Composite noise source separation method

Abstract

The invention relates to the technical field of hovercars, in particular to a composite noise source separation method. Comprising the following steps: performing short-time Fourier transform processing on an acquired noise signal to obtain a time-frequency domain signal; performing preliminary separation by adopting an improved FastICA algorithm, initializing a rotor noise component by utilizing a rotor rotating speed, and initializing an engine noise component by utilizing an engine working condition; spatial optimization is carried out by adopting a frequency domain joint diagonalization algorithm based on independent vector analysis (IVA), sound source position probability distribution is calculated by combining microphone array position information, a TDOA technology and a beam forming technology, and the number of noise sources is adaptively detected; and a CNN-RNN hybrid model is constructed to carry out fine trimming and classification on the separated noise features, the CNN-RNN hybrid model adopts a deep separable convolution and BiLSTM network structure, and flight attitude parameters are used as additional input. According to the technical scheme, the composite noise source of the hovercar can be separated more accurately and effectively.

Description

Technical Field

[0001] This invention relates to the field of flying car technology, and more specifically to a method for separating complex noise sources. Background Technology

[0002] Flying cars, as a new type of transportation that integrates aviation and automotive technologies, offer new possibilities for people's travel with their unique flight modes and diverse application scenarios. They are expected to play an important role in alleviating urban traffic congestion and achieving rapid point-to-point transportation.

[0003] However, noise has become one of the key obstacles to be addressed in the process of commercializing flying cars. Flying cars generate extremely complex noise under different flight conditions, such as high-speed rotor rotation to provide lift during takeoff, maintaining a stable attitude during hovering, precisely controlling descent speed during landing, and changing flight direction during translation. This noise is not generated by a single source, but rather is the result of a mixture and superposition of multiple noise sources, including rotor noise from the high-speed rotation of the rotor interacting with the air, engine noise from engine operation, and wind noise from airflow disturbances during flight.

[0004] Traditional noise testing methods have revealed numerous limitations when dealing with the complex, multi-source mixed noise of flying cars. On one hand, traditional methods often rely on simple signal processing techniques, assuming that noise sources are independent or have specific linear relationships. However, the actual noise sources in flying cars exhibit complex nonlinear coupling relationships, making it difficult for traditional methods to accurately capture the characteristics of each noise source and effectively separate composite noise sources. On the other hand, traditional methods lack effective utilization of the spatial information of noise sources, failing to fully consider factors such as microphone array layout and time delays during sound propagation, resulting in insufficient ability to locate and identify noise sources in complex environments.

[0005] Because it is impossible to effectively separate the composite noise sources, the optimization of noise from flying cars lacks specificity. It is difficult to accurately determine the contribution of different noise sources to the overall noise level, making it impossible to formulate precise and effective noise reduction strategies. Summary of the Invention

[0006] The purpose of this invention is to propose a method for separating composite noise sources, which can more accurately and effectively separate composite noise sources from flying cars.

[0007] To achieve the above objectives, the present invention provides a method for separating composite noise sources, comprising: The acquired noise signal is processed by short-time Fourier transform to obtain the time-frequency domain signal; An improved FastICA algorithm is used for initial separation, and the rotor noise component is initialized using rotor speed and the engine noise component is initialized using engine operating conditions. Spatial optimization is performed using a frequency domain joint diagonalization algorithm based on Independent Vector Analysis (IVA). Combined with microphone array position information, TDOA technology, and beamforming technology, the probability distribution of sound source location is calculated, and the number of noise sources is adaptively detected. A CNN-RNN hybrid model is constructed to refine and classify the separated noise features. The CNN-RNN hybrid model adopts a depthwise separable convolution and BiLSTM network structure, and takes flight attitude parameters as additional input.

[0008] The beneficial effects of the basic scheme are: by using short-time Fourier transform (STFT) to convert the noise signal into a time-frequency domain signal, the limitation of time domain signals being unable to simultaneously reflect the frequency change over time is overcome, and the time-varying characteristics of flying car noise (such as rotor periodic noise and engine transient noise) can be captured more accurately, providing a richer information basis for subsequent separation.

