Dynamic adjustment real-time parachute landing state prediction method and system

Abstract

The invention discloses a dynamically-adjusted real-time parachute landing state prediction method, which comprises the following steps of: collecting flight state data in real time through a plurality of sensors arranged on a parachutist, and wirelessly transmitting the flight state data to a data processing unit; preprocessing the collected data, extracting key features by using a convolutional neural network, and generating a one-dimensional feature vector; the parachuting task is divided into a plurality of sub-tasks, each sub-task corresponds to different environments or states, initial weight distribution is obtained based on historical data, feature weighting is carried out on each sub-task, the weight is adjusted in real time, and a weighted comprehensive feature vector is generated; performing time sequence analysis on the weighted comprehensive feature vector by using a long-short memory neural network, and generating a prediction result of a future time point; according to the method, the environment or state change is responded in real time through a dynamic adjustment mechanism, so that the prediction model is more suitable for a complex and changeable parachuting environment.

Description

Technical Field

[0001] The present invention relates to the technical field of skydiving safety protection, and particularly to a method and system for predicting real-time parachute descent state with dynamic adjustment. Background Art

[0002] Skydiving, as an extremely challenging and exciting high-altitude sport, its safety is the focus of attention. Especially during the parachute descent process, skydivers may encounter complex and unpredictable airflow environments, among which eddy currents are a particularly important influencing factor. The existence of eddy currents may cause the skydiver to stay in the air for a longer time, increasing the time under the action of the synthetic wind in the air, thus disrupting the original planned landing trajectory. It may even cause the skydiver to drift out of the landing field, resulting in landing in complex terrains, seriously threatening landing safety. In addition, the instability and rotation of eddy currents are extremely likely to cause the parachute to swing and jolt, and in severe cases, it may cause symptoms such as dizziness and vomiting in the skydiver, and even result in operation errors, endangering the life safety of the skydiver.

[0003] In order to predict whether eddy currents will be encountered during skydiving, traditional methods mainly rely on static prediction methods based on physical models. These methods usually use fixed parameters and formulas to model and predict the key characteristics during the skydiving process. However, the actual skydiving environment is complex and variable, involving many non-linear factors and dynamic changes, such as non-linear airflow, temperature changes, rotational fluctuations, etc. These static models often have difficulty effectively describing these sudden changes or state evolutions across time steps, resulting in serious limitations on the accuracy and reliability of the prediction results. Summary of the Invention

[0004] The purpose of the present invention is to address the problems existing in the prior art, and provide a method and system for predicting real-time parachute descent state with dynamic adjustment. Through a dynamic adjustment mechanism, it can respond to changes in the environment or state during the skydiving process in real time, dynamically adjust the feature weights, and generate a weighted comprehensive feature vector, making the prediction model more adaptable to the complex and variable environment during the actual skydiving process, thereby improving the prediction accuracy and reliability.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is:

[0006] A method for predicting real-time parachute descent state with dynamic adjustment, comprising the following steps:

[0007] S1. Real-time collect a plurality of flight state data during the skydiving process through a plurality of sensors arranged on the skydiver, and wirelessly transmit these data to the data processing unit;

[0008] S2. Preprocess the collected data, and use a convolutional neural network to extract key features from the preprocessed data to generate a one-dimensional feature vector representing the skydiving environment or state;

[0009] S3. Divide the skydiving task into multiple subtasks, each subtask corresponding to a different environment or state. Based on the initial weight assignment obtained from historical data, perform feature weighting for each subtask, and adjust the weight assignment of each feature in real time based on the changes in the environmental or state features during skydiving to generate a weighted comprehensive feature vector.

[0010] S4. Use a long short-term memory neural network to perform time series analysis on the weighted comprehensive feature vector, analyze the relationship between different time steps, and generate a prediction result for future time points.

[0011] Step S1 includes:

[0012] Obtain the airflow direction around the skydiver by calculating the difference between pressure sensors at different body parts.

[0013] Monitor the environmental pressure during skydiving through an air pressure sensor.

[0014] Monitor the temperature during skydiving through a temperature sensor.

[0015] Measure the body posture of the skydiver through an attitude angle sensor, including the tilt angle and pitch angle.

[0016] Measure the descent speed of the skydiver through a speed sensor.

[0017] In step S2, the preprocessing of the collected data includes:

[0018] S2.1. Check whether there are missing values in the original data collected from each sensor on the skydiver. For the missing data records, if the records are incomplete, discard these records, and then detect and replace the outliers based on the rules of historical data.

[0019] S2.2. Align the data of different sensors to the same time reference through timestamp matching.

