Fatigue alarm method fusing physiological signals of driver and flight state data of unmanned aerial vehicle

Abstract

The invention provides a fatigue alarm method fusing physiological signals of a driver and flight state data of an unmanned aerial vehicle, which comprises the following steps: synchronously acquiring physiological signals such as heart rate variability and electroencephalogram alpha waves and flight state characteristics such as control frequency and communication interruption through wearable equipment and a flight control system, and generating a multi-modal fusion sample by means of data alignment and cleaning; according to the method, physical complexity, decision pressure and time urgency are modeled by using a comprehensive load index, self-adaptive adjustment and optimization of parameters of a meta-learning driving classification model are combined with a driver historical response mode, individualized dynamic judgment of a fatigue state is realized, physiological behavior changes of a driver under different task loads can be accurately reflected, and the driver experience is improved. The real-time performance, the sensitivity and the individuation level of fatigue detection are improved, and the method has important significance for guaranteeing the operation safety of the unmanned aerial vehicle in a complex operation scene.

Description

Technical Field

[0001] This invention relates to the field of intelligent human-computer interaction and flight mission status perception technology, and in particular to a fatigue alarm method that integrates pilot physiological signals and UAV flight status data. Background Technology

[0002] Currently, drone pilot fatigue monitoring and alerting technologies mainly revolve around physiological signal analysis and operational behavior feature extraction, using multimodal fusion models to automatically identify and classify pilot fatigue states. Mainstream solutions generally employ statically parameterized fatigue judgment models, meaning that during system deployment, discrimination thresholds and key feature weights are set according to predetermined task scenarios and the general physiological response patterns of the pilot group. For example, some technologies collect statistical features of heart rate variability, eye movement frequency, and electroencephalogram (EEG) signals, combined with flight operation frequency and behavioral event counts, to establish fixed fusion models based on support vector machines, random forests, or multilayer neural networks, thereby achieving preliminary classification of low, medium, and high fatigue levels. These solutions possess a certain level of recognition accuracy under standardized, highly consistent tasks and have been practically applied in fields such as intelligent human-machine interaction and flight safety monitoring.

[0003] In recent years, with the rapid increase in the complexity and diversity of UAV missions, the technological development trend has gradually shifted towards multi-level fatigue assessment systems that integrate dynamic changes in mission load and individual differences in response. Some studies have quantified and modeled factors such as mission type (e.g., reconnaissance, strike, formation), mission urgency, and operational density, and attempted to introduce these as auxiliary criteria into the fatigue identification process. Many publicly available technical solutions still rely on static modeling, making it difficult to adjust model parameters in real time according to mission status and load changes. Therefore, in complex scenarios, problems such as false fatigue assessments, missed assessments, or decreased adaptability often occur. Existing technologies typically employ the following approach: pre-setting fatigue judgment logic rules, statically configuring abnormal physiological signal thresholds and operational behavior weights. Some solutions support simple compensation based on historical behavior patterns, but the overall framework lacks task load sensitivity and online parameter tuning capabilities. These technologies are suitable for standardized flight scenarios and collective pilot average models, but they struggle to meet the demands for dynamic identification of individual fatigue under varying task conditions. When UAVs perform high-intensity or time-sensitive emergency missions, the pilot's physiological and behavioral response patterns may deviate significantly from the basic model. Fixed threshold judgment mechanisms cannot promptly capture risk states, impacting the robustness and safety of the monitoring system. Summary of the Invention

[0004] In order to solve the above-mentioned technical problems, the present invention provides a fatigue alarm method that integrates pilot physiological signals and UAV flight status data.

[0005] The technical solution of this invention is implemented as follows: a fatigue alarm method that integrates pilot physiological signals and UAV flight status data, comprising: S1: Acquire the pilot's physiological signal data and the UAV's flight status data. The physiological signal data includes heart rate variability, eye movement frequency, and brainwave alpha power. The flight status data includes control input frequency, number of route changes, number of target locks, and communication interruption frequency. All data are aligned and marked with a unified timestamp. S2: Based on the preset task type knowledge base and flight status data, the comprehensive load index of the current task is calculated by using a rule engine combined with a lightweight neural network model. This index reflects the superimposed effect of the task on three dimensions: physical complexity, decision pressure and time urgency, and is output as an adjustment factor for dynamic optimization. S3: Input the comprehensive load index into the parameter tuning unit of the fatigue judgment model. This tuning unit is constructed based on the model-independent fine-tuning idea under the meta-learning framework. It obtains the basic parameter space adaptability through historical multi-task-fatigue sample training and generates the corresponding discrimination boundary adjustment vector according to the currently input comprehensive load index. S4: Based on the discrimination boundary adjustment vector, the classification threshold and feature weight distribution of the original fatigue classifier are dynamically corrected to form a real-time fatigue judgment sub-model that adapts to the current task load conditions. Under high load conditions, the tolerance for physiological signal variation is reduced and the contribution ratio of behavioral features is increased. S5: Input the aligned physiological signal data and flight status data into the real-time fatigue judgment sub-model, perform feature-level fusion analysis, and output the preliminary fatigue score at the current moment. This score is a continuous numerical index that represents the quantitative result of the driver's fatigue level. S6: Call the driver's individual historical response pattern memory bank, retrieve the average value of the driver's past physiological-behavioral response characteristics under the same or similar comprehensive load index, calculate the deviation value between the current preliminary fatigue score and the historical benchmark, and generate a personalized deviation compensation vector. S7: The personalized deviation compensation vector is weighted and fused with the preliminary fatigue score to output the final fatigue level judgment result after individual adaptation correction. The result is divided into three discrete states: low, medium and high, which are used for subsequent alarm strategy matching. S8: Determine whether the final fatigue level determination result reaches the preset high fatigue alarm threshold. If it does, trigger the multimodal alarm mechanism and record the context information of this determination to the system log for subsequent model iteration optimization and abnormal scenario tracing.

[0006] The fatigue alarm method that integrates pilot physiological signals and UAV flight status data provided by this invention has the following beneficial effects: (1) This invention effectively overcomes the technical defects of traditional UAV pilot fatigue monitoring systems, which rely on static thresholds, fixed weight fusion, and general discrimination models, resulting in poor adaptability and high misjudgment rate, by introducing a task load modeling submodule and a context-aware dynamic parameter tuning mechanism. It obtains multi-dimensional operational data such as flight trajectory, control frequency, route change, target lock, and communication interruption from the flight state perception module, and combines it with a task type knowledge base. It uses a rule engine and a lightweight neural network to collaboratively calculate a comprehensive load index, which comprehensively describes the time pressure, decision density, and execution complexity of the task. This index is used as a key context input to the fatigue judgment model based on the meta-learning framework, triggering real-time adjustment of its internal discrimination boundary. For example, in high-load situations, it automatically reduces the tolerance threshold for heart rate variability and increases the weight of behavioral indicators such as eye movement lag, thereby achieving more sensitive and accurate identification of fatigue state. Compared with the response delay of hundreds of milliseconds or even seconds in traditional methods, this solution can significantly advance the fatigue risk reassessment after a sudden change in the task, greatly improving the dynamic response capability and context fit of the monitoring system. (2) This solution innovatively constructs a driver historical response pattern memory mechanism, realizing a leap from "group common modeling" to "personalized deviation compensation". During the online operation of this invention, the physiological-behavioral response feature sequence of the same driver under similar task loads is continuously accumulated to form a personalized response baseline template library, and a deviation compensation vector is generated based on this to correct the current fatigue score output. This mechanism effectively alleviates the model generalization difficulty caused by individual differences (such as stress tolerance, operating habits, and biological rhythms), and significantly improves the judgment stability and consistency across time periods and task scenarios without the need for additional calibration steps. At the same time, combined with the MAML idea in meta-learning, the basic fatigue classifier has a fast transfer capability and can quickly complete the initialization adaptation on a small amount of new user data, avoiding the disadvantages of long training cycles and high resource consumption of traditional deep models. It is particularly suitable for combat-oriented UAV operation environments that require rapid deployment and multi-machine collaboration. Attached Figure Description

[0007] Figure 1 This is a flowchart of the fatigue alarm method that integrates pilot physiological signals and UAV flight status data according to the present invention; Figure 2 This is a sub-flowchart of the fatigue alarm method that integrates pilot physiological signals and UAV flight status data according to the present invention. Figure 3 This is another sub-flowchart of the fatigue alarm method of the present invention, which integrates the pilot's physiological signals and the UAV flight status data. Detailed Implementation

