Aircraft airfoil pressure self-adaptive calibration method for wind tunnel environment

Abstract

The invention discloses an aircraft airfoil pressure self-adaptive calibration method for a wind tunnel environment, belongs to the technical field of pressure sensors, and aims to solve the problem that the sensitivity of the sensors is reduced during aircraft airfoil pressure testing. Comprising the steps that a flexible pressure sensor array is installed in a matched mode according to the airfoil shape of a to-be-tested aircraft, the to-be-tested aircraft is placed in a wind tunnel, wind tunnel parameters are adjusted, and pressure signals in a multi-scene environment are generated; each pressure sensor unit is regarded as an independent agent, and a calibration model is constructed by using an MADDPG algorithm; an Actor-Critic network architecture of multi-agent cooperative training is designed and obtained; and performing model compression on the trained Actor network, deploying the Actor network to an embedded system, fitting a nonlinear mapping relation between output of a pressure sensor unit and a real pressure value in real time, correcting a nonlinear error, and performing dynamic compensation on zero drift. The method is used for aircraft airfoil pressure distribution testing.

Description

TECHNICAL FIELD

[0001] The application relates to an aircraft wing surface pressure self-adaptive calibration method and belongs to the technical field of pressure sensors. BACKGROUND

[0002] Aircraft wing surface pressure distribution testing is a key link in aircraft design and aerodynamic performance research. When a flexible pressure sensor array is used to test the pressure of an aircraft wing surface, the following problems exist: 1. Each aircraft usually has thousands of pressure test points. After long-term use, the sensitivity of the pressure test units in the pressure sensor array will decrease due to material fatigue, and signal drift and other phenomena will occur.

[0003] 2. The existing system relies on traditional data processing methods and cannot cope with real-time data streams of high-density sensor arrays. In actual application, time is consumed and the testing ability and experience of the testing personnel are highly dependent.

[0004] Therefore, it is of great practical significance to research an aircraft wing surface pressure self-adaptive calibration method. SUMMARY

[0005] The application aims to solve the problem of sensor sensitivity decrease when testing the pressure of an aircraft wing surface, and provides an aircraft wing surface pressure self-adaptive calibration method for a wind tunnel environment.

[0006] The aircraft wing surface pressure self-adaptive calibration method for a wind tunnel environment comprises the following steps: A flexible pressure sensor array is fixedly installed on the wing surface of an aircraft to be tested, and the array structure is matched with the shape of the wing surface of the aircraft to be tested. The aircraft to be tested is placed in a wind tunnel, and the parameters of the wind tunnel are adjusted to generate pressure signals under multiple scene environments. Each pressure sensor unit in the flexible pressure sensor array is regarded as an independent intelligent agent, and a calibration model is constructed by using a MADDPG algorithm. An Actor-Critic network architecture for multi-agent collaborative training is designed, wherein the Actor network is used to train a calibration strategy, and the Critic network is used to train and evaluate the global calibration effect. The weights of the Actor-Critic network architecture are optimized by minimizing a long-term accumulated error objective function. The trained Actor network is model compressed and deployed to an embedded system. The nonlinear mapping relationship between the output of the pressure sensor unit and the real pressure value is fitted in real time, the nonlinear error is corrected, and the zero drift is dynamically compensated.

[0007] Preferably, the flexible pressure sensor array is packaged with piezoelectric sensors and piezoresistive sensors, and flexible electrodes are sprayed on the flexible PDMS substrate.

[0008] Preferably, the wind tunnel parameters include wind speed and temperature. The generated pressure signals in multiple scenarios are achieved by changing the attack angle, sideslip angle and roll angle of the aircraft.

[0009] Preferably, the construction of the calibration model includes: constructing a state space, an action space and a reward function. The state space includes pressure signal test values, wind tunnel parameters, historical zero-point drift errors and historical sensitivity attenuation errors. The action space includes calibration parameter adjustment instructions. The reward function is constructed according to the reduction amplitude of the pressure error after calibration.

