Driver fatigue state data determination method and device and computer equipment

Abstract

The invention relates to a driver fatigue state data determination method and device and computer equipment. The method comprises the following steps: acquiring environment state data corresponding to a running automobile and driver action data corresponding to a driver; according to the environment state data, risk events of the running automobile in a future preset time domain are predicted and analyzed, and task complexity analysis data are obtained; performing mutual information analysis on a corresponding relation between the environment state data and the driver action data to obtain driver-man mutual information analysis data; according to the task complexity analysis data and the driver mutual information analysis data, calculating a fatigue index of the driver at the current moment to obtain an information efficiency attenuation fatigue index; and performing sequential statistical test analysis on the information efficiency attenuation fatigue index to obtain fatigue state data corresponding to the driver. By adopting the method, the generalization ability and the stability of determining the fatigue state of the driver can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus and computer equipment for determining driver fatigue status data. Background Technology

[0002] Traditional driver fatigue assessment technologies typically begin by using basic in-vehicle sensors to roughly observe the driver's state and driving behavior, as well as any abnormalities in vehicle operation. These signals are then used to make an overall trend judgment. When the judgment indicates a significant tendency towards fatigue, a conservative strategy is employed, issuing mild prompts such as audio or interface prompts, and providing appropriate assistance when necessary, while gradually adapting to individual habits. However, traditional driver fatigue assessment technologies heavily rely on fixed rules and limited scenarios, making it difficult to adapt to individual differences and environmental changes, resulting in insufficient generalization ability and stability in driver fatigue assessment. Summary of the Invention

[0003] Therefore, it is necessary to provide a method, apparatus, and computer equipment for determining driver fatigue state data that can effectively improve the generalization ability and stability of driver fatigue state determination, in order to address the above-mentioned technical problems.

[0004] Firstly, this application provides a method for determining driver fatigue state data, including:

[0005] Acquire environmental status data for the running vehicle and driver action data for the driver;

[0006] Based on the environmental state data, the risk events of the operating vehicle in the future preset time domain are predicted and analyzed to obtain task complexity analysis data.

[0007] Mutual information analysis is performed on the correspondence between the environmental state data and the driver action data to obtain vehicle occupant mutual information analysis data.

[0008] Based on the task complexity analysis data and the driver-passenger mutual information analysis data, the driver's fatigue index at the current moment is calculated to obtain the information efficiency decay fatigue index.

[0009] Sequential statistical tests and analyses were performed on the information efficiency decay fatigue index to obtain the fatigue state data corresponding to the driver.

[0010] Secondly, this application also provides a device for determining driver fatigue state data, comprising:

[0011] The vehicle data acquisition module is used to acquire environmental status data corresponding to the running vehicle and driver action data corresponding to the driver.

[0012] The vehicle risk prediction module is used to predict and analyze risk events of the operating vehicle in a future preset time domain based on the environmental state data, and obtain task complexity analysis data.

[0013] The mutual information analysis module is used to perform mutual information analysis on the correspondence between the environmental state data and the driver action data to obtain vehicle occupant mutual information analysis data.

[0014] The fatigue index calculation module is used to calculate the driver's fatigue index at the current moment based on the task complexity analysis data and the driver-passenger mutual information analysis data, and obtain the information efficiency decay fatigue index.

[0015] The fatigue index analysis module is used to perform sequential statistical test analysis on the information efficiency decay fatigue index to obtain the fatigue state data corresponding to the driver.

[0016] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step of a method for determining driver fatigue state data.

[0017] The aforementioned method, apparatus, and computer equipment for determining driver fatigue status quantify the external environmental load into a task complexity curve through future-oriented risk event prediction. Simultaneously, it uses mutual information to accurately characterize the intensity and direction of information transmission between driver actions and environmental elements, constructing an information efficiency decay fatigue index that reflects changes in human-machine information processing capabilities. Based on this, sequential statistical tests are introduced to set significance and stopping rules, achieving second-level online discrimination and continuous updates. This not only provides early warnings before significant behavioral degradation but also effectively controls false alarms and missed alarms. It effectively reduces equipment dependence and improves adaptability to individual differences, thereby significantly improving the generalization ability and stability of driver fatigue status determination, ultimately leading to a significant improvement in driving safety, driving comfort, and overall system operational efficiency. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a diagram illustrating the application environment of a method for determining driver fatigue status data in one embodiment.

[0020] Figure 2This is a flowchart illustrating a method for determining driver fatigue status data in one embodiment;

[0021] Figure 3 This is a flowchart illustrating a method for obtaining the information efficiency decay fatigue index in one embodiment.

[0022] Figure 4 This is a flowchart illustrating a method for obtaining information cost β data in one embodiment;

[0023] Figure 5 This is a flowchart illustrating a method for obtaining relative entropy control strategy stream data in one embodiment;

[0024] Figure 6 This is a flowchart illustrating a method for obtaining data from the interaction information analysis between vehicle occupants in one embodiment.

[0025] Figure 7 This is a flowchart illustrating a method for obtaining target interval data of sequence mutual information in one embodiment;

[0026] Figure 8 This is a flowchart illustrating a method for obtaining inequality shrinkage data in one embodiment;

[0027] Figure 9 This is a flowchart illustrating a method for obtaining fatigue state data in one embodiment;

[0028] Figure 10 A structural block diagram of a driver fatigue state data determination device in one embodiment;

[0029] Figure 11 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0031] The driver fatigue state data determination method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on other network servers. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0032] In one exemplary embodiment, such as Figure 2As shown, a method for determining driver fatigue state data is provided, which can be applied to... Figure 1 Taking the server in the example, the explanation includes the following steps 202 to 210. Wherein:

[0033] Step 202: Obtain the environmental status data corresponding to the running vehicle and the driver's action data corresponding to the driver.

[0034] Step 204: Based on the environmental status data, predict and analyze the risk events of the operating vehicle in the future preset time domain to obtain task complexity analysis data.

[0035] Step 206: Perform mutual information analysis on the correspondence between environmental state data and driver action data to obtain vehicle and driver mutual information analysis data.

[0036] Step 208: Based on the task complexity analysis data and the driver-driver mutual information analysis data, calculate the driver's fatigue index at the current moment to obtain the information efficiency decay fatigue index.

[0037] Step 210: Perform sequential statistical test analysis on the information efficiency decay fatigue index to obtain the driver's corresponding fatigue state data.

[0038] Among them, environmental state data is time-series data obtained from vehicle perception and external information, used to describe scene elements such as road geometry, traffic participants, traffic signs / speed limits, road surface adhesion, and weather.

[0039] Among them, driver action data is time-series data of the control inputs generated by the driver through the human-machine interface (such as steering wheel angle / angular velocity / hand torque, accelerator / brake pedal travel and rate of change, turn signal / lane change control signals, etc.).

[0040] Among them, the future preset time domain is the analysis interval of a future time window (fixed length or adaptive length) pre-set for risk prediction.

[0041] Among them, risk events are a set of events that may occur within the time window and affect driving safety, such as lane departure, time-distance violation, longitudinal and lateral collisions, speeding, low adhesion / sudden drop in visibility, etc.

[0042] Among them, the task complexity analysis data is a set of decision-oriented metrics based on risk prediction, including event occurrence probability and severity, minimum intervention cost, hazard uncertainty / risk weight, state transition prior and time window weight, etc.

[0043] Mutual information analysis is a process of directional and robust estimation of the information dependency relationship between "environmental state → driver action" under unpredictable constraints (which may include variational upper and lower bounds, cross-domain robustness, and information inequality constraints).

[0044] Among them, the driver-passenger mutual information analysis data is the result data package of mutual information analysis, which includes mutual information point values ​​and intervals at the step-by-step and sequence levels, confidence and quality labels, as well as the marginal information contribution of key environmental factors to the action.

