A method for determining the starting point of an airdrop and its application

Abstract

This application belongs to the field of aerial object delivery technology, and particularly relates to a method and application for determining airdrop starting points. The method includes: collecting historical airdrop cargo data; simultaneously acquiring the time-wind field matrix of historical airdrop cargo from cargo ejection to landing; standardizing the historical airdrop cargo data, and constructing a training set and a validation set based on the time-wind field matrix; training a machine learning model using the training set, and then validating and iteratively optimizing it using the validation set to obtain an airdrop starting point prediction model; using the airdrop starting point prediction model to predict the potential landing points corresponding to airdrops at various airdrop starting points; recording potential landing points within a preset landing range as backup landing points; recording the airdrop starting points corresponding to the backup landing points as backup airdrop starting points, and outputting the backup airdrop starting points. This application can significantly improve the accuracy of landing point prediction.

Description

Technical Field

[0001] This application belongs to the field of aerial object delivery technology, and in particular relates to a method for determining the starting point of an airdrop and its application. Background Technology

[0002] Aerial object delivery (hereinafter referred to as airdrop), compared to water and land transportation, has advantages such as high mobility, timeliness, high efficiency, and less susceptibility to terrain. Airdrops can be divided into powered and unpowered airdrops based on whether they are equipped with guidance equipment. The unpowered parachute airdrop process includes stages such as exiting the aircraft, opening the tow parachute, inflating and deploying the main parachute, and the slow descent of the airdropped object until landing. The opening of the tow parachute and the inflation and deployment of the main parachute are closely related to the type of parachute used, and the movement of the cargo after the main parachute deploys is mainly affected by the laminar winds at different altitudes.

[0003] For unpowered airdrops, existing technologies typically determine a preset landing point based on the location of the cargo recipient, then determine a preset landing range based on the preset landing point, and finally subjectively determine the airdrop starting point based on wind measurement data from wind-measuring radar and training experience.

[0004] However, when airdropping cargo, after exiting the cargo hold, the pilot parachute pulls out the parachute lines and canopy to guide the main parachute to deploy properly, generating air resistance. During the descent, because the main parachute is large when fully deployed, it is easily affected by the laminar winds at different altitudes, the airdrop starting point, and the transport aircraft's attitude (including the transport aircraft's flight speed and attitude at the time of deployment). Therefore, the actual landing point corresponding to the airdrop starting point determined by existing technology is mostly outside the preset landing range. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of the prior art and provide a method for determining the airdrop starting point, which can significantly improve the accuracy of the landing point prediction.

[0006] To achieve the above objectives, this application adopts the following technical solution: A method for determining the starting point of an airdrop includes the following steps: Step 1: Collect historical airdrop cargo data; simultaneously obtain the time-wind field matrix of historical airdrop cargo from cargo leaving the cargo hold to landing. Step 2: After standardizing the historical airdrop cargo data, construct the training set and validation set based on the time-wind field matrix; Step 3: After training the machine learning model using the training set, validate and iteratively optimize it using the validation set to obtain the airdrop starting point prediction model; Step 4: Use the airdrop start point prediction model to predict the potential landing point of the airdropped cargo at each airdrop start point. Step 5: Record potential landing points within the preset landing range as backup landing points; record the airdrop starting point corresponding to the backup landing point as the backup airdrop starting point, and output the backup airdrop starting point.

[0007] Preferably, replace step 5 with step 5': In step 5', the potential landing point that is within the preset landing range and is closest to the preset landing point in a straight line is recorded as the best landing point; the airdrop starting point corresponding to the best landing point is recorded as the best airdrop starting point, and the best airdrop starting point is output.

[0008] Preferably, in step 1, the time-wind field matrix corresponding to the trajectory point of a historical airdropped cargo during the sampling period is obtained through a wind field feature extraction model, including the following steps: Step 11: After pre-training the convolutional neural network model using historical wind field observation data, a wind field feature extraction model is obtained. Step 12: Input the historical wind field observation data of a certain trajectory point during the sampling period into the wind field feature extraction model. The wind field feature extraction model outputs several key wind field vectors within the sampling period in chronological order to form the time-wind field matrix corresponding to the trajectory point during the sampling period.

[0009] Preferably, step 11 includes the following sub-steps: Step 111: After obtaining historical wind field observation data of several adjacent vertical wind layers within a continuous historical sampling period, the historical wind field observation data is preprocessed. Step 112: Divide the preprocessed historical wind field observation data into a first dataset and a second dataset; Step 113: After training the CNN convolutional neural network model using the first dataset, the first model is obtained; Step 114: After validating the first model using the second dataset, the wind field feature extraction model is obtained.

[0010] Preferably, step 113 includes the following sub-steps: Step 1131: After labeling each historical wind field observation data in the first dataset with key wind field features, the key wind field vector corresponding to each historical wind field observation data is obtained; the key wind field vector serves as the label of the corresponding historical wind field observation data, and together with the corresponding historical wind field observation data, constitutes the sample in the first training set. Step 1132: Train the CNN convolutional neural network model using the first training set, so that the CNN convolutional neural network model has the function of converting wind field observation data into key wind field vectors after training; denote the trained CNN convolutional neural network model as the first model; Step 114 also includes the following: inputting wind field observation data from the second dataset into the first model to test the accuracy of the first model in converting wind field observation data into key wind field vectors; then verifying the accuracy of the key wind field vectors output by the first model; if the accuracy is above C1, then the first model at this time is recorded as the wind field feature extraction model; if the accuracy is less than C1, then the parameters of the first model are optimized along the gradient reduction direction of the mean squared error loss function, and then returning to S132, retraining and validating the CNN convolutional neural network model using the first training set and the second dataset until the accuracy of the first model is above C1.

[0011] Preferably, in step 2: the airdrop trajectory data of historical airdropped goods is subjected to Min-Max standardization to eliminate the dimensional differences between different airdrop trajectory data, and then a third training set and a third validation set are constructed; Step 3 also includes the following sub-steps: Step 31: Use the third training set to train the machine learning model so that after training, the machine learning model can predict the airdrop trajectory data of the next trajectory point of the airdrop cargo based on the airdrop cargo quality, weather, umbrella type, deployment attitude, airdrop trajectory data of the known trajectory points of the airdrop cargo, and the time-wind field matrix corresponding to the trajectory points; the trained machine learning model is denoted as the third model. Step 32: Use the third validation set to validate the third model, and compare the validation data labels output by the third model with the real labels of the corresponding validation data. If the accuracy is ≥ A1, then the third model at this time is recorded as the airdrop starting point prediction model and can be officially put into use; otherwise, after optimizing the model parameters along the direction of decreasing gradient of the mean square error loss function, return to step 31 and retrain and test the third model using the training set and validation set until the accuracy of the third model reaches the target. Step 4 also includes the following sub-steps: Step 41: Obtain the longitude, latitude, and altitude of each air position point within the range of the transport aircraft's forward movement, and calculate the time corresponding to the current arrival of the transport aircraft at each air position point; use these times as the timestamps of the corresponding air position points. Step 42: Each aerial location is designated as the cargo drop point. Based on the cargo drop point, cargo mass, weather, parachute type, drop attitude, and the future time-wind field matrix during the descent of cargo from different drop points, the airdrop origin prediction model is used to predict the trajectory data of the cargo falling from different drop points. The landing points corresponding to the predicted cargo drop points are recorded as potential landing points.

[0012] Preferably, in step 42, the future time-wind field matrix of each trajectory point during the descent of the airdropped cargo is predicted using a spatiotemporal wind prediction model, including the following steps: S1. After pre-training the wind field feature prediction model using a multi-layer long short-term memory network model based on historical wind field observation data, a wind field feature prediction model is obtained. The wind field feature extraction model and the wind field feature prediction model are connected in series to form a spatiotemporal wind prediction model. S2 collects wind field observation data of the predicted trajectory points in real time and sends it to the spatiotemporal wind prediction model. The spatiotemporal wind prediction model outputs the future time-wind field matrix of the corresponding trajectory points.