[0020] S2.3. For the missing points in the data, use dynamic interpolation method to fill them.

[0021] S2.4. Select key features from the sensor data, including attitude angle, speed, and wind speed, and construct these data features at each time point into a row. As time progresses, form a two-dimensional matrix.

[0022] Step S2.3 includes:

[0023] S2.3.1. Select the adjacent points g0, g1, g2, g3 around the missing point gt and their corresponding time points, where g1 and g2 are the adjacent valid values before and after gt respectively.

[0024] S2.3.2. Initialize the weights of g1 and g2 to the mean value, and the calculation formula is:

[0025] w1(t - 1) = 0.5, w2(t - 1) = 0.5;

[0026] where w1(t - 1) and w2(t - 1) are the initial weight values of g1 and g2 respectively.

[0027] S2.3.3. Calculate the change amount using the difference between adjacent point feature values:

[0028] Δg1 = |g1 - g0|, Δg2 = |g2 - g3|;

[0029] S2.3.4. Calculate the weight increments of g1 and g2 according to the change amount, and the calculation method is:

[0030] Δw1 = λfront·Δg1, Δw2 = λrear·Δg2;

[0031] where λfront and λrear are sensitivity parameters set according to historical experience;

[0032] S2.3.5. Dynamically update the weights of g1 and g2:

[0033] w1(t) = w1(t - 1) + Δw1, w2(t) = w2(t - 1) + Δw2;

[0034] S2.3.6. Normalize the weights:

[0035] w1(t) = w1(t) / (w1(t) + w2(t)), w2(t) = w2(t) / (w1(t) + w2(t));

[0036] S2.3.7. Dynamically interpolate the missing point gt according to the normalized weights, and the calculation formula is:

[0037] gt = w1(t)·g1 + w2(t)·g2.

[0038] In step S2, the preprocessing of the collected data also includes:

[0039] S2.5. Extract the minimum and maximum values of each feature from the collected raw data;

[0040] S2.6. Perform scaling processing on each data, and the calculation method is:

[0041] x' = (x - min) / (max - min),

[0042] Where x is the original data, min is the minimum value of this feature, max is the maximum value of this feature, and x' is the scaled data.

[0043] S2.7. Verify the scaled dataset to ensure that the data values of each feature are within the range of [0, 1].

[0044] Step S3 includes:

[0045] S3.1. Divide the state prediction task during skydiving into multiple subtasks f = [f1, f2, …, fn], where each subtask corresponds to a different environment or state change;

[0046] S3.2. Conduct inner-loop training on the subtasks to obtain the general initial attention weights α0 = [α1, α2, …, αn] based on historical data;

[0047] S3.3. For the features extracted by the convolutional neural network, perform feature weighting based on the initial weight assignment obtained from historical data;

[0048] S3.4. When the environment or state features during skydiving change, dynamically adjust the attention weights of the corresponding features in real time;

[0049] S3.5. Summarize and analyze the inner-loop results of multiple subtasks through the outer loop, extract the common information across tasks, and readjust the weights.

[0050] Step S3.4 includes:

[0051] S3.4.1. When the airspeed changes drastically, calculate the change value Δf1 of the air current feature;

[0052] When the relative positions and angles of the skydiver's head and limbs change sharply, calculate the change value Δf2 of the attitude feature;

[0053] When the vertical descent speed or the lateral drift speed changes suddenly, calculate the change value Δf3 of the speed feature;

[0054] When the rotation rate of the skydiver changes suddenly, calculate the change value Δf4 of the rotation feature;

[0055] S3.4.2. Normalize each change amount, and the calculation formula is as follows:

[0056]

[0057] ; where max(Δfi) is the maximum change amount of all features;

[0058] S3.4.3. Calculate the dynamic adjustment value Δαi of each feature based on the change amount, and the calculation formula is:

[0059]

[0060] Among them, λi is the sensitivity parameter for the i-th feature.

[0061] S3.4.4. Dynamically adjust the attention weights according to the difference between the current feature f(t) and the historical feature f(t-1). The calculation formula is:

[0062] α i (t) = α i (t - 1) + Δα i

[0063] ; where αi(t - 1) is the attention weight of the i-th feature at the previous moment;

[0064] S3.4.5. After dynamically adjusting the attention weights of each feature, multiply all features by their respective dynamic weights and then perform weighted summation to generate a comprehensive weighted feature vector f. The calculation formula is:

[0065]

[0066] Step S3.5 includes:

[0067] S3.5.1. According to the feature change values of each subtask, determine whether there are tasks that progress or regress together. If the change values of two tasks are always positive or negative at the same time, they belong to tasks that progress or regress together;

