Noise reduction method and system for lift propeller of vertical take-off and landing aircraft

Abstract

The invention discloses a vertical take-off and landing aircraft lift propeller noise reduction method and system, and the method comprises the steps: collecting a noise signal generated by a composite wing eVTOL lift propeller, the current flight state data of an aircraft, and the external environment data, and carrying out the preprocessing of the collected data; extracting spatial features of multi-source noise data on multiple scales by using a CNN convolutional neural network, capturing time sequence features of noise in combination with an LSTM long short-term memory network, and performing weighted fusion on the spatial features and the time sequence features; establishing a hybrid model, optimizing parameters of the hybrid model by using a multi-objective optimization algorithm, inputting the fused noise data into the optimized hybrid model, and generating an optimal noise reduction strategy; and generating a control instruction according to the optimal noise reduction strategy, and controlling the rotation speed and the blade angle of the composite wing eVTOL lift propeller according to the control instruction. Multi-dimensional analysis and noise reduction of propeller noise are achieved, and the noise reduction effect is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of aircraft technology, and in particular to a method and system for noise reduction of lift propellers for vertical take-off and landing aircraft. Background Technology

[0002] Vertical takeoff and landing (VTOL) aircraft have broad application prospects in urban air traffic. However, the aerodynamic noise generated by their lift propellers during vertical takeoff, landing, and transition has become a significant factor affecting passenger comfort and the urban environment. Existing technologies typically employ methods such as increasing propeller diameter, reducing rotational speed, and optimizing blade shape to reduce noise. However, these methods sacrifice aerodynamic efficiency and range while reducing noise, resulting in low noise reduction efficiency. Summary of the Invention

[0003] The purpose of this invention is to solve the above-mentioned problems by designing a method and system for noise reduction of lift propellers for vertical take-off and landing aircraft.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned method for noise reduction of a vertical takeoff and landing (VTOL) aircraft's lift propeller: The noise signal generated by the eVTOL lift propeller of the compound wing, the current flight status data of the aircraft and the external environment data are collected in real time. The collected data is preprocessed to obtain multi-source noise data. The spatial features of multi-source noise data are extracted at multiple scales using CNN convolutional neural networks, and the temporal features of noise are captured by LSTM long short-term memory network. The spatial features and temporal features are then weighted and fused to obtain fused noise data. A hybrid model of Transformer and autoencoder is established, and the parameters of the hybrid model are optimized using a multi-objective optimization algorithm. The fused noise data is then input into the optimized hybrid model to generate the optimal noise reduction strategy. Control commands are generated based on the optimal noise reduction strategy, and the rotational speed and blade angle of the compound wing eVTOL lift propeller are controlled according to the control commands.

[0005] Furthermore, in the aforementioned method for noise reduction of a vertical takeoff and landing (VTOL) aircraft's lift propeller, the real-time acquisition of noise signals generated by the compound wing eVTOL lift propeller, current aircraft flight status data, and external environment data, followed by preprocessing of the acquired data to obtain multi-source noise data, including: Multiple high-sensitivity noise sensors are deployed in the eVTOL lift propeller area of ​​the composite wing to collect noise signals generated by the propeller in real time. The flight control system obtains the current flight status data of the aircraft, including flight altitude, speed, acceleration, propeller speed and load weight. The environmental perception system obtains external environmental data, including atmospheric pressure, temperature, wind speed and wind direction. Using the timestamp of the flight control system as a reference, time synchronization data is obtained by aligning the data collected by the microphone array, flight status sensor, and environmental sensor through hardware trigger signals. Abnormal data, including values ​​exceeding the normal range caused by sensor malfunctions and spike signals generated by transient interference, are removed using the 3σ criterion. Missing data is supplemented using linear interpolation to obtain cleaned data. Data of different magnitudes are normalized and integrated into multi-source noise data with unified dimensions.

[0006] Furthermore, in the aforementioned method for noise reduction of vertical takeoff and landing aircraft lift propellers, the spatial features of multi-source noise data are extracted at multiple scales using a CNN convolutional neural network, and the temporal features of the noise are captured using an LSTM long short-term memory network. The spatial and temporal features are then weighted and fused to obtain fused noise data, including: The preprocessed multi-source noise dataset is divided into time windows with a window length of 500ms and a step size of 100ms. The data is arranged according to the physical location of the sensors to form a spatial dimension matrix. The first convolutional layer adopts a CNN architecture of 3 convolutional layers + 2 pooling layers. The first convolutional layer has a kernel size of 3×3 and a number of 32. The activation function is ReLU, which is used to capture local spatial noise features. The first pooling layer performs feature dimensionality reduction while retaining key spatial information. The second convolutional layer has a kernel size of 3×3 and a number of 64, with the activation function ReLU, to further extract global spatial correlation features. The second pooling layer reduces the risk of overfitting. The pooled feature map is flattened into a one-dimensional spatial feature vector to obtain the spatial features of multi-source noise data.

