A data-driven flight risk assessment method

Abstract

The application discloses a kind of based on data-driven flight risk assessment method, the method collects the multidimensional flight state parameters in the process of aircraft flight, input flight state depth clustering network obtains the flight state data set of different clustering categories;The entropy weight value and the closeness to ideal solution of each clustering category are calculated using entropy weight approximation ideal solution sequencing algorithm;Based on aviation collision risk model, construct flight risk assessment initial model framework in combination with entropy weight value and closeness;The clustering result and related numerical value are substituted into initial model framework to train and optimize;Real-time flight state parameters are processed using the optimized model, and the risk assessment result of the current flight phase is output.The method classifies flight state by refinement, scientifically determines parameter weight and fuses multiple technologies, improves the evaluation accuracy and reliability, and can effectively guarantee the flight safety of aircraft.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of flight risk assessment, and particularly relates to a flight risk assessment method based on data driving. BACKGROUND

[0002] With the rapid development of the aviation transportation industry, the safety of aircraft flight is increasingly valued, and the demand for precise and efficient flight risk assessment methods is urgent. Currently, the flight process of an aircraft involves a large number of complex flight state parameters, including flight path, aircraft type, engine operation, navigation and environment, and other multi-dimensional information. The dynamic changes of these parameters directly affect the flight risk level. Traditional flight risk assessment relies on experience or simple models, which are difficult to fully integrate multi-source parameters and capture their complex correlations. The rise of data-driven technology provides a possible solution to this problem. Through the combination of flight state deep clustering network, entropy weight approximation ideal solution sorting algorithm and aviation collision risk model, more accurate risk assessment is expected to be achieved.

[0003] The existing technology has two obvious shortcomings in flight risk assessment. On the one hand, the existing method does not classify the flight state finely, and mostly uses a simple division method, which fails to fully explore the internal differences between different flight states, resulting in insufficient pertinence of risk assessment and difficulty in accurately reflecting the risk characteristics of different clustering categories of flight states. On the other hand, the existing model lacks scientificity in determining the weight of each parameter when integrating multi-dimensional parameters for risk assessment, and fails to effectively combine the information entropy characteristics of the parameters and the closeness to the ideal safe state, making the model insufficiently sensitive to key risk factors and the reliability and accuracy of the assessment results need to be improved. SUMMARY

[0004] In order to overcome the shortcomings and deficiencies of the existing technology, the present application provides a flight risk assessment method based on data driving.

[0005] The technical solution adopted by the present application is a flight risk assessment method based on data driving, comprising:

[0006] S1. Collect multi-dimensional flight state parameters generated by the aircraft during flight. These parameters include flight path parameters, aircraft configuration parameters, engine parameters, navigation parameters, and environmental parameters. Flight path parameters include altitude, airspeed, vertical speed, heading, pitch angle, roll angle, and yaw angle. Aircraft configuration parameters include flap position, landing gear status, and spoiler deployment status. Engine parameters include engine operating status. Navigation parameters include lateral and vertical deviations from the expected flight path. Environmental parameters include wind speed, wind direction, temperature, and atmospheric pressure. S2. Input the collected flight state parameters into a deep clustering network for flight state analysis. This network performs clustering analysis on the flight state to obtain different cluster categories. S3. Process the flight status datasets of different cluster categories using the entropy weight approximation ideal solution sorting algorithm, and calculate the entropy weight of the flight status of each cluster category and its closeness to the ideal solution; S4. Based on the aviation collision risk model, combine the obtained entropy weight and closeness to construct the initial model framework for flight risk assessment; S5. Substitute the clustering results output by the flight status deep clustering network and the relevant values ​​calculated by the entropy weight approximation ideal solution sorting algorithm into the initial model framework to train and optimize the flight risk assessment model; S6. Use the trained and optimized flight risk assessment model to process the real-time flight status parameters of the aircraft and output the risk assessment results of the aircraft in the current flight phase.

[0007] Furthermore, in step S3, when calculating the entropy weight of the flight state of each cluster category and its closeness to the ideal solution using the entropy weight approximation ideal solution sorting algorithm, the following model formula is adopted: ,in, w j Indicates the first j Entropy weights of each flight state parameter, , , x ij For the first i In the cluster category, the first j Standardized values ​​of each flight state parameter m For the number of cluster categories, n The number of flight status parameters; ,in, C j Indicates the first i The closeness of cluster categories, , , For the first j The optimal values ​​of each parameter For the first j The worst value of each parameter.

[0008] Furthermore, in step S4, when constructing the initial model framework for flight risk assessment, the following model formula is adopted based on the aviation collision risk model combined with entropy weights and proximity: ,in, R Indicates the flight risk value. α , β These are the weighting coefficients. f ( v ij ) is the first i In the cluster category, the first j Risk impact function of each flight state parameter v ij For the first i In the cluster category, the first j The actual values ​​of each flight status parameter g ( d ij ) is the risk impact function of the distance between aircraft. d ij For the first i aircraft and the first j The real-time distance between aircraft.

[0009] Furthermore, in step S5, during the training and optimization of the flight risk assessment model, the model parameters are adjusted based on the clustering results output by the flight state deep clustering network. The relevant model formula is as follows: ,in, θ k+1 For the first k Model parameters for +1 iteration θ k For the first k Model parameters for the next iteration γ For learning rate, For loss function L ( θ k )exist θ k gradient at, , R i This represents the actual risk value. To predict risk values ​​for the model, c i For the first The weight of each sample is related to the clustering result output by the flight state deep clustering network.

[0010] Furthermore, in step S6, when outputting the risk assessment results of the aircraft in the current flight phase, the following model formula is used to classify the risk level: Where Level indicates the risk level.T 1. T 2 represents the threshold for risk level classification. The flight risk value is calculated by combining the initial model framework constructed in step S4 with the parameters trained and optimized in step S5, and is related to the clustering category output by the flight state deep clustering network and the closeness obtained by the entropy weight approximation ideal solution sorting algorithm.

[0011] Furthermore, in step S2, the model formula used by the flight state deep clustering network to perform clustering analysis on the flight state is: ,in, z i For the first i Latent feature representation of a set of flight state parameters h (·) is the mapping function for the flight state deep clustering network. θ For network parameters, x i For the first i A set of flight state parameters; during clustering, through calculation Determine sample attribution, where, q ij For the first i The sample belongs to the first j The probability of a cluster center u j For the first j Cluster centers, K This represents the number of clusters.

[0012] Further, step S3 includes the following sub-steps: S3.1, standardizing the flight state data sets of different cluster categories output by the flight state deep clustering network, converting the values ​​of each flight state parameter to the same data magnitude, and eliminating the impact of differences in the units of different parameters; S3.2, based on the standardized flight state data, calculating the information entropy of each flight state parameter in each cluster category, using the information entropy to reflect the dispersion of the parameter and its impact on risk assessment; S3.3, based on the calculated information entropy, determining the entropy weight of each flight state parameter according to the entropy weight calculation rules, the magnitude of the entropy weight is positively correlated with the importance of the parameter in risk assessment; S3.4, constructing a weighted decision matrix based on the entropy weight of each parameter, and obtaining the proximity of each cluster category to the ideal solution and the worst solution by calculating the distance between each cluster category and the ideal solution, the proximity is used to measure the degree of closeness between the flight state of the cluster category and the ideal safe state.

[0013] Further, step S4 includes the following sub-steps: S4.1, obtaining the structure and parameters of the aviation collision risk model, which includes the correlation between aircraft position, speed, heading parameters and collision probability; S4.2, using the entropy weights and proximity of the flight states of each cluster category obtained in step S3 as input variables, integrating them into the structure of the aviation collision risk model to establish new model variable relationships; S4.3, determining the weight coefficients of each input variable in the model, the setting of which is based on the actual needs of flight risk assessment and the historical statistical results of the impact of each variable on risk; S4.4, integrating the variable relationships and weight coefficients to form an initial model framework for flight risk assessment, which can receive the results of flight state parameters after preprocessing and output the corresponding intermediate risk assessment values.

