurement data with timestamps to generate multimodal joint features specifically includes: after aligning the eye movement sequence signals with the carriage inertial measurement data with timestamps, using early feature splicing or cross-modal cross-attention mechanisms to perform feedforward fusion of multi-source data during the forward propagation of the lightweight neural network model to obtain multimodal joint features.

3. The eye-tracking cognitive state reasoning method according to claim 1, characterized in that, After performing phase space reconstruction on the historical eye-tracking time series signal, the method further includes extracting perceptual decoupling early warning labels, specifically including the following steps: Determine the physical anchor time of perceptual decoupling in historical eye-tracking time series signals; Within a preset time window prior to the physical anchor point time, calculate the time-shifted variance and the first-order autocorrelation coefficient of the state transition energy index. The critical slowdown precursor interval is defined based on the abrupt increase characteristics of time-shifted variance and lag first-order autocorrelation coefficient. This critical slowdown precursor interval is then incorporated as a supervisory soft label into the training process of the lightweight neural network model, enabling the lightweight neural network model to implicitly internalize the nonlinear fluctuation characteristics of the time series before the collapse.

4. The eye-tracking cognitive state reasoning method according to claim 1, characterized in that, The absolute state health is calculated using an electrodynamic isomorphic decoupling model based on stateless assessment, specifically including the following steps: Topological deterministic indices characterizing the degree of degradation of the manifold structure in eye movement phases are extracted and global calibration values ​​are used in the driver's awake baseline state. Construct a capacity ceiling Φ(t) to represent the absolute cognitive reserve; Extract the state transition energy index, which represents the actual kinetic energy consumed by the driver's sympathetic nervous system activation, and the sample entropy, which represents the complexity of the representational signal, within the current sliding time window. SampEn ; Introducing a preset physical penalty coefficient, based on the state transition energy index STEI With topological determinism indicators D top The ratio of is used to construct a damping operator that characterizes the dynamic workload discounting; Multiply the capacity ceiling by the sample entropy and divide by the damping operator Γt to obtain the absolute state health within the current time window, which is free from compensatory artifacts and independent of historical states.

5. The eye-tracking cognitive state reasoning method according to claim 4, characterized in that, The absolute state health Ψ Tonic(t) Represented as: ； Where γ is the physical penalty coefficient; Capacity ceiling; τ is the damping operator; τ is the absolute critical threshold for triggering decoupling warning.

6. The eye-tracking cognitive state reasoning method according to claim 1, characterized in that, Phase space reconstruction of historical eye-tracking time series is performed using a nonlinear dynamics physics engine; The process of reconstructing the phase space also includes extracting topological dynamic features in the multidimensional phase space that characterize the capacity and compensatory kinetic energy evolution of the underlying structure of the eye-movement manifold as another hidden layer supervision signal to decouple transient mental load and absolute fatigue state from external inducements; adding this hidden layer supervision signal to the joint loss function for supervised training, so that the lightweight neural network model can adaptively identify transient high cognitive input triggered by any factor and has the ability to block system false alarms caused by transient compensatory illusions from the bottom layer.

7. The eye-tracking cognitive state reasoning method according to claim 1, characterized in that, The joint loss function is expressed as: ； in, Predicting losses for the main task. D top For the physical truth value of topological determinism, The structural capacity prediction value is the output of the network-assisted output. STEI The physical true value of the state transition energy index is... λ1 and λ2 are the predicted compensatory kinetic energy output by the network, and are the preset loss balance weight coefficients.

8. The eye-tracking cognitive state reasoning method according to claim 6, characterized in that, The topological dynamics characteristics that characterize the evolution of the capacity and compensatory kinetic energy of the eye-moving manifold's underlying structure include at least one of topological determinism, state transition energy index, or transient spatial extensibility.

9. The eye-tracking cognitive state reasoning method according to claim 1, characterized in that, The perceptual decoupling early warning signal is specifically determined by extracting baseline features representing the pure physiological background of the system, topological geometric features representing the local manifold distortion law of eye movement phase space, and dynamic features representing the evolution law of the system's compensatory kinetic energy. Among them, the dynamic features include at least one of the following: the recent mean, time-shifted variance, or first-order evolutionary gradient of the transient mental load quantification index within a preset time window; the baseline features include the health descent gradient calculated from the first difference of the absolute state health quantification index; the topological geometric features include at least one of the following: the recent mean of the topological deterministic index, the time-shifted variance of spatial extension, the second-order evolutionary acceleration of spatial extension, or the nonlinear product cross term of the transient mental load quantification index and the topological deterministic index.

10. An eye-tracking cognitive state reasoning system, characterized in that, include: The acquisition module is used to collect the driver's historical eye movement time series signals and synchronous inertial measurement data of the carriage during the driving process; The extraction module is used to reconstruct the phase space of historical eye movement time series signals to extract absolute state health to characterize current cognitive reserve capacity and topological deterministic indicators to characterize the degree of degradation of the eye movement phase space manifold structure. The model training module is used to align historical eye-tracking time-series signals and inertial measurement data of the carriage with timestamps to generate multimodal joint features. These multimodal joint features are then input into a lightweight neural network model for training. Absolute state health is used as the primary task label, and topological determinism is used as the auxiliary supervision label to obtain the primary task prediction output and auxiliary output of the network model. Based on the primary task prediction output and auxiliary output, a joint loss function is constructed, comprising primary task prediction loss and physical auxiliary loss. The parameters of the lightweight neural network model are updated by calculating the joint loss function and through backpropagation until the convergence condition is met, thus completing the training of the lightweight neural network model. The output module is used to input the multimodal joint features generated by aligning the real-time acquired driver eye movement time series signal with the synchronous inertial measurement data of the carriage with the timestamps into the trained lightweight neural network model for forward propagation, and output the driver's cognitive state inference results; the inference results include absolute state health, transient mental load and perceptual decoupling early warning signal.

Images