[0010] Preferably, the state space is:

[0011] wherein, represents the current pressure normalized value, represents the time domain differential calculated by five-point central difference, represents the energy proportion of the 50-200Hz frequency band, , , respectively represent the mean, variance and kurtosis of the 60-second sliding window, represents the surface temperature of each pressure sensor, represents the wing surface vibration acceleration, represents the attack angle of the aircraft, represents the historical gain adjustment amount, represents the historical bias compensation amount, represents the reference sensor pressure value, represents the neighbor pressure similarity.

[0012] Preferably, the reward function is:

[0013] wherein, a cosine similarity constraint term representing the pressure distribution of adjacent sensors; an absolute error term representing the absolute error of calibrated pressure value and reference benchmark, a square penalty term representing the gain adjustment amount, a square penalty term representing the offset compensation amount, an exponential decay function representing the error-reward mapping relationship.

[0014] Preferably, the design obtains the Actor-Critic network architecture of multi-agent collaborative training, specifically comprising: each pressure sensor unit in the flexible pressure sensor array is regarded as an independent agent, and the state space is a 12-dimensional feature vector; the independent agent outputs the gain adjustment amount through the Actor network and the offset compensation amount two continuous action parameters, the Critic network adopts a Bayesian neural network to output the value estimate and the confidence of parameter adjustment.

[0015] Preferably, the action parameter limit range is .

[0016] Preferably, the non-linear error is corrected and the zero drift is dynamically compensated by fitting the non-linear mapping relationship between the output of the pressure sensor unit and the real pressure value in real time, specifically comprising: outputting the parameter confidence through the MonteCarlo Dropout layer of the Critic network ; when , a three-level response is started: cross-verification of redundant sensors is performed; Kalman filter data fusion is performed; online adjustment is performed to increase the learning rate to 3 times the regular value.

[0017] Preferably, the dropout rate of the MonteCarlo Dropout layer is .

[0018] The application has the advantages that the application provides an aircraft airfoil pressure adaptive calibration method for a wind tunnel environment, the flexible sensor array is made according to the shape of the aircraft airfoil to be measured, the artificial intelligence algorithm is used to realize joint strategy optimization of each pressure test unit of the flexible pressure sensor array in a continuous action space, the calibration parameters are dynamically adjusted through collaborative learning to cope with dynamic pressure changes in a complex flow field, and parameter calibration of the pressure test unit of the flexible pressure sensor array is realized.

[0019] The aircraft airfoil pressure adaptive calibration method based on artificial intelligence is used for adaptive calibration of the pressure unit, so that the long-term stability, dynamic adaptability and data processing efficiency of the sensor array are maximally improved. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a flow block diagram of the aircraft airfoil pressure adaptive calibration method for a wind tunnel environment according to the application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0022] It should be noted that the embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0023] The application will be further described below with reference to the drawings and specific embodiments, but the application is not limited by the embodiments.

[0024] Embodiment 1 The application will be further described below with reference to the drawings and specific embodiments, but the application is not limited by the embodiments. Figure 1 The application will be further described below with reference to the drawings and specific embodiments, but the application is not limited by the embodiments. The flexible pressure sensor array is fixedly installed on the aircraft airfoil to be measured, and the array structure is matched with the shape of the aircraft airfoil to be measured; The aircraft to be measured is placed in the wind tunnel, the wind tunnel parameters are adjusted, and pressure signals under multiple scenes are generated; Each pressure sensor unit in the flexible pressure sensor array is regarded as an independent intelligent agent, and a calibration model is constructed by using a MADDPG algorithm; An Actor-Critic network architecture for multi-agent collaborative training is designed, in which the Actor network is used to train the generation of calibration strategy, and the Critic network is used to train the evaluation of global calibration effect; and the weights of the Actor-Critic network architecture are optimized by minimizing the long-term accumulated error objective function; The trained Actor network is model compressed and deployed to an embedded system, which corrects the non-linear error and dynamically compensates the zero-point drift by fitting the non-linear mapping relationship between the output of the pressure sensor unit and the real pressure value in real time.

[0025] Further, the flexible pressure sensor array is packaged with piezoelectric sensors and piezoresistive sensors, and flexible electrodes are sprayed onto the flexible PDMS substrate.

[0026] Further, the wind tunnel parameters include wind speed and temperature. The generated pressure signal in multiple scene environments is realized by changing the attack angle, side slip angle and roll angle of the aircraft.