[0008] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0009] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0010] like Figure 1 As shown, this invention provides a fatigue alarm method that integrates pilot physiological signals and UAV flight status data, specifically including: S1: Acquire the pilot's physiological signal data and the UAV's flight status data. The physiological signal data includes heart rate variability, eye movement frequency, and brainwave alpha power. The flight status data includes control input frequency, number of route changes, number of target locks, and communication interruption frequency. All data are aligned and marked with a unified timestamp. S2: Based on the preset task type knowledge base and flight status data, the comprehensive load index of the current task is calculated by using a rule engine combined with a lightweight neural network model. This index reflects the superimposed effect of the task on three dimensions: physical complexity, decision pressure and time urgency, and is output as an adjustment factor for dynamic optimization. S3: Input the comprehensive load index into the parameter tuning unit of the fatigue judgment model. This tuning unit is constructed based on the model-independent fine-tuning idea under the meta-learning framework. It obtains the basic parameter space adaptability through historical multi-task-fatigue sample training and generates the corresponding discrimination boundary adjustment vector according to the currently input comprehensive load index. S4: Based on the discrimination boundary adjustment vector, the classification threshold and feature weight distribution of the original fatigue classifier are dynamically corrected to form a real-time fatigue judgment sub-model that adapts to the current task load conditions. Under high load conditions, the tolerance for physiological signal variation is reduced and the contribution ratio of behavioral features is increased. S5: Input the aligned physiological signal data and flight status data into the real-time fatigue judgment sub-model, perform feature-level fusion analysis, and output the preliminary fatigue score at the current moment. This score is a continuous numerical index that represents the quantitative result of the driver's fatigue level. S6: Call the driver's individual historical response pattern memory bank, retrieve the average value of the driver's past physiological-behavioral response characteristics under the same or similar comprehensive load index, calculate the deviation value between the current preliminary fatigue score and the historical benchmark, and generate a personalized deviation compensation vector. S7: The personalized deviation compensation vector is weighted and fused with the preliminary fatigue score to output the final fatigue level judgment result after individual adaptation correction. The result is divided into three discrete states: low, medium and high, which are used for subsequent alarm strategy matching. S8: Determine whether the final fatigue level determination result reaches the preset high fatigue alarm threshold. If it does, trigger the multimodal alarm mechanism and record the context information of this determination to the system log for subsequent model iteration optimization and abnormal scenario tracing.

[0011] Step S1: Acquire the pilot's physiological signal data and the UAV's flight status data. The physiological signal data includes heart rate variability, eye movement frequency, and EEG alpha wave power. The flight status data includes control input frequency, number of route changes, number of target locks, and communication interruption frequency. All data is aligned and marked with a unified timestamp. Specifically, this includes: S1.1: Based on the raw bioelectrical signals of the driver collected by wearable physiological sensing devices, the electrocardiogram (ECG), electroencephalogram (EEG), and eye movement (EOG) signals are preprocessed using bandpass filtering and baseline drift correction algorithms to extract effective heart rate variability (HRV), EEG alpha power, and eye movement frequency (Blink Rate) characteristic parameter sequences as input conditions for the quantitative analysis of physiological state. The input conditions are raw bioelectric signals from wearable physiological sensing devices, including electrocardiogram (ECG), electroencephalogram (EEG), and eye movement (EOG) signals. The signal sampling rate is maintained at millisecond-level accuracy, and the sampling frequency is specifically set to 1-500Hz, preferably 128-256Hz. A bandpass filtering method was used (parameters: ECG signal passband 0.5–40Hz, EEG signal passband 8–13Hz, EOG signal passband 0.1–15Hz) to achieve interference suppression and spectrum range limitation for ECG, EEG and eye movement signals respectively. Furthermore, by using a baseline drift correction algorithm (parameters: polynomial fitting order of 3, sliding window length of 10s), low-frequency trends in ECG, EEG, and EOG signals are removed, and pure physiological waveform data without background drift are obtained. Furthermore, a peak detection and time-domain interval calculation method is adopted (parameter: ECG signal R wave detection threshold is 0.6mV) to achieve accurate extraction of heart rate variability sequence and generate RR interval time series in milliseconds; Furthermore, using the power spectral density estimation method (parameters: Welch method, window length 2s, 50% overlap), energy integration was performed on the alpha wave band (8–13Hz) of the EEG signal to calculate the EEG alpha wave power index. And as a quantitative characteristic reflecting the trend of central fatigue; Furthermore, an eye-tracking waveform zero-crossing and pulse width combination detection algorithm (parameters: amplitude threshold 120μV, minimum pulse width 50ms) is used to calculate the eye-tracking frequency and obtain the Blink Rate feature value expressed in minutes. By using feature encapsulation processing, the results of the previous step are transformed into a sequence of feature parameters for heart rate variability (HRV), brainwave alpha power, and eye movement frequency (Blink Rate), thereby achieving a unified and formatted input condition for quantitative analysis of physiological state. For example, in a reconnaissance mission scenario, the driver's ECG sensor sampling rate is configured to 500Hz, the EEG sampling rate to 256Hz, and the EOG sampling rate to 128Hz. The ECG passband filter cutoff frequencies are set to 0.5Hz and 40Hz, the EEG filter to 8Hz and 13Hz, and the EOG filter to 0.1Hz and 15Hz. After baseline drift correction, the ECG signal drift amplitude is reduced to within 0.05mV. The RR interval calculation result is: average RR interval. ms, standard deviation The calculated power of the EEG α wave is ms. microvolts squared, eye movement frequency is The frequency is calculated per minute. In this task scenario, the HRV sequence, alpha wave power index, and Blink Rate feature value are all encapsulated into a three-channel feature matrix with a unified timestamp, providing high-precision physiological state input for subsequent task load index calculation and fatigue judgment model dynamic optimization. S1.2: Real-time acquisition of flight operation log data through the UAV flight control system bus interface; analysis of key operation behavior records such as control input frequency, number of route changes, number of target locks, and communication interruption frequency based on event recognition algorithm; generation of structured flight status event stream as basic input data for task complexity modeling. The input conditions include the bus data interface output from the UAV flight control system. The raw data consists of various operation log entries generated during the flight mission, covering indicators such as control input frequency, number of route changes, number of targets locked, and communication interruption frequency. The timestamp accuracy must meet the millisecond requirement. The flight control bus data monitoring method (parameters: interface address, data packet parsing protocol, refresh period set to 10ms~100ms, preferably 10ms) is adopted to achieve real-time access to the raw data stream of the flight operation log; Furthermore, through a protocol parsing algorithm (parameters: field mapping table, event code definition), the operation codes and status codes in the original data packets are decomposed to obtain the separated basic behavior record set; Furthermore, an event recognition algorithm (parameters: sliding window length 60s, trigger condition threshold set) is used to filter and label events with operational meaning in the basic behavior record set, and generate a list of event types; the event types include control actions, route adjustments, target locking operations, and communication link status changes; Furthermore, by using a frequency statistics algorithm (parameters: statistics window 60s, frequency transformation function), the calculation of the input frequency is realized, and the numerical result of the index within the current statistics window is generated; Furthermore, by using the difference comparison method (parameters: continuous waypoint sequence, segment change detection threshold), the number of route changes is identified, and the corresponding segment change count value is generated; Furthermore, by using the target locking event parser (parameters: lock initiation signal mode, switching judgment rule), the cumulative statistics of the number of locked targets are realized, and the lock count index within the current task cycle is obtained; Furthermore, a communication interruption monitoring algorithm (parameters: link status sampling frequency 10Hz, link breakage determination delay 500ms) is applied to measure the frequency of communication interruptions and generate a set of interruption event identifiers associated with timestamps. By using a structured encapsulation process, the four types of indicators identified in the previous step—control input frequency, number of route changes, number of target locks, and frequency of communication interruptions—are transformed into a time-series flight status event stream, thereby providing high-precision and traceable raw input data for the mission complexity modeling module. For example, in a nighttime reconnaissance mission, the flight control bus interface was set to a listening port UDP 14550, using the MAVLink protocol for packet parsing with a refresh cycle of 10ms. The event recognition algorithm had a sliding window length of 60s, a control input frequency trigger threshold of 50 times per minute, and a route change determination threshold of more than 5% of the waypoint segments changing. The target lock event parser used radar lock signal sequence pattern matching, and the communication interruption monitoring algorithm defined a continuous link disconnection for more than 500ms as an interruption. During the parsing process, the frequency statistics algorithm calculated the control input frequency to be 62 times per minute within a continuous 60s window, the difference comparison method identified 3 route changes, 4 target locks, and 2 communication interruptions. After structured encapsulation, the output was an event flow matrix containing four types of indicators and their unified timestamps, which served as the input for the subsequent task load index calculation unit. This significantly improved the stability and accuracy of event parsing during the task complexity modeling process. S1.3: Deploy a unified clock synchronization mechanism within the edge computing unit. Use the Network Time Protocol (NTP) combined with a hardware pulse alignment method to mark the data streams from the physiological acquisition module and the flight status perception module with a unified timestamp at the millisecond level. This will eliminate the time misalignment problem caused by cross-device acquisition delay and obtain a set of multi-source data frames after time alignment. S1.4: Perform missing value detection and outlier filtering on the time-aligned multi-source data frames. Use linear interpolation and isolated forest algorithms to fill in transient signal loss and remove outliers that deviate significantly from the normal range, so as to improve data integrity and reliability and output the cleaned and standardized physiological-flight joint data sequence. S1.5: The cleaned physiological signal feature parameters and flight status event stream are encapsulated into feature vectors matched by timestamps to generate multi-dimensional fusion data samples containing heart rate variability, eye movement frequency, EEG alpha wave power, control input frequency, number of route changes, number of target locks, and communication interruption frequency. These samples serve as the input basis for the next stage of task load index calculation and fatigue model dynamic optimization.

