Intelligent diagnosis method and system for potential faults of helicopter rotor based on acoustic signals

Abstract

The present application belongs to the technical field of helicopter fault diagnosis, and relates to a helicopter rotor potential fault intelligent diagnosis method and system based on acoustic signals. The method comprises: implanting typical fault characteristics into characteristic parameters to generate rotor models of multiple fault states; solving flow field parameters to obtain dynamic pressure loads on the surface of rotor blades; collecting time-varying sequence signals of acoustic pressure through a virtual acoustic sensor array; extracting fault characteristics; constructing a fault classification and discrimination model based on machine learning; and outputting the fault state and fault type of the rotor. The present application can comprehensively reflect the influence of rotor faults on acoustic radiation characteristics, effectively solve the problem of difficult acquisition of helicopter rotor fault samples through high-fidelity geometric modeling, fault implantation and CFD and CSD coupling strategies, and realize automatic identification and type positioning of faults through extraction of multi-dimensional fault characteristics and construction of a fault diagnosis model based on machine learning algorithms.

Description

Technical Field

[0001] This invention belongs to the field of helicopter fault diagnosis technology, and more specifically, relates to an intelligent diagnostic method and system for potential faults in helicopter rotors based on acoustic signals. Background Technology

[0002] As a core piece of equipment in general aviation, helicopters play an irreplaceable role in emergency rescue, medical transport, police patrol, agricultural and forestry operations, and offshore oil services due to their unique vertical takeoff and landing and hovering capabilities. However, helicopters typically operate in extremely harsh environments (such as strong winds, sandstorms, and high salt spray at sea), and their mechanical structures are extremely complex. The rotor system, in particular, as the helicopter's lifting and control surfaces, bears enormous alternating aerodynamic loads, centrifugal forces, and Coriolis forces, making it highly susceptible to fatigue cracks, mass imbalances, and loose connections. Statistics show that helicopter accident rates are far higher than those of fixed-wing aircraft, and rotor system failure is one of the main causes of catastrophic accidents.

[0003] Traditional helicopter maintenance primarily relies on scheduled maintenance and reactive maintenance. Scheduled maintenance presents a dilemma of over- or under-maintenance, while reactive maintenance often implies that an accident has already occurred. To meet the high-intensity, high-reliability operational demands of modern aviation, there is an urgent need to shift towards condition-based and predictive maintenance.

[0004] Currently, helicopter health and usage monitoring systems (HUMS) are widely used, primarily relying on vibration signal analysis. While vibration diagnostic technology is relatively mature, it has revealed significant limitations in practical applications:

[0005] (1) When using contact measurement, the vibration sensor (accelerometer) must be physically installed on the component being measured. For high-speed rotating rotor blades, installing the sensor not only requires complex slip ring actuators or wireless transmission systems, but may also change the mass distribution and aerodynamic shape of the blade, or even cause the sensor itself to fall off, resulting in safety hazards.

[0006] (2) The vibration signal generated by the fault needs to be transmitted through multiple layers of structure such as blades, hubs, main reducers, and fuselage before it can be received by the fuselage sensor. The complex path leads to severe attenuation of high frequency features and is easily interfered with by other vibration sources such as engines and transmission systems, resulting in low signal-to-noise ratio and difficulty in feature extraction.

[0007] (3) Vibration sensors can only reflect the local conditions near the installation point. They are not sensitive enough to early local faults (such as micro cracks) of large flexible rotor blades and have limited spatial coverage.

[0008] (4) Building a data-driven diagnostic model requires a large amount of sample data with fault labels. However, conducting destructive fault tests on real helicopters (such as artificially creating cracks during flight) is extremely risky and costly, which makes it difficult for existing diagnostic algorithms to acquire data.

[0009] Current technologies are limited to vibration signal analysis, failing to address the fundamental challenges of contact measurement and neglecting the enormous potential of acoustic signals in reflecting rotor aerodynamic characteristics and early structural changes. Acoustic signals, as a non-contact measurement method, contain rich global modal information, are extremely sensitive to aerodynamic shape changes (such as icing and deformation), and offer flexible and inexpensive sensor deployment, making them an important direction for next-generation helicopter health monitoring. Summary of the Invention

[0010] To address the aforementioned technical problems, this invention provides a method and system for intelligent diagnosis of potential faults in helicopter rotors based on acoustic signals.

[0011] In a first aspect, the present invention provides an intelligent diagnostic method for potential faults in helicopter rotors based on acoustic signals, comprising:

[0012] A three-dimensional geometric model of a helicopter rotor is constructed, and typical fault features are implanted into the feature parameters of the three-dimensional geometric model to generate rotor models with various fault states.

[0013] The dynamic rotation process of a rotor model under fault conditions is simulated using an overlapping mesh method. The flow field parameters are solved by a CFD and CSD coupling strategy to obtain the dynamic pressure load on the rotor blade surface.