[0009] Traditional Independent Component Analysis (ICA) algorithms are susceptible to slow convergence or biased separation results when processing multi-source mixed signals due to the influence of initial values. This method initializes the corresponding noise components by rotor speed and engine operating conditions, and uses prior information to constrain the separation process. This reduces the number of iterations while avoiding interference from irrelevant signals, thus improving the targeting and efficiency of the initial separation.

[0010] Independent Vector Analysis (IVA) performs joint diagonalization of multi-channel signals in the frequency domain, effectively utilizing the statistical independence of noise sources at different frequencies and addressing the shortcomings of traditional methods in handling frequency correlation. By combining microphone array position information, TDOA (Time Difference of Arrival), and beamforming techniques, sound source localization is achieved through spatial orientation constraints. The probability distribution of sound source locations is calculated, realizing multi-dimensional joint separation of "time-frequency-space," further distinguishing spatially overlapping noise sources (such as the mixing of rotor and engine noise in the same area). Adaptive detection of the number of noise sources avoids errors introduced by manual settings, improving the algorithm's adaptability to complex operating conditions (such as noise source changes during acceleration and turning in flying cars).

[0011] Depthwise separable convolutional neural networks (CNNs) excel at extracting local features from time-frequency maps (such as noise peaks at specific frequencies), while bidirectional LSTM (RNNs) can capture the temporal dependencies of noise signals (such as the periodic fluctuations in rotor rotation). Combining the two can comprehensively mine the time-frequency and temporal-series features of noise, enabling the refinement of residual interference after separation. Introducing flight attitude parameters (such as speed and pitch angle) as additional inputs correlates noise features with operating conditions, improving classification accuracy (such as distinguishing engine noise under different flight conditions), and providing a precise basis for subsequent noise source tracing and noise reduction design.

[0012] As a feasible and preferred approach, the acquired noise signal also includes noise reduction and normalization processing; the noise reduction processing employs filtering techniques, including low-pass filtering or high-pass filtering.

[0013] As a feasible and preferred option, the optimization criteria for the FastICA algorithm are:

[0014] in, , These are standard Gaussian variables.

[0015] As a feasible and preferred solution, the rotor speed is extracted using the Z-axis signal from the gyroscope, and a speed-noise level regression model is established:

[0016] in, The rotor speed, The frequency at which the blade passes through.

[0017] As a feasible and preferred option, preliminary separation also includes: The analysis window length is dynamically adjusted based on flight altitude. Window length = Basic window length × (1 + Height compensation coefficient × Current height / Maximum height).

[0018] As a feasible and preferred approach, the IVA algorithm employs frequency domain joint diagonalization to optimize the objective function:

[0019] in, For the separation matrix, Let be the autocorrelation matrix of the signal.

[0020] As a feasible and preferred approach, this method combines microphone array location information, TDOA technology, and beamforming technology to calculate the probability distribution of sound source locations and adaptively detect the number of noise sources, including the following: By combining the microphone array position as spatial prior information, a sound source-array distance constraint matrix is ​​established; The formula for calculating the probability distribution of sound source location is as follows:

[0021] Adaptive source number detection based on eigenvalue thresholding The number of sources is constrained based on the flight status.

[0022] As a feasible and preferred option, the CNN part of the CNN-RNN hybrid model includes 5 layers of depthwise separable convolutions, BatchNorm and LeakyReLU activation functions, and incorporates a spatial attention mechanism; The RNN part uses a bidirectional 3-layer LSTM with 256 hidden units and uses flight attitude parameters as additional input gating. The CNN feature map is concatenated with the RNN hidden state, and residual connections are used to bypass the noisy feature extraction path.

[0023] As a feasible and preferred approach, the training strategy includes multi-stage training: the first stage fixes the ICA-BSS parameters and trains the neural network; the second stage performs end-to-end joint fine-tuning.