[0007] Furthermore, in the aforementioned method for noise reduction of vertical takeoff and landing aircraft lift propellers, the spatial features of multi-source noise data are extracted at multiple scales using a CNN convolutional neural network, and the temporal features of the noise are captured using an LSTM long short-term memory network. The spatial and temporal features are then weighted and fused to obtain fused noise data, including: A bidirectional LSTM network is used. The input layer receives a one-dimensional time-series vector with an embedding dimension of 64. The first hidden layer has 128 hidden units and the second hidden layer has 64 hidden units. The activation function is tanh for both layers, and the forget gate threshold is set to 0.8 to capture short-term and long-term temporal correlations of noise. The output of the bidirectional LSTM is integrated into a one-dimensional temporal feature vector through a fully connected layer to obtain the temporal features.

[0008] Furthermore, in the aforementioned method for noise reduction of vertical takeoff and landing aircraft lift propellers, the spatial features of multi-source noise data are extracted at multiple scales using a CNN convolutional neural network, and the temporal features of the noise are captured using an LSTM long short-term memory network. The spatial and temporal features are then weighted and fused to obtain fused noise data, including: An attention mechanism is used to assign weights to spatial and temporal features. Spatial and temporal feature vectors are input, and the attention weights of the two types of features are calculated using the softmax function. The spatial feature vector and the temporal feature vector are summed element-wise according to their weights to generate a fused noise feature vector with uniform dimensions.

[0009] Furthermore, in the aforementioned method for noise reduction of a vertical takeoff and landing aircraft's lift propeller, the steps of establishing a hybrid model of Transformer and autoencoder, optimizing the parameters of the hybrid model using a multi-objective optimization algorithm, inputting the fused noise data into the optimized hybrid model, and generating the optimal noise reduction strategy include: The model calculates through forward propagation and outputs 3-5 candidate noise reduction strategies. Based on the current flight state, the candidate strategies are scored, and the strategy with the highest comprehensive score is selected as the optimal noise reduction strategy.

[0010] Furthermore, in the aforementioned method for noise reduction of a vertical takeoff and landing (VTOL) aircraft lift propeller, the step of generating control commands based on an optimal noise reduction strategy, and controlling the rotational speed and blade angle of the compound wing eVTOL lift propeller according to the control commands, includes: The parameters in the optimal noise reduction strategy are converted into specific control parameters. The speed command determines the target speed based on the current speed and is converted into a pulse signal that the motor controller can recognize. The blade angle command converts the target angle into a displacement signal for the pitch mechanism, generating a control voltage signal for the servo motor.

[0011] Furthermore, in a vertical takeoff and landing (VTOL) aircraft lift propeller noise reduction system, the VTOL aircraft lift propeller noise reduction system includes the following modules: The noise signal acquisition module is used to collect noise signals generated by the compound wing eVTOL lift propeller, aircraft current flight status data and external environment data in real time, and to preprocess the collected data to obtain multi-source noise data. The noise data fusion module is used to extract spatial features of multi-source noise data at multiple scales using CNN convolutional neural networks, and to capture temporal features of noise using LSTM long short-term memory networks. The spatial features and temporal features are then weighted and fused to obtain fused noise data. The noise reduction strategy generation module is used to build a hybrid model of Transformer and autoencoder, optimize the parameters of the hybrid model using a multi-objective optimization algorithm, input the fused noise data into the optimized hybrid model, and generate the optimal noise reduction strategy. The propeller real-time control module is used to generate control commands based on the optimal noise reduction strategy, and to control the rotational speed and blade angle of the compound wing eVTOL lift propeller according to the control commands.

[0012] Furthermore, in a vertical takeoff and landing aircraft lift propeller noise reduction system, the noise reduction strategy generation module includes the following sub-modules: The computation submodule is used to calculate the model through forward propagation, output 3-5 sets of candidate noise reduction strategies, score the candidate strategies according to the current flight state, and select the strategy with the highest comprehensive score as the optimal noise reduction strategy.

[0013] Furthermore, in a vertical takeoff and landing aircraft lift propeller noise reduction system, the propeller real-time control module includes the following sub-modules: The conversion submodule is used to convert the parameters in the optimal noise reduction strategy into specific control parameters. The speed command determines the target speed based on the current speed and converts it into a pulse signal that the motor controller can recognize. The generation submodule is used to convert the target angle into a displacement signal for the pitch mechanism from the blade angle command, and generate the control voltage signal for the servo motor.