[0012] Step S2: Based on a preset task type knowledge base and flight status data, a rule engine combined with a lightweight neural network model is used to calculate the comprehensive load index of the current task. This index reflects the cumulative effect of the task in three dimensions: physical complexity, decision-making pressure, and time urgency, and is output as an adjustment factor for dynamic optimization. Specifically, it includes: S2.1: Acquire UAV flight status data, including control input frequency, number of route changes, number of target locks, and communication interruption frequency, and align the markers based on a unified timestamp to form a structured task behavior feature sequence, which serves as the raw input data for task load modeling; S2.2: Based on the preset task type knowledge base, perform pattern matching and semantic parsing on the structured task behavior feature sequence to identify the type label of the currently executed task. The task types include reconnaissance, strike, formation flight, etc. Each type of task is associated with a predefined complexity baseline weight vector, which is used to initialize the task load calculation framework. The input is a structured sequence of task behavior features formed by step S2.1. This sequence covers the frequency of control inputs, the number of route changes, the number of targets locked, and the frequency of communication interruptions, and is time-aligned according to a unified timestamp. A pattern matching algorithm based on hash index (parameter: pre-stored operation feature template vector set in task type knowledge base) is used to realize the similarity calculation function between feature sequence and template of each task type; Furthermore, the input features and template features are numerically calculated using the cosine similarity formula, as follows:

[0013] in, For the input feature vector, the first dimensional components, For the template feature vector, the first dimensional components, This represents the similarity value. Furthermore, a threshold discrimination method (parameter: similarity judgment threshold for each task type) is used to initially screen task types, and candidate types are stored in the task type candidate set; Furthermore, a semantic parsing algorithm based on multilayer perceptron (MLP) is adopted (parameters: input is template features and structured task behavior features in the candidate set; the number of hidden layer nodes is set according to the number of task types) to achieve deep semantic matching of candidate types in multiple feature dimensions and generate task type confidence vectors; Furthermore, the task type label with the highest confidence is output through the Top-1 selection mechanism (parameter: confidence threshold); Furthermore, the task type knowledge base is invoked to index the complexity baseline weight vector corresponding to the tag, and this vector is used as the parameter baseline for initializing the task load calculation framework. By using a chain of pattern matching and semantic parsing algorithms, the structured task behavior feature sequence from the previous step is transformed into a unique task type label and its associated complexity baseline weight, thus achieving the expected technical effect of task type identification and computational framework initialization. For example, in a reconnaissance flight mission, the structured mission behavior feature sequence is as follows: control input frequency 15 times / minute, route change frequency 3 times / mission, target lock frequency 2 times, and communication interruption frequency 1 time. The pre-stored reconnaissance mission template feature vector in the mission type knowledge base is [14, 3, 2, 1], the strike mission template feature vector is [20, 5, 4, 0], and the formation mission template feature vector is [10, 2, 0, 2]. Cosine similarity is used to calculate the similarity between the input feature vector X=[15,3,2,1] and the reconnaissance mission template Y=[14,3,2,1]. The calculation result is significantly higher than other mission templates, indicating that the current mission type is a reconnaissance mission. The knowledge base is called to return the reconnaissance mission complexity baseline weight vector [0.6, 0.2, 0.15, 0.05], which serves as the initial weight baseline for subsequent mission load calculations. The task type labels and weight baselines output by this process significantly improve the accuracy of the physical complexity sub-index and the contextual adaptability of the task load index generation in subsequent calculations. S2.3: Based on the identified task type labels and their corresponding complexity baseline weight vectors, a weighted summation algorithm is used to perform primary load component fusion on the frequency of control inputs, the number of route changes, and the number of targets locked, generating a task physical complexity sub-index. This sub-index represents the operational burden borne by the driver in terms of spatial operation and action density. S2.4: Using the frequency of communication interruptions and the degree of disruption of temporal continuity as input, the time distribution characteristics of sudden interference events are extracted through a lightweight long short-term memory network (LSTM) model. Combined with the inverse ratio factor of the remaining task time, the task time urgency sub-index is calculated. This sub-index reflects the intensity of psychological pressure brought about by the accelerated pace of information processing under the task deadline constraint. S2.5: Based on the rule engine, the task physical complexity sub-index and the task time urgency sub-index are integrated, and the target locking switching frequency is introduced as the decision pressure modulation term. The three-dimensional coupling operation is performed through the nonlinear activation function to output the normalized comprehensive load index. This index serves as the external adjustment factor of the parameter tuning unit of the subsequent fatigue judgment model, driving the dynamic reconstruction of the classification boundary.

[0014] like Figure 2 As shown, step S3 involves inputting the comprehensive load index into the parameter tuning unit of the fatigue judgment model. This tuning unit is constructed based on the model-independent fine-tuning idea under the meta-learning framework. It obtains the basic parameter space adaptability through historical multi-task fatigue sample training and generates the corresponding discrimination boundary adjustment vector based on the currently input comprehensive load index. Specifically, this includes: S3.1: Based on the historically collected multi-task fatigue-labeled sample dataset, a basic training set is constructed for training the meta-learning parameter tuning unit. Each sample contains the comprehensive load index under the corresponding task scenario, the synchronously collected physiological signal feature vector and behavioral state feature vector, and the fatigue level label calibrated by experts, so as to support the model to learn the cross-task generalization mapping relationship between task load and fatigue response. Based on the historical multi-task fatigue-labeled sample dataset, the data preprocessing module is called to load the raw collected data covering multi-task scenarios. The data includes task metadata, comprehensive load index, synchronously collected physiological signal feature vectors and behavioral state feature vectors, as well as fatigue level labels calibrated by experts, to achieve raw data integration. A time series slicing algorithm (parameters: uniform time window length, sliding step size) is used to perform synchronous segmentation on the physiological signal channel and the behavioral state channel to ensure that the feature vectors of different modalities are aligned within the same time frame and to eliminate time offset in task records. Furthermore, by using the feature normalization method (parameters: Z-score standardization, outlier limiting threshold), the various feature vectors are transformed into dimensionless forms according to their dimensions, and the isolated forest algorithm is used to remove significant outlier feature points, thereby improving the robustness of the training sample set. Furthermore, the feature selection algorithm is invoked (parameters: maximum correlation minimum redundancy mRMR, correlation threshold set to 0.6) to filter out features that are weakly correlated with fatigue level labels or have high redundancy, so that the retained feature subset can maintain high discriminative power in cross-task scenarios. The label encoding method is used to convert the discrete fatigue level labels identified by experts into numerical classification labels, forming a standardized training label matrix. A task cluster index is constructed based on the meta-information of multi-task scenarios for inter-task sampling in subsequent meta-learning algorithms. Through the above processing method, the original sampling data of the previous step is transformed into a basic training set with unified structure, carefully selected features, and label encoding, so as to realize the input preparation for the cross-task generalization mapping capability of the meta-learning parameter tuning unit. For example, a historical fatigue sample set containing 120 task scenarios is loaded. For each task scenario, the following data are stored: comprehensive load index (range 0.25~0.88), heart rate variability (ms), eye movement frequency (Hz), EEG alpha wave power (μV²), and behavioral characteristics such as control input frequency and number of flight path changes. Fatigue level labels are determined by three senior flight experts, with the majority consensus value being the final label. A time window length of 5000ms and a sliding step of 1000ms are set, and the feature vector sequence is sliced ​​according to the window to generate time-series segments. Z-score standardization is performed on the features of each channel, calculated using the following formula:

[0015] in, These are the original eigenvalues. This is the mean of the feature within the current task cluster. The standard deviation is set, and the outlier limit interval is set to [-3, 3]. The mRMR algorithm iteratively selects the optimal 30-dimensional feature subset, encoding the fatigue level labels "low, medium, high" as values ​​0, 1, and 2 to form a label matrix. The sample set is then divided into three task clusters: reconnaissance, strike, and formation, forming a training benchmark set for task-level sampling by the MAML algorithm. Under this configuration, the generated training set maintains consistent feature distribution across tasks and exhibits high label mapping stability, significantly enhancing the cross-task adaptability of the meta-learning parameter tuning unit. S3.2: Using the Model Independent Fine-Tuning (MAML) algorithm framework, meta-training is performed on the initial fatigue judgment parameter tuning network. By performing inner loop fast adaptation and outer loop gradient update on multiple task clusters, the initial parameters of the network are optimized, enabling it to have the meta-knowledge representation ability to efficiently adjust the discrimination boundary under new task load conditions with only a few gradient steps. The input condition is the historical multi-task fatigue labeled sample dataset constructed in step S3.1. The samples contain comprehensive load index, synchronous physiological signal feature vector, behavioral state feature vector and corresponding fatigue level label, which serve as the input set for the initial training of the fatigue judgment parameter tuning network. The Model Independent Fine-Tuning (MAML) algorithm framework (parameters: task batch size B, inner loop learning rate α, outer loop learning rate β) is adopted to perform a meta-training process on the network to fine-tune the initial fatigue judgment parameters, thereby optimizing the cross-task generalization ability. Furthermore, through the rapid adaptation phase of the inner loop, for each task cluster in the task batch, the network parameters are updated on a small number of samples using the gradient descent method, the comprehensive load index and fatigue response data are mapped to the preliminary estimate of the discrimination boundary adjustment vector, and the updated task-specific parameters are retained. Furthermore, in the outer loop gradient update phase, based on the retained task-specific parameters, the gradients of the loss function for all task clusters are accumulated and summed, and the update direction of the global initial parameters is calculated using the following formula:

[0016] in, This is the global initial parameter vector. The learning rate for the outer loop. For task batch size, For task cluster indexing, Let i be the loss function for the i-th task cluster. This is the parameter vector updated by the inner loop; Furthermore, through multiple iterations, combined with the distribution characteristics of the comprehensive load index in different task clusters, the initial parameter space is gradually adjusted so that the network can complete efficient adjustments with only a few gradient steps when encountering new task load conditions; By using the inner and outer loop parameter optimization processing method executed by the MAML algorithm, the cross-task feature mapping capability of the samples in the training phase is transformed into a global meta-knowledge representation, thereby enabling the fatigue judgment model to quickly adapt to new task load conditions and optimize the discrimination boundary. For example, in a UAV flight mission scenario, the training set contains B=4 mission clusters, corresponding to reconnaissance, strike, formation flight, and emergency avoidance missions, respectively. 20 sample data points are randomly selected from each cluster. The inner loop learning rate α is set to 0.01, and the outer loop learning rate β is set to 0.001. In the inner loop update, each mission cluster calculates a loss function based on the comprehensive load index and the corresponding fatigue level label, and updates the parameters using gradient descent. For example, in the reconnaissance mission cluster, the comprehensive load index ranges from 0.35 to 0.45, and the corresponding fatigue level labels are mainly distributed in the low and medium ranges. The mission-specific parameters for this cluster are obtained through a single parameter update. In the outer loop, the loss gradients are calculated for the updated parameters of the four task clusters, and the averages are accumulated. The global initial parameters are then updated using the formula described above. After 50 rounds of training, the network can complete the discrimination boundary adjustment within 3 gradient updates in the case of a sudden communication interruption task with new input, and the accuracy of the judgment output is significantly improved, meeting the fatigue recognition requirements under dynamic task loads; S3.3: Deploy the trained meta-learning parameter tuning unit to the fusion analysis module and connect its input interface to the output of the task load modeling submodule to ensure that it can receive the comprehensive load index jointly calculated by the rule engine and the lightweight neural network as a conditional input in real time; S3.4: The current moment comprehensive load index is embedded and encoded, and it is mapped into a load representation vector in a high-dimensional semantic space using a preset nonlinear transformation function. This vector serves as the key context information for generating personalized discrimination boundary adjustment instructions in the driving parameter tuning unit. S3.5: Based on the load representation vector, forward inference is performed within the meta-learning parameter tuning unit to generate a set of discrimination boundary adjustment vectors for the original fatigue classifier. These vectors include classification threshold offset, feature weight redistribution coefficients, and activation function slope adjustment factors, which are used to guide the correction of structured parameters of the fatigue judgment model in subsequent steps.

[0017] like Figure 3As shown, step S4 involves dynamically adjusting the classification threshold and feature weight distribution of the original fatigue classifier based on the discrimination boundary adjustment vector, forming a real-time fatigue judgment sub-model adapted to the current task load conditions. Under high load conditions, this sub-model reduces the tolerance for physiological signal variations and increases the contribution ratio of behavioral features. Specifically, this includes: S4.1: Obtain the discrimination boundary adjustment vector output by the parameter tuning unit. This vector is a multi-dimensional numerical tensor, and each dimension corresponds to the adjustment direction and magnitude of a specific discrimination parameter in the original fatigue classifier. Based on this vector, perform parameter mapping parsing operation, and use a predefined parameter index mapping table to associate each dimension component with five adjustable parameter sets: classification threshold offset, heart rate variability feature weight gain coefficient, eye movement frequency sensitivity adjustment factor, EEG alpha power normalization baseline bias term, and control input frequency fusion weight ratio, so as to establish an explicit correspondence between the adjustment vector and the model structure parameters. The input condition is the discrimination boundary adjustment vector output by the parameter tuning unit in sub-step S3.5. This vector structure is a multi-dimensional numerical tensor, which contains component values ​​representing the adjustment direction and magnitude of each key discrimination parameter of the original fatigue classifier. The tensor channel parsing method (parameters: the discrimination boundary adjustment vector is N-dimensional, N≥5 and each component corresponds to a single discrimination parameter) is adopted to realize the independent reading and type identification of each component of the adjustment vector, and the tensor is decomposed into a set of parameter elements with clear engineering semantics; Furthermore, through a predefined parameter index mapping table (parameter: the number of index entries is equal to and corresponds one-to-one with the types of discrimination parameters), the association matching between the discrimination boundary adjustment vector components and the adjustable parameter set inside the fatigue classifier is realized, and the mapping data of component and parameter type pairing is obtained; Furthermore, a parameter classification and semantic annotation method is adopted (parameters: category labels include classification threshold offset, heart rate variability feature weight gain coefficient, eye movement frequency sensitivity adjustment factor, EEG alpha power normalized baseline bias term and control input frequency fusion weight ratio) to achieve classified storage of mapping data and generate a parameter component table with semantic labels; Furthermore, the vector component values ​​are directly assigned to the corresponding parameter objects through index-value parsing operations (formula below):