[0013] Preferably, step S1 further includes the following sub-steps: S11, using a wind field feature extraction model, the historical wind field observation data in the first dataset is transformed into corresponding key wind field vectors; then, a time sliding window is used to package the key wind field vectors at the same location point into sliding window vectors in chronological order, which are then recorded as the second training set; each sliding window vector contains k key wind field vectors at the same location point arranged in chronological order, and the sliding window slides in the direction of increasing time, with each step of the sliding window being m key wind field vectors; simultaneously, the wind field feature extraction model is used to transform the wind field observation data in the second dataset into a second validation set containing several consecutive sliding window vectors; k is a positive integer; S12, Use the second training set to train the multilayer long short-term memory network model, so that the multilayer long short-term memory network model can predict the future key wind field vector at a certain location point based on the sliding window vector after training; denote the trained multilayer long short-term memory network model as the second model; S13, Determine the wind field feature prediction model: Input the sliding window vector of the second validation set into the second model in chronological order, and calculate the prediction accuracy of the second model for the next key wind field vector; if the accuracy is above C2, then the second model at this time is recorded as the wind field feature prediction model; if the accuracy is less than C2, then optimize the model parameters along the gradient decreasing direction of the cross-entropy loss function and return to S12, and retrain and test the multilayer long short-term memory network model using the second training set and the second validation set until the accuracy of the second model is above C2.

[0014] Preferably, in S12, the multilayer long short-term memory network model is trained using the second training set, including the following: Let the key wind field vector at a certain location at the current moment be . The time sliding window formed by the current moment and the k key wind field vectors before that moment... ,in Let J represent the j-th key wind field vector in sliding window Y, where 1 ≤ j ≤ k and j and k are both positive integers; in the key wind field vector There is ,in Represents the key wind field vector Data in the i-th dimension; Let the current time be t, the previous time be t-1, and the next time be t+1. The sliding window corresponding to a certain position at the current time is: Then the current time window will slide. As the input to the multilayer long short-term memory network model at the current time step; based on the following formula, the multilayer long short-term memory network model is iteratively trained using the second training set, continuously updating the cell states and hidden layer states in the multilayer long short-term memory network model: ; In the multilayer long short-term memory network model, the cell state at the previous time step is: The hidden state at the previous time step was ; This indicates the output of the forget gate at the current moment; (·) represents the sigmoid activation function; The weight matrix representing the forget gate; Represents the hidden state at the previous time step. Slide window at the current moment splicing; This indicates that the bias term of the door has been forgotten; This represents the input gate output at the current time t; Represents the candidate cell state at time t; tanh(·) represents the hyperbolic tangent activation function; A weight matrix representing the state of candidate cells; Bias terms representing the state of candidate cells; This indicates the output of the gate at the current time t; This represents the weight matrix of the output gate; The output gate's bias term; the hidden state at the current time step. As the output of the multilayer long short-term memory network model at the current time; the cell state at the current time t within the multilayer long short-term memory network model is: .

[0015] This application also provides an application of an airdrop starting point determination method: when the transport aircraft driver reaches the backup airdrop starting point or the optimal airdrop starting point, the transport aircraft driver performs a non-powered airdrop; the backup airdrop starting point or the optimal airdrop starting point is obtained using an airdrop starting point determination method as described above.

[0016] The beneficial effects of this application are as follows: (1) The method for determining the airdrop starting point of this application predicts the landing point corresponding to different aerial locations. Based on the relationship between the predicted landing point and the preset landing range, it inversely deduces which aerial locations can be used as airdrop starting points. By using historical airdrop cargo data and the corresponding future time-wind field matrix, a machine learning model is trained to obtain an airdrop starting point prediction model. Based on the airdrop cargo quality, weather, umbrella type, and deployment attitude, the airdrop starting point prediction model can predict the trajectory data of the next trajectory point based on the trajectory data of the previous trajectory point and the corresponding future time-wind field matrix, thereby predicting the landing point of the current airdrop cargo. Landing points within the preset landing range are recorded as potential landing points, and there may be multiple potential landing points. Among the potential landing points, the one closest to the preset landing point in a straight line is recorded as the best landing point.

[0017] (2) The airdrop starting point determination method of this application efficiently and accurately determines the airdrop starting point: In the training and prediction process, not only the trajectory data of the aerial position point, the quality of the dropped goods, the weather, the umbrella type, and the drop posture are considered, but also the influence of the wind field spatial feature vector that changes with time during the airdrop process (i.e., the importance of the time-wind field matrix); Therefore, in the process of predicting the landing point, this application also includes the prediction of the future time-wind field matrix, which is not available in the prior art, further improving the accuracy of the airdrop starting point prediction model in predicting the trajectory data of the next trajectory point; since each trajectory point is the basis for predicting the next trajectory point, this application improves the accuracy of the landing point prediction on the premise of ensuring the accurate prediction of each trajectory point; thus, it can ensure that the output airdrop starting point meets the requirements.

[0018] (3) The method for determining the airdrop starting point in this application does not rely on meteorological sensors and kinematic models. It constructs an airdrop starting point prediction model based solely on historical airdrop cargo data and the corresponding time-wind field matrix, and predicts the landing point corresponding to airdrops at multiple spatial locations. Furthermore, it uses the actual data obtained from airdrops based on the prediction results to further optimize the prediction model, thereby improving the airdrop accuracy of subsequent transport aircraft and significantly reducing the complexity and cost of the prediction system.

[0019] (4) The prediction of all trajectory points in the landing process in this application is always accompanied by the coupling between the output of the airdrop starting point prediction model and the output of the spatiotemporal wind prediction model.

[0020] (5) This application uses a spatiotemporal wind prediction model to predict the future time-wind field matrix of each trajectory point during the descent of the airdropped cargo from the cargo drop point: ① First, a CNN convolutional neural network model is trained into a wind field feature extraction model. Then, the wind field feature extraction model is used to obtain the training data (sliding window vectors) needed to train a multilayer long short-term memory network model. Finally, the sliding window vectors are used to train the multilayer long short-term memory network model into a wind field feature prediction model. The entire training process is compact and interconnected, significantly reducing the time and manpower costs for technicians to manually construct training and validation sets, and significantly improving the training efficiency of the wind field feature prediction model.

[0021] ② Achieve accurate prediction in both time and space dimensions: In the technical solution of this application, the wind field prediction data output for a future period of time has the validity of both time and space dimensions: In the time dimension, only key wind field vectors with a probability of C3 or higher are output to ensure prediction accuracy; In the spatial dimension, through the fusion and dimensionality reduction of the wind field feature extraction model, the core features of different wind layers are retained, so that the wind field prediction data includes the features of the wind field in different spaces, that is, to achieve wind feature prediction in both time and space dimensions.

[0022] (6) This application can predict wind field data for a future period of time at different vertical wind layers in a timely and accurate manner, and the resulting future time-wind field matrix for each trajectory point also balances lightweightness and accuracy. To ensure prediction accuracy, this application requires that the input wind field observation data include dense wind layers (the original data is accurate and contains a large amount of information); at the same time, through the fusion and dimensionality reduction of the wind field feature extraction model, the number of wind layers in the key wind field vectors is reduced without losing core information, thereby achieving lightweight prediction data: ① At the transmission level: Lightweight data can avoid problems such as transmission interruption and excessive time consumption, ensuring timeliness and integrity.

[0023] ② At the computational level: Subsequent calculations based on lightweight data (such as airdrop trajectory simulation) have significantly reduced overhead and time consumption. Moreover, because the core features are not lost, the calculation results still have a high accuracy rate and can directly support parameter decisions for air cargo transportation.

[0024] (7) The application of the method for determining the airdrop starting point in this application can greatly improve the accuracy of airdropped goods falling into the preset landing range. Attached Figure Description

[0025] Figure 1 This is a flowchart of a method for determining the airdrop starting point according to this application; Figure 2 The spatial trajectory curves of the same airdropped cargo obtained by airdropping the same cargo at the same exit point under different transport aircraft flight speeds; Figure 3 The spatial trajectory curves obtained by airdropping the same cargo at different altitude exit points under the same transport aircraft speed; Figure 4 This is a schematic diagram of the structure of a random forest regression model. Detailed Implementation

[0026] To make the technical solution of this application clearer and more explicit, the application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Solutions derived by those skilled in the art through equivalent substitution and conventional reasoning of the technical features of the technical solution of this application without creative effort all fall within the protection scope of this application.