[0014] Further, step S5 includes the following sub-steps: S5.1, using the flight state data corresponding to the clustering results output by the flight state deep clustering network, the entropy weights calculated by the entropy weight approximation ideal solution sorting algorithm, and the closeness as training sample data, dividing them into training set and validation set; S5.2, inputting the training set data into the initial model framework constructed in step S4, calculating the model's predicted output through forward propagation, and comparing it with the actual risk label to obtain the loss value; S5.3, based on the loss value, using the backpropagation algorithm to adjust the model parameters, and monitoring the model performance in conjunction with the validation set data during the adjustment process; S5.4, repeating the forward and backpropagation parameter adjustment process of the training set until the model's performance on the validation set reaches the preset standard, completing the training and optimization of the flight risk assessment model.

[0015] Beneficial Effects: This invention proposes a data-driven flight risk assessment method. Addressing the shortcomings of existing methods, such as coarse flight state classification and insufficient targeting, this method employs a deep flight state clustering network to finely cluster multi-dimensional flight state parameters, fully exploring the inherent differences between different flight states. This allows the risk assessment to accurately reflect the risk characteristics of flight states in different cluster categories. For existing models that lack scientific rigor in parameter weight determination and are insufficiently sensitive to key risk factors, this method combines an entropy weight approximation ideal solution ranking algorithm to calculate the entropy weight and proximity of each parameter, scientifically determining parameter weights and making the model more sensitive to key risk factors, thus improving the reliability and accuracy of the assessment results. Furthermore, by combining the deep flight state clustering network, the entropy weight approximation ideal solution ranking algorithm, and the aviation collision risk model to construct an assessment model, this method achieves comprehensive integration of multi-source parameters and captures complex correlations, significantly improving the accuracy and efficiency of flight risk assessment and providing stronger protection for aircraft flight safety. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method steps of the present invention;

[0017] Figure 2 This is a diagram showing the unit composition for implementing the method of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, a data-driven flight risk assessment method includes:

[0020] Step S1: Collect multi-dimensional flight status parameters generated by the aircraft during flight. These parameters include flight path parameters, aircraft configuration parameters, engine parameters, navigation parameters, and environmental parameters. The flight path parameters include altitude, airspeed, vertical speed, heading, pitch angle, roll angle, and yaw angle. The aircraft configuration parameters include flap position, landing gear status, and spoiler deployment status. The engine parameters include engine operating status. The navigation parameters include lateral and vertical deviations from the expected flight path. The environmental parameters include wind speed, wind direction, temperature, and air pressure altitude.

[0021] Specifically, the collection of multi-dimensional flight status parameters is the primary step in flight risk assessment, and its completeness and accuracy directly determine the effectiveness of subsequent assessments. These parameters cover multiple key areas, including flight path, aircraft configuration, engine, navigation, and environment. Flight path parameters reflect the aircraft's spatial motion state; aircraft configuration parameters demonstrate the impact of the aircraft's structural configuration on flight performance; engine parameters characterize the operating status of the power system; navigation parameters show the degree to which the aircraft deviates from its planned route; and environmental parameters reflect the interference of external weather conditions on flight. By collecting these parameters, the state characteristics of the aircraft at different flight stages can be comprehensively captured, providing raw data support for subsequent cluster analysis, weight calculation, and model building, thus forming the foundation for accurate risk assessment.

[0022] In practical implementation, it is necessary to first clarify the specific indicators and value ranges of each dimension parameter. Among the flight path parameters, the altitude is set from -100 meters to 15,000 meters, which can cover the entire stage of ground taxiing (-100 meters to 0 meters), takeoff (0 meters to 3,000 meters), cruise (3,000 meters to 12,000 meters), and landing (12,000 meters to 0 meters); the airspeed range is 0 to 900 kilometers per hour, which is suitable for the takeoff, cruise, and landing speed requirements of different aircraft types; the vertical speed is between -5,000 meters per minute and 5,000 meters per minute, with negative values ​​indicating descent and positive values ​​indicating ascent; the heading is a circular angle from 0 to 360 degrees; the pitch angle range is -20 degrees to 30 degrees, which is used to describe the degree of aircraft pitching up or down; the roll angle is between -45 degrees and 45 degrees, which reflects the aircraft's tilt state; and the yaw angle is between -10 degrees and 10 degrees, which reflects the aircraft's deviation from the heading to the left or right. In the aircraft configuration parameters, flap positions are divided into different levels such as 0 degrees, 10 degrees, 20 degrees, 30 degrees, and 40 degrees, corresponding to different lift requirements; the landing gear status has only two discrete states: retracted and extended; the spoiler deployment status is expressed as a percentage from 0% to 100%, used to adjust drag. Engine parameters mainly record operating states such as idle, climb, cruise, and maximum thrust. In navigation parameters, lateral deviation is controlled within -500 meters to 500 meters, and vertical deviation within -100 meters to 100 meters, used to measure the degree of deviation from the flight path. Regarding environmental parameters, wind speed is set to 0 to 60 meters per second, wind direction to 0 to 360 degrees, temperature to -50 degrees Celsius to 50 degrees Celsius, and air pressure altitude is adjusted in real time according to altitude. The data collection is carried out using high-precision sensors mounted on the aircraft. The collection frequency for core parameters such as altitude and airspeed is set to 10 times per second, and environmental parameters are collected once per minute. During the collection process, outliers are removed through a preset threshold judgment mechanism, such as data where the altitude suddenly exceeds the range of -100 meters to 15,000 meters. Finally, the filtered parameters are stored in the database according to timestamps and categories to prepare data for the cluster analysis in step S2.

[0023] Step S2: Input the collected flight status parameters into the flight status deep clustering network, and perform cluster analysis on the flight status through the network to obtain flight status data sets of different cluster categories;

[0024] Specifically, deep clustering networks for flight states perform clustering analysis on collected parameters, which is the core technology for achieving refined classification of flight states. This network utilizes the multi-layer nonlinear mapping capabilities of deep neural networks to extract hidden feature patterns from high-dimensional flight parameters, identifying subtle differences between different flight states. Compared to traditional clustering methods, it exhibits stronger classification accuracy and robustness. By grouping flight state datasets of different categories obtained through clustering, flight states with similar characteristics are grouped together. This allows subsequent risk assessments to employ consistent analytical logic for similar states, avoiding assessment biases caused by mixed states. This improves the targeting and accuracy of risk assessments and provides a structured data foundation for subsequent weight calculations and model construction.

[0025] In practice, the parameters collected and stored in step S1 are first formatted and constructed into multi-dimensional arrays according to the categories of flight path, aircraft configuration, engine, navigation, and environment. Each array element includes a parameter name, value, and timestamp. Then, a deep clustering network architecture for flight status is built. The number of neurons in the input layer is consistent with the parameter dimension (e.g., 7 parameters for flight path, 3 parameters for aircraft configuration, etc., totaling 16 parameters, corresponding to 16 input neurons). The hidden layer consists of three fully connected layers: the first layer has 32 neurons, the second layer has 16 neurons, and the third layer has 8 neurons, all using the ReLU activation function for non-linear feature transformation. The number of neurons in the output layer is determined according to the preset clustering categories. For example, if divided into four categories—ground taxiing, takeoff, cruise, and landing—there are four output neurons, using the Softmax activation function to output the probability values ​​for each category. During the network training phase, 100,000 sets of historical flight data from the past 12 months were used as training samples, including various normal and abnormal flight states. The network weights and biases were iteratively adjusted using the Adam optimizer, with a learning rate of 0.001 and 500 iterations, until the loss function (using cross-entropy loss) converged to below 0.01. After training, real-time collected parameters were input into the network in the same format. The network calculated the probability values ​​of each category and assigned the samples to the cluster with the highest probability. For example, when the input parameters were altitude between 8000 and 10000 meters, airspeed between 750 and 850 km / h, and vertical velocity between -100 and 100 m / min, the probability of the network outputting the cruise category exceeded 0.9, thus classifying it into the cruise cluster. When the altitude was between 0 and 500 meters, airspeed between 200 and 300 km / h, and vertical velocity between -1000 and -500 m / min, the probability of the outputting the landing category exceeded 0.9, thus classifying it into the landing cluster. The final cluster category dataset includes at least 5,000 samples, and the similarity of parameters within each group exceeds 0.85, providing clear data support for the entropy weight calculation in step S3.

