A trajectory prediction method for a downward-throwing type detection assembly based on an LSTM neural network

Abstract

The application discloses a trajectory prediction method of a lower-throwing type detection assembly based on an LSTM neural network, and belongs to the technical field of lower-throwing type detection assemblies and high-altitude detection. The application is based on the characteristics of LSTM neural network processing time series, adopts a data-based method, learns historical trajectory data of the lower-throwing type detection assembly, combines the influence of environmental factors on the trajectory of the lower-throwing type detection assembly, and combines three environmental parameters of temperature, humidity and air pressure into the trajectory data of the lower-throwing type detection assembly; a high-precision trajectory prediction model of the lower-throwing type detection assembly is established, parameters collected by a sensor are preprocessed, compensation is performed on temperature and air pressure data by using a corresponding compensation algorithm, and the reliability and authenticity of data input into the LSTM neural network are ensured; and a recursive prediction method is adopted to predict the future motion trajectory of the lower-throwing type detection assembly. The application can assist in selecting appropriate lower-throwing time and a lower-throwing point, and improve the efficiency and precision of a detection task of the lower-throwing type detection assembly.

Description

Technical Field

[0001] This invention belongs to the field of drop-down detection components and high-altitude detection technology, and particularly relates to a trajectory prediction method for drop-down detection components based on LSTM neural networks. Background Technology

[0002] In recent years, drop-in detectors have attracted widespread attention and continued to develop both domestically and internationally. They are primarily used for in-situ detection in aerospace scenarios, serving as payloads deployed into target airspace to perform high-precision detection and real-time transmission of environmental parameters. Drop-in detectors play a crucial role in many scenarios. For example, typhoons are among the world's most severe natural disasters, posing a serious threat to people's lives and property. Measuring changes in environmental parameters through meteorological monitoring instruments and modeling the typhoon's path and trends in advance can help people effectively issue warnings, take preventative measures, and significantly reduce losses. Currently, most detection methods can only obtain data from the outer periphery of typhoons, with limited detection depth, breadth, and precision, restricting accurate predictions of typhoon intensity and development trends. By deploying drop-in detectors to different areas of a typhoon, refined in-situ detection can be carried out in each area. The detection data can then be used to achieve high-precision modeling of the typhoon, effectively mitigating the losses caused by typhoon disasters.

[0003] Drop-down probes have also made significant contributions to near-space environment detection. Near space refers to the airspace 20-100 km above the ground, a transitional zone between aerospace and space travel, and its important strategic position has attracted widespread international attention. Connecting aerospace and space, near space is becoming a focus of attention for major military powers, who are vying to develop near-space vehicles capable of flying within it. Near-space vehicles will introduce new combat spaces into future warfare, profoundly impacting the mechanisms and models of modern warfare. They can also serve as research platforms for meteorological observation, biological experiments, atmospheric sampling, special experiments, and high-altitude research. However, the aerodynamic shape, trajectory design, and landing accuracy of these vehicles heavily depend on the near-space atmospheric environment. Therefore, research on high-precision detection methods for the near-space atmospheric environment is extremely important. By deploying drop-down probes, it is possible to better detect the near-space environment with high precision, laying the foundation for gaining the initiative in near-space.

[0004] Therefore, predicting the trajectory of drop-in detection components allows for the selection of appropriate drop timing, ensuring that the components land in the corresponding airspace for in-situ detection and guaranteeing the smooth progress of the detection mission. It can also, in turn, improve the performance and aerodynamic parameters of parachutes. Furthermore, it provides insights for predicting the trajectories of many unpowered flight missions, such as ensuring that paratroopers land in designated areas, which is of great significance.

[0005] Currently, flight target trajectory prediction models can be categorized into dynamic models, Markov models, and machine learning models. Due to the complex and variable aerospace environment, it is difficult to establish accurate dynamic models for drop-in probes, resulting in low trajectory prediction accuracy. Markov models, when performing high-order calculations, have high complexity and do not meet the requirements for real-time trajectory prediction of drop-in probes. Machine learning models have been widely used in trajectory prediction and have achieved good results. Since aircraft trajectory prediction is essentially a prediction of time series data, using data-driven methods to transform the trajectory prediction problem into predicting time series data can avoid the high uncertainty of aircraft trajectories caused by the complexity of the aerospace environment. Currently, trajectory prediction for drop-in probes is still at the stage of establishing dynamic models, with a limited amount of fitted data and no consideration of the influence of environmental parameters, making it difficult to accurately predict the drop trajectory of drop-in probes. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, the main objective of this invention is to provide a trajectory prediction method for drop-in probe components based on an LSTM neural network. For complex aerospace environments, the method trains the LSTM neural network using trajectory data from the drop-in probe component. Based on the LSTM neural network, a data-driven approach is used to accurately predict the trajectory of the drop-in probe component, selecting an appropriate drop-in timing to ensure the component lands in the corresponding airspace for in-situ detection, thereby improving the accuracy and efficiency of the detection mission.

[0007] To achieve the above objectives, the present invention adopts the following solution:

[0008] This invention discloses a trajectory prediction method for a drop-type probe component based on an LSTM neural network, comprising the following steps:

[0009] S1: Obtain a segment of trajectory data from historical missions in which drop-in detection components equipped with multiple sensors are deployed on an airborne platform.

