Deep learning-based parabolic trajectory prediction method and system

Abstract

The invention relates to the technical field of deep learning, in particular to a parabolic trajectory prediction method and system based on deep learning, and the method comprises the following steps: collecting parabolic trajectory data under a differentiated environment condition, subdividing the initial speed, emission angle and position of each trajectory data point, carrying out the standardized sorting by taking a timestamp as a sequence, and obtaining a target trajectory data point; and calculating an optimal matching point between the sequences by using feature extraction. According to the method, by refining and re-matching the parabolic trajectory data, the time sequence and accuracy of data processing are improved, prediction errors caused by environmental changes are reduced, the prediction result is made to better conform to the actual situation, the bidirectional learning method is adopted, processing of the dependency relationship between data points is optimized, and the prediction accuracy is improved. The coping capacity of the model under complex change conditions is enhanced, the model can automatically correct deviation in the prediction process due to introduction of a feedback loop, the prediction accuracy is further improved, and the data processing efficiency and the trajectory prediction precision are remarkably improved through the measures.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a parabola trajectory prediction method and system based on deep learning. Background Art

[0002] Deep learning is a technique that achieves machine learning by simulating the way the human brain works. It is mainly based on a type of computational model called artificial neural networks, especially those designed with multi-layer (deep) structures. These models are trained with large amounts of data to automatically recognize complex patterns and features. The applications of deep learning are very wide, covering a wide range of fields, including image and speech recognition, natural language processing, and autonomous vehicle technology. Its key advantage is its ability to process and learn from large amounts of unstructured data, thereby making efficient predictions and decisions on this data without explicit programming instructions.

[0003] Among them, the parabola trajectory prediction method based on deep learning refers to the use of deep learning technology to predict the trajectory of an object under the influence of external forces such as gravity. This prediction technology is usually used in physics experiments, motion analysis, military and aerospace fields, and other scenarios that require accurate calculation of the trajectory of projectiles. By training neural networks to recognize trajectory data under different initial velocities, angles, and environmental conditions, this technology can help scientists and engineers more accurately simulate and predict the motion behavior of objects.

[0004] In the current deep learning technology, the commonly used network structures and algorithms in parabolic trajectory prediction show insufficient flexibility and adaptability when dealing with complex dynamic environments. Especially in scenarios that require extremely high precision, such as physical experiments and aerospace applications, traditional methods may fail to fully capture the influence of large-scale initial conditions, such as speed and angle, resulting in insufficient prediction accuracy. This limitation may cause errors in precision tasks, affecting the safety and efficiency of task execution. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings existing in the prior art and propose a parabolic trajectory prediction method and system based on deep learning.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a parabola trajectory prediction method based on deep learning, comprising the following steps:

[0007] S1: Collect parabolic trajectory data under differentiated environmental conditions, subdivide the initial speed, launch angle and position of each trajectory data point, standardize and sort them in timestamp order, use feature extraction to calculate the optimal matching points between sequences, adjust the data points to match the time series, and generate time-synchronized sequence data;

[0008] S2: Input the time-synchronized sequence data into the bidirectional LSTM network, learn from both ends of the sequence simultaneously, convert the dependency of each data point with the adjacent data point into a state update, merge the updated states, and generate a merged feature vector;

[0009] S3: Based on the merged feature vector, analyze the weight influence of the features at each time point, adjust the weight of each feature, correct the model prediction deviation through a feedback loop, and generate optimized weight data;

[0010] S4: Apply the optimized weight data, adjust the time warp parameters in the bidirectional LSTM network, fine-tune the time perception layer in the network structure, adapt to environmental changes, output the predicted trajectory, and establish a predicted trajectory model.

