Pilot stress state prediction model construction method and stress state prediction method

Abstract

The application provides a pilot stress state prediction model construction method and a stress state prediction method, and is applied to the field of aerospace artificial intelligence technology. For each flight scene in multiple flight scenes, a stress label is defined, multiple test variables related to the stress state of the pilot are determined, the multiple test variables include engine abnormalities, foreign object impact, communication failures, abnormal weather, abnormal tasks, abnormal psychological characteristics and abnormal physiological characteristics, and the respective sub-feature vectors of these data are obtained; conditional independence test is performed on each sub-feature vector, and sub-feature vectors having a causal relationship with the stress label are screened out to construct a causal feature set; the gain importance score of each sub-feature vector in the causal feature set is calculated; and the sub-feature vectors corresponding to the gain importance scores greater than the preset score threshold are used to construct a prediction model corresponding to the flight scene, the prediction model has strong interpretability and excellent generalization performance, and is used for accurately evaluating the stress state of the pilot and the attribution analysis result.

Description

Technical Field

[0001] This application relates to the field of aerospace artificial intelligence technology, and in particular to a method for constructing a pilot stress state prediction model and a stress state prediction method. Background Technology

[0002] In modern aerospace safety systems, the psychological and physiological stress state of pilots is a key factor affecting safe flight. Faced with complex flight environments, sudden system failures, or high-intensity mission loads, pilots may enter a state of high stress, leading to cognitive resource exhaustion, narrowed attention, delayed decision-making, and even operational errors, which can result in serious safety accidents. Therefore, real-time and accurate prediction and intervention of pilot stress states has become an urgent need to improve aerospace safety levels.

[0003] Traditional methods for predicting pilot stress states often rely on multimodal physiological signals (such as electrocardiogram, electroencephalogram, conductance of skin, and eye movement), but they still face two major challenges: First, most methods rely on statistical correlation to build prediction models, and the selected features lack causal explanations for stress states, resulting in poor model generalization ability and easy failure in unknown scenarios; Second, the features are highly dimensional and redundant, and the stress patterns of different flight scenarios (such as takeoff, cruise, and landing) are significantly different, making it difficult for a single prediction model to achieve robust predictions across scenarios.

[0004] Therefore, developing a new predictive system that can adapt to different flight scenarios and accurately predict pilot stress states has become a key breakthrough direction for ensuring the reliability and safety of human-machine systems in complex flight environments. Summary of the Invention

[0005] This application provides a method for constructing a pilot stress state prediction model and a stress state prediction method. Through a dual screening mechanism of causal inference and gain importance, it can screen key sub-feature vectors with high importance and true causal explanatory power for pilot stress states from multi-source features such as engine anomalies, foreign object impacts, communication failures, abnormal weather, abnormal missions, abnormal psychological characteristics, and abnormal physiological characteristics. This allows for the construction of a prediction model with strong interpretability and excellent generalization performance. Furthermore, dedicated prediction models are constructed for different flight scenarios, effectively overcoming prediction biases caused by differences in stress response patterns across scenarios, achieving accurate and stable assessment of pilot stress states. In addition, it can output attribution analysis of stress states, clarifying the contribution of each key sub-feature vector to the prediction results, providing a reliable basis for subsequent refined pilot state monitoring and real-time operational assistance.

[0006] In a first aspect, embodiments of this application provide a method for constructing a pilot stress state prediction model, including: Multiple flight scenarios are identified, and for each flight scenario, corresponding stress labels are defined and multiple experimental variables associated with the pilot's stress state are determined. The multiple experimental variables include at least engine abnormality, foreign object impact, communication failure, abnormal weather, abnormal mission, abnormal psychological characteristics, and abnormal physiological characteristics. Obtain the sub-feature vectors corresponding to the engine anomaly, the foreign object impact, the communication failure, the abnormal weather, the abnormal task, the abnormal psychological characteristics, and the abnormal physiological characteristics respectively; Perform a conditional independence test on each of the sub-feature vectors in all sub-feature vectors, and select the sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, and construct a causal feature set; and calculate the gain importance score of each sub-feature vector in the causal feature set when assessing the pilot's stress state. Based on the sub-feature vectors corresponding to the gain importance scores that are greater than a preset scoring threshold, a prediction model corresponding to the flight scenario is constructed. The prediction model is used to evaluate the stress state of the pilot in the flight scenario and the attribution analysis results of the stress state.

[0007] According to an embodiment of this application, a method for constructing a pilot stress state prediction model is provided. The method involves performing a conditional independence test on each sub-feature vector among all sub-feature vectors to select sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, and constructing a causal feature set. This includes: obtaining the observation distribution corresponding to the stress label; for each sub-feature vector among all sub-feature vectors, using the sub-feature vector as the intervention variable when assessing the pilot's stress state and the stress label as the outcome variable, determining the intervention distribution corresponding to the sub-feature vector by controlling a preset set of confounding variables and performing backdoor adjustments, wherein the set of confounding variables is used to describe and control known interference factors that simultaneously affect the intervention variable and the outcome variable in causal inference; and selecting sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state based on the observation distribution and the intervention distribution corresponding to each sub-feature vector, thus constructing the causal feature set.

[0008] According to an embodiment of this application, a method for constructing a pilot stress state prediction model is provided. The method involves using the sub-feature vector as the intervention variable for assessing the pilot's stress state and the stress label as the outcome variable. By controlling a preset set of confounding variables and performing backdoor adjustments, the method determines the intervention distribution corresponding to the sub-feature vector. The method includes: identifying a backdoor path between the intervention variable and the outcome variable based on the set of confounding variables. The backdoor path is a non-causal path connecting the intervention variable and the outcome variable, and the non-causal path begins with an arrow pointing to the intervention variable; determining a subset of confounding variables to be controlled based on the backdoor path; and determining the intervention distribution corresponding to the sub-feature vector based on the subset of confounding variables.

[0009] According to an embodiment of this application, a method for constructing a pilot stress state prediction model is provided. The method involves selecting sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, based on the observed distribution and the intervention distribution corresponding to each sub-feature vector, and constructing the causal feature set. This includes: for each sub-feature vector, using the difference between the observed distribution and the intervention distribution corresponding to the sub-feature vector as an estimate of the causal effect of the probability of the pilot being in different stress states; performing a statistical significance test on the causal effect estimate to obtain a causal significance index for the sub-feature vector; confirming that the sub-feature vector has a causal relationship with the stress label when assessing the pilot's stress state if the causal significance index is less than a preset significance threshold; and constructing the causal feature set based on all sub-feature vectors that have a causal relationship with the stress label.

[0010] According to an embodiment of this application, a method for constructing a pilot stress state prediction model is provided. The step of calculating the gain importance score of each sub-feature vector in the causal feature set when evaluating the pilot's stress state includes: using all sub-feature vectors in the causal feature set as model input and the stress label as model output to train an XGBoost classifier for preliminary evaluation of the pilot's stress state; after the XGBoost classifier is trained, traversing all decision trees constructed by the XGBoost classifier, and statistically analyzing the information gain brought by each sub-feature vector in the causal feature set during the node splitting process of the decision tree to distinguish different stress states; and calculating the gain importance score of each sub-feature vector when evaluating the pilot's stress state based on the sum or average of the information gains corresponding to each sub-feature vector in the causal feature set. The gain importance score is used to characterize the discriminative contribution of each sub-feature vector to the XGBoost classifier.