[0010] The drop-down detection component is deployed at a designated altitude by an airborne platform, with a parachute opening simultaneously. The trajectory data of the drop-down detection component after deployment is acquired via sensors. The drop-down detection component is equipped with a GPS / BeiDou positioning system, an inertial navigation system, a barometric pressure sensor, a temperature sensor, a humidity sensor, and a communication system.

[0011] The air-based platform includes airships and rockets.

[0012] The communication system has a sampling rate of 1Hz and sends the collected signal to the ground receiving system every 1 second.

[0013] The data used for trajectory prediction includes 16 dimensions: longitude, latitude, altitude, three-dimensional velocity, three-dimensional acceleration, two-dimensional angular velocity, two-dimensional angular acceleration, temperature, humidity, and air pressure. Due to the small size and light weight of the drop-down probe, aerodynamic forces play a significant role in its movement. Environmental factors have a crucial influence on aerodynamic forces; therefore, in addition to the position, attitude, and velocity of the drop-down probe, the impact of environmental parameters on trajectory prediction must also be characterized. These environmental parameters include temperature, humidity, and air pressure. A segment of trajectory data is selected from the collected data stored in the ground receiving system and used as a training sample for the neural network.

[0014] S2: Preprocess the collected trajectory data, fill in missing values ​​and smooth the data, while removing invalid and outlier values, and normalize the data range to the [0,1] interval.

[0015] Because the detection airspace environment is quite complex, the data acquired by the sensors may contain noise or be incomplete, and cannot be directly input into the neural network. Preprocessing of the data is required.

[0016] For three-axis angles, data within the range of (-180°, 180°) are retained. For eastward and northward velocities measured by the GPS / BeiDou positioning system, the velocity change Δv between two adjacent data points is calculated. If Δv is significantly greater than a set threshold, the current velocity data is discarded. For the outputs of the barometer and temperature / humidity sensor, the data filtering method is the same as that for velocity data, i.e., outliers exceeding the threshold are removed. After removing outliers, maximum likelihood estimation is used to impute missing values. First, the missing values ​​are replaced with estimated values; then, parameter estimation is performed on the complete data; finally, the missing values ​​are estimated again based on the parameter estimates. The likelihood function is as follows:

[0017]

[0018] Where, D = {x1, x2, ..., x} N Let {x1, x2, ..., x} be the sample set, and the joint probability density function p(D|θ) is called the probability density function relative to {x1, x2, ..., x}. N Let l(θ) be the likelihood function of θ. Find the value of θ that maximizes the likelihood function. This value is the maximum likelihood estimate and can be used to replace the missing value.

[0019] Due to the influence of environmental convection, radiation, conduction, and the electronic components used in the measurement, compensation is required for the temperature and pressure parameters measured by the sensor. The temperature parameter errors acquired by the down-projection detection component include Joule thermal error, aerodynamic thermal error, and irradiation thermal error. The correction method for the temperature measurement data is as follows:

[0020]

[0021] Where T is the actual temperature, T a For actual measured temperature, The Joule thermal correction error is given, where U is the voltage across the platinum thin film, t is the time of action, A1 is the cross-sectional area of ​​the object perpendicular to the thermal conduction direction, and R... T λ is the resistance value of the platinum thin film, and λ is the thermal conductivity. The irradiation correction error is represented by ε, where ε is the emissivity of the object, ranging from 0 to 1, A2 is the irradiated surface area, and σ is the Stefan-Boltzmann constant, which is 5.67 × 10⁻⁶. -8 W / (m 2 ·K 4 ), T r The blackbody thermodynamic temperature; For aerodynamic thermal correction error, r is the temperature correction coefficient, V is the relative air velocity, and c P This is the isobaric specific heat capacity of air.

[0022] When the detection component encounters air obstruction during its descent, pressure will be generated on its surface. Therefore, the pressure measured by the barometric pressure sensor includes not only static atmospheric pressure but also the pressure generated by air obstruction during its movement (called wind pressure). The relationship between the pressure magnitude and velocity is expressed as follows:

[0023] w p =0.5*ro*v 2

[0024] Where w p Let w represent wind pressure, ro represent air density, and v represent the falling velocity of the detection component. Therefore, during the falling process of the detection component, the measured value of the air pressure sensor is w. d =w s +w p , where w s The static air pressure value is the actual value that needs to be measured. However, due to the superposition of wind pressure, the air pressure sensor reading is larger than the actual value. Therefore, it is necessary to subtract the wind pressure to obtain the actual air pressure value.

[0025] Because there are many input parameters, each with a different value range, zero-mean normalization is used to prevent some parameters from becoming too large and causing unstable calculation results. This normalization ensures that the data values ​​fall within the [0,1] interval, allowing the network to converge quickly and making it easier to train. The zero-mean normalization transformation function is as follows:

[0026]

[0027] In the formula, x is the actual value of a certain data, μ is the mean of the original data, δ is the standard deviation, and x′ is the normalized value.