[0014] Dynamic pressure loads are mapped onto an acoustic field, far-field radiated noise is calculated based on the FW-H acoustic analogy equation, and a sequence signal of sound pressure changing with time is collected through a virtual acoustic sensor array.

[0015] The sequence signal is preprocessed, and fault features are extracted using signal processing algorithms to obtain an acoustic feature matrix reflecting the rotor operating status, and the corresponding fault labels are recorded.

[0016] A fault classification and discrimination model based on machine learning is constructed. The model is trained using acoustic feature matrices and corresponding fault labels to establish a mapping relationship between acoustic features and fault modes.

[0017] The system collects real-time acoustic signals from the rotor of the helicopter to be diagnosed, inputs them into a trained fault classification and discrimination model, and outputs the fault status and fault type of the rotor.

[0018] Secondly, the present invention provides an intelligent diagnostic system for potential faults of helicopter rotors based on acoustic signals, including a geometric modeling and fault implantation unit, a coupled solution unit, an acoustic analysis unit, a feature extraction unit, a model building and training unit, and a fault diagnosis unit.

[0019] The geometric modeling and fault implantation unit is used to construct a three-dimensional geometric model of a helicopter rotor and implant typical fault features into the feature parameters of the three-dimensional geometric model to generate rotor models with various fault states.

[0020] The coupled solver unit is used to simulate the dynamic rotation process of the rotor model under fault conditions using the overlapping mesh method. The flow field parameters are solved by the CFD and CSD coupling strategy to obtain the dynamic pressure load on the rotor blade surface.

[0021] The acoustic analysis unit is used to map dynamic pressure loads onto the acoustic field, calculate far-field radiated noise based on the FW-H acoustic analogy equation, and collect time-varying sequence signals of sound pressure through a virtual acoustic sensor array.

[0022] The feature extraction unit is used to preprocess the sequence signal, extract fault features using signal processing algorithms, obtain an acoustic feature matrix reflecting the rotor operating status, and record the corresponding fault labels.

[0023] The model building and training unit is used to build a fault classification and discrimination model based on machine learning. It uses the acoustic feature matrix and the corresponding fault labels to train the fault classification and discrimination model and establishes the mapping relationship between acoustic features and fault modes.

[0024] The fault diagnosis unit is used to collect real-time acoustic signals from the rotor of the helicopter to be diagnosed, input the trained fault classification and discrimination model, and output the fault status and fault type of the rotor.

[0025] Based on the above technical solution, the present invention can be further improved as follows.

[0026] Furthermore, the characteristic parameters include blade properties, blade rotation speed, blade diameter, blade chord length, basic blade pass frequency, number of blades, and blade torsional distribution.

[0027] Furthermore, typical fault characteristics include blade cracks, counterweight imbalance, and rotor blade icing. Blade cracks are simulated by setting a pre-defined depth and width of cuts at the blade's span. Material parameter corrections or stiffness degradation modeling are performed on the cut area in the CSD model to reflect the structural dynamics effect of blade cracks. Counterweight imbalance is simulated by adding a mass block of a set size to the blade tip.

[0028] Furthermore, the flow field parameters are solved using a CFD and CSD coupling strategy to obtain the dynamic pressure loads on the rotor blade surface, including:

[0029] Establish CFD solution model and CSD calculation model;

[0030] Calculate the CSD computational model and obtain the initial deformation data of the rotor blades;

[0031] Input the initial deformation data into the CFD solution model to update the blade mesh;

[0032] The CFD solution model calculates the flow field parameters, obtains the dynamic pressure load, and returns it to the CSD calculation model;

[0033] Iterative calculations are performed until the rotor blade structural deformation and dynamic pressure load converge.

[0034] Furthermore, an overlapping mesh method is used to simulate the dynamic rotation process of the rotor model under fault conditions, including: the overlapping mesh includes a background mesh and a component mesh; the background mesh is a fixed mesh covering the entire computational domain; the component mesh is a moving mesh that wraps around the rotor blades; the moving mesh rotates with the blades; the overlapping area between the component mesh and the background mesh is set to a set ratio of the rotor diameter; and linear interpolation is used for data transfer between the component mesh and the background mesh.

[0035] Furthermore, when solving the flow field parameters using a CFD and CSD coupling strategy, the LES model is used to solve the turbulent structure in the core fluid, and the RANS model is used to cover the wall boundary layer. The governing equations of the RANS model are in differential form, and the set time step satisfies the Nyquist sampling theorem and the dynamic mesh time scale relationship. For density, For average speed, For average pressure, For time, For unit tensors, The resultant force of volume forces, For the mean viscous stress tensor, Let Reynolds stress tensor be the stress tensor. The total energy of filtering per unit mass. For filtering heat flux, If we use a vector differential operator, then the differential form of the governing equations of the RANS model is expressed as:

[0036] ;

[0037] ;

[0038] .