[0011] Secondly, embodiments of this application provide a method for predicting pilot stress states, including: Given the current flight scenario, a test feature vector with the same structure as the initial feature vector is constructed based on the external stressor characteristics of the aircraft and the internal psychological and physiological characteristics of the pilot. The initial feature vector is determined based on multiple experimental variables under the current flight scenario. The feature vector to be tested is input into the prediction model corresponding to the current flight scenario to obtain the pilot's stress state and the attribution analysis results of the stress state. The prediction model is obtained based on any of the pilot stress state prediction model construction methods in the first aspect.

[0012] Thirdly, embodiments of this application also provide a system for constructing a pilot stress state prediction model, comprising: The vector acquisition module is used to determine multiple flight scenarios, define corresponding stress labels for each flight scenario, and determine multiple experimental variables associated with the pilot's stress state. The multiple experimental variables include at least engine anomaly, foreign object impact, communication failure, abnormal weather, abnormal mission, abnormal psychological characteristics, and abnormal physiological characteristics. The module acquires sub-feature vectors corresponding to each of the engine anomaly, foreign object impact, communication failure, abnormal weather, abnormal mission, abnormal psychological characteristics, and abnormal physiological characteristics. The model building module performs conditional independence tests on all sub-feature vectors, filters out sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, and constructs a causal feature set; it also calculates the gain importance score of each sub-feature vector in the causal feature set when assessing the pilot's stress state; and constructs a prediction model corresponding to the flight scenario based on the sub-feature vectors whose gain importance scores are greater than a preset score threshold. The prediction model is used to assess the pilot's stress state in the flight scenario and the attribution analysis results of the stress state.

[0013] Fourthly, embodiments of this application also provide a pilot stress state prediction system, including: The vector construction module is used to construct a test feature vector with the same structure as the initial feature vector, based on the external stressor characteristics of the aircraft and the internal psychological and physiological characteristics of the pilot, given the current flight scenario. The initial feature vector is determined based on multiple experimental variables under the current flight scenario. The state prediction module is used to input the feature vector to be tested into the prediction model corresponding to the current flight scenario to obtain the stress state of the pilot and the attribution analysis results of the stress state. The prediction model is obtained based on any of the pilot stress state prediction model construction methods in the first aspect.

[0014] Fifthly, embodiments of this application also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the pilot stress state prediction model construction method as described in any of the first aspects above, or the pilot stress state prediction system as described in the second aspect.

[0015] Sixthly, embodiments of this application also provide a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the pilot stress state prediction model construction method as described in any of the first aspects above, or the pilot stress state prediction system as described in the second aspect.

[0016] In a seventh aspect, embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements the pilot stress state prediction model construction method as described in any of the first aspects above, or the pilot stress state prediction system as described in the second aspect.

[0017] The pilot stress state prediction model construction method and stress state prediction method provided in this application embodiment determine multiple flight scenarios, define corresponding stress labels for each flight scenario, and determine multiple experimental variables associated with the pilot's stress state. The multiple experimental variables include at least engine anomaly, foreign object impact, communication failure, abnormal weather, abnormal mission, abnormal psychological characteristics, and abnormal physiological characteristics. The method obtains sub-feature vectors corresponding to each of the engine anomaly, foreign object impact, communication failure, abnormal weather, abnormal mission, abnormal psychological characteristics, and abnormal physiological characteristics. It performs conditional independence tests on each sub-feature vector in all sub-feature vectors, filters out sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, and constructs a causal feature set. It then calculates the gain importance score of each sub-feature vector in the causal feature set when assessing the pilot's stress state. Based on the sub-feature vectors corresponding to gain importance scores greater than a preset score threshold, a prediction model corresponding to the flight scenario is constructed. The prediction model is used to assess the pilot's stress state in the flight scenario and the attribution analysis results of the stress state. This method employs a dual screening mechanism of causal inference and gain importance to identify key sub-feature vectors with high importance and genuine causal explanatory power for pilot stress states from multi-source features such as engine anomalies, foreign object impacts, communication failures, abnormal weather, abnormal missions, abnormal psychological characteristics, and abnormal physiological characteristics. This allows for the construction of a predictive model with strong interpretability and excellent generalization performance. Furthermore, dedicated predictive models are built for different flight scenarios, effectively overcoming prediction biases caused by differences in stress response patterns across scenarios, achieving accurate and stable assessment of pilot stress states. In addition, attribution analysis of stress states can be output, clarifying the contribution of each key sub-feature vector to the prediction results, providing a reliable basis for subsequent refined pilot state monitoring and real-time operational assistance. Attached Figure Description

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

[0019] Figure 1 This is a flowchart illustrating the pilot stress state prediction model construction method provided in this application embodiment; Figure 2 This is a flowchart illustrating the pilot stress state prediction method provided in the embodiments of this application; Figure 3This is a schematic diagram of the structure of the pilot stress state prediction model construction system provided in this application embodiment; Figure 4 This is a schematic diagram of the pilot stress state prediction system provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] It should be noted that the execution entity involved in the embodiments of this application can be a pilot stress state prediction model construction system or an electronic device. Optionally, the electronic device may include: a computer / laptop, a mobile terminal, a server, and an aircraft central control device, etc.

[0022] The following section uses electronic devices as an example to illustrate in detail the method for constructing a pilot stress state prediction model provided in this application: Figure 1 This is a flowchart illustrating the method for constructing a pilot stress state prediction model provided in an embodiment of this application. Figure 1 As shown, the method may include the following steps 101-104.

[0023] Step 101: Identify multiple flight scenarios. For each flight scenario, define corresponding stress labels and determine multiple experimental variables associated with the pilot's stress state. These multiple experimental variables should include at least engine malfunction, foreign object impact, communication failure, abnormal weather, abnormal mission, abnormal psychological characteristics, and abnormal physiological characteristics.

[0024] Among them, multiple flight scenarios can also be referred to as multiple flight phases. Optionally, these multiple flight scenarios may include at least the takeoff, cruise, and landing of the aircraft.

[0025] Stress labels are indicators used to quantify or classify the level of psychological and physiological stress experienced by pilots during a specific flight mission. Optionally, these stress labels can typically be categorized into discrete levels such as "normal / low stress," "moderate stress," and "excessive / high stress."

[0026] Experimental variables refer to internal and external disturbances and state characteristics that may induce stress responses in pilots during flight simulation or actual flight, and are used to influence the stress state of pilots.

[0027] Engine malfunctions refer to various failures in an aircraft's power system, such as engine misfire, abnormal speed, thrust asymmetry, or in-flight engine failure, which result in impaired flight performance.

[0028] Foreign object impact refers to a sudden event in which an aircraft is struck by birds, hail, drones, or other space debris during flight or gliding, resulting in damage to the aircraft's structure or sensor failure.