[0026] Step S3: Use the entropy weight approximation ideal solution sorting algorithm to process the flight state data sets of different cluster categories, and calculate the entropy weight of the flight state of each cluster category and its closeness to the ideal solution.

[0027] Specifically, the entropy-weighted approximation-ideal-solution ranking algorithm, when processing clustered data, can scientifically quantify the importance of each parameter and the safety level of each category state, serving as a crucial link between cluster analysis and risk model construction. The entropy-weighted method determines weights by calculating the information entropy of parameters; the higher the dispersion of a parameter, the lower its information entropy, and the greater its corresponding weight. This allows the weight allocation to objectively reflect the actual contribution of parameters in risk assessment. The approximation-ideal-solution ranking algorithm obtains a proximity index by measuring the distance between each cluster category and the ideal safe state (ideal solution) and the dangerous state (worst-case solution). This index can be directly used to measure the safety level of the category state. The combined use of these two methods not only solves the problem of subjective assignment of parameter weights but also achieves a quantitative assessment of the safety level of cluster categories, providing a reliable quantitative basis for subsequent risk assessment model construction and contributing to improving the objectivity and accuracy of model evaluation.

[0028] In practice, the datasets for each cluster category (such as cruise, landing, etc.) are processed separately. First, the parameters in the dataset are standardized, converting parameters of different dimensions to a range of 0 to 1. The conversion formula is (parameter value - minimum value) / (maximum value - minimum value). For example, the altitude parameter is converted from -100 meters to 15000 meters to a standardized value of 0 to 1. Next, the information entropy of each parameter is calculated. For a given parameter, its proportion in all samples of that cluster is calculated, i.e., the ratio of the parameter value of each sample to the sum of all sample values ​​for that parameter. Then, the entropy value is calculated using the information entropy formula, where the sample size is set to the total number of samples in that cluster category (e.g., 5000 samples for the cruise category). The entropy weight is calculated based on the entropy value using the formula (1 - entropy value) / (sum of (1 - entropy values) of all parameters). For example, the calculated entropy weight for the altitude parameter in the cruise category is 0.18, airspeed is 0.15, and vertical speed is 0.12, indicating that altitude is more important in risk assessment during the cruise phase. Next, the ideal and worst solutions for each parameter are determined. The ideal solution is the optimal value of the parameter in historical safe flight data. For example, the ideal solution for altitude during the cruise phase is 9000 meters (normalized value 0.6), and the ideal solution for airspeed is 800 km / h (normalized value 0.89). The worst solution is the critical value in historical accident data. For example, the worst solutions for altitude are 12000 meters (normalized value 0.8) and 3000 meters (normalized value 0.2), and the worst solutions for airspeed are 900 km / h (normalized value 1.0) and 700 km / h (normalized value 0.78). Then, the weighted Euclidean distance between each cluster category and the ideal solution and the worst solution is calculated. The weights are the entropy weights calculated above. For example, the distance between the cruise category and the ideal solution is 0.35, and the distance between it and the worst solution is 0.65. According to the proximity formula (proximity = distance to ideal solution / (distance to ideal solution + distance to worst solution)), the proximity is calculated to be 0.35 / (0.35 + 0.65) = 0.35, indicating that the safety level of this cluster category is at a medium level. Finally, all parameters, entropy weights, and proximity for each cluster category are organized into a matrix and stored in a data file to provide input data for building the initial model framework in step S4.

[0029] Step S4: Based on the aviation collision risk model, and combining the obtained entropy weights and proximity, construct the initial model framework for flight risk assessment.

[0030] Specifically, based on the aviation collision risk model, integrating the entropy weights and proximity scores obtained in step S3 to construct the initial model framework is a crucial step in achieving multi-source information fusion. The aviation collision risk model itself calculates collision probabilities based on parameters such as aircraft position, speed, and heading, reflecting the physical collision risk between aircraft. The entropy weights reflect the differences in importance of various flight parameters in risk assessment, while the proximity score quantifies the basic safety level of flight states in different cluster categories. Organically combining these three elements allows the initial model framework to consider not only physical collision factors but also the differences in parameter weights and the safety baseline of the states, thereby comprehensively capturing various risk factors during flight and forming a structurally complete and logically clear risk assessment framework, providing a solid structural foundation for subsequent model training and optimization.

[0031] In practical implementation, the core structure of the aviation collision risk model is first analyzed. This model includes a position input module, a relative motion calculation module, and a collision probability output module. The position input module receives the aircraft's longitude, latitude, and altitude parameters. The relative motion calculation module calculates the relative distance and approach rate based on speed and heading. The collision probability output module outputs a collision probability value from 0 to 1 (1 indicates a certainty of collision) based on the relative distance and approach rate. Subsequently, a parameter weighting module and a state correction module are added to this model. The parameter weighting module receives the entropy weights calculated in step S3 and weights the input flight parameters. For example, the altitude parameter is multiplied by its entropy weight of 0.18, and the airspeed parameter is multiplied by its entropy weight of 0.15, so that the model gives more emphasis to the influence of high-weight parameters when calculating relative motion. The state correction module receives the proximity value and corrects the result of the collision probability output module. The correction formula is: corrected probability = original collision probability × (1 - proximity). For example, when the proximity is 0.35, the original collision probability of 0.2 is corrected to 0.2 × (1 - 0.35) = 0.13. That is, a state with a good safety foundation will appropriately reduce the assessed value of the collision probability. Next, the data transmission path between modules is designed. Flight parameters first enter the parameter weighting module for weighting. The weighted parameters are then input into the relative motion calculation module to calculate the relative motion index with the parameters of another aircraft. The relative motion index is input into the collision probability output module to obtain the original probability. The original probability enters the state correction module, is corrected by proximity, and then outputs the final intermediate risk assessment value. Simultaneously, the model's input is set as the multi-dimensional flight parameters of the two aircraft and their proximity to their respective cluster categories, with the output being a risk assessment median value ranging from 0 to 1 (a higher value indicates higher risk). The parameter ranges for each module are defined, such as a relative distance calculation range of 0 to 100 kilometers and a proximity rate range of -1000 meters / second to 1000 meters / second (negative values ​​indicate distance). Through the integration process, the resulting initial model framework includes five functional modules and clearly defined input-output relationships. It can process the data output from step S3 and generate the risk assessment median value, providing an adjustable structural foundation for model training in step S5.

[0032] Step S5: Substitute the clustering results output by the flight state deep clustering network and the relevant values ​​calculated by the entropy weight approximation ideal solution sorting algorithm into the initial model framework to train and optimize the flight risk assessment model.

[0033] Specifically, incorporating clustering results, entropy weights, and proximity scores into the initial model framework for training and optimization is a core step in improving model evaluation accuracy. While the initial model framework integrates multiple types of information, the parameter configurations between modules (such as thresholds for collision probability calculation and correction coefficients) are still based on empirical settings, potentially deviating from actual flight risk characteristics. Through the training process, iterative adjustments to model parameters using risk cases from historical flight data ensure that the model's output aligns with actual risk conditions, enhancing the model's adaptability to different flight scenarios. The training process employs a data-driven optimization approach, continuously reducing prediction errors to optimize parameters. The resulting model more accurately reflects the mapping relationship between flight parameters and risk levels, providing reliable algorithmic support for real-time risk assessment.