[0039] Furthermore, the sequence signal is preprocessed, including: using wavelet soft thresholding to denoise the sequence signal, normalizing the denoised sequence signal, and segmenting the acoustic signal by time.

[0040] Furthermore, signal processing algorithms are used to extract fault features and obtain an acoustic feature matrix reflecting the rotor's operating state. This includes: using fast Fourier transform to convert the time-domain signal into a frequency-domain signal and analyzing the rotor's passing frequency and energy distribution; using wavelet packet transform to filter non-feature frequencies and extracting the energy, entropy value, and peak factor of each frequency band as fault-sensitive feature quantities, thus converting the acoustic data into a fault feature matrix.

[0041] Furthermore, the machine learning model includes one or more combinations of support vector machines, random forests, convolutional neural networks, and recurrent neural networks; the parameters of the machine learning model are adjusted using Adam or SGD optimization algorithms.

[0042] The beneficial effects of this invention are:

[0043] (1) The present invention uses acoustic signals as a diagnostic information source. Compared with traditional vibration signals, it can more comprehensively reflect the impact of rotor failure on acoustic radiation characteristics and improve the characterization ability of early failures and aerodynamic failures. At the same time, the acoustic sensor is non-contact installed, which avoids the installation problem of traditional vibration sensors on rotating parts and reduces the complexity and cost of the sensing system.

[0044] (2) This invention, through high-fidelity geometric modeling and fault implantation, as well as the CFD and CSD coupling strategy, can generate a large amount of acoustic signal data with clear fault labels without relying on real fault tests. This effectively solves the problem of difficulty in obtaining helicopter rotor fault samples and provides a sufficient data foundation for training fault diagnosis models.

[0045] (3) This invention extracts multi-dimensional fault features in the time domain, frequency domain and time-frequency domain to construct a highly recognizable fault feature matrix, and combines it with machine learning algorithms to construct a fault diagnosis model. The diagnosis accuracy is high, and it can realize automatic fault identification and type positioning, thus meeting the needs of predictive maintenance of helicopters. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the intelligent diagnostic method for potential helicopter rotor faults based on acoustic signals provided in Embodiment 1 of the present invention.

[0047] Figure 2 This is a schematic diagram of the structural principle of the intelligent diagnostic system for potential faults in helicopter rotors based on acoustic signals provided in Embodiment 1 of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0049] Example 1

[0050] As an example, see the attached document. Figure 1 As shown, to solve the above-mentioned technical problems, this embodiment provides an intelligent diagnosis method for potential helicopter rotor faults based on acoustic signals, including the following steps:

[0051] Step 110: Construct a three-dimensional geometric model of the helicopter rotor, and implant typical fault features into the feature parameters of the three-dimensional geometric model to generate rotor models with multiple fault states.

[0052] In some optional embodiments, a high-fidelity geometric model of the helicopter rotor is created using ANSYS 3D modeling software.

[0053] Optional characteristic parameters include blade properties, blade rotation speed, blade diameter, blade chord length, basic blade pass frequency, number of blades, and blade torsional distribution.

[0054] Taking the Robinson R44 Raven II helicopter as an example, the rotor blade parameters are shown in Table 1.

[0055] Table 1 Blade Parameters

[0056]

[0057] While maintaining the rotor airfoil, geometric defects of different scales are constructed at specific locations on the blades by modifying characteristic parameters. For example, a slit with a depth of 'd' and a width of 'l' is set at one-third of the blade's span to simulate a narrow crack, and a mass block is added at the blade tip to simulate uneven counterweight or rotor blade icing. High-density body-fitted meshes are used for the rotor surface and near-field region to capture high-frequency acoustic components.

[0058] Optionally, typical fault characteristics include blade cracks, counterweight imbalance, and rotor blade icing. Blade cracks are simulated by setting a pre-defined depth and width of notch at the blade's span. The material parameters of the notch area are corrected or stiffness degradation is modeled in the CSD model to reflect the structural dynamics effect of the blade crack. Counterweight imbalance is simulated by adding a mass block of a predetermined size to the blade tip. For each typical fault, hierarchical modeling is performed based on the fault severity, spatial location, and dimensional parameters to form multiple sub-fault states, while the fault-free normal state is treated as a separate state.

[0059] Weight imbalance faults are mainly caused by uneven mass distribution of the rotor blades, blade icing, and loose components. These faults generate additional centrifugal force during rotor rotation, leading to abnormal vibration and noise. Rotor blade icing is a common fault for helicopters operating in low-temperature, high-humidity environments. It alters the aerodynamic shape and mass distribution of the blades, seriously affecting flight safety.

