Low-altitude aircraft noise map generation and route optimization method and system, and storage medium
By deeply integrating physical acoustics and data-driven models, a high-precision real-time noise map is generated and a rolling silent path planning is performed, which solves the shortcomings of existing noise prediction and path planning technologies and realizes high-precision and dynamic adaptive noise assessment and path optimization in complex environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA AERO GEOPHYSICAL SURVEY & REMOTE SENSING CENT FOR LAND & RESOURCES
- Filing Date
- 2026-02-12
- Publication Date
- 2026-06-02
Smart Images
Figure CN122135597A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent management of low-altitude airspace, aircraft noise prediction and route planning, and specifically to a method, system and storage medium for generating noise maps and optimizing routes for low-altitude aircraft. Background Technology
[0002] With the rapid development of urban air mobility (UAM), drone logistics, and general aviation, low-altitude aircraft activity is becoming increasingly frequent, and the noise they generate is having a growing impact on urban residents' lives, wildlife conservation, and acoustically sensitive areas. Accurate and real-time prediction of aircraft noise and the planning of quiet flight paths based on this prediction are crucial for achieving green and sustainable operation of low-altitude transportation.
[0003] Existing technologies have the following main shortcomings: In terms of noise prediction, traditional methods mostly rely on simplified physical models or static empirical formulas, which are difficult to adapt to complex and ever-changing real-time meteorological environments, terrain features, and dynamic operating conditions of aircraft, resulting in limited prediction accuracy; In terms of path optimization, existing route planning focuses more on the shortest path or the shortest time, rarely taking noise pollution as a core optimization objective, or only using simple rules such as static noise buffers, which cannot achieve dynamic and refined quiet path generation; In terms of data fusion, it fails to deeply integrate aircraft status, high-precision environmental data, and physical acoustic models, and lacks accurate modeling of the spatiotemporal propagation characteristics of noise, resulting in poor real-time performance and low resolution of the generated noise maps, making it difficult to support real-time online path replanning.
[0004] Therefore, there is an urgent need for an intelligent method that can deeply integrate physical acoustic principles and data-driven models to achieve high-precision, real-time noise map generation, and on this basis, use deep reinforcement learning algorithms for online, rolling silent path planning, in order to solve the shortcomings of traditional methods in terms of dynamic environment adaptability, prediction accuracy, and proactive optimization. Summary of the Invention
[0005] The purpose of this invention is to provide a method, system, and storage medium for generating noise maps and optimizing routes for low-altitude aircraft, in order to solve the technical problems in the prior art, such as inaccurate noise assessment and limited quiet route planning effect caused by insufficient accuracy of noise prediction models, insufficient consideration of dynamic noise impact in route planning, and lack of deep fusion of multi-source information and real-time closed-loop optimization capabilities of the system.
[0006] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:
[0007] A method for generating noise maps and optimizing routes for low-altitude aircraft includes the following steps:
[0008] Step 1: Collect real-time data from multiple sources, including aircraft status data, environmental characteristic data, and basic acoustic data;
[0009] Step 2: Rasterize and standardize the environmental feature data to construct an environmental feature tensor;
[0010] Step 3: Construct and train a real-time noise map generation and path quieting optimization model, which includes an environmental feature encoding network, a physical-data hybrid noise prediction network, and a quieting path optimization network;
[0011] Step 4: Input real-time data into the model, extract deep environmental features through the environmental feature encoding network; calculate a real-time noise map centered on the aircraft's current position through the physical-data hybrid noise prediction network, combining the physical acoustic model and neural network correction terms.
[0012] Step 5: Based on the current noise map, aircraft status, and target point, generate a sequence of quiet optimized flight maneuvers in the future time domain through the quiet path optimization network;
[0013] Step 6: Execute the first action in the action sequence, update the aircraft state, and return to step 4 to perform the next rolling optimization until the target point is reached.
[0014] In a preferred embodiment of the present invention, in step 1: the aircraft status data includes position. ,speed Engine parameters and model parameters ;
[0015] The environmental characteristic data includes atmospheric parameters and a three-dimensional wind speed field. Atmospheric parameters include air density. ,temperature ,humidity ;
[0016] The acoustic baseline data is the reference sound pressure level. .