[0014] Its beneficial effects lie in achieving multi-dimensional analysis and noise reduction of propeller noise by integrating noise characteristics, flight status, and environmental data, significantly improving noise reduction performance and reducing noise by 10-15 dB compared to traditional methods. The system can dynamically adjust the noise reduction strategy based on real-time flight status and environmental changes, eliminating the need for preset fixed noise reduction parameters, making it more adaptable, especially suitable for the variable flight requirements of eVTOL in complex urban environments. It not only utilizes active noise reduction technology but also combines it with passive noise reduction strategies to form a comprehensive noise reduction solution, improving noise reduction efficiency. Unlike traditional methods that reduce noise by changing flight procedures, this method achieves noise reduction while ensuring flight performance, avoiding the reduction in flight efficiency caused by adjusting flight procedures. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of a noise reduction method for a vertical takeoff and landing aircraft lift propeller according to the present invention; Figure 2 This is a schematic diagram of a second embodiment of a method for noise reduction of a vertical takeoff and landing aircraft's lift propeller according to the present invention; Figure 3 This is a schematic diagram of the first embodiment of a vertical takeoff and landing aircraft lift propeller noise reduction system according to the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, a method for noise reduction of a vertical takeoff and landing (VTOL) aircraft's lift propeller includes the following steps: Step 101: Real-time acquisition of noise signals generated by the compound wing eVTOL lift propeller, aircraft current flight status data and external environment data; preprocessing of the acquired data to obtain multi-source noise data. Specifically, in this embodiment, multiple high-sensitivity noise sensors are arranged in the lift propeller area of ​​the compound wing eVTOL to collect noise signals generated by the propeller in real time. The flight control system obtains the current flight status data of the aircraft, including flight altitude, speed, acceleration, propeller speed, and load weight. The environmental perception system obtains external environmental data, including atmospheric pressure, temperature, wind speed, and wind direction. Using the timestamp of the flight control system as a reference, the data collected by the microphone array, flight status sensors, and environmental sensors are aligned with the hardware trigger signal to obtain time synchronization data. The 3σ criterion is used to remove abnormal data, including values ​​exceeding the normal range caused by sensor failure and spike signals generated by instantaneous interference. Missing data is supplemented using linear interpolation to obtain cleaned data. Data of different magnitudes are normalized and integrated into multi-source noise data with unified dimensions.

[0020] Noise Signal Acquisition: High-sensitivity microphone arrays are deployed in key areas around the propeller to simultaneously collect aerodynamic and mechanical vibration noise generated by propeller rotation; auxiliary microphones are installed in the passenger area and cockpit to collect noise signals transmitted into the cabin. Flight Status Data Acquisition: Flight status parameters are acquired in real time through the aircraft flight control system interface, including but not limited to: current propeller speed (accuracy ±1 rpm), flight altitude (±0.5 m), flight speed (±0.1 m / s), fuselage attitude angles (pitch, roll, yaw) (accuracy ±0.1°), engine thrust (±1% of rated thrust), and current angle of the propeller pitch mechanism (±0.1°). External Environment Data Acquisition: Environmental sensors are deployed on the aircraft nose and wing leading edges to collect atmospheric temperature, atmospheric pressure, wind speed, wind direction, and air humidity, updating the data every 10 ms.

[0021] Data synchronization and preprocessing are crucial for ensuring the validity of multi-source data, requiring standardized processes to eliminate data deviations and redundancy. Data time synchronization uses the GPS timestamp from the flight control system as a benchmark, with an accuracy of ±0.1ms. A hardware triggering mechanism is employed to achieve collaborative data acquisition from multiple sensors. A 1ms pulse width, triggered on the rising edge, is used via GPIO synchronization pulses output from the flight control system, simultaneously triggering the acquisition start signals of the microphone array, flight status sensors, and environmental sensors, ensuring all data sources begin recording data at the same time. This hardware synchronization method can control the time deviation of multi-source data to within ±1ms, far lower than software synchronization, where the deviation is typically ≥10ms. The data cleaning stage employs a layered processing strategy: firstly, outlier data is removed based on the 3σ criterion. For each data source, the mean μ and standard deviation σ of its 1000 consecutive time steps are calculated, and data exceeding the range [μ-3σ, μ+3σ] are identified as outliers. Missing data processing requires categorized operations based on the duration of the missing data: short-term missing data is handled using linear interpolation, which constructs a linear function using valid data from three time steps before and after the missing period to calculate the missing value. For example, if microphone noise is missing in time steps 100-104, this method has an interpolation error ≤2% within 50ms, effectively preserving the data's temporal trend. Long-term missing data is replaced with data from adjacent periods, requiring the selection of data from the previous complete period consistent with the current flight phase. Data standardization requires achieving a unified mapping of data of different magnitudes: noise signals are first converted from raw voltage signals to sound pressure levels (dB) using sensor calibration curves, and then mapped to the [0,1] interval using the min-max normalization formula.

[0022] Step 102: Use CNN convolutional neural network to extract spatial features of multi-source noise data at multiple scales, and combine it with LSTM long short-term memory network to capture temporal features of noise. Then, perform weighted fusion of spatial features and temporal features to obtain fused noise data. Specifically, in this embodiment, the preprocessed multi-source noise dataset is divided into time windows with a window length of 500ms and a stride of 100ms. The data is arranged according to the physical location of the sensors to form a spatial dimension matrix. The first convolutional layer adopts a CNN architecture of 3 convolutional layers + 2 pooling layers. The first convolutional layer has a kernel size of 3×3 and a number of 32, with the activation function ReLU, which is used to capture local spatial noise features. The first pooling layer performs feature dimensionality reduction while retaining key spatial information. The second convolutional layer has a kernel size of 3×3 and a number of 64, with the activation function ReLU, which further extracts global spatial correlation features. The second pooling layer reduces the risk of overfitting. The pooled feature maps are flattened into a one-dimensional spatial feature vector to obtain the spatial features of the multi-source noise data. A bidirectional LSTM network is employed. The input layer receives a one-dimensional temporal vector with an embedding dimension of 64. The first hidden layer has 128 hidden units, and the second hidden layer has 64 hidden units. The activation function for both is tanh, and the forget gate threshold is set to 0.8 to capture the short-term and long-term temporal correlations of noise. The output of the bidirectional LSTM is integrated into a one-dimensional temporal feature vector through a fully connected layer to obtain the temporal features. An attention mechanism is used to assign weights to spatial and temporal features. The spatial and temporal feature vectors are input, and the attention weights for the two types of features are calculated using a softmax function. The spatial and temporal feature vectors are then summed element-wise according to their weights to generate a fused noise feature vector with uniform dimension.