[0029] A communication failure refers to a disruption of the communication link between the pilot and ground control, air traffic control system or other crew members, noise interference or incorrect instruction transmission, which hinders information exchange.

[0030] Abnormal weather refers to complex environments that exceed normal flight weather conditions, such as strong turbulence, low visibility, wind shear, severe thunderstorms, or wing icing, which increase the difficulty of handling.

[0031] An unusual mission refers to a mission instruction that is temporarily changed during flight, such as diverting to another location, emergency avoidance, performing high-difficulty maneuvers, or a mission that is urgent due to time pressure.

[0032] Abnormal psychological characteristics refer to the negative psychological states exhibited by pilots during flight missions, such as narrow attention (tunnel vision), spatial disorientation, cognitive overload, or decision-making hesitation caused by excessive tension.

[0033] Abnormal physiological characteristics refer to abnormal fluctuations in the pilot's biosignals collected by sensors, such as a decrease in heart rate variability (HRV), a sudden increase in Galvanic Skin Response (GSR), an imbalance in the ratio of theta waves to alpha waves in the electroencephalogram (EEG), or disordered eye movement trajectories.

[0034] In step 101, the electronic device first combines flight data recordings and audio / video data to divide a complete flight mission into different flight phases such as takeoff, cruise, and landing. Then, the electronic device can define stress labels corresponding to each flight phase. Specifically: for the takeoff phase, since takeoff is a high-load, irreversible process, the stress label for this phase focuses on reflecting the pilot's instantaneous physiological arousal when facing thrust imbalance, extreme crosswinds, or changes in ground instructions. Based on this, the stress label for this takeoff phase is a stress state grading defined based on instantaneous load and operational precision. For the cruise phase, since the task in the cruise phase is relatively simple and lasts a long time, the stress label for this phase focuses on reflecting the pilot's stress level in a low-stimulation environment. Based on the level of alertness and sensitivity to sudden failures (such as communication failures or system failures) in the cruise phase, the stress label for this cruise phase is a stress state classification defined by alert level and monotonous fatigue. For the landing phase, since the landing phase is the riskiest and visual information processing phase of the flight mission, the stress label for this landing phase focuses on reflecting the pilot's spatial orientation ability and psychological threshold under complex weather or runway changes. Based on this, the stress label for this landing phase is a stress state classification defined by multidimensional decision-making pressure and attitude control robustness.

[0035] Meanwhile, the electronic device can identify multiple experimental variables associated with the pilot's stress state at each stage of flight. Specifically, the experimental variables for the takeoff stage focus on power redundancy and takeoff safety, which directly reflect the psychological impact on the pilot within a very short decision window; the experimental variables for the cruise stage focus on environmental monotonicity and system reliability, which capture the damage to the pilot's alertness level caused by sudden disturbances under long-term low-load conditions; and the experimental variables for the landing stage focus on precision approach and multi-source information fusion, which reflect the cognitive load on the pilot when dealing with high-frequency visual input and attitude micro-manipulation in complex landing environments.

[0036] The content recorded in step 101 addresses the prediction bias caused by the differences in stress patterns across different flight phases by defining stress labels and experimental variables in a differentiated manner. This achieves accurate characterization and benchmark alignment of pilot stress states under different mission backgrounds.

[0037] Step 102: Obtain the sub-feature vectors corresponding to engine abnormality, foreign object impact, communication failure, abnormal weather, abnormal task, abnormal psychological characteristics, and abnormal physiological characteristics.

[0038] Optionally, when the target test variable is any one of engine malfunction, foreign object impact, communication failure, abnormal weather, abnormal task, abnormal psychological characteristics, and abnormal physiological characteristics, the electronic device may acquire the sub-feature vector corresponding to the target test variable. This may include: the electronic device acquiring first data corresponding to the target test variable; the electronic device performing denoising and normalization on the first data to obtain second data; and the electronic device performing single-source feature extraction and / or cross-source feature extraction on the second data to construct a structured array, which is an initial feature vector, including the sub-feature vector corresponding to the target test variable.

[0039] In this embodiment, after acquiring the first data corresponding to the target experimental variable, the electronic device can use algorithms such as median filtering and wavelet transform to remove environmental noise or physiological artifacts from the first data, and eliminate dimensional differences using deviation standardization or standard deviation standardization methods to obtain the second data. Then, the electronic device performs single-source feature extraction on the second data, specifically performing time-domain (e.g., mean, standard deviation), frequency-domain (e.g., power spectral density), or nonlinear (e.g., information entropy) analysis on a single variable, and / or performs cross-source feature extraction on the second data, specifically extracting coupling features between different variables through cross-correlation analysis (e.g., cross-correlation, synchronous compression transform). All extracted features are then concatenated according to a preset sequence to form a structured array, which is an initial feature vector. The entire process, through standardized preprocessing of multi-source experimental variables and cross-source coupling feature extraction, achieves a multi-dimensional and in-depth characterization of the underlying information affecting the pilot's stress state, providing a high-quality data foundation for subsequent accurate screening of causal features.

[0040] Optionally, the preset sequence may include at least one of the following: the order of physical dimensions, the order of time dimensions, and the combination rules of cross-source features, etc.

[0041] The physical dimensions are arranged in a fixed order, with features of different properties joined together at fixed positions. For example, [positions 1-5]: time-domain features (e.g., mean, standard deviation, range, skewness, kurtosis); [positions 6-10]: frequency-domain features (e.g., low-frequency energy, high-frequency energy, centroid frequency); [positions 11-15]: nonlinear features (e.g., sample entropy, approximate entropy).

[0042] The chronological order of time represents the arrangement of time steps. For example, arranging the physiological data at the current time t with the physiological data at the previous 5 times (t−1 to t−5) in reverse or forward order forms a sequence that reflects the dynamic trend of change.

[0043] The rules for combining cross-source features specify how "main variables" and "related variables" are arranged. For example, "engine thrust features" can be placed first, followed by the corresponding "pilot heart rate response features." This paired arrangement helps the model capture the coupling relationship between variables.

[0044] Optionally, the first data corresponding to the engine abnormality may include at least: engine speed, exhaust temperature, fuel flow rate, lubricating oil pressure, and vibration frequency and amplitude data of the engine casing.

[0045] Optionally, the first data corresponding to the foreign object impact may include at least: the impact pulse signal from the wing or fuselage strain gauge, the distance information of the unidentified space target scanned by radar, and the abnormal collision sound signal recorded in the cockpit audio and video.

[0046] Optionally, the first data corresponding to the communication failure may include at least: signal-to-noise ratio, data packet loss rate, silence duration, and instruction repetition rate or semantic inconsistency information detected by the speech recognition system.

[0047] Optionally, the first data corresponding to abnormal weather may include at least: atmospheric turbulence index, crosswind speed and direction changes, atmospheric visibility, wing icing thickness detection signal, and lightning activity intensity.

[0048] Optionally, the first data corresponding to the abnormal mission may include at least: the coordinates of the alternate airport, the emergency go-around instruction, the deviation between the expected trajectory and the actual trajectory of the conflict avoidance maneuver, and the remaining fuel / time margin of the mission.