[0028] S3: The preprocessed data is segmented using the sliding window method to construct training and testing samples. The trajectory data of the drop-down probe component has continuous and temporal characteristics. The hidden layers of the LSTM neural network have feedback connections, ensuring that the output at any given moment is not only related to the input at that moment but also incorporates information from previous moments, effectively enabling its application in trajectory prediction for the drop-down probe component. The dataset is segmented using the sliding window method and sequentially used as input to the neural network. The number of data points in the dataset is N = {x0, x1, ..., x...}. N-1 The first M points are designated as training samples, and the last K points are designated as test samples. The window length is m, meaning it passes through the data {x0, x1, ..., xn} from time 0 to m-1. m-1 Predict the trajectory information of the drop-in probe component at time m, using data from time 1 to time m {x1, x2, ..., x...} m The trajectory information at time m+1 is predicted, and so on, resulting in a total of M-m+1 points as training sample inputs and K-m+1 points as test sample inputs. Therefore, the dimension of the input data for LSTM neural network training is (M-m+1, m, 16), where M is the number of training sample points, m is the window length, and 16 represents the 16-dimensional data information contained in each point, as follows:

[0029] Longitude, latitude, altitude, three-dimensional velocity, three-dimensional acceleration, two-dimensional angular velocity, two-dimensional angular acceleration, temperature, humidity, air pressure.

[0030] S4: Establish a trajectory prediction model for the drop-down probe component based on LSTM. Minimize the error between the output and the true value by iteratively adjusting the network parameters to determine the network structure. The LSTM network incorporates three gate controllers, drawing inspiration from human memory and forgetting patterns based on information importance. After data is input into the neural network, it passes through the forget gate, update gate, and output gate to finally obtain the output.

[0031] The formula for data processing after inputting into the LSTM neural network is as follows:

[0032] f t =σ(W hf h t-1 +W xf x t +b f )

[0033] i t =σ(W hi h t-1 +W xi x t +b i )

[0034] o t=σ(W ho h t-1 +W xo x t +b o )

[0035]

[0036]

[0037] h t =o t tanh(c t )

[0038] Where x t Given 16-dimensional data at time t, f t For the Gate of Oblivion, i t To update the door, o t For the output gate, h t For the hidden layer output, c t Let σ represent the state of the LSTM neuron cell, σ be the sigmoid activation function, tanh be the tanh activation function, W be the weight matrix, and b be the bias parameter matrix.

[0039] The forget gate adjusts the degree of information retention f from the previous unit by regulating the weights. t When the weight is 0, the information is completely discarded; when the weight is 1, it is completely retained. The update gate determines how much of the input data at the current moment needs to be retained by setting a threshold, thus obtaining i. t Combined with the temporary state of the cell unit Update to the current cell unit state c t In order to continue the backward propagation; the output gate will take the current input data x. t Compared with the output data h from the previous time step t-1 The weight o for the next time step is obtained by combining the calculations. t Then, combined with the current cell unit state c t Finally, the output h at the current time is obtained. t .

[0040] The training samples constructed in S3 are input into the LSTM neural network, which is then trained on a GPU. The cuDNN library is used to accelerate the convolutional neural network, and the Adam optimization algorithm is employed to update parameters and weights, enabling the network model to converge faster. The root mean square error between the predicted and actual data is chosen as the loss function to optimize the parameters of the LSTM neural network for better prediction results. The root mean square error function is:

[0041]

[0042] Where y i Let y′ be the true value of the parameter at time i. i Let be the predicted value at time i, and n be the number of data points input for one training iteration.

[0043] The LSTM neural network outputs trajectory predictions for M-m+1 points. Each point contains 13-dimensional trajectory data, namely: longitude, latitude, altitude, three-dimensional velocity, three-dimensional acceleration, two-dimensional angular velocity, and two-dimensional angular acceleration.

[0044] S5: The test sample is input into the trained LSTM neural network, and the output is denormalized to obtain the trajectory prediction result of the drop probe component;

[0045] By inputting test samples of length K into a pre-trained trajectory prediction model for a drop-in probe based on an LSTM neural network, and performing zero-mean inverse normalization on the output, the model is compared with the actual trajectory to verify the accuracy of the LSTM neural network's trajectory prediction. The trained trajectory prediction model for the drop-in probe is saved and can be directly used in subsequent trajectory prediction tasks to improve prediction efficiency and ensure real-time trajectory prediction.

[0046] S6: In actual detection, environmental information of the detection airspace is first collected, and then integrated with the trajectory information of the drop-type detection component after deployment. This information is then input into the LSTM neural network to predict the future trajectory of the detection component.