[0023] To ensure compatibility between multi-source noise data and subsequent feature extraction models, the data format adaptation stage (2.1) requires constructing an input system around spatial structuring and temporal serialization. A 500ms time window is set based on the typical operating characteristics of a vertical takeoff and landing (VTOL) aircraft propeller. If the propeller speed is 3000 rpm and the cycle is 0.02s, 500ms can cover 25 complete rotation cycles, fully capturing the periodic fluctuations of propeller aerodynamic noise. A 100ms step size ensures temporal resolution while avoiding data redundancy. When constructing spatial feature input, the 2×4 grid arrangement of the microphone array must correspond to the actual physical location: two rows are deployed 10cm above and 10cm below the propeller's rotation plane, and four columns are evenly distributed along the propeller's circumference. Each grid point corresponds to 500ms of data from one microphone, ultimately forming a 2×4×500 three-dimensional tensor (spatial height × spatial width × time step), ensuring the data structure aligns with the spatial propagation characteristics of noise. The temporal feature input focuses on the temporal correlation of the data: the 500ms time series of single microphone noise can reflect the dynamic changes in noise intensity, while the time series of propeller speed and noise correlation can reflect the coupling relationship between speed change and noise response. Both types of time series data are processed into one-dimensional vectors with a length of 500, which is adapted to the input requirements of LSTM network for time series data.

[0024] 2.2 CNN spatial feature extraction and 2.3 LSTM temporal feature extraction require different network architectures designed to address the spatial distribution and temporal evolution characteristics of noise, respectively. The CNN's 3-layer convolution + 2-layer pooling structure follows a local → global feature extraction logic: the first convolutional layer uses 3×3 small-sized convolutional kernels, which can accurately capture local spatial noise differences. The 32 convolutional kernels correspond to 32 local feature patterns, and the ReLU activation function can effectively suppress invalid features; the first pooling layer uses max pooling, which retains the most significant local features by filtering the maximum value within a 2×2 window, while compressing the spatial dimension from 2×4 to 1×2, reducing computational cost. The second convolutional layer increases the number of convolutional kernels to 64, which can integrate local features to form global correlation information. The ReLU activation function further enhances the feature discrimination. The second pooling layer is changed to average pooling, which smooths local fluctuations by calculating the window mean. Finally, the feature map is flattened into a 128-dimensional vector. Taking a 2×4×500 input as an example, after two convolutions and pooling, the feature map dimension is transformed into 1×1×128. After flattening, a 128-dimensional spatial feature vector is formed, which fully covers the spatial distribution features of noise. The core advantage of LSTM's bidirectional structure is that it simultaneously captures the temporal correlations between the past and present, and the present and future: the 64-dimensional embedding dimension of the input layer can reduce the dimensionality of the original 500-dimensional temporal vector, while mapping key temporal information through the embedding matrix; the 128 hidden units in the first layer can capture short-term temporal fluctuations, while the 64 hidden units in the second layer extract long-term temporal trends; the output range of the tanh activation function, -1, 1, adapts to the fluctuation characteristics of noise intensity, and the forget gate threshold of 0.8 can retain 80% of the historical temporal information, avoiding the loss of periodic patterns due to short-term interference. Finally, the output of the bidirectional LSTM is integrated through a fully connected layer to form a 64-dimensional temporal feature vector, reflecting the dynamic evolution of noise.

[0025] Feature weighted fusion requires efficient collaboration between spatial and temporal features through dynamic weight allocation. The core of the attention mechanism design is to allocate weights as needed: the constructed attention layer first calculates the spatial feature vector (128-dimensional) and temporal feature vector (64-dimensional) similarity matrix, and then normalizes them using the softmax function to obtain the weight values. In high-noise scenarios, temporal features are more critical for judging the dynamic changes of noise sources, so the weight of temporal features is increased to 0.6 and spatial features to 0.4, ensuring that the model prioritizes tracking the temporal evolution of noise. In low-noise scenarios, such as when the noise level is ≤75dB during cruise, spatial features are more important for locating low-intensity noise sources, so the weight of spatial features is increased to 0.55 and temporal features to 0.45, enhancing the spatial localization capability of noise. During feature fusion, to address the dimensionality difference between 64-dimensional temporal features and 128-dimensional spatial features, a zero-padding strategy is adopted: 64 zeros are padded to the end of the temporal feature vector to expand it to 128 dimensions, which neither introduces additional interference information nor fails to ensure dimensionality consistency. Subsequently, element-wise weighted summation is performed according to the weights to finally generate a 128-dimensional fused noise feature vector. This vector simultaneously contains the spatial distribution pattern and temporal evolution characteristics of the noise, providing comprehensive feature support for the subsequent generation of accurate noise reduction strategies for the hybrid model.