[0045] The fatigue index is a scalar indicator used to quantify the current level of alertness / fatigue of the driver. It is obtained by mapping behavioral information features to an individual baseline.

[0046] Among them, the Information Efficiency Decay Fatigue Index is a dedicated fatigue indicator based on the decline in information usage cost or efficiency (relative to the sober baseline), reflecting the degree of decay in the amount of effective information that the driver obtains from the environment and uses for action.

[0047] Among them, sequential statistical test analysis is a process of online testing and confirmation of fatigue index over continuous time, using readily available boundary and evidence measures to control false alarms and locate the moment of change.

[0048] Among them, fatigue status data is the final output of the driver's status result, which includes meta-information such as fatigue level / label, confirmation time and duration, confidence level (or strength of evidence) and triggering basis.

[0049] Specifically, the system synchronously collects initial environmental state data (lane geometry, relative traffic flow, speed limits / signs, road surface and weather, etc.) and initial driver action data (steering wheel angle / angular velocity / hand torque, accelerator / brake pedal travel and rate of change, lane change signals, etc.) corresponding to the running vehicle from the vehicle perception and control bus. This data is unified to the same time base, and clock drift and jitter of each ECU are corrected. Then, the raw data undergoes unit standardization, zero-drift calibration, and outlier removal. Median / low-pass filtering and wavelet denoising are used to suppress the effects of road impacts and electrical glitches. Multi-source data is resampled to a unified frequency and sliced ​​using a fixed sliding window, while unpredictable pruning is performed, retaining only information currently and historically usable. Based on torque / motor current and power assist status, driver manual inputs and auxiliary control quantities are separated. Active intervention periods such as ACC / LKA / AEB are automatically masked. Boundary conservative interpolation is applied to missing or obscured segments, with quality scoring and masking. Key event markers (lane change, braking, sharp turn, etc.) are retained, resulting in environmental state data corresponding to the running vehicle and driver action data corresponding to the driver.

[0050] Based on environmental state data, scene semantic enhancement (such as high-precision map features, speed limits and slopes, weather and road surface adhesion, traffic density and relative speed distribution) and condition segmentation (such as straight / curved roads, ramps, urban / highway, sparse / congested) are performed. The processed data is then used to jointly extrapolate key indicators within a predetermined time domain using a multi-model rolling prediction approach (vehicle trajectory, adjacent vehicle trajectory, cutting / deceleration intention, lane available bandwidth and curvature changes, visibility attenuation). For potential events (lane departure, time-distance violation, forward / lateral conflict, speed limit exceeding, sudden drop in adhesion), the probability of occurrence, severity, minimum intervention cost, and time to event are calculated, and uncertainty assessment (model ensemble and conformal intervals) is used to provide confidence boundaries. By combining sensor health and road grade, risk weighting and temporal attenuation are applied to events, generating task complexity analysis data.

[0051] First, the environmental state data and the driver's action data undergo unpredictable processing, retaining only information currently and historically available. An information bottleneck method is used to extract control representations directly related to action generation to reduce irrelevant interference. Based on the pre-processed data, an information metric for "environment → action" is established using directional variational upper and lower bound estimation. Simultaneously, cross-domain partial Bruker correction is implemented across environmental subdomains such as different road types, traffic densities, and lighting / weather conditions to suppress spurious correlations and mismatches caused by scene migration. After obtaining the initial stepwise and sequential mutual information intervals, information theory constraints such as data processing inequalities and directed information decomposition are introduced. Feasibility domain projection is performed to tighten the intervals and eliminate solutions inconsistent with the causal direction, generating a driver-driver mutual information analysis data package. This package includes stepwise and sequential mutual information point values ​​and target intervals, confidence and quality labels, marginal information contribution vectors of key environmental elements to actions (based on contrast masking / information Shapley approximation), and their temporal location.

[0052] Using task complexity analysis data and driver-driver mutual information analysis data as constraints, an information-cost dual model of "driving cost versus information cost trade-off" is constructed. Within a fixed time window, the reference action distribution and risk score are jointly transformed into path priors. Subsequently, a Schrödinger bridge entropy regularized forward-backward proportional scaling iteration is adopted to obtain the driver-oriented strategy path distribution sequence under the principle of minimum relative entropy deviation, and a monotonic mapping between the temperature parameter β and the mutual information of the model sequence is formed accordingly. Combining the mutual information target interval, a one-dimensional monotonic root solution and confidence convergence determination are performed on this mapping to obtain information cost β data and its uncertainty. The information cost β data is compared with the individual sobriety baseline and trend and risk weights are superimposed to generate an information efficiency decay fatigue index (including point values, intervals, and quality labels).

[0053] The information efficiency decay fatigue index is time-varyingly weighted and self-regularized for scenario risk and data quality. An evidence stream (e-value sequence) based on betting tests is constructed, and a judgment boundary that is effective at any time is generated simultaneously to ensure that the false alarm rate is controllable under arbitrary stop time. Furthermore, "dual threshold irreversible confirmation + multi-scale change point positioning" is implemented on a single time axis. That is, an early warning is first triggered with a sensitive threshold, and then irreversible confirmation is completed with a higher threshold. At the same time, hysteresis and cooling time are used to avoid frequent jitter. For windows with ACC / LKA / AEB intervention, missing data, or low quality, a masking and robust aggregation strategy is adopted, and a conservative judgment is rolled back when necessary. Finally, fatigue state data including fatigue level, confirmation time and duration, evidence strength, false alarm control commitment, and triggering basis are obtained.

[0054] In the aforementioned method for determining driver fatigue status data, the external environmental load is quantified into a task complexity curve through future-oriented risk event prediction. Simultaneously, mutual information is used to accurately characterize the intensity and direction of information transmission between driver actions and environmental elements, constructing an information efficiency decay fatigue index that reflects changes in human-machine information processing capabilities. Based on this, sequential statistical tests are introduced to set significance and stopping rules, enabling second-level online discrimination and continuous updates. This not only provides early warnings before significant behavioral degradation but also effectively controls false alarms and missed alarms. Furthermore, it effectively reduces equipment dependence and improves adaptability to individual differences, thereby significantly enhancing the generalization ability and stability of driver fatigue status determination, ultimately leading to a significant improvement in driving safety, driving comfort, and overall system operational efficiency.

[0055] In one exemplary embodiment, such as Figure 3 As shown, based on task complexity analysis data and driver-vehicle interaction information analysis data, the driver's fatigue index at the current moment is calculated to obtain the information efficiency decay fatigue index, including steps 302 to 304. Wherein:

[0056] Step 302: Based on the task complexity analysis data and the driver-passenger mutual information analysis data, the driver's information-cost dual model is solved by inversion to obtain the information cost β data.

[0057] Step 304: Calculate the driver's fatigue index at the current moment based on the information cost β data to obtain the information efficiency decay fatigue index.

[0058] Among them, the information-cost dual model unifies the cost of driving performance and the cost of information use into a trade-off framework. It links "strategy deviation" with "required information" through duality and is used to inversely determine the driver's information use tendency.

[0059] Among them, the information cost β data is the parameter value and its time series / confidence level that measures the "unit information cost" under the above model, which is used to quantify the driver's willingness to use environmental information under the current working conditions.

[0060] Among them, the fatigue index is a normalized scalar score that monotonically represents the driver's current fatigue level based on the change in information cost or information use efficiency relative to the sober baseline.