[0011] As a further solution of the present invention, the step of acquiring the time synchronization sequence data is specifically:

[0012] S111: Collect parabola trajectory data from multiple environmental conditions, including initial velocity, launch angle, and position information of each data point, sort the data according to timestamps, and generate a preliminary trajectory data set;

[0013] S112: Based on the preliminary trajectory data set, analyze the fit between data points through feature extraction, identify and calculate the optimal matching point, and generate a feature matching data set;

[0014] S113: Using the feature matching data set, time series adjustment is performed on the data points using the formula:

[0015]

[0016] Calculate the time offset and generate an adjusted time series dataset;

[0017] Where Δt represents the time offset, t obs,i Represents the observation data timestamp, t model,i Represents the model prediction timestamp, w i represents the weight of the matching point in time adjustment, and n represents the number of data points;

[0018] S114: using the adjusted time series data set, synchronizing the time series of all trajectory data points to ensure that multiple data points are consistent in time, and obtaining time-synchronized series data.

[0019] As a further solution of the present invention, the step of obtaining the merged feature vector is specifically as follows:

[0020] S211: inputting the time-synchronized sequence data into a bidirectional LSTM network, synchronously processing data points from both ends of the sequence, analyzing the dependency of each data point with adjacent data points, and generating forward and backward dependency analysis results;

[0021] S212: Calculate the state update of each data point according to the forward and backward dependency analysis results, using the formula:

[0022] u i =α·f i +β·b i

[0023] Generate status update results;

[0024] Among them, u i represents the updated state of the i-th data point, f i represents the forward state, b i represents the backward state, α and β are the weight adjustment coefficients of the forward and backward states respectively;

[0025] S213: Using the status update result, merge the updated statuses and integrate them through a weighted formula:

[0026] s i =γ·(u f,i +u b,i )

[0027] Generate merged status results;

[0028] Among them, s i represents the merged state vector, u f,i 、u b,i represents the updated forward and backward states, and γ is the state merging weight;

[0029] S214: Extract feature vectors from the merged state result, each vector comprehensively representing the time and space characteristics of a group of data points, to generate a merged feature vector set.

[0030] As a further solution of the present invention, the step of obtaining the optimization weight data is specifically as follows:

[0031] S311: Starting from the combined feature vector, analyzing the influence of features at multiple time points, using statistical methods to calculate initial weights of multiple features, determining a baseline weight for each feature, and generating an initial weight analysis result;

[0032] S312: Based on the initial weight analysis result, the weight of each feature is refined and adjusted using the formula:

[0033] w i ′ =wi ·(1+k·(δ i -e i ))

[0034] Calculate the current weight and generate the adjusted weight result;

[0035] Among them, w i ′ represents the adjusted weight, w i represents the original weight, δ i represents the adjustment factor, e i represents the prediction deviation, and k represents the value of the adjustment coefficient;

[0036] S313: Using the adjusted weight result set in combination with a real-time feedback loop to correct deviations in model predictions, the accuracy of each feature weight is ensured through dynamic adjustment to generate optimized weight data.

[0037] As a further solution of the present invention, the step of acquiring the predicted trajectory model is specifically as follows:

[0038] S411: based on the optimized weight data, reconfigure the time warp parameters in the bidirectional LSTM network, refine and adjust the processing logic of each time point, and generate a time warp parameter adjustment result;

[0039] S412: Using the time warp parameter adjustment result, fine-tuning the time perception layer in the network structure is performed using the formula:

[0040] L new =L old ·(1+η·T adj )

[0041] Get the fine-tuned time perception layer result;

[0042] Among them, L new represents the adjusted layer parameters, L old is the original layer parameter, T adj is the time distortion adjustment amount, η is the fine-tuning coefficient;

[0043] S413: Apply the fine-tuned time perception layer to process real-time data, output a predicted trajectory in combination with environmental changes, and establish a predicted trajectory model.