[0034] In practice, a training dataset is first constructed, comprising 200,000 sets of flight data from the past 5 years. Each set includes flight parameters (16 items such as altitude and airspeed) for two aircraft, entropy weights and proximity scores for their respective clusters, and corresponding actual risk labels (0 for no risk, 1 for low risk, 2 for medium risk, and 3 for high risk). The dataset is divided into a training set (160,000 sets) and a validation set (40,000 sets) in an 8:2 ratio, and the risk labels are one-hot encoded. The training set data is then input into the initial model framework. The model outputs intermediate risk assessment values ​​(0 to 1), which are converted into probability distributions for the four risk classes using a softmax function. The cross-entropy loss is then calculated between these distributions and the one-hot encoded actual labels. The model parameters were optimized using stochastic gradient descent, including the threshold for collision probability calculation (initially set to 5 km) and the coefficients of the state correction module (initially set to 1.0). The learning rate was set to 0.0001, and each iteration used a batch of 1024 samples, with 1000 iterations. After the 100th iteration, the loss value on the validation set decreased from the initial 1.5 to 0.8, at which point the learning rate was adjusted to 0.00005. After the 500th iteration, the loss value decreased to 0.4, and iterations continued until the 800th iteration, where the loss value stabilized at around 0.35. The difference between the validation set loss and the training set loss was less than 0.05, indicating that the model was not overfitting. In the optimized model parameters, the threshold for collision probability calculation was adjusted to 3 km, and the coefficients of the state correction module dynamically changed according to the clustering category (e.g., 0.9 for the cruise category and 1.2 for the landing category). Finally, the optimized model was tested using a validation set. The prediction accuracy for various risks was 92% for no risk, 88% for low risk, 85% for medium risk, and 80% for high risk, all of which met the preset standard of over 80%. The training and optimization of the model were completed, and the optimized model parameters were stored as a binary file to provide a basis for real-time evaluation in step S6.

[0035] S6 uses the trained and optimized flight risk assessment model to process the real-time flight status parameters of the aircraft and outputs the risk assessment results of the aircraft in the current flight phase.

[0036] Specifically, the final application of the entire assessment method is to process real-time flight parameters using the trained and optimized model and output risk assessment results, directly serving flight safety decision-making. The trained model possesses high predictive accuracy and can quickly respond to changes in real-time parameters, calculating the risk level for the current flight phase. The output risk assessment results include specific risk values ​​and levels, providing pilots with immediate risk warnings and ground control personnel with decision-making support. This helps them take timely measures such as adjusting course and altitude to mitigate risks, thereby keeping flight risks within acceptable limits and fully leveraging the practical value of this assessment method in ensuring flight safety.

[0037] In practice, a real-time communication interface between the model and the aircraft data bus is first established. The ARINC429 protocol is used to receive real-time flight parameters, with a data transmission rate set to 100kbps to ensure parameter update latency does not exceed 100 milliseconds. Real-time parameters include flight path parameters of the aircraft and other aircraft within a 50km radius (updated 10 times per second), aircraft configuration parameters (updated once per second), engine parameters (updated twice per second), navigation parameters (updated five times per second), and environmental parameters (updated once per minute). Simultaneously, the entropy weights and proximity scores of each aircraft's cluster category are obtained through the interface (updated once every 10 seconds). These parameters are then encapsulated into input vectors according to the model's requirements. The aircraft's parameters and the parameters of each surrounding aircraft form one set of inputs (e.g., if there are three surrounding aircraft, three sets of inputs are generated). Each set of inputs includes 32 parameters (16 for the aircraft itself + 16 for other aircraft) and two proximity scores. The trained and optimized model is used to perform parallel computation on the input vector, with the computation time for a single set of parameters controlled within 50 milliseconds. The output is a risk value of 0 to 1 and its corresponding risk level (0 to 3). A risk value less than 0.2 is classified as level 0 (no risk); 0.2 to 0.4 as level 1 (low risk); 0.4 to 0.6 as level 2 (medium risk); and greater than 0.6 as level 3 (high risk). The assessment results are presented in numerical and color format on the cockpit's multi-function display: green for level 0, yellow for level 1, orange for level 2, and red for level 3, accompanied by corresponding voice prompts (e.g., "High risk warning, please evacuate immediately" for level 3 risk). The ground monitoring system receives the assessment results synchronously via data link, stores them in the server database, and generates a real-time risk heatmap. For example, when the relative distance between the pilot and another aircraft decreases to 2 kilometers, with a proximity rate of 500 meters per second, the model calculates a risk value of 0.75 and outputs a level 3 risk. The pilot can immediately climb 500 meters based on the warning, reducing the risk value to 0.15 after 10 seconds, restoring it to level 0 risk. This real-time assessment and response mechanism enables dynamic monitoring and timely handling of flight risks, effectively ensuring aircraft flight safety.

[0038] Preferably, in step S3, when calculating the entropy weight of the flight state of each cluster category and its closeness to the ideal solution using the entropy weight approximation ideal solution sorting algorithm, the following model formula is adopted: ,in, w j Indicates the first j Entropy weights of each flight state parameter, , , x ij For the first i In the cluster category, the first j Standardized values ​​of each flight state parameter m For the number of cluster categories, n The number of flight status parameters; ,in, C j Indicates the first i The closeness of cluster categories, , , For the first j The optimal values ​​of each parameter For the first j The worst value of each parameter.

[0039] Specifically, step S3 clarifies the implementation details and related technical aspects of the entropy weight approximation ideal solution sorting algorithm when calculating entropy weights and proximity. This step is crucial for the accuracy of subsequent risk assessment. During implementation, it is first necessary to clarify the various technical parameters involved and their value ranges. The number of flight state parameters is determined based on the actual dimensions collected, typically including multiple aspects such as flight path, aircraft configuration, engine, navigation, and environment, totaling 15 to 20. The number of clusters is divided into 4 to 6 categories based on flight phases and state characteristics, such as ground taxiing, takeoff, cruise, and landing. When calculating entropy weights, the values ​​of each parameter in each cluster must first be processed and converted into standardized values ​​using a specific method. The standardized values ​​are typically controlled between 0 and 1 to eliminate the influence of different parameter units and magnitudes. Subsequently, based on the standardized values, the proportion of each parameter in different clusters is calculated. The proportion calculation must ensure that the sum of the proportions of the same parameter in all clusters is 1. Next, information entropy is calculated based on the proportions. Information entropy ranges from 0 to 1; a smaller value indicates greater dispersion and richer information. Entropy weights are then calculated based on the information entropy. The sum of the entropy weights is 1, and their magnitude directly reflects the importance of the corresponding parameter in risk assessment. A larger value indicates a more significant impact of the parameter on risk assessment. When calculating the closeness, the optimal and worst values ​​for each parameter must first be determined. The optimal value is typically selected as the typical value of the parameter in historical safe flight data, while the worst value is selected as the critical value in historical accident or near-miss data. These values ​​are determined based on statistical analysis of a large amount of historical flight data. Then, the weighted distance between each cluster category and the optimal and worst values ​​is calculated. The calculation of the weighted distance must be combined with the previously obtained entropy weights to highlight the influence of important parameters. Finally, the closeness is calculated based on the weighted distances. The closeness ranges from 0 to 1; a larger value indicates that the flight status of that cluster category is closer to the ideal safety state. This process enables the scientific and objective quantification of the importance of each parameter and the safety level of each cluster category, providing a reliable quantitative basis for the subsequent construction of flight risk assessment models and ensuring the scientific validity and accuracy of the assessment results.

[0040] Preferably, in step S4, when constructing the initial model framework for flight risk assessment, the following model formula is adopted based on the aviation collision risk model combined with entropy weights and proximity: ,in, R Indicates the flight risk value. α , β These are the weighting coefficients. f ( v ij ) is the first i In the cluster category, the first j Risk impact function of each flight state parameter vij For the first i In the cluster category, the first j The actual values ​​of each flight status parameter g ( d ij ) is the risk impact function of the distance between aircraft. d ij For the first i aircraft and the first j The real-time distance between aircraft.