[0061] Specifically, based on the task complexity analysis data and the driver-passenger mutual information analysis data, the reference action distribution, risk score, and state transition prior are integrated into a driver-oriented path prior. Then, the target interval of the sequence mutual information from the driver-passenger mutual information analysis data is used as the path mutual information constraint. Within a fixed time window, a forward-backward scaling iteration using Schrödinger bridge entropy regularization is performed on the driver's state-action path probability flow to obtain the strategy path distribution sequence with minimum relative entropy deviation. Based on this sequence, a monotonic mapping between temperature parameters and model sequence mutual information is constructed. One-dimensional monotonic root solving is used, combined with homotopy extension and convergence criteria, to ensure that the model mutual information is consistent with the target interval, thereby obtaining information cost β data and its uncertainty and quality label (while simultaneously shielding active intervention periods such as ACC / LKA / AEB, processing missing segments, and recording iterative convergence and constraint satisfaction).

[0062] Aligning information cost β data with individual baselines and permissible fluctuation bands established under driver alert conditions, the process achieves time synchronization, anomaly masking, quality grading, and scenario risk weighting. Then, it robustly aggregates the current level, short-term slope, volatility, and persistence of β, combining task complexity weights and data integrity weights to form a monotonic mapping. This transforms the relationship of "increased information cost equals decreased efficiency" into a single-scale fatigue score. To suppress jitter and false triggers, the process incorporates hysteresis and cooldown time strategies. It implements deweighting or rollback strategies for ACC / LKA / AEB interventions and low-confidence segments, and sets sensitive and inertial channels for sudden jumps and slow climbs, respectively. Finally, it obtains the information efficiency decay fatigue index and uncertainty range, along with interpretable meta-information (such as main contributing factors, trigger scenario subdomains, data quality labels, and mapping versions).

[0063] In this embodiment, by using task complexity analysis data and vehicle-person mutual information analysis data as constraints to invert the information-cost dual model, interpretable information cost β data is output, and the information efficiency decay fatigue index is obtained accordingly. This enables low-cost deployment that does not rely on camera / physiology and active stimulation and directly reuses vehicle-end signals. It facilitates one-dimensional root solution and timely and effective sequential judgment, takes into account low false alarms and auditability, and has the advantages of privacy friendliness and cross-vehicle / operating condition portability.

[0064] In one exemplary embodiment, such as Figure 4As shown, based on task complexity analysis data and driver-driver mutual information analysis data, the driver's information-cost dual model is inverted and solved to obtain information cost β data, including steps 402 to 408. Wherein:

[0065] Step 402: Based on the task complexity analysis data, analyze the driver's unreferenced action distribution and driving risk potential to obtain path prior data.

[0066] Step 404: Analyze the mutual information constraints of the vehicle and passenger mutual information analysis data to obtain path mutual information constraint data.

[0067] Step 406: Based on the path prior data and path mutual information constraint data, perform Schrödinger bridge entropy regular iterative analysis on the driver's state-action path probability flow to obtain relative entropy control strategy flow data.

[0068] Step 408: Based on the relative entropy control strategy flow data, solve for the one-dimensional monotonic root of the mutual information-temperature monotonic operator for the driver to obtain the information cost β data.

[0069] Among them, the non-environmental reference action distribution is the baseline distribution of driving actions given only based on safety and controllability constraints without utilizing environmental state information.

[0070] Among them, driving risk potential is a risk scoring function that maps factors such as the severity of events, arrival time and road conditions in future scenarios into comparable risk scoring functions.

[0071] Among them, the path prior data is a set of path-level initialization information composed of reference action distribution, risk potential parameters, state transition priors and time window weights.

[0072] Among them, the mutual information constraint is the specification of the target value or target range and its confidence level of the information intensity of "environmental state → driver action".

[0073] Among them, path mutual information constraint data is constraint data used for path-level optimization, which is formed by packaging mutual information objectives, intervals, weights and consistency check results.

[0074] Among them, the state-action path probability flow is the probability distribution and flow process of the driver's state and action sequence evolving over time.

[0075] Among them, the Schrödinger Bridge Entropy Regular Iterative Analysis is a solution process that uses entropy regularized forward-backward scaling iteration to match path priors with mutual information constraints in order to reshape the path distribution.

[0076] Among them, the relative entropy control strategy flow data is the strategy path distribution sequence and its normalization factor and constraint satisfaction log obtained under the principle of minimum relative entropy deviation.

[0077] Among them, the mutual information-temperature monotonic operator maps temperature parameters to a monotonic function relationship of mutual information of the model sequence under fixed constraints.

[0078] Among them, the one-dimensional monotonic root is the parametric solution obtained by one-dimensional search on the monotonic relation, which makes the model consistent with the objective.

[0079] Specifically, based on task complexity analysis data, without considering environmental conditions, a referenceless action distribution is first constructed from the set of safe operations and the minimum intervention cost (for example, parameters are set according to the principle that the higher the risk, the more convergent the reference distribution). The driving risk potential is then modeled (unifying factors such as the severity of events in the future time domain, arrival time, road grade, and sensor health into the same risk scoring framework). Subsequently, time window anchor point alignment and state transition prior extraction are performed on the above referenceless action distribution and the modeled driving risk potential to obtain path prior data containing reference action distribution parameters, risk potential parameters, state transition kernel, time window weights, and quality labels.

[0080] The driver-vehicle mutual information analysis data is aligned with time windows and subjected to confidence processing. The stepwise and sequential mutual information targets and their confidence intervals for "environmental state → driver action" are extracted, and domain weights and tolerances are given in combination with scene subdomains (straight / curved roads, traffic density, lighting and weather). At the same time, unpredictable constraints and data quality masks are applied to remove biases caused by ACC / LKA / AEB interventions and missing segments, resulting in path mutual information constraint data that includes target intervals, domain weights, tolerances, and consistency check results.

[0081] The prior path data and path mutual information constraint data are jointly input into the Schrödinger Bridge Entropy Regular Iterative Model. The model uses forward-backward scaling to constrain and shape the driver's state-action path probability flow. The forward step follows state transitions and risk potential to ensure reachability and safety, while the backward step matches the mutual information target interval to meet information usage intensity requirements. In each iteration, convergence and constraint satisfaction checks, anomaly window masks, and robust reweighting are performed. When the iteration satisfies the residual threshold and stability criterion, the strategy path distribution sequence, normalization factor, and constraint satisfaction log under the meaning of minimum relative entropy deviation are output as relative entropy control strategy flow data.

[0082] Based on the relative entropy control strategy streaming data, a monotonic relationship of "temperature parameter β - model sequence mutual information" is constructed. Then, homotopy extension is used to progressively advance from the steady-state interval, combined with a one-dimensional root solution strategy, to find a solution consistent with the observed mutual information at the boundary of the target interval. During the solution process, monotonicity violations and multiple solution cases are protected (e.g., interval shrinkage, subgradient clues, and breakpoint restart). The output information cost β data includes point values, confidence ranges, and solution quality markers.

[0083] In this embodiment, the actual degree of driver utilization of environmental information is inverted into an interpretable information cost β through a path prior + mutual information constraint → Schrödinger bridge entropy regularization iteration → one-dimensional monotonic root link, thereby obtaining a stable and traceable fatigue quantification result. It can directly reuse vehicle-side perception and control bus signals without the need for cameras / physiological and active stimulation, has low deployment cost, and is more robust to nighttime, backlight, and occlusion. At the same time, it can output parameterized evidence and constraint satisfaction logs according to the individual driver baseline. The results are auditable, interpretable, privacy-friendly, and have good transferability to different vehicle models and operating conditions.

[0084] In one exemplary embodiment, such as Figure 5 As shown, based on the path prior data and path mutual information constraint data, a Schrödinger bridge entropy regularized iterative analysis is performed on the driver's state-action path probability flow to obtain the relative entropy control strategy flow data, including steps 502 to 508. Wherein:

[0085] Step 502: Based on the path prior data, construct the driver's transition kernel under the state-action path to obtain uncontrolled path kernel data.