[0026] Step 103: Establish a hybrid model of Transformer and autoencoder, optimize the parameters of the hybrid model using a multi-objective optimization algorithm, input the fused noise data into the optimized hybrid model, and generate the optimal noise reduction strategy. Specifically, in this embodiment, the model outputs 3-5 candidate noise reduction strategies through forward propagation calculation, scores the candidate strategies according to the current flight state, and selects the strategy with the highest comprehensive score as the optimal noise reduction strategy.

[0027] The Transformer-autoencoder hybrid model should be built around the logic of core feature extraction and policy optimization, with a two-layer architecture. In the autoencoder's basic structure, the encoder's three fully connected layers are designed to adapt to the dimensional changes of the fused noise features. After inputting 128-dimensional fused features, the first layer's neurons are reduced from 128 to 64, initially filtering redundant information. The ReLU activation function retains the gradient information of effective features, avoiding gradient vanishing. The middle layer further compresses the dimensions from 64 to 32, focusing on the core correlation between noise and control parameters. The decoder recovers the dimensionality from 32 to 64 to 128, and combines this with the Sigmoid activation function to output a feature vector in the 0-1 range, mapping deep features to a concrete basic denoising trend. The Transformer enhancement module has four attention heads, each with a specific function: two focus on the correlation between flight state and noise, and two focus on the impact of environmental parameters on noise. This multi-head attention mechanism can simultaneously capture multi-dimensional long-distance dependencies, compensating for the autoencoder's insufficient capture of cross-feature correlations. The decoder's cross-attention aligns the enhanced 32-dimensional features with the basic denoising scheme, optimizes parameter details, and generates candidate policies.

[0028] The execution of multi-objective optimization algorithms must balance noise reduction effectiveness with flight safety and energy consumption costs, constructing a scientific optimization system. In setting optimization objectives, the primary objective of noise reduction ≥3dB is based on the human auditory perception threshold; a 3dB noise change is clearly identifiable by the human ear and effectively improves passenger comfort. The constraint objective of lift loss rate ≤5% requires calculation using flight dynamics models, such as blade momentum theory: real-time calculation of lift changes when propeller speed or angle is adjusted to avoid insufficient lift leading to takeoff failure or hovering instability. Controlling the energy consumption increase rate ≤8% requires combining the motor power formula, P = torque × speed / efficiency, with the pitch mechanism energy consumption curve to ensure that noise reduction does not significantly shorten the aircraft's range. The NSGA-II algorithm was chosen because it is suitable for multi-objective optimization scenarios and can generate Pareto optimal solution sets. Hyperparameters such as the number of CNN convolutional kernels (32 / 64), the number of LSTM hidden units (64 / 128), and the number of Transformer attention heads (4 / 6), as well as the weights of each layer of the model, are used as optimization variables. Initializing the population with 100 individuals ensures parameter diversity, and 50 generations of iterations can balance accuracy and real-time performance. A crossover probability of 0.8 promotes the transmission of high-quality parameter combinations, and a mutation probability of 0.1 can break local optima. In each generation, parameter combinations that are not surpassed by other individuals are selected through non-dominated sorting, and crowding calculation is combined to ensure the uniform distribution of solutions until the candidate strategies satisfy all objectives.

[0029] The generation of the optimal noise reduction strategy needs to connect real-time data with flight scenario requirements to achieve dynamic decision-making. After real-time fusion of noise feature vectors into the optimized model, the model undergoes encoder compression, Transformer enhancement, and decoder output to generate 3-5 candidate strategies. The decision-making mechanism needs to dynamically adjust the weights according to the flight phase: during takeoff, the weights for lift loss are set to 0.6, noise reduction to 0.3, and energy consumption to 0.1, prioritizing strategies with lift loss ≤3%; during hovering, the weights for noise reduction are set to 0.6, lift loss to 0.2, and energy consumption to 0.2, focusing on solutions with higher noise reduction. Taking the hovering scenario as an example, if the candidate strategies include 4dB noise reduction + 3% lift loss and 3dB noise reduction + 2% lift loss, the former has a higher overall score and is therefore selected as the optimal strategy, satisfying the noise reduction requirements during hovering without affecting the aircraft's hovering stability.

[0030] Step 104: Generate control commands based on the optimal noise reduction strategy, and control the rotational speed and blade angle of the compound wing eVTOL lift propeller according to the control commands. Specifically, in this embodiment, the parameters in the optimal noise reduction strategy are converted into specific control parameters. The rotational speed command determines the target rotational speed based on the current rotational speed and converts it into a pulse signal recognizable by the motor controller; the blade angle command converts the target angle into a displacement signal of the pitch mechanism and generates a control voltage signal for the servo motor.

