A method for full-cycle perception and early warning of wing icing

Abstract

A method for full-cycle perception and warning of wing icing belongs to the field of aircraft icing detection and warning. First, cloud particle data, flight parameter sensing data, and planar electrode icing detection data are obtained; then, the cloud particle data and flight parameter data are preprocessed to obtain the impact efficiency of cloud particles hitting the wing surface and the icing mass flow rate; finally, a hybrid model based on BP neural network and recurrent neural network is constructed, and time series data is processed through long and short-term memory units to achieve prediction of icing conditions. The present invention can monitor multi-scale dynamic cloud fields in actual flight, output predicted ice thickness and icing type in real time, and set multi-level warning thresholds to achieve gradient warning output of potential risk prompts; it can realize the correction of icing prediction results during real-time flight and improve the prediction accuracy; it has the advantages of high-precision prediction, real-time monitoring, strong generalization, etc., and is suitable for icing safety protection under complex meteorological conditions.

Description

Technical Field

[0001] The present invention belongs to the field of aircraft icing detection and early warning, and relates to a method for full-cycle perception and early warning of wing icing. Background Art

[0002] In aviation, aircraft icing is a common but dangerous phenomenon that poses a serious threat to flight safety. Because the spatiotemporal distribution of icing conditions (such as supercooled water droplets, ice crystals, and mixed-phase clouds) is highly dynamic, relying solely on single-stage detection cannot cope with the sudden changes in the microphysical environment along the route. Therefore, dynamic monitoring of all stages of the icing cycle is crucial to ensuring flight safety.

[0003] Current research on full-cycle icing detection and warning for aircraft primarily assesses icing risk indirectly, lacking key parameters for ice type and thickness, or focusing solely on identifying ice type and thickness. This lacks dynamic monitoring of the entire process from cloud particles floating in the atmosphere to impacting and subsequently adhering to the aircraft surface. Therefore, it is inadequate for the complex, real-time environmental changes that occur during flight. For example, Commercial Aircraft Corporation of China, Ltd. (Chinese invention patent CN117644978A) has proposed an icing warning method based on a mapping of flight parameters and total temperature. This method indirectly determines icing risk by inverting the relationship between aircraft surface temperature and flight parameters. However, its reliance on temperature thresholds to trigger warnings, rather than directly monitoring ice formation, can result in delayed warnings, potentially posing the possibility that ice has formed but the total temperature has not yet reached the threshold. Wuhan University of Technology (Chinese invention patent CN118329087A) has proposed a wing de-icing and icing status monitoring system that uses embedded optical fibers inscribed with gratings to sense wing ice and temperature changes. Taiyuan University of Technology (Chinese invention patent CN118013359A) has proposed a method for identifying ice types and predicting ice thickness on wing icing. This method detects the icing state through the temperature, frequency, and capacitance signals of a planar electrode icing sensor. After network model training, it can identify different icing types and predict ice thickness.

[0004] In fact, the full-cycle perception of wing icing involves the detection of cloud particles and the full-process monitoring of the collision and adhesion of cloud particles to the wings. This can not only match the complexity of the multi-scale dynamic cloud field in real flight, but also obtain the ice coverage status of the wing surface in real time, which is an important guarantee for achieving accurate and efficient early warning. Summary of the Invention

[0005] To address the shortcomings of existing technologies or the need for improvement, this paper proposes a full-cycle wing icing perception and warning method. By integrating sensors, a real-time data acquisition and processing system, an accurate icing physics model, and intelligent machine learning algorithms, this method aims to achieve real-time monitoring, assessment, and early warning of icing conditions throughout the entire flight process. Specifically, through multi-sensor fusion and intelligent algorithms, this method achieves full coverage of the entire process, from sensing icing meteorological conditions to icing warning.

[0006] To achieve the above object, the method scheme adopted by the present invention is as follows:

[0007] A method for full-cycle perception and early warning of wing icing is described. In this method, the full wing icing cycle encompasses the entire monitoring process, from cloud particle detection to cloud particle impact to ice adhesion to the aircraft surface. First, cloud particle data, flight parameter sensing data, and planar electrode icing detection data are acquired. Then, the acquired cloud particle data and flight parameter data are preprocessed to obtain the impact efficiency of cloud particles impacting the wing surface and the ice mass flow rate. Finally, a hybrid model based on a BP neural network and a recurrent neural network (RNN) is constructed, processing time series data through long short-term memory (LSTM) units to predict icing conditions. Specifically, the following steps are included:

[0008] Step 1: During flight, a cloud particle imaging sensor is used to collect cloud particle data in real time. The cloud particle data includes the particle size distribution and phase distribution of cloud particles in front of the wing. A flight parameter sensing module is used to obtain flight parameter sensing data, including aircraft flight parameter data and wing surface temperature. The aircraft flight parameters include speed, altitude, atmospheric density, air pressure, speed of sound, and dynamic viscosity. A planar electrode ice detector is used to obtain planar electrode ice detection data, including ice type and thickness data, excitation frequency, capacitance at the corresponding excitation frequency, and planar electrode ice detector surface temperature. The ice type and thickness data include ice type and ice thickness changes of ice on the aircraft wing.