[0086] Step 504: Based on the path mutual information constraint data, perform dualization modeling on the mutual information confidence interval of the vehicle and driver mutual information analysis data to obtain the mutual information Lagrange variables and the initial value data of the dual potential.

[0087] Step 506: Based on the uncontrolled path kernel data, mutual information Lagrange variables, and initial values ​​of dual potentials, perform a forward-backward scaling iterative analysis of the state-action path probability flow using Schrödinger bridge entropy regularization to obtain the policy path distribution sequence data.

[0088] Step 508: Based on the strategy path distribution sequence data, analyze the driver's strategy density under causal coupling constraints to obtain relative entropy control strategy flow data.

[0089] Among them, the state-action path is a joint trajectory composed of the environmental state sequence and the driver action sequence arranged in chronological order within a preset time window.

[0090] The transition kernel is a mechanism that describes how the current state and action evolve to the state at the next time step with a given probability.

[0091] Among them, uncontrolled path kernel data is the state-action path transition kernel and its metadata constructed solely based on task complexity and reachability without imposing mutual information or control constraints.

[0092] The mutual information confidence interval is the range of statistical uncertainty given for the information strength of "environmental state pointing to driver action".

[0093] Dual modeling involves rewriting the objective and constraints as dual problems, and using variables such as Lagrange multipliers to express the constraint strength for iterative solution.

[0094] Among them, the mutual information Lagrange variable is the multiplier that adjusts the strength of the mutual information target (a weighted schedule arranged by time and scenario subdomain).

[0095] The initial value data of the dual potential is the potential field configuration initialized for the forward-backward solution, which includes rules such as boundary treatment, anomaly repair and homotopy tightening.

[0096] Among them, the Schrödinger Bridge Entropy Regularization is a method that uses entropy regularization as a stabilizer to reshape the path distribution between the prior and the target through the principle of minimum relative entropy.

[0097] Among them, the forward-backward scaling iterative analysis is an iterative process that alternately performs forward propagation and backward weighting to simultaneously match edge and target constraints.

[0098] Among them, the strategy path distribution sequence data is the probability distribution sequence of driver state-action paths covering the entire time window, along with its normalization factor and convergence log.

[0099] Among them, the causal coupling constraint requires that the strategy only depends on past and current information and is consistent with the state evolution model, and must not utilize future states.

[0100] Among them, strategy density is the conditional probability distribution of the driver's possible actions at each moment based on the observed history, reflecting his decision-making tendency.

[0101] Specifically, based on prior path data, the time axis is discretized into fixed sliding windows. State and action features that represent driving semantics (such as lane geometry, relative traffic flow, attachment and visibility, steering wheel / pedal micro-components, etc.) are selected. Feasibility domains and safety constraint encoding are performed on reachable state-action pairs according to road level and risk potential function. Smoothing estimation and confidence correction are performed using historical priors and simulation playback. A "state-action path transition kernel" that follows reachability, conservation and sparsity is constructed. Simultaneously, version number, quality label and mask (masking ACC / LKA / AEB intervention sections) are output to obtain uncontrolled path kernel data.

[0102] First, based on the path mutual information constraint data, the solution time window and scene subdomain are aligned to the computation grid one by one. Then, the mutual information target value, upper and lower limits of the interval, and quality label of the corresponding cell are read from the mutual information confidence interval of the vehicle and crew mutual information analysis data to form a gridded mutual information target-interval table. On this basis, the target and interval of each grid cell in the above table are mapped to the constraint strength and weight plan according to the quality label, and a mask matrix is ​​generated to mask low confidence and missing segments. Under the constraints of unpredictability and causal consistency, the constraint strength, weights, and masks of the mask matrix are compiled grid by grid into an initial mutual information Lagrange variable schedule (including initial values, update upper limits, and convergence thresholds) arranged by time and subdomain. Simultaneously, a dual potential initial value grid (including boundary handling, anomaly repair, and homotopy tightening rules) compatible with the uncontrolled path kernel is constructed. If a conflict is found between the target or interval and the prior, out-of-bounds pruning and feasible region widening are triggered and written back to the variable schedule and potential grid. The final output is mutual information Lagrange variables and dual potential initial value data that can directly drive forward-backward scaling iterations.

[0103] The uncontrolled path kernel data, mutual information Lagrange variables, and initial values ​​of the dual potential are fed into the Schrödinger Bridge Entropy Regularized Model. The model employs a forward-backward proportional scaling iteration. In the forward step, the state-action path probability flow is advanced according to the uncontrolled path kernel, and a risk potential is superimposed to ensure reachability and safety. In the backward step, the path weights are multiplicatively redistributed based on the Lagrange variables and initial values ​​of the dual potential to approximate the mutual information target, while maintaining marginal consistency and unpredictable constraints. Homotopy tightening, damping, and trust domain step size control are enabled during iteration to stabilize convergence. Masking and robust reweighting are applied to ACC / LKA / AEB interventions and anomaly windows. The target deviation, marginal residuals, and convergence rate are used as stopping criteria to obtain policy path distribution sequence data covering a preset time window, along with their normalization factors, constraint satisfaction, and convergence logs.

[0104] After extracting driver-oriented policy density stepwise from the policy path distribution sequence data in chronological order, strict causal coupling constraints are applied (relying only on current and historical information, and consistent with state transition and safe reach constraints). Future information leakage is removed through marginalization and conditionalization, and compliance checks are performed using edge consistency verification, KL budget threshold, and risk potential constraints. Specifically, ACC / LKA / AEB interventions, missing information, and low-confidence windows are masked and robustly weighted, and if necessary, reverted to an action distribution without environmental reference to ensure conservatism. For abrupt transitions, segmented smoothing and breakpoint restarts are used to maintain temporal continuity. Homotopy tightening and trust domain step size control are performed for cross-scenario switching to suppress oscillations. Simultaneously, normalization, unit and timestamp alignment, and version marking are completed to obtain "minimum relative entropy control policy flow data" containing policy density time series, applicable scenario labels, constraint satisfaction, quality scores, and audit logs.

[0105] In this embodiment, by relying solely on vehicle-side perception and control bus data without requiring cameras / physiological or external stimuli during the driver's operation, the path prior and mutual information confidence interval are unified through Schrödinger bridge entropy regularization and dualization to form a robust one-dimensional monotonic root inversion. This enables earlier, lower false alarm rate, and auditable output of the driver's fatigue state.

[0106] In one exemplary embodiment, such as Figure 6 As shown, mutual information analysis is performed on the correspondence between environmental state data and driver action data to obtain vehicle-occupant mutual information analysis data, including steps 602 to 608. Wherein:

[0107] Step 602: Perform unpredictable cropping on the environmental state data and driver action data to obtain unpredictable control cropping data.

[0108] Step 604: Perform information bottleneck control representation modeling on the unpredictable control pruning data to obtain aligned sequence data and unpredictable control representation data.

[0109] Step 606: Based on the aligned sequence data and the unpredictable control representation data, perform cross-domain robust directional variational mutual information analysis on the mutual information between the environmental state data and the driver's action data to obtain the target interval data of the sequence mutual information.

[0110] Step 608: Based on the target interval data of sequence mutual information, perform conformal confidence calibration analysis on the components of mutual information to obtain the vehicle crew mutual information analysis data.

[0111] Among them, non-predictive pruning is the process of retaining only currently available and historical information while shielding any information leaked or inferred from future states.

[0112] Among them, the unpredictable control clipping data is a dataset of state and action sequences that have been time-aligned, quality-marked, and masked, and that meet the unpredictable constraints.