[0024] As a feasible preferred option, it also includes using a flight condition dynamic identification system based on multi-sensor fusion to establish a database of mapping relationships between flight attitude and noise characteristics. Attached Figure Description

[0025] Figure 1 This is a logic diagram of a method for separating complex noise sources; Figure 2 A schematic diagram of the architecture of the electronic device provided by the present invention.

[0026] Reference numerals: Electronic device 500, processor 501, communication interface 502, memory 503, bus 504. Detailed Implementation

[0027] To make the technical solution and advantages of this application clearer, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. It is understood that the specific embodiments described herein are only some embodiments of the present invention, and are only used to explain this application, not to limit it. It should be noted that the technical features or combinations of technical features described in the following embodiments should not be considered isolated; they can be combined with each other to achieve better technical effects. The same reference numerals appearing in the accompanying drawings of the following embodiments represent the same features or components, and can be applied to different embodiments.

[0028] Furthermore, unless otherwise defined, the technical or scientific terms used in this invention description shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains.

[0029] The present invention will now be described in further detail with reference to the accompanying drawings: Reference Figure 1 A method for separating complex noise sources includes the following steps.

[0030] Step S100, Test Preparation, includes: Step S101, sensor arrangement, including: Sensor placement in the driver's cab: A three-dimensional coordinate H-dummy is placed in the cockpit to simulate the head positions of the driver and passengers. The H-dummy should conform to ISO or SAE standards, have three-dimensional coordinate positioning capabilities, and be placed in the cockpit at the head positions of the driver's seat and passenger seat, respectively.

[0031] Omnidirectional microphones are placed near the left and right ears of the H-dummy to measure the noise levels in the ears of the driver and passenger. The omnidirectional microphones are about 2 cm away from the ears, and are of the omnidirectional type with a frequency response range of 20Hz - 20kHz. The number of microphones for the driver dummy is preferably 2 (1 for the left ear and 1 for the right ear), and the number of microphones for the passenger dummy is preferably 2 (1 for the left ear and 1 for the right ear).

[0032] An omnidirectional microphone is positioned at the center of the top of the cockpit, approximately 50 centimeters from the dummy's head, to measure the overall noise level within the cockpit. The preferred frequency response range of the omnidirectional microphone is 20Hz - 20kHz, and one microphone is preferred.

[0033] An omnidirectional microphone is positioned approximately 20 centimeters from the feet of the dummy in the center of the cockpit floor to measure noise from the floor and chassis. The preferred frequency response range of the omnidirectional microphone is 20Hz - 20kHz, and one microphone is preferred.

[0034] An omnidirectional microphone is positioned approximately 30 centimeters in front of the dummy in the center of the dashboard to measure noise from the dashboard and electronic devices. The preferred frequency response range of the omnidirectional microphone is 20Hz - 20kHz, and the preferred number of microphones is one.

[0035] External sensor arrangement for flying car: Microphone arrays are deployed near the rotor, engine compartment, and fuselage surface of the flying car to measure external noise. The microphone arrays should have high sensitivity and wide bandwidth response characteristics to accurately collect noise signals from outside the flying car.

[0036] Simultaneously, omnidirectional microphones are deployed in the test area and / or outside the flying car to measure ambient noise, record the time-domain signal and spectral characteristics of the background noise, and generate a background noise database.

[0037] The microphone array is preferably a spherical or circular array.

[0038] Connect all sensors to the data acquisition system, ensuring a stable and reliable connection. Use standard signal transmission cables to avoid signal interference and attenuation.

[0039] Calibrate the microphone and sensors to ensure measurement accuracy. The calibration process should be performed according to relevant standards, using standard sound sources and calibration equipment to calibrate the microphone's sensitivity and frequency response.

[0040] Step S200, Test Implementation, includes: Step S201, data acquisition, including: The flying car is activated, and during flight, gyroscopes and accelerometers are used to collect real-time data on the car's acceleration direction and motion status. The gyroscopes should have high precision and a high sampling rate to accurately measure the flying car's angular velocity and attitude changes; the accelerometers should have high sensitivity and a wide range to accurately measure the flying car's linear acceleration.

[0041] Noise data of the flying car, including rotor noise, engine noise, and wind noise, is collected through a microphone array.