[0068] S3.5.2. For the subtasks identified as tasks that progress or regress together, apply a reduction strategy to reduce the weight difference between these subtasks. This is achieved through the following formula:

[0069] α i (t) = α i (t - 1) × (1 - β)

[0070] α j (t) = α j (t - 1) × (1 - β);

[0071] where αi(t) and αj(t) are two tasks that progress or regress together; β is the adjustment ratio, which is adjustable and is initially set to 0.2;

[0072] S3.5.3. Normalize the adjusted weights so that the sum of the weights of all subtasks is 1. The normalization calculation formula is as follows:

[0073]

[0074] S3.5.4. After the adjustment is completed at each time step, enter the next time step t+1, and re-detect the feature changes and make adjustments.

[0075] Step S4 includes:

[0076] S4.1. Input the weighted comprehensive feature vectors into the LSTM network in the order of time steps;

[0077] S4.2. The LSTM network generates the prediction results of the eddy current at future time points according to the input time series features; among them, the prediction results include identifying and predicting the eddy current area formed around the parachutist, evaluating the intensity of the eddy current and its impact on the parachutist.

[0078] A dynamically adjusted real-time parachuting state prediction system includes:

[0079] A data acquisition module, which is used to collect multiple flight state data during the parachuting process in real time through multiple sensors arranged on the parachutist, and wirelessly transmit these data to the data processing unit;

[0080] A feature extraction module, which uses a convolutional neural network to extract key features from the preprocessed data and generates a one-dimensional feature vector representing the state of the parachutist;

[0081] A dynamic attention adjustment module divides the parachuting task into multiple subtasks, each subtask corresponding to a different environment or state. Based on the initial weight assignment obtained from historical data, it weights the features for each subtask, and adjusts the weight assignment of each feature in real time based on the changes in the environmental or state features during the parachuting process, and generates a weighted comprehensive feature vector;

[0082] A time series analysis and prediction module uses a long short-term memory neural network to perform time series analysis on the weighted comprehensive feature vector, analyzes the relationship between different time steps, and generates prediction results for future time points based on the results of the time series analysis.

[0083] A computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the above method steps.

[0084] Compared with the prior art, the beneficial effects of the present invention are:

[0085] 1. Through the dynamic adjustment mechanism, it can respond to the changes in the environment or state during the parachuting process in real time, dynamically adjust the feature weights, and generate a weighted comprehensive feature vector, making the prediction model more adaptable to the complex and changeable environment during the actual parachuting process, thereby improving the prediction accuracy and reliability;

[0086] 2. Divide the state prediction task during skydiving into multiple subtasks. Through a hierarchical loop training strategy, various situations during skydiving can be captured more meticulously, improving the adaptability of the model.

[0087] 3. The prediction results generated by the LSTM network include identifying and predicting the vortex regions formed around the skydiver, evaluating the intensity of the vortices and their impact on the skydiver. The skydiver can more accurately judge the current safety situation, make the next safety decision, and reduce the risks during skydiving. Description of the Drawings

[0088] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0089] Figure 1 It is the overall flowchart of the prediction method in the embodiments of the present application;

[0090] Figure 2 It is the schematic diagram of the method for extracting one-dimensional feature vectors in the embodiments of the present application;

[0091] Figure 3 It is the schematic diagram of the method for the dynamic attention adjustment mechanism in the embodiments of the present application. Detailed Embodiments

[0092] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0093] The sequence numbers of the steps in the specification of the present application do not indicate the order of execution. The execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0094] In the description of the specification and the appended claims of the present application, the terms "first", "second", "third", etc. are only used for distinguishing descriptions and should not be construed as indicating or implying relative importance. It should also be understood that although the terms "first", "second", etc. are used in the text in some embodiments of the present application to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, the first table can be named the second table, and similarly, the second table can be named the first table without departing from the scope of the various described embodiments. The first table and the second table are both tables, but they are not the same table.

[0095] The reference to "one embodiment" or "some embodiments" etc. described in the specification of the present application means that a specific feature, structure or characteristic described in connection with the embodiment is included in one or more embodiments of the present application. Thus, the statements "in one embodiment", "in some embodiments", "in other some embodiments", "in still other embodiments" etc. that appear in different places in this specification do not necessarily all refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in another way. The terms "comprising", "including", "having" and their variants all mean "including but not limited to", unless otherwise specifically emphasized in another way.