[0113] Among them, information bottleneck control representation modeling is the process of extracting a compact representation that retains only the information necessary for generating driver actions under unpredictable conditions, in order to suppress irrelevant interference.

[0114] Among them, the aligned sequence data is the state and action time sequence after synchronization with the scene subdomain according to a unified time base, along with its labels and quality markers.

[0115] Among them, the unpredictable control representation data is a set of state control representation features that satisfy unpredictability and are oriented towards action prediction.

[0116] Among them, cross-domain robust directional variational mutual information analysis is an analytical method that estimates mutual information by taking "environmental state → driver action" as the causal direction and combining variational upper and lower bounds with cross-domain robust correction.

[0117] Mutual information is an indicator that measures the strength of statistical dependence between two variables, i.e., the amount by which the uncertainty of one variable is reduced for the other.

[0118] Among them, the target interval data of sequence mutual information is the directional mutual information estimate at the sequence level within a preset time window, along with its uncertainty range and related metadata.

[0119] Among them, the component is a single part of the environmental element (such as lane geometry, traffic density, etc.) used to assess its marginal contribution to motion mutual information.

[0120] Among them, conformal confidence calibration analysis is a process of performing distribution-independent calibration of interval coverage using conformal inference under the conditions of limited samples and unknown distribution.

[0121] Specifically, after performing unified time-base alignment and quality screening on environmental state data and driver action data, only information currently and historically available needs to be retained within a preset time window, while shielding any direct or indirect leakage of future states. Further, based on the above processing results, masking and robust reweighting are applied to ACC / LKA / AEB interventions, sensor jitter, and missing segments. After resampling, anomaly removal, and unit normalization, unpredictable control pruning data is output.

[0122] Unpredictable control data is consistently aligned and quality-graded at time anchors and scene subdomains. Then, control representations for action generation are extracted using only current and historical information. Specifically, factors unrelated to driving decisions (lighting, appearance, background traffic fluctuations, sensor jitter, ADAS intervention times, etc.) are masked, desensitized, and robustly reweighted to enhance responses to key controllable elements (lane geometry, relative traffic flow, adhesion and visibility changes, speed limits, and sign events). Temporal smoothing and subdomain consistency constraints are introduced during representation learning to suppress spurious correlations during scene transitions, and dimensional adaptation and drift monitoring are implemented to ensure long-term stability. Two types of results are obtained: one is aligned sequence data with a unified time base, subdomain labels, and quality markers; the other is unpredictable control representation data that retains only the necessary information for action generation and meets the unpredictability requirements.

[0123] Based on aligned sequence data and unpredictable control representation data, a variational estimation with parallel upper and lower bounds is employed to align the causal direction of "environmental state pointing to driver action." A sliding window is used to output time-continuous mutual information point values ​​and intervals. To suppress spurious correlations caused by scene migration and data heterogeneity, hierarchical modeling is performed according to subdomains such as road type, traffic density, and lighting / weather. Cross-domain robustness is achieved through adversarial reweighting, importance resampling, and uncertainty set constraints. Masking and robust aggregation are performed for low-confidence, missing, and ADAS intervention periods. Finally, residual consistency and monotonicity checks are conducted under multi-model and multi-window cross-validation. Hysteresis and cooling time are combined to mitigate short-term jitter, generating target interval data for sequence mutual information with subdomain weights, tolerances, and quality labels.

[0124] Starting with the target interval data of sequential mutual information, environmental elements are decoupled at the component level (such as lane geometry, relative traffic flow, attachment and visibility, speed limits and signs, etc.). Then, the marginal impact trajectory of each component is obtained by comparing masking and interpretable feature perturbations. Conformal inference is introduced, and the marginal impact trajectory is subjected to coverage checks and adaptive tightening of component-level and overall mutual information intervals in a distribution-independent manner within sliding windows and scene subdomains. During the processing, a reserved calibration set and cross-time rolling updates are adopted, and masks and weight reduction are enabled for low confidence, missing, and ADAS intervention windows. At the same time, anomaly handling such as directional consistency and monotonicity checks, hysteresis and cooldown time stabilization, breakpoint restart and conservative expansion are performed to obtain vehicle-occupant mutual information analysis data, including stepwise and sequential mutual information point values ​​and target intervals, component contribution vectors and their confidence bands, subdomain weights and quality labels, and timestamps and version metadata.

[0125] In this embodiment, through a chain-like processing of unpredictable pruning → information bottleneck control representation → cross-domain robust directional variational mutual information → component-level conformal confidence calibration, without adding any cameras / physiological sensors, the vehicle-occupant mutual information analysis data can be obtained solely from existing vehicle-end signals. This data is causally consistent, robust across scenarios, has guaranteed coverage, and is interpretable. It can significantly reduce false alarms / false negatives and improve portability and audit availability, providing highly reliable constraint inputs for fatigue index calculation.

[0126] In one exemplary embodiment, such as Figure 7 As shown, based on the aligned sequence data and the unpredictable control representation data, cross-domain robust directional variational mutual information analysis is performed on the mutual information between environmental state data and driver action data to obtain the target interval data of the sequence mutual information, including steps 702 to 706. Wherein:

[0127] Step 702: Based on the aligned sequence data and the unpredictable control representation data, perform directional variational upper and lower bound analysis on the mutual information to obtain directional variational mutual information upper and lower bound data.

[0128] Step 704: Perform cross-domain robust correction on the upper and lower bounds of the directional variational mutual information data in different environmental subdomains to obtain cross-domain robust directional mutual information interval data.

[0129] Step 706: Based on the aligned sequence data and the unpredictable control representation data, project the cross-domain robust directional mutual information interval data into the feasible region of the information inequality to obtain the sequence mutual information target interval data.

[0130] Among them, directional variational upper and lower bound analysis uses variational methods to simultaneously estimate the lower and upper bounds of mutual information in the causal direction of "environmental state → driver action" to obtain a time-continuous interval metric.

[0131] Among them, the directional variational mutual information upper and lower bound data are a set of data consisting of mutual information point values, upper and lower bounds, residuals and diagnostic indicators output by the above analysis in each time window.

[0132] The environmental subdomain is a homogeneous data area divided according to scene elements such as road type, traffic density, and lighting and weather, which is used for hierarchical modeling and evaluation.

[0133] Among them, cross-domain sub-Bruker correction is a process that addresses the distribution differences between different environmental subdomains by using methods such as reweighting, resampling, and uncertain set constraints to suppress spurious correlations and improve generalization.

[0134] Among them, the cross-domain robust directional mutual information interval data is the mutual information interval and its subdomain weights and quality labels obtained in the causal direction after cross-domain robust correction.

[0135] Among them, the feasible region projection of information inequality is an operation that combines constraints such as data processing inequalities, strong contraction coefficients and directed information decomposition into a feasible region, and projects the mutual information interval into the feasible region to obtain a consistent and tightest result.

[0136] Specifically, based on aligned sequence data and unpredictable control representation data, a causal direction metric for "environmental state → driver action" is established using a sliding window scrolling method. A variational estimation pipeline with parallel upper and lower bounds is adopted, and interpretable discriminant / density approximators, regularization, and early stop strategies are configured for the lower and upper bounds, respectively. Masking and robust reweighting are enabled for ACC / LKA / AEB interventions, missing, and anomalous segments. Mutual information point estimates, upper and lower bounds, residuals, and diagnostic indicators are output in each window to form time-continuous directional variational mutual information upper and lower bound data.