[0044] A parabola trajectory prediction system based on deep learning, wherein the parabola trajectory prediction system based on deep learning is used to execute the above-mentioned parabola trajectory prediction method based on deep learning, and the system comprises:

[0045] The data synchronization module collects parabolic trajectory data from differentiated environmental conditions, sorts the initial velocity, launch angle and position of each data point by timestamp, and synchronizes the time series data to generate standardized time series data;

[0046] The feature vector synthesis module inputs the standardized time series data into a bidirectional LSTM network, calculates the dependency of each data point with its adjacent data points, updates the state and merges the state information to obtain a merged feature vector;

[0047] The weight optimization module analyzes and adjusts the feature weights at each time point based on the merged feature vector, corrects the prediction deviation through a feedback mechanism, and generates optimized weight data;

[0048] The time warp adjustment module applies the optimized weight data to adjust the time warp parameters in the bidirectional LSTM network, fine-tune the time perception layer to match the environmental changes, and generate adjusted network parameters;

[0049] The trajectory prediction model building module uses the adjusted network parameters to perform trajectory prediction operations, output predicted trajectories, and establish a parabolic trajectory prediction model.

[0050] Compared with the prior art, the advantages and positive effects of the present invention are:

[0051] In the present invention, by refining and re-matching the parabolic trajectory data, the timing and accuracy of data processing are improved, the prediction error caused by environmental changes is reduced, and the prediction results are more in line with the actual situation. The two-way learning method is adopted to optimize the processing of dependencies between data points and enhance the model's ability to cope with complex changing conditions. The introduction of the feedback loop enables the model to self-correct deviations during the prediction process, further improving the accuracy of the prediction. These measures significantly improve the data processing efficiency and the accuracy of trajectory prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic diagram of the workflow of the present invention;

[0053] Figure 2 This is a flow chart of the steps for obtaining time synchronization sequence data of the present invention;

[0054] Figure 3 A flow chart of the steps for obtaining the combined feature vector of the present invention;

[0055] Figure 4 A flow chart of the steps for obtaining weight data for optimization of the present invention;

[0056] Figure 5 This is a flow chart of the steps for obtaining the prediction trajectory model of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0058] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.

[0059] Embodiment 1

[0060] See also Figure 1 The present invention provides a technical solution: a parabola trajectory prediction method based on deep learning, comprising the following steps:

[0061] S1: Collect parabolic trajectory data under differentiated environmental conditions, subdivide the initial speed, launch angle and position of each trajectory data point, standardize and sort them in timestamp order, use feature extraction to calculate the optimal matching points between sequences, adjust the data points to match the time series, and generate time-synchronized sequence data;

[0062] S2: Input time-synchronized sequence data to the bidirectional LSTM network, learn from both ends of the sequence simultaneously, convert the dependency of each data point with adjacent data points into state updates, merge the updated states, and generate a merged feature vector;

[0063] S3: Based on the merged feature vector, analyze the weight impact of the features at each time point, adjust the weight of each feature, correct the model prediction deviation through the feedback loop, and generate optimized weight data;

[0064] S4: Apply optimized weight data, adjust the time warp parameters in the bidirectional LSTM network, fine-tune the time perception layer in the network structure, adapt to environmental changes, output the predicted trajectory, and establish a predicted trajectory model.

[0065] The time-synchronized sequence data includes initial velocity, launch angle, position and timestamp. The merged feature vector includes bidirectional LSTM forward features, backward features and fusion representation. The optimized weight data includes feature weights and feedback adjustment parameters. The predicted trajectory is specifically the trajectory point sequence and velocity information.

[0066] See also Figure 2 ,The specific steps for obtaining time-synchronized series data are:

[0067] S111: Collect parabola trajectory data from multiple environmental conditions, including initial velocity, launch angle, and position information of each data point, sort the data according to timestamps, and generate a preliminary trajectory data set;

[0068] Parabolic trajectory data is collected from multiple environmental conditions, including the initial speed, launch angle and position information of each data point. The data is sorted according to timestamps. By measuring the data points, including using radar and optical sensors to capture the speed and angle of each parabola, and recording the launch point and time information of each parabola, a preliminary trajectory data set is generated. This data set can be used for iterative trajectory analysis and pattern recognition. This method can ensure that the collected data is both accurate and complete for trajectory matching and feature analysis.

[0069] S112: Based on the preliminary trajectory data set, analyze the fit between data points through feature extraction, identify and calculate the optimal matching points, and generate a feature matching data set;

[0070] Feature extraction technology is applied to the preliminary trajectory data set. By analyzing the fit between data points, the optimal matching points are identified and calculated. The trajectory data is analyzed using machine learning algorithms to identify similar or repeated trajectory patterns, including using clustering algorithms to group trajectory points, calculating the fit of data points in each group, and determining the trajectory points with the highest matching degree. This process involves the comprehensive application of multiple algorithms, from data cleaning to feature extraction, to ensure that each matching point is the optimal representative based on the actual trajectory data, thereby generating a feature matching data set. This data set provides the basis for subsequent time series adjustment.

[0071] S113: Using the feature matching data set, adjust the data points in time series using the formula:

[0072]

[0073] Calculate the time offset and generate an adjusted time series dataset;

[0074] Where Δt represents the time offset, t obs,i Represents the observation data timestamp, t model,i Represents the model prediction timestamp, w i represents the weight of the matching point in time adjustment, and n represents the number of data points;

[0075] formula:

[0076]

[0077] The benefit of the formula is that by introducing the weight factor w i , so that each data point has a different influence according to its reliability or accuracy, thus calibrating the time series;

[0078] Detailed explanation of the formula and the process of formula calculation and derivation:

[0079] Set the observation timestamp t of three data points obs They are 1.5s, 2.5s, 3.5s and the model prediction timestamp t model They are 1s, 2s, and 3s respectively, and the weights w are 0.5, 1, and 1.5 respectively. Insert them into the formula for calculation:

[0080]

[0081] The results show that the average time offset after adjustment is about 0.167s, indicating that the observed time series needs to be adjusted forward by about 0.167 seconds to better align with the model prediction and thus synchronize the time series.

[0082] S114: Using the adjusted time series data set, synchronize the time series of all trajectory data points to ensure that multiple data points are consistent in time, and obtain time-synchronized series data.

[0083] Using the adjusted time series data set, the time series of all trajectory data points are synchronized to ensure that multiple data points are consistent in time and obtain time-synchronized series data. By adjusting the timestamps of the data points, the observed data is aligned with the time series of the prediction model. This process involves recalculating and correcting the timestamp of each data point to ensure that all data points can be compared and analyzed within the same time frame. This not only improves the availability of the data, but also enhances the accuracy of data analysis. The obtained time-synchronized series data provides a solid foundation for the next step of analysis.

[0084] See also Figure 3 , the specific steps for obtaining the merged feature vector are:

[0085] S211: input the time-synchronized sequence data into the bidirectional LSTM network, synchronously process the data points from both ends of the sequence, analyze the dependency of each data point with adjacent data points, and generate forward and backward dependency analysis results;

[0086] The time-synchronized sequence data is input into the bidirectional LSTM network. The network processes data points synchronously from both ends of the sequence and analyzes the dependency of each data point with adjacent data points. The network uses forward and backward recurrent neural network structures to capture long-term dependencies in time series data. This structure enables the network to learn information from both ends of the sequence to the center, optimize the time dependency analysis between data points, and improve the model's ability to capture time series features by dynamically adjusting the network's weights and biases. This method is particularly suitable for scenarios that require extracting time characteristics from sequence data, such as speech recognition and stock price prediction. Through bidirectional learning, the network can more comprehensively parse the before-and-after relationships of the data structure, thereby generating more accurate dependency analysis results.

[0087] S212: Calculate the state update of each data point based on the forward and backward dependency analysis results, using the formula:

[0088] u i =α·f i +β·b i

[0089] Generate status update results;

[0090] Among them, u i represents the updated state of the i-th data point, f i represents the forward state, b i represents the backward state, α and β are the weight adjustment coefficients of the forward and backward states respectively;

[0091] formula:

[0092] u i =α·f i +β·b i

[0093] The benefit of the formula is that by introducing weight parameters α and β, the forward state f can be adjusted i and the backward state b i The influence of the weighted method makes the model more flexible to match differentiated data features. This weighted method helps to balance the time dynamics in sequence data and improve the accuracy of state updates.

[0094] Detailed explanation of the formula and the process of formula calculation and derivation:

[0095] Set the forward state f i The value is 0.6, the backward state b i The value of is 0.4, and the weights α and β are set to 0.5 and 0.5 respectively. The calculation process is as follows:

[0096] u i =0.5·0.6+0.5·0.4=0.3+0.2=0.5

[0097] The results show that the state update result u i It is 0.5, which means that under a given weight, the forward and backward states are referenced evenly, reflecting the average effect of the data point updating the state, which helps to more accurately merge the states in subsequent steps and improve the prediction performance of the model.

[0098] S213: Using the status update result, merge the updated status and integrate them through the weighted formula:

[0099] s i =γ·(u f,i +u b,i )

[0100] Generate merged status results;

[0101] Among them, s i represents the merged state vector, u f,i 、u b,i represents the updated forward and backward states, and γ is the state merging weight;

[0102] formula:

[0103] s i =γ·(u f,i +u b,i )

[0104] The benefit of the formula is that by adjusting the merge weight γ, the forward and backward update states u can be effectively balanced f,i and u b,i The contribution of makes the state merging not only refer to the information at both ends of the time series, but also can cope with the time series data with differentiated characteristics by matching and adjusting γ, which enhances the flexibility and accuracy of the model in processing data with differentiated time dependence intensity.

[0105] Detailed explanation of the formula and the process of formula calculation and derivation:

[0106] Set the forward update state u f,i The value is 0.7, and the state u is updated backward b,i The value of is 0.3, the merge weight γ is set to 0.8, and the calculation process is as follows:

[0107] s i =0.8·(0.7+0.3)=0.8·1.0=0.8

[0108] The results show that the merged state vector s i It is 0.8, which means that under the current weight, the model effectively integrates the state information from both ends of the sequence, improves the integration effect of features, and provides a basis for generating feature vector sets.

[0109] S214: Extract feature vectors from the merged state result, each vector comprehensively represents the time and space characteristics of a group of data points, and generate a merged feature vector.

[0110] Feature vectors are extracted from the merged state results. Each vector comprehensively represents the time and space characteristics of a group of data points. This process involves high-dimensional feature transformation of the merged state, which is carried out through the stacking of activation functions and network layers in the machine learning model. The extraction of feature vectors not only includes the time information of the data points, but also integrates the spatial relationships. This processing method enables the feature vector to more comprehensively describe the intrinsic characteristics of the data, and provide subsequent machine learning tasks including classification and prediction with richer and more useful input information. Through the analysis of feature vectors, time series prediction or classification can be more accurately performed, which helps to improve the accuracy of decision support systems, thus playing a key role in practical applications.

[0111] See also Figure 4 , the specific steps for obtaining the optimized weight data are:

[0112] S311: Starting from the merged feature vector, analyze the influence of features at multiple time points, use statistical methods to calculate the initial weights of multiple features, determine the baseline weight of each feature, and generate initial weight analysis results;

[0113] In the process of analyzing and merging feature vectors, the influence of each feature at a time point is quantified through statistical methods. Not only is the performance of multiple features in the model referred to, but also its weight is adjusted through real-time data feedback to ensure the accuracy and practicality of the analysis. In this way, the influence of each feature on the model output is obtained, and then one by one is adjusted according to the actual performance of the model. The statistical method compares the deviation between the predicted values ​​and the actual values ​​of multiple features, and adjusts the weights of multiple features according to the size of the deviation to ensure that the model after weight adjustment can more accurately predict trends in future time periods. The implementation effect of this method will directly affect the optimization and accuracy of the model.

[0114] S312: Based on the initial weight analysis results, the weight of each feature is refined and adjusted using the formula:

[0115] w ′ i =w i ·(1+k·(δ i -e i ))

[0116] Calculate the current weight and generate the adjusted weight result;

[0117] Among them, w ′ i represents the adjusted weight, wi represents the original weight, δ i represents the adjustment factor, e i represents the prediction deviation, and k represents the value of the adjustment coefficient;

[0118] formula:

[0119] w ′ i =w i ·(1+k·(δ i -e i ))

[0120] The benefit of the formula is that by introducing the adjustment coefficient k and the dynamic adjustment factor δ i -e i , the weights can be flexibly adjusted according to the deviations predicted by the actual model, thus enhancing the model’s adaptability to data fluctuations and prediction accuracy.

[0121] Detailed explanation of the formula and the process of formula calculation and derivation:

[0122] Set the initial weight w i =0.5, adjustment coefficient k = 0.1, adjustment factor δ i =0.05, prediction deviation e i =0.02;

[0123] The calculation process is as follows:

[0124] w ′ i =0.5·(1+0.1·(0.05-0.02))=0.5·(1+0.1·0.03)=0.5·1.003=0.5015

[0125] The results show that by fine-tuning the weights, the model can more precisely adjust its response to the target features and improve the prediction performance.

[0126] S313: Use the adjusted weight result set in combination with a real-time feedback loop to correct deviations in model predictions, ensure the accuracy of each feature weight through dynamic adjustment, and generate optimized weight data.

[0127] The adjusted weights are combined with a real-time feedback loop to correct the model's prediction bias. This dynamic correction mechanism makes the adjustment of each feature weight not only based on data, but also combined with real-time data feedback. By cyclically adjusting and optimizing weight parameters, it ensures that the model can show higher flexibility and adaptability when facing current data. It not only improves the model's prediction accuracy, but also enhances the model's ability to respond to emergencies. By performing real-time optimization of weights, the generated optimized weight data can more accurately reflect the actual impact of each feature on the model output, thereby making the model prediction closer to the actual situation.

[0128] See also Figure 5 , the specific steps for obtaining the prediction trajectory model are:

[0129] S411: based on the optimized weight data, reconfigure the time warp parameters in the bidirectional LSTM network, refine and adjust the processing logic of each time point, and generate a time warp parameter adjustment result;

[0130] The process of reconfiguring the time warp parameters of the bidirectional LSTM network based on the optimized weight data requires analyzing the adaptability of the dynamic changes of data points. This process involves the analysis of batch real-time monitoring data to determine the most appropriate parameters for each time point. By collecting and analyzing data fluctuations in differentiated time periods, the adjustment value of the time warp parameter can be calculated. This value reflects the impact of environmental changes on network performance. Adjusting the time warp parameters can make the network better match future data changes, and real-time adjustment ensures the network's flexibility and ability to respond to emergencies. The generated time warp parameter adjustment results provide the network with more accurate time perception capabilities to match the cyclically changing input data characteristics.

[0131] S412: Using the time warp parameter adjustment result, fine-tuning the time perception layer in the network structure is performed using the formula:

[0132] L new =L old ·(1+η·T adj )

[0133] Get the fine-tuned time perception layer result;

[0134] Among them, L new represents the adjusted layer parameters, L old is the original layer parameter, T adj is the time distortion adjustment amount, η is the fine-tuning coefficient;

[0135] formula:

[0136] L new =L old ·(1+η·T adj )

[0137] The benefit of the formula is that by adjusting the coefficient η and the time warp parameter T adj , the response speed and sensitivity of the network layer can be flexibly adjusted to match differentiated operating environments and data changes, thereby improving the adaptability and accuracy of the model.

[0138] Detailed explanation of the formula and the process of formula calculation and derivation:

[0139] Set the original time perception layer parameter L old is 0.5, the adjustment coefficient η is set to 0.05, and the time warp parameter adjustment value T adj Based on data analysis, it is determined to be 2, then

[0140] L new =0.5·(1+0.05·2)=0.5·1.1=0.55

[0141] The results show that by fine-tuning, the parameters of the time-aware layer are increased from 0.5 to 0.55, which enhances the layer’s ability to match time warps and improves the network’s sensitivity and reaction speed to temporal changes.

[0142] S413: Apply the fine-tuned time perception layer to process the real-time data, output the predicted trajectory in combination with the environmental changes, and establish a predicted trajectory model.

[0143] The fine-tuned time perception layer is processed using real-time data. During this process, data on environmental changes are integrated to adjust the network output. First, real-time data is collected through environmental monitoring equipment, and then the data is input into the fine-tuned LSTM network. The network calculates the predicted trajectory based on the characteristics and data patterns of the input data and the newly adjusted time perception layer parameters. This calculation process includes not only the forward propagation of data, but also reverse error propagation to optimize predictions in future time periods. Through data processing and network adjustment, it is ensured that the output prediction trajectory is as close to the actual situation as possible. The resulting prediction trajectory model will be used for prediction tasks in similar environments in the future to ensure that the model can continue to provide high-precision outputs.

[0144] A parabolic trajectory prediction system based on deep learning, which is used to execute the above-mentioned parabolic trajectory prediction method based on deep learning, comprises:

[0145] The data synchronization module collects parabolic trajectory data from differentiated environmental conditions, sorts the initial velocity, launch angle and position of each data point by timestamp, and synchronizes the time series data to generate standardized time series data;

[0146] The feature vector synthesis module inputs the standardized time series data into the bidirectional LSTM network, calculates the dependency of each data point with its adjacent data points, updates the state and merges the state information to obtain the merged feature vector;

[0147] The weight optimization module analyzes and adjusts the feature weights at each time point based on the merged feature vectors, corrects the prediction deviation through the feedback mechanism, and generates optimized weight data;

[0148] The time warp adjustment module applies optimized weight data, adjusts the time warp parameters in the bidirectional LSTM network, fine-tunes the time perception layer to match the environmental changes, and generates adjusted network parameters;

[0149] The trajectory prediction model building module uses the adjusted network parameters to perform trajectory prediction operations, output the predicted trajectory, and build a parabolic trajectory prediction model.

[0150] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A parabola trajectory prediction method based on deep learning, characterized in that: The following steps are involved: Collect parabolic trajectory data under differentiated environmental conditions, subdivide the initial speed, launch angle and position of each trajectory data point, standardize and sort them in order of timestamp, use feature extraction to calculate the optimal matching points between sequences, adjust the data points to match the time series, and generate time-synchronized sequence data; Input the time-synchronized sequence data into a bidirectional LSTM network, learn from both ends of the sequence simultaneously, convert the dependency of each data point with adjacent data points into a state update, merge the updated states, and generate a merged feature vector; Based on the combined feature vector, the weight influence of the features at each time point is analyzed, the weight of each feature is adjusted, the model prediction deviation is corrected through a feedback loop, and optimized weight data is generated; The optimized weight data is applied to adjust the time warp parameters in the bidirectional LSTM network, fine-tune the time perception layer in the network structure, adapt to environmental changes, output the predicted trajectory, and establish a predicted trajectory model.

2. The parabola trajectory prediction method based on deep learning according to claim 1, characterized in that: The steps for acquiring the time synchronization sequence data are specifically as follows: Collect parabolic trajectory data from multiple environmental conditions, including initial velocity, launch angle, and position information for each data point, sort the data by timestamp, and generate a preliminary trajectory dataset; Based on the preliminary trajectory data set, analyzing the fit between data points through feature extraction, identifying and calculating the optimal matching points, and generating a feature matching data set; Using the feature matching dataset, the data points are adjusted in time series using the formula: Calculate the time offset and generate an adjusted time series dataset; Where Δt represents the time offset, t obs,i Represents the observation data timestamp, t model,i Represents the model prediction timestamp, w i represents the weight of the matching point in time adjustment, and n represents the number of data points; The adjusted time series data set is used to synchronize the time series of all trajectory data points, ensuring that multiple data points are consistent in time, and obtaining time-synchronized series data.

3. The parabola trajectory prediction method based on deep learning according to claim 2 is characterized in that: The steps of obtaining the merged feature vector are specifically as follows: Inputting the time-synchronized sequence data into a bidirectional LSTM network, synchronously processing data points from both ends of the sequence, analyzing the dependency of each data point with adjacent data points, and generating forward and backward dependency analysis results; Based on the forward and backward dependency analysis results, the state update of each data point is calculated using the formula: you i =α·f i +b·b i Generate status update results; Among them, u i represents the updated state of the i-th data point, f i represents the forward state, b i represents the backward state, α and β are the weight adjustment coefficients of the forward and backward states respectively; Using the state update results, the updated states are merged and integrated using a weighted formula: with i =γ·(in f,i +in b,i ) Generate merged status results; Among them, s i represents the merged state vector, u f,i 、u b,i represents the updated forward and backward states, and γ is the state merging weight; Feature vectors are extracted from the merged state results, each vector comprehensively representing the temporal and spatial characteristics of a group of data points, to generate a merged feature vector set.

4. The parabola trajectory prediction method based on deep learning according to claim 3 is characterized in that: The steps for obtaining the optimization weight data are specifically as follows: Starting from the combined feature vector, analyzing the influence of features at multiple time points, using statistical methods to calculate initial weights of multiple features, determining a baseline weight for each feature, and generating initial weight analysis results; Based on the initial weight analysis results, the weight of each feature is refined and adjusted using the formula: w i ′ =w i ·(1+k·(δ i -e i )) Calculate the current weight and generate the adjusted weight result; Among them, w i ′ represents the adjusted weight, w i represents the original weight, δ i represents the adjustment factor, e i represents the prediction deviation, and k represents the value of the adjustment coefficient; The adjusted weight result set is used in combination with a real-time feedback loop to correct deviations in model predictions, and the accuracy of each feature weight is ensured through dynamic adjustment to generate optimized weight data.

5. The parabola trajectory prediction method based on deep learning according to claim 4 is characterized in that: The steps for obtaining the prediction trajectory model are specifically as follows: Based on the optimized weight data, reconfigure the time warp parameters in the bidirectional LSTM network, refine and adjust the processing logic of each time point, and generate a time warp parameter adjustment result; Using the time warp parameter adjustment result, fine-tuning of the time perception layer in the network structure is performed using the formula: L new =L old ·(1+η·T adj ) Get the fine-tuned time perception layer result; Among them, L new represents the adjusted layer parameters, L old is the original layer parameter, T adj is the time distortion adjustment amount, η is the fine-tuning coefficient; The fine-tuned time perception layer is applied to process real-time data, and the predicted trajectory is output in combination with environmental changes to establish a predicted trajectory model.

6. The parabola trajectory prediction system based on deep learning is characterized by: According to the parabola trajectory prediction method based on deep learning according to any one of claims 1 to 5, the system comprises: The data synchronization module collects parabolic trajectory data from differentiated environmental conditions, sorts the initial velocity, launch angle and position of each data point by timestamp, and synchronizes the time series data to generate standardized time series data; The feature vector synthesis module inputs the standardized time series data into a bidirectional LSTM network, calculates the dependency of each data point with its adjacent data points, updates the state and merges the state information to obtain a merged feature vector; The weight optimization module analyzes and adjusts the feature weights at each time point based on the merged feature vector, corrects the prediction deviation through a feedback mechanism, and generates optimized weight data; The time warp adjustment module applies the optimized weight data to adjust the time warp parameters in the bidirectional LSTM network, fine-tune the time perception layer to match the environmental changes, and generate adjusted network parameters; The trajectory prediction model building module uses the adjusted network parameters to perform trajectory prediction operations, output predicted trajectories, and establish a parabolic trajectory prediction model.

Images