Low-altitude non-cooperative target flight conflict early warning method based on motion trail prediction

Abstract

The invention discloses a low-altitude non-cooperative target flight conflict early warning method based on motion trail prediction, and belongs to the technical field of low-altitude flight safety. Comprising the steps of obtaining non-cooperative target trajectory data and performing feature extraction and vectorization; inputting the trajectory feature vector sequence into a pre-trained multi-hidden-layer long short-term memory recurrent neural network LSTM, and outputting a trajectory prediction vector and probability distribution; constructing a three-dimensional probability reachable set representing the flight influence range based on the prediction probability distribution; performing spatial geometric intersection analysis on the three-dimensional probability reachable set and a three-dimensional flight protection area of the cooperative aircraft, and judging a flight conflict risk; and performing real-time rolling updating on the trajectory data through a sliding time window, and iteratively executing trajectory prediction and conflict judgment to realize rolling early warning. According to the method, the trajectory prediction precision and the conflict early warning accuracy are improved, the missing alarm and false alarm rate is reduced, the sufficient early warning lead is guaranteed, and the method is suitable for low-altitude traffic management of the plateau complex terrain.

Description

Technical Field

[0001] This invention relates to the field of low-altitude flight safety technology, and more specifically, to a method for early warning of low-altitude non-cooperative target flight conflicts based on motion trajectory prediction. Background Technology

[0002] With the rapid development of the low-altitude economy, the number of low-altitude aircraft, represented by drones, has increased dramatically. Among them, non-cooperative targets (such as unauthorized "black flight" drones) pose a serious security threat to air traffic and are known as "mobile aerial reefs." Especially in complex terrain conditions such as plateaus, the low-altitude environment is characterized by severe terrain obstruction, variable weather conditions, and complex electromagnetic environments, which makes early warning of flight conflicts with non-cooperative targets extremely challenging.

[0003] In existing low-altitude flight conflict early warning technologies, trajectory prediction of non-cooperative targets is a core component. Currently, mainstream methods mostly employ recurrent neural networks (RNNs) to process time-series data of non-cooperative target trajectories. However, RNNs and their variants have inherent limitations when processing long time series: when the series is long, the network is prone to gradient vanishing or gradient exploding problems during backpropagation, causing the network to lose its ability to learn historical trajectory information from long distances and failing to fully mine and utilize the trajectory's contextual information, resulting in low trajectory prediction accuracy. As trajectory prediction is the foundation of early warning, insufficient accuracy directly affects the reliability of subsequent conflict assessment.

[0004] Furthermore, non-cooperative targets are characterized by high maneuverability, random movement intentions, and high trajectory uncertainty. Traditional deterministic trajectory prediction methods (i.e., outputting only the most probable predicted trajectory) are insufficient to accurately characterize their actual flight impact range. If conflict assessment is based solely on a single, deterministic trajectory prediction result, it is highly likely to lead to underreporting of potential flight conflicts (failure to identify real risks) or false alarms (incorrectly reporting non-existent risks), failing to meet the needs of actual safety monitoring.

[0005] Furthermore, existing early warning methods typically lack dynamic iterative prediction mechanisms for highly maneuverable non-cooperative targets. This means the system cannot dynamically correct and update trajectory predictions based on the target's real-time motion status, resulting in insufficient lead time for warnings and making it difficult to allow air traffic controllers or nearby cooperative aircraft sufficient time for conflict avoidance decisions. This problem is particularly pronounced in complex terrain such as plateaus, making it difficult to meet the urgent requirements for low-altitude traffic safety control in this scenario.

[0006] While there have been attempts to combine trajectory prediction with conflict assessment, a complete method integrating trajectory prediction and conflict early warning has yet to be developed that is specifically designed for highly maneuverable non-cooperative targets, adaptable to the complex low-altitude environment of plateau regions, and capable of adapting to both. Existing technologies have significant shortcomings in terms of trajectory prediction accuracy, conflict early warning accuracy, and warning timeliness, necessitating a more efficient and precise method for low-altitude flight conflict early warning. Summary of the Invention

[0007] The purpose of this invention is to address the aforementioned problems by providing a low-altitude non-cooperative target flight conflict early warning method based on motion trajectory prediction. This method achieves accurate prediction of non-cooperative target trajectories through an improved long short-term memory recurrent neural network, and combines this with a three-dimensional probabilistic reachability set to make scientific judgments about flight conflicts. It enables iterative trajectory prediction and rolling conflict early warning for highly maneuverable non-cooperative targets, improving the accuracy and lead time of low-altitude non-cooperative target flight conflict early warning in complex terrain such as plateaus, and providing reliable technical support for low-altitude traffic safety control.

[0008] The technical solution of the present invention is as follows: A method for early warning of low-altitude non-cooperative target flight conflicts based on motion trajectory prediction includes the following steps: Acquire real-time trajectory data of non-cooperative targets, and perform feature extraction and vectorization on the real-time trajectory data to obtain a sequence of trajectory feature vectors; The trajectory feature vector sequence is input into a pre-trained multi-hidden-layer long short-term memory recurrent neural network (LSTM). The multi-hidden-layer long short-term memory recurrent neural network (LSTM) outputs the trajectory prediction vector sequence of the non-cooperative target in a future preset time period and the corresponding trajectory state prediction probability distribution. Based on the trajectory state prediction probability distribution and combined with the flight dynamics model of non-cooperative targets, a three-dimensional probability reachability set characterizing the flight influence range of non-cooperative targets is constructed. Obtain the planned flight path and three-dimensional flight protection zone of the cooperative aircraft. Perform spatial geometric intersection analysis on the three-dimensional probability reachable set and the three-dimensional flight protection zone. If the intersection probability is greater than the preset safety probability threshold, it is determined that there is a risk of flight conflict. By using a preset sliding time window to continuously update the trajectory data of non-cooperative targets, and iteratively performing trajectory prediction and conflict judgment based on the updated trajectory data, rolling conflict warnings for non-cooperative targets are achieved.

[0009] By organically integrating an improved multi-hidden-layer long short-term memory recurrent neural network (LSTM), a probabilistic conflict judgment mechanism, and a dynamic iterative early warning mechanism, this method not only improves the accuracy of trajectory prediction but also transforms the prediction results into quantifiable conflict risk judgments. Combined with a real-time updating rolling mechanism, it effectively solves the comprehensive technical problems of low prediction accuracy, high false alarm rate, and insufficient early warning lead time in existing technologies, providing a complete and reliable technical solution for flight safety monitoring of low-altitude non-cooperative targets.

[0010] Furthermore, each LSTM unit in the multi-hidden-layer long short-term memory recurrent neural network (LSTM) includes a three-level gating structure of forget gate, input gate, and output gate, as well as a cell state with a dynamic weight self-looping mechanism; the internal operation logic of the LSTM unit is as follows: Forgotten Gate: Control the proportion of cell state information retained from the previous moment; Input Gate: Generate candidate cell states Control the proportion of new information incorporated; Cell status update: This enables linear transmission of trajectory information across time points; Output gate: Hidden layer output Filter and output effective information about cell status; in, The trajectory feature vector input at the current moment. Output the hidden state from the previous time step. Output the hidden state at the current moment. This represents the cell state at the previous moment. This represents the cell state updated at the current moment. This represents the current state of the candidate cells. The gating activation value for the forget gate. This is the gate activation value for the input gate. This is the gate activation value for the output gate. This is the weight matrix corresponding to the forget gate. This is the weight matrix corresponding to the input gate. The weights corresponding to the candidate cell states. This is the weight matrix corresponding to the output gate. For the bias term corresponding to the forget gate, The bias term corresponding to the input gate. This represents the bias term corresponding to the candidate cell state. For the bias term corresponding to the output gate, The Sigmoid activation function is used to map the gate value to... interval, The hyperbolic tangent activation function is used to generate candidate cell states and hidden layer outputs. For element-wise multiplication, This involves concatenating the hidden layer output from the previous time step with the input vector from the current time step.

[0011] By introducing a three-level gating structure and a linear summation update mechanism for cell states, the network acquires the ability to selectively retain or forget historical information, fundamentally solving the inherent defects of gradient vanishing / expansion in RNNs when processing long-term sequences. This enables the multi-hidden-layer Long Short-Term Memory Recurrent Neural Network (LSTM) to fully mine long-distance contextual information in the historical trajectories of non-cooperative targets, significantly improving the trajectory prediction accuracy for highly maneuverable targets and providing a reliable data foundation for subsequent conflict assessment.

[0012] Furthermore, during the training process, the multi-hidden-layer long short-term memory recurrent neural network (LSTM) employs a backpropagation time-based (BPTT) algorithm that combines gradient clipping and gradient smoothing mechanisms for parameter updates. Gradient clipping is used to scale gradients exceeding a preset threshold proportionally. The gradient smoothing mechanism mitigates gradient decay by weighted averaging of gradients at adjacent time points, where the time closest to the current time point is designated as the first time point, and the time farthest from the current time point is designated as the second time point. The gradient weight coefficient at the first time point is greater than the gradient weight coefficient at the second time point.

[0013] Furthermore, the preset threshold for gradient clipping is 1.0; the weight coefficient for gradient smoothing decays over time, with a weight of 0.8 at the first time step and a weight of 0.2 at the second time step.

[0014] Gradient pruning limits the gradient magnitude, preventing parameter oscillations caused by excessively large gradients during training. Gradient smoothing, by weighted averaging of gradients from adjacent time steps, mitigates gradient decay during backpropagation. Together, these mechanisms ensure that the network can stably converge to optimal parameters under conditions of multiple hidden layers and complex trajectory samples, fully learning trajectory features and further improving the model's prediction generalization ability and accuracy in complex high-altitude environments.

[0015] Furthermore, the three-dimensional probabilistic reachable set is constructed based on the Zeno polyhedron, specifically including: setting the dynamic constraint range of the control variables in the plateau environment according to the flight dynamics model of the non-cooperative target; deriving the set of all possible flight states of the non-cooperative target in the future time period according to the trajectory state prediction probability distribution, and transforming it into a three-dimensional probabilistic reachable set represented by the Zeno polyhedron.

[0016] Furthermore, the spatial geometric intersection analysis includes: using a three-dimensional spatial geometric intersection algorithm to calculate the intersection region between the three-dimensional probabilistic reachable set and the three-dimensional flight protection zone of the cooperative aircraft, and determining the conflict probability of the intersection region based on the probability distribution of the three-dimensional probabilistic reachable set.

[0017] By using Zeno polyhedra to transform the probabilistic trajectory prediction results of non-cooperative targets into a three-dimensional probabilistic reachable set in three-dimensional space, the uncertain flight influence range of non-cooperative targets is quantitatively characterized, which better reflects the actual motion characteristics of non-cooperative targets than traditional deterministic trajectory points. Simultaneously, incorporating the three-dimensional flight protection zone of cooperative aircraft into spatial intersection analysis enables a more comprehensive identification of potential conflict risks. The combination of these two methods effectively reduces the false alarm and missed warning rates for conflict warnings caused by the randomness of target maneuvers, thereby improving the accuracy of conflict assessment.

[0018] Furthermore, the preset safety probability threshold is dynamically adjusted and set according to the geographical scene where the non-cooperative target is currently located: a first threshold is used in open areas, and a second threshold lower than the first threshold is used in high-risk areas such as mountains or canyons.

[0019] Furthermore, the first threshold is 8%, and the second threshold is 3%.

[0020] By dynamically adjusting the safety threshold for conflict assessment in different plateau scenarios (such as mountainous areas, canyons, and open plains), the early warning system can adapt to the actual risk levels under complex terrain. A stricter threshold (3%) is used in high-risk areas, enabling earlier detection of potential risks and compensating for safety hazards caused by environmental factors such as terrain obstruction and variable weather. A relatively lenient threshold (8%) is used in open areas, avoiding over-warning. This differentiated configuration enhances the practicality and adaptability of this method in complex low-altitude environments on plateaus.

[0021] Furthermore, the length of the sliding time window is equal to a future preset time period, and the sliding step size is less than the window length; when updating the trajectory data through the sliding time window, the trajectory data at the earliest moment in the window is removed, and the latest collected trajectory data is incorporated to ensure that the data in the window always reflects the latest motion state of the non-cooperative target.

[0022] Furthermore, the future preset time period is 30 seconds, the window length of the sliding time window is 30 seconds, and the sliding step size is 1 second; the update frequency of the rolling conflict warning is once per second.

[0023] By setting the window length to equal the prediction period (30 seconds), a 30-second lead time is ensured for each warning, providing sufficient time for controllers and cooperative aircraft to make conflict avoidance decisions. With a 1-second sliding step and a warning update frequency of once per second, rapid acquisition and continuous tracking of the motion state of non-cooperative targets are achieved. This ensures the system can respond promptly to sudden maneuvers of targets, maintaining the timeliness and accuracy of warning information and effectively solving the problems of slow warning updates and inability to track dynamic changes in traditional methods.

[0024] Compared with existing technologies, the advantages of this invention are: 1. Regarding the accuracy and reliability of conflict early warning, this invention abandons the traditional model of using a single deterministic trajectory prediction result for conflict judgment. Instead, it innovatively transforms the trajectory prediction result into a three-dimensional probabilistic reachable set in three-dimensional space. By introducing Zeno polyhedra to quantify the uncertain flight influence range of non-cooperative targets, and combining this with spatial geometric intersection analysis using the three-dimensional flight protection zone of cooperative aircraft, the potential motion space of non-cooperative targets can be reflected more scientifically and realistically. Simultaneously, by dynamically adjusting the safety probability threshold in high-altitude scenarios, the false alarm and missed alarm rates of flight conflict early warning caused by the high maneuverability and random intentions of targets are effectively reduced, significantly improving the reliability of the early warning system.

[0025] 2. Regarding the real-time and timely nature of early warning, this invention achieves real-time rolling updates, iterative trajectory prediction, and dynamic assessment of conflict risks by constructing a high-frequency sliding time window (e.g., a window length of 30 seconds and a sliding step size of 1 second). This dynamic iterative mechanism ensures that the system can immediately correct the prediction results and early warning level based on the latest motion state of the target, providing controllers and cooperative aircraft with sufficient time for conflict avoidance decisions and solving the technical problems of insufficient early warning lead and inability to effectively respond to sudden maneuvers in existing technologies.

[0026] 3. Regarding environmental adaptability and system versatility, this invention fully considers the impact of complex plateau terrain (such as terrain obstruction and variable climate) on low-altitude flight safety. By embedding a plateau environment feature attention module in the feature extraction stage and introducing a dynamic safety probability threshold in the conflict judgment stage, this method can well adapt to the low-altitude traffic operation needs in complex environments such as plateaus. Furthermore, the core algorithm of this method is highly modular, does not depend on specific detection equipment, is easily integrated with existing low-altitude monitoring systems, and is applicable to various types of non-cooperative targets such as UAVs and unidentified flying objects, demonstrating significant potential for widespread application. Attached Figure Description

[0027] Figure 1 This is a flowchart of the low-altitude non-cooperative target flight conflict early warning method of this application.

[0028] Figure 2 A schematic diagram of the improved trajectory prediction network structure.

[0029] Figure 3 A schematic diagram for determining conflict between non-cooperative targets and cooperative aircraft. Detailed Implementation

[0030] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0031] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0032] Please see Figure 1-3 This paper presents a low-altitude non-cooperative target flight conflict early warning method based on motion trajectory prediction. Addressing the shortcomings of traditional RNN networks, an improved Long Short-Term Memory (LSTM) recurrent neural network is designed. By modifying the input, forget, and output thresholds, the vanishing / expansion gradient problem is solved, enabling it to possess both short-term and long-term trajectory memory capabilities and fully exploit historical trajectory context information. The trajectory state change feature descriptor output by the trajectory matching network is vectorized through an embedding layer and then input into the multi-hidden-layer LSTM recurrent neural network to achieve accurate prediction of non-cooperative target trajectories. The trajectory prediction results are then transformed into a three-dimensional probability set. Based on the three-dimensional probability reachability set, the flight impact range of the non-cooperative target is quantitatively represented to assess the flight conflict risk between the non-cooperative target and the cooperative aircraft. Finally, a dynamic iteration mechanism is constructed through a sliding event window to achieve iterative trajectory prediction and rolling conflict early warning for non-cooperative targets, ensuring the real-time performance and accuracy of the early warning. Figure 1 As shown, the specific steps include: S1 Improved Structure Design of Multi-Hidden-Layer Long Short-Term Memory Recurrent Neural Network (LSTM): This paper proposes technical improvements to address the problems of gradient vanishing / dilation and the inability to learn long-distance trajectory information in traditional RNN networks. Specific improvements and implementation methods include: (1) Modify the three-level gating structure and abandon the traditional single neuron transmission structure of RNN. In the single hidden layer of the multi-hidden layer long short-term memory recurrent neural network LSTM, modify the three-level gating structure of forget gate, input gate and output gate. All three gates adopt the sigmoid activation function and the output value is mapped to the [0,1] interval. The selective transmission of trajectory event sequence information is realized through the "filtering" function of the gating, so as to avoid the information over long distances being infinitely attenuated or amplified during transmission.

[0033] (2) Design a dynamic weight self-loop mechanism for cell state: Set the cell state as the core carrier of trajectory information. Its operation logic is linear addition and gated product, without additional nonlinear transformation, reducing information transmission loss; Combine three-level gating to design a dynamic weight self-loop mechanism for cell state, replacing the traditional fixed weight self-loop method of RNN, and solving the problem of gradient disappearing or expanding rapidly in the self-loop.

[0034] (3) Optimize the complete operation logic inside the single hidden layer: Clarify the input, output and operation relationship of each module in the single hidden layer. The input is the trajectory feature vector at the current time. Hidden layer output at the previous time step The output is the hidden layer output at the current time. The core operation is: Forgotten Gate: Control the proportion of cell state information retained from the previous moment; Input Gate: Generate candidate cell states Control the proportion of new information incorporated; Cell status update: This enables linear transmission of trajectory information across time points; Output gate: Hidden layer output Filter and output effective information about cell status; in, The trajectory feature vector input at the current moment. Output the hidden state from the previous time step. Output the hidden state at the current moment. This represents the cell state at the previous moment. This represents the cell state updated at the current moment. This represents the current state of the candidate cells. The gating activation value for the forget gate. This is the gate activation value for the input gate. This is the gate activation value for the output gate. This is the weight matrix corresponding to the forget gate. This is the weight matrix corresponding to the input gate. The weights corresponding to the candidate cell states. This is the weight matrix corresponding to the output gate. For the bias term corresponding to the forget gate, The bias term corresponding to the input gate. This represents the bias term corresponding to the candidate cell state. For the bias term corresponding to the output gate, The Sigmoid activation function is used to map the gate value to... interval, The hyperbolic tangent activation function is used to generate candidate cell states and hidden layer outputs. For element-wise multiplication, This involves concatenating the hidden layer output from the previous time step with the input vector from the current time step.

[0035] (4) Constructing a multi-hidden-layer long short-term memory recurrent neural network LSTM: using An improved LSTM single hidden layer is linearly connected in series, with the output of the previous hidden layer serving as the input of the next hidden layer. Through deep stacking, the network's ability to learn and fit complex maneuver trajectories of non-cooperative targets is enhanced, adapting to the trajectory characteristics of complex high-altitude environments.

[0036] S2 extracts the trajectory features of non-cooperative targets and performs feature vectorization.

[0037] To address the technical challenges of high noise levels, inconsistent feature dimensions, and insufficient extraction of effective trajectory features in complex high-altitude environments, as well as the standardized requirements of the multi-hidden-layer long short-term memory recurrent neural network (LSTM) after the S1 improvement, this paper proposes multi-dimensional improvements to achieve accurate extraction and standardized processing of trajectory features. This provides high-quality input features for subsequent network training and trajectory prediction. Specific improvements and implementation methods are as follows: (1) Feature extraction optimization of trajectory matching network: The trajectory matching network is improved by lightweighting and adding a plateau environment feature attention module. This module strengthens the weight of trajectory disturbance features corresponding to complex terrain (mountains, canyons) and climate (clouds, strong winds) in the plateau environment, so that the network can accurately capture the trajectory state change feature descriptors of non-cooperative targets in the plateau environment (including position change rate, speed trend, maneuver characteristics, environmental disturbance adaptation features, etc.), discard invalid background features, and improve the correlation between feature descriptors and target flight state.

[0038] (2) Standardization improvement of the embedding layer: Three major functional modules, namely normalization, dimension unification and dimensionality compression, are added to the traditional embedding layer to solve the problems of dimension differences and dimensional redundancy of different detection modalities and different feature indicators: ① Dimensional unification: By using max-min standardization, feature descriptors of different dimensions are mapped to the [0,1] interval, eliminating dimensional interference of indicators such as speed, position, and heading; ② Dimensionality compression: Principal component analysis (PCA) is used to reduce the dimensionality of high-dimensional feature descriptors, retaining effective feature information and reducing the computational complexity of subsequent multi-hidden-layer long short-term memory recurrent neural networks (LSTM). ③ Feature Vector Fixation: The processed feature descriptors are transformed into trajectory feature vectors of fixed length and uniform dimension, forming a sequence of trajectory feature vectors that can be directly recognized by a multi-hidden-layer long short-term memory recurrent neural network (LSTM). .

[0039] (3) Outlier removal in feature preprocessing: After feature vectorization, add... The outlier removal mechanism identifies and removes outliers in the trajectory feature vector sequence, avoiding interference from sudden detection noise in the high-altitude environment on subsequent network training and improving the quality of input features.

[0040] S3 trains a multi-hidden-layer Long Short-Term Memory (LSTM) recurrent neural network to achieve non-cooperative target trajectory prediction, such as... Figure 2 As shown, this demonstrates the overall network structure from the output feature descriptor of the trajectory matching network, through the embedding layer (embedding layer) for feature vectorization, and then input into the n-layer multi-hidden-layer long short-term memory recurrent neural network LSTM, finally outputting the trajectory prediction vector sequence and the trajectory state prediction probability distribution.

[0041] To address the technical challenges of traditional network training samples lacking plateau environment characteristics, slow training algorithm convergence, and poor generalization ability of prediction models, this paper proposes an improved multi-hidden-layer long short-term memory recurrent neural network (LSTM) and extracted standardized trajectory features. Through improvements to the entire training strategy, accurate prediction of non-cooperative target trajectories is achieved. The specific improvements and implementation methods are as follows: (1) Training sample construction: Cut the sample trajectories of all non-cooperative targets into fixed time windows. A sequence of length is used as a training sample. The length is determined based on the time the controller instructs the evacuation (30 seconds is used in this method to meet the advance warning requirement). (2) Network training: Input the trajectory feature vector sequence output by the Embedding layer into the network. In a multi-hidden-layer Long Short-Term Memory Recurrent Neural Network (LSTM) with linear connections between its hidden layers, the error term at each time step is calculated using the Backpropagation in Time (BPTT) algorithm. This error term is then propagated to the next layer, and the gradient of each weight is calculated based on the error term to update the model parameters. However, when the traditional BPTT algorithm handles gradient backpropagation in a multi-hidden-layer LSTM, the gradient decays layer by layer along both the time and network layer dimensions. This decay is particularly pronounced in high-altitude environments where trajectory samples exhibit perturbations and complex features, making it difficult for the network to converge to optimal parameters and accurately learn trajectory features. This invention introduces a gradient clipping mechanism into the BPTT algorithm, setting a gradient threshold (1.0 in this method). When the absolute value of the backpropagation gradient exceeds this threshold, the gradient is scaled proportionally to avoid gradient oscillation caused by excessively large gradients. At the same time, a gradient smoothing factor is introduced to perform a weighted average of the gradients at adjacent time steps (the weight coefficient decays over time, with a weight of 0.8 at the first time step and 0.2 at the second time step), alleviating the gradient decay problem along the time dimension and ensuring that the gradient can be effectively propagated to the shallow layers of the network, allowing the network to fully learn the long-distance trajectory context features.

[0042] (3) Trajectory prediction: After training is completed, the current trajectory prediction will be... The time length is The non-cooperative target trajectory feature vector sequence is input into a multi-hidden-layer long short-term memory recurrent neural network (LSTM). After reading the vector sequence, the network predicts the feature descriptor for the next time step and calculates... Trajectory prediction vector sequence at time step Simultaneously, the hidden layer vectors output by the multi-hidden-layer Long Short-Term Memory Recurrent Neural Network (LSTM) are dimension-mapped, and then obtained through a softmax classifier. The trajectory state prediction probability distribution within a time period is used to determine the final non-cooperative target trajectory prediction result after sorting by probability.

[0043] S4 determines flight conflicts between non-cooperative targets and cooperative aircraft based on three-dimensional probabilistic reachability sets.

[0044] To address the core technical problems of high false alarm / missed alert rates in conflict assessment and the inability to characterize the uncertainty of non-cooperative target trajectories caused by traditional deterministic trajectory prediction, this paper improves the conflict assessment method through probabilistic and quantitative approaches. This enables scientific assessment of flight conflicts and solves the technical problem of high false alarm / missed alert rates in conflict warnings. Furthermore, by incorporating the low-altitude traffic characteristics of complex plateau terrain, the paper enhances the practicality and adaptability of conflict assessment. Specific improvements and implementation methods are as follows: (1) Construction of a three-dimensional probabilistic reachability set based on Zeno polyhedra: Abandoning the traditional deterministic trajectory influence range representation method, and combining the probabilistic trajectory prediction results output by S3, Zeno polyhedra are introduced to quantitatively represent the flight influence range of non-cooperative targets: ① Combine the flight dynamics model of non-cooperative targets (dynamic equations of UAVs / low-altitude aircraft) to set the constraint range of the model control variables (velocity, acceleration, heading) in the plateau environment (adapting to the maneuver restrictions of strong winds and terrain obstruction in the plateau). ② Based on the probability distribution of trajectory prediction, the set of all possible flight states of a non-cooperative target in the next 30 seconds is derived. The set is then transformed into a three-dimensional probabilistic reachable set in three-dimensional space using a Zeno polyhedron, which quantitatively characterizes the flight influence range of the non-cooperative target and accurately reflects the uncertainty of the target trajectory.

[0045] (2) Conflict judgment of cooperative aircraft flight protection zone: In view of the problem that traditional conflict judgment only considers the flight route and ignores the flight protection zone, the planned flight route of the cooperative aircraft + three-dimensional flight protection zone is included in the conflict judgment scope, which is in line with the actual safety regulations of low-altitude traffic operation and avoids the problem of missed reporting due to the intersection of trajectories in the protection zone.

[0046] (3) Set dynamic safety probability thresholds for plateau scenarios: Abandon the traditional fixed conflict judgment thresholds and combine them with the low-altitude traffic characteristics of the complex terrain of the plateau to set dynamic safety probability thresholds (such as a basic threshold of 5%, which is lowered to 3% in high-risk areas such as mountains and canyons, and raised to 8% in open dam areas). The thresholds are dynamically adjusted according to the plateau scenario where the non-cooperative target is located to improve the scientificity and adaptability of conflict judgment.

[0047] (4) Refined analysis of spatial geometric intersection: Using a three-dimensional spatial geometric intersection algorithm, a refined intersection analysis is performed on the three-dimensional probability reachable set of non-cooperative targets and the predetermined flight path and protection zone of cooperative aircraft. The conflict probability of the intersection area is calculated. If the intersection probability is greater than the dynamic safety probability threshold of the current scenario, it is determined that there is a risk of flight conflict.

[0048] S5 uses a sliding time window to implement iterative trajectory prediction and rolling conflict warning, such as Figure 3 As shown, the figure uses a three-dimensional coordinate system to display the flight influence range (probability set, represented by a sphere) of non-cooperative targets and the predetermined flight path of cooperative aircraft, intuitively demonstrating the conflict judgment logic when the two spaces intersect.

[0049] To address the technical issues of existing methods, such as the lack of a dynamic iterative early warning mechanism, insufficient early warning lead time, and inability to adapt to highly maneuverable non-cooperative targets, a sliding time window mechanism is implemented to improve the dynamic, continuous, and real-time rolling conflict early warning for non-cooperative targets. Specific improvements and implementation methods are as follows: (1) Construction of high-frequency sliding time window: Construct a high-frequency sliding time window with a window length of 30 seconds and a sliding step of 1 second. The window length matches the advance warning requirement, and the 1-second step ensures the rapid capture of the real-time motion state of non-cooperative targets, adapting to their high maneuverability characteristics and solving the problems of slow early warning updates and inability to track sudden maneuvers of targets in traditional methods.

[0050] Explanation of the rationality and practicality of the 30-second lead time: Adapted to the motion characteristics of low-altitude non-cooperative targets: The flight speed of low-altitude non-cooperative targets such as drones is usually 10–20 m / s. A 30-second lead time can cover a warning distance of 300–600 m, meeting the full-process implementation needs of controllers for rapid decision-making and emergency avoidance of small aircraft (hovering, small-angle turning, climb / descent), which is in line with the short-time and high-maneuverability control characteristics of the low-altitude field.

[0051] Matching the constraints of the complex environment of the plateau: The terrain of the plateau and mountainous areas is obstructed and the airflow disturbance is strong. Excessive lead time will lead to a surge in trajectory prediction error and an excessive expansion of the probability reachable set, resulting in a large number of false alarms. The 30-second window achieves the optimal balance between prediction accuracy and early warning timeliness, ensuring prediction reliability while reserving sufficient time for implementation.

[0052] The rolling early warning mechanism ensures timeliness: This application adopts a rolling early warning with a 1-second sliding step and updates every second. Even with an initial 30-second lead time, the system will continuously iterate and correct the prediction results, compress the expected time of conflict occurrence in real time, and dynamically supplement the time for avoidance implementation, thus avoiding the problem of insufficient timeliness of a single early warning.

[0053] (2) Real-time rolling update of trajectory data: The trajectory data of non-cooperative targets is collected in real time once every second. The trajectory data within the sliding time window is rolled and updated. The earliest 1-second trajectory data in the window is removed and the latest 1-second trajectory data is incorporated to ensure that the trajectory data in the window always reflects the latest motion state of the target, providing a real-time and accurate data source for iterative trajectory prediction.

[0054] (3) Iterative trajectory prediction: Using the trained multi-hidden-layer long short-term memory recurrent neural network LSTM trajectory prediction model, an iterative trajectory prediction is completed within 1 second to ensure the real-time performance of the iterative prediction, meet the high-frequency update requirements of the sliding window, and realize continuous iterative trajectory prediction for non-cooperative targets.

[0055] (4) Dynamic update mechanism for rolling conflict warning: Based on the updated trajectory data within the window and the iterative prediction results of the lightweight model, the three-dimensional probability reachable set construction and conflict judgment process of S4 are repeated to realize rolling conflict warning for non-cooperative targets: If the conflict probability changes, the conflict risk level is upgraded / downgraded, or the expected time of conflict is advanced / delayed, the warning system will push the updated warning information in real time to ensure dynamic and continuous monitoring of the conflict situation and provide continuous and sufficient decision-making time for conflict avoidance.

[0056] To verify the trajectory prediction effect, a simulation environment with complex terrain such as plateau cities and mountainous areas was built. 120 sets of non-cooperative target trajectory test samples were selected, including maneuver types such as constant speed, variable speed, turning, and climbing. The improved multi-hidden-layer LSTM network of this application was used to predict the trajectory for the next 30 seconds. The accuracy criterion was that the trajectory point position error was ≤1.5m. The test results showed that the number of accurate prediction samples was 104, and the overall trajectory prediction accuracy reached 86.67%, which meets the technical effect of ≥85%. This verifies the trajectory prediction accuracy of the improved LSTM network in plateau environments.

[0057] To verify the accuracy of conflict early warning, this application constructs 100 sets of simulated test samples of flight conflicts between non-cooperative targets and cooperative aircraft (including real conflict samples and near-conflict samples). The three-dimensional probabilistic reachability set + dynamic safety threshold method of this application is used for conflict early warning determination, and the specific definitions are as follows: (1) Real Conflict A real conflict is defined as a situation where, within a 30-second warning window, the three-dimensional probability sets (covering position errors and maneuver uncertainties) of a non-cooperative target and a cooperative aircraft substantially overlap with each other's three-dimensional safety protection zone (set according to low-altitude flight safety regulations).

[0058] (2) Proximity to conflict If the flight trajectories of a non-cooperative target and a cooperative aircraft do not substantially overlap with each other's three-dimensional security protection zones within a 30-second warning window, but the distance is less than the dynamic security threshold (50m for plateau cities / open areas, and 30m for mountainous / canyon areas), it is considered an approach conflict.

[0059] Based on the actual conflict results, the test results are as follows: Test type Sample size Correct number of early warning samples Early warning accuracy Real Conflict 50 48 96% Approaching Conflict 50 49 98% Overall test 100 97 97% The method proposed in this application achieves an overall accuracy of 97% in conflict early warning, meeting the technical requirement of ≥95%, and verifying the accuracy of the conflict judgment method that combines probability reachable sets and dynamic thresholds.

[0060] Compared with the prior art, the beneficial effects of the present invention are reflected in the following aspects: (1) Solved the gradient problem of traditional RNN network and improved trajectory prediction accuracy: By improving the design of multi-hidden-layer long short-term memory recurrent neural network LSTM network and modifying the three thresholds of input, forgetting and output, the gradient vanishing / expansion problem when RNN processes trajectory time series is solved, so that the network has both short-term and long-term trajectory memory capabilities, can fully explore the contextual information of historical trajectories, and greatly improve the trajectory prediction accuracy of non-cooperative targets. The trajectory prediction accuracy within the next 30 seconds is ≥85%.

[0061] (2) Reduced the false alarm rate and missed alarm rate of conflict warning: The trajectory prediction results are transformed into a probability set in three-dimensional space. The flight impact range of non-cooperative targets is represented by the three-dimensional probability reachable set, instead of using the traditional deterministic trajectory prediction results for conflict judgment. This is more in line with the characteristics of non-cooperative targets being highly maneuverable and having random motion intentions, effectively reducing the false alarm rate and missed alarm rate of flight conflict warning. The accuracy rate of conflict warning is ≥95%.

[0062] (3) It realizes dynamic early warning of highly maneuverable targets and ensures sufficient decision-making time: The iterative trajectory prediction and rolling conflict early warning mechanism with sliding time window can dynamically update the trajectory prediction results and conflict judgment conclusions according to the real-time motion status of non-cooperative targets. The early warning lead time is ≥30 seconds, providing controllers and cooperative aircraft with sufficient conflict avoidance decision-making time.

[0063] (4) Adaptable to the low-altitude environment of complex plateau terrain, with strong practicality: This method does not rely on single detection mode data, but carries out prediction and early warning based on trajectory data estimated by non-cooperative target motion state. It can adapt to the low-altitude environment of complex plateau terrain with terrain obstruction and variable climate, and meet the actual needs of low-altitude traffic safety control in this scenario.

[0064] (5) Strong versatility and easy integration and promotion: The core algorithm module of this method can be independently packaged and is easy to integrate with existing low-altitude monitoring systems and multimodal fusion detection systems. It can also be adapted to different types of non-cooperative targets such as UAVs and unidentified flying objects. It has a wide range of application value in low-altitude traffic management in complex terrains such as plateaus, mountains and canyons.

[0065] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

Claims

1. A method for early warning of low-altitude non-cooperative target flight conflicts based on motion trajectory prediction, characterized in that, Includes the following steps: Acquire real-time trajectory data of non-cooperative targets, and perform feature extraction and vectorization on the real-time trajectory data to obtain a sequence of trajectory feature vectors; The trajectory feature vector sequence is input into a pre-trained multi-hidden-layer long short-term memory recurrent neural network (LSTM). The multi-hidden-layer long short-term memory recurrent neural network (LSTM) outputs the trajectory prediction vector sequence of the non-cooperative target in a future preset time period and the corresponding trajectory state prediction probability distribution. Based on the trajectory state prediction probability distribution and combined with the flight dynamics model of non-cooperative targets, a three-dimensional probability reachability set characterizing the flight influence range of non-cooperative targets is constructed. Obtain the planned flight path and three-dimensional flight protection zone of the cooperative aircraft. Perform spatial geometric intersection analysis on the three-dimensional probability reachable set and the three-dimensional flight protection zone. If the intersection probability is greater than the preset safety probability threshold, it is determined that there is a risk of flight conflict. By using a preset sliding time window to continuously update the trajectory data of non-cooperative targets, and iteratively performing trajectory prediction and conflict judgment based on the updated trajectory data, rolling conflict warnings for non-cooperative targets are achieved.

2. The method for early warning of low-altitude non-cooperative target flight conflict based on motion trajectory prediction according to claim 1, characterized in that, Each LSTM unit in the multi-hidden-layer Long Short-Term Memory Recurrent Neural Network (LSTM) includes a three-level gating structure of forget gate, input gate, and output gate, as well as a cell state with a dynamic weight self-looping mechanism; the internal operation logic of the LSTM unit is as follows: Forgotten Gate: Control the proportion of cell state information retained from the previous moment; Input Gate: Generate candidate cell states Control the proportion of new information incorporated; Cell status update: This enables linear transmission of trajectory information across time points; Output gate: Hidden layer output Filter and output effective information about cell status; in, The trajectory feature vector input at the current moment. Output the hidden state from the previous time step. Output the hidden state at the current moment. This represents the cell state at the previous moment. This represents the cell state updated at the current moment. This represents the current state of the candidate cells. The gating activation value for the forget gate. This is the gate activation value for the input gate. This is the gate activation value for the output gate. This is the weight matrix corresponding to the forget gate. This is the weight matrix corresponding to the input gate. The weights corresponding to the candidate cell states. This is the weight matrix corresponding to the output gate. For the bias term corresponding to the forget gate, The bias term corresponding to the input gate. This represents the bias term corresponding to the candidate cell state. For the bias term corresponding to the output gate, The Sigmoid activation function is used to map the gate value to... interval, The hyperbolic tangent activation function is used to generate candidate cell states and hidden layer outputs. For element-wise multiplication, This involves concatenating the hidden layer output from the previous time step with the input vector from the current time step.

3. The method for early warning of low-altitude non-cooperative target flight conflict based on motion trajectory prediction according to claim 2, characterized in that, During training, the multi-hidden-layer long short-term memory recurrent neural network (LSTM) employs a backpropagation time-based (BPTT) algorithm that combines gradient clipping and gradient smoothing mechanisms for parameter updates. Gradient clipping is used to scale gradients exceeding a preset threshold proportionally. The gradient smoothing mechanism mitigates gradient decay by weighted averaging of gradients from adjacent time points, where the time point closest to the current time is designated as the first time point, and the time point furthest from the current time point is designated as the second time point. The gradient weight coefficient of the first time point is greater than that of the second time point.

4. The method for early warning of low-altitude non-cooperative target flight conflict based on motion trajectory prediction according to claim 3, characterized in that, The preset threshold for gradient clipping is 1.0; the weight coefficient for gradient smoothing decays over time, with a weight of 0.8 at the first time step and 0.2 at the second time step.

5. The method for early warning of low-altitude non-cooperative target flight conflict based on motion trajectory prediction according to claim 1, characterized in that, The three-dimensional probabilistic reachable set is constructed based on the Zeno polyhedron, specifically including: setting the dynamic constraint range of the control variables in the plateau environment according to the flight dynamics model of the non-cooperative target; deriving the set of all possible flight states of the non-cooperative target in the future time period according to the trajectory state prediction probability distribution, and transforming it into a three-dimensional probabilistic reachable set represented by the Zeno polyhedron.

6. The method for early warning of low-altitude non-cooperative target flight conflict based on motion trajectory prediction according to claim 5, characterized in that, The spatial geometric intersection analysis includes: using a three-dimensional spatial geometric intersection algorithm to calculate the intersection region between the three-dimensional probabilistic reachable set and the three-dimensional flight protection zone of the cooperative aircraft, and determining the conflict probability of the intersection region based on the probability distribution of the three-dimensional probabilistic reachable set.

7. The method for early warning of low-altitude non-cooperative target flight conflict based on motion trajectory prediction according to claim 1, characterized in that, The preset safety probability threshold is dynamically adjusted and set according to the geographical scene where the non-cooperative target is currently located: a first threshold is used in open areas, and a second threshold lower than the first threshold is used in high-risk areas such as mountains or canyons.

8. A method for early warning of low-altitude non-cooperative target flight conflict based on motion trajectory prediction according to claim 7, characterized in that, The first threshold is 8%, and the second threshold is 3%.

9. A method for early warning of low-altitude non-cooperative target flight conflict based on motion trajectory prediction according to claim 1, characterized in that, The length of the sliding time window is equal to a future preset time period, and the sliding step size is less than the window length. When updating the trajectory data through the sliding time window, the trajectory data at the earliest moment in the window is removed, and the latest collected trajectory data is incorporated to ensure that the data in the window always reflects the latest motion state of the non-cooperative target.

10. A method for early warning of low-altitude non-cooperative target flight conflict based on motion trajectory prediction according to claim 1, characterized in that, The preset time period is 30 seconds, the window length of the sliding time window is 30 seconds, and the sliding step size is 1 second; the update frequency of the rolling conflict warning is once per second.

Images