[0017] As a preferred embodiment of the present invention, the environmental feature encoding network in step 3 adopts a three-dimensional convolution and attention mechanism, and its processing procedure is as follows:
[0018] Input: Environmental feature tensor ,in, The number of horizontal grid lines. The number of vertical layers. The number of feature channels;
[0019] Processing: Spatial features are extracted through multi-layer 3D convolution, and key region features are enhanced through an attention module;
[0020] Output: Deep environmental features .
[0021] As a preferred embodiment of the present invention, the physical-data hybrid noise prediction network in step 3 includes a physical basis model and a neural network correction module:
[0022] Physical basis model for calculating reference sound pressure level :
[0023] ;
[0024] in, For transmission distance, For the observation direction angle, For the speed of sound, For the characteristic area, For reference sound intensity, For directional correction, Atmospheric absorption and attenuation;
[0025] The neural network correction module uses environmental features Aircraft status and the calculated Input, output correction value ;
[0026] Final predicted sound pressure level ;
[0027] Real-time noise maps calculate all grid points within the coverage area. of The value is calculated and accumulated exponentially over time to obtain the equivalent sound level. And generated.
[0028] As a preferred embodiment of the present invention, the silent path optimization network in step 3 employs a near-end strategy optimization deep reinforcement learning algorithm, the elements of which include:
[0029] state Includes aircraft status. Flattened noise map Deep environmental characteristics and target location ;
[0030] action Defined as incremental control commands for aircraft speed, heading, and thrust. ;
[0031] reward function :
[0032] ;
[0033] in, As a reward for moving towards the goal, As a punishment for excessive noise in sensitive areas of the ground, For smoothness of motion reward, To terminate the reward, These are the weighting coefficients.
[0034] In a preferred embodiment of the present invention, in step 5, the silent path optimization network generates the action in the following manner:
[0035] ;
[0036] in, For policy networks, For network parameters, For exploration, random noise is used; the network maximizes the expected cumulative reward. Training was conducted, among which For trajectory.
[0037] As a preferred embodiment of the present invention, it further includes a model training step, employing a combined loss function:
[0038] ;
[0039] in, This is the mean square error loss for noise prediction. To enhance the gradient loss of the learning strategy, To balance the hyperparameters.
[0040] As a preferred embodiment of the present invention, the present invention provides a low-altitude aircraft noise map generation and route optimization system, applied to a low-altitude aircraft noise map generation and route optimization method, comprising:
[0041] Data acquisition and preprocessing unit: used to acquire aircraft status, environmental and acoustic data in real time, and to perform standardization and rasterization processing;
[0042] Noise map generation unit: Built-in physical-data hybrid noise prediction network, used to generate high-resolution noise maps based on real-time data;
[0043] Silent Path Optimization Unit: Built-in silent path optimization network, used to calculate silent optimization flight commands based on the current state and noise map;
[0044] Rolling Execution and Feedback Unit: Used to execute optimization instructions, update system status, and achieve closed-loop rolling optimization.
[0045] As a preferred embodiment of the present invention, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method such as low-altitude aircraft noise map generation and route optimization.
[0046] Compared with the prior art, the present invention has the following advantages:
[0047] This invention achieves high-precision, real-time noise distribution prediction by deeply integrating physical and data-driven models. It ensures the correctness of acoustic principles and uses neural networks to compensate for complex environmental factors. Using a noise map as the core input, it plans paths online through deep reinforcement learning to actively avoid noise-sensitive areas and time periods, thus achieving a leap from passive assessment to active optimization. It effectively integrates multimodal data such as aircraft dynamic parameters, high-precision meteorological data, and terrain data, making noise prediction and path planning closer to real complex environments. It realizes a complete automated closed loop from noise perception and map generation to path optimization and command execution, providing an end-to-end solution for low-altitude quiet flight. Attached Figure Description
[0048] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.
[0049] Figure 1 This is a flowchart of a method for generating noise maps and optimizing flight routes for low-altitude aircraft, provided in an embodiment of the present invention.
[0050] Figure 2 This is a structural diagram of the noise map real-time generation and path quieting optimization model provided in an embodiment of the present invention;
[0051] Figure 3 A block diagram of a low-altitude aircraft noise map generation and route optimization system provided in an embodiment of the present invention. Detailed Implementation
[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0053] like Figure 1 As shown, this invention provides a method for generating noise maps and optimizing routes for low-altitude aircraft, comprising the following steps:
[0054] Step 1: Collect real-time data from multiple sources, including aircraft status data, environmental characteristic data, and basic acoustic data;
[0055] Step 2: Rasterize and standardize the environmental feature data to construct an environmental feature tensor;
[0056] Step 3: Construct and train a real-time noise map generation and path quieting optimization model. The model includes an environmental feature encoding network, a physical-data hybrid noise prediction network, and a quieting path optimization network.