[0041] Specifically, the implementation of step S4, which constructs the initial model framework for flight risk assessment, is significant in that it organically integrates the aviation collision risk model with the entropy weights and proximity values ​​obtained in step S3, forming an assessment framework that can comprehensively reflect multiple risk factors. During implementation, the definitions and value ranges of various technical parameters in the model must be clearly defined. The weight coefficients are determined based on the actual needs of flight risk assessment and historical data statistics, typically ranging from 0 to 1, with the sum of the two being 1, to balance the influence of different risk factors. The actual values ​​of flight state parameters are determined based on collected real-time data, including multiple dimensions such as altitude, airspeed, and vertical speed. Each parameter has its specific physical range; for example, altitude can range from -100 meters to 15,000 meters, and airspeed can range from 0 to 900 km / h. The design of the risk impact function needs to be based on extensive historical flight data and accident case analysis. Its output reflects the degree to which the actual parameter values ​​deviate from the ideal values ​​and their impact on risk. Generally, when the actual parameter values ​​are close to the ideal values, the function output is smaller, indicating a smaller impact on risk; conversely, the output is larger. The real-time distance between aircraft is calculated based on their position parameters, ranging from 0 to 100 kilometers. The closer the distance, the larger the corresponding risk impact function output, indicating a higher collision risk. When constructing the model framework, entropy weights are first used as weights for each flight state parameter, allowing the model to highlight the influence of important parameters during calculation. Simultaneously, proximity is combined with the risk impact function of each cluster category to reflect the basic safety level of flight states in different cluster categories. The final flight risk value is obtained by weighting and summing the two risk impacts using weighting coefficients. The flight risk value typically ranges from 0 to 10, with higher values ​​indicating higher risk. This initial model framework comprehensively considers the importance of flight state parameters, the safety level of different cluster categories, and the collision risk between aircraft, laying the foundation for subsequent model training and optimization, and ensuring that the final risk assessment model can comprehensively and accurately reflect the actual risk level during flight.

[0042] Preferably, in step S5, when training and optimizing the flight risk assessment model, the model parameters are adjusted based on the clustering results output by the flight state deep clustering network. The relevant model formula is as follows: ,in, θ k+1 For the first k Model parameters for +1 iteration θ k For the first k Model parameters for the next iteration γ For learning rate, For loss function L ( θ k )exist θ k gradient at, , R i This represents the actual risk value. To predict risk values ​​for the model, c i For the first The weight of each sample is related to the clustering result output by the flight state deep clustering network.

[0043] Specifically, step S5, model training and optimization, is described. Its core significance lies in adjusting model parameters through a data-driven approach to improve the model's accuracy in predicting flight risks. The technical parameters involved include model parameters, learning rate, loss function, actual risk value, predicted risk value, and sample weights. Model parameters include various coefficients and thresholds in the initial model framework, with initial values ​​set empirically and continuously adjusted through training. The learning rate typically ranges from 0.0001 to 0.01; a value that is too large may lead to model training instability, while a value that is too small will prolong training time. Both the actual and predicted risk values ​​range from 0 to 10. The actual risk value is determined based on historical accident data and expert evaluation, while the predicted risk value is the model's output. Sample weights are related to the clustering results output by the flight state deep clustering network. For clusters containing important risk features, the corresponding sample weights are larger, typically between 1.0 and 1.5, to improve the model's learning effect on these categories. During training, the training set data is first input into the initial model framework. Predicted risk values ​​are calculated through forward propagation. These predicted risk values ​​are then compared with actual risk values ​​to calculate the loss function. The magnitude of the loss function reflects the deviation between the prediction and the actual situation; the initial value may be large, gradually decreasing as training progresses. Based on the loss function value, the backpropagation algorithm is used to adjust the model parameters. The adjustment magnitude is determined by both the learning rate and the gradient of the loss function; a larger gradient results in a larger parameter adjustment. In each iteration, the loss function value needs to be recalculated, and its changes are used to determine if the model has converged. When the loss function value stabilizes within a small range (e.g., less than 0.5), it indicates that the model training has achieved the expected results. Simultaneously, to avoid overfitting, validation set data is used to monitor model performance. When the loss function value on the validation set stops decreasing or even increases, training should be stopped immediately, and model parameters or training strategies should be adjusted. Through the training and optimization process, the model continuously learns the intrinsic relationship between flight parameters and risk, making the prediction results closer to the actual risk level and providing reliable model support for subsequent real-time risk assessment.

[0044] Preferably, in step S6, when outputting the risk assessment results of the aircraft in the current flight phase, the following model formula is used to classify the risk level: Where Level indicates the risk level. T 1. T 2 represents the threshold for risk level classification. The flight risk value is calculated by combining the initial model framework constructed in step S4 with the parameters trained and optimized in step S5, and is related to the clustering category output by the flight state deep clustering network and the closeness obtained by the entropy weight approximation ideal solution sorting algorithm.

[0045] Specifically, the method for classifying risk levels when outputting the risk assessment results in step S6 is significant in converting abstract risk values ​​into intuitive risk levels, facilitating pilots and ground control personnel to quickly understand and respond to flight risks. During implementation, the threshold for risk level classification and the range of risk values ​​must be clearly defined. The risk values ​​are calculated using the initial model framework built in step S4 combined with the parameters trained and optimized in step S5, typically ranging from 0 to 10. Risk levels are divided into three levels, each corresponding to a different degree of risk, facilitating the adoption of different response measures. Determining the risk level classification threshold requires based on extensive historical flight data and accident statistical analysis. Typically, the low-risk threshold is set between 2 and 3, and the medium-risk threshold between 5 and 6. Specific values ​​need to be adjusted according to factors such as aircraft type, flight phase, and operating environment. When calculating risk values, the clustering categories output by the deep clustering network of flight status and the proximity obtained by the entropy weight approximation algorithm for ideal solutions must be considered comprehensively. Different clustering categories correspond to different basic risk levels, while the proximity reflects how close the flight status of that category is to the ideal safe state. The higher the proximity, the lower the risk value calculated under the same parameter conditions. When the risk value is less than the low-risk threshold, it is determined to be at a low-risk level, indicating that the current flight status is relatively safe and no special measures are required. When the risk value is between the low-risk threshold and the medium-risk threshold, it is determined to be at a medium-risk level, prompting pilots and ground control personnel to closely monitor changes in flight status and prepare for contingencies. When the risk value is greater than or equal to the medium-risk threshold, it is determined to be at a high-risk level, requiring immediate measures such as adjusting flight attitude, changing course, and maintaining a safe distance to reduce flight risk. This risk level classification method simplifies and visualizes complex risk assessment results, improves the efficiency of risk information transmission and the timeliness of response decisions, and effectively ensures aircraft flight safety.

[0046] Preferably, in step S2, when the flight state deep clustering network performs clustering analysis on the flight state, the model formula used is as follows: ,in, z i For the first i Latent feature representation of a set of flight state parameters h (·) is the mapping function for the flight state deep clustering network. θ For network parameters, x i For the first i A set of flight state parameters; during clustering, through calculation Determine sample attribution, where, q ij For the first i The sample belongs to the first jThe probability of a cluster center u j For the first j Cluster centers, K This represents the number of clusters.

[0047] Specifically, step S2 involves using a deep clustering network to perform cluster analysis on flight states. Through the feature extraction and clustering capabilities of deep neural networks, a refined classification of flight states is achieved, providing an accurate data foundation for subsequent risk assessment. The technical parameters involved include the flight state parameter set, network parameters, latent feature representations, cluster centers, and the number of clusters. The flight state parameter set includes collected multi-dimensional parameters, each with its specific physical meaning and value range. Network parameters include the number of neural network layers, the number of neurons per layer, and the activation function type. Typically, the number of network layers is set to 3 to 5, with the number of neurons per layer gradually decreasing from the input layer to the output layer. For example, the input layer has 16 neurons, and the hidden layers have 32, 16, and 8 neurons respectively. The ReLU activation function is used to improve feature extraction capabilities. The number of clusters is determined based on the flight stage and state characteristics, typically ranging from 4 to 6, such as ground taxiing, takeoff, cruise, and landing. The cluster centers are feature representatives of each category determined during the training process, and their values ​​are continuously optimized as training progresses. In cluster analysis, the flight state parameter set is first input into a deep clustering network. Through multi-layer nonlinear mapping, the high-dimensional parameter set is transformed into a low-dimensional latent feature representation. This latent feature representation can more effectively capture the inherent correlations and hidden patterns between parameters. Then, the distance between the latent feature representation of each sample and each cluster center is calculated using the Euclidean distance formula; a smaller distance indicates a higher similarity between the sample and the cluster center. The probability of a sample belonging to each cluster center is calculated based on the distance, with the probability ranging from 0 to 1. The sum of the probabilities of all cluster centers is 1, and the cluster center with the highest probability is the category to which the sample belongs. During network training, the network parameters and cluster centers are continuously adjusted to maximize the similarity of samples within the same cluster category and minimize the similarity between samples from different cluster categories. This is typically achieved by minimizing a specific loss function, which gradually decreases as training progresses until it converges to a small value (e.g., less than 0.1). Cluster analysis can accurately classify complex and diverse flight states into different categories. Flight states within each category have a high degree of similarity, providing a clear and accurate data foundation for subsequent steps such as entropy weight calculation and model building, which helps improve the accuracy and reliability of the entire flight risk assessment method.