[0137] The upper and lower bounds of the directional variational mutual information data are stratified according to environmental subdomains such as road type, traffic density, and lighting / weather. Then, cross-domain offsets are handled based on importance reweighting, adversarial reweighting, and uncertain set constraints. Outlier pruning, trust region step size, and homotopy tightening are combined to suppress overfitting and spurious correlations. After confidence-weighted and robust aggregation of the results for each subdomain, a robust cross-domain directional mutual information interval data with subdomain weights and quality labels is obtained.

[0138] By utilizing aligned sequence data and causal structures extracted from unpredictable control representation data, consistency conditions such as data processing inequalities, strong contraction coefficient constraints, directed information chain decomposition, and error-information inequalities are simultaneously applied to cross-domain robust intervals to construct feasible regions of mutual information. When an interval conflicts with a feasible region, it is tightened and written back according to convex / Bregman projection and out-of-bounds pruning rules, and the sequence mutual information target interval data with consistency certificates and version metadata is output.

[0139] In this embodiment, by using parallel upper and lower bounds to avoid bias from a single lower bound, cross-domain correction suppresses spurious correlations caused by different road / flow / lighting subdomains and improves generalization. Finally, information theory constraints such as data processing inequalities are used to project the feasible region of the interval to obtain a tighter and more reliable target interval for sequence mutual information. Thus, without adding camera / physiological sensors, it is possible to provide key constraint inputs for fatigue determination with low false alarms, high robustness, and auditability, relying solely on vehicle-side signals.

[0140] In one exemplary embodiment, such as Figure 8 As shown, based on the aligned sequence data and the unpredictable control representation data, the cross-domain robust directional mutual information interval data is projected onto the feasible region of the information inequality to obtain the sequence mutual information target interval data, including steps 802 to 810. Wherein:

[0141] Step 802: Based on the aligned sequence data and the unpredictable control representation data, analyze the driver's state-action condition distribution to obtain inequality contraction data.

[0142] Step 804: Based on the aligned sequence data, perform a measurement analysis on the driver's action prediction error to obtain action error measurement data.

[0143] Step 806: Perform constraint mapping between the Fano bound and the Piensk bound of the motion error metric data and mutual information to obtain the consistency constraint data of error mutual information.

[0144] Step 808: Based on the cross-domain robust directional mutual information interval data, inequality shrinkage data, and error mutual information consistency constraint data, perform joint constraint construction analysis on the feasible region of mutual information to obtain the feasible region data of sequence mutual information.

[0145] Step 810: Based on the feasible domain data of sequence mutual information, perform Bregman projection on the cross-domain robust directional mutual information interval data to obtain the target interval data of sequence mutual information.

[0146] Among them, the state-action conditional distribution is a conditional probability model of a driver's actions relative to the environmental state, given past and current information.

[0147] Among them, the inequality shrinkage data are hard shrinkage parameters and applicable range markings given for the range of achievable mutual information based on information theory constraints such as data processing inequalities and strong shrinkage coefficients.

[0148] Among them, motion prediction error is a measure of the deviation generated when a model that relies solely on historical and current information predicts the driver's next action.

[0149] Among them, metric analysis is a systematic evaluation process that involves windowing, domaining, statistics, and diagnosis of target quantities (such as error, mutual information, and stability).

[0150] Among them, the motion error metric data is a time-seriesd error information set consisting of prediction residuals, dispersion, trend change rate and their quality labels.

[0151] Among them, the Fano bound and the Piensk bound are classic information inequalities that relate the reachability prediction error rate to the amount of information, and are used to limit the upper and lower bounds of mutual information.

[0152] Among them, constraint mapping is the process of converting error indices and information inequalities into interval restrictions and penalty weights that can be directly used for solving.

[0153] Among them, the error mutual information consistency constraint data is a set of constraints such as the mutual information reachable interval, penalty coefficient and mask matrix obtained by mapping the error-information inequality.

[0154] The feasible region is the entire area where the mutual information values ​​are allowed to take the condition that both hard and soft constraints are met.

[0155] Among them, joint constraint construction analysis is a modeling step that integrates cross-domain robust results, inequality shrinkage and error-information constraints into a unified feasible domain according to priority.

[0156] Among them, the sequence mutual information feasible domain data is a feasible domain description organized by time window and scenario subdomain, including interval boundaries, constraint satisfaction and consistency markers.

[0157] Among them, the Bregman projection is a projection method that, while maintaining convexity and stability, projects the current mutual information interval into the feasible region to obtain the most compact solution that satisfies all constraints.

[0158] Specifically, based on aligned sequence data and unpredictable control representation data, a state-action condition distribution that "only depends on the present and the past" is constructed and verified according to time windows and scenario subdomains. The construction and verification process first ensures the causal structure is valid by detecting future information leakage and checking edge consistency and direction consistency. Then, within each time window / subdomain, robust statistical sampling and comparative perturbation are used to evaluate the maximum contraction amplitude and effective boundary of mutual information under the causal structure. At the same time, masking and weighting are enabled for missing, abnormal and active security intervention segments and diagnostic information is recorded to obtain inequality contraction data containing contraction coefficient, scope of application, subdomain label, quality level and consistency certificate.

[0159] Based on aligned sequence data, short-term predictions of driver actions are made using only current and historical information under unpredictable constraints, generating residual sequences that are one-to-one aligned with actual actions. The residual sequences are then evaluated hierarchically by fixed sliding windows and scene subdomains, producing robust metrics such as point-by-point error, in-window dispersion, and trend change rate. Masking and weighting are applied to ACC / LKA / AEB interventions, sensor jitter, and missing segments, and short-term jitter is suppressed through hysteresis and cooling time. Error consistency checks and baseline drift monitoring are performed by combining data quality grading and cross-model cross-validation. Breakpoint restarts and conservative rollbacks are executed for abnormal mutations, resulting in action error metrics including error time series, subdomain weights, confidence interval labels, masks, and diagnostic logs.

[0160] The motion error metric data is aligned to the solution grid according to time windows and scene subdomains. Then, based on information inequalities (Feno and Piensk boundaries), the error index of each grid cell is converted into an executable mutual information reachability range and violation penalty weight. Conservative relaxation, masking, and robust reweighting are applied to low-confidence and missing segments using quality labels. The above processing results are then checked for continuity and monotonicity, and smoothing and drift detection are performed by subdomain and adjacent time windows. If a conflict occurs with the inequality shrinkage data output in the first step, out-of-bounds pruning and backoff strategies are triggered, and the diagnosis is recorded. Error mutual information consistency constraint data, arranged by time and subdomain, is generated, containing interval suggestions, penalty coefficients, mask matrices, and consistency markers.

[0161] The cross-domain robust directional mutual information interval data, inequality shrinkage data, and error mutual information consistency constraint data are unified into the same time window and scene subdomain grid. After aligning the caliber and units, constraints are applied layer by layer according to the priority from "hard" to "soft". In the process of applying constraints layer by layer according to priority, hard constraints such as unpredictability and data processing inequalities / strong shrinkage coefficients are first implemented to determine the achievable upper and lower limits. Then, soft constraints and penalty weights derived from action prediction errors are superimposed to tighten the interval. At the same time, chain decomposition consistency, edge matching, and cross-window continuity checks are performed. Conflicting segments are pruned according to preset rules, the trust domain is relaxed or rolled back, and the diagnosis is recorded. Masking and conservative imputation are enabled for low-confidence and missing grids to ensure that the results are both robust and interpretable in time and subdomain, resulting in sequence mutual information feasible domain data arranged by time and subdomain.