[0027] like Figure 1 The diagram shown is a flowchart of a method for determining the airdrop starting point according to this application, including: Step 1: Collect historical airdrop cargo data; simultaneously obtain the time-wind field matrix of historical airdrop cargo from cargo leaving the cabin to landing.

[0028] Step 2: After standardizing the historical airdrop cargo data, construct the third training set and the third validation set based on the time-wind field matrix.

[0029] Step 3: After training the machine learning model using the third training set, the model is then validated and iteratively optimized using the third validation set to obtain the airdrop starting point prediction model.

[0030] Step 4: Use the airdrop start point prediction model to predict the potential landing point of the airdropped cargo at each airdrop start point.

[0031] Step 5: Record potential landing points within the preset landing range as backup landing points; record the airdrop starting point corresponding to the backup landing point as the backup airdrop starting point, and output the backup airdrop starting point.

[0032] Optionally, replace step 5 with step 5': Step 5': Record the potential landing point that is within the preset landing range and is closest to the preset landing point in a straight line as the best landing point; record the airdrop starting point corresponding to the best landing point as the best airdrop starting point, and output the best airdrop starting point.

[0033] Step 1 also includes the following sub-steps: By installing data collection equipment on historical airdropped cargo, we can obtain historical airdropped cargo data such as the quality of the cargo, the weather at the time of the airdrop, the type of parachute, the deployment attitude, and the airdrop trajectory data from the cargo leaving the cabin to landing.

[0034] The airdrop trajectory data includes data from several trajectory points, namely the exit point, landing point, key points at several range states, the moment of full parachute deployment, and several key points after full parachute deployment (before landing). These key points are set by technical personnel. The trajectory data includes longitude, latitude, altitude, and timestamp. Additionally, the time-wind field matrix corresponding to these trajectory points for historical airdrop cargo within the sampling period needs to be obtained.

[0035] Each point on the trajectory of an airdropped cargo during its descent has a unique timestamp. Because descent takes time, the airdropped cargo cannot reach a specific trajectory point before the designated timestamp has elapsed.

[0036] The sampling time period refers to a period of time from the corresponding timestamp of a certain trajectory point. The specific duration is set by technical personnel and generally does not exceed the time between the airdropped cargo's departure from the trajectory point and its landing.

[0037] The time-wind field matrix of a certain trajectory point refers to several key wind field vectors arranged in chronological order within the vertical wind layer space (the two ends of the vertical wind layer space are the ground and the wind layer at a certain height, respectively) of the trajectory point within the sampling time period. The first key wind field vector in the time-wind field matrix is ​​the key wind field vector of the trajectory point at the corresponding timestamp. The values ​​of each dimension of the key wind field vector are the key wind field features of the corresponding wind layer (such as radial wind speed, vertical wind speed, wind direction, etc.).

[0038] The time-wind field matrix corresponding to the trajectory points of historical airdropped cargo during the sampling period can be obtained through manual annotation by technicians.

[0039] When the total number of trajectory points of historical airdropped cargo is large, the time-wind field matrix corresponding to the trajectory points of historical airdropped cargo in the sampling period can also be obtained by using a wind field feature extraction model.

[0040] The process of obtaining the time-wind field matrix corresponding to the trajectory points of a historical airdropped cargo within a sampling period using a wind field feature extraction model includes the following steps: Step 11: After pre-training the convolutional neural network model using historical wind field observation data, a wind field feature extraction model is obtained.

[0041] Step 12: Input the historical wind field observation data of a certain trajectory point during the sampling period into the wind field feature extraction model. The wind field feature extraction model outputs several key wind field vectors within the sampling period in chronological order to form the time-wind field matrix corresponding to the trajectory point during the sampling period.

[0042] It should be noted that the historical wind field observation data for a certain historical airdrop cargo trajectory point within the sampling period are all known quantities and can be collected through a wind measurement radar system.

[0043] Step 11 also includes the following sub-steps: Step 111: After obtaining historical wind field observation data of several adjacent vertical wind layers within a continuous historical sampling period, the historical wind field observation data is preprocessed.

[0044] The acquisition of historical wind field observation data is affected by the location and sampling time. For example, with a sampling interval of 1 minute, within a continuous historical sampling period, the next data point after the wind field observation data at 15:00 on September 24, 2025, is the wind field observation data at 15:01 on September 24, 2025. Each wind field observation data includes pre-collected layer wind observation data for all wind layers. Layer wind observation data for different wind layers can be collected through a wind measurement radar system. For example, technicians pre-set a certain point on the ground plane as the base point in the current sampling time location. Starting from the base point, every 10 meters in the vertical direction is a wind layer, and the wind within the wind layer is the layer wind. Collecting layer wind observation data corresponding to 300 wind layers upward from the base point, that is, the 300 layer wind observation data collected at that location point at 15:00 on September 24, 2025, constitutes one wind field observation data.

[0045] That is, at a certain historical sampling moment, a wind field observation data consists of several consecutive layers of wind observation data starting from a certain base point on the ground plane.

[0046] The wind observation data includes, but is not limited to, radial wind speed (unit: m / s), vertical wind speed (unit: m / s), wind direction (unit: °, north is 0° and measured clockwise), the vertical height range of the wind layer (unit: m; the vertical height of the base point is 0), and relative humidity (unit: % RH).

[0047] Preprocessing includes removing gross errors, time synchronization processing, spatial interpolation, and data normalization to remove dirty data from historical wind field observation data and ensure the quality and consistency of all historical wind field observation data.

[0048] Remove gross errors: Use the 3δ criterion (data that deviates from the mean by more than 3 times the standard deviation is removed).

[0049] Time synchronization processing: unify data acquisition timestamps to eliminate device time deviations.

[0050] Spatial interpolation: Kriging interpolation is used to supplement missing wind layer data to ensure the continuity of wind layer observation data in vertical space; Data normalization: The Min-Max method is used to map data of the same type to the [0,1] interval to eliminate the influence of units.

[0051] Step 112: Divide the preprocessed historical wind field observation data into a first dataset and a second dataset.

[0052] This embodiment contains 1200 wind field observation data. These 1200 wind field observation data are divided into two datasets in a 7:3 ratio: the first 840 wind field observation data are used as the first dataset, and the last 360 wind field observation data are used as the second dataset.

[0053] Following step 112, step 113 is also included: Step 113: After training the CNN convolutional neural network model using the first dataset, the first model is obtained; Step 114: After validating the first model using the second dataset, the wind field feature extraction model is obtained.

[0054] Step 113 also includes the following sub-steps: Step 1131: After the technicians annotate each historical wind field observation data in the first dataset with key wind field features (such as radial wind speed, vertical wind speed, and wind direction of the core wind layer), they obtain the key wind field vector corresponding to each historical wind field observation data. The key wind field vector serves as the label of the corresponding historical wind field observation data and together with the corresponding historical wind field observation data, constitutes the sample in the first training set. Step 1132: Train the CNN convolutional neural network model using the first training set, so that the CNN convolutional neural network model can transform each high-dimensional wind field observation data into a low-dimensional key wind field vector with fewer dimensions after training; the trained CNN convolutional neural network model is denoted as the first model.

[0055] The CNN convolutional neural network model in this application adopts the VGG16 model, which contains 13 convolutional layers and 3 fully connected layers.

[0056] For example, a wind field observation data point is 300-dimensional (corresponding to 300 wind layers), but the corresponding key wind field vector is 100-dimensional. In this case, the data of each dimension in the key wind field vector can be used to represent the wind data of the original three consecutive wind layers. This ensures the representativeness of the key wind field vector while reducing the amount of data in the key wind field vector.