[0048] Preferably, step S3 includes the following sub-steps: S3.1, standardizing the flight status data sets of different cluster categories output by the flight status deep clustering network, converting the values ​​of each flight status parameter to the same data volume, and eliminating the impact of differences in the units of different parameters; S3.2, based on the standardized flight status data, calculating the information entropy of each flight status parameter in each cluster category, using the information entropy to reflect the dispersion of the parameter and its impact on risk assessment; S3.3, determining the entropy weight of each flight status parameter according to the entropy weight calculation rules based on the calculated information entropy, the magnitude of the entropy weight is positively correlated with the importance of the parameter in risk assessment; S3.4, constructing a weighted decision matrix based on the entropy weight of each parameter, and obtaining the proximity of each cluster category to the ideal solution and the worst solution by calculating the distance between each cluster category and the ideal safety state, the proximity is used to measure the degree of closeness between the flight status of the cluster category and the ideal safety state.

[0049] Specifically, step S3, through four clear sub-step operations, ensures that the entropy weight approximation ideal solution sorting algorithm can accurately quantify the weights of each parameter and the safety level of the cluster category, providing solid data support for the risk assessment model. In implementation, the first step standardizes the different cluster category data output by the flight state deep clustering network, uniformly mapping all flight state parameter values ​​to the range of 0 to 1. For example, parameters of different magnitudes, such as altitude from -100 meters to 15,000 meters and airspeed from 0 to 900 km / h, are converted to the same range, completely eliminating calculation bias caused by unit differences and laying the foundation for a fair comparison of the impact of each parameter in the future. The second step calculates the information entropy of each parameter in each cluster category based on the standardized data. By statistically analyzing the distribution frequency of the parameter across all samples, the entropy value is determined based on the uniformity of distribution. The entropy value ranges from 0 to 1; when parameter values ​​are concentrated in a certain range, the entropy value approaches 1, and the more dispersed the distribution, the closer the entropy value is to 0, thus intuitively reflecting the richness of risk information included in the parameters. The third step is to calculate the entropy weights based on information entropy. The weights are determined by the ratio of (1 - entropy value) to the sum of (1 - entropy values) of all parameters, ensuring that the sum of the entropy weights for all parameters is 1. For example, if the entropy value of a parameter is 0.3, its entropy weight is (1 - 0.3) divided by the sum of (1 - entropy values) of all parameters, ensuring that the entropy weights accurately reflect the importance ranking of parameters in risk assessment. The fourth step is to construct a weighted decision matrix, using entropy weights to calculate the weighted distance between each cluster category and the ideal solution and the worst-case solution. The ideal solution is the optimal value of each parameter based on historical safety data, and the worst-case solution is the accident threshold value. The distance calculation uses the Euclidean distance formula and incorporates entropy weights, ultimately obtaining a proximity score from 0 to 1. A higher score indicates that the cluster category is closer to the ideal safety state. Through this series of steps, the importance of parameters and the safety level of the cluster category are scientifically quantified.

[0050] Preferably, step S4 includes the following sub-steps: S4.1, obtaining the structure and parameters of the aviation collision risk model, which includes the correlation between aircraft position, speed, heading parameters and collision probability; S4.2, using the entropy weights and proximity of the flight states of each cluster category obtained in step S3 as input variables, integrating them into the structure of the aviation collision risk model to establish new model variable relationships; S4.3, determining the weight coefficients of each input variable in the model, the setting of the weight coefficients is based on the actual needs of flight risk assessment and the historical data statistical results of the impact of each variable on risk; S4.4, integrating the variable relationships and weight coefficients to form an initial model framework for flight risk assessment, which can receive the results of flight state parameters after preprocessing and output the corresponding intermediate risk assessment values.

[0051] Specifically, step S4 involves four sub-steps that systematically integrate the aviation collision risk model with previous calculations to construct an initial model framework that comprehensively reflects risk factors, providing a structural foundation for subsequent optimization. In implementation, the first step involves acquiring the basic structure and core parameters of the aviation collision risk model. This model includes positional parameters such as longitude, latitude, and altitude (accuracy up to 0.1 meters), motion parameters such as speed (0 to 900 km / h) and heading (0 to 360 degrees), and collision probability correlation indicators such as relative distance (0 to 100 km) and proximity rate (-1000 to 1000 m / s) calculated based on these parameters. Clarifying the mathematical relationships between these parameters is a prerequisite for integration. The second step incorporates the entropy weights and proximity obtained in step S3 as input variables into the model, establishing new variable relationships. The entropy weights are used to adjust the weights of each flight parameter in the collision probability calculation, and the proximity serves as a basic safety factor to correct the final risk value. This ensures that the model not only considers physical collision factors but also reflects the importance of parameters and the impact of the state safety baseline. The third step is to determine the weighting coefficients of each input variable. By analyzing flight accident data and safety records from the past 10 years, statistical regression methods are used to calibrate the coefficient values, which range from 0 to 1 and sum to 1. For example, during the cruise phase, the weighting coefficient for the relative distance parameter might be set to 0.4, and the proximity correction coefficient to 0.3, ensuring that the coefficients accurately reflect the actual risk contribution. The fourth step is to integrate the variable relationships and weighting coefficients to form an initial model framework. The inputs are defined as processed flight parameters, entropy weights, and proximity, and the outputs are intermediate risk assessment values ​​from 0 to 10. The framework includes functional units such as a parameter weighting module, a collision probability calculation module, and a safety correction module. The modules are connected through data flow, providing a clear calculation path and adjustable parameter interfaces for subsequent training.

[0052] Preferably, step S5 includes the following sub-steps: S5.1, using the flight state data corresponding to the clustering results output by the flight state deep clustering network, the entropy weights calculated by the entropy weight approximation ideal solution sorting algorithm, and the closeness as training sample data, dividing them into a training set and a validation set; S5.2, inputting the training set data into the initial model framework constructed in step S4, calculating the model's predicted output through forward propagation, and comparing it with the actual risk label to obtain the loss value; S5.3, based on the loss value, using the backpropagation algorithm to adjust the model parameters, and monitoring the model performance in conjunction with the validation set data during the adjustment process to avoid overfitting; S5.4, repeating the forward and backpropagation parameter adjustment process of the training set until the model's performance on the validation set reaches the preset standard, completing the training and optimization of the flight risk assessment model.

[0053] Specifically, step S5, through a standardized training optimization process, enables the model to accurately learn the intrinsic relationship between flight status and risk, improving prediction accuracy to meet real-time assessment needs. In implementation, the first step involves assembling a training sample set using clustering results, entropy weights, and proximity scores. This set includes 50,000 historical data sets, divided in an 8:2 ratio into a training set (40,000 sets) and a validation set (10,000 sets). The samples cover all flight phases, including ground taxiing, takeoff, cruise, and landing. Each set includes 16 flight parameters, a corresponding entropy weight vector, a proximity score, and an actual risk level label from 0 to 3 (0 for no risk, 3 for extremely high risk), ensuring sample diversity and representativeness. The second step inputs the training set data into the initial model framework, calculates the predicted risk value through forward propagation, compares the predicted value with the actual label, and calculates the loss value using the cross-entropy loss function. The initial loss value is typically between 1.5 and 2.0, directly reflecting the degree of prediction bias and serving as the core basis for parameter adjustment. The third step involves adjusting the model parameters using backpropagation based on the loss value. This includes adjusting the weight coefficients and thresholds of each module, with a learning rate set to 0.0001. Each iteration adjusts the parameters to reduce the loss value, and a validation set is used for testing every 100 iterations. If the validation set loss value increases for three consecutive iterations, it is considered overfitting, and the current adjustment is stopped. At this point, the training set loss value has typically dropped below 0.3. The fourth step repeats the forward and backpropagation process until the model achieves an accuracy of over 85% on the validation set and the loss value stabilizes below 0.25. The optimized model parameters include 128 weight coefficients and 32 thresholds, all stored as binary files. Through this process, the model can accurately capture the non-linear relationship between parameters and risk, ensuring that the error during real-time evaluation is controlled within an acceptable range.