[0049] Optionally, the first data corresponding to the abnormal physiological characteristics may include at least: electrocardiogram, original waveform, skin conduction current intensity, potential of each channel of electroencephalogram, and fixation point coordinates, saccade speed and pupil diameter data recorded by eye tracker.

[0050] By refining the first data composition corresponding to each experimental variable, it is ensured that the input features can comprehensively cover all dimensions of physical environment, system state and physiological response, providing original materials with high information entropy for causal screening mechanism, thereby improving the predictive model's ability to characterize complex flight conditions.

[0051] Step 103: Perform conditional independence tests on all sub-feature vectors, select sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, construct a causal feature set, and calculate the gain importance score of each sub-feature vector in the causal feature set when assessing the pilot's stress state.

[0052] In step 103, the electronic device performs conditional independence tests on each sub-feature vector in all sub-feature vectors, eliminates spurious correlations, and achieves causal structure learning. This allows it to select sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, thus constructing a causal feature set. Then, the electronic device utilizes a feature evaluation mechanism based on gradient boosting decision trees (such as Extreme Gradient Boosting (XGBoost) or Light Gradient Boosting Machine (LightGBM)) to calculate gain calculation logic and quantify contribution. This allows it to calculate the gain importance score of each sub-feature vector in the causal feature set when assessing the pilot's stress state. The entire process, through causal screening followed by gain quantification, effectively addresses the drawbacks of "black box prediction" in traditional machine learning models. Causal testing ensures the robustness of the prediction model (avoiding interference from irrelevant variables), while gain importance scoring ensures its efficiency (retaining the most discriminative features). This allows the constructed prediction model to maintain extremely high prediction accuracy and reliable attribution analysis capabilities even when facing novel and unseen emergencies.

[0053] In some embodiments, the electronic device performs a conditional independence test on each sub-feature vector among all sub-feature vectors, filters out sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, and constructs a causal feature set. This may include: the electronic device acquiring the observation distribution corresponding to the stress label; for each sub-feature vector among all sub-feature vectors, the electronic device uses the sub-feature vector as the intervention variable when assessing the pilot's stress state and the stress label as the result variable, and determines the intervention distribution corresponding to the sub-feature vector by controlling a preset set of confounding variables and performing backdoor adjustment; the electronic device filters out sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state based on the observation distribution and the intervention distribution corresponding to each sub-feature vector, and constructs a causal feature set.

[0054] The observed distribution refers to the statistical correlation actually observed, which may include "spurious correlations".

[0055] The confounding set of variables is used to describe and control for known confounding factors that simultaneously affect intervention and outcome variables in causal inference, such as ambient temperature, mission complexity, individual pilot differences (e.g., age, experience), simulator status, and basic visibility of the flight environment.

[0056] An intervention distribution is a distribution simulated using the do-operator, representing the true causal effect of how the outcome will change if the intervention variable changes.

[0057] In this embodiment, the electronic device constructs a joint feature-label distribution. Specifically, based on historical flight datasets, it statistically analyzes the probability distribution of stress labels at various levels (e.g., high, medium, low) under natural conditions (before intervention) to obtain the observed distribution corresponding to the stress label. ,in, The stress label is used to represent the stress level. Then, for each sub-feature vector among all sub-feature vectors, the electronic device uses the sub-feature vector as the intervention variable for assessing the pilot's stress state and the stress label as the outcome variable. To eliminate the influence of other confounding factors, the intervention distribution corresponding to the sub-feature vector can be calculated by controlling a preset set of confounding variables Z and performing backdoor adjustments. ,in, It also represents the outcome variable; This represents a sub-feature vector, which also represents the intervention variable. Based on this, the electronic device can obtain the corresponding intervention distribution for each of the sub-feature vectors. Finally, the electronic device analyzes the observation distribution and the corresponding intervention distribution for each of the sub-feature vectors, and calculates... ≠ To determine if a causal relationship exists between the selected features and the stress label when assessing pilot stress, a causal feature set is constructed, consisting of multiple sub-feature vectors. This process, employing a backdoor-based adjustment causal inference mechanism, effectively eliminates complex confounding interference in the flight environment, ensuring that the selected features are the true causes of pilot stress. This significantly improves the generalization accuracy of the prediction model under extreme and rare flight conditions.

[0058] It should be noted that, in ≠ If true, confirm the sub-feature vector. A causal relationship exists between the assessment of pilot stress state and stress labeling, and this sub-feature vector is preserved. .

[0059] In some embodiments, the electronic device uses a sub-feature vector as the intervention variable and a stress label as the outcome variable when assessing the pilot's stress state. By controlling a preset set of confounding variables and performing backdoor adjustments, the electronic device determines the intervention distribution corresponding to the sub-feature vector. This can include: the electronic device identifying a backdoor path between the intervention variable and the outcome variable based on the set of confounding variables; the electronic device determining a subset of confounding variables that needs to be controlled based on the backdoor path; and the electronic device determining the intervention distribution corresponding to the sub-feature vector based on the subset of confounding variables.

[0060] The backdoor path is a non-causal path that connects the intervention variable and the outcome variable, and the non-causal path begins with an arrow pointing to the intervention variable.

[0061] A subset of confounding variables refers to the smallest set of variables that must be observed and controlled within an identified backdoor path to block all non-causal associations. According to the backdoor criterion, this subset of confounding variables not only blocks the flow of pseudo-related information from the intervention variable through confounding factors to the outcome variable, but also does not contain any successor nodes of the intervention variable, thus ensuring the unbiasedness of causal effect calculation.

[0062] In this embodiment, during the process of determining the intervention distribution corresponding to the sub-feature vector, the electronic device first constructs a directed acyclic graph (DAG) reflecting the correlation between the flight environment, pilot characteristics, and stress state. Then, the electronic device uses a topology search algorithm (such as breadth-first search) to identify all non-causal paths (i.e., backdoor paths) from the DAG that originate from the intervention variable X and terminate at the outcome variable Y. Next, based on the d-separation principle, the electronic device filters out a subset of confounding variables from the aforementioned set of confounding variables that can block these backdoor paths. In practice, the electronic device adjusts according to the backdoor formula: For a subset of mixed variables Perform a full probability weighted average, where, Represents a subset of mixed variables One of the specific values. This means that the electronic device calculates the sub-feature vector. Corresponding to the stimulation tag When considering the impact, the distribution of the intervention will be observed under different mixed backgrounds (such as different age groups of flight crew members and different original visibility levels). With observed distribution The conditional probability distribution is then integrated according to the prior probability of the presence of mixed backgrounds. In this way, the electronic device can smooth out background differences, thereby accurately determining the intervention distribution corresponding to the sub-feature vector. The entire process, through the identification of backdoor paths and the control of confounding subsets of variables, not only identifies the numerical correlations between variables but also quantifies the marginal contribution of a single factor (such as weak engine vibration or specific psychological characteristics) to the pilot's stress state in complex aerospace coupled environments. This deep causal analysis effectively avoids erroneous attributions caused by the covariation of environmental factors (such as weather conditions) and pilot physiological performance, significantly enhancing the system's decision robustness under novel and unseen operating conditions.