[0057] Step 114 also includes the following: The wind field observation data in the second dataset is input into the first model to test the accuracy of the first model in converting the wind field observation data into key wind field vectors; then the accuracy of the key wind field vectors output by the first model is manually verified. If the accuracy is above C1, then the first model at this time is recorded as the wind field feature extraction model, which can be put into formal use in the future. If the accuracy is less than C1, optimize the parameters of the first model (such as the convolution kernel weights) along the gradient reduction direction of the mean squared error loss function, and then return to step 132 to retrain and validate the CNN convolutional neural network model using the first training set and the second dataset until the accuracy of the first model is above C1.

[0058] In this embodiment, C1=85%.

[0059] During the training of the CNN convolutional neural network model using the first training set, the network parameters are optimized using the backpropagation algorithm.

[0060] Taking a single wind field observation dataset as an example, the wind field feature extraction model extracts all wind field spatial features (i.e., radial wind speed, vertical wind speed, wind direction, wind layer height range, and humidity) from this wind field observation dataset using multi-scale convolutional kernels to construct a wind field spatial feature vector. At this point, the number of dimensions of the wind field spatial feature vector is equal to the total number of wind layers. Then, global average pooling is used to retain the key features (i.e., radial wind speed, vertical wind speed, and wind direction) in the wind field spatial feature vector and reduce the dimensionality, transforming the wind field spatial feature vector into a key wind field vector with fewer dimensions. At this point, the number of dimensions of the key wind field vector is much smaller than the total number of wind layers.

[0061] The method for obtaining wind field prediction data in this application does not directly select 100 wind layers from 300 wind layers in a single wind field observation dataset to construct the key wind field vector. This is because such a method would only represent the wind characteristics of the selected layers, failing to represent the wind characteristics of the unselected layers, resulting in information loss and reduced representativeness. Instead, this application uses a wind field feature extraction model to learn the potential relationships between continuous (or adjacent) wind layers, fusing several consecutive wind layer data into a new wind layer dataset. This new wind layer dataset is highly representative of the fused wind layers. In this way, the 300-dimensional wind field spatial feature vector is transformed into a 100-dimensional key wind field vector. This ensures both the representativeness and overall effectiveness of the key wind field vector while also reducing the weight of the wind field observation data.

[0062] By using a wind field feature extraction model to obtain the time-wind field matrix corresponding to each trajectory point of historical airdropped cargo during the sampling period, the accuracy is high, and the process is convenient and efficient, greatly saving manpower.

[0063] It should be noted that the timestamps of the several trajectory points included in each airdrop trajectory data are different, and the vertical height gradually decreases as time goes by.

[0064] When cargo leaves the cabin but the main parachute attached to the cargo is not fully deployed, the cargo is within range of the missile.

[0065] The key points during the firing range and after the parachute is fully deployed can be sampling points pre-set by technicians during the corresponding process. For example, sampling can be performed every 10ms, and the trajectory points corresponding to the sampling are the key points.

[0066] In step 2: The airdrop trajectory data of historical airdropped cargo is standardized (using the Min-Max standardization method) to eliminate the dimensional differences between different airdrop trajectory data, and then the third training set and the third validation set are constructed.

[0067] Construct a third training set, including the following: For a given historical airdrop cargo, let's say that the historical airdrop cargo contains k+1 trajectory points, then we obtain k training data points: The quality, weather, parachute type, and drop attitude of the historical airdropped cargo, as well as the airdrop trajectory data and time-wind field matrix of the cargo's exit point (the first trajectory point), are used as the features of the first training data. The trajectory data of the historical airdropped cargo at the second trajectory point is used as the label of the first training data, thus obtaining the first training data of the historical airdropped cargo.

[0068] The features of the historical airdrop cargo's quality, weather, umbrella type, deployment attitude, and the airdrop trajectory data of the first two trajectory points of the historical airdrop cargo, along with the corresponding time-wind field matrix, are used as the features of the second training data. The trajectory data of the historical airdrop cargo at the third trajectory point is used as the label of the second training data, thus obtaining the second training data of the historical airdrop cargo.

[0069] The features of the historical airdrop cargo's quality, weather, umbrella type, deployment attitude, and the airdrop trajectory data of the first three trajectory points of the historical airdrop cargo, along with the corresponding time-wind field matrix, are used as the features of the third training data. The trajectory data of the historical airdrop cargo at the fourth trajectory point is used as the label of the third training data, thus obtaining the third training data of the historical airdrop cargo.

[0070] This process continues until the p-th training data point for the historical airdrop cargo is obtained. The label of the p-th training data point is the landing trajectory data of the historical airdrop cargo.

[0071] The same method was used to process several other historical airdropped cargoes, resulting in a training set containing n training data points.

[0072] Construct a third validation set, including the following: For a specific historical airdrop cargo, let's say that the historical airdrop cargo contains j+1 trajectory points; then we obtain j training data points, constructed as follows: The quality, weather, parachute type, and deployment attitude of the historical airdropped cargo, as well as the airdrop trajectory data and time-wind field matrix of the cargo's exit point (the first trajectory point), are used as the features of the first verification data. The trajectory data of the historical airdropped cargo at the second trajectory point is used as the true label of the first verification data, thus obtaining the first verification data of the historical airdropped cargo.

[0073] The quality, weather, umbrella type, and deployment attitude of the historical airdropped cargo, as well as the airdrop trajectory data of the first two trajectory points of the historical airdropped cargo, and the corresponding time-wind field matrix are used as the features of the second verification data; the trajectory data of the historical airdropped cargo at the third trajectory point is used as the true label of the second verification data, thus obtaining the second verification data of the historical airdropped cargo.

[0074] The quality, weather, umbrella type, and deployment attitude of the historical airdropped cargo, as well as the airdrop trajectory data of the first three trajectory points of the historical airdropped cargo, and the corresponding time-wind field matrix are used as the features of the third verification data; the trajectory data of the historical airdropped cargo at the fourth trajectory point is used as the true label of the third verification data, thus obtaining the third verification data of the historical airdropped cargo.

[0075] This process continues until the g-th verification data point for the historical airdrop cargo is obtained. The true label of the g-th verification data point is the landing point trajectory data of the historical airdrop cargo.

[0076] In this embodiment, the data in the third training set and the third validation set do not overlap.

[0077] Step 3 also includes the following sub-steps: Step 31: Use the third training set to train the machine learning model (in this scheme, a random forest regression model that uses attention and self-attention mechanisms is adopted) so that after training, the machine learning model can predict the airdrop trajectory data of the next trajectory point of the airdrop cargo based on the airdrop cargo quality, weather, umbrella type, deployment attitude, airdrop trajectory data of the known trajectory points of the airdrop cargo, and the time-wind field matrix corresponding to the trajectory points.

[0078] The trained machine learning model is denoted as the third model.

[0079] Step 32: Use the third validation set to validate the third model, and compare the validation data labels output by the third model with the real labels of the corresponding validation data. If the accuracy is ≥ A1, then the third model at this time is recorded as the airdrop start point prediction model and can be officially put into use; otherwise, the model parameters are optimized along the direction of decreasing gradient of the mean squared error loss function, and after returning to step 31, the training set and validation set are used again to train and test the third model until the accuracy of the third model reaches the target.

[0080] The number of dimensions of the wind field spatial feature vector is equal to the total number of wind layers. Therefore, whether in the training or in the actual use of the airdrop start point prediction model, once we know / predict the trajectory data of the trajectory point, we can determine which wind layer the trajectory point belongs to by the height of the trajectory point. This makes it easier for the airdrop start point prediction model to learn the influence of the wind feature data of the wind layer and the adjacent wind layers on the airdrop cargo from the current trajectory point to the next trajectory point, and use this influence to accurately predict the trajectory data of the next trajectory point.

[0081] Step 4 also includes the following sub-steps: Step 41: Obtain the longitude, latitude, and altitude of each air position point within the range of the transport aircraft's forward movement, and calculate the time corresponding to the current arrival of the transport aircraft at each air position point; use these times as the timestamps of the corresponding air position points.

[0082] It should be noted that "forward range" does not refer only to the front of the transport aircraft; rather, it refers to the spatial range that the transport aircraft can reach by flying over a period of time.