[0096] In order to predict whether a vortex will be encountered during a parachute jump, traditional methods mainly rely on static prediction methods based on physical models. These methods usually use fixed parameters and formulas to model and predict the key characteristics during the parachute jump. However, the actual parachute jump environment is complex and variable, involving many non-linear factors and dynamic changes, such as non-linear airflows, temperature changes, rotational fluctuations, etc. These static models often have difficulty effectively describing these sudden changes or the state evolution across time steps, resulting in serious limitations on the accuracy and reliability of the prediction results.

[0097] In view of the above technical problems, as Figure 1 shown, an embodiment of the present application provides a method for predicting the real-time parachute descent state with dynamic adjustment, including the following steps:

[0098] S1. Real-time collect a plurality of flight state data during the parachute jump through a plurality of sensors arranged on the parachutist, and wirelessly transmit these data to the data processing unit.

[0099] S2. Preprocess the collected data, and use a convolutional neural network to extract key features from the preprocessed data to generate a one-dimensional feature vector representing the parachute jump environment or state.

[0100] S3. Divide the parachuting mission into multiple subtasks, where each subtask corresponds to a different environment or state. Based on the initial weight allocation obtained from historical data, perform feature weighting for each subtask, and adjust the weight allocation of each feature in real time based on the changes in the environmental or state features during the parachuting process to generate a weighted comprehensive feature vector.

[0101] S4. Use a long short-term memory neural network to perform time series analysis on the weighted comprehensive feature vector, analyze the relationships between different time steps, and generate a prediction result for future time points.

[0102] In the above method of this embodiment, through the dynamic adjustment mechanism, it responds to the changes in the environment or state during the parachuting process in real time, dynamically adjusts the feature weights, and generates a weighted comprehensive feature vector, making the prediction model more adaptable to the complex and changeable environment in the actual parachuting process, thereby improving the prediction accuracy and reliability.

[0103] Step S1 includes:

[0104] The direction of the airflow around the parachutist can be obtained by calculating the difference between the pressure sensors at different parts of the body.

[0105] The environmental pressure during the parachuting process can be monitored by an air pressure sensor.

[0106] The temperature during the parachuting process can be monitored by a temperature sensor.

[0107] The body posture of the parachutist, including the tilt angle and pitch angle, can be measured by an attitude angle sensor.

[0108] The descending speed of the parachutist can be measured by a speed sensor.

[0109] In step S2, the preprocessing of the collected data includes:

[0110] S2.1. Check whether there are missing values in the original data collected from each sensor on the parachutist. For the missing data records, if the records are incomplete, discard these records, and then detect and replace the outliers based on the rules of historical data.

[0111] S2.2. Align the data from different sensors to the same time base through timestamp matching; to eliminate the time misalignment problem caused by different sensor sampling rates or transmission delays.

[0112] S2.3. For the missing points in the data, use dynamic interpolation to fill them to ensure the continuity and integrity of the data.

[0113] S2.4. Select key features from the sensor data, including attitude angles, speed, and wind speed. Construct these data features at each time point into a row, and as time progresses, form a two-dimensional matrix, thereby providing structured data input for subsequent feature extraction and time series analysis.

[0114] Through the above preprocessing process, the handling of missing values and outliers ensures the accuracy of the data, the timestamp matching eliminates the problem of time misalignment, and the dynamic interpolation method fills in the missing data points, ensuring the continuity of the data. The selection of key features and the construction of the two-dimensional matrix make the data structured, facilitating subsequent feature extraction by the convolutional neural network and time series analysis by the long short-term memory neural network.

[0115] Step S2.3 includes:

[0116] S2.3.1. Select adjacent points g0, g1, g2, g3 around the missing point gt, and their corresponding time points, where g1 and g2 are the adjacent valid values before and after gt respectively;

[0117] S2.3.2. Initialize the weights of g1 and g2 to the mean, and the calculation formula is:

[0118] w1(t - 1) = 0.5, w2(t - 1) = 0.5;

[0119] where w1(t - 1) and w2(t - 1) are the initial weight values of g1 and g2 respectively; assume that the weights are equal in the case of no feature change.

[0120] S2.3.3. Calculate the change amount using the difference between the feature values of adjacent points:

[0121] Δg1 = |g1 - g0|, Δg2 = |g2 - g3|;

[0122] Thereby evaluating the change trend of the data near the missing point.

[0123] S2.3.4. Calculate the weight increments of g1 and g2 according to the change amount, and the calculation method is:

[0124] Δw1 = λfront·Δg1, Δw2 = λrear·Δg2;

[0125] where λfront and λrear are sensitivity parameters set according to historical experience to reflect the degree of emphasis on data changes in historical experience.