[0060] Optionally, for crack failures, the constitutive relation of the material can be modified by introducing a local damage factor, thereby locally weakening the material parameters or equivalent stiffness, and thus changing the CSD (Computational Structural Dynamics) mode and response of the structure. As an alternative implementation, a geometric gap can be directly pre-set in the structural model, that is, a piece of the structure can be directly removed from the geometric model, and the characteristics of the real crack can be simulated through geometric discontinuity.

[0061] Step 120: The dynamic rotation process of the rotor model under fault conditions is simulated using the overlapping mesh method. The flow field parameters are solved by the CFD (Computational Fluid Dynamics) and CSD coupling strategy to obtain the dynamic pressure load on the rotor blade surface.

[0062] Optionally, an overlapping mesh method is used to simulate the dynamic rotation process of the rotor model under fault conditions, including: the overlapping mesh includes a background mesh and a component mesh; the background mesh is a fixed mesh covering the entire computational domain; the component mesh is a moving mesh that wraps around the rotor blades; the moving mesh rotates with the blades; the overlapping area between the component mesh and the background mesh is set to a set ratio of the rotor diameter; and linear interpolation is used for data transfer between the component mesh and the background mesh.

[0063] Optionally, the flow field parameters can be solved using a CFD and CSD coupling strategy to obtain the dynamic pressure loads on the rotor blade surface, including:

[0064] A CFD solution model and a CSD calculation model were established. A loose coupling strategy between CFD and CSD was applied to achieve information transfer between the fluid and the structure. Specifically, a structural dynamics model of the rotor blade was established using ANSYS Mechanical (the module for structural mechanics analysis in the ANSYS simulation platform), with the blade material parameters set to the performance parameters of glass fiber reinforced epoxy resin. The first 10 natural frequencies and mode shapes of the blade were obtained through modal analysis and used as the initial parameters for the CSD calculation model.

[0065] Calculate the CSD calculation model and obtain the initial deformation data of the rotor blades; specifically, apply centrifugal force load (generated by rotor rotation) to the CSD calculation model and calculate the initial deformation data of the blades, including displacement and rotation along the span and chord direction;

[0066] Input the initial deformation data into the CFD solution model to update the blade mesh; import the blade deformation data calculated by the CSD calculation model into ANSYS Fluent (fluid solver), and update the shape of the component mesh through mesh deformation to make the blade mesh consistent with the deformed blade structure;

[0067] The CFD solution model calculates the flow field parameters, obtains the dynamic pressure load, and returns it to the CSD calculation model. The LES-RANS hybrid turbulence model (Large Eddy Simulation-Reynolds-Averaged Navier-Stokes Hybrid Turbulence Model) is used to solve the flow field parameters. The core fluid region (the region far from the blade surface) uses the LES model, which can accurately capture the turbulent fluctuation characteristics. The wall boundary layer region (the region less than a set distance from the blade surface) uses the RANS model to reduce the computational load.

[0068] Iterative calculations are performed until the rotor blade structural deformation and dynamic pressure load converge.

[0069] Based on the Nyquist sampling theorem and the relationship between the time scale of the moving mesh, the time step is set as follows: , To analyze the highest frequency and capture the surface aerodynamic forces during rotor rotation, the following conditions must be met:

[0070] .

[0071] Optionally, when solving for the flow field parameters using a CFD and CSD coupling strategy, the LES model is used to solve for the turbulent structure in the core fluid, and the RANS model is used to cover the wall boundary layer; the governing equations of the RANS model are in differential form, and the set time step satisfies the Nyquist sampling theorem and the dynamic mesh time scale relationship; assuming For density, For average speed, For average pressure, For unit tensors, The resultant force of volume forces, For the mean viscous stress tensor, Let Reynolds stress tensor be the stress tensor. The total energy of filtering per unit mass. For the filtered heat flux, the differential form of the governing equations of the RANS model is expressed as:

[0072] ;

[0073] ;

[0074] .

[0075] Step 130: Map the dynamic pressure load onto the acoustic field, calculate the far-field radiated noise based on the FW-H acoustic analogy equation, and collect the time-varying sequence signal of sound pressure through a virtual acoustic sensor array.

[0076] Using the rotor blade surface as an impermeable integral surface, the extracted dynamic pressure load is mapped onto the acoustic field. Based on the FW-H acoustic analogy equation, the far-field radiated noise generated by rotational motion and aerodynamic load disturbance is calculated. At the same time, the position of the virtual acoustic sensor array is selected in the simulation space, and finally the sequence signal of the sound pressure received by the sensor under various working conditions as a function of time is obtained.

[0077] In practical applications, to comprehensively collect the radiated noise signal of the rotor, a virtual acoustic sensor array is arranged in the acoustic simulation calculation domain. The array is a ring array, and the specific parameters are as follows:

[0078] (1) Array radius: 10m (distance from the rotor rotation center). This distance ensures that the sensor is in the far field region and can capture clear fault characteristic signals.