[0162] Using the feasible domain data of sequence mutual information as the projection space, the cross-domain robust directional mutual information interval data is taken as the object to be tightened. Bregman projection is used to perform one or more constraint unification and tightening iterations. During the constraint unification and tightening iteration process, the hard constraints are first satisfied by projecting grid by grid according to the time window and scene subdomain. Then, continuity and monotonicity coupling are applied between adjacent time windows and similar subdomains. When necessary, alternating projection and trust domain step size control are used to stabilize convergence. For cases such as boundary reach, interval inversion, and conflict reproduction, out-of-bounds pruning, conservative backoff, and breakpoint restart strategies are used, and quality labels, masks, and diagnostic logs are retained to obtain the sequence mutual information target interval data that is tightened and satisfies all information inequalities and causal consistency requirements, along with convergence status, constraint satisfaction, and version metadata.

[0163] In this embodiment, the empirical mutual information estimation is transformed into a sequence mutual information target interval that is guaranteed by information theory, robust across scenarios, and auditable, through the link of inequality contraction of state-action condition distribution → Feno / Piensk constraint mapping based on action prediction error → joint construction of mutual information feasible region → Bregman projection tightening interval. This enables the provision of highly reliable input for subsequent β inversion and fatigue determination with lower variance and lower false alarms, and requires no additional sensors and is easy to deploy and trace online.

[0164] In one exemplary embodiment, such as Figure 9 As shown, a sequential statistical test analysis is performed on the information efficiency decay fatigue index to obtain the driver's corresponding fatigue state data, including steps 902 to 908. Wherein:

[0165] Step 902: Perform time-varying risk weighting and self-regular standardization on the information efficiency decay fatigue index to obtain standardized fatigue index sequence data.

[0166] Step 904: Perform test martingale construction analysis based on betting test on the exponential increment of the standardized fatigue index sequence data to obtain fatigue evidence e-value sequence data.

[0167] Step 906: Construct the time-consistent confidence sequence boundary of the fatigue evidence e-value sequence data to obtain the ever-effective discrimination boundary data.

[0168] Step 908: Based on the fatigue evidence e-value sequence data and the ever-effective discrimination boundary data, perform dual-threshold irreversible confirmation and change point location analysis on the information efficiency decay fatigue index to obtain the driver's corresponding fatigue state data.

[0169] Among them, time-varying risk weighting assigns weights to the fatigue index according to the changes in scenario risk and data quality over time, so that high-risk periods contribute more to the judgment.

[0170] Among them, self-regularization is an adaptive process that eliminates scale differences, slow drift and heteroscedasticity without relying on external calibration, making the sequences comparable and stable.

[0171] The standardized fatigue index sequence data is a fatigue index time series that has undergone time-varying weighting and self-regularization, and includes timestamps and quality labels.

[0172] Among them, the exponential increment is the change in the standardized fatigue index between adjacent time points, which is used to construct the subsequent evidence stream.

[0173] Among them, the betting test is an evidence accumulation framework that describes sequence monitoring as a "betting" counterhypothesis, and measures the strength of evidence with an interpretable e-value.

[0174] Among them, test martingale construction analysis is a process of transforming exponential increments into multiplicative evidence that satisfies arbitrary stopping time properties, in order to generate a robust e-value sequence.

[0175] Among them, the fatigue evidence e-value sequence data is a time series of e-values ​​that accumulate monotonically over time, representing the strength of evidence supporting the fatigue hypothesis.

[0176] Among them, the time-consistent confidence sequence boundary is a dynamic threshold upper and lower bound that effectively controls the false alarm rate at any observation time.

[0177] Among them, the boundary data that can be effectively identified at any time is the boundary trajectory obtained by instantiating the time-consistent boundary and can be judged by comparing the e-value in real time.

[0178] Among them, the dual-threshold irreversible confirmation first uses a low threshold to trigger an early warning, and then uses a higher threshold to complete the irreversible confirmation in order to reduce false alarms and improve the confidence level.

[0179] Among them, change point location analysis identifies the time and stage of fatigue state abrupt change in time series, and is used to mark the start and duration intervals.

[0180] Specifically, the information efficiency decay fatigue index, risk weights given by task complexity, data quality labels, and vehicle status are unified to the same time base. Then, masks are set for ACC / LKA / AEB interventions, sensor anomalies, and missing segments, and robust weight reduction is performed. After completion, a dynamic baseline is established and drift elimination is performed by sliding window. Anomaly removal, variance stabilization, and scaling are comprehensively performed. The index is weighted and aggregated according to the time-varying risk curve and quality weights, and self-regularization is applied to eliminate the influence of vehicle model differences and dimensions, suppress heteroscedasticity and short-term jitter, and obtain standardized fatigue index sequence data with timestamps, window numbers, risk and quality weights, mask trajectories, and processing logs.

[0181] Based on the incremental standardization of the fatigue index sequence, an online-updating martingale evidence stream is established according to the betting test concept. This generates an additive evidence increment for each time step, sets anomaly protection and trust domain step size, and supports arbitrary pauses and online updates. Specifically, it first adaptively assigns "betting weights" to each time step based on the scenario risk, data quality, and short-term volatility of the evidence increment, and uses masking and robust weight reduction for ACC / LKA / AEB interventions, missing segments, and anomalous segments. Then, a hybrid strategy and trust domain step size control are used to suppress excessive amplification, combined with truncation and numerical normalization to avoid overflow. Breakpoint restarts and hysteresis cooling are provided during cross-scenario switching to stabilize the direction of evidence accumulation, generating a fatigue evidence e-value sequence that monotonically accumulates over time. It also outputs key derived quantities (such as subsequence peaks, pullbacks, and trigger candidate indices) and processing logs.

[0182] Based on fatigue evidence e-value sequence data, a false positive rate budget and error cost strategy are set according to task complexity and data quality, initializing two boundary channels: early-stage sensitive and late-stage robust. Based on these two boundary channels, the boundary slope and stride are adaptively updated using historical fluctuations and current activity levels. Combined with self-regulation and empirical variance-based tightening rules, the boundaries are widened promptly when evidence accumulates rapidly and automatically tightened when noise dominates. Hysteresis and cooldown times are set for short-term jitter and pullbacks, and continuity constraints and breakpoint restarts are implemented for cross-scenario switching to ensure consistent discrimination criteria under different road conditions, lighting, and traffic flow. Masking and conservative expansion are enabled for anomalies and missing windows throughout the process, and boundary versions, budget consumption, trigger candidates, and consistency logs are recorded, resulting in readily available and effective discrimination boundary data that can be directly compared at any time.

[0183] Based on fatigue evidence e-value sequence data and readily available boundary data, an online discrimination process of "early warning → confirmation → tracking / recovery" is constructed. Evidence is compared with two boundaries in real time. When a low threshold is reached, an early warning is issued, and hysteresis and cooling-off periods are initiated to suppress fluctuations. Irreversible confirmation is performed after a high threshold is met and the minimum dwell time is consistently reached. Simultaneously, multi-scale change point localization is performed to mark the initial transition point and subsequent stages of deterioration / relief. Masking and conservative backoff strategies are employed for ACC / LKA / AEB interventions, missing or low-confidence windows. Automatic degradation to safe mode is implemented when boundary budgets are exhausted or consistency is compromised. Causal consistency and data quality verification results are recorded throughout the process. Fatigue state data, including fatigue level, confirmation time and duration, evidence strength, triggering basis, mask trajectory, and audit logs, is output.

[0184] In this embodiment, the fatigue index is robustly standardized by time-varying weighting and self-regularization. The false alarm rate is controllable under arbitrary stop time by using the e-value evidence stream based on betting test and the confidence boundary that is valid at any time. Then, the dual threshold irreversible confirmation + change point location takes into account both early warning and confirmation. At the same time, the ADAS intervention, abnormal and missing segments are masked and robustly reduced in weight. It can obtain earlier, more stable and auditable online fatigue judgment and event location capabilities without increasing the camera / physiological sensing, effectively reducing false alarms / missed alarms and improving cross-scene consistency.