[0063] In some embodiments, the electronic device, based on the observation distribution and the intervention distribution corresponding to each sub-feature vector, selects sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, and constructs a causal feature set. This may include: for each sub-feature vector, the electronic device uses the difference between the observation distribution and the intervention distribution corresponding to the sub-feature vector as an estimate of the causal effect of the probability of the pilot being in different stress states; performs a statistical significance test on the causal effect estimate to obtain a causal significance index for the sub-feature vector; if the causal significance index is less than a preset significance threshold, it confirms that the sub-feature vector has a causal relationship with the stress label when assessing the pilot's stress state; the electronic device constructs a causal feature set based on all sub-feature vectors that have a causal relationship with the stress label.

[0064] In this embodiment, the electronic device calculates the observation distribution for each sub-feature vector among all sub-feature vectors. With intervention distribution The electronic device calculates the statistical distance between the sub-feature vectors (e.g., Hellinger distance or the absolute value of the probability difference) and uses this distance as an estimate of the causal effect of the probability of the pilot being in different stress states. A larger causal effect estimate indicates a more significant actual impact of the intervention (variable change) on the stress state. Then, the electronic device performs a statistical significance test (e.g., Bootstrap test or t-test) on this causal effect estimate, calculating the probability that the causal effect is not caused by random sampling error, thus obtaining a causal significance index (e.g., p-value) for the sub-feature vector. This index is then compared to a preset significance threshold: if the index is greater than or equal to the threshold, it confirms that the sub-feature vector has no causal relationship with the stress label when assessing the pilot's stress state; if the index is less than the threshold, the null hypothesis is eliminated, confirming a statistically significant causal relationship between the sub-feature vector and the stress label when assessing the pilot's stress state. Finally, the electronic device aggregates all sub-feature vectors that pass the test to construct a causal feature set. The entire process is achieved by calculating the observed distribution. With intervention distribution By combining the bias with rigorous significance testing, a second stripping of "spurious correlation features" was achieved, ensuring that each sub-feature variable selected into the causal feature set has a substantial driving effect on the stress state, thereby supporting the high interpretability of the prediction model.

[0065] Among them, the p-value measures the observed effect. ≠ How likely is it due to randomness?

[0066] It should be noted that the observed distribution With intervention distribution The difference between these values ​​can quantify the extent to which the observed distribution of stress states would change if the sub-feature vector (such as engine vibration) were removed or altered. Furthermore, the causal features selected through significance testing in the entire process exhibit strong cross-sample stability. This means that the explanatory power of these features for stress states is reliable and repeatable across different flight phases, such as takeoff, cruise, or landing, and the model predictions will not fail due to minor perturbations in the sample data.

[0067] In some embodiments, the electronic device calculates the gain importance score of each sub-feature vector in the causal feature set when assessing the pilot's stress state. This may include: the electronic device using all sub-feature vectors in the causal feature set as model input and stress labels as model output to train an XGBoost classifier for preliminary assessment of the pilot's stress state; after the XGBoost classifier is trained, the electronic device traverses all decision trees constructed by the XGBoost classifier and calculates the information gain brought by each sub-feature vector in the causal feature set during the node splitting process of the decision tree to distinguish different stress states; the electronic device calculates the gain importance score of each sub-feature vector when assessing the pilot's stress state based on the sum or average of the information gains corresponding to each sub-feature vector in the causal feature set, and the gain importance score is used to characterize the discriminative contribution of each sub-feature vector in the XGBoost classifier.

[0068] Among them, each sub-feature vector in the causal feature set can also be called a causal feature.

[0069] In this embodiment, the electronic device uses all sub-feature vectors in the causal feature set as model input and stress labels as model output to train an XGBoost classifier for preliminary assessment of pilot stress state. During this process, multiple decision trees are iteratively generated by minimizing a loss function (such as logistic regression loss). After the XGBoost classifier is trained, the electronic device traverses all decision trees constructed by the XGBoost classifier and calculates the information gain of each sub-feature vector in the causal feature set during node splitting in all decision trees, aimed at reducing sample category uncertainty (distinguishing different stress states). Information gain represents the increase in the score of the left and right subtrees relative to the original node score after introducing the sub-feature at a node for splitting. Finally, the electronic device calculates the gain importance score of each sub-feature vector in assessing pilot stress state based on the sum or average of the information gains corresponding to each sub-feature vector in the full decision tree (i.e., the aforementioned XGBoost classifier). This gain importance score establishes the prediction weight (i.e., the magnitude of influence) of the corresponding sub-feature vector. The entire process accurately quantifies the prediction weights of each causal feature through the gradient boosting mechanism of the XGBoost classifier, realizing the scientific ranking of the causal influence of the pilot's stress state. While eliminating low-contribution redundant features, it provides a quantitative basis for achieving high interpretability and refined real-time attribution of the prediction model.

[0070] Step 104: Based on the sub-feature vectors corresponding to the gain importance scores that are greater than the preset scoring threshold, construct a prediction model corresponding to the flight scenario. The prediction model is used to evaluate the stress state of the pilot in the flight scenario and the attribution analysis results of the stress state.

[0071] In step 104, the electronic device compares the gain importance score of each sub-feature vector with a preset score threshold, eliminating weakly correlated causal features with low scores and retaining only key sub-feature vectors that significantly contribute to the prediction results. Specifically, it retains sub-feature vectors with gain importance scores greater than the preset score threshold (e.g., 0.05), simplifying the model input dimension, reducing onboard computing resource consumption, and improving the real-time performance of the prediction. Furthermore, these retained sub-feature vectors have a real causal impact on the pilot's stress state and contribute significantly to the prediction model. Next, the electronic device uses these retained sub-feature vectors as input and the stress state as a label to train and fine-tune the parameters of the original prediction model, obtaining a trained prediction model. Based on this, the electronic device can determine prediction models for multiple flight scenarios, namely, prediction models for the takeoff phase, cruise phase, and landing phase. The entire process, through feature simplification based on gain scoring thresholds and scenario-based model construction, not only ensures that the prediction model has extremely high real-time response speed in the computing-constrained airborne environment, but also breaks the invisibility of traditional black box models by outputting a combination of "prediction + attribution" results, providing transparent and scientific decision support for subsequent precise flight interventions (such as automated mission degradation or remote ground takeover).

[0072] It should be noted that, since the gain importance score quantifies the impact of each sub-feature vector on the prediction result, after assessing the pilot's stress state in the flight scenario, the electronic equipment can generate an attribution analysis result for the stress state in real time based on the gain importance score ratio of each sub-feature vector. For example, it can be clearly stated that the stress state was mainly induced by the sudden change in heart rate (contribution 60%) in "abnormal physiological characteristics" and the wind shear (contribution 30%) in "abnormal weather".

[0073] In this embodiment, the technical solution described in steps 101-104 above utilizes a dual screening mechanism of causal inference and gain importance to select key sub-feature vectors with high importance and real causal explanatory power for pilot stress states from multi-source features such as engine anomalies, foreign object impacts, communication failures, abnormal weather, abnormal missions, abnormal psychological characteristics, and abnormal physiological characteristics. This allows for the construction of a predictive model with strong interpretability and excellent generalization performance. Furthermore, dedicated predictive models are constructed for different flight scenarios, effectively overcoming prediction biases caused by differences in stress response patterns across scenarios, and achieving accurate and stable assessment of pilot stress states. In addition, attribution analysis of stress states can be output, clarifying the contribution of each key sub-feature vector to the prediction results, providing a reliable basis for subsequent refined pilot state monitoring and real-time operational assistance.