[0054] The flight state deep clustering network of this invention is a flight state classification tool based on deep neural networks. Its core lies in extracting latent features of flight state parameters through multi-layer nonlinear mapping, achieving refined classification of complex flight states. The implementation process first requires building a network architecture including an input layer, multiple hidden layers, and an output layer. The number of neurons in the input layer matches the dimensionality of the collected flight state parameters. The hidden layers extract features through operations such as convolution and fully connected layers. The output layer corresponds to the preset number of cluster categories. During the training phase, a large amount of historical flight data is used to train the network. By adjusting parameters such as network weights and biases, the network can accurately identify the feature patterns of different flight states. A specific loss function is used to measure the classification effect during training until the loss function converges to a preset range. After training, the real-time collected flight state parameters are input into the network. The network calculates the probability values ​​of each category and assigns the samples to the cluster category with the highest probability. Its function is to divide multi-dimensional and complex flight state parameters into clusters with similar features, enabling subsequent risk assessments to use consistent analytical logic for similar states and avoiding assessment bias caused by mixed states. The significance of this network lies in its ability to overcome the limitations of traditional clustering methods in feature extraction. It can more accurately capture subtle differences between flight states, providing a structured and clearly categorized data foundation for the entire flight risk assessment method. This improves the pertinence and accuracy of risk assessment and lays a solid data foundation for subsequent weight calculation and model building.

[0055] The entropy-weighted approximation-ideal-solution ranking algorithm is a comprehensive algorithm used in this invention to quantify the importance of flight state parameters and the safety level of cluster categories. Its implementation includes two key steps: entropy weight calculation and approximation-ideal-solution ranking. In the entropy weight calculation part, the flight state parameters of different cluster categories are first standardized, mapping parameter values ​​to the range of 0 to 1 to eliminate unit differences. Then, the distribution frequency of each parameter in different samples is calculated, and the information entropy is calculated based on the frequency. The entropy weight value of each parameter is then determined based on the information entropy. The magnitude of the entropy weight value is positively correlated with the importance of the parameter in risk assessment, and the sum of the entropy weight values ​​of all parameters is 1. In the approximation-ideal-solution ranking part, the ideal solution (optimal safety value) and the worst-case solution (critical danger value) for each parameter need to be determined. Then, the weighted distance between each cluster category and the ideal and worst-case solutions is calculated using the entropy weight values. Finally, the proximity degree is calculated based on the distance. The proximity degree value is between 0 and 1; a larger value indicates that the flight state of that cluster category is closer to the ideal safety state. This algorithm scientifically quantifies the influence weights of various parameters and the safety levels of each cluster category, resolving the issue of subjective parameter weight assignment and enabling an intuitive assessment of the safety level of each cluster category. Its significance lies in providing an objective and reliable quantitative basis for flight risk assessment models, enabling the models to highlight the influence of important parameters when calculating risks, accurately reflecting the safety foundation of flight states in different cluster categories, thereby improving the objectivity and accuracy of risk assessment and providing crucial support for constructing scientifically sound flight risk assessment models.

[0056] The aircraft collision risk model is the core model in this invention used to assess the collision risk between aircraft. It is implemented based on the aircraft's physical parameters such as position, speed, and heading, calculating risk by establishing the correlation between these parameters and collision probability. The model first acquires the aircraft's positional parameters (accuracy up to 0.1 meters), speed (0 to 900 km / h), and heading (0 to 360 degrees), etc., and then calculates indicators such as relative distance (0 to 100 km) and proximity rate (-1000 to 1000 m / s) between aircraft based on these parameters. These indicators are then converted into collision probabilities using specific mathematical formulas. The collision probability ranges from 0 to 1, with higher values ​​indicating higher collision risk. In this invention, this model is not used independently but integrated with entropy weights and proximity scores. Entropy weights are used to adjust the weights of each flight parameter in the collision probability calculation, and proximity scores serve as a basic safety factor to correct the final risk value, forming a comprehensive initial model framework for flight risk assessment. Its role is to assess the likelihood of collisions between aircraft from a physical perspective, serving as a crucial foundation for flight risk assessment. The significance of this model lies in providing a core basis for flight risk assessment that reflects physical collision risk. By integrating it with other parameters and algorithms, it can comprehensively consider various factors affecting flight risk, enabling the constructed risk assessment model to not only reflect physical collision risk but also the impact of parameter importance and state safety baseline. This provides a structural foundation for subsequent model training and optimization, ultimately improving the comprehensiveness and reliability of flight risk assessment and providing key technical support for ensuring aircraft flight safety.

[0057] like Figure 2 The aforementioned data-driven flight risk assessment method, implemented through different units, includes:

[0058] The real-time acquisition and transmission unit for multi-dimensional flight status parameters is equipped with a multi-type sensor group to simultaneously collect flight path parameters (altitude, airspeed, vertical speed, etc.), aircraft configuration parameters (flap position, landing gear status, etc.), engine parameters (engine operating status, etc.), navigation parameters (deviation from expected flight path, etc.), and environmental parameters (wind speed, wind direction, etc.) of the aircraft during flight. The parameter data is transmitted in real time to the flight status deep clustering analysis unit through an encrypted transmission protocol. At the same time, the collected data is initially verified to remove obvious outliers.

[0059] The flight state deep clustering analysis unit is connected to the flight state multi-dimensional parameter real-time acquisition and transmission unit through a data interface. After receiving the transmitted parameters, it calls the built-in flight state deep clustering network model to perform feature extraction and nonlinear mapping processing on the parameters, generate potential feature representations of the flight state, and then classify the potential features according to the clustering rules to obtain flight state data sets of different cluster categories. The set is then transmitted to the entropy weight approximation ideal solution sorting calculation unit.

[0060] The entropy weight approximation ideal solution sorting calculation unit is connected to the flight state deep clustering analysis unit through a data bus. After receiving the clustering results, it processes the flight state data of each cluster category. First, it calculates the information entropy of each parameter to determine the entropy weight, then constructs a weighted decision matrix, and obtains the closeness by calculating the distance to the ideal solution and the worst solution. The entropy weight and closeness data are packaged and sent to the aviation collision risk model framework construction and parameter integration unit.

[0061] The aviation collision risk model framework construction and parameter integration unit is connected to the entropy weight approximation ideal solution sorting calculation unit via wired communication. After receiving the entropy weight and proximity data, it retrieves the pre-stored aviation collision risk model structure, incorporates the received data as input variables, determines the weight coefficients of each variable in combination with the influence characteristics of flight risk parameters, integrates them to form the initial model framework for flight risk assessment, and transmits the framework to the flight risk assessment model training and optimization unit.

[0062] The flight risk assessment model training and optimization unit is connected to the aviation collision risk model framework construction and parameter integration unit through a data interaction module. After receiving the initial model framework, it imports the clustering results output by the flight state deep clustering network and the relevant values ​​calculated by the entropy weight approximation ideal solution sorting algorithm as training samples. It uses an iterative optimization algorithm to adjust the model parameters and monitors the model performance through a loss function until the model reaches the preset accuracy standard. The optimized model is then transmitted to the flight risk real-time assessment and result output unit.

[0063] The real-time flight risk assessment and result output unit is connected to the flight risk assessment model training and optimization unit and the real-time acquisition and transmission unit of multi-dimensional flight status parameters. It receives the optimized model from the former and obtains real-time flight parameters from the latter. The real-time parameters are input into the model for calculation to obtain the risk assessment results of the current flight phase. The risk value and risk level are displayed through a visual interface, and the results are synchronously stored in the local database.