[0027] Further, the construction of the calibration model includes: constructing a state space, an action space and a reward function. The state space includes pressure signal test values, wind tunnel parameters, historical zero-point drift errors and historical sensitivity decay errors. The action space includes calibration parameter adjustment instructions. The reward function is constructed according to the reduction amplitude of the pressure error after calibration.

[0028] Further, the state space is:

[0029] wherein, the current pressure normalized value is represented by, the five-point central difference calculation time domain differential is represented by, the 50-200Hz frequency band energy ratio is represented by, , , the mean, variance and kurtosis of the 60-second sliding window are represented by, the surface temperature of each pressure sensor is represented by, the wing surface vibration acceleration is represented by, the aircraft attack angle is represented by, the historical gain adjustment amount is represented by, denotes the history bias compensation amount, denotes the reference sensor pressure value, denotes the neighbor pressure similarity.

[0030] Further, the reward function is:

[0031] wherein, denotes the cosine similarity constraint term of the adjacent sensor pressure distribution; denotes the absolute error of the calibrated pressure value and the reference benchmark, denotes the square penalty of the gain adjustment amount, denotes the square penalty of the bias compensation amount, denotes an exponential decay function for establishing an error-reward mapping relationship.

[0032] Further, the design obtains an Actor-Critic network architecture for multi-agent collaborative training, specifically comprising: Each pressure sensor unit in the flexible pressure sensor array is regarded as an independent agent, and the state space is a 12-dimensional feature vector; The independent agent outputs the gain adjustment amount and the offset compensation amount through the Actor network, The Critic network adopts a Bayesian neural network to output the value estimate and the confidence of parameter adjustment.

[0033] Further, the action parameter limit range is .

[0034] Further, the non-linear error is corrected and the zero drift is dynamically compensated by fitting the non-linear mapping relationship between the output of the pressure sensor unit and the real pressure value in real time, specifically comprising: The parameter confidence is output by the MonteCarlo Dropout layer of the Critic network; When , a three-level response is started: Cross-validation of redundant sensors is performed; Kalman filter data fusion is performed; Online adjustment is performed to increase the learning rate to 3 times the regular value.

[0035] Still further, the dropout rate of the Monte Carlo Dropout layer is .

[0036] In the application, the flexible pressure sensor array is composed of piezoelectric sensors, piezoresistive sensors and silicone encapsulation, the flexible electronic device patterned electrode is prepared by a classic spraying process, the flexible electrode substrate material is PDMS, the flexible sensor array is made according to the shape of the aircraft wing surface to be measured, and the sensor array is installed to the aircraft wing surface to be measured.

[0037] In the application, the aircraft wing surface to be measured is placed in a wind tunnel, the wind tunnel parameters (wind speed and temperature) are adjusted, the pressure signals under different scenes (aircraft angle of attack, aircraft sideslip and aircraft roll) are tested, based on the flexible pressure sensor array of step 1, each pressure test unit of the sensor is regarded as an independent agent, the state space thereof includes the current pressure test value, the wind tunnel parameters and the historical error data of zero drift and sensitivity attenuation, the calibration action space is designed and the reward function is constructed.

[0038] In the application, the Actor-Critic network for multi-agent collaborative training is designed, the output calibration parameter adjustment strategy is realized, the global calibration effect is evaluated, the training target of minimizing the long-term cumulative error and optimizing the robustness of the calibration strategy is achieved.

[0039] In the application, the embedded system carrying the deployed compressed Actor network can realize low-delay calibration, the nonlinear mapping between the sensor output and the true pressure is realized through the algorithm strategy and the deep network fitting, and the zero dynamic compensation and nonlinear error correction are realized.

[0040] In the application, the flexible pressure sensor array: a plurality of sensitive units are integrated on a flexible PDMS substrate, each sensitive unit can independently measure a pressure signal, and the sensitive units are integrated on the flexible substrate. The reference sensor: a high-precision static pressure sensor, serving as a calibration reference. The signal processing module: filtering, amplifying and digitizing the collected signals. The data acquisition module: synchronously acquiring the original signals of the sensor and environmental parameters (wind speed, temperature and vibration). The FPGA: deploying the reinforcement learning model in a multi-core FPGA. The dynamic database: storing the associated data of the historical drift mode and the environmental parameters.