[0018] in, The fatigue classifier parameter value corresponding to index i. To determine the boundary, adjust the component values ​​of the vector at index i; Furthermore, a parameter consistency verification algorithm is adopted (parameter: allowable deviation range ε ​​is set according to different parameter types) to realize the legality verification of the adjustment vector after it is assigned to the model parameters, and to generate a verified parameter mapping result set; Through the tensor parsing, index matching, semantic annotation, assignment operation and consistency verification mentioned above, the discrimination boundary adjustment vector is transformed into a structured parameter set with explicit parameter types and values, realizing the direct mapping between the adjustment vector and the internal structural parameters of the fatigue classifier; For example, in a single flight mission, the discrimination boundary adjustment vector output by the parameter tuning unit is a 5-dimensional tensor. The tensor channel analysis method was used to determine the corresponding classification threshold offset, heart rate variability feature weight gain coefficient, eye movement frequency sensitivity adjustment factor, EEG alpha wave power normalized baseline bias term, and control input frequency fusion weight ratio for each component. Based on the parameter index mapping table, the matching relationship between the components and the parameter set was established, for example, index 0 → classification threshold offset, index 1 → heart rate variability feature weight gain coefficient, etc. After the assignment operation, the classification threshold offset P(0) = 0.12, the heart rate variability feature weight gain coefficient P(1) = 1.08, the eye movement frequency sensitivity adjustment factor P(2) = 0.95, the EEG alpha wave power normalized baseline bias term P(3) = -0.03, and the control input frequency fusion weight ratio P(4) = 0.87. The parameter consistency check showed that each component was in the preset legal range, realizing the direct binding of the adjustment vector to the internal structural parameters of the fatigue classifier. In this embodiment, the mapping results are sent to steps S4.2-S4.4 in real time to ensure that the model can immediately adjust the discrimination threshold and feature weight distribution according to changes in task load, thereby improving the sensitivity and accuracy of fatigue recognition in high-load scenarios. S4.2: The classification threshold offset is processed by nonlinear transformation and mapped to the preset dynamic adjustment range [-Δth, +Δth] using the Sigmoid compression function to obtain the classification boundary offset parameter after amplitude limiting; based on this parameter, the fatigue state discrimination threshold of the original fatigue classifier is pulled down and corrected, especially in high comprehensive load index scenarios, the threshold is shifted down to reduce the system's tolerance to mild physiological abnormalities and activate the fatigue response mechanism in advance; The Sigmoid compression function (parameters: compression coefficient κ, center shift parameter b) is used to perform a nonlinear transformation mapping on the input data of the classification threshold offset in order to achieve a smooth limit on excessively large or small offsets. Furthermore, the original offset δ is mapped to a preset dynamic adjustment range using a compression function. To obtain the classification boundary offset parameter β after amplitude limiting, so as to ensure that the threshold adjustment range is within a controllable range; Using a threshold correction algorithm (parameter: comprehensive load index) The segmented adjustment coefficient γ), in the high comprehensive load index scenario, affects the discrimination threshold of the original fatigue classifier. Perform a dropdown correction to generate a new threshold. When β is negative, it indicates that the decision threshold is lowered to trigger the fatigue response earlier; Furthermore, a load state weighting function is adopted. The β value is adjusted by weighting based on the current comprehensive load index value, so as to achieve a dynamic effect of significantly lowering the threshold under high load and slightly correcting it under low load. The correction amount output by this weighting function is superimposed on the β after the amplitude limiting to form the final downward offset. This is used to update the discriminant function input of the fatigue classifier; Perform model consistency verification on the updated discriminant function (parameter: recent physiological signal statistical mean). mean behavioral response To ensure that threshold changes do not cause significant misjudgment drift, if a consistency error is detected to exceed the allowable threshold ε, a portion of the offset is automatically rolled back to maintain model stability. By combining dynamic limiting and offset pull-down processing, the threshold component in the discrimination boundary adjustment vector of the previous step is transformed into a correction parameter that can be directly applied in the current task load context, thereby achieving an adaptive adjustment effect of fatigue judgment sensitivity. For example, in a single operation with an input frequency of 45 times per minute, a route change frequency of 12 times, and a comprehensive load index of... Under a high-load task condition with a parameter value of 0.82, the classification threshold offset δ output by the parameter tuning unit is -0.15. The Sigmoid compression function is used. Setting κ=10 and b=0, mapping δ to the interval [-Δth, +Δth], and Δth=0.2, we obtain β=-0.13. This is based on the high-load state weighting function. =1.25, adjust β-weighted to =-0.1625. Update the original threshold. =0.6 is set as the new threshold θ=0.4375, realizing the earlier fatigue judgment trigger point. In the consistency verification, the deviation between the fatigue judgment output and the historical high load scenario is detected to be within the allowable range, and the correction amount remains unchanged. Finally, in the feature-level fusion analysis stage, the system's sensitivity to mild HRV anomalies is significantly improved, and fatigue warning results are output in a timely manner. S4.3: Perform a product coupling operation on the heart rate variability feature weight gain coefficient and the eye movement frequency sensitivity adjustment factor to generate a physiological-behavioral dual-modal feature enhancement factor; input this factor into the feature weight redistribution module, and perform an affine transformation on the original feature weight distribution based on this factor to obtain an updated feature weight vector, which significantly improves the relative contribution ratio of behavior-related features (such as eye movement frequency) under high task load, realizing the judgment logic shift from 'physiological dominance' to 'behavioral enhancement'; S4.4: Based on the baseline bias term of the normalized alpha wave power of EEG, the background drift compensation processing of the original EEG signal features is performed, the deviation of the current alpha band energy relative to the individual baseline level is calculated, and the trend component is extracted by the moving average filter; the trend component is superimposed on the input of the bias term of the classifier to dynamically correct the baseline drift phenomenon caused by long-term task accumulation and maintain the discrimination stability of the model in continuous work scenarios. S4.5: Integrate all modified classification thresholds, feature weighting vectors, and bias term parameters to construct a real-time fatigue judgment sub-model with current task context awareness. This sub-model inherits the network architecture or decision rule framework of the original fatigue classifier, but its discrimination boundary has been adaptively deformed according to the task load state, and the output form remains consistent to ensure interface compatibility with subsequent fusion analysis processes. This provides a context-adaptive inference engine foundation for performing feature-level fusion analysis in the next step.

[0019] Step S5: The aligned physiological signal data and flight status data are input into the real-time fatigue determination sub-model, feature-level fusion analysis is performed, and a preliminary fatigue score for the current moment is output. This score is a continuous numerical index, representing the quantitative result of the pilot's fatigue level. Specifically, it includes: S5.1: Acquire multi-source heterogeneous data streams aligned with timestamps, including heart rate variability (HRV), eye movement frequency (Blink Rate), and EEG alpha power from the physiological signal acquisition module, as well as control input frequency, number of route changes, number of target locks, and communication interruption frequency from the flight status perception module; segment and synchronize the data of each channel based on a unified time window slicing mechanism to eliminate sampling delay differences and generate a structured temporal feature matrix as the input condition for feature-level fusion; For the multi-source heterogeneous data access channel configured at the input end of the real-time fatigue judgment sub-model after the S4 step correction, the heart rate variability (HRV), eye movement frequency (Blink Rate), and EEG alpha power from the physiological signal acquisition module, as well as the control input frequency, number of route changes, number of target locks, and communication interruption frequency from the flight status perception module, are selected as the raw data input objects for this sub-step. A unified timestamp alignment mechanism (parameters: millisecond-level synchronization accuracy, NTP network time synchronization protocol configuration delay threshold <2ms) is adopted to achieve consistency in the output timing of each data channel and eliminate the interference of sampling time offset between different acquisition modules on subsequent fusion analysis; Furthermore, using the time window slicing method (parameter: window length) =5s, window slide step size =1s), the synchronized channel data is divided into a series of fixed-length segments according to the time period, realizing the synchronous segmentation processing of multi-source heterogeneous data within the same time window, and obtaining multi-channel feature data segments within the window; Furthermore, for each time window, data structured encoding processing is performed within each channel segment to construct a matrix with time steps as rows and feature channels as columns, where each column corresponds to a physiological or behavioral feature, forming a structured time series feature matrix for subsequent normalization and nonlinear fusion modeling. By using the timestamp alignment and window slicing methods described above, the original heterogeneous data stream from the previous step is transformed into a structured temporal feature matrix, thus laying the foundation for seamless fusion of multi-source data in the time dimension. For example, in a UAV cruise mission with a medium-to-high workload, the sampling rate for heart rate variability data was set to 250Hz, the sampling rate for eye movement frequency was set to 60Hz, the sampling rate for EEG alpha wave power was set to 256Hz, and the sampling period for flight status features was uniformly set to 1Hz. The NTP synchronization mechanism was configured with a delay threshold of 2ms, and the hardware pulse signal trigger time synchronization interface delay was less than 1ms. The time window slicing method selected a window length of 5s and a sliding step size of 1s. Within a certain time window, the mean values ​​of 5 heart rate variability segments, the total number of 5 eye movement frequency segments, the mean values ​​of 5 EEG alpha wave powers, and the corresponding 5 control input frequencies, number of route changes, number of target locks, and frequency of communication interruptions could be obtained. After matrix encoding, a structured temporal feature matrix was obtained. This matrix is ​​the direct input for subsequent S5.2 normalization processing, ensuring the consistency of different modal data in time and structure, and significantly improving the stability and robustness of feature fusion analysis. S5.2: Perform multimodal feature normalization processing on the structured time-series feature matrix, and use the Z-score normalization method to perform dimensionless transformation on physiological and behavioral features of different dimensions to suppress the interference of numerical scale differences between modes on the fusion results; at the same time, introduce a sliding window mean filtering algorithm to smooth out abrupt outliers and obtain a standardized feature vector set as the data basis for subsequent nonlinear fusion modeling. S5.3: Input the standardized feature vector set into the real-time fatigue judgment sub-model that has been constructed in step S4. This sub-model is a bidirectional long short-term memory network (Attention-BiLSTM) enhanced by the attention mechanism. Its classification threshold boundary and feature weight distribution have been dynamically corrected according to the task load index. The network calculates the hidden state of each time step through forward propagation and uses the attention weights to adaptively focus on key fatigue-related feature segments to extract highly discriminative temporal representations. The input conditions are a standardized feature vector set obtained by multimodal feature normalization, including heart rate variability, eye movement frequency, EEG alpha power and flight control behavior features, and have been matched with the classification threshold and feature weight dynamic correction parameters corresponding to the task load state. An Attention-BiLSTM network structure (parameters: 128 bidirectional hidden layer units, time step size is uniform slice window length, and attention layer weight initialization strategy is uniform distribution) is adopted to realize temporal bidirectional encoding of the standardized feature vector set. The forward and backward memory units capture the sequential dependencies of features on the time axis and generate a bidirectional hidden layer state sequence. Furthermore, through an attention mechanism (parameters: the attention scoring function adopts Scaled Dot-Product, and the scaling factor is calculated based on the square root of the hidden layer dimension), the correlation model of each time step in the bidirectional hidden layer state sequence is realized, and the attention weight distribution of each time step is obtained; Furthermore, the bidirectional hidden state is weighted and aggregated by attention weight distribution to highlight feature segments that are highly related to the fatigue state and suppress noise features that are irrelevant to the judgment target, thereby obtaining a highly discriminative temporal representation vector. Furthermore, a feature discarding regularization method (parameter: discarding rate 0.3) is adopted to perform structural sparsification on the highly discriminative time series representation vector, so as to alleviate the risk of overfitting of the model under high-dimensional feature input and enhance the generalization ability under the scenario of changing task load. By using the Attention-BiLSTM encoding and feature aggregation processing method described above, the standardized feature vector set from the previous step is transformed into a highly discriminative temporal representation with adaptive capability under task load conditions, thereby enabling accurate capture of fatigue-related patterns in variable flight mission scenarios. For example, in a low-altitude reconnaissance mission, the input standardized feature vector has a dimension of 7, where HRV is... Standardized unit, eye movement frequency (IF) Standardized unit, alpha wave power in EEG is Standardized units; the rest are flight control frequencies and mission event parameters. The number of bidirectional hidden layer units in the Attention-BiLSTM network is configured as follows: The time step is The scaling factor is calculated as follows: In the hidden layer coding stage, the sequence lengths output by both the forward and backward units are... The attention scoring function calculates a correlation score with the global fatigue pattern at each time step, distributed across... Within the specified interval, highly relevant segments receive higher weight values. In the weighted aggregation operation, the HRV and eye-tracking frequency feature contributions of high-weight segments are increased, forming segments of length [missing value]. The time-series representation vector, after being dropped at a rate of After regularization and sparsification, the representation vector in the test set shows significant sensitivity to sudden changes in fatigue state, and can accurately capture the linkage changes in physiological and behavioral characteristics when the task load increases exponentially. S5.4: Based on the high discriminative temporal representation, perform fully connected mapping and Sigmoid activation transformation to generate a continuous fatigue confidence output between 0 and 1. This output reflects the comprehensive fatigue trend intensity of the driver in the current task context. By calibrating the mapping relationship through historical validation datasets, the confidence value range is divided into reference intervals corresponding to low, medium and high levels to form an interpretable preliminary fatigue score. The input condition is the highly discriminative temporal representation output from step S5.3. This representation is extracted by a bidirectional long short-term memory network with enhanced attention mechanism under dynamically corrected classification threshold and feature weight distribution conditions. It covers multimodal fatigue-related features such as heart rate variability, eye movement frequency, EEG alpha power, control input frequency, number of route changes, number of target locks, and communication interruption frequency. A fully connected mapping method (parameters: the number of output nodes is set to 1, and the weight matrix is ​​initialized with a uniform distribution of Xavier) is adopted to reduce the dimensionality of the highly discriminative time series representation to a single fatigue confidence prediction space, so as to ensure that the output dimension matches the subsequent scoring mapping module. Furthermore, by processing the fully connected mapping results through the Sigmoid activation function, the fatigue confidence score is limited to the numerical range [0,1], and a continuous fatigue confidence score index is obtained, which ensures the standardization of the value range and enhances interpretability. Furthermore, the mapping relationship was calibrated using historical verification datasets, and the continuous fatigue confidence score was divided into intervals using a piecewise linear fitting method. The value range was mapped to three reference intervals: low, medium, and high, to ensure the consistency between the scoring interval division and the actual fatigue level distribution, and to form an interpretable preliminary fatigue score. Furthermore, the boundary values ​​of the three reference intervals are fine-tuned by using an interval endpoint optimization algorithm (parameter: endpoint adjustment step size is set to 0.01) to improve the correlation between the classification boundary and the individual driver fatigue response mode. By using fully connected mapping and Sigmoid activation, the high discriminative time series representation of the previous step is transformed into a continuous fatigue confidence level and divided into low, medium and high intervals, thus generating a preliminary fatigue score with interpretability and numerical stability. For example, in a patrol reconnaissance mission, the high-discriminative temporal representation has a dimension of 128. A fully connected mapping method is used, configuring the number of output nodes to 1. The weight matrix is ​​initialized with a Xavier uniform distribution, and the mapped output value is... Activated by the Sigmoid function Calculate the confidence level as Using historical validation datasets, mapping relationships are established, mapping confidence values ​​in the range [0, 0.4] to low fatigue, (0.4, 0.7] to medium fatigue, and >0.7 to high fatigue. The current confidence value corresponds to the medium fatigue range, and the initial fatigue score is output as follows. In multiple tests within the same scenario, the interval endpoint optimization algorithm fine-tuned the upper fatigue limit to [value missing]. This improves the consistency between the scoring and the driver's actual fatigue report, significantly increasing the accuracy of the judgment; S5.5: Output the preliminary fatigue score as a continuous numerical indicator, and attach the current timestamp, task load index label and the sub-model version identifier used, and encapsulate it into a structured data unit for use in step S6; at the same time, cache the score to a short-term memory queue to support subsequent sliding window trend analysis and dynamic threshold backtracking verification, and improve the time consistency and anti-jitter capability of fatigue judgment.