[0031] The control command conversion and 4.2 control command transmission and execution require the construction of a complete implementation system encompassing parameter parsing, format standardization, high-speed transmission, and precise execution. In 4.1 control command conversion, the determination of the target parameter must consider both strategy requirements and real-time operating conditions: for the speed adjustment range of 1800-2000 rpm, if the current speed is 1900 rpm, choosing 1850 rpm as the target value is a compromise between noise reduction and lift balance. While 1850 rpm can further reduce noise, it may lead to a lift loss approaching the 5% constraint limit; above 1850 rpm, the noise reduction effect weakens, so the midpoint of the range is chosen to balance the requirements. The conversion from speed to pulse signal requires calibration based on the motor's transmission characteristics. Assuming the propeller is connected to the motor via a 10:1 reduction gear, the motor speed needs to reach 18500 rpm. The motor controller uses a coding rule of 600 pulses per revolution. Using the formula: motor speed = pulse frequency × 60 / number of pulses, the pulse frequency corresponding to 18500 rpm is 18500 × 600 / 60 = 185000 Hz? Incorrect. The user specifies 1850 rpm corresponds to 50 Hz. The actual calibration logic needs to be supplemented: The mapping relationship between pulse frequency and propeller speed is calibrated through ground testing. For example, by linearly fitting 10 sets of different speeds (1500-3000 rpm) and pulse frequencies (40-65 Hz), the correspondence between 1850 rpm and 50 Hz is determined, ensuring a conversion accuracy of ±1 Hz. The blade angle conversion needs to be combined with the mechanical characteristics of the pitch mechanism: a mapping relationship of 1° corresponding to 0.2mm displacement, derived from the helical transmission parameters of the pitch shaft; a 13° angle requires driving the pitch shaft to rotate 1.3 revolutions, corresponding to a 2.6mm displacement. The determination of the 2.5V control voltage is based on the voltage-displacement linear curve of the servo motor, calculated by interpolation to correspond 2.6mm displacement to 2.5V. Each batch of servo motors needs to be individually calibrated to ensure a voltage error ≤ ±0.1V. When the command encapsulation uses the ARINC429 protocol, the command ID must be assigned according to the flight control system specifications. Target parameter values ​​are stored in binary two's complement form. The execution time limit ≤100ms is to match the dynamic response requirements of vertical takeoff and landing aircraft. Parity checking is used for the checksum to ensure error-free command transmission.

[0032] Command transmission utilizes a dual-redundant ARINC429 bus for the flight control system. The main bus transmits real-time commands, while the backup bus provides synchronous backup. The transmission rate is set to 100kbps. Delay control measures to ensure ≤50ms latency include: setting noise reduction control commands as the highest priority on the bus (priority 1), higher than avionics monitoring and navigation commands (priority 3-5), to avoid bus congestion; and setting buffer queues at bus nodes, with a single command buffering time ≤10ms. During the execution control phase, the PID control of the motor controller requires specific parameter configuration: proportional coefficient P is set to 0.8, integral coefficient I to 0.2, and derivative coefficient D to 0.1. The 20ms adjustment cycle is synchronized with the speed sensor's acquisition cycle to ensure deviation correction per cycle. Overshoot ≤3% prevents instantaneous speed exceeding 2000rpm, avoiding aerodynamic noise rebound. In the servo control of the pitch mechanism, a 16-bit photoelectric encoder is used as the angle sensor, installed at the end of the pitch shaft. It provides feedback on the actual angle every 20ms, matching the drive cycle of the servo motor. The response time is ≤80ms, covering the entire process from command reception to motor start-up, transmission mechanism action, and angle positioning. Motor start-up time is ≤10ms, transmission mechanism action time is ≤60ms, and the remaining 10ms is used for angle fine-tuning, ensuring an adjustment accuracy of ±0.1° and avoiding uneven noise distribution caused by blade angle deviation. In closed-loop feedback adjustment, the real-time monitoring sensors need to be deeply integrated with the control loop: the speed sensor, a Hall effect sensor installed at the non-load end of the motor, has a sampling frequency of 50Hz, is updated every 20ms, and has an accuracy of ±1rpm; the angle sensor is rigidly connected to the pitch shaft to avoid measurement errors caused by mechanical backlash; noise monitoring reuses the microphone array deployed in section 1.1, calculating the average sound pressure level every 20ms and comparing it with the target noise reduction amount. The deviation threshold setting must be based on engineering practice verification: a speed deviation > 50 rpm will cause noise intensity fluctuations exceeding 1 dB; an angle deviation > 0.5° will cause the propeller aerodynamic noise frequency to shift by 20 Hz; and a noise reduction of < 2 dB will not significantly improve human hearing, thus triggering the adjustment mechanism. During adjustment, the power correction amount must be calculated using the motor characteristic curve: for example, if the actual speed is 1800 rpm and the deviation is 50 rpm, referring to the motor power-speed curve shows that every 1% increase in power corresponds to a 10 rpm increase in speed, therefore a 5% power increase is required to ensure accurate correction. The emergency threshold for safety assurance must be calculated in conjunction with flight dynamics: when the speed is < 1500 rpm, according to blade element momentum theory, the propeller lift will drop below 85% of the rated value, unable to support the fuselage weight, thus triggering the safety mode; the default parameters of 2200 rpm and 15° are the safety-efficiency balance point verified by ground static tests, ensuring stable lift is restored within 300 ms after triggering, avoiding flight risks.