[0047] Before the actual deployment of the drop-in detector component, environmental parameters such as temperature, humidity, and air pressure in the deployment airspace are obtained in advance using a prototype detector component. These 3D environmental parameters are combined with the predicted 13D trajectory information of the drop-in detector component, resulting in a total of 16 parameters. This ensures the input dimensionality of the LSTM neural network is consistent. A recursive prediction method with a sliding window is used: the prediction result from the first step is fed into the neural network, and then the prediction from the second step is performed. This process is repeated to continuously predict the future trajectory. In short:

[0048] Prediction Step 1: X = [X0, X1, ..., X m-1 The prediction result is: Combining 3D environmental parameters Expanded to 16 dimensions;

[0049] Prediction step 2: X = [X1, X2, ..., X m The prediction result is: Combining 3D environmental parameters Expanded to 16 dimensions;

[0050] …

[0051] Predicting step L: X = [X L-1 ,X L ,…,Xm+L-2 The prediction result is: Obtain the complete trajectory prediction curve

[0052] Based on the predicted trajectory, select the appropriate timing and point for dropping the drop-type detection component to ensure that it falls into the corresponding airspace for in-situ detection, thereby improving the efficiency of the detection mission.

[0053] Beneficial effects:

[0054] 1. This invention discloses a trajectory prediction method for a drop-type probe component based on an LSTM neural network. By leveraging the time-series processing characteristics of the LSTM neural network and employing a data-driven approach, the method learns from the historical trajectory data of the drop-type probe component to predict its future trajectory, establishing a high-precision trajectory prediction model for the drop-type probe component. This eliminates the need for a complex aerodynamic model of the drop-type probe component, thus solving the problem that aerodynamic models cannot perform high-precision trajectory prediction. This method can effectively assist in selecting appropriate drop timing and drop points, improving the efficiency and accuracy of the drop-type probe component's detection mission.

[0055] 2. The present invention discloses a trajectory prediction method for a drop-down probe component based on an LSTM neural network. By characterizing the influence of environmental factors on the trajectory of the drop-down probe component, the method combines three environmental parameters—temperature, humidity, and air pressure—into the trajectory data of the drop-down probe component for trajectory prediction, thereby improving the accuracy of trajectory prediction for the drop-down probe component.

[0056] 3. The present invention discloses a trajectory prediction method for a drop-type detection component based on an LSTM neural network. The method preprocesses the parameters collected by the sensor and compensates the temperature and air pressure data with corresponding compensation algorithms to ensure the reliability and authenticity of the data input into the LSTM neural network and improve the accuracy of trajectory prediction for the drop-type detection component.

[0057] 4. The present invention discloses a trajectory prediction method for a drop-type detection component based on an LSTM neural network. It adopts a recursive prediction method to continuously predict future trajectories, thereby improving the efficiency of trajectory prediction for drop-type detection components and assisting in the effective selection of appropriate drop timing and drop points. Attached Figure Description

[0058] The following is a further explanation with reference to the accompanying drawings:

[0059] Figure 1 This is a flowchart of the trajectory prediction process for the drop-type detection component based on an LSTM neural network according to an embodiment of the present invention.

[0060] Figure 2 This is a schematic diagram of the coordinate system of the downward-projecting detection component according to an embodiment of the present invention;

[0061] Figure 3 This is a schematic diagram of the LSTM neural network prediction model according to an embodiment of the present invention. Detailed Implementation

[0062] To better illustrate the purpose and advantages of the present invention, the invention will be further described below in conjunction with the accompanying drawings and examples.

[0063] Example 1:

[0064] like Figure 1 As shown in the figure, this embodiment discloses a trajectory prediction method for a drop-type detection component based on an LSTM neural network. The specific implementation steps are as follows:

[0065] S1: Obtain a segment of trajectory data from historical missions in which drop-in detection components equipped with multiple sensors are deployed on an airborne platform.

[0066] The drop-in detection module is deployed at a designated altitude by an airborne platform, while a parachute opens simultaneously. A schematic diagram of the coordinate system of the drop-in detection module after deployment is shown below. Figure 2 As shown. The trajectory data of the drop-down detection component after being deployed is obtained through sensors. The drop-down detection component is equipped with a GPS / BeiDou positioning system, an inertial navigation system, a barometric pressure sensor, a temperature sensor, a humidity sensor, and a communication system.

[0067] The air-based platform includes airships and rockets.

[0068] The communication system has a sampling rate of 1Hz and sends the collected signal to the ground receiving system every 1 second.

[0069] The data used for trajectory prediction includes 16 dimensions: longitude, latitude, altitude, three-dimensional velocity, three-dimensional acceleration, two-dimensional angular velocity, two-dimensional angular acceleration, temperature, humidity, and air pressure. Due to the small size and light weight of the drop-down probe, aerodynamic forces play a significant role in its movement. Environmental factors have a crucial influence on aerodynamic forces; therefore, in addition to the position, attitude, and velocity of the drop-down probe, the impact of environmental parameters on trajectory prediction must also be characterized. These environmental parameters include temperature, humidity, and air pressure. A segment of trajectory data is selected from the collected data stored in the ground receiving system and used as a training sample for the neural network.

[0070] S2: Preprocess the collected trajectory data, fill in missing values ​​and smooth the data, while removing invalid and outlier values, and normalize the data range to the [0,1] interval.

[0071] Because the detection airspace environment is quite complex, the data acquired by the sensors may contain noise or be incomplete, and cannot be directly input into the neural network. Preprocessing of the data is required.