[0057] Step 4: Input real-time data into the model, extract deep environmental features through the environmental feature encoding network; calculate a real-time noise map centered on the aircraft's current position through the physical-data hybrid noise prediction network, combining the physical acoustic model with neural network correction terms;
[0058] Step 5: Based on the current noise map, aircraft status, and target point, generate a sequence of quiet optimized flight maneuvers for the future time domain through a quiet path optimization network;
[0059] Step 6: Execute the first action in the action sequence, update the aircraft state, and return to step 4 to perform the next rolling optimization until the target point is reached.
[0060] In step 1: the aircraft status data includes position. ,speed Engine parameters and model parameters ;
[0061] Environmental characteristic data include atmospheric parameters and three-dimensional wind speed fields. Atmospheric parameters include air density. ,temperature ,humidity ;
[0062] The basic acoustic data is the reference sound pressure level. .
[0063] The environmental feature encoding network in step 3 employs a three-dimensional convolution and attention mechanism, and its processing procedure is as follows:
[0064] Input: Environmental feature tensor ,in, The number of horizontal grid lines. The number of vertical layers. The number of feature channels;
[0065] Processing: Spatial features are extracted through multi-layer 3D convolution, and key region features are enhanced through an attention module;
[0066] Output: Deep environmental features .
[0067] The physics-data hybrid noise prediction network in step 3 includes a physics-based model and a neural network correction module:
[0068] Physical basis model for calculating reference sound pressure level :
[0069] ;
[0070] in, For transmission distance, For the observation direction angle, For the speed of sound, For the characteristic area, For reference sound intensity, For directional correction, Atmospheric absorption and attenuation;
[0071] The neural network correction module uses environmental features Aircraft status and the calculated Input, output correction value ;
[0072] Final predicted sound pressure level ;
[0073] Real-time noise maps calculate all grid points within the coverage area. of The value is calculated and accumulated exponentially over time to obtain the equivalent sound level. And generated.
[0074] The silent path optimization network in step 3 employs a proximal strategy optimization deep reinforcement learning algorithm, the elements of which include:
[0075] state Includes aircraft status. Flattened noise map Deep environmental characteristics and target location ;
[0076] action Defined as incremental control commands for aircraft speed, heading, and thrust. ;
[0077] reward function :
[0078] ;
[0079] in, As a reward for moving towards the goal, As a punishment for excessive noise in sensitive areas of the ground, For smoothness of motion reward, To terminate the reward, These are the weighting coefficients.
[0080] In step 5, the silent path optimization network generates actions in the following way:
[0081] ;
[0082] in, For policy networks, For network parameters, For exploration, random noise is used; the network maximizes the expected cumulative reward. Training was conducted, among which For trajectory.
[0083] It also includes a model training step, using a combined loss function:
[0084] ;
[0085] in, This is the mean square error loss for noise prediction. To enhance the gradient loss of the learning strategy, To balance the hyperparameters.
[0086] This invention provides an embodiment including the following steps:
[0087] S101. Real-time acquisition and preprocessing of multi-source data.
[0088] The aircraft uses onboard sensors and downlink to... Real-time reporting of status data at a high frequency Simultaneously, gridded analytical fields and short-term forecast data covering the flight area were obtained from meteorological services, including temperature. ,humidity Air pressure (used to calculate air density) ) and three-dimensional wind speed Digital elevation model (DEM) data is acquired from a geographic information system. All environmental data is then uniformly interpolated to a horizontal range centered on the aircraft. Vertical layering On the three-dimensional mesh of the layer, the initial environment feature tensor is formed. (8 feature channels). All data were standardized. ,in represents the mean and standard deviation of the historical data for each feature.
[0089] S102. Deep coding of environmental features.