[0041] In the application, the spatial distribution characteristics of the sensor array are introduced into the system to realize multi-node data fusion through the space-time attention mechanism. The 64-dimensional feature vector extracted by each sensor node calculates the spatial correlation weight through the self-attention layer, and the influence of the sensor data in the key area of the flow field (such as the leading edge separation zone of the airfoil) is highlighted. At the same time, the environmental parameters (temperature, vibration, and attack angle) are predicted through the LSTM network to predict their coupling effect on the sensor output, and a dynamic compensation equation is established. Adaptive filtering: according to the real-time acquisition of signal characteristics, the filter parameters are dynamically adjusted to effectively remove noise and crosstalk. The median filtering and wavelet transform fusion technology is adopted, the median filtering can suppress impulse noise, and is used for smoothing processing of dynamic pressure signals, and the wavelet transform can decompose the signal frequency domain characteristics, and filter the noise in the specific frequency band.

[0042] In the application, each sensor node is an independent intelligent agent, and its state space includes 12-dimensional feature vectors such as preprocessed signals, time domain differential characteristics, environmental parameters, and historical drift patterns. The intelligent agent outputs two continuous action parameters, gain adjustment (ΔG) and bias compensation (ΔO), through the Actor network, and the action range is limited to [-0.1, 0.1] to ensure the stability of the adjustment process. The Critic network integrates global environmental information and uses a Bayesian neural network structure to output value estimates and parameter adjustment confidence. The reward function design adopts a multi-objective optimization strategy to constrain the parameter adjustment amplitude while ensuring calibration accuracy, avoiding overcompensation.

[0043] In the application, a flexible pressure sensor array is installed on the airfoil of an aircraft, which includes a composite sensing unit of piezoelectric sensors and piezoresistive sensors. The sensing unit forms a patterned electrode on a PDMS flexible substrate through a spraying process. The airfoil with the installed sensor array is placed in a wind tunnel environment to collect pressure signals and environmental parameters, including wind speed, temperature, vibration frequency, and attack angle. A MADDPG multi-agent deep reinforcement learning algorithm is used to model each pressure test unit as an independent intelligent agent, and its state space includes the current pressure value, historical drift characteristics, and environmental parameters. The Actor-Critic network is used to cooperatively optimize the calibration strategy. The trained Actor network is deployed on an FPGA hardware to output gain adjustment ΔG and bias compensation ΔO in real time, and dynamically correct the sensor output signal.

[0044] In the application, the FPGA deployment adopts a three-level optimization scheme: Network quantization: 32-bit floating point → 8-bit fixed point (dynamic range quantization DRQ algorithm) Parameter pruning: remove weights with absolute value <1e-4 (compression ratio ≥63%) Layer fusion: merge the FC3+Tanh layer of the Actor network into a pre-computed lookup table operation In the application, the online trusted calibration mechanism: Parameter confidence σ output by Monte Carlo Dropout layer (dropout rate p=0.2) of Critic network When σ>0.1, start three-level response: Redundant sensor cross-validation Kalman filter data fusion Online fine-tuning (learning rate increased to 3 times the regular value) In the present application, multi-agent collaborative perception: break through the limitation of traditional single-point calibration, use the spatial correlation characteristics of sensor array to improve the robustness of the system; Physical-data hybrid driving: integrate fluid mechanics priori knowledge into reinforcement learning reward function to accelerate model convergence; Double time scale adaptation: millisecond level real-time compensation combined with day level continuous learning to adapt to material aging nonlinear process.

[0045] While the application has been described with reference to particular embodiments, it will be understood that the examples are merely illustrative of the principles and applications of the present application. It will be understood that various modifications can be made to the illustrative embodiments, and other arrangements can be devised without departing from the spirit and scope of the present application as defined by the appended claims. It will be understood that the features of the various embodiments can be combined with each other, in different ways than as described herein. It will be understood that features described with reference to individual embodiments can be used in other embodiments described herein.