[0072] For three-axis angles, data within the range of (-180°, 180°) are retained. For eastward and northward velocities measured by the GPS / BeiDou positioning system, the velocity change Δv between two adjacent data points is calculated. If Δv is significantly greater than a set threshold, the current velocity data is discarded. For the outputs of the barometer and temperature / humidity sensor, the data filtering method is the same as that for velocity data, i.e., outliers exceeding the threshold are removed. After removing outliers, maximum likelihood estimation is used to impute missing values. First, the missing values ​​are replaced with estimated values; then, parameter estimation is performed on the complete data; finally, the missing values ​​are estimated again based on the parameter estimates. The likelihood function is as follows:

[0073]

[0074] Where, D = {x1, x2, ..., x} N Let {x1, x2, ..., x} be the sample set, and the joint probability density function p(D|θ) is called the probability density function relative to {x1, x2, ..., x}. N Let l(θ) be the likelihood function of θ. Find the value of θ that maximizes the likelihood function. This value is the maximum likelihood estimate and can be used to replace the missing value.

[0075] Due to the influence of environmental convection, radiation, conduction, and the electronic components used in the measurement, compensation is required for the temperature and pressure parameters measured by the sensor. The temperature parameter errors acquired by the down-projection detection component include Joule thermal error, aerodynamic thermal error, and irradiation thermal error. The correction method for the temperature measurement data is as follows:

[0076]

[0077] Where T is the actual temperature, T a For actual measured temperature, The Joule thermal correction error is given, where U is the voltage across the platinum thin film, t is the time of action, A1 is the cross-sectional area of ​​the object perpendicular to the thermal conduction direction, and R... T λ is the resistance value of the platinum thin film, and λ is the thermal conductivity. The irradiation correction error is represented by ε, where ε is the emissivity of the object, ranging from 0 to 1, A2 is the irradiated surface area, and σ is the Stefan-Boltzmann constant, which is 5.67 × 10⁻⁶. -8 W / (m 2 ·K 4 ), T r The blackbody thermodynamic temperature; For aerodynamic thermal correction error, r is the temperature correction coefficient, V is the relative air velocity, and c P This is the isobaric specific heat capacity of air.

[0078] When the detection component encounters air obstruction during its descent, pressure will be generated on its surface. Therefore, the pressure measured by the barometric pressure sensor includes not only static atmospheric pressure but also the pressure generated by air obstruction during its movement (called wind pressure). The relationship between the pressure magnitude and velocity is expressed as follows:

[0079] w p =0.5*ro*v 2

[0080] Where w p Let w represent wind pressure, ro represent air density, and v represent the falling velocity of the detection component. Therefore, during the falling process of the detection component, the measured value of the air pressure sensor is w. d =w s +w p , where w s The static air pressure value is the actual value that needs to be measured. However, due to the superposition of wind pressure, the air pressure sensor reading is larger than the actual value. Therefore, it is necessary to subtract the wind pressure to obtain the actual air pressure value.

[0081] Because there are many input parameters, each with a different value range, zero-mean normalization is used to prevent some parameters from becoming too large and causing unstable calculation results. This normalization ensures that the data values ​​fall within the [0,1] interval, allowing the network to converge quickly and making it easier to train. The zero-mean normalization transformation function is as follows:

[0082]

[0083] In the formula, x is the actual value of a certain data, μ is the mean of the original data, δ is the standard deviation, and x′ is the normalized value.

[0084] S3: The preprocessed data is segmented using the sliding window method to construct training and testing samples. The trajectory data of the drop-down probe component has continuous and temporal characteristics. The hidden layers of the LSTM neural network have feedback connections, ensuring that the output at any given moment is not only related to the input at that moment but also incorporates information from previous moments, effectively enabling its application in trajectory prediction for the drop-down probe component. The dataset is segmented using the sliding window method and sequentially used as input to the neural network. The number of data points in the dataset is N = {x0, x1, ..., x...}. N-1 The first M points are designated as training samples, and the last K points are designated as test samples. The window length is m, meaning it passes through the data {x0, x1, ..., xn} from time 0 to m-1. m-1 Predict the trajectory information of the drop-in probe component at time m, using data from time 1 to time m {x1, x2, ..., x...} mThe trajectory information at time m+1 is predicted, and so on, resulting in a total of M-m+1 points as training sample inputs and K-m+1 points as test sample inputs. Therefore, the dimension of the input data for LSTM neural network training is (M-m+1, m, 16), where M is the number of training sample points, m is the window length, and 16 represents the 16-dimensional data information contained in each point, as follows:

[0085] Longitude, latitude, altitude, three-dimensional velocity, three-dimensional acceleration, two-dimensional angular velocity, two-dimensional angular acceleration, temperature, humidity, air pressure.

[0086] In this embodiment of the invention, the sliding window method is used to divide the dataset into segments and use them as inputs to the neural network. The dataset has 2000 points and the window length is 25. The first 1800 points are divided into training samples and the last 200 points are divided into test samples. The dimension of the input data during the training of the LSTM neural network is (176, 25, 16).

[0087] S4: Establish an LSTM-based trajectory prediction model for the drop-down probe component. By iteratively adjusting the network parameters, minimize the error between the output and the true value, and determine the network structure. A schematic diagram of the LSTM neural network prediction model in this embodiment is shown below. Figure 3 As shown, the LSTM network borrows from the human tendency to remember and forget information based on its importance, incorporating three "gate" controllers into the network. After data is input into the neural network, it passes through the forget gate, update gate, and output gate, ultimately yielding the output.