[0090] Standardized environmental feature tensor Input environment feature encoding network. This network consists of two 3D convolutional layers (3×3×3 kernels, with the number of channels increasing from 8 to 32) and a 3D spatial attention module. The attention module calculation process is as follows:
[0091] ;
[0092] ;
[0093] ;
[0094] in, This is an intermediate feature map. The key vector dimension is used. The network ultimately outputs deep environmental features. It encapsulates environmental information that has a critical impact on noise propagation.
[0095] The environment feature encoding network is an encoder that deeply integrates 3D convolution and attention mechanisms. Its specific architecture and data flow are as follows:
[0096] 1. Input layer: Receives the standardized environmental feature tensor. .
[0097] 2. Feature extraction backbone: contains two consecutive 3D convolutional layers.
[0098] First convolutional layer: Convolutional kernel size is 3×3×3, input channels are 8, output channels are 16, stride is 1, padding is 1, followed by ReLU activation function.
[0099] The second convolutional layer has a 3×3×3 kernel, 16 input channels, 32 output channels, a stride of 2, padding of 1, and is followed by a ReLU activation function. This step is used to initially extract multi-level spatial environment features and output a feature map. .
[0100] 3. Three-dimensional spatial attention module:
[0101] Objective: To enable the network to focus on environmental areas that significantly affect sound wave propagation (such as temperature inversion layers and densely built-up areas).
[0102] Implementation: Perform three independent 1×1×1 convolution operations to generate three tensors: query (Q), key (K), and value (V).
[0103] Calculation: Attention Weights ,in Let K be the channel dimension, set to 32.
[0104] Residual connection: In order to preserve the original feature information.
[0105] 4. Output layer: After a final 1×1×1 convolutional layer, the number of channels is consolidated to 64, outputting deep environmental features. The output encodes high-dimensional environmental features containing information about key acoustically affected areas, providing structured input for subsequent noise prediction.
[0106] S103. Real-time generation of physical-data mixed noise maps.
[0107] For ground evaluation grid points (Assuming the ground height is 0):
[0108] Calculate the propagation vector: ,distance Direction angle .
[0109] Physical model calculations: based on the aircraft model and engine parameters Obtain the reference sound power from the table. Substitute it into the formula to calculate. .
[0110] Neural Network Correction: Constructing Feature Vectors The input is fed into a multilayer perceptron (MLP), and the output is a correction value. .in It is obtained by bilinear interpolation from The environmental characteristics of that point were obtained from the data.
[0111] Synthesis and accumulation: Instantaneous sound pressure level at this point To reflect the cumulative effect of noise, the equivalent sound level is calculated. :
[0112] ;
[0113] in, The attenuation factor represents the exponential decay of sound energy. A real-time noise map is generated by traversing all ground grid points. .
[0114] The neural network correction module is a five-layer fully connected network (MLP), and its structure and function are as follows:
[0115] Input layer: receives feature vectors . The 64-dimensional deep environmental feature vector corresponding to the ground grid point is obtained from $E_{feat}$ through trilinear interpolation, which ensures the accurate alignment of spatial features.
[0116] Hidden layers: There are three layers with 256, 128 and 64 neurons respectively. All layers use the ReLU activation function and introduce Dropout (dropout rate of 0.1) to prevent overfitting.
[0117] Output layer: Linear layer, outputs a scalar value. , representing the correction value of the sound pressure level predicted by the physical model at that point.
[0118] This MLP learns complex nonlinear effects not covered by the physical model, such as diffraction from irregular terrain and acoustic refraction distortion caused by non-uniform wind fields. Its input features... Integrating geometry ( ),dynamics( ) and environment ( The information fully depicts the relationship between the sound source, the propagation path, and the receiving point.
[0119] This module requires pre-training before joint training with the underlying physical model. The pre-training dataset consists of historical flight data and synchronous ground noise monitoring data. The loss function is the mean square error (MSE) between the predicted sound pressure level Lp and the measured value: noise loss. N is the total number of samples. Let be the predicted value for the i-th sample. Let be the true value of the i-th sample. The optimizer uses Adam, with an initial learning rate of 1e-3 and a learning rate decay strategy.
[0120] S104. Optimization of silent path scrolling based on deep reinforcement learning.
[0121] The Silent Path Optimization Network (PON) employs the PPO algorithm framework. (Policy Network) and value network All are MLPs.
[0122] State construction: Noise map Flattened, in relation to the aircraft's state Target relative position and deep environmental characteristics The global average pooling vectors are concatenated to form the state vector. .