[0074] It should be noted that the execution entity involved in the embodiments of this application can be a pilot stress state prediction system or an electronic device. Optionally, the electronic device may include: a computer / laptop, a mobile terminal, a server, and an aircraft central control device, etc.

[0075] The following uses an electronic device as an example to illustrate in detail the pilot stress state prediction method provided in this application: Figure 2 This is a flowchart illustrating the pilot stress state prediction method provided in an embodiment of this application. Figure 2 As shown, the method may include the following steps 201-202.

[0076] Step 201: Given the current flight scenario, construct a test feature vector with the same structure as the initial feature vector, based on the external stressor characteristics of the aircraft and the internal psychological and physiological characteristics of the pilot. The initial feature vector is determined based on multiple experimental variables under the current flight scenario.

[0077] The current flight scenario is one of the following: takeoff, cruise, and landing.

[0078] Optionally, the characteristics of external stressors may include at least objective environmental factors that cause pilots to experience mission stress or safety concerns, specifically including abnormal engine signals (such as instantaneous thrust drop), foreign object impact monitoring data, communication failure link indicators, abnormal meteorological parameters (such as wind shear intensity), and abnormal mission instructions (such as diversion or emergency go-around).

[0079] Optionally, internal psychological characteristics may include at least anomalous psychological characteristics, such as attention narrowing index, cognitive processing delay time, spatial orientation stability score, and logical error rate in task execution.

[0080] Optionally, the internal physiological characteristics may include at least abnormal physiological characteristics, such as time-frequency domain indices of heart rate variability, arousal peak of skin conductance response, power spectrum ratio of electroencephalogram, and disorder of eye saccade pathways.

[0081] It should be noted that the tested feature vector has the same structure as the initial feature vector, indicating that the tested feature vector is highly consistent with the initial feature vector used in the model training phase in terms of the number of feature dimensions, the order of the feature dimensions (i.e., following the same preset sequence), and the normalization scale of the data. This structural alignment ensures that the real-time collected multi-source heterogeneous data can be accurately parsed by the prediction model, avoiding prediction failures caused by feature misalignment or inconsistencies in units, thereby guaranteeing the real-time performance and accuracy of stress state assessment.

[0082] Furthermore, when some features of the acquired external stressor features, internal psychological features, and internal physiological features are missing, the electronic device compensates for the missing dimensions through time-series interpolation, mean filling, or preset placeholders to ensure that the generated feature vector to be tested is completely isomorphic to the initial feature vector in terms of spatial dimension and arrangement order, thereby achieving robust analysis of incomplete input samples by the prediction model.

[0083] In step 201, by constructing a test feature vector that is completely isomorphic to the training phase, a standardized mapping from multi-source raw signals to the model input is achieved, ensuring that the predefined causal relationships and weight parameters can be directly applied to the current flight conditions, and providing a data structure guarantee for achieving accurate assessment of pilot stress state at the second level.

[0084] Step 202: Input the feature vector to be tested into the prediction model corresponding to the current flight scenario to obtain the pilot's stress state and the attribution analysis results of the stress state. The prediction model is based on, for example... Figure 1 The pilot stress state prediction model is constructed using the method shown.

[0085] In step 202, the electronic device inputs the feature vector to be tested as input data into the prediction model corresponding to the current flight scenario. The prediction model performs a nonlinear mapping operation on the input data and outputs the pilot's current prediction result (i.e., the pilot's stress state) in real time. Simultaneously with outputting the stress state, the electronic device uses the gain importance score recorded internally by the prediction model to trace the contribution of each dimension of the feature vector to be tested. Specifically, it calculates which abnormal fluctuations in feature factors led to the stress state, obtains the attribution analysis results for the stress state, and generates a comprehensive report containing the prediction results and attribution analysis results. For example, the comprehensive report shows: "Current stress state: High; Dominant attribution: Abnormal weather - wind shear (contribution 65%), internal physiological characteristics - heart rate surge (contribution 25%)". This entire process, by inputting the structure-aligned feature vector to be tested into the scenario-based prediction model, not only achieves millisecond-level accurate determination of the stress state, but more importantly, by synchronously outputting the attribution analysis results, it solves the problem of traditional artificial intelligence "knowing what but not why". This provides pilots with a clear direction for psychological / operational adjustments, and also provides a transparent and reliable logical basis for airborne automated safety systems to perform targeted interventions (such as automatically adjusting flight parameters to compensate for pilot cognitive lag).

[0086] In this embodiment of the application, the technical solution described in steps 201-202 above achieves instantaneous and accurate assessment of stress states and in-depth tracing of causes by constructing a test feature vector isomorphic to the training phase in real time and inputting it into a scenario-based prediction model. This provides a quantitative basis with high real-time performance and high interpretability for human-machine co-pilot decision-making in complex flight environments.

[0087] The pilot stress state prediction model construction system provided in the embodiments of this application is described below. The pilot stress state prediction model construction system described below and the pilot stress state prediction model construction method described above can be referred to in correspondence.

[0088] Figure 3 This is a schematic diagram of the pilot stress state prediction model construction system provided in this application embodiment. Figure 3 As shown, the system includes a vector acquisition module 301 and a model building module 302.

[0089] The vector acquisition module 301 is used to determine multiple flight scenarios, define corresponding stress labels for each flight scenario, and determine multiple experimental variables associated with the pilot's stress state. These multiple experimental variables include at least engine anomaly, foreign object impact, communication failure, abnormal weather, abnormal mission, abnormal psychological characteristics, and abnormal physiological characteristics. The module acquires the sub-feature vectors corresponding to each of the following: engine anomaly, foreign object impact, communication failure, abnormal weather, abnormal mission, abnormal psychological characteristics, and abnormal physiological characteristics. The model building module 302 is used to perform conditional independence tests on each sub-feature vector in all sub-feature vectors, screen out the sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, and construct a causal feature set; and calculate the gain importance score of each sub-feature vector in the causal feature set when assessing the pilot's stress state; based on the sub-feature vectors corresponding to gain importance scores greater than a preset score threshold, construct a prediction model corresponding to the flight scenario, which is used to assess the pilot's stress state in the flight scenario and the attribution analysis results of the stress state.

[0090] Optionally, the model building module 302 is specifically used to obtain the observation distribution corresponding to the stress label; for each sub-feature vector among all the sub-feature vectors, taking the sub-feature vector as the intervention variable when assessing the pilot's stress state and the stress label as the outcome variable, by controlling a preset set of confounding variables and performing backdoor adjustment, the intervention distribution corresponding to the sub-feature vector is determined. The set of confounding variables is used to describe and control known interference factors that simultaneously affect the intervention variable and the outcome variable in causal inference; based on the observation distribution and the intervention distribution corresponding to each sub-feature vector, the sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state are selected, and the causal feature set is constructed.