[0126] S2.3.5. Dynamically update the weights of g1 and g2:

[0127] w1(t) = w1(t - 1) + Δw1, w2(t) = w2(t - 1) + Δw2;

[0128] S2.3.6. Normalized weight:

[0129] w1(t) = w1(t) / (w1(t) + w2(t)), w2(t) = w2(t) / (w1(t) + w2(t));

[0130] S2.3.7. According to the normalized weight, perform dynamic interpolation on the missing point gt, and the calculation formula is:

[0131] gt = w1(t)·g1 + w2(t)·g2.

[0132] By using the above dynamic interpolation method, the change trend is reflected by dynamically adjusting the weights of adjacent points, so that the interpolation result is more in line with the change law of the actual data.

[0133] In step S2, the preprocessing of the collected data also includes:

[0134] S2.5. Extract the minimum and maximum values of each feature from the collected original data;

[0135] S2.6. Perform scaling processing on each data, and the calculation method is:

[0136] x' = (x - min) / (max - min),

[0137] where x is the original data, min is the minimum value of the feature, max is the maximum value of the feature, and x' is the scaled data.

[0138] S2.7. Verify the scaled data set to ensure that the data values of each feature are within the range of [0, 1].

[0139] Through the above scaling processing of the feature values, the influence caused by the differences in dimension and value range between different features can be eliminated, and the consistency of the data is improved.

[0140] As Figure 2 shown, in step S2, the convolutional neural network is used to extract key features from the preprocessed data to generate a one-dimensional feature vector representing the skydiving environment or state, including:

[0141] Apply the first layer of convolution operation to the data of each time step after preprocessing, so as to calculate the local feature map.

[0142] After the first convolutional layer, stack multiple convolutional layers to gradually extract higher-level features.

[0143] Each convolutional layer performs a convolution operation on the feature map of the previous layer to extract more abstract and higher-level features.

[0144] After each convolution layer, a batch normalization operation is performed to normalize the data distribution, accelerate the training process, and prevent the model from overfitting.

[0145] The max pooling layer is used to downsample the feature maps output by the convolution layer.

[0146] Max pooling reduces the size of the feature maps by selecting the maximum value within a local region while retaining important information.

[0147] Multiple pooling operations further reduce the dimension of the feature maps, reducing the computational amount while retaining key features.

[0148] The feature maps processed by multiple convolutional and pooling operations are flattened into a one-dimensional vector.

[0149] The high-dimensional feature vector is further reduced in dimension through the fully connected layer to generate the final one-dimensional feature vector.

[0150] In the fully connected layer, the Dropout regularization technique is applied.

[0151] Dropout prevents the model from overfitting by randomly discarding a certain proportion of neurons.

[0152] After multiple convolutional, pooling, and dimensionality reduction operations, a one-dimensional feature vector is generated.

[0153] This one-dimensional feature vector contains the key features and information in the input data and is ready to be input into the subsequent dynamic attention mechanism.

[0154] As Figure 3 shown, step S3 includes:

[0155] S3.1. Divide the state prediction task during the parachuting process into multiple subtasks f = [f1, f2, …, fn], where each subtask corresponds to a different environment or state change;

[0156] S3.2. Conduct an inner loop training on the subtasks to obtain the general initial attention weights α0 = [α1, α2, …, αn] based on historical data;

[0157] S3.3. For the features extracted by the convolutional neural network, perform feature weighting based on the initial weight assignment obtained from historical data;

[0158] S3.4. When the environmental or state features change during the parachuting process, dynamically adjust the attention weights of the corresponding features in real time;

[0159] S3.5. Through the outer loop, summarize and analyze the inner loop results of multiple subtasks, extract the common information across tasks, and readjust the weights.

[0160] Through the hierarchical loop training strategy in the above steps, the overall task is divided into multiple subtasks, which can capture various situations during the parachuting process in more detail, so as to more accurately predict the state changes.

[0161] Step S3.4 includes:

[0162] S3.4.1. When the air flow velocity changes violently, calculate the change value Δf1 of the air flow characteristics;

[0163] When the relative positions and angles of the head and limbs of the parachutist change sharply, calculate the change value Δf2 of the attitude characteristics;

[0164] When the vertical falling speed or the lateral drift speed changes suddenly, calculate the change value Δf3 of the speed characteristics;

[0165] When the rotation rate of the parachutist changes suddenly, calculate the change value Δf4 of the rotation characteristics;

[0166] S3.5.2. Normalize each change amount, and the calculation formula is as follows:

[0167]

[0168] Among them, max(Δfi) is the maximum change amount of all characteristics;

[0169] S3.4.3. Calculate the dynamic adjustment value Δαi based on the change amount, and the calculation formula is:

[0170]

[0171] Among them, λi is the sensitivity parameter for the i-th characteristic.