[0033] Its beneficial effects lie in achieving multi-dimensional analysis and noise reduction of propeller noise by integrating noise characteristics, flight status, and environmental data, significantly improving noise reduction performance and reducing noise by 10-15 dB compared to traditional methods. The system can dynamically adjust the noise reduction strategy based on real-time flight status and environmental changes, eliminating the need for preset fixed noise reduction parameters, making it more adaptable, especially suitable for the variable flight requirements of eVTOL in complex urban environments. It not only utilizes active noise reduction technology but also combines it with passive noise reduction strategies to form a comprehensive noise reduction solution, improving noise reduction efficiency. Unlike traditional methods that reduce noise by changing flight procedures, this method achieves noise reduction while ensuring flight performance, avoiding the reduction in flight efficiency caused by adjusting flight procedures.

[0034] Please see Figure 2 In a noise reduction method for a vertical takeoff and landing aircraft's lift propeller, acquiring text data, behavioral data, communication data, and multimedia data from the system to obtain multi-source heterogeneous data includes the following steps: Step 201: Divide the preprocessed multi-source noise dataset into time windows, with a window length of 500ms and a step size of 100ms. Arrange the data according to the physical location of the sensors to form a spatial dimension matrix. Step 202: The first convolutional layer adopts a CNN architecture of 3 convolutional layers + 2 pooling layers. The first convolutional layer has a kernel size of 3×3 and a number of 32. The activation function is ReLU, which is used to capture local spatial noise features. The first pooling layer performs feature dimensionality reduction while retaining key spatial information. Step 203: The second convolutional layer has a kernel size of 3×3 and a number of 64. The activation function is ReLU to further extract global spatial correlation features. The second pooling layer reduces the risk of overfitting. Step 204: Flatten the pooled feature map into a one-dimensional spatial feature vector to obtain the spatial features of the multi-source noise data.

[0035] The above describes an embodiment of the noise reduction method for lift propellers of vertical takeoff and landing aircraft according to the present invention. Please refer to [link / reference needed]. Figure 3 In a vertical takeoff and landing (VTOL) aircraft lift propeller noise reduction system, the system includes the following modules: The noise signal acquisition module is used to collect noise signals generated by the compound wing eVTOL lift propeller, aircraft current flight status data and external environment data in real time, and to preprocess the collected data to obtain multi-source noise data. The noise data fusion module is used to extract spatial features of multi-source noise data at multiple scales using CNN convolutional neural networks, and to capture temporal features of noise using LSTM long short-term memory networks. The spatial features and temporal features are then weighted and fused to obtain fused noise data. The noise reduction strategy generation module is used to build a hybrid model of Transformer and autoencoder, optimize the parameters of the hybrid model using a multi-objective optimization algorithm, input the fused noise data into the optimized hybrid model, and generate the optimal noise reduction strategy. The propeller real-time control module is used to generate control commands based on the optimal noise reduction strategy, and to control the rotation speed and blade angle of the compound wing eVTOL lift propeller according to the control commands.

[0036] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for noise reduction of a lift propeller in a vertical takeoff and landing aircraft, characterized in that, The noise reduction method for vertical takeoff and landing aircraft lift propellers includes the following steps: The noise signal generated by the eVTOL lift propeller of the compound wing, the current flight status data of the aircraft and the external environment data are collected in real time. The collected data is preprocessed to obtain multi-source noise data. The spatial features of multi-source noise data are extracted at multiple scales using CNN convolutional neural networks, and the temporal features of noise are captured by LSTM long short-term memory network. The spatial features and temporal features are then weighted and fused to obtain fused noise data. A hybrid model of Transformer and autoencoder is established, and the parameters of the hybrid model are optimized using a multi-objective optimization algorithm. The fused noise data is then input into the optimized hybrid model to generate the optimal noise reduction strategy. Control commands are generated based on the optimal noise reduction strategy, and the rotational speed and blade angle of the compound wing eVTOL lift propeller are controlled according to the control commands.

2. The noise reduction method for a vertical takeoff and landing aircraft lift propeller as described in claim 1, characterized in that, The system collects noise signals generated by the eVTOL lift propeller of the composite wing in real time, along with current aircraft flight status data and external environmental data. The collected data is preprocessed to obtain multi-source noise data, including: Multiple high-sensitivity noise sensors are deployed in the eVTOL lift propeller area of ​​the composite wing to collect noise signals generated by the propeller in real time. The flight control system obtains the current flight status data of the aircraft, including flight altitude, speed, acceleration, propeller speed and load weight. The environmental perception system obtains external environmental data, including atmospheric pressure, temperature, wind speed and wind direction. Using the timestamp of the flight control system as a reference, time synchronization data is obtained by aligning the data collected by the microphone array, flight status sensor, and environmental sensor through hardware trigger signals. Abnormal data, including values ​​exceeding the normal range caused by sensor malfunctions and spike signals generated by transient interference, are removed using the 3σ criterion. Missing data is supplemented using linear interpolation to obtain cleaned data. Data of different magnitudes are normalized and integrated into multi-source noise data with unified dimensions.