[0088] The formula for data processing after inputting into the LSTM neural network is as follows:

[0089] f t =σ(W hf h t-1 +W xf x t +b f )

[0090] i t =σ(W hi h t-1 +W xi x t +b i )

[0091] o t =σ(W ho h t-1 +W xo x t +b o )

[0092]

[0093]

[0094] h t =o t tanh(c t )

[0095] Where x t Given 16-dimensional data at time t, f t For the Gate of Oblivion, i t To update the door, o t For the output gate, h t For the hidden layer output, c t Let σ represent the state of the LSTM neuron cell, σ be the sigmoid activation function, tanh be the tanh activation function, W be the weight matrix, and b be the bias parameter matrix.

[0096] The forget gate adjusts the degree of information retention f from the previous unit by regulating the weights. t When the weight is 0, the information is completely discarded; when the weight is 1, it is completely retained. The update gate determines how much of the input data at the current moment needs to be retained by setting a threshold, thus obtaining i. t Combined with the temporary state of the cell unit Update to the current cell unit state c t In order to continue the backward propagation; the output gate will take the current input data x. t Compared with the output data h from the previous time step t-1 The weight o for the next time step is obtained by combining the calculations. t Then, combined with the current cell unit state c t Finally, the output h at the current time is obtained. t .

[0097] The training samples constructed in S3 are input into the LSTM neural network, which is then trained on a GPU. The cuDNN library is used to accelerate the convolutional neural network, and the Adam optimization algorithm is employed to update parameters and weights, enabling the network model to converge faster. The initial learning rate is set to 0.001, the batch size to 64, the number of LSTM hidden layers to 2, and the number of hidden layer nodes to 12. The root mean square error between the predicted and actual data is chosen as the loss function to optimize the parameters of the LSTM neural network for better prediction performance. The root mean square error function is:

[0098]

[0099] Where y i Let y′ be the true value of the parameter at time i. i Let be the predicted value at time i, and n be the number of data points input for one training iteration, which is 2000 in this embodiment.

[0100] As the number of training iterations increases, the error function value gradually decreases until it approaches 0, indicating that the model training effect is good.

[0101] The LSTM neural network outputs trajectory predictions for 176 points, with each point containing 13-dimensional trajectory data: longitude, latitude, altitude, three-dimensional velocity, three-dimensional acceleration, two-dimensional angular velocity, and two-dimensional angular acceleration.

[0102] S5: The test sample is input into the trained LSTM neural network, and the output is denormalized to obtain the trajectory prediction result of the drop probe component;

[0103] By inputting 200 test samples into a pre-trained LSTM neural network-based trajectory prediction model for a drop-type probe component, and performing zero-mean inverse normalization on the output, the results were compared with the actual trajectories. The absolute errors all approached zero, indicating stable prediction results. This verifies the accuracy of the LSTM neural network trajectory prediction. The trained trajectory prediction model for the drop-type probe component can be saved and directly used in subsequent trajectory prediction tasks, improving prediction efficiency and ensuring real-time trajectory prediction.

[0104] S6: In actual detection, environmental information of the detection airspace is first collected, and then integrated with the trajectory information of the drop-type detection component after deployment. This information is then input into the LSTM neural network to predict the future trajectory of the detection component.

[0105] Before the actual deployment of the drop-in detector component, environmental parameters such as temperature, humidity, and air pressure in the deployment airspace are obtained in advance using a prototype detector component. These 3D environmental parameters are combined with the predicted 13D trajectory information of the drop-in detector component, resulting in a total of 16 parameters. This ensures the input dimensionality of the LSTM neural network is consistent. A recursive prediction method with a sliding window is used: the prediction result from the first step is fed into the neural network, and then the prediction from the second step is performed. This process is repeated to continuously predict the future trajectory. In short:

[0106] Prediction Step 1: X = [X0, X1, ..., X m-1 The prediction result is: Combining 3D environmental parameters Expanded to 16 dimensions;

[0107] Prediction step 2: X = [X1, X2, ..., X m The prediction result is: Combining 3D environmental parameters Expanded to 16 dimensions;

[0108] …

[0109] Predicting step L: X = [X L-1 ,X L,…,X m+L-2 The prediction result is: Obtain the complete trajectory prediction curve

[0110] Based on the predicted trajectory, the appropriate timing and point for dropping the drop-type probe are selected to ensure that the probe lands in the corresponding airspace for in-situ detection, thus improving the efficiency of the detection mission.