[0042] Step S202, perform working condition identification, including: A flight condition dynamic identification system based on multi-sensor fusion (gyroscope, accelerometer, and barometer) is employed to establish a database mapping the relationship between flight attitude and noise characteristics. Specifically, this includes the following: By fusing sensor data using an extended Kalman filter (EKF) framework, parameters such as the flying car's attitude angle, velocity, and altitude are obtained. The state vector is designed as follows:

[0043] in, This is the roll angle. The pitch angle, Yaw angle , , These are the three-axis velocities. For height.

[0044] The prediction steps of the extended Kalman filter are as follows:

[0045]

[0046] The update steps for the extended Kalman filter are as follows:

[0047]

[0048]

[0049]

[0050]

[0051] in, Here is the state transition matrix. Let Q be the observation matrix, Q be the process noise covariance matrix, and R be the observation noise covariance matrix.

[0052] Based on the fused sensor data, the time-domain, frequency-domain, and spatial features of the flying car are extracted. The time-domain features include the standard deviation of acceleration (0.5s window) and angular velocity energy (20-100Hz band); the frequency-domain features include the dominant frequency component (rotor passing frequency and its harmonics) analyzed by FFT and the wavelet packet energy entropy (4-level decomposition); the spatial features include the radius of curvature of the trajectory and the climb / descent gradient.

[0053] The extracted features are input into a hybrid classification architecture for work condition classification. This architecture includes 1D CNN feature extraction, LSTM temporal modeling, and a random forest classifier. The specific process is as follows: The raw data is first processed by a 1D CNN to extract local features.

[0054] The raw data is input into an LSTM for time series modeling to capture the time dependencies of the data.

[0055] The features extracted by 1D CNN and LSTM are concatenated and input into a random forest classifier for working condition classification.

[0056] Step S202: Under different operating conditions of the flying car, the noise data of the flying car is recorded throughout the entire process using a microphone array and a data acquisition system. The data acquisition system should have a high sampling rate and high resolution, capable of accurately acquiring the noise signal of the flying car. The sampling rate should be no less than 48kHz, and the resolution should be no less than 16 bits. Simultaneously, the operating condition information of the flying car, such as takeoff, hovering, landing, and translation, is recorded for subsequent classification and analysis of the noise data.

[0057] Step S300, noise separation, includes: Step S301, data preprocessing, including: The acquired noise signals are subjected to noise reduction and normalization processing. Noise reduction processing employs filtering techniques, such as low-pass filtering and high-pass filtering, to remove high-frequency or low-frequency interference; normalization processing normalizes the multi-channel noise signals to ensure data consistency.

[0058] The short-time Fourier transform (STFT) is used to convert the time-domain signal into a time-frequency domain signal. The window function of the STFT can be selected from Hanning windows, flat-top windows, or Kaiser windows, and the window length can be adjusted according to the flight state. For example, in hovering, the window length can be set to a shorter time to focus on short-time stationarity analysis; in cruise, the window length can be set to a longer time to focus on fine spectral analysis.

[0059] Step S302, noise source separation, including: Step S302-1, preliminary separation based on independent component analysis (ICA), including: An improved FastICA algorithm is used for initial noise source separation. The optimization criteria for the FastICA algorithm are:

[0060] in, , These are standard Gaussian variables.

[0061] Flight parameter-guided initialization strategy: Initialize rotor noise components using rotor speed and engine noise components using engine operating conditions. For example, extract rotor speed using gyroscope Z-axis signals and establish a speed-noise level regression model.

[0062] in, The rotor speed, This represents the frequency at which the blades pass through. Engine noise initialization follows the same principle.

[0063] The analysis window length is dynamically adjusted based on flight altitude. Window length = base window length × (1 + altitude compensation coefficient × current altitude / maximum altitude). Nonlinear preprocessing is performed according to flight phase, such as focusing on high-frequency component enhancement during takeoff and full-band equalization during cruise.