[0172] S3.4.4. Dynamically adjust the attention weight according to the difference between the current feature f(t) and the historical feature f(t - 1), and the calculation formula is:

[0173] α i (t) = α i (t - 1) + Δα i ;

[0174] Among them, αi(t - 1) is the attention weight at the previous moment;

[0175] S3.4.5. After dynamically adjusting the attention weight of each feature, multiply all features by their respective dynamic weights and then perform weighted summation to generate a comprehensive weighted feature vector f, and the calculation formula is:

[0176]

[0177] In the above steps, the dynamic adjustment mechanism ensures that the system can respond to changes in the environment or state in real time, and the synthesized weighted feature vector f provides more comprehensive data information.

[0178] Step S3.5 includes:

[0179] S3.5.1. Determine whether there are tasks that progress or regress together according to the feature change values of each subtask. If the change values of two tasks are always both positive or both negative, they belong to tasks that progress or regress together.

[0180] Specifically, after obtaining the feature change values of each subtask, further analyze the relationships between these change values to determine whether there are some subtasks whose feature change values are always both positive or both negative, that is, their change trends are consistent. If such an association exists, then these subtasks are regarded as "tasks that progress or regress together".

[0181] S3.5.2. For the subtasks identified as tasks that progress or regress together, calculate the current weights and apply a reduction strategy to reduce the weight differences between these subtasks, where the reduction strategy is implemented by the following formula:

[0182] α i (t) = α i (t - 1) × (1 - β)

[0183] α j (t) = α j (t - 1) × (1 - β);

[0184] Where αi(t) and αj(t) are two tasks that progress or regress together; β is the adjustment ratio, which is adjustable and is initially set to 0.2.

[0185] For the subtasks identified as tasks that progress or regress together, it is necessary to adjust their feature weights to reduce the weight differences between these subtasks to reflect the commonality or association between them.

[0186] S3.5.3. Normalize the adjusted weights so that the sum of the weights of all subtasks is 1. Among them, the normalization calculation formula is as follows:

[0187]

[0188] Through normalization, it can be ensured that the weight of each subtask is between 0 and 1, and the sum of the weights of all subtasks is 1.

[0189] S3.5.4. After the adjustment is completed at each time step, enter the next time step t + 1, and re-detect the feature changes and make adjustments.

[0190] The entire dynamic adjustment process is a continuous process. At each time step, the feature change value is recalculated according to the current environmental or state feature changes, the same-in or same-out tasks are judged, the weights are readjusted, and a weighted comprehensive feature vector is generated. Then, it enters the next time step and repeats the above process, so as to be able to respond to the environmental or state changes during the parachuting process in real time, and improve the accuracy and reliability of the prediction model.

[0191] Step S4 includes:

[0192] S4.1. Input the weighted comprehensive feature vector into the LSTM network in the order of time steps;

[0193] S4.2. The LSTM network generates the prediction result of the eddy current at future time points according to the input time series features; wherein, the prediction result includes identifying and predicting the eddy current region formed around the parachutist, evaluating the intensity of the eddy current and its impact on the parachutist.

[0194] By using the LSTM network for time series analysis, the memory unit and gating mechanism of the LSTM network can process long sequence data and capture the long-term dependence relationship between features, thereby improving the accuracy and reliability of the prediction.

[0195] By identifying the eddy current region, evaluating the intensity, and predicting the impact on the parachutist, real-time environmental information is provided for the parachutist, providing a basis for making the next safety decision.

[0196] Specifically, the training process of the LSTM network is as follows:

[0197] According to the complexity of the data and the requirements of the prediction task, design the number of layers of the LSTM network and the number of neurons in each layer.

[0198] Introduce the forgetting gate, input gate and output gate mechanisms to enhance the network's ability to capture long-term dependence relationships.

[0199] Extract the key features related to the eddy current around the parachutist, such as speed, pressure, temperature, and eddy current intensity, and use these features as the input of the LSTM network.

[0200] During the training process of the LSTM network, by learning the formation, development and dissipation patterns of eddy currents in historical data, an understanding of the dynamic changes of eddy currents is established.

[0201] Use the backpropagation algorithm and optimizer to adjust the network parameters to minimize the prediction error.