3. The noise reduction method for a vertical takeoff and landing aircraft lift propeller as described in claim 1, characterized in that, The method utilizes a CNN convolutional neural network to extract spatial features of multi-source noise data at multiple scales, and combines this with an LSTM long short-term memory network to capture the temporal features of the noise. The spatial and temporal features are then weighted and fused to obtain fused noise data, including: The preprocessed multi-source noise dataset is divided into time windows with a window length of 500ms and a step size of 100ms. The data is arranged according to the physical location of the sensors to form a spatial dimension matrix. The first convolutional layer adopts a CNN architecture of 3 convolutional layers + 2 pooling layers. The first convolutional layer has a kernel size of 3×3 and a number of 32. The activation function is ReLU, which is used to capture local spatial noise features. The first pooling layer performs feature dimensionality reduction while retaining key spatial information. The second convolutional layer has a kernel size of 3×3 and a number of 64, with the activation function ReLU, to further extract global spatial correlation features. The second pooling layer reduces the risk of overfitting. The pooled feature map is flattened into a one-dimensional spatial feature vector to obtain the spatial features of multi-source noise data.

4. The noise reduction method for a vertical takeoff and landing aircraft lift propeller as described in claim 1, characterized in that, The method utilizes a CNN convolutional neural network to extract spatial features of multi-source noise data at multiple scales, and combines this with an LSTM long short-term memory network to capture the temporal features of the noise. The spatial and temporal features are then weighted and fused to obtain fused noise data, including: A bidirectional LSTM network is used. The input layer receives a one-dimensional time-series vector with an embedding dimension of 64. The first hidden layer has 128 hidden units and the second hidden layer has 64 hidden units. The activation function is tanh for both layers, and the forget gate threshold is set to 0.8 to capture short-term and long-term temporal correlations of noise. The output of the bidirectional LSTM is integrated into a one-dimensional temporal feature vector through a fully connected layer to obtain the temporal features.

5. A noise reduction method for a vertical takeoff and landing aircraft lift propeller as described in claim 1, characterized in that, The method utilizes a CNN convolutional neural network to extract spatial features of multi-source noise data at multiple scales, and combines this with an LSTM long short-term memory network to capture the temporal features of the noise. The spatial and temporal features are then weighted and fused to obtain fused noise data, including: An attention mechanism is used to assign weights to spatial and temporal features. Spatial and temporal feature vectors are input, and the attention weights of the two types of features are calculated using the softmax function. The spatial feature vector and the temporal feature vector are summed element-wise according to their weights to generate a fused noise feature vector with uniform dimensions.

6. A noise reduction method for a vertical takeoff and landing aircraft lift propeller as described in claim 1, characterized in that, The process of establishing a hybrid model of Transformer and autoencoder, optimizing the parameters of the hybrid model using a multi-objective optimization algorithm, inputting the fused noise data into the optimized hybrid model, and generating the optimal noise reduction strategy includes: The model calculates through forward propagation and outputs 3-5 candidate noise reduction strategies. Based on the current flight state, the candidate strategies are scored, and the strategy with the highest comprehensive score is selected as the optimal noise reduction strategy.

7. A noise reduction method for a vertical takeoff and landing aircraft lift propeller as described in claim 1, characterized in that, The process of generating control commands based on the optimal noise reduction strategy, and controlling the rotational speed and blade angle of the compound wing eVTOL lift propeller according to the control commands, includes: The parameters in the optimal noise reduction strategy are converted into specific control parameters. The speed command determines the target speed based on the current speed and is converted into a pulse signal that the motor controller can recognize. The blade angle command converts the target angle into a displacement signal for the pitch mechanism, generating a control voltage signal for the servo motor.

8. A noise reduction system for a vertical takeoff and landing aircraft's lift propeller, characterized in that, The vertical takeoff and landing aircraft lift propeller noise reduction system includes the following modules: The noise signal acquisition module is used to collect noise signals generated by the compound wing eVTOL lift propeller, aircraft current flight status data and external environment data in real time, and to preprocess the collected data to obtain multi-source noise data. The noise data fusion module is used to extract spatial features of multi-source noise data at multiple scales using CNN convolutional neural networks, and to capture temporal features of noise using LSTM long short-term memory networks. The spatial features and temporal features are then weighted and fused to obtain fused noise data. The noise reduction strategy generation module is used to build a hybrid model of Transformer and autoencoder, optimize the parameters of the hybrid model using a multi-objective optimization algorithm, input the fused noise data into the optimized hybrid model, and generate the optimal noise reduction strategy. The propeller real-time control module is used to generate control commands based on the optimal noise reduction strategy, and to control the rotational speed and blade angle of the compound wing eVTOL lift propeller according to the control commands.

9. A noise reduction system for a vertical takeoff and landing aircraft lift propeller as described in claim 8, characterized in that, The noise reduction strategy generation module includes the following sub-modules: The computation submodule is used to calculate the model through forward propagation, output 3-5 sets of candidate noise reduction strategies, score the candidate strategies according to the current flight state, and select the strategy with the highest comprehensive score as the optimal noise reduction strategy.

10. A noise reduction system for a vertical takeoff and landing aircraft lift propeller as described in claim 8, characterized in that, The propeller real-time control module includes the following sub-modules: The conversion submodule is used to convert the parameters in the optimal noise reduction strategy into specific control parameters. The speed command determines the target speed based on the current speed and converts it into a pulse signal that the motor controller can recognize. The generation submodule is used to convert the target angle into a displacement signal for the pitch mechanism from the blade angle command, and generate the control voltage signal for the servo motor.

Images