Claims

1. A method for adaptively calibrating aircraft wing surface pressure in a wind tunnel environment, characterized in that: It includes: The flexible pressure sensor array is fixedly mounted on the wing surface of the aircraft to be tested, and the array structure matches the shape of the wing surface of the aircraft to be tested; Place the aircraft under test in a wind tunnel, adjust the wind tunnel parameters, and generate pressure signals under multiple scenarios; Each pressure sensor unit in the flexible pressure sensor array is regarded as an independent intelligent agent, and the calibration model is constructed using the MADDPG algorithm; We designed an actor-critic network architecture for multi-agent collaborative training, where the actor network is used to train the generated calibration strategy, and the critic network is used to train and evaluate the global calibration effect. We also optimized the weights of the actor-critic network architecture by minimizing the long-term accumulated error objective function. The trained Actor network model is compressed and deployed to the embedded system. By real-time fitting the nonlinear mapping relationship between the output of the pressure sensor unit and the actual pressure value, the nonlinear error is corrected and the zero drift is dynamically compensated.

2. The method for adaptively calibrating aircraft wing surface pressure in a wind tunnel environment according to claim 1, characterized in that: The flexible pressure sensor array is formed by packaging piezoelectric sensors and piezoresistive sensors, and flexible electrodes are sprayed onto a flexible PDMS substrate.

3. The method for adaptively calibrating aircraft wing surface pressure in a wind tunnel environment according to claim 1, characterized in that: The wind tunnel parameters include wind speed and temperature; The generation of the pressure signal in the multi-scenario environment is achieved by changing the aircraft's angle of attack, sideslip angle, and roll angle.

4. The method for adaptively calibrating aircraft wing surface pressure in a wind tunnel environment according to claim 1, wherein: The constructing of the calibration model includes: constructing a state space, an action space and a reward function; The state space includes pressure signal test values, wind tunnel parameters, historical zero drift errors, and historical sensitivity attenuation errors; The action space includes calibration parameter adjustment instructions; The reward function is constructed based on the reduction in pressure error after calibration.

5. The method for adaptively calibrating aircraft wing surface pressure in a wind tunnel environment according to claim 4, characterized in that: The state space for: in, Indicates the current normalized pressure value, represents the time domain differential of the five-point central difference calculation, Indicates the energy proportion of the 50~200Hz frequency band, 、 、 Represent the mean, variance, and kurtosis of the 60-second sliding window, represents the surface temperature of each pressure sensor, represents the wing surface vibration acceleration, represents the aircraft's angle of attack, Indicates the historical gain adjustment amount, Indicates the historical bias compensation amount, Indicates the reference sensor pressure value, Indicates the neighbor pressure similarity.

6. The method for adaptively calibrating aircraft wing surface pressure in a wind tunnel environment according to claim 4, characterized in that: The reward function for: in, The cosine similarity constraint term represents the pressure distribution of adjacent sensors; Indicates the absolute error between the calibrated pressure value and the reference standard. represents the square penalty of the gain adjustment, represents the square penalty of bias compensation, Represents an exponential decay function, which is used to establish the error-reward mapping relationship.

7. The method for adaptively calibrating aircraft wing surface pressure in a wind tunnel environment according to claim 1, characterized in that: The Actor-Critic network architecture designed for multi-agent collaborative training specifically includes: Each pressure sensor unit in the flexible pressure sensor array is regarded as an independent intelligent agent, and the state space is a 12-dimensional feature vector; Independent agents output gain adjustments through the Actor network and offset compensation Two continuous action parameters, The critic network uses a Bayesian neural network to output confidence in value estimation and parameter adjustment.

8. The method for adaptively calibrating aircraft wing surface pressure in a wind tunnel environment according to claim 7, characterized in that: The action parameter limit range is .

9. The method for adaptively calibrating aircraft wing surface pressure in a wind tunnel environment according to claim 1, wherein: The method of correcting the nonlinear error and dynamically compensating the zero drift by real-time fitting the nonlinear mapping relationship between the output of the pressure sensor unit and the actual pressure value specifically includes: Output parameter confidence through the Monte Carlo Dropout layer of the Critic network ; when When a three-level response is initiated: Perform redundant sensor cross-validation; Perform Kalman filter data fusion; Perform online adjustments and increase the learning rate to 3 times the normal value.

10. The method for adaptively calibrating aircraft wing surface pressure in a wind tunnel environment according to claim 9, characterized in that: The dropout rate of the Monte Carlo Dropout layer .

Images