[0079] (2) Number of sensors: 16, evenly distributed on a ring array, with an angle of 22.5° between adjacent sensors;

[0080] (3) Sensor height: flush with the rotor rotation plane to ensure that the sensor can receive the main radiated noise generated by the rotor rotation;

[0081] (4) Sampling parameters: sampling frequency 48kHz, sampling duration 10s, number of sampling points 480,000, sampling accuracy 16-bit, to ensure that high-frequency fault characteristic signals can be captured.

[0082] The sound pressure sequence signals acquired by the virtual acoustic sensor array are stored in WAV or CSV format. Each sensor corresponds to a set of sound pressure time series data, including sound signal data under normal conditions and different fault conditions.

[0083] Furthermore, considering that the acoustic signals generated by numerical simulation methods may differ from the rotor acoustic signals collected under actual flight or test conditions in terms of environmental noise, sound transmission path, and system response, this invention introduces consistency calibration and domain adaptation processing steps between simulated and measured acoustic signals to improve the robustness and applicability of the acoustic signal-based fault diagnosis model in practical applications:

[0084] 1) In the implementation, the acoustic sensors used to collect rotor acoustic signals are calibrated, including but not limited to microphone sensitivity calibration, frequency response compensation and phase consistency correction, in order to reduce the influence of different sensors and acquisition systems on the amplitude and spectral characteristics of the acoustic signals.

[0085] 2) Environmental noise suppression processing is performed on the measured sound signal. By using spectral subtraction, adaptive filtering, or array-based beamforming methods, the interference of background noise, wind noise, airframe scattering sound, and non-target sound sources on the rotor sound signal is reduced, making it closer to the simulated sound signal in statistical characteristics.

[0086] 3) In the feature extraction stage, the simulated sound signal and the measured sound signal are normalized, for example, based on the order analysis of the rotor speed, the blade is normalized by frequency alignment or feature scale, so as to reduce the difference in the distribution of sound signal features under different operating conditions.

[0087] Step 140: Preprocess the sequence signal, extract fault features using signal processing algorithms, obtain an acoustic feature matrix reflecting the rotor operating state, and record the corresponding fault labels.

[0088] Optionally, wavelet soft thresholding denoising method is used to denoise the sequence signal, the denoised sequence signal is normalized, and the acoustic signal is segmented by time.

[0089] Optionally, signal processing algorithms are used to extract fault features to obtain an acoustic feature matrix reflecting the rotor's operating state. This includes: using fast Fourier transform to convert the time-domain signal into a frequency-domain signal and analyzing the rotor's passing frequency and energy distribution; using wavelet packet transform to filter non-feature frequencies and extracting the energy, entropy, and peak factor of each frequency band as fault-sensitive features, thus converting the acoustic data into a fault feature matrix.

[0090] Step 150: Construct a fault classification and discrimination model based on machine learning. Train the fault classification and discrimination model using the acoustic feature matrix and the corresponding fault labels to establish the mapping relationship between acoustic features and fault modes.

[0091] Optionally, the machine learning model includes one or more combinations of support vector machines, random forests, convolutional neural networks, and recurrent neural networks; the parameters of the machine learning model are adjusted using Adam (Adaptive Moment Estimation) or SGD (Stochastic Gradient Descent) optimization algorithms.

[0092] The specific model structure is as follows:

[0093] (1) Support Vector Machine: The kernel function is a radial basis function with a parameter of 0.01; the penalty coefficient is 10, which is determined by grid search. Multi-class strategy: One-to-one strategy, which transforms the 10-class classification problem into 45 binary classification problems;

[0094] (2) Random Forest Model: Number of decision trees: 100; Maximum depth of each decision tree: 15 layers; Number of split features per decision tree: 8; Minimum number of splits per sample: 2; Minimum number of leaf nodes per sample: 1;

[0095] (3) Convolutional Neural Network: Input layer: 78-dimensional feature vector, reshaped as a one-dimensional sequence of (78,1); Convolutional layer 1: 16 convolutional kernels, kernel size 3×1, stride 1, activation function ReLU; Pooling layer 1: max pooling, pooling size 2×1, stride 2; Convolutional layer 2: 32 convolutional kernels, kernel size 3×1, length 1, activation function ReLU; Pooling layer 2: max pooling, pooling size 2×1, stride 2; Fully connected layer 1: 128 neurons, activation function ReLU, dropout rate 0.5; Fully connected layer 2: 64 neurons, activation function ReLU, dropout rate 0.5; Output layer: 10 neurons, activation function Softmax, corresponding to 10 fault labels.