[0064] A data-driven flight risk assessment method is proposed, which achieves refined classification of flight states, effectively overcoming the problems of coarse classification and insufficient specificity in existing methods. Through a deep clustering network of flight states, in-depth analysis and clustering of multi-dimensional flight state parameters, including flight path, aircraft configuration, engine, navigation, and environment, are performed to fully explore the inherent differences between different flight states and divide them into more granular cluster categories. This refined classification enables risk assessment to accurately match the characteristics of each flight state category and accurately reflect the risk characteristics under different cluster categories, laying a precise foundation for subsequent risk assessment.

[0065] Secondly, this method scientifically determines the weights of each flight state parameter, overcoming the shortcomings of existing models in lacking scientific rigor in parameter weight determination and insufficient sensitivity to key risk factors. By employing an entropy-weighted algorithm to approximate the ideal solution, the entropy weight of each parameter and its closeness to the ideal solution are calculated. The weights are then determined based on the information entropy characteristics and closeness of the parameters. This process fully considers the actual impact of each parameter in risk assessment, making the weight allocation more reasonable and the model more sensitive to key risk factors, thereby significantly improving the reliability and accuracy of risk assessment results.

[0066] Furthermore, this method achieves effective fusion of multiple technologies and comprehensive integration of multi-source parameters, significantly improving the overall effectiveness of flight risk assessment. It organically combines a deep clustering network for flight states, an entropy-weighted algorithm for approximating ideal solutions, and an aviation collision risk model to construct an integrated flight risk assessment model. This fusion not only comprehensively captures the complex correlations between multi-source parameters but also fully leverages the advantages of each technology, achieving complementary strengths. This makes risk assessment both accurate and efficient, providing stronger technical support for aircraft flight safety.

[0067] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data-driven flight risk assessment method, characterized in that, include: Step S1: Collect multi-dimensional flight state parameters generated by the aircraft during flight. These parameters include flight path parameters, aircraft configuration parameters, engine parameters, navigation parameters, and environmental parameters. Flight path parameters include altitude, airspeed, vertical speed, heading, pitch angle, roll angle, and yaw angle. Aircraft configuration parameters include flap position, landing gear status, and spoiler deployment status. Engine parameters include engine operating status. Navigation parameters include lateral and vertical deviations from the expected flight path. Environmental parameters include wind speed, wind direction, temperature, and atmospheric pressure. Step S2: Input the collected flight state parameters into a deep clustering network for flight state analysis. This network performs cluster analysis on the flight state to obtain flight data for different cluster categories. Step S3: Process the flight state data sets of different cluster categories using the entropy weight approximation ideal solution sorting algorithm, and calculate the entropy weight of the flight state of each cluster category and its closeness to the ideal solution; Step S4: Based on the aviation collision risk model, and combined with the obtained entropy weight and closeness, construct the initial model framework for flight risk assessment; Step S5: Substitute the clustering results output by the flight state deep clustering network and the relevant values ​​calculated by the entropy weight approximation ideal solution sorting algorithm into the initial model framework to train and optimize the flight risk assessment model; Step S6: Use the trained and optimized flight risk assessment model to process the real-time flight state parameters of the aircraft and output the risk assessment results of the aircraft in the current flight phase; In step S3, when calculating the entropy weight of the flight state of each cluster category and its closeness to the ideal solution using the entropy weight approximation ideal solution sorting algorithm, the following model formula is adopted: ,in, w j Indicates the first j Entropy weights of each flight state parameter, , , x ij For the first i In the cluster category, the first j Standardized values ​​of each flight state parameter m For the number of cluster categories, n The number of flight status parameters; ,in, Ci Indicates the first i The closeness of each cluster category , , For the first i In the cluster category, the first j The optimal values ​​of each flight state parameter For the first i In the cluster category, the first j The worst value of each flight status parameter; In step S4, when constructing the initial model framework for flight risk assessment, the following model formula is adopted based on the aviation collision risk model combined with entropy weights and proximity: ,in, R Indicates the flight risk value. α , β These are the weighting coefficients. f ( v ij ) is the first i In the cluster category, the first j Risk impact function of each flight state parameter v ij For the first i In the cluster category, the first j The actual values ​​of each flight status parameter g ( d ij ) is the risk impact function of the distance between aircraft. d ij For the first i aircraft and the first j Real-time distance between aircraft; In step S2, the flight state deep clustering network uses the following model formula to perform clustering analysis on the flight state: ,in, z i For the first i Latent feature representation of a set of flight state parameters h (·) is the mapping function for the flight state deep clustering network. θ For network parameters, x i For the first i A set of flight state parameters; during clustering, through calculation Determine sample attribution, where, q ij For the first i The sample belongs to the first j The probability of a cluster center u j For the first j Cluster centers, K This represents the number of clusters.

2. The data-driven flight risk assessment method according to claim 1, characterized in that, In step S5, during the training and optimization of the flight risk assessment model, the model parameters are adjusted based on the clustering results output by the flight state deep clustering network. The relevant model formulas are as follows: ,in, θ k+1 For the first k Model parameters for +1 iteration θ k For the first k Model parameters for the next iteration γ For learning rate, For loss function L ( θ k )exist θ k gradient at, , R i This represents the actual risk value. To predict risk values ​​for the model, c i For the first The weight of each sample is related to the clustering result output by the flight state deep clustering network.

3. The data-driven flight risk assessment method according to claim 1, characterized in that, In step S6, when outputting the risk assessment results of the aircraft in the current flight phase, the following model formula is used to classify the risk level: Where Level indicates the risk level. T 1. T 2 represents the threshold for risk level classification. The flight risk value is calculated by combining the initial model framework constructed in step S4 with the parameters trained and optimized in step S5, and is related to the clustering category output by the flight state deep clustering network and the closeness obtained by the entropy weight approximation ideal solution sorting algorithm.

4. The data-driven flight risk assessment method according to claim 1, characterized in that, Step S3 includes the following sub-steps: S3.1, standardize the flight status data sets of different cluster categories output by the flight status deep clustering network, convert the values ​​of each flight status parameter to the same data volume, and eliminate the impact of differences in the units of different parameters; S3.2, based on the standardized flight status data, calculate the information entropy of each flight status parameter in each cluster category, and use the information entropy to reflect the dispersion of the parameter and its impact on risk assessment. S3.3, Based on the calculated information entropy, determine the entropy weight of each flight state parameter according to the entropy weight calculation rules. The magnitude of the entropy weight is positively correlated with the importance of the parameter in the risk assessment. S3.4, Based on the entropy weight of each parameter, construct a weighted decision matrix. By calculating the distance between each cluster category and the ideal solution and the worst solution, obtain the proximity of each cluster category. The proximity is used to measure the degree to which the flight state of the cluster category is close to the ideal safe state.

5. The data-driven flight risk assessment method according to claim 1, characterized in that, Step S4 includes the following sub-steps: S4.1, obtaining the structure and parameters of the aviation collision risk model, which includes the correlation between aircraft position, speed, heading parameters and collision probability; S4.2, The entropy weights and proximity of the flight states of each cluster category obtained in step S3 are used as input variables and integrated into the structure of the aviation collision risk model to establish new model variable relationships; S4.3, The weight coefficients of each input variable in the model are determined. The weight coefficients are set based on the actual needs of flight risk assessment and the historical data statistics of the impact of each variable on risk. S4.4 Integrate variable relationships and weighting coefficients to form an initial model framework for flight risk assessment. This framework can receive the results of flight state parameters after preprocessing and output the corresponding intermediate risk assessment values.

6. The data-driven flight risk assessment method according to claim 1, characterized in that, Step S5 includes the following sub-steps: S5.1, using the flight state data corresponding to the clustering results output by the flight state deep clustering network, the entropy weights calculated by the entropy weight approximation ideal solution sorting algorithm, and the closeness as training sample data, dividing them into training set and validation set; S5.2, inputting the training set data into the initial model framework constructed in step S4, calculating the model's predicted output through forward propagation, and comparing it with the actual risk label to obtain the loss value; S5.3, based on the loss value, using the backpropagation algorithm to adjust the model parameters, and monitoring the model performance in conjunction with the validation set data during the adjustment process; S5.4, repeating the forward and backpropagation parameter adjustment process of the training set until the model's performance on the validation set reaches the preset standard, completing the training and optimization of the flight risk assessment model.

Images