[0083] Step 42: Each aerial location is designated as the cargo drop point. Based on the cargo drop point, cargo mass, weather, parachute type, drop attitude, and the future time-wind field matrix during the descent of cargo from different drop points, the airdrop origin prediction model is used to predict the trajectory data of the cargo falling from different drop points. The landing points corresponding to the predicted cargo drop points are recorded as potential landing points.

[0084] In step 42, it should be noted that: ① Obtain the "time-wind field matrix corresponding to different airdrop cargo exit points": After knowing the timestamp of a certain airdrop cargo exit point and the wind field observation data of the airdrop cargo exit point at the current time, the key wind field vector of the airdrop cargo when it is dropped at the specific airdrop cargo exit point can be obtained by using the wind field feature extraction model.

[0085] ② The specific process of "predicting the landing point corresponding to each airdrop cargo exit point": Taking a certain airdrop cargo exit point α and airdrop cargo β as an example, the airdrop starting point prediction model predicts the trajectory data of the second trajectory point of cargo β based on the mass, weather, parachute type, deployment attitude, airdrop trajectory data of airdrop cargo exit point α, and the future time-wind field matrix corresponding to exit point α. The airdrop starting point prediction model then predicts the trajectory data of the third trajectory point of cargo β based on the mass, weather, parachute type, deployment attitude, airdrop trajectory data of the second trajectory point, and the future time-wind field matrix corresponding to the second trajectory point. This process continues, predicting the next trajectory point based on the previous one, until the trajectory data of the trajectory point with an altitude of 0 is obtained. This is the landing point corresponding to airdrop cargo exit point α, thus obtaining the potential landing point corresponding to airdrop cargo exit point α.

[0086] The future time-wind field matrix refers to a number of key wind field vectors arranged in chronological order over a period of time, starting from the timestamp corresponding to the trajectory point.

[0087] As can be seen from the above, in step 4, before the airdropped cargo is actually dropped, each trajectory point in its descent process after being dropped from different airdrop cargo drop points is obtained through prediction. Therefore, the time-wind field matrix corresponding to each trajectory point naturally refers to a period of time in the future.

[0088] This application uses a spatiotemporal wind prediction model to predict the future time-wind field matrix of each trajectory point during the descent of airdropped cargo from the cargo drop point, including the following steps: S1. After pre-training using a multi-layer long short-term memory network model based on historical wind field observation data, a wind field feature prediction model is obtained. The wind field feature extraction model and the wind field feature prediction model are connected in series to form a spatiotemporal wind prediction model.

[0089] S2 collects wind field observation data of the predicted trajectory points in real time and sends it to the spatiotemporal wind prediction model. The spatiotemporal wind prediction model outputs the future time-wind field matrix of the corresponding trajectory points.

[0090] S1 also includes the following sub-steps: S11, using a wind field feature extraction model, the historical wind field observation data in the first dataset are transformed into corresponding key wind field vectors. Then, a time sliding window is used to package the key wind field vectors at the same vertical wind layer into sliding window vectors in chronological order, which are denoted as the second training set. Each sliding window vector contains k key wind field vectors at the same vertical wind layer arranged in chronological order. The sliding window slides in the direction of increasing time, and the step size of each slide window is m key wind field vectors. Similarly, the wind field feature extraction model is used to transform the 360 ​​wind field observation data in the second dataset into a second validation set containing several consecutive sliding window vectors. k is a positive integer.

[0091] In this embodiment, k=10; m=1, and the rightward direction represents the increasing time.

[0092] S12, use the second training set to train the multilayer long short-term memory network model, so that the multilayer long short-term memory network model can predict the key wind field vector of a certain vertical wind layer space in the future based on the sliding window vector after training; the trained multilayer long short-term memory network model is denoted as the second model.

[0093] The multilayer long short-term memory network model in this embodiment adopts the SWLSTM + attention mechanism.

[0094] The training logic in S12 is: using the sliding window vector as input, learn the potential relationship between the current "sliding window" and the "next key wind field vector" in the same vertical wind layer space.

[0095] The first critical wind field vector on the right side of each sliding window is the next critical wind field vector for that window. For example, if the critical wind field vectors in a sliding window are numbered from 1 to 10 from left to right, then the first critical wind field vector on the right side of the right boundary of that window, i.e., the 11th critical wind field vector, is the next critical wind field vector for that window, and the 11th critical wind field vector is also the target we want to predict.

[0096] During training using the second training set, after the multilayer long short-term memory network model learns the potential relationships between adjacent sliding window vectors and between each sliding window and the next key wind field vector, it can predict the key wind field vector for a certain sliding window in the future. The "key wind field vector in the future" mentioned here may be more than one, which is not intended to limit this application.

[0097] In S12, “Training the multilayer long short-term memory network model using the second training set” also includes the following: Let the key wind field vector at a certain location at the current moment be . The time sliding window formed by the k key wind field vectors at the current time and before (including the key wind field vector at the current time) ,in Let represent the j-th key wind field vector in the sliding window Y, where 1 ≤ j ≤ k and j and k are both positive integers. In the key wind field vector... There is ,in Represents the key wind field vector The data in the i-th dimension.

[0098] The structure of a cell unit in a multilayer long short-term memory network model includes: ① Forget gate: determines what information to discard from the cell state at the previous time step. ② Input gate: used to update the current cell state. ③ Output gate: used to determine the value of the hidden state at the current time step.

[0099] Let the current time be t, the previous time be t-1, and the next time be t+1. The sliding window corresponding to a certain vertical wind layer at the current time is... The input to the multilayer long short-term memory network model at the current time step is: the sliding window at the current time step. Based on the following formula, the multilayer long short-term memory network model is iteratively trained using the second training set, continuously updating the cell states and hidden layer states in the multilayer long short-term memory network model: ; In the multilayer long short-term memory network model, the cell state at the previous time step is: The hidden state at the previous time step was The cell state at time t in the multilayer long short-term memory network model is: . The output of the forget gate at the current time step is a vector of length 0 to 1; it is used to determine the cell state at the previous time step. The proportion of "retained / discarded information". (·) represents the sigmoid activation function, which maps the input to the (0,1) interval to implement the "gating" logic of "on / off". This represents the weight matrix of the forget gate. Represents the hidden state at the previous time step. Slide window at the current moment splicing. This indicates the bias term of the forget gate. If the value is a large positive value, the old information in the cell state at the previous time step is retained first; if the value is a large negative value, the old information in the cell state at the previous time step is discarded first. The input gate output at the current time t is a vector of length 0 to 1; The amount of "new information (i.e., candidate cell states)" that is added to the current cell state is determined. middle. This represents the candidate cell state at the current time t, with a value range of (-1, 1), and includes the new information currently input. The input gate controls whether it is added to the current cell state. In the middle, tanh(·) represents the hyperbolic tangent activation function, which maps the input to the interval (-1, 1). The weight matrix represents the state of the candidate cells. Bias terms representing candidate cell states. The cell state at the current moment. It is the core "memory unit" of the multi-layered long and short-term memory network model, consisting of two superimposed parts: "old memories" filtered by the forget gate (i.e., . ) and the "new memory" after input gate filtering (i.e. ). The output of the gate at time t is a vector of length 0 to 1, which determines the cell state at the current time. How much information is output to the hidden state at the current moment? middle. This represents the weight matrix of the output gate. This represents the bias term of the output gate. The hidden state at the current time step. As the output at the current moment; obtained from the cell state after filtering by the output gate, and compressed to the range of (-1,1) by the tanh function.

[0100] Current cell state and the hidden state at the current moment Simultaneously, it is passed to the next time step t+1 for use with the sliding window of the next time step. We train a multilayer long short-term memory network model together to update the cell state and hidden layer state of the multilayer long short-term memory network model at the next time step.

[0101] Following S12, S13 is also included: S13, Determine the wind field characteristic prediction model: Input the sliding window vectors of the second validation set into the second model in chronological order, and calculate the prediction accuracy of the second model for the "next key wind field vector". If the accuracy is above C2, then the second model at this time is recorded as the wind field feature prediction model and can be used in subsequent applications. If the accuracy is less than C2, optimize the model parameters along the direction of decreasing gradient of the cross-entropy loss function (such as optimizing the forget gate weights of the multilayer long short-term memory network model) and return to S12. Retrain and test the multilayer long short-term memory network model using the second training set and the second validation set until the accuracy of the second model is above C2.