[0111] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A trajectory prediction method for a drop-type detection component based on an LSTM neural network, characterized in that: Includes the following steps, S1: Obtain a segment of trajectory data from historical missions in which drop-down detection components equipped with multiple sensors were deployed on an airborne platform; S2: Preprocess the collected trajectory data, fill in missing values ​​and smooth the data, while removing invalid and outlier values, and normalize the data range to the [0,1] interval. S3: The preprocessed data is segmented using the sliding window method to construct training and testing samples. The trajectory data of the drop-down probe component has continuous and temporal characteristics, while the hidden layers of the LSTM neural network have feedback connections, so the output at this moment is not only related to the input at this moment, but also can combine information from previous moments, which can be effectively used in the trajectory prediction of the drop-down probe component. The dataset is segmented using the sliding window method and used as the input of the neural network in sequence. The number of data points in the dataset is N = {x0, x1, ..., x...} N-1 The first M points are designated as training samples, and the last K points are designated as test samples; the window length is m, that is, the data {x0, x1, ..., xn} from time 0 to m-1 are used. m-1 Predict the trajectory information of the drop-in probe component at time m, using data from time 1 to time m {x1, x2, ..., x...} m The trajectory information at time m+1 is predicted, and so on, resulting in a total of M-m+1 points as training sample inputs and K-m+1 points as test sample inputs. Therefore, the dimension of the input data during LSTM neural network training is (M-m+1, m, 16), where M is the number of training sample points, m is the window length, and 16 represents the 16-dimensional data information contained in each point, as follows: Longitude, latitude, altitude, three-dimensional velocity, three-dimensional acceleration, two-dimensional angular velocity, two-dimensional angular acceleration, temperature, humidity, air pressure; S4: Establish a trajectory prediction model for the drop-down probe component based on LSTM. By iteratively adjusting the network parameters, minimize the error between the output result and the true value, and determine the network structure. The LSTM network incorporates three "gate" controllers by referencing the human characteristic of remembering and forgetting information based on its importance. After the data is input into the neural network, it passes through the forget gate, update gate, and output gate to finally obtain the neural network output. S5: The test sample is input into the trained LSTM neural network, and the output is denormalized to obtain the trajectory prediction result of the drop probe component; By inputting test samples of length K into the trained trajectory prediction model of the drop-type probe component based on LSTM neural network, the output results are inversely normalized to zero-mean and compared with the actual trajectory to verify the accuracy of LSTM neural network trajectory prediction. The trained trajectory prediction model of the drop-type probe component is saved and can be used directly in subsequent new trajectory prediction tasks to improve prediction efficiency and ensure the real-time performance of trajectory prediction. S6: In actual detection, environmental information of the detection airspace is first collected, and then integrated with the trajectory information of the deployed drop-type detection component and input into the LSTM neural network to predict the future trajectory of the detection component.

2. The trajectory prediction method for a drop-type detection component based on an LSTM neural network as described in claim 1, characterized in that: In step S1, The drop-down detection component is deployed at a designated altitude by an airborne platform, while a parachute opens above it. The trajectory data of the drop-down detection component after deployment is acquired by sensors. The drop-down detection component is equipped with a GPS / BeiDou positioning system, an inertial navigation system, a barometric pressure sensor, a temperature sensor, a humidity sensor, and a communication system. The air-based platform includes airships and rockets; The data used for trajectory prediction includes 16 dimensions: longitude, latitude, altitude, three-dimensional velocity, three-dimensional acceleration, two-dimensional angular velocity, two-dimensional angular acceleration, temperature, humidity, and air pressure. Because the drop-down probe is small and lightweight, aerodynamic forces play a significant role in its movement. Environmental factors have a crucial influence on aerodynamic forces; therefore, in addition to the position, attitude, and velocity of the drop-down probe, the influence of environmental parameters on trajectory prediction must also be characterized. These environmental parameters include temperature, humidity, and air pressure. A segment of trajectory data is selected from the collected data stored in the ground receiving system and used as a training sample for the neural network.

3. The trajectory prediction method for a drop-type detection component based on an LSTM neural network as described in claim 2, characterized in that: In step S2, For three-axis angles, data within the range of (-180°, 180°) are retained. For eastward and northward velocities measured by the GPS / BeiDou positioning system, the velocity change Δv between two adjacent data points is calculated. If Δv is significantly greater than a set threshold, the current velocity data is discarded. For the outputs of the barometer and temperature / humidity sensor, the data filtering method is the same as that for velocity data, i.e., outliers exceeding the threshold are removed. After removing outliers, maximum likelihood estimation is used to fill in missing values. That is, first, the missing values ​​are replaced with estimated values, then the parameters of the complete data are estimated, and then the missing values ​​are estimated again based on the above parameter estimates. The likelihood function is as follows: Where, D = {x1, x2, ..., x} N Let {x1, x2, ..., x} be the sample set, and the joint probability density function p(D|θ) is called the probability density function relative to {x1, x2, ..., x}. N Let l(θ) be the likelihood function of θ. Find the value of θ that maximizes the likelihood function. This value is the maximum likelihood estimate and can be used to replace the missing value. Due to the influence of environmental convection, radiation, conduction, and the electronic components used in the measurement, compensation is required for the temperature and pressure parameters measured by the sensor. The temperature parameter errors acquired by the down-projection detection component include Joule thermal error, aerodynamic thermal error, and irradiation thermal error. The correction method for the temperature measurement data is as follows: Where T is the actual temperature, T a For actual measured temperature, The Joule thermal correction error is given, where U is the voltage across the platinum thin film, t is the time of action, A1 is the cross-sectional area of ​​the object perpendicular to the thermal conduction direction, and R... T λ is the resistance value of the platinum thin film, and λ is the thermal conductivity. The irradiation correction error is given by ε, where ε is the emissivity of the object, ranging from 0 to 1, A2 is the irradiated surface area, σ is the Stefan-Boltzmann constant, and T... r The blackbody thermodynamic temperature; For aerodynamic thermal correction error, r is the temperature correction coefficient, V is the relative air velocity, and c P The specific heat capacity of air at isobaric pressure; When the detection component encounters air obstruction during its descent, pressure will be generated on its surface. Therefore, the pressure measured by the barometric pressure sensor includes not only static atmospheric pressure but also the pressure generated by air obstruction during its movement (called wind pressure). The relationship between the pressure magnitude and velocity is expressed as follows: w p =0.5*ro*v 2 Where w p Let w represent wind pressure, ro represent air density, and v represent the falling velocity of the detection component. Therefore, during the falling process of the detection component, the measured value of the air pressure sensor is w. d =w s +w p , where w s The static air pressure value is the actual value that needs to be measured. However, due to the superposition of wind pressure, the air pressure sensor reading is larger than the actual value. Therefore, it is necessary to subtract the wind pressure to obtain the actual air pressure value. Due to the large number of input parameters, each with a different value range, zero-mean normalization is used to prevent some parameters from becoming too large and causing unstable calculation results. This ensures that the data values ​​fall within the [0,1] interval, allowing the network to converge quickly and making it easier to train. The zero-mean normalization transformation function is as follows: In the formula, x is the actual value of a certain data, μ is the mean of the original data, δ is the standard deviation, and x′ is the normalized value.