[0020] Step S6: Invoke the driver's individual historical response pattern memory database, retrieve the average physiological-behavioral response characteristics of the driver under the same or similar comprehensive load index, calculate the deviation between the current preliminary fatigue score and the historical benchmark, and generate a personalized deviation compensation vector. Specifically, this includes: S6.1: Based on the driver's identity information, retrieve the driver's historical record set from the individual historical response pattern memory bank. The historical record set contains multiple historical task segments. Each historical task segment is labeled with the corresponding comprehensive load index, synchronously collected physiological signal characteristics (including heart rate variability, eye movement frequency and EEG alpha wave power), and corresponding behavioral response characteristics (including control input frequency and number of route changes) to build the data foundation for individualized modeling. S6.2: Perform comprehensive load matching processing on the retrieved historical task segments. Based on the comprehensive load index at the current moment as the query key, use the K-nearest neighbor algorithm to filter out several historical segments with similar load states within a preset threshold range to form a matching historical response subset, so as to ensure that the referenced historical response pattern is comparable to the current task situation. Using the current comprehensive load index as the query key for matching and retrieval, a feature selection preprocessing algorithm (parameter: consistency constraint of comprehensive load feature vector dimension) is adopted to calculate the difference between the comprehensive load index of each record in the historical task fragment set and the query key. Furthermore, the comprehensive load index is sorted by distance using the Euclidean distance metric formula, as follows:

[0021] in, For the first The load difference of each historical segment, This is the comprehensive load index for this segment. The query result is the comprehensive load index at the current moment; Furthermore, the K-Nearest Neighbors (KNN) matching algorithm (parameters: number of neighbors K, distance metric type Euclidean distance, distance weight mode set to uniform weight) is used to filter the top K historical task segments from the sorted results based on the minimum distance, and construct a candidate matching set; Furthermore, a threshold filtering method (parameter: preset load similarity threshold ΔL) is used to perform difference discrimination on each segment in the candidate matching set, eliminating any... Records greater than ΔL are used to control the load index similarity of the matching set; Furthermore, the candidate matching set after threshold filtering is defined as a subset of matching historical responses, and its index list is associated with and stored with the original historical record set to ensure that it can be directly accessed in subsequent feature time-series alignment and mean calculation steps; By using the K-nearest neighbor matching algorithm and threshold filtering method described above, the historical task fragments retrieved in the previous step are transformed into a matching historical response subset with high comprehensive load similarity, thereby achieving comparability between the reference sample and the current task context in the load feature space. For example, a driver's current overall load index is 0.78. The historical response pattern memory contains 200 task segment records, with overall load index values ​​ranging from 0.35 to 0.92 for each segment. The KNN algorithm is set to have 5 neighbors (K), Euclidean distance is calculated based on the one-dimensional difference in overall load index, and a preset similarity threshold ΔL is set to 0.05. The calculation is performed for each historical segment. The results are then sorted in ascending order, and the five records with the smallest differences are selected as the candidate matching set. For example, records ranked 1st to 5th... The values ​​were 0.01, 0.02, 0.03, 0.04, and 0.06, respectively. After filtering with a threshold ΔL=0.05, records with a difference of 0.06 were removed, resulting in a final output subset of matching historical responses containing four task segments. These four segments will be used in subsequent steps for feature mean calculation and baseline vector generation, significantly improving the adaptability and accuracy of individualized fatigue assessment in similar task scenarios. S6.3: Perform time-series alignment and normalization on each physiological-behavioral feature sequence in the matched historical response subset, and then calculate the statistical mean of each type of feature on the matched subset to generate the standardized historical response baseline vector of the driver under the current load level, as an expected reference model for individualized fatigue response. S6.4: Map the preliminary fatigue score output by S5 to the same quantization space as the historical response baseline vector, calculate the numerical deviation between the score and the expected fatigue baseline value derived from the historical response baseline vector, and generate an initial deviation value using the weighted Euclidean distance metric method to characterize the degree to which the current response deviates from the individual's normal state. Using the preliminary fatigue score output from step S5 as input, a quantization space mapping method (parameter: the target space of the mapping is the feature scale range defined by the historical response benchmark vector) is adopted to transform the score into a measurement system consistent with the historical response benchmark vector, so as to eliminate the differences in numerical scale between different data sources. Furthermore, by deriving the expected fatigue baseline value through the historical response baseline vector (parameter: calculation formula based on the statistical mean of the matched historical response subset), the expected fatigue response value of individual drivers under the corresponding load level is generated, and a baseline score reflecting the normal state is obtained. Furthermore, a weighted Euclidean distance metric (parameter: the weight coefficient vector is derived from the contribution ratio of each physiological-behavioral feature in the fatigue assessment model) is used to calculate the deviation between the current preliminary fatigue score and the expected fatigue baseline value, and to generate an initial deviation value that quantifies the degree of deviation. The calculation formula is as follows:

[0022] in, The weighting coefficients are for the i-th type of feature. This represents the mapping value of the initial fatigue score to this feature dimension. This is the baseline value of the historical response baseline vector in this feature dimension, where n is the total number of feature dimensions; By using the weighted Euclidean distance metric described above, the quantification spatial mapping result from the previous step is transformed into a deviation metric that can be used for subsequent stability adjustment, thereby achieving precise quantification of the degree to which the current response deviates from the individual's normal state. For example, in a high-load UAV reconnaissance mission, the initial fatigue score is 0.72. After mapping to the three-dimensional parameter space of the historical response baseline vector, the values ​​of each feature dimension are 0.68 (heart rate variability mapping value), 0.75 (eye movement frequency mapping value), and 0.70 (EEG alpha wave power mapping value). The statistical mean baseline vector of the matched historical response subset is 0.65, 0.73, and 0.69, and the weighting coefficients of each feature are set to 0.4, 0.35, and 0.25. Substituting the above data into the weighted Euclidean distance formula, the calculation process is as follows: the difference in heart rate dimension is 0.03, multiplied by a weight of 0.4 and squared, resulting in 0.000144; the difference in eye movement frequency is 0.02, multiplied by a weight of 0.35 and squared, resulting in 0.000049; the difference in EEG alpha wave power is 0.01, multiplied by a weight of 0.25 and squared, resulting in 0.00000625. The sum of these three values ​​is 0.00019925, and the square root yields an initial deviation value of 0.0141. This deviation value represents a slight deviation between the driver's fatigue response under the current task and their individual normal state. In S6.5, this value can be nonlinearly corrected based on a stability adjustment factor, thereby improving the robustness and accuracy of the final personalized compensation. S6.5: Introduce an individual stability adjustment factor to the initial deviation value for nonlinear correction. This adjustment factor is dynamically generated based on the standard deviation of the driver's historical response and is used to suppress the interference of abnormal historical samples on deviation estimation. Finally, a robust and personalized deviation compensation vector is output as the correction input for weighted fusion in S7.

[0023] Step S7: The personalized deviation compensation vector is weighted and fused with the preliminary fatigue score to output the final fatigue level determination result after individual adaptation correction. This result is divided into three discrete states: low, medium, and high, for subsequent alarm strategy matching. Specifically, it includes: S7.1: Obtain the preliminary fatigue score and the personalized deviation compensation vector. The preliminary fatigue score is a continuous numerical index output by S5, which represents the quantitative result of the fatigue level at the current moment before individual adaptation correction. The personalized deviation compensation vector is the deviation value calculated by S6 based on the mean of physiological and behavioral characteristics stored in the driver's individual historical response pattern memory bank and the current preliminary fatigue score, reflecting the typical response deviation characteristics of the driver under the same or similar comprehensive load index. S7.2: The weighted fusion coefficient is calculated based on a dynamic weight allocation mechanism. This mechanism automatically adjusts the contribution ratio of the compensation vector according to the current range of the comprehensive load index: when the comprehensive load index is higher than the preset high load threshold, the weight of the individual deviation compensation vector is increased to enhance the system's ability to capture the nonlinear response characteristics of individual drivers under high-pressure tasks; conversely, in the low load range, the influence of the compensation vector is reduced, and the basic fatigue score obtained from group training is given priority. S7.3: Perform a weighted summation operation on the preliminary fatigue score and the personalized deviation compensation vector to generate an individual-adapted fatigue comprehensive score: Specifically, multiply the preliminary fatigue score by its corresponding first fusion weight coefficient, multiply the personalized deviation compensation vector by its corresponding second fusion weight coefficient, add the two together and normalize them to obtain an individual-adapted fatigue comprehensive score in the range of [0,1], which is used as the optimized evaluation value of the fusion individual's historical response characteristics. S7.4: Discretize and classify the individual's adaptive fatigue comprehensive score based on a preset adaptive classification threshold. The classification threshold is dynamically adjusted according to the distribution statistics of the driver's historical fatigue events to ensure that the classification boundary conforms to the driver's actual fatigue evolution pattern. When the individual's adaptive fatigue comprehensive score is lower than the first threshold, it is determined to be a "low" fatigue level; when it is between the first and second thresholds, it is a "medium" fatigue level; and when it is higher than the second threshold, it is determined to be a "high" fatigue level. The input condition is the individual adaptive fatigue comprehensive score output from step S7.3, whose numerical range has been normalized to [0,1]. The classification threshold set needs to be dynamically determined based on the statistical results of the driver's historical fatigue event distribution. The classification threshold set includes a first threshold and a second threshold, used for three-level discretization classification. A historical event distribution modeling method (parameters: driver fatigue level label sequence, historical task load index sequence) is used to perform statistical analysis on the probability of individual driver fatigue levels. Furthermore, a percentile calculation method (parameter: cumulative distribution function CDF) is used to initially estimate the first and second thresholds, obtaining a reference set of thresholds that satisfy historical distribution patterns. Furthermore, a sliding window weighted update algorithm (parameters: window length W, decay coefficient λ) is used to dynamically adjust the threshold set and generate real-time values ​​for the first and second thresholds adapted to the current task, improving the ability of the classification boundary to track recent task states. Furthermore, a boundary smoothing function (parameters: smoothing order n, smoothing interval Δ) is employed to suppress noise on the first and second thresholds, preventing classification results from jumping due to sudden changes in instantaneous task load. Furthermore, through conditional discrimination logic, the individual's overall fatigue score is used to determine interval assignment based on a dynamically adjusted set of thresholds. When the condition is met... Achieve low fatigue level output; when the following conditions are met The fatigue level output is realized in real time; when the condition is met... It can achieve high fatigue level output and generate corresponding discrete state labels; By using adaptive threshold division and dynamic interval discrimination processing, the normalized individual adaptive fatigue comprehensive score of the previous step is transformed into three discrete fatigue level states, achieving a high degree of consistency between the classification boundary and the individual fatigue evolution law. For example, in the past 120 task segments, the probabilities of low, medium, and high fatigue levels for a driver are 0.35, 0.45, and 0.20, respectively. The initial estimate of the first threshold, calculated using the cumulative distribution function, is... The initial estimate of the second threshold is The sliding window length is set to... The attenuation coefficient λ is The smoothing order is The smoothing interval Δ is In the current task, the individual adaptation fatigue score is: The first threshold after smoothing is The second threshold is The discriminant logic calculation satisfies The conditions are met, and a "high" fatigue level label is output, which drives the subsequent multimodal alarm mechanism to activate; S7.5: Output the final fatigue level determination result, which is a discrete state identifier containing three levels: "low", "medium" and "high", along with a confidence label and determination timestamp. This is used to drive the matching of differentiated response strategies in the subsequent multimodal alarm mechanism, and at the same time, to provide a traceable decision basis for system log recording.

[0024] Step S8: Determine whether the final fatigue level determination result reaches the preset high fatigue alarm threshold. If it does, trigger the multimodal alarm mechanism and record the context information of this determination to the system log for subsequent model iteration optimization and abnormal scenario tracing. Specifically, this includes: S8.1: Based on the preset high fatigue alarm threshold, the final fatigue level determination result is subjected to threshold comparison processing, wherein the determination result is a low, medium and high three-level discrete state output after individual adaptation correction, and the alarm triggering conditions are determined by logical judgment to achieve qualitative identification of fatigue risk. S8.2: After confirming that the high fatigue alarm threshold conditions are met, the multimodal alarm mechanism is activated. Based on the UAV human-machine interaction system architecture, the sound and light warning unit and tactile feedback device are activated simultaneously to execute multiple alarm signal outputs such as visual flashing, audio prompts and seat vibration, so as to improve the driver's perception reliability in complex operating environments. S8.3: Call the system event log recording module to generate a context snapshot of this fatigue alarm event. This snapshot includes the current comprehensive load index, raw physiological signal data, flight status data, preliminary fatigue score, personalized deviation compensation vector and final fatigue level judgment result, which is persistently stored as a data encapsulation of the complete judgment chain. S8.4: Upload the context snapshot to the cloud data analysis platform, and use the incremental learning pipeline in the edge-cloud collaborative architecture to perform label archiving and feature space mapping processing on the sample in order to expand the historical task-fatigue database and provide new training instances for model-independent fine-tuning under the subsequent meta-learning framework; S8.5: Based on the recorded context information, execute the abnormal pattern detection algorithm, statistically model the time series distribution of continuous alarm events, the gradient of sudden changes in task load, and the trend of individual response deviations, identify potential system misjudgments or driver adaptive degradation, and generate a model optimization suggestion report for remote operation and maintenance center to review.