[0102] In this embodiment, C2=90%.

[0103] In S13, "optimizing model parameters along the direction of decreasing gradient of the cross-entropy loss function" also includes the following: using a second validation set. The objective function for testing the generalization of the optimal hyperparameters within the second model on the second validation set is the average error loss of the model on the second validation set: ; in, The sample size of the second validation set; Represents the 2-norm; Indicates the second verification set The actual key wind field vector of the q-th sample at the next time step. The second model indicates that it is based on the samples. The predicted key wind field vector for the next moment.

[0104] Optionally, in formal use, the wind field feature prediction model first predicts several key wind field vectors at the current vertical wind layer for the next moment, calculates the probability of each predicted key wind field vector, and outputs only key wind field vectors with a probability of C3 or higher. If the wind field feature prediction model is used to predict key wind field vectors at the current location for a future period of time (referring to the period starting from the next moment after the current moment), then at least two key wind field vectors will be output: the key wind field vector for the next moment (i.e., the first key wind field vector in the future) and the key wind field vector for the moment after that (i.e., the second key wind field vector in the future).

[0105] In this embodiment, C3 = 80%.

[0106] During training, we ensure that the accuracy of the next key wind field vector predicted by the wind field feature prediction model based on the sliding window vector is above C2. However, the accuracy of other key wind field vectors corresponding to the same sliding window vector in the future will gradually decrease (for example, for the same sliding window vector, the accuracy of the predicted second and third key wind field vectors will become increasingly lower). However, when the wind field feature prediction model is officially put into use, it first predicts several consecutive key wind field vectors in the future based on a certain sliding window vector, then calculates the probability of these key wind field vectors, and finally outputs all key wind field vectors with a probability of C3 or higher (including at least two key wind field vectors). In other words, the number of key wind field vectors predicted by the wind field feature prediction model each time, as well as the time interval between adjacent key wind field vectors, are dynamically changing (for example, if the probability of the second key wind field vector in the future is less than C3, the wind field feature prediction model only outputs the first and third key wind field vectors in the future, which results in a long time interval between two adjacent key wind field vectors in the prediction results).

[0107] Optionally, after obtaining the true key wind field vector at the next prediction time, the wind field feature prediction model calculates the loss function based on the predicted key wind field vector at the next time and the true key wind field vector at the next time, and iteratively optimizes the wind field feature prediction model based on the direction of the decrease in the gradient of the loss function.

[0108] Optimized wind field feature prediction model parameters, including weights for spatial features and time series.

[0109] After training, the multilayer long short-term memory network model, namely the shared weight long short-term memory network (SWLSTM) in the wind field feature prediction model, combined with the attention mechanism, can learn the evolution rules between key wind field vectors in the sliding window vector (i.e., within a certain period of time), and then obtain the key wind field vectors in the future period of time to form the future time-wind field matrix of the corresponding trajectory point. Moreover, the first key wind field vector predicted in the future has extremely high accuracy.

[0110] S2 also includes the following sub-steps: S21, In the process of predicting the landing trajectory point, the airdrop starting point prediction model predicts the trajectory data of the next trajectory point based on the trajectory data of the previous predicted trajectory point and the corresponding future time-wind field matrix, where the timestamp of the airdropped cargo arriving at the next trajectory point is ζ.

[0111] S22. After collecting wind field observation data for the next predicted trajectory point between timestamp ξ and timestamp ζ, the data is sent to the wind field feature extraction model. The wind field feature extraction model outputs the key wind field vector corresponding to the wind field observation data to the wind field feature prediction model. The time interval between timestamp ξ and timestamp ζ is preset by the technicians.

[0112] S23, the wind field feature prediction model uses several key wind field vectors (including the key wind field vectors of timestamps ξ and ζ) between timestamp ξ and timestamp ζ as a time sliding window to predict the key wind field vectors corresponding to the next predicted trajectory point in the future to form the corresponding future time-wind field matrix, which is the output of the entire spatiotemporal wind prediction model.

[0113] Optionally, output data concatenation: The output of the entire spatiotemporal wind prediction model can be a concatenation of all key wind field vectors in the current time window (i.e., "the 10 key wind field vectors in the current window") and key wind field vectors for a corresponding future period, which facilitates the simultaneous acquisition of historical, current, and future wind feature data in subsequent scenarios (such as airdrop trajectory calculation).

[0114] Optional, wind field prediction data decoding: The spatiotemporal wind prediction model also includes a "decoding module" (using a "fully connected layer + deconvolution layer" structure) connected after the wind field feature prediction model. The decoding module is used to decode the current key wind field vector and the key wind field vector to be predicted in the future into wind field prediction data.

[0115] The wind field prediction data includes the predicted wind data for all wind layers. The wind prediction data for different wind layers includes, but is not limited to, information such as radial wind speed, vertical wind speed, wind direction, and the height range of the wind layer.

[0116] Lightweight characteristics: The number of wind layers in the wind field prediction data is less than that in the corresponding original wind field observation data (e.g., 300 layers in the original data → 100 layers in the prediction data); however, the height of each wind layer in the wind field prediction data is higher than the height of a single wind layer in the wind field observation data, and the height coverage remains unchanged (e.g., still 0-3000 meters); thus improving transmission and calculation efficiency.

[0117] It's important to clarify that the future time-wind field matrix of a trajectory point is predicted only after the initial prediction. This is done so that the next trajectory point can be predicted using the airdrop initiation point prediction model based on the trajectory data and the future time-wind field matrix. Therefore, S2 and its sub-steps do not predict all trajectory points from the moment the cargo is dropped from a specific cargo release point until landing all at once. Instead, the output of the spatiotemporal wind prediction model for the previous trajectory point is needed to obtain the next trajectory point using the airdrop initiation point prediction model. Similarly, because all trajectory points during landing are not predicted all at once, the future time-wind field matrices for each trajectory point cannot be obtained all at once either. The prediction of all trajectory points during landing in this application is always coupled with the output of the airdrop initiation point prediction model and the output of the spatiotemporal wind prediction model. This is a major innovation of this application.

[0118] This application uses a spatiotemporal wind prediction model to predict the future time-wind field matrix of each trajectory point during the descent of airdropped cargo from the cargo drop point: ① First, a CNN convolutional neural network model is trained into a wind field feature extraction model. Then, the wind field feature extraction model is used to obtain the training data (sliding window vectors) needed to train a multilayer long short-term memory network model. Finally, the sliding window vectors are used to train the multilayer long short-term memory network model into a wind field feature prediction model. The entire training process is compact and interconnected, significantly reducing the time and manpower costs for technicians to manually construct training and validation sets, and significantly improving the training efficiency of the wind field feature prediction model.

[0119] ② Achieve accurate prediction in both time and space dimensions: In the technical solution of this application, the wind field prediction data output for a future period of time has the validity of both time and space dimensions: In the time dimension, only key wind field vectors with a probability of C3 or higher are output to ensure prediction accuracy; In the spatial dimension, through the fusion and dimensionality reduction of the wind field feature extraction model, the core features of different wind layers are retained, so that the wind field prediction data includes the features of the wind field in different spaces, that is, to achieve wind feature prediction in both time and space dimensions.

[0120] This application enables timely and accurate prediction of wind field data over a future period at different vertical wind layers. The resulting future time-wind field matrix for each trajectory point balances lightweight design and accuracy. To ensure prediction accuracy, this application requires the input wind field observation data to include dense wind layers (the original data is accurate and information-rich). Simultaneously, through fusion and dimensionality reduction using a wind field feature extraction model, the number of wind layers in key wind field vectors is reduced (e.g., from 300 dimensions to 100 dimensions) without losing core information, thus achieving lightweight prediction data. ①At the transmission level: Lightweight data can avoid problems such as transmission interruption and excessive time consumption, ensuring timeliness and integrity; ② At the computational level: Subsequent calculations based on lightweight data (such as airdrop trajectory simulation) have significantly reduced overhead and time consumption. Moreover, because the core features are not lost, the calculation results still have a high accuracy rate and can directly support parameter decisions for air cargo transportation.