[0064] Step S302-2, spatial optimization based on blind source separation (BSS), includes: A spatial optimization algorithm based on IVA is employed. The objective function for frequency domain joint diagonalization optimization is:

[0065] in, For the separation matrix, Let be the autocorrelation matrix of the signal.

[0066] Using the microphone array position as spatial prior information, a sound source-array distance constraint matrix is ​​established.

[0067] Combining TDOA (Time Difference of Arrival) and beamforming techniques, the probability distribution of sound source location is calculated:

[0068] Adaptive source number detection based on eigenvalue thresholding The number of sources is constrained according to the flight status, such as 3-4 main sources in ground mode and 5-6 main sources in flight mode.

[0069] Step S302-3, feature refinement and classification based on the CNN-RNN hybrid model, including: The front-end uses CNN feature extraction, with the input being a time-frequency image (STFT or wavelet transform result). The network structure includes 5 layers of depthwise separable convolutions, each layer is equipped with BatchNorm and LeakyReLU (α = 0.2), and a spatial attention mechanism is added to enhance key frequency bands.

[0070] The temporal RNN is used to process the data. The BiLSTM network structure is adopted, which is a bidirectional 3-layer LSTM with 256 hidden units. Skip connections are used to prevent gradient vanishing, and the attitude parameters are used as additional input gating to realize the working condition switching detection mechanism.

[0071] The hybrid connection strategy concatenates the CNN feature maps with the RNN hidden states and uses residual connections to bypass the noisy feature extraction path.

[0072] The training strategy includes multi-stage training. The first stage fixes the ICA-BSS parameters and trains the neural network; the second stage involves end-to-end joint fine-tuning. Simultaneously, flight data augmentation is performed, including noise simulation based on flight dynamics, such as rotor noise Doppler effect simulation, engine load-related harmonic generation, and environmental noise synthesis.

[0073] Step S400, environmental noise compensation, includes: Step S401: Utilize the time-domain signal and spectral characteristics of the background noise to generate a background noise database. During the measurement process, ensure that the microphone array is arranged reasonably to avoid interference from flying car noise.

[0074] Step S402, perform data processing, including: Step S401, Time Domain Analysis: Perform time domain analysis on the collected environmental noise and flying car noise data, calculate their mean, variance and other statistics to assess the overall level of environmental noise.

[0075] Step S402, frequency domain analysis: Use Short Time Fourier Transform (STFT) to convert the time-domain signal into a frequency-domain signal to generate a spectrum of environmental noise. Analyze the spectral characteristics to determine the main noise components and their frequency distribution.

[0076] Step S403, environmental noise compensation, includes: Step S403-1, compensation method matching: Based on the time-domain and frequency-domain analysis results of environmental noise and flying car noise, different compensation methods are selected. Specifically: When environmental noise and flying car noise do not overlap in the time domain, a time-domain compensation method is selected for matching. When environmental noise and flying car noise do not overlap in the frequency domain, a frequency domain compensation method is selected for matching. When there is some overlap between environmental noise and flying car noise in both the time and frequency domains, an adaptive filtering compensation method is selected.

[0077] Step S403-2, frequency domain compensation implementation steps, including: The mixed noise signal and the background noise signal are sampled to obtain discrete signal sequences x(n) and y(n), where n is the sampling point number.

[0078] Calculate the signal length N, ensuring x(n) and y(n) The lengths are consistent.

[0079] Perform time-domain alignment, that is x(n )and y(n) The starting points are aligned, and the alignment method can be calculated using a cross-correlation function to find the optimal time delay between the two.

[0080] By directly subtracting background noise, the true noise signal of the flying car is obtained. z(n)=x(n)-y(n) When subtracting, the consistency of signal amplitude and phase must be considered to avoid introducing new errors.

[0081] Step S403-3, frequency domain compensation implementation steps, including: For mixed noise signals x(n) and background noise signal y(n) STFT was performed separately to obtain the frequency domain signal. X(f,n) and Y(f,n) ,in f For frequency point number, n This is the frame number.

[0082] In the frequency domain, the spectral components of environmental noise are subtracted. During this subtraction, the consistency of spectral resolution and frequency response must be considered to ensure accuracy and obtain the true noise spectrum of the flying car. Z(f,n)=X(f,n)-Y(f,n) .