[0096] (4) Recurrent Neural Network: Input layer: 78-dimensional feature vector, reshaped as a one-dimensional sequence of (78,1); LSTM layer 1: 64 neurons, dropout rate of 0.3; LSTM layer 2: 32 neurons, dropout rate of 0.3; Fully connected layer 1: 64 neurons, activation function is ReLU, dropout rate of 0.5; Output layer: 10 neurons, activation function is Softmax, corresponding to 10 fault labels.

[0097] Model training:

[0098] Support Vector Machine: Using the SVC class from Scikit-learn, with the training duration set;

[0099] Random Forest Model: Uses a random forest classifier for classification, and sets the training duration;

[0100] Convolutional Neural Network: The optimizer uses Adam with a learning rate of 0.001, and the loss function is calculated;

[0101] Recurrent Neural Network: The optimizer uses Adam with a learning rate of 0.001. The loss function is calculated, and the training time is set.

[0102] The performance of the four trained models was evaluated using test set data, with evaluation metrics including accuracy, precision, recall, and F1 score. In practical applications, the recurrent neural network demonstrated the best overall performance and the highest accuracy, meeting the precision requirements for helicopter rotor fault diagnosis.

[0103] Step 160: Collect real-time acoustic signals from the rotor of the helicopter to be diagnosed, input them into the trained fault classification and discrimination model, and output the fault status and fault type of the rotor.

[0104] The proposed intelligent diagnostic method for potential helicopter rotor faults based on acoustic signals achieves non-contact and high-precision diagnosis of helicopter rotor faults through high-fidelity geometric modeling and fault implantation, CFD and CSD coupling strategies, multi-dimensional feature extraction, and machine learning model construction. This method has the following core advantages:

[0105] (1) The present invention uses acoustic signals as a diagnostic information source. Compared with traditional vibration signals, it can more comprehensively reflect the impact of rotor failure on acoustic radiation characteristics and improve the characterization ability of early failures and aerodynamic failures. At the same time, the acoustic sensor is non-contact installed, which avoids the installation problem of traditional vibration sensors on rotating parts and reduces the complexity and cost of the sensing system.

[0106] (2) This invention, through high-fidelity geometric modeling and fault implantation, as well as the CFD and CSD coupling strategy, can generate a large amount of acoustic signal data with clear fault labels without relying on real fault tests. This effectively solves the problem of difficulty in obtaining helicopter rotor fault samples and provides a sufficient data foundation for training fault diagnosis models.

[0107] (3) This invention extracts multi-dimensional fault features in the time domain, frequency domain and time-frequency domain to construct a highly recognizable fault feature matrix, and combines it with machine learning algorithms to construct a fault diagnosis model. The diagnosis accuracy is high, and it can realize automatic fault identification and type positioning, thus meeting the needs of predictive maintenance of helicopters.

[0108] Example 2

[0109] Based on the same principle as the method shown in Embodiment 1 of the present invention, as illustrated in the appendix. Figure 2 As shown, the embodiments of the present invention also provide an intelligent diagnostic system for potential faults of helicopter rotors based on acoustic signals, including a geometric modeling and fault implantation unit, a coupled solution unit, an acoustic analysis unit, a feature extraction unit, a model building and training unit, and a fault diagnosis unit;

[0110] The geometric modeling and fault implantation unit is used to construct a three-dimensional geometric model of a helicopter rotor and implant typical fault features into the feature parameters of the three-dimensional geometric model to generate rotor models with various fault states.

[0111] The coupled solver unit is used to simulate the dynamic rotation process of the rotor model under fault conditions using the overlapping mesh method. The flow field parameters are solved by the CFD and CSD coupling strategy to obtain the dynamic pressure load on the rotor blade surface.

[0112] The acoustic analysis unit is used to map dynamic pressure loads onto the acoustic field, calculate far-field radiated noise based on the FW-H acoustic analogy equation, and collect time-varying sequence signals of sound pressure through a virtual acoustic sensor array.

[0113] The feature extraction unit is used to preprocess the sequence signal, extract fault features using signal processing algorithms, obtain an acoustic feature matrix reflecting the rotor operating status, and record the corresponding fault labels.

[0114] The model building and training unit is used to build a fault classification and discrimination model based on machine learning. It uses the acoustic feature matrix and the corresponding fault labels to train the fault classification and discrimination model and establishes the mapping relationship between acoustic features and fault modes.

[0115] The fault diagnosis unit is used to collect real-time acoustic signals from the rotor of the helicopter to be diagnosed, input the trained fault classification and discrimination model, and output the fault status and fault type of the rotor.

[0116] Optional characteristic parameters include blade properties, blade rotation speed, blade diameter, blade chord length, basic blade pass frequency, number of blades, and blade torsional distribution.

[0117] Optional, typical fault characteristics include blade cracks, counterweight imbalance, and rotor blade icing; blade cracks are simulated by setting a pre-defined depth and width of cuts at the blade span, and the material parameters of the cut area are corrected or stiffness degradation is modeled in the CSD model to reflect the structural dynamics effect of blade cracks; counterweight imbalance is simulated by adding a mass block of a set size to the blade tip.