[0202] On the other hand, an embodiment of the present application provides a real-time parachuting state prediction system for dynamic adjustment, including:

[0203] A data acquisition module, configured to collect multiple flight state data during a parachuting process in real time through a plurality of sensors arranged on a parachutist, and wirelessly transmit the data to a data processing unit;

[0204] A feature extraction module, which uses a convolutional neural network to extract key features from the preprocessed data and generates a one-dimensional feature vector representing the state of the parachutist;

[0205] A dynamic attention adjustment module divides a parachuting task into multiple subtasks, each subtask corresponding to a different environment or state. Based on the initial weight assignment obtained from historical data, feature weighting is performed for each subtask, and based on the changes in the environmental or state features during the parachuting process, the weight assignment of each feature is adjusted in real time to generate a weighted comprehensive feature vector;

[0206] A time series analysis and prediction module uses a long short-term memory neural network to perform time series analysis on the weighted comprehensive feature vector, analyzes the relationship between different time steps, and based on the results of the time series analysis, generates a prediction result for a future time point.

[0207] On the other hand, an embodiment of the present application provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the above method steps are implemented.

[0208] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dynamically adjusted real-time parachute state prediction method, characterized in that: The following steps are involved: S1, collecting multiple flight status data during the parachuting process in real time through multiple sensors arranged on the parachutist, and wirelessly transmitting these data to the data processing unit; S2. Preprocess the collected data, extract key features from the preprocessed data using a convolutional neural network, and generate a one-dimensional feature vector representing the skydiving environment or state; S3, dividing the parachuting task into multiple subtasks, each subtask corresponding to a different environment or state, using the initial weight distribution obtained from historical data as a basis, weighting the features of each subtask, and adjusting the weight distribution of each feature in real time based on the changes in the environment or state features during the parachuting process, to generate a weighted comprehensive feature vector; S4. Use the long short-term memory neural network to perform time series analysis on the weighted comprehensive feature vector, analyze the relationship between different time steps, and generate prediction results for future time points.

2. A dynamically adjusted real-time parachute state prediction method according to claim 1, characterized in that: Step S1 includes: By calculating the difference between the pressure sensors in different parts of the body, the direction of the airflow around the skydiver is obtained; Monitor the ambient pressure during skydiving through air pressure sensors; Monitor the temperature during skydiving through temperature sensors; The attitude angle sensor is used to measure the body posture of the skydiver, including the tilt angle and pitch angle. The speed sensor measures the parachutist's descent speed.

3. A dynamically adjusted real-time parachute state prediction method according to claim 1, characterized in that: In step S2, preprocessing the collected data includes: S2.

1. Check whether there are missing values ​​in the raw data collected from the sensors on the skydivers. For missing data records, if the records are incomplete, discard these records and detect and replace abnormal values ​​based on the rules of historical data. S2.2, align the data of different sensors to the same time base through timestamp matching; S2.

3. For missing points in the data, dynamic interpolation is used to fill them; S2.

4. Select key features from the sensor data, including attitude angle, speed, and wind speed, and construct these data features at each time point into a row, forming a two-dimensional matrix as time progresses.

4. A dynamically adjusted real-time parachute state prediction method according to claim 3, characterized in that: Step S2.3 includes: S2.3.1, select the adjacent points g0, g1, g2, g3 on the left and right of the missing point gt, and their corresponding time points, where g1 and g2 are the previous and next adjacent valid values ​​of gt respectively; S2.3.

2. Initialize the weights of g1 and g2 to the mean, and the calculation formula is: w1(t-1)=0.5, w2(t-1)=0.5; Among them, w1(t-1) and w2(t-1) are the initial weight values ​​of g1 and g2 respectively. S2.3.

3. Calculate the change using the difference between the eigenvalues ​​of adjacent points: Δg1=|g1-g0|, Δg2=|g2-g3|; S2.3.

4. Calculate the weight increment of g1 and g2 according to the change. The calculation method is: Δw1=before λ·Δg1, Δw2=after λ·Δg2; Among them, λ before and λ after are sensitivity parameters set according to historical experience; S2.3.

5. Dynamically update the weights of g1 and g2: w1(t)=w1(t-1)+Δw1, w2(t)=w2(t-1)+Δw2; S2.3.6, Normalized weights: w1(t)=w1(t) / (w1(t)+w2(t)), w2(t)=w2(t) / (w1(t)+w2(t)); S2.3.

7. According to the normalized weights, the missing points gt are dynamically interpolated. The calculation formula is: gt=w1(t)·g1+w2(t)·g2.

5. A dynamically adjusted real-time parachute state prediction method according to claim 3, characterized in that: In step S2, preprocessing the collected data further includes: S2.5, extract the minimum and maximum values ​​of each feature from the collected raw data; S2.