[0083] right Z(f,n) Perform inverse STFT to obtain the real noise signal of the flying car. z(n) .

[0084] The subtracted frequency domain signal is restored to the time domain signal through inverse short-time Fourier transform (ISTFT) to obtain the real noise data of the flying car.

[0085] Step S403-4, Adaptive filtering compensation implementation steps, including: Design adaptive filters, such as least mean square (LMS) filters or normalized least mean square (NLMS) filters.

[0086] Specifically, let the input signal be:

[0087] The filter coefficients are:

[0088] The filter output is then:

[0089] The error signal is:

[0090] in, This is the desired signal.

[0091] Initialize the filter coefficients to random or zero values, and set the filter length L and step size factor. The selection should be based on the signal characteristics. For flying car noise testing, the filter length can be set to 32, 64, or 128, and the step factor can be set to 0.01, 0.001, or 0.0001.

[0092] During testing, mixed noise signals and ambient noise signals are acquired in real time and input into an adaptive filter. The filter coefficients are updated according to the LMS algorithm to minimize the mean square error. The update rule is as follows:

[0093] in, This is the step size factor, which controls the convergence speed and stability. This is the filter coefficient vector at time n; This is the error signal (i.e., the difference between the desired signal and the filter output). The input signal vector.

[0094] The step size factor μ is dynamically adjusted based on the magnitude of the error signal. When the error signal is large, a larger μ is used to accelerate convergence; when the error signal is small, a smaller μ is used to improve stability. The dynamic adjustment strategy employs the Variable Step Size LMS (VSLMS) algorithm.

[0095] in, This is the initial step size factor. This is the error threshold.

[0096] Step S500, Result Analysis and Verification, includes: Step S501: Noise data analysis. This involves performing spectral analysis and source localization on the separated noise sources. Spectral analysis can be performed using methods such as Fast Fourier Transform (FFT) to generate a noise spectrum. Source localization can be achieved using techniques such as beamforming, combined with spatial information from the microphone array, to determine the location of the noise sources. The contribution of each noise source to the total noise is calculated, providing a basis for noise optimization. Contribution analysis can be achieved by calculating the proportion of each noise source's power in the total power.

[0097] Step S502 involves subjective and objective testing, simultaneously testing both the noise levels inside the driver's cab and external noise levels, combining subjective evaluation with objective data analysis. Subjective evaluation equipment includes questionnaires and a scoring system; after the test, the driver and passengers complete the subjective evaluation questionnaire. Evaluation indicators include noise intensity, noise frequency, comfort, and interference level, using a 5-point or 10-point scale to quantify the subjective feelings of the driver and passengers. Objective data analysis includes statistical analysis of data such as noise spectrum, sound source localization, and contribution analysis, generating an objective evaluation report.

[0098] Step S503, Verification and Calibration: Compare and analyze the noise data before and after compensation to verify the compensation effect. Generate noise spectrum and time-domain waveform diagrams, and use subjective evaluation and objective indicators (such as signal-to-noise ratio SNR) to verify the compensation effect. Simultaneously, regularly calibrate the sensors and testing equipment to ensure the accuracy and reliability of the test results.

[0099] This application also provides a composite noise source separation system, which utilizes the aforementioned composite noise source separation method.

[0100] This application also provides an electronic device 500 that utilizes the aforementioned method for separating complex noise sources. The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the aforementioned method for separating complex noise sources. In this application embodiment, the processor is the control center of the computer method and can be a physical machine processor or a virtual machine processor.

[0101] Reference Figure 2 The electronic device 500 includes at least one processor 501, at least one communication interface 502, at least one memory 503, and at least one bus 504. The bus 504 is used for communication between these components, the communication interface 502 is used for signaling or data communication with other node devices, and the memory 503 stores machine-readable instructions executable by the processor 501. When the electronic device 500 is running, the processor 501 communicates with the memory 503 via the bus 504, and when the machine-readable instructions are invoked by the processor 501, the steps of the composite noise source separation method described above are executed.