[0123] Action generation and execution: at each decision moment The policy network is based on the current state. Output action distribution mean and variance Sampling obtains action .implement Afterwards, the aircraft entered a new state. .
[0124] Reward Calculation: The reward function is designed as follows:
[0125] ;
[0126] Total Rewards .
[0127] Online learning: The system continuously collects transfer tuples during operation. Store in the experience replay pool. Periodically sample small batches of data to update network parameters. To maximize accumulated rewards.
[0128] The Silent Path Optimization Network is implemented based on the Proximal Policy Optimization (PPO) algorithm, and its core components include:
[0129] Policy Network A four-layer MLP (input layer → 128-dimensional hidden layer (ReLU) → 64-dimensional hidden layer (ReLU) → output layer). The input is a state vector. The output is an action. The mean of each component and logarithmic standard deviation (For exploration purposes).
[0130] Value Network A three-layer MLP (input layer → 64-dimensional hidden layer (ReLU) → output layer). The input is... The output is a state value estimate.
[0131] State construction concretization: Flattened noise map This refers to expanding a two-dimensional noise map matrix into a one-dimensional vector in row-major order. time t , A vector operator that stretches a 2D / high-dimensional matrix into a 1D column vector using column-major order (default). Deep environmental features. The input is a 64-dimensional vector obtained by global average pooling, which provides a global context of the environment and helps the policy network understand the acoustic properties of the region.
[0132] The basis for determining the weights of the reward function: weight coefficients. , , This was determined through multi-objective optimization simulation experiments. For example, while ensuring mission success, the effect of adjusting weights on peak noise (L) was observed. eq The maximum value of the noise reduction factor (NRF) and the influence of flight time are considered, and finally a set of balance values (e.g., "noise reduction benefit / time increment") that optimizes the overall utility is selected. , , This demonstrates the direct connection between the design of the reward function and the technical goal of "quiet and efficient flight".
[0133] S105. Model training process.
[0134] The training is divided into two phases:
[0135] Noise prediction network pre-training: using historical flight data and synchronous ground noise monitoring data, with mean squared error... The noise prediction network is trained separately, using the loss function.
[0136] Joint fine-tuning and path optimization training: In the simulator, a pre-trained noise prediction network is fixed, and a silent path optimization network is trained. This stage employs a combined loss mechanism. ,in The clipped surrogate objective loss of the PPO algorithm, i.e. Joint training enables the path optimization network to learn to predict the network's output using noise.
[0137] Online training was conducted using an actor-critic framework. In a simulated environment, the agent interacts and collects trajectory data based on its current policy, storing this data in an experience replay pool. After accumulating a certain amount of data, the pruning objective function of PPO was applied. Update the policy network and the value network using mean squared error loss. Training continues until the policy converges.
[0138] The core innovation of the PPO algorithm is the introduction of a policy objective function with "Clip," designed to stabilize training and improve data efficiency. Its standard form is as follows:
[0139] ;
[0140] Parameters of the policy network.
[0141] E t : Indicates the expected value (average value) of the current batch of sample data.
[0142] r t(θ): Importance sampling ratio, which measures the difference in probability between the new strategy and the old strategy in taking the same action.
[0143] A t : In state s t Next, action a t The advantage function estimate (calculated by the value network). It represents how "good" or "bad" the action is relative to the average level.
[0144] A very small hyperparameter (e.g., 0.1 or 0.2) is used to define the clipping interval. .
[0145] This is the basic policy gradient objective. If A^t > 0 (good action), we increase... (That is, make the new strategy more inclined to take this action); conversely, reduce it.
[0146] This is the cropped target. It will have an importance sampling ratio. Limited to Within the range.
[0147] The min operation ultimately selects the smaller of the two objectives. Its brilliance lies in:
[0148] When the movement is good ( >0) and When attempting to become very large (i.e., the new strategy tries to over-exploit this good move), the pruning item will cap its upper limit. This is to prevent the strategy update from taking too big a step.
[0149] When the movement is very poor ( <0) and When the value becomes very small (i.e., the new strategy tries to avoid this bad action), the clipping item will limit its lower bound to [a specific value]. This prevents the strategy from deviating excessively.