[0118] Optionally, the flow field parameters can be solved using a CFD and CSD coupling strategy to obtain the dynamic pressure loads on the rotor blade surface, including:

[0119] Establish CFD solution model and CSD calculation model;

[0120] Calculate the CSD computational model and obtain the initial deformation data of the rotor blades;

[0121] Input the initial deformation data into the CFD solution model to update the blade mesh;

[0122] The CFD solution model calculates the flow field parameters, obtains the dynamic pressure load, and returns it to the CSD calculation model;

[0123] Iterative calculations are performed until the rotor blade structural deformation and dynamic pressure load converge.

[0124] Optionally, an overlapping mesh method is used to simulate the dynamic rotation process of the rotor model under fault conditions, including: the overlapping mesh includes a background mesh and a component mesh; the background mesh is a fixed mesh covering the entire computational domain; the component mesh is a moving mesh that wraps around the rotor blades; the moving mesh rotates with the blades; the overlapping area between the component mesh and the background mesh is set to a set ratio of the rotor diameter; and linear interpolation is used for data transfer between the component mesh and the background mesh.

[0125] Optionally, when solving for the flow field parameters using a CFD and CSD coupling strategy, the LES model is used to solve for the turbulent structure in the core fluid, and the RANS model is used to cover the wall boundary layer; the governing equations of the RANS model are in differential form, and the set time step satisfies the Nyquist sampling theorem and the dynamic mesh time scale relationship; assuming For density, For average speed, For average pressure, For unit tensors, The resultant force of volume forces, For the mean viscous stress tensor, Let Reynolds stress tensor be the stress tensor. The total energy of filtering per unit mass. For the filtered heat flux, the differential form of the governing equations of the RANS model is expressed as:

[0126] ;

[0127] ;

[0128] .

[0129] Optionally, the sequence signal is preprocessed, including: denoising the sequence signal using a wavelet soft thresholding method, normalizing the denoised sequence signal, and segmenting the acoustic signal by time.

[0130] Optionally, signal processing algorithms are used to extract fault features to obtain an acoustic feature matrix reflecting the rotor's operating state. This includes: using fast Fourier transform to convert the time-domain signal into a frequency-domain signal and analyzing the rotor's passing frequency and energy distribution; using wavelet packet transform to filter non-feature frequencies and extracting the energy, entropy, and peak factor of each frequency band as fault-sensitive features, thus converting the acoustic data into a fault feature matrix.

[0131] Optionally, the machine learning model includes one or more combinations of support vector machines, random forests, convolutional neural networks, and recurrent neural networks; the parameters of the machine learning model are adjusted using the Adam or SGD optimization algorithm.

[0132] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent diagnosis of potential helicopter rotor faults based on acoustic signals, characterized in that, include: A three-dimensional geometric model of a helicopter rotor is constructed, and typical fault features are implanted into the feature parameters of the three-dimensional geometric model to generate rotor models with various fault states. The dynamic rotation process of a rotor model under fault conditions is simulated using an overlapping mesh method. The flow field parameters are solved using a CFD and CSD coupling strategy to obtain the dynamic pressure loads on the rotor blade surface. The process includes: establishing a CFD solution model and a CSD calculation model; calculating the CSD calculation model and obtaining the initial deformation data of the rotor blade; inputting the initial deformation data into the CFD solution model to update the blade mesh; calculating the flow field parameters using the CFD solution model to obtain the dynamic pressure loads and returning them to the CSD calculation model; and iterative calculation until the rotor blade structural deformation and dynamic pressure loads converge. Dynamic pressure loads are mapped onto an acoustic field, far-field radiated noise is calculated based on the FW-H acoustic analogy equation, and a sequence signal of sound pressure changing with time is collected through a virtual acoustic sensor array. The sequence signal is preprocessed, and fault features are extracted using signal processing algorithms to obtain an acoustic feature matrix reflecting the rotor operating status, and the corresponding fault labels are recorded. A fault classification and discrimination model based on machine learning is constructed. The model is trained using acoustic feature matrices and corresponding fault labels to establish a mapping relationship between acoustic features and fault modes. The system collects real-time acoustic signals from the rotor of the helicopter to be diagnosed, inputs them into a trained fault classification and discrimination model, and outputs the fault status and fault type of the rotor.

2. The intelligent diagnostic method for potential helicopter rotor faults based on acoustic signals according to claim 1, characterized in that, The characteristic parameters include blade properties, blade rotation speed, blade diameter, blade chord length, basic blade passing frequency, number of blades, and blade torsional distribution.