[0121] like Figure 2 The figure shows the spatial trajectory curves of the same airdropped cargo at the same exit point (1200m altitude) under different transport aircraft flight speeds.

[0122] like Figure 3 The figure shows the spatial trajectory curves of the same airdropped cargo at different altitude exit points under the same transport aircraft flight speed (150km / h).

[0123] Taking the training process in step 31 as an example, various features of a training dataset (such as historical airdrop cargo quality, weather, umbrella type, drop attitude, and airdrop trajectory data and wind field spatial feature vector of the historical airdrop cargo exit point) are one-hot encoded to generate computer-recognizable 0-1 codes, which then constitute the feature matrix of the training dataset. Similarly, the features of the verification data are also fed into the third model after forming the feature matrix. The input to the airdrop origin point prediction model is also the feature matrix.

[0124] This approach employs a random forest regression model that combines attention and self-attention mechanisms. Self-attention captures global dependencies in the input data, making it suitable for regression tasks involving high-dimensional, temporal, or spatial data. The regression model is typically based on an encoder structure, with the output mapped to continuous values ​​through fully connected layers. These layers include: an input embedding layer, a positional encoding layer, a multi-head self-attention layer, a feedforward neural network, and a regression output layer. For example... Figure 4 As shown.

[0125] The airdrop starting point determination method of this application predicts the landing point corresponding to different aerial locations. Based on the relationship between the predicted landing point and the preset landing range, it inversely deduces which aerial locations can serve as airdrop starting points. An airdrop starting point prediction model is trained on a machine learning model using historical airdrop cargo data and the corresponding future time-wind field matrix. This model, based on airdrop cargo quality, weather, parachute type, and deployment attitude, can predict the trajectory data of the next trajectory point based on the trajectory data of the previous trajectory point and the corresponding future time-wind field matrix, thus predicting the landing point of the current airdrop cargo. Landing points within the preset landing range are recorded as potential landing points (there may be multiple potential landing points). The potential landing point that is closest to the preset landing point in a straight line is recorded as the optimal landing point. The airdrop starting point corresponding to the potential or optimal landing point is then output to the transport aircraft pilot.

[0126] The airdrop starting point determination method of this application efficiently and accurately determines the airdrop starting point: During the training and prediction process, it considers not only the trajectory data of the aerial position point, the quality of the dropped goods, weather, umbrella type, and drop posture, but also the influence of the wind field spatial feature vector that changes over time during the airdrop process (i.e., the importance of the time-wind field matrix); therefore, this application also predicts the future time-wind field matrix during the prediction of the landing point, which is not found in the prior art, further improving the accuracy of the airdrop starting point prediction model in predicting the trajectory data of the next trajectory point; since each trajectory point is the basis for predicting the next trajectory point, this application improves the accuracy of the landing point prediction only by ensuring the accurate prediction of each trajectory point; thus, it ensures that the output airdrop starting point meets the requirements.

[0127] The airdrop starting point determination method of this application can use real data obtained in actual application to optimize the airdrop starting point prediction model, thereby improving the model optimization efficiency; and make up for the negative impact on the prediction accuracy of the airdrop starting point prediction model when the training set and validation set data are insufficient.

[0128] The method for determining the airdrop starting point in this application does not rely on meteorological sensors and kinematic models. It constructs an airdrop starting point prediction model solely based on historical airdrop cargo data and the corresponding time-wind field matrix, predicting the landing point corresponding to airdrops at multiple spatial locations. Furthermore, it uses actual data obtained from airdrops based on the prediction results to further optimize the prediction model, improving the airdrop accuracy of subsequent transport aircraft. This significantly reduces the complexity and cost of the prediction system.

[0129] This application also provides an application of the airdrop starting point determination method: the transport aircraft driver performs a non-powered airdrop when the transport aircraft reaches the backup airdrop starting point (when the airdrop starting point prediction model finally outputs the backup airdrop starting point) or the optimal airdrop starting point (when the airdrop starting point prediction model finally outputs the optimal airdrop starting point).

[0130] The alternative or optimal airdrop starting point is obtained by using one of the airdrop starting point determination methods described above.

[0131] This application can significantly increase the probability that unpowered airdropped cargo will land within a predetermined landing area.

[0132] By utilizing real data obtained during practical applications, the airdrop start point prediction model is optimized through feedback, thereby improving the model optimization efficiency. This also compensates for the negative impact on the prediction accuracy of the airdrop start point prediction model when the initial training and validation sets are insufficient.

[0133] Specifically, in the practical application of the airdrop origin prediction model, the actual landing point of the airdropped cargo is obtained to update the relevant model parameters of the airdrop origin prediction model, thereby further improving the prediction accuracy.

[0134] An incremental learning method is used to dynamically update the relevant model parameters of the airdrop start point prediction model. Combined with a sliding window data fusion strategy (retaining the most recent data and using exponential decay weights), the influence of new and old data is balanced to achieve continuous optimization of model parameters.

[0135] This application also provides an airdrop starting point determination system, including: a historical cargo time-wind field matrix construction module, a dataset construction module, an airdrop starting point prediction model construction module, a potential landing point prediction module, and a backup airdrop starting point acquisition module.

[0136] The historical cargo time-wind field matrix construction module is used to collect historical airdrop cargo data and obtain the time-wind field matrix of historical airdrop cargo from the time it leaves the cabin to the time it lands before sending it into the dataset construction module. The dataset construction module is used to standardize historical airdrop cargo data and, based on the time-wind field matrix, construct a third training set and a third validation set before sending them into the airdrop starting point prediction model construction module. The airdrop starting point prediction model building module is used to train the machine learning model using the third training set, and then use the third validation set for validation and iterative optimization to obtain the airdrop starting point prediction model, which is then sent to the potential landing point prediction module. The potential landing point prediction module is used to use the airdrop start point prediction model to predict the potential landing points of airdropped cargo at each airdrop start point and send them to the backup airdrop start point acquisition module. The backup airdrop starting point acquisition module is used to record potential landing points within a preset landing range as backup landing points, record the airdrop starting points corresponding to the backup landing points as backup airdrop starting points, and output the backup airdrop starting points.

[0137] Each module is programmed or configured to perform the steps of an airdrop starting point determination method as described above.

[0138] This application also provides a computer-readable storage medium storing a computer program programmed or configured to perform an airdrop starting point determination method as described above.

[0139] This application also provides a computer program product, including a computer program / instructions, which are executed by a processor to implement the steps of the airdrop starting point determination method described above.

[0140] The technologies, shapes, and structures not described in detail in this application are all well-known technologies. It should also be noted that the above are merely preferred embodiments of this application and are not intended to limit the scope of this application. The components or steps in the embodiments of this application can be decomposed and / or recombined, and these decompositions and / or recombinations should be considered as equivalent solutions of this application and should all fall within the protection scope of this application.

Claims

1. A method for determining the starting point of an airdrop, characterized in that, Includes the following steps: Step 1: Collect historical airdrop cargo data; simultaneously obtain the time-wind field matrix of historical airdrop cargo from cargo leaving the cargo hold to landing. Step 2: After standardizing the historical airdrop cargo data, construct the training set and validation set based on the time-wind field matrix; Step 3: After training the machine learning model using the training set, validate and iteratively optimize it using the validation set to obtain the airdrop starting point prediction model; Step 4: Use the airdrop start point prediction model to predict the potential landing point of the airdropped cargo at each airdrop start point. Step 5: Record potential landing points within the preset landing range as backup landing points; record the airdrop starting point corresponding to the backup landing point as the backup airdrop starting point, and output the backup airdrop starting point.

2. The method for determining the airdrop starting point according to claim 1, characterized in that, Replace step 5 with step 5': In step 5', the potential landing point that is within the preset landing range and is closest to the preset landing point in a straight line is recorded as the best landing point; the airdrop starting point corresponding to the best landing point is recorded as the best airdrop starting point, and the best airdrop starting point is output.