6. Scale each data, the calculation method is: x' = (x-min) / (max-min); Where x is the original data, min is the minimum value of the feature, max is the maximum value of the feature, and x' is the scaled data. S2.

7. Verify the scaled dataset to ensure that the data value of each feature is in the range [0,1].

6. A dynamically adjusted real-time parachute state prediction method according to claim 1, characterized in that: Step S3 includes: S3.1, dividing the state prediction task during parachuting into multiple subtasks f = [f1, f2, ..., fn], each subtask corresponds to a different environment or state change; S3.2, perform inner loop training on the subtasks, and obtain the general initial attention weight α0=[α1,α2,…,αn] based on historical data; S3.

3. For the features extracted by the convolutional neural network, feature weighting is performed based on the initial weight distribution obtained from historical data; S3.

4. When the environment or state characteristics change during the parachuting process, the attention weight of the corresponding characteristics is adjusted dynamically in real time; S3.

5. Summarize and analyze the inner loop results of multiple subtasks through the outer loop, extract the common information across tasks, and readjust the weights.

7. A dynamically adjusted real-time parachute state prediction method according to claim 6, characterized in that: Step S3.4 includes: S3.4.

1. When the airflow velocity changes dramatically, calculate the change value Δf1 of the airflow characteristic; When the relative positions and angles of the head and limbs of the parachutist change dramatically, the change value Δf2 of the posture feature is calculated; When the vertical falling speed or the lateral drifting speed changes suddenly, the change value Δf3 of the speed characteristic is calculated; When the skydiver's rotation rate suddenly changes, the change value Δf4 of the rotation characteristic is calculated; S3.4.

2. Normalize each change amount and calculate it using the following formula: Among them, max(Δfi) is the maximum change of all features; S3.4.

3. Calculate the dynamic adjustment value Δαi of each feature based on the change. The calculation formula is: where λi is the sensitivity parameter for the i-th feature. S3.4.

4. Dynamically adjust the attention weight according to the difference between the current feature f(t) and the historical feature f(t-1). The calculation formula is: a i (t)=a i (t-1)+Da i Among them, αi(t-1) is the attention weight of the i-th feature at the previous moment; S3.4.

5. After dynamically adjusting the attention weight of each feature, all features are multiplied by their respective dynamic weights and weighted summed to generate a comprehensive weighted feature vector f. The calculation formula is:

8. A dynamically adjusted real-time parachute state prediction method according to claim 6, characterized in that: Step S3.5 includes: S3.5.

1. Based on the characteristic change value of each subtask, determine whether there are tasks that advance or retreat together. If the change values ​​of two tasks are always positive or negative, they belong to tasks that advance or retreat together. S3.5.

2. For subtasks identified as co-advancing or co-regressing tasks, a reduction strategy is applied to reduce the weight difference between these subtasks, which is achieved through the following formula: a i (t)=a i (t-1)×(1-β) a j (t)=a j (t-1)×(1-β); Among them, αi(t) and αj(t) are two tasks that advance or retreat at the same time; β is the adjustment ratio, which is adjustable and initially set to 0.2; S3.5.

3. Normalize the adjusted weights so that the sum of the weights of all subtasks is 1. The normalization calculation formula is as follows: S3.5.

4. After completing the adjustment at each time step, proceed to the next time step t+1, re-detect feature changes and make adjustments.

9. A dynamically adjusted real-time parachute state prediction method according to claim 1, characterized in that: Step S4 includes: S4.1, input the weighted comprehensive feature vector into the LSTM network in time step order; S4.

2. The LSTM network generates vortex prediction results for future time points based on the input time series characteristics; the prediction results include identifying and predicting the vortex area formed around the skydiver, evaluating the intensity of the vortex and its impact on the skydiver.

10. A dynamically adjusted real-time parachute state prediction system, characterized in that: include: A data acquisition module is used to collect multiple flight status data during the parachuting process in real time through multiple sensors arranged on the parachutist, and transmit these data wirelessly to a data processing unit; The feature extraction module uses a convolutional neural network to extract key features from the preprocessed data and generate a one-dimensional feature vector representing the state of the skydiver; The dynamic attention adjustment module divides the parachuting task into multiple subtasks. Each subtask corresponds to a different environment or state. Based on the initial weight distribution obtained from historical data, the feature weighting is performed for each subtask. Based on the changes in the environment or state characteristics during the parachuting process, the weight distribution of each feature is adjusted in real time to generate a weighted comprehensive feature vector. The time series analysis and prediction module uses a long short-term memory neural network to perform time series analysis on the weighted comprehensive feature vector, analyzes the relationship between different time steps, and generates prediction results for future time points based on the results of the time series analysis.

Images