[0150] To quantitatively evaluate the effectiveness of the method of this invention, we set up a comparative experiment in a simulated low-altitude logistics scenario in a suburban city. The test route was 5 kilometers long, passing through two types of noise-sensitive areas: residential areas (highly sensitive) and parks (moderately sensitive). The meteorological conditions included both typical calm days and days with sudden crosswinds. This invention compared three methods:
[0151] Baseline A (traditional shortest path method): The A* algorithm path with the sole objective of minimizing flight time.
[0152] Baseline B (Static Noise Avoidance Method): During the route planning phase, the RRT* algorithm path is used to bypass the noise-sensitive area buffer zone with a fixed radius (300 meters) in advance based on historical noise data.
[0153] This method (dynamic silent path optimization) employs the real-time noise map generation and reinforcement learning path optimization system of this invention.
[0154] First, a comparison of overall path planning performance:
[0155] We conducted Monte Carlo simulations (100 times each) on the three methods under two typical meteorological conditions, and the statistical results of the key performance indicators are shown in Table 1 below.
[0156]
[0157] As shown in the table above, the present invention can achieve optimal noise reduction performance under both meteorological conditions, and in particular, it can control the peak noise within the ideal range, which proves the accuracy of the physical-data hybrid noise prediction model and the effectiveness of the reinforcement learning optimizer active noise reduction strategy.
[0158] This invention achieves a significant noise reduction (approximately 10 dB) in exchange for an acceptable increase in flight time and energy consumption (approximately 8-9%), and its overall utility (noise reduction gain per unit time / energy consumption) is far greater than that of the two baseline methods.
[0159] Furthermore, in the "sudden crosswind day" scenario, the performance index fluctuation range (standard deviation) of this invention is the smallest, and the task success rate is the highest. This proves that the closed-loop mechanism of rolling optimization using a real-time noise map has a strong adaptability to environmental disturbances, while static methods perform poorly in this scenario.
[0160] The above comparison results fully demonstrate that the present invention achieves more accurate sound field modeling through a physical-data hybrid noise prediction network and achieves dynamic response and active avoidance of noise maps through a silent path optimization network based on deep reinforcement learning. The synergistic effect of the two forms a rolling optimization closed loop, thus maintaining excellent noise reduction performance and task robustness under complex weather conditions.
[0161] The second method is to verify the accuracy of the noise prediction model:
[0162] To separately evaluate the performance of the core component of this invention—the physics-data hybrid noise prediction network—we compared it with a pure physics model and a pure data-driven model (DNN) on an independent test set. The test set contained 1000 sets of noise monitoring data under different flight conditions, weather, and terrain combinations, and the results are shown in Table 2.
[0163]
[0164] As shown in Table 2 above, the physical-data hybrid noise prediction network of this invention achieves optimal results in terms of MAE, RMSE, and R². This confirms the effectiveness of the architecture that combines physical acoustic priors with data-driven correction: it overcomes the shortcomings of pure physical models in complex environments and avoids the physical unreliability predictions that may occur with pure black-box data models, providing a core guarantee for achieving high-precision and high-reliability real-time noise maps.
[0165] Therefore, through the above systematic quantitative comparative analysis, it is fully demonstrated that the "Low-Altitude Aircraft Noise Map Generation and Route Optimization Method" provided by this invention is significantly superior to traditional static planning methods in terms of noise prediction accuracy, quiet path planning effect, system dynamic adaptability, and overall mission efficiency. This invention provides an effective technical solution for effectively reducing low-altitude flight noise pollution and achieving green and intelligent low-altitude traffic management.
[0166] The multi-source data collected and processed in this invention, such as aircraft status, publicly available meteorological grid data, and digital elevation models, are all technical data necessary for achieving noise prediction and path optimization. Their sources are legal and compliant, and do not involve the acquisition or use of illegal or unauthorized data such as personal privacy or trade secrets. They comply with relevant laws, regulations, and ethical norms.
[0167] like Figure 2 As shown, this invention provides a low-altitude aircraft noise map generation and route optimization system, applied to a low-altitude aircraft noise map generation and route optimization method, comprising:
[0168] Data acquisition and preprocessing unit: used to acquire aircraft status, environmental and acoustic data in real time, and to perform standardization and rasterization processing;
[0169] Noise map generation unit: Built-in physical-data hybrid noise prediction network for generating high-resolution noise maps based on real-time data;
[0170] Silent Path Optimization Unit: Built-in silent path optimization network, used to calculate silent optimization flight commands based on the current state and noise map;
[0171] Rolling Execution and Feedback Unit: Used to execute optimization instructions, update system status, and achieve closed-loop rolling optimization.