3. The method for determining the airdrop starting point as described in claim 1, characterized in that: In step 1, the time-wind field matrix corresponding to the trajectory point of a historical airdropped cargo within the sampling period is obtained through a wind field feature extraction model, including the following steps: Step 11: After pre-training the convolutional neural network model using historical wind field observation data, a wind field feature extraction model is obtained. Step 12: Input the historical wind field observation data of a certain trajectory point during the sampling period into the wind field feature extraction model. The wind field feature extraction model outputs several key wind field vectors within the sampling period in chronological order to form the time-wind field matrix corresponding to the trajectory point during the sampling period.

4. The method for determining the airdrop starting point as described in claim 3, characterized in that, Step 11 includes the following sub-steps: Step 111: After obtaining historical wind field observation data of several adjacent vertical wind layers within a continuous historical sampling period, the historical wind field observation data is preprocessed. Step 112: Divide the preprocessed historical wind field observation data into a first dataset and a second dataset; Step 113: After training the CNN convolutional neural network model using the first dataset, the first model is obtained; Step 114: After validating the first model using the second dataset, the wind field feature extraction model is obtained.

5. The method for determining the airdrop starting point as described in claim 4, characterized in that: Step 113 includes the following sub-steps: Step 1131: After labeling each historical wind field observation data in the first dataset with key wind field features, the key wind field vector corresponding to each historical wind field observation data is obtained; the key wind field vector serves as the label of the corresponding historical wind field observation data, and together with the corresponding historical wind field observation data, constitutes the sample in the first training set. Step 1132: Train the CNN convolutional neural network model using the first training set, so that the CNN convolutional neural network model has the function of converting wind field observation data into key wind field vectors after training; denote the trained CNN convolutional neural network model as the first model; Step 114 also includes the following: inputting wind field observation data from the second dataset into the first model and testing the accuracy of the first model in converting the wind field observation data into key wind field vectors; Then verify the accuracy of the key wind field vectors output by the first model; if the accuracy is above C1, then the first model at this time is recorded as the wind field feature extraction model. If the accuracy is less than C1, optimize the parameters of the first model along the direction of decreasing gradient of the mean squared error loss function, and then return to S132 to retrain and validate the CNN convolutional neural network model using the first training set and the second dataset until the accuracy of the first model is above C1.

6. The method for determining the airdrop starting point as described in claim 5, characterized in that: In step 2: the airdrop trajectory data of historical airdropped cargo is processed by Min-Max standardization to eliminate the dimensional differences between different airdrop trajectory data, and then the third training set and the third validation set are constructed. Step 3 also includes the following sub-steps: Step 31: Use the third training set to train the machine learning model so that after training, the machine learning model can predict the airdrop trajectory data of the next trajectory point of the airdrop cargo based on the airdrop cargo quality, weather, umbrella type, deployment attitude, airdrop trajectory data of the known trajectory points of the airdrop cargo, and the time-wind field matrix corresponding to the trajectory points; the trained machine learning model is denoted as the third model. Step 32: Use the third validation set to validate the third model, and compare the validation data labels output by the third model with the real labels of the corresponding validation data. If the accuracy is ≥ A1, then the third model at this time is recorded as the airdrop starting point prediction model and can be officially put into use; otherwise, after optimizing the model parameters along the direction of decreasing gradient of the mean square error loss function, return to step 31 and retrain and test the third model using the training set and validation set until the accuracy of the third model reaches the target. Step 4 also includes the following sub-steps: Step 41: Obtain the longitude, latitude, and altitude of each air position point within the range of the transport aircraft's forward movement, and calculate the time corresponding to the current arrival of the transport aircraft at each air position point; use these times as the timestamps of the corresponding air position points. Step 42: Each aerial location is designated as the cargo drop point. Based on the cargo drop point, cargo mass, weather, parachute type, drop attitude, and the future time-wind field matrix during the descent of cargo from different drop points, the airdrop origin prediction model is used to predict the trajectory data of the cargo falling from different drop points. The landing points corresponding to the predicted cargo drop points are recorded as potential landing points.

7. The method for determining the airdrop starting point as described in claim 6, characterized in that: In step 42, the future time-wind field matrix of each trajectory point during the descent of the airdropped cargo is predicted using a spatiotemporal wind prediction model, including the following steps: S1. After pre-training the wind field feature prediction model using a multi-layer long short-term memory network model based on historical wind field observation data, a wind field feature prediction model is obtained. The wind field feature extraction model and the wind field feature prediction model are connected in series to form a spatiotemporal wind prediction model. S2 collects wind field observation data of the predicted trajectory points in real time and sends it to the spatiotemporal wind prediction model. The spatiotemporal wind prediction model outputs the future time-wind field matrix of the corresponding trajectory points.

8. The method for determining the airdrop starting point according to claim 7, characterized in that, S1 also includes the following sub-steps: S11, using a wind field feature extraction model, the historical wind field observation data in the first dataset is transformed into corresponding key wind field vectors; then, a time sliding window is used to package the key wind field vectors at the same location point into sliding window vectors in chronological order, which are then recorded as the second training set; each sliding window vector contains k key wind field vectors at the same location point arranged in chronological order, and the sliding window slides in the direction of increasing time, with each step of the sliding window being m key wind field vectors; simultaneously, the wind field feature extraction model is used to transform the wind field observation data in the second dataset into a second validation set containing several consecutive sliding window vectors; k is a positive integer; S12, Use the second training set to train the multilayer long short-term memory network model, so that the multilayer long short-term memory network model can predict the future key wind field vector at a certain location point based on the sliding window vector after training; denote the trained multilayer long short-term memory network model as the second model; S13, Determine the wind field feature prediction model: Input the sliding window vector of the second validation set into the second model in chronological order, and calculate the prediction accuracy of the second model for the next key wind field vector; if the accuracy is above C2, then the second model at this time is recorded as the wind field feature prediction model; if the accuracy is less than C2, then optimize the model parameters along the gradient decreasing direction of the cross-entropy loss function and return to S12, and retrain and test the multilayer long short-term memory network model using the second training set and the second validation set until the accuracy of the second model is above C2.

9. The method for determining the airdrop starting point as described in claim 8, characterized in that, In S12, the multilayer long short-term memory network model is trained using the second training set, including the following: Let the key wind field vector at a certain location at the current moment be . The time sliding window formed by the current moment and the k key wind field vectors before that moment... ,in Let J represent the j-th key wind field vector in sliding window Y, where 1 ≤ j ≤ k and j and k are both positive integers; in the key wind field vector There is ,in Represents the key wind field vector Data in the i-th dimension; Let the current time be t, the previous time be t-1, and the next time be t+1. The sliding window corresponding to a certain position at the current time is: Then the current time window will slide. As the input to the multilayer long short-term memory network model at the current time step; based on the following formula, the multilayer long short-term memory network model is iteratively trained using the second training set, continuously updating the cell states and hidden layer states in the multilayer long short-term memory network model: ； In the multilayer long short-term memory network model, the cell state at the previous time step is: The hidden state at the previous time step was ; This indicates the output of the forget gate at the current moment; (·) represents the sigmoid activation function; The weight matrix representing the forget gate; Represents the hidden state at the previous time step. Slide window at the current moment splicing; This indicates that the bias term of the door has been forgotten; This represents the input gate output at the current time t; Represents the candidate cell state at time t; tanh(·) represents the hyperbolic tangent activation function; A weight matrix representing the state of candidate cells; Bias terms representing the state of candidate cells; This indicates the output of the gate at the current time t; This represents the weight matrix of the output gate; This represents the bias term of the output gate; Hidden state at the current moment As the output of the multilayer long short-term memory network model at the current time; the cell state at the current time t within the multilayer long short-term memory network model is: .

10. An application of a method for determining the starting point of an airdrop, characterized in that: The transport aircraft pilot performs a non-powered airdrop when the transport aircraft reaches the backup airdrop starting point or the optimal airdrop starting point; the backup airdrop starting point or the optimal airdrop starting point is obtained by any one of the airdrop starting point determination methods described in claims 1-9.

Images