[0091] Optionally, the model building module 302 is specifically used to identify, based on the set of confounding variables, a backdoor path between the intervention variable and the outcome variable, wherein the backdoor path is a non-causal path connecting the intervention variable and the outcome variable, and the non-causal path begins with an arrow pointing to the intervention variable; based on the backdoor path, determine a subset of confounding variables that need to be controlled; and based on the subset of confounding variables, determine the intervention distribution corresponding to the sub-feature vector.

[0092] Optionally, the model building module 302 is specifically used to take the difference between the observed distribution and the intervention distribution corresponding to the sub-feature vector as the causal effect estimate of the probability of the pilot being in different stress states for each sub-feature vector; perform a statistical significance test on the causal effect estimate to obtain the causal significance index of the sub-feature vector; if the causal significance index is less than a preset significance threshold, confirm that the sub-feature vector has a causal relationship with the stress label when assessing the pilot's stress state; and construct the causal feature set based on all sub-feature vectors that have a causal relationship with the stress label.

[0093] Optionally, the model building module 302 is specifically used to train an XGBoost classifier for preliminary evaluation of the pilot's stress state by using all sub-feature vectors in the causal feature set as model input and the stress label as model output; after the XGBoost classifier is trained, it traverses all decision trees constructed by the XGBoost classifier and calculates the information gain brought by each sub-feature vector in the causal feature set to distinguish different stress states during the node splitting process of the decision tree; based on the sum or average of the information gains corresponding to each sub-feature vector in the causal feature set, it calculates the gain importance score of each sub-feature vector in evaluating the pilot's stress state, and the gain importance score is used to characterize the discriminative contribution of each sub-feature vector in the XGBoost classifier.

[0094] The pilot stress state prediction system provided in the embodiments of this application is described below. The pilot stress state prediction system described below can be referred to in correspondence with the pilot stress state prediction method described above.

[0095] Figure 4 This is a schematic diagram of the pilot stress state prediction system provided in an embodiment of this application. Figure 4 As shown, the system includes a vector construction module 401 and a state prediction module 402.

[0096] The vector construction module 401 is used to construct a test feature vector with the same structure as the initial feature vector, based on the external stressor characteristics of the aircraft and the internal psychological and physiological characteristics of the pilot, when the current flight scenario is determined. The initial feature vector is determined based on multiple experimental variables under the current flight scenario. State prediction module 402 is used to input the feature vector to be measured into the prediction model corresponding to the current flight scenario to obtain the pilot's stress state and the attribution analysis results of the stress state. The prediction model is based on, for example, Figure 1 The pilot stress state prediction model is constructed using the method shown.

[0097] Figure 5 This is a schematic diagram of the structure of the electronic device provided in an embodiment of this application. For example... Figure 5As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other through the communications bus 540. The processor 510 can call logical instructions in the memory 530 to execute a method for constructing a pilot stress state prediction model. This method includes: determining multiple flight scenarios; for each flight scenario, defining corresponding stress labels and determining multiple experimental variables associated with the pilot's stress state, wherein the multiple experimental variables include at least engine anomaly, foreign object impact, communication failure, abnormal weather, abnormal mission, abnormal psychological characteristics, and abnormal physiological characteristics; acquiring sub-feature vectors corresponding to each of the engine anomaly, foreign object impact, communication failure, abnormal weather, abnormal mission, abnormal psychological characteristics, and abnormal physiological characteristics; performing conditional independence tests on each sub-feature vector in all sub-feature vectors, filtering out sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, and constructing a causal feature set; calculating the gain importance score of each sub-feature vector in the causal feature set when assessing the pilot's stress state; and constructing a prediction model corresponding to the flight scenario based on the sub-feature vectors whose gain importance scores are greater than a preset score threshold. The prediction model is used to assess the pilot's stress state in the flight scenario and the attribution analysis results of the stress state. Alternatively, a pilot stress state prediction method may be implemented, comprising: given a current flight scenario, constructing a test feature vector with the same structure as an initial feature vector based on the external stressor characteristics of the aircraft and the internal psychological and physiological characteristics of the pilot, wherein the initial feature vector is determined based on multiple experimental variables under the current flight scenario; inputting the test feature vector into a prediction model corresponding to the current flight scenario to obtain the pilot's stress state and the attribution analysis results of the stress state, wherein the prediction model is based on the pilot stress state prediction model construction method described above.

[0098] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0099] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a method for constructing a pilot stress state prediction model. This method includes: determining multiple flight scenarios; defining corresponding stress labels for each flight scenario and determining multiple experimental variables associated with the pilot's stress state, wherein the multiple experimental variables include at least engine anomaly, foreign object impact, communication failure, abnormal weather, abnormal mission, abnormal psychological characteristics, and abnormal physiological characteristics; and acquiring the engine anomaly, the foreign object impact, the communication failure, and the abnormal... Sub-feature vectors corresponding to weather, the abnormal task, the abnormal psychological characteristics, and the abnormal physiological characteristics are generated. Conditional independence tests are performed on each sub-feature vector to select those that have a causal relationship with the stress label when assessing the pilot's stress state, thus constructing a causal feature set. The gain importance score of each sub-feature vector in the causal feature set when assessing the pilot's stress state is calculated. Based on the sub-feature vectors corresponding to gain importance scores greater than a preset score threshold, a prediction model corresponding to the flight scenario is constructed. This prediction model is used to assess the pilot's stress state in the flight scenario and the attribution analysis results of the stress state. Alternatively, a pilot stress state prediction method may be implemented, comprising: given a current flight scenario, constructing a test feature vector with the same structure as an initial feature vector based on the external stressor characteristics of the aircraft and the internal psychological and physiological characteristics of the pilot, wherein the initial feature vector is determined based on multiple experimental variables under the current flight scenario; inputting the test feature vector into a prediction model corresponding to the current flight scenario to obtain the pilot's stress state and the attribution analysis results of the stress state, wherein the prediction model is based on the pilot stress state prediction model construction method described above.

[0100] In another aspect, embodiments of this application also provide a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, this computer program implements a method for constructing a pilot stress state prediction model. The method includes: determining multiple flight scenarios; for each flight scenario, defining corresponding stress labels and determining multiple experimental variables associated with the pilot's stress state, wherein the multiple experimental variables include at least engine anomaly, foreign object impact, communication failure, abnormal weather, abnormal mission, abnormal psychological characteristics, and abnormal physiological characteristics; and acquiring the engine anomaly, the foreign object impact, the communication failure, the abnormal weather, the abnormal mission, and the abnormal... Sub-feature vectors corresponding to psychological characteristics and abnormal physiological characteristics are generated. Conditional independence tests are performed on each sub-feature vector to identify those that have a causal relationship with the stress label when assessing the pilot's stress state, thus constructing a causal feature set. Gain importance scores are calculated for each sub-feature vector in the causal feature set when assessing the pilot's stress state. Based on the sub-feature vectors corresponding to gain importance scores greater than a preset threshold, a prediction model for the flight scenario is constructed. This prediction model is used to assess the pilot's stress state in the flight scenario and the attribution analysis results for the stress state. Alternatively, a pilot stress state prediction method may be implemented, comprising: given a current flight scenario, constructing a test feature vector with the same structure as an initial feature vector based on the external stressor characteristics of the aircraft and the internal psychological and physiological characteristics of the pilot, wherein the initial feature vector is determined based on multiple experimental variables under the current flight scenario; inputting the test feature vector into a prediction model corresponding to the current flight scenario to obtain the pilot's stress state and the attribution analysis results of the stress state, wherein the prediction model is based on the pilot stress state prediction model construction method described above.