[0185] Based on the same inventive concept, this application also provides a driver fatigue state data determination device for implementing the driver fatigue state data determination method described above. For example... Figure 10 As shown, the device includes: a vehicle data acquisition module 1002, a vehicle risk prediction module 1004, a mutual information analysis module 1006, a fatigue index calculation module 1008, and a fatigue index analysis module 1010. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more driver fatigue state data determination device embodiments provided below can be found in the limitations of a driver fatigue state data determination method described above, and will not be repeated here.

[0186] The various modules in the aforementioned driver fatigue state data determination device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0187] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 11As shown. This computer device includes a processor, memory, input / output interfaces (I / O), and communication interfaces.

[0188] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0189] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0190] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0191] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.

[0192] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0193] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0194] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for determining driver fatigue state data, characterized in that, The method includes: Acquire environmental status data for the running vehicle and driver action data for the driver; Based on the environmental state data, the risk events of the operating vehicle in the future preset time domain are predicted and analyzed to obtain task complexity analysis data. Mutual information analysis is performed on the correspondence between the environmental state data and the driver action data to obtain vehicle occupant mutual information analysis data. Based on the task complexity analysis data and the driver-passenger mutual information analysis data, the driver's fatigue index at the current moment is calculated to obtain the information efficiency decay fatigue index. Sequential statistical tests and analyses were performed on the information efficiency decay fatigue index to obtain the fatigue state data corresponding to the driver.

2. The method according to claim 1, characterized in that, The step of calculating the driver's fatigue index at the current moment based on the task complexity analysis data and the driver-passenger mutual information analysis data, to obtain the information efficiency decay fatigue index, includes: Based on the task complexity analysis data and the driver mutual information analysis data, the information-cost dual model of the driver is solved by inversion to obtain the information cost β data. Based on the information cost β data, the driver's fatigue index at the current moment is calculated to obtain the information efficiency decay fatigue index.

3. The method according to claim 2, characterized in that, The step of inverting and solving the driver's information-cost dual model based on the task complexity analysis data and the driver-passenger mutual information analysis data to obtain information cost β data includes: Based on the task complexity analysis data, the driver's unreferenced action distribution and driving risk potential are analyzed to obtain path prior data. The mutual information constraints of the vehicle and passenger mutual information analysis data are analyzed to obtain path mutual information constraint data; Based on the path prior data and the path mutual information constraint data, Schrödinger bridge entropy regular iterative analysis is performed on the driver's state action path probability flow to obtain relative entropy control strategy flow data; Based on the relative entropy control strategy flow data, the one-dimensional monotonic root of the mutual information-temperature monotonic operator for the driver is solved to obtain the information cost β data.

4. The method according to claim 3, characterized in that, The step of performing Schrödinger bridge entropy regularized iterative analysis on the driver's state-action path probability flow based on the path prior data and path mutual information constraint data to obtain relative entropy control strategy flow data includes: Based on the path prior data, the driver's transition kernel under the state action path is constructed to obtain uncontrolled path kernel data; Based on the path mutual information constraint data, the mutual information confidence interval of the vehicle driver mutual information analysis data is dualized and modeled to obtain the mutual information Lagrange variables and the initial value data of the dual potential. Based on the uncontrolled path kernel data, the mutual information Lagrange variables, and the initial value data of the dual potential, a forward-backward scaling iterative analysis of the state-action path probability flow is performed using Schrödinger bridge entropy regularization to obtain policy path distribution sequence data. Based on the strategy path distribution sequence data, the strategy density of the driver under causal coupling constraints is analyzed to obtain the relative entropy control strategy flow data.

5. The method according to claim 1, characterized in that, The mutual information analysis of the correspondence between the environmental state data and the driver action data is performed to obtain vehicle occupant mutual information analysis data, including: Unpredictable cropping is performed on the environmental state data and the driver action data to obtain unpredictable control cropping data; Information bottleneck control representation modeling is performed on the unpredictable control pruning data to obtain aligned sequence data and unpredictable control representation data; Based on the aligned sequence data and the unpredictable control representation data, cross-domain robust directional variational mutual information analysis is performed on the mutual information between the environmental state data and the driver action data to obtain the sequence mutual information target interval data. Based on the target interval data of the sequence mutual information, conformal confidence calibration analysis is performed on the components of the mutual information to obtain the vehicle crew mutual information analysis data.

6. The method according to claim 5, characterized in that, The step involves performing cross-domain robust directional variational mutual information analysis on the mutual information between the environmental state data and the driver's action data, based on the aligned sequence data and the unpredictable control representation data, to obtain the sequence mutual information target interval data, including: Based on the aligned sequence data and the unpredictable control representation data, directional variational upper and lower bound analysis is performed on the mutual information to obtain directional variational mutual information upper and lower bound data. Cross-domain robust directional mutual information interval data are obtained by performing cross-domain directional variational mutual information upper and lower bound data on the data in different environmental subdomains. Based on the aligned sequence data and the unpredictable control representation data, the cross-domain robust directional mutual information interval data is projected onto the feasible region of the information inequality to obtain the sequence mutual information target interval data.

7. The method according to claim 6, characterized in that, The step of projecting the cross-domain robust directional mutual information interval data onto the information inequality feasible region based on the aligned sequence data and the unpredictable control representation data to obtain the sequence mutual information target interval data includes: Based on the aligned sequence data and the unpredictable control representation data, the state-action condition distribution of the driver is analyzed to obtain inequality contraction data; Based on the alignment sequence data, the driver's motion prediction error is measured and analyzed to obtain motion error measurement data; Constraint mapping is performed on the Fano bound and Piensk bound between the action error measurement data and the mutual information to obtain error mutual information consistency constraint data; Based on the cross-domain robust directional mutual information interval data, the inequality shrinkage data, and the error mutual information consistency constraint data, a joint constraint construction analysis is performed on the feasible region of the mutual information to obtain the sequence mutual information feasible region data. Based on the feasible domain data of the sequence mutual information, the cross-domain robust directional mutual information interval data is subjected to Bregman projection to obtain the target interval data of the sequence mutual information.

8. The method according to claim 1, characterized in that, The sequential statistical test analysis of the information efficiency decay fatigue index yields the fatigue state data corresponding to the driver, including: The information efficiency decay fatigue index is subjected to time-varying risk weighting and self-regularization to obtain standardized fatigue index sequence data. The test martingale construction analysis based on betting test is performed on the exponential increment of the standardized fatigue index sequence data to obtain fatigue evidence e-value sequence data; Construct the time-consistent confidence sequence boundary of the fatigue evidence e-value sequence data to obtain the ever-effective discrimination boundary data; Based on the fatigue evidence e-value sequence data and the real-time effective discrimination boundary data, the information efficiency decay fatigue index is subjected to dual-threshold irreversible confirmation and change point location analysis to obtain the fatigue state data corresponding to the driver.

9. A device for determining driver fatigue status data, characterized in that, The device includes: The vehicle data acquisition module is used to acquire environmental status data corresponding to the running vehicle and driver action data corresponding to the driver. The vehicle risk prediction module is used to predict and analyze risk events of the operating vehicle in a future preset time domain based on the environmental state data, and obtain task complexity analysis data. The mutual information analysis module is used to perform mutual information analysis on the correspondence between the environmental state data and the driver action data to obtain vehicle occupant mutual information analysis data. The fatigue index calculation module is used to calculate the driver's fatigue index at the current moment based on the task complexity analysis data and the driver-passenger mutual information analysis data, and obtain the information efficiency decay fatigue index. The fatigue index analysis module is used to perform sequential statistical test analysis on the information efficiency decay fatigue index to obtain the fatigue state data corresponding to the driver.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images