[0025] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0026] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fatigue alarm method integrating pilot physiological signals and UAV flight status data, characterized in that, Includes the following steps: S1: Acquire the pilot's physiological signal data and the UAV flight status data, and align and mark the physiological signal data and the UAV flight status data according to a unified timestamp; S2: Calculate the comprehensive load index of the current task based on the preset task type knowledge base and the UAV flight status data; S3: Input the comprehensive load index into the parameter tuning unit of the fatigue judgment model. The parameter tuning unit is constructed based on the model-independent fine-tuning idea under the meta-learning framework. It obtains the basic parameter space adaptability through historical multi-task-fatigue sample training and generates the corresponding discrimination boundary adjustment vector according to the comprehensive load index. S4: Based on the discrimination boundary adjustment vector, the classification threshold and feature weight distribution of the original fatigue classifier are dynamically corrected to form a real-time fatigue judgment sub-model; S5: Input the aligned physiological signal data and UAV flight status data into the real-time fatigue judgment sub-model, perform feature-level fusion analysis, and output the preliminary fatigue score at the current moment; S6: Call the driver's individual historical response pattern memory bank, retrieve the average value of the driver's past physiological-behavioral response characteristics under the same or similar comprehensive load index, calculate the deviation value between the preliminary fatigue score and the historical benchmark, and generate a personalized deviation compensation vector. S7: The personalized deviation compensation vector is weighted and fused with the preliminary fatigue score to output the final fatigue level determination result.

2. The fatigue alarm method according to claim 1, which integrates pilot physiological signals and UAV flight status data, is characterized in that... Following step S7, the following is also included: S8: Determine whether the final fatigue level determination result reaches the preset high fatigue alarm threshold. If it does, trigger the multimodal alarm mechanism and record the determination context information to the system log.

3. The fatigue alarm method according to claim 1, which integrates pilot physiological signals and UAV flight status data, is characterized in that... Step S1 specifically includes: Based on the raw bioelectrical signals of the driver collected by wearable physiological sensing devices, the ECG, EEG and eye movement signals were preprocessed using bandpass filtering and baseline drift correction algorithms to extract effective heart rate variability, EEG alpha wave power and eye movement frequency characteristic parameter sequences. Flight operation log data is acquired in real time through the UAV flight control system bus interface, and key operation behavior records are analyzed based on event recognition algorithms to generate a structured flight status event stream. A unified clock synchronization mechanism is deployed within the edge computing unit. A network time protocol combined with a hardware pulse alignment method is used to mark the data streams from the physiological acquisition module and the flight status perception module with a unified timestamp at the millisecond level, thereby obtaining a time-aligned set of multi-source data frames. Missing value detection and outlier filtering are performed on the time-aligned multi-source data frames. Linear interpolation and isolated forest algorithms are used to fill in transient signal loss and remove outliers that deviate significantly from the normal range, respectively, and the cleaned and standardized physiological-flight joint data sequence is output. The cleaned, standardized physiological-flight combined data sequences are encapsulated into feature vectors matched by timestamps to generate multidimensional fusion data samples.

4. The fatigue alarm method according to claim 3, which integrates pilot physiological signals and UAV flight status data, is characterized in that... The wearable physiological sensing device is configured to synchronously acquire ECG, EEG and EOG signals at a sampling frequency of 1 to 500 Hz, and the bus refresh cycle of the UAV flight control system is set to 10 ms to 100 ms.

5. The fatigue alarm method according to claim 1, which integrates pilot physiological signals and UAV flight status data, is characterized in that... Step S2 specifically includes: Acquire UAV flight status data, including control input frequency, number of route changes, number of targets locked, and communication interruption frequency, and form a structured sequence of task behavior features based on unified timestamp alignment markers; Based on a pre-defined task type knowledge base, pattern matching and semantic parsing are performed on the structured task behavior feature sequence to identify the type label of the currently executing task; Based on the identified task type labels and their corresponding complexity baseline weight vectors, a weighted summation algorithm is used to perform primary load component fusion on the frequency of control inputs, the number of route changes, and the number of targets locked, generating a task physical complexity sub-index. Using the frequency of communication interruptions and the degree of disruption in time sequence as inputs, the temporal distribution characteristics of sudden interference events are extracted through a long short-term memory network model, and the task time urgency sub-index is calculated by combining the task remaining time inverse ratio factor. Based on the rule engine, the task physical complexity sub-index and the task time urgency sub-index are integrated, and the target locking switching frequency is introduced as a decision pressure modulation term. A three-dimensional coupling operation is performed through a nonlinear activation function to output a normalized comprehensive load index.

6. The fatigue alarm method according to claim 5, which integrates pilot physiological signals and UAV flight status data, is characterized in that... The physical complexity sub-index of the generated task characterizes the operational burden borne by the driver in terms of spatial operation and action density; the task time urgency sub-index reflects the intensity of psychological pressure brought about by the accelerated pace of information processing under task deadline constraints.

7. The fatigue alarm method according to claim 1, which integrates pilot physiological signals and UAV flight status data, is characterized in that... Step S3 specifically includes: Based on a historically collected multi-task fatigue labeled sample dataset, a basic training set is constructed for training meta-learning parameter tuning units. Using a model-independent fine-tuning algorithm framework, meta-training is performed on the initial fatigue judgment parameter tuning network, and inner loop fast adaptation and outer loop gradient update are performed on multiple task clusters. Deploy the trained meta-learning parameter tuning unit to the fusion analysis module and connect its input interface to the output of the task load modeling submodule. The current moment comprehensive load index is embedded and encoded, and then mapped into a load representation vector in a high-dimensional semantic space using a preset nonlinear transformation function. Based on the load representation vector, forward inference is performed within the meta-learning parameter tuning unit to generate a discrimination boundary adjustment vector.

8. The fatigue alarm method according to claim 7, which integrates pilot physiological signals and UAV flight status data, is characterized in that... Each sample in the basic training set used to train the meta-learning parameter tuning unit contains a comprehensive load index for the corresponding task scenario, synchronously collected physiological signal feature vectors and behavioral state feature vectors, and fatigue level labels calibrated by experts.

9. The fatigue alarm method according to claim 1, which integrates pilot physiological signals and UAV flight status data, is characterized in that... Step S4 specifically includes: Obtain the discrimination boundary adjustment vector output by the parameter tuning unit, perform parameter mapping parsing operation based on the discrimination boundary adjustment vector, and use a predefined parameter index mapping table to associate each dimension component with five adjustable parameter sets: classification threshold offset, heart rate variability feature weight gain coefficient, eye movement frequency sensitivity adjustment factor, EEG alpha power normalized baseline bias term, and control input frequency fusion weight ratio. The classification threshold offset is subjected to nonlinear transformation processing and mapped to a preset dynamic adjustment range to obtain the classification boundary offset parameter after amplitude limiting. Based on the classification boundary offset parameter, the fatigue state discrimination threshold of the original fatigue classifier is pulled down for correction. The heart rate variability feature weight gain coefficient and the eye movement frequency sensitivity adjustment factor are subjected to a product coupling operation to generate a physiological-behavioral bimodal feature enhancement factor. The physiological-behavioral bimodal feature enhancement factor is input into the feature weight redistribution module, and the original feature weight distribution is subjected to affine transformation to obtain the updated feature weight vector. Based on the normalized baseline bias term of the EEG alpha wave power, the original EEG signal features are processed to compensate for background drift. The deviation of the current alpha band energy relative to the individual baseline level is calculated, and the trend component is extracted. The trend component is superimposed on the input of the classifier's bias term to dynamically correct the baseline drift phenomenon caused by long-term task accumulation. By integrating all the modified classification thresholds, feature weighting vectors, and bias term parameters, a real-time fatigue determination sub-model with current task context awareness is constructed.

10. The fatigue alarm method according to claim 1, which integrates pilot physiological signals and UAV flight status data, is characterized in that... Step S5 further includes employing a bidirectional long short-term memory network enhanced by an attention mechanism to perform temporal bidirectional encoding on the standardized feature vector set, using attention weights to focus on key fatigue-related segments, outputting a highly discriminative temporal representation, outputting a continuous fatigue confidence score in the [0,1] interval after fully connected and Sigmoid transformation, and segmenting it into low, medium, and high intervals based on historical validation datasets.

Images