4. The trajectory prediction method for a drop-type detection component based on an LSTM neural network as described in claim 3, characterized in that: In step S4, The formula for data processing after inputting into the LSTM neural network is as follows: f t =σ(W hf h t-1 +W xf x t +b f ) i t =σ(W hi h t-1 +W xi x t +b i ) o t =σ(W ho h t-1 +W xo x t +b o ) h t = no t fishy(c) t ) Where x t Given 16-dimensional data at time t as input, f t For the Gate of Oblivion, i t To update the door, o t For the output gate, h t For the hidden layer output, c t Let σ represent the state of the LSTM neuron cell unit, σ be the sigmoid activation function, tanh be the tanh activation function, W be the weight matrix, and b be the bias parameter matrix. The forget gate adjusts the degree of retention f of information from the previous unit by adjusting the weights. t When the weight is 0, the information is completely discarded; when the weight is 1, it is completely retained. The update gate determines how much of the input data at the current time needs to be retained by setting a threshold, thus obtaining i. t Combined with the temporary state of the cell unit c t Update to the current cell unit state c t In order to continue the backward transmission; The output gate will output the current input data x. t Compared with the output data h from the previous time step t-1 The weight o for the next time step is obtained by combining the calculations. t Then, combined with the current cell unit state c t Finally, the output h at the current time is obtained. t ; The training samples constructed in S3 are input into the LSTM neural network, which is then trained on a GPU. The cuDNN library is used to accelerate the convolutional neural network, and the Adam optimization algorithm is employed to update parameters and weights, enabling the network model to converge faster. The root mean square error between the predicted and actual data is chosen as the loss function to optimize the parameters of the LSTM neural network for better prediction results. The root mean square error function is as follows: Where y i Let y′ be the true value of the parameter at time i. i Let be the predicted value at time i, and n be the number of data points in the training input. The LSTM neural network outputs trajectory predictions for M-m+1 points. Each point contains 13-dimensional trajectory data, namely: longitude, latitude, altitude, three-dimensional velocity, three-dimensional acceleration, two-dimensional angular velocity, and two-dimensional angular acceleration.

5. The trajectory prediction method for a drop-type detection component based on an LSTM neural network as described in claim 4, characterized in that: In step S6, Before the actual deployment of the drop-in probe, environmental parameters such as temperature, humidity, and air pressure in the deployment airspace are obtained in advance using a prototype probe. These 3D environmental parameters are then combined with the predicted 13D trajectory information of the drop-in probe, resulting in a total of 16 parameters to ensure consistency with the dimension of the input LSTM neural network. A recursive prediction method is used, employing a sliding window to input the prediction results from the first step, followed by the prediction of the second step. This process is repeated to continuously predict the future trajectory. Prediction Step 1: X = [X0, X1, ..., X m-1 The prediction result is: Combining 3D environmental parameters Expanded to 16 dimensions; Prediction step 2: X = [X1, X2, ..., X m The prediction result is: Combining 3D environmental parameters Expanded to 16 dimensions; … Predicting step L: X = [X L-1 ,X L ,…,X m+L-2 The prediction result is: Obtain the complete trajectory prediction curve Based on the predicted trajectory, select the appropriate timing and point for dropping the drop-type detection component to ensure that it falls into the corresponding airspace for in-situ detection, thereby improving the efficiency of the detection mission.

6. A trajectory prediction method for a drop-type detection component based on an LSTM neural network as described in claim 2, 3, 4, or 5, characterized in that: The communication system has a sampling rate of 1Hz and sends the collected signal to the ground receiving system every 1 second.

Images