[0102] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor of an electronic device, can implement the steps of a composite noise source separation method as described above.

[0103] Those skilled in the art will understand that implementing all or part of the processes in a method for separating complex noise sources can be accomplished by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium. When executed, the program can include the processes of various embodiments of the method for separating complex noise sources. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0104] This embodiment provides a method for separating composite noise sources, specifically addressing the difficulty of boundary identification in topographic mapping scenarios. It achieves efficient and accurate triangulation network construction through inertial equipment combined with a dynamic insertion algorithm. The following detailed description of the specific implementation steps is provided: The above content is merely an embodiment of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, and are capable of applying conventional experimental methods prior to that date. Those skilled in the art can improve and implement this solution based on the guidance provided in this application and their own capabilities. Typical well-known structures or systems should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention. These modifications and improvements should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for separating complex noise sources, characterized in that, include: The acquired noise signal is processed by short-time Fourier transform to obtain the time-frequency domain signal; An improved FastICA algorithm is used for initial separation, and the rotor noise component is initialized using rotor speed and the engine noise component is initialized using engine operating conditions. Spatial optimization is performed using a frequency domain joint diagonalization algorithm based on Independent Vector Analysis (IVA). Combined with microphone array position information, TDOA technology, and beamforming technology, the probability distribution of sound source location is calculated, and the number of noise sources is adaptively detected. A CNN-RNN hybrid model is constructed to refine and classify the separated noise features. The CNN-RNN hybrid model adopts a depthwise separable convolution and BiLSTM network structure, and takes flight attitude parameters as additional input.

2. The method for separating a complex noise source according to claim 1, characterized in that, The acquired noise signals also undergo noise reduction and normalization processing; the noise reduction process employs filtering techniques, including low-pass filtering or high-pass filtering.

3. The method for separating a complex noise source according to claim 1, characterized in that, The optimization criteria for the FastICA algorithm are: in, , These are standard Gaussian variables.

4. The method for separating a complex noise source according to claim 2, characterized in that, Rotor speed is extracted using the Z-axis signal from the gyroscope, and a speed-noise level regression model is established: in, The rotor speed, The frequency at which the blade passes through.

5. The method for separating a complex noise source according to claim 1, characterized in that, Preliminary separation also includes: The analysis window length is dynamically adjusted based on flight altitude. Window length = Basic window length × (1 + Height compensation coefficient × Current height / Maximum height).

6. The method for separating a complex noise source according to claim 1, characterized in that, The IVA algorithm employs frequency domain joint diagonalization to optimize the objective function. in, For the separation matrix, Let be the autocorrelation matrix of the signal.

7. A method for separating a complex noise source according to claim 1 or 6, characterized in that, By combining microphone array location information, TDOA technology, and beamforming technology, the probability distribution of sound source locations is calculated, and the number of noise sources is adaptively detected, including the following: By combining the microphone array position as spatial prior information, a sound source-array distance constraint matrix is ​​established; The formula for calculating the probability distribution of sound source location is as follows: Adaptive source number detection based on eigenvalue thresholding The number of sources is constrained based on the flight status.

8. The method for separating a complex noise source according to claim 1, characterized in that, The CNN part of the CNN-RNN hybrid model includes 5 layers of depthwise separable convolutions, BatchNorm and LeakyReLU activation functions, and incorporates a spatial attention mechanism; The RNN part uses a bidirectional 3-layer LSTM with 256 hidden units and uses flight attitude parameters as additional input gating. The CNN feature map is concatenated with the RNN hidden state, and residual connections are used to bypass the noisy feature extraction path.

9. A method for separating complex noise sources according to claim 8, characterized in that, The training strategy includes multi-stage training. The first stage fixes the ICA-BSS parameters and trains the neural network. The second stage involves end-to-end joint fine-tuning.

10. The method for separating a complex noise source according to claim 1, characterized in that, It also includes the use of a flight condition dynamic identification system based on multi-sensor fusion to establish a database of mapping relationships between flight attitude and noise characteristics.

Images