[0009] Step 2: Perform preliminary data processing on the cloud particle data and aircraft flight parameters obtained in Step 1. First, use computational fluid dynamics (CFD) to obtain the ambient flow field distribution. Then, based on the air flow field, solve the trajectory equation for cloud particle motion to obtain the cloud particle impact efficiency. Finally, combined with the cloud particle impact efficiency, calculate the mass flow rate of cloud particles impacting the wing surface. The details are as follows:

[0010] In step 2.1, the air flow field is first calculated based on CFD computational fluid dynamics theory. The turbulence model uses the k-ε equation, and the fluid dynamics control theory uses the 3-dimensional Navier-Stokes (NS) equation.

[0011] The k-ε equation of the turbulence model is:

[0012]

[0013] in, is the fluid density; is the turbulent kinetic energy; It’s time; is the average speed; is the fluid dynamic viscosity; is the turbulent viscosity; is related to turbulent kinetic energy The corresponding Prandtl number; is the turbulent kinetic energy due to the mean velocity gradient The generated items; is the turbulent kinetic energy due to buoyancy The generated items; is the turbulent dissipation rate; is the contribution of pulsation expansion in turbulent flow; is a user-defined source term; is the spatial coordinate component in the j direction; is the spatial coordinate component in the i direction;

[0014] The 3-dimensional Navier-Stokes (NS) equations are:

[0015]

[0016] in, is the fluid density; It’s time; is the fluid dynamic viscosity; u, v, and w are velocity vectors Components in the x, y, and z directions; p is the pressure on the fluid element; div() is the divergence; grad() is the gradient; is the generalized source term in the u direction of the momentum conservation equation; is the generalized source term in the v direction of the momentum conservation equation; It is the generalized source term in the w direction of the momentum conservation equation.

[0017] Substituting the cloud particle data and aircraft flight parameters obtained in step 1 into the turbulence model and the three-dimensional Navier-Stokes equations can solve the environmental flow field distribution to obtain the air flow field, wherein the environmental flow field distribution includes air pressure distribution, velocity distribution, temperature distribution, and shear stress distribution;

[0018] Step 2.2: Based on the air flow field obtained in step 2.1, the cloud particle Euler governing equation is established as follows:

[0019]

[0020] in, is the cloud particle volume fraction; is the velocity vector of the cloud particles; is the dynamic viscosity of the cloud particles; is the cloud particle density; is the diameter of the cloud particles; is the drag coefficient; is the Reynolds number; is the velocity vector of the air; is the air density; is the acceleration due to gravity; represents divergence;

[0021] By solving the cloud particle data in step 1, the aircraft flight parameters and the ambient flow field distribution in step 2.1, equations (6) and (7) can be used to obtain the impact efficiency of cloud particles on the wing surface. The impact efficiency for:

[0022]

[0023] in, is the velocity of the impacting cloud particles; is the unit normal vector of the surface; is the far-field velocity; is the volume fraction of cloud particles on the surface; is the cloud particle volume fraction in the far-field free stream.

[0024] Step 2.3, combine the cloud particle impact efficiency calculated in step 2.2 with the wing surface The ice mass flow rate per unit time on the wing surface m is expressed as:

[0025]

[0026] Where S is the effective area of cloud particles impacting the wing, is the density of cloud particles, is the velocity vector of the cloud particles, is the concentration of cloud particles, V(D) is the volume of the particles, D is the diameter of the cloud particles, and d represents the sign of the variable of integration.

[0027] Step 3: Construct a hybrid model based on BP neural network and recurrent neural network (RNN), including BP neural network and RNN neural network. The BP neural network is used to process multidimensional data, including cloud particle data and aircraft flight parameters; the RNN neural network uses long short-term memory (LSTM) units to model time series data to capture the temporal evolution of the icing process. In particular, the ice thickness and ice type data observed in real time by the planar electrode icing detector is used as a physical constraint condition, and a constraint loss function based on the physical equation is constructed. The real-time ice thickness and ice type data measured by the planar electrode icing detector and the predicted ice thickness and ice type data finally output by the hybrid model are jointly optimized. Specifically:

[0028] In step 3.1, the BP neural network extracts key feature vectors through a multi-layer feedforward structure. The data obtained in steps 1 and 2 are initialized and processed by the BP neural network and input into the BP neural network. The initialization process involves converting the cloud particle data, aircraft flight parameter data, capacitance, and excitation frequency from step 1, and the cloud particle impact efficiency and ice mass flow rate from step 2 into multidimensional input vectors. A five-layer fully connected structure is set up, the multidimensional input vectors are input, and the predicted ice type and ice thickness data are output. The BP neural network uses an adaptive activation function, which is:

[0029]

[0030] Among them, a and b in the activation function are learnable parameters that change with the BP neural network training process; is the activation function, is the output of the current neuron, and z is the weighted linear combination of the current neuron, that is, the input value of the activation function.

[0031] In step 3.2, a recurrent neural network (RNN) is constructed. Long short-term memory (LSTM) units are used to model the time series data and capture the temporal evolution of the icing process. The predicted ice type and thickness data from the BP neural network in step 3.1 and the ice type and thickness data from the planar electrode ice detector in step 1 are used as input features of the RNN. A dynamic weighted fusion of the predicted ice type and thickness data and the ice type and thickness data from the planar electrode ice detector is performed. The time series data is then processed using LSTM (Long Short-Term Memory) units, with a sigmoid activation function used to map the input data to a value between 0 and 1.

[0032] In step 3.3, the hybrid model composed of the BP neural network and the RNN neural network introduces the ice type and thickness data of the planar electrode ice detector in step 1 as physical constraints, and constructs a constraint loss function based on the physical equation as a dynamic correction module. The constraint loss function shown is as follows:

[0033]

[0034] in, is the loss of physical equation constraints; T is the total detection time; For ice thickness; is the detection time; is the freezing mass flow rate; is the ice type shape factor function;

[0035] The physical constraints are weightedly fused with the data-driven loss, which is a measure of the difference between the prediction results of the hybrid model and the actual observation data obtained by the planar electrode ice detector, reflecting the hybrid model's ability to fit the existing data. Balance data fitting and physical consistency. Form the total loss function:

[0036]

[0037] Among them, Loss is the loss function of the total output; is a weight hyperparameter that needs to be adjusted according to the data signal-to-noise ratio and the confidence of the physical model; KL is a divergence term used to constrain the consistency of the predicted distribution and the measured distribution; is the predicted ice thickness output by the neural network model; Ice thickness data for flat electrode ice detection; Ice type distribution predicted by the neural network model; is the ice type distribution actually observed by the planar electrode ice detector; represents the mean square error.

[0038] Step 4: Set up multi-level warning rules. When the prediction result output by the hybrid model exceeds the threshold, a wing icing warning corresponding to the icing risk level will be issued. Specifically:

[0039] The icing risk levels are divided into low risk, medium risk, high risk and extremely high risk according to the ice thickness, ice type and temperature. The classification principle is: ice thickness <5 mm, frost ice, below -10℃ is low risk; ice thickness 5-15 mm, frost ice / mixed ice, -10℃ to -5℃ is medium risk; ice thickness 15-30 mm, clear ice, -5℃ to -1℃ is high risk; ice thickness >30 mm, clear ice, -1℃ to 0℃ is extremely high risk.

[0040] When the prediction output by the hybrid model exceeds a threshold, a low-risk warning, medium-risk warning, high-risk warning, or extremely high-risk warning is triggered based on the icing risk level, enabling real-time monitoring and early warning of wing icing. The thresholds are based on the ice thickness, ice type, and temperature range corresponding to each icing risk level.

[0041] The beneficial effects of the present invention are:

[0042] (1) The present invention acquires data through cloud particle imaging sensors, flight parameter acquisition modules and planar electrode icing detectors, and predicts ice type and thickness through a hybrid neural network. It can not only monitor multi-scale dynamic cloud fields in real flight, but also obtain the icing status of the wing surface in real time, and output the predicted ice thickness and icing type in real time. At the same time, it sets multi-level warning thresholds to achieve gradient warning output of potential risk prompts.

[0043] (2) The present invention introduces the ice type and thickness data of the wing surface obtained by the planar electrode ice detector into the loss function of the hybrid neural network model, thereby realizing the correction of the icing prediction results during real-time flight and improving the prediction accuracy.

[0044] In summary, the present invention has the advantages of high-precision prediction, real-time monitoring, and strong generalization, and is suitable for icing safety protection of multiple models under complex meteorological conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A flow chart of a full-cycle wing icing perception and warning method provided by a specific embodiment of the present invention;

[0046] Figure 2 A full-cycle perception diagram of cloud particle icing provided by a specific embodiment of the present invention;

[0047] Figure 3 A data processing and neural network model training flow chart provided for a specific embodiment of the present invention;

[0048] Figure 4 A schematic diagram of a cloud particle icing generation mechanism provided by a specific embodiment of the present invention;

[0049] Figure 5 A neural network flow chart provided for a specific embodiment of the present invention. DETAILED DESCRIPTION

[0050] To facilitate understanding of the present invention, the present invention is described in more detail below with reference to the accompanying drawings and specific embodiments. The accompanying drawings provide preferred embodiments of the present invention. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in this specification. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0051] The present invention provides a method for full cycle perception and early warning of wing icing, such as Figure 1 As shown in the figure, the entire process from cloud particle detection to cloud particle impact to ice adhesion on the aircraft surface is monitored (e.g. Figure 2 Specific implementation measures are as follows:

[0052] Step 1: During the flight of the aircraft, use a cloud particle imaging sensor to collect cloud particle data in real time. The cloud particle data includes the particle size distribution and phase distribution of cloud particles in front of the wing; use a flight parameter sensing module to obtain flight parameter sensing data, including aircraft flight parameter data and wing surface temperature. The aircraft flight parameters include speed, altitude, atmospheric density, air pressure, speed of sound, and dynamic viscosity; use a planar electrode ice detector to obtain planar electrode ice detection data, including ice type and thickness data, excitation frequency, capacitance at the corresponding excitation frequency, and planar electrode ice detector surface temperature. The ice type and thickness data include the ice type and ice thickness changes of the aircraft wing ice. Specifically:

[0053] The cloud particle data, aircraft flight parameters, ice type and thickness data, capacitance, and frequency data obtained in step 1 are processed and the neural network model is trained. The specific process is as follows: Figure 3 As shown, that is, steps 2, 3, and 4.

[0054] Step 2: Perform preliminary data processing on the cloud particle data and aircraft flight parameters obtained in Step 1. First, use the Computational Fluid Dynamics (CFD) method to obtain the ambient flow field distribution. Then, based on the Euler model and the air flow field, solve the trajectory equation of the cloud particle motion to obtain the cloud particle impact efficiency. Finally, based on the cloud particle impact efficiency, calculate the mass flow rate of the cloud particles impacting the wing surface. The details are as follows:

[0055] In step 2.1, the environmental flow field is first calculated based on CFD computational fluid dynamics theory. The turbulence model uses the k-ε equation, and the fluid dynamics control theory uses the 3-dimensional Navier-Stokes (NS) equation.

[0056] The k-ε equation of the turbulence model is:

[0057]

[0058] in, is the fluid density; is the turbulent kinetic energy; It’s time; is the average speed; is the fluid dynamic viscosity; is the turbulent viscosity; is related to turbulent kinetic energy The corresponding Prandtl number; is the turbulent kinetic energy due to the mean velocity gradient The generated items; is the turbulent kinetic energy due to buoyancy The generated items; is the turbulent dissipation rate; is the contribution of pulsation expansion in turbulent flow; is a user-defined source term; is the spatial coordinate component in the j direction; is the spatial coordinate component in the i direction;

[0059] The 3-dimensional Navier-Stokes (NS) equations are:

[0060]

[0061] in, is the fluid density; It’s time; is the fluid dynamic viscosity; u, v, and w are velocity vectors Components in the x, y, and z directions; p is the pressure on the fluid element; div() is the divergence; grad() is the gradient; is the generalized source term in the u direction of the momentum conservation equation; is the generalized source term in the v direction of the momentum conservation equation; It is the generalized source term in the w direction of the momentum conservation equation.

[0062] Substituting the cloud particle data and aircraft flight parameters obtained in step 1 into the turbulence model and the three-dimensional Navier-Stokes equations can solve the environmental flow field distribution to obtain the air flow field, wherein the environmental flow field distribution includes air pressure distribution, velocity distribution, temperature distribution, and shear stress distribution;

[0063] Step 2.2: Based on the Euler model and the air flow field, the cloud particle Euler control equation is established as follows:

[0064]

[0065] in, is the cloud particle volume fraction; is the velocity vector of the cloud particles; is the dynamic viscosity of the cloud particles; is the cloud particle density; is the diameter of the cloud particles; is the drag coefficient; is the Reynolds number; is the velocity vector of the air; is the air density; is the acceleration due to gravity; represents divergence;

[0066] By solving equations (6) and (7) using the cloud particle data in step 1, the aircraft flight parameters, and the flow field distribution in step 2.1, we can obtain the impact efficiency of cloud particles onto the wing surface. The impact efficiency is:

[0067]

[0068] in, is the velocity of the impacting cloud particles; is the unit normal vector of the surface; is the far-field velocity; is the volume fraction of cloud particles on the surface; is the