3. The intelligent diagnostic method for potential helicopter rotor faults based on acoustic signals according to claim 1, characterized in that, Typical failure characteristics include blade cracks, counterweight imbalance, and rotor blade icing. Blade cracks are simulated by setting a pre-defined depth and width of cuts at the blade's span. The material parameters of the cut area are corrected or stiffness degradation is modeled in the CSD model to reflect the structural dynamics effect of blade cracks. Counterweight imbalance is simulated by adding a mass block of a set size to the blade tip.

4. The intelligent diagnostic method for potential helicopter rotor faults based on acoustic signals according to claim 1, characterized in that, The dynamic rotation process of a rotor model under fault conditions is simulated using an overlapping mesh method, including: the overlapping mesh consists of a background mesh and component meshes; the background mesh is a fixed mesh covering the entire computational domain; the component mesh is a moving mesh that wraps around the rotor blades; the moving mesh rotates with the blades; the overlapping area between the component mesh and the background mesh is set to a predetermined ratio of the rotor diameter; and linear interpolation is used for data transfer between the component mesh and the background mesh.

5. The intelligent diagnostic method for potential helicopter rotor faults based on acoustic signals according to claim 1, characterized in that, When solving for flow field parameters using a CFD and CSD coupling strategy, the LES model is used to solve for the turbulent structure in the core fluid, and the RANS model is used to cover the wall boundary layer. The governing equations of the RANS model are in differential form, and the set time step satisfies the Nyquist sampling theorem and the dynamic mesh time scale relationship. For density, For average speed, For average pressure, For unit tensors, The resultant force of volume forces, For the mean viscous stress tensor, Let Reynolds stress tensor be the stress tensor. The total energy of filtering per unit mass. For the filtered heat flux, the differential form of the governing equations of the RANS model is expressed as: ； ； 。 6. The intelligent diagnostic method for potential helicopter rotor faults based on acoustic signals according to claim 1, characterized in that, The sequence signal is preprocessed, including: using wavelet soft thresholding to denoise the sequence signal, normalizing the denoised sequence signal, and segmenting the acoustic signal by time.

7. The intelligent diagnostic method for potential helicopter rotor faults based on acoustic signals according to claim 1, characterized in that, The fault features are extracted using signal processing algorithms to obtain an acoustic feature matrix reflecting the rotor's operating state. This includes: converting the time-domain signal into a frequency-domain signal using fast Fourier transform to analyze the rotor's passing frequency and energy distribution; filtering non-feature frequencies using wavelet packet transform to extract the energy, entropy, and peak factor of each frequency band as fault-sensitive features, and converting the acoustic data into a fault feature matrix.

8. The intelligent diagnostic method for potential helicopter rotor faults based on acoustic signals according to claim 1, characterized in that, Machine learning models include one or more combinations of support vector machines, random forests, convolutional neural networks, and recurrent neural networks; the parameters of the machine learning model are adjusted using Adam or SGD optimization algorithms.

9. An intelligent diagnostic system for potential helicopter rotor faults based on acoustic signals, characterized in that, It includes a geometric modeling and fault implantation unit, a coupled solution unit, an acoustic analysis unit, a feature extraction unit, a model building and training unit, and a fault diagnosis unit; The geometric modeling and fault implantation unit is used to construct a three-dimensional geometric model of a helicopter rotor and implant typical fault features into the feature parameters of the three-dimensional geometric model to generate rotor models with various fault states. The coupled solver unit is used to simulate the dynamic rotation process of a rotor model under fault conditions using an overlapping mesh method. It solves for flow field parameters through a CFD and CSD coupling strategy to obtain the dynamic pressure loads on the rotor blade surface. The process includes: establishing the CFD solver model and the CSD calculation model; calculating the CSD calculation model and obtaining the initial deformation data of the rotor blade; inputting the initial deformation data into the CFD solver model to update the blade mesh; calculating the flow field parameters using the CFD solver model to obtain the dynamic pressure loads and returning them to the CSD calculation model; and iteratively calculating until the rotor blade structural deformation and dynamic pressure loads converge. The acoustic analysis unit is used to map dynamic pressure loads onto the acoustic field, calculate far-field radiated noise based on the FW-H acoustic analogy equation, and collect time-varying sequence signals of sound pressure through a virtual acoustic sensor array. The feature extraction unit is used to preprocess the sequence signal, extract fault features using signal processing algorithms, obtain an acoustic feature matrix reflecting the rotor operating status, and record the corresponding fault labels. The model building and training unit is used to build a fault classification and discrimination model based on machine learning. It uses the acoustic feature matrix and the corresponding fault labels to train the fault classification and discrimination model and establishes the mapping relationship between acoustic features and fault modes. The fault diagnosis unit is used to collect real-time acoustic signals from the rotor of the helicopter to be diagnosed, input the trained fault classification and discrimination model, and output the fault status and fault type of the rotor.

Images