[0101] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0102] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for constructing a pilot stress state prediction model, characterized in that, include: Multiple flight scenarios are identified, and for each flight scenario, corresponding stress labels are defined and multiple experimental variables associated with the pilot's stress state are determined. The multiple experimental variables include at least engine abnormality, foreign object impact, communication failure, abnormal weather, abnormal mission, abnormal psychological characteristics, and abnormal physiological characteristics. Obtain the sub-feature vectors corresponding to the engine anomaly, the foreign object impact, the communication failure, the abnormal weather, the abnormal task, the abnormal psychological characteristics, and the abnormal physiological characteristics respectively; Perform a conditional independence test on each of the sub-feature vectors in all sub-feature vectors, and select the sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, and construct a causal feature set; and calculate the gain importance score of each sub-feature vector in the causal feature set when assessing the pilot's stress state. Based on the sub-feature vectors corresponding to the gain importance scores that are greater than a preset scoring threshold, a prediction model corresponding to the flight scenario is constructed. The prediction model is used to evaluate the stress state of the pilot in the flight scenario and the attribution analysis results of the stress state.

2. The method for constructing a pilot stress state prediction model according to claim 1, characterized in that, The conditional independence test is performed on each sub-feature vector in all sub-feature vectors to filter out the sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, and to construct a causal feature set, including: Obtain the observation distribution corresponding to the stress labels; For each of the sub-feature vectors, the sub-feature vector is used as the intervention variable when assessing the pilot's stress state, and the stress label is used as the outcome variable. By controlling a preset set of confounding variables and performing backdoor adjustment, the intervention distribution corresponding to the sub-feature vector is determined. The set of confounding variables is used to describe and control known confounding factors that simultaneously affect the intervention variable and the outcome variable in causal inference. Based on the observation distribution and the intervention distribution corresponding to each sub-feature vector, sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state are selected, and the causal feature set is constructed.

3. The method for constructing a pilot stress state prediction model according to claim 2, characterized in that, The process of determining the intervention distribution corresponding to the sub-feature vector by using the sub-feature vector as the intervention variable for assessing the pilot's stress state and the stress label as the outcome variable, through controlling a preset set of confounding variables and performing backdoor adjustments, includes: Based on the set of confounding variables, a backdoor path is identified between the intervention variable and the outcome variable. The backdoor path is a non-causal path connecting the intervention variable and the outcome variable, and the non-causal path begins with an arrow pointing to the intervention variable. Based on the backdoor path, determine the subset of promiscuous variables that need to be controlled; Based on the subset of confounding variables, determine the intervention distribution corresponding to the sub-feature vector.

4. The method for constructing a pilot stress state prediction model according to claim 2, characterized in that, The step involves selecting sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, based on the observed distribution and the intervention distribution corresponding to each sub-feature vector, and constructing the causal feature set, including: For each sub-feature vector, the difference between the observed distribution and the intervention distribution corresponding to the sub-feature vector is used as the causal effect estimate of the probability of the pilot being in different stress states; the causal effect estimate is subjected to a statistical significance test to obtain the causal significance index of the sub-feature vector; if the causal significance index is less than a preset significance threshold, it is confirmed that the sub-feature vector has a causal relationship with the stress label when assessing the pilot's stress state; The causal feature set is constructed based on all sub-feature vectors that have a causal relationship with the stress label.

5. The method for constructing a pilot stress state prediction model according to any one of claims 1-4, characterized in that, The calculation of the gain importance score of each sub-feature vector in the causal feature set when assessing the pilot's stress state includes: Using all sub-feature vectors in the causal feature set as model input and the stress label as model output, an XGBoost classifier is trained to preliminarily assess the pilot's stress state. After the XGBoost classifier is trained, all decision trees constructed by the XGBoost classifier are traversed, and the information gain brought by each sub-feature vector in the causal feature set to distinguish different stress states during the node splitting process of the decision tree is calculated. Based on the sum or average of the information gains corresponding to each sub-feature vector in the causal feature set, the gain importance score of each sub-feature vector in evaluating the pilot's stress state is calculated. The gain importance score is used to characterize the discriminative contribution of each sub-feature vector to the XGBoost classifier.

6. A method for predicting pilot stress states, characterized in that, include: Given the current flight scenario, a test feature vector with the same structure as the initial feature vector is constructed based on the external stressor characteristics of the aircraft and the internal psychological and physiological characteristics of the pilot. The initial feature vector is determined based on multiple experimental variables under the current flight scenario. The feature vector to be tested is input into the prediction model corresponding to the current flight scenario to obtain the pilot's stress state and the attribution analysis results of the stress state. The prediction model is obtained based on the pilot stress state prediction model construction method as described in any one of claims 1-5.

7. A system for constructing a pilot stress state prediction model, characterized in that, include: The vector acquisition module is used to determine multiple flight scenarios. For each flight scenario, it defines corresponding stress labels and determines multiple experimental variables associated with the pilot's stress state. The multiple experimental variables include at least engine abnormality, foreign object impact, communication failure, abnormal weather, abnormal mission, abnormal psychological characteristics, and abnormal physiological characteristics. Obtain the sub-feature vectors corresponding to the engine anomaly, the foreign object impact, the communication failure, the abnormal weather, the abnormal task, the abnormal psychological characteristics, and the abnormal physiological characteristics respectively; The model building module performs conditional independence tests on all sub-feature vectors, filters out sub-feature vectors that have a causal relationship with the stress label when assessing the pilot's stress state, and constructs a causal feature set; it also calculates the gain importance score of each sub-feature vector in the causal feature set when assessing the pilot's stress state; and constructs a prediction model corresponding to the flight scenario based on the sub-feature vectors whose gain importance scores are greater than a preset score threshold. The prediction model is used to assess the pilot's stress state in the flight scenario and the attribution analysis results of the stress state.

8. A pilot stress state prediction system, characterized in that, include: The vector construction module is used to construct a test feature vector with the same structure as the initial feature vector, based on the external stressor characteristics of the aircraft and the internal psychological and physiological characteristics of the pilot, given the current flight scenario. The initial feature vector is determined based on multiple experimental variables under the current flight scenario. The state prediction module is used to input the feature vector to be tested into the prediction model corresponding to the current flight scenario to obtain the stress state of the pilot and the attribution analysis results of the stress state. The prediction model is obtained based on the pilot stress state prediction model construction method as described in any one of claims 1-5.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the pilot stress state prediction model construction method as described in any one of claims 1 to 5, or the pilot stress state prediction method as described in claim 6.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the pilot stress state prediction model construction method as described in any one of claims 1 to 5, or the pilot stress state prediction method as described in claim 6.

Images