[0172] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method such as low-altitude aircraft noise map generation and route optimization.
[0173] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.
Claims
1. A method for generating noise maps and optimizing routes for low-altitude aircraft, characterized in that, Includes the following steps: Step 1: Collect real-time data from multiple sources, including aircraft status data, environmental characteristic data, and basic acoustic data; Step 2: Rasterize and standardize the environmental feature data to construct an environmental feature tensor; Step 3: Construct and train a real-time noise map generation and path quieting optimization model, which includes an environmental feature encoding network, a physical-data hybrid noise prediction network, and a quieting path optimization network; Step 4: Input real-time data into the model, extract deep environmental features through the environmental feature encoding network; calculate a real-time noise map centered on the aircraft's current position through the physical-data hybrid noise prediction network, combining the physical acoustic model and neural network correction terms. Step 5: Based on the current noise map, aircraft status, and target point, generate a sequence of quiet optimized flight maneuvers in the future time domain through the quiet path optimization network; Step 6: Execute the first action in the action sequence, update the aircraft state, and return to step 4 to perform the next rolling optimization until the target point is reached.
2. The method according to claim 1, characterized in that, In step 1: the aircraft status data includes position. ,speed Engine parameters and model parameters ; The environmental characteristic data includes atmospheric parameters and a three-dimensional wind speed field. Atmospheric parameters include air density. ,temperature ,humidity ; The acoustic baseline data is the reference sound pressure level. .
3. The method according to claim 2, characterized in that, The environmental feature encoding network in step 3 employs a three-dimensional convolution and attention mechanism, and its processing procedure is as follows: Input: Environmental feature tensor ,in, The number of horizontal grid lines. The number of vertical layers. The number of feature channels; Processing: Spatial features are extracted through multi-layer 3D convolution, and key region features are enhanced through an attention module; Output: Deep environmental features .
4. The method according to claim 3, characterized in that, The physics-data hybrid noise prediction network in step 3 includes a physics-based model and a neural network correction module: Physical basis model for calculating reference sound pressure level : ； in, For transmission distance, For the observation direction angle, For the speed of sound, For the characteristic area, For reference sound intensity, For directional correction, Atmospheric absorption and attenuation; The neural network correction module uses environmental features Aircraft status and the calculated Input, output correction value ; Final predicted sound pressure level ; Real-time noise maps calculate all grid points within the coverage area. of The value is calculated and accumulated exponentially over time to obtain the equivalent sound level. And generated.
5. The method according to claim 4, characterized in that, The silent path optimization network in step 3 employs a near-end strategy optimization deep reinforcement learning algorithm, the elements of which include: state Includes aircraft status. Flattened noise map Deep environmental characteristics and target location ; action Defined as incremental control commands for aircraft speed, heading, and thrust. ; reward function : ； in, As a reward for moving towards the goal, As a punishment for excessive noise in sensitive areas of the ground, For smoothness of motion reward, To terminate the reward, These are the weighting coefficients.
6. The method according to claim 5, characterized in that, In step 5, the silent path optimization network generates actions in the following manner: ； in, For policy networks, For network parameters, For exploration, random noise is used; the network maximizes the expected cumulative reward. Training was conducted, among which For trajectory.
7. The method according to claim 6, characterized in that, It also includes a model training step, using a combined loss function: ； in, This is the mean square error loss for noise prediction. To enhance the gradient loss of the learning strategy, To balance the hyperparameters.
8. A system for generating noise maps and optimizing routes for low-altitude aircraft, characterized in that, The method applied to any one of claims 1 to 7 includes: Data acquisition and preprocessing unit: used to acquire aircraft status, environmental and acoustic data in real time, and to perform standardization and rasterization processing; Noise map generation unit: Built-in physical-data hybrid noise prediction network, used to generate high-resolution noise maps based on real-time data; Silent Path Optimization Unit: Built-in silent path optimization network, used to calculate silent optimization flight commands based on the current state and noise map; Rolling Execution and Feedback Unit: Used to execute optimization instructions, update system status, and achieve closed-loop rolling optimization.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for generating noise maps and optimizing routes for low-altitude aircraft as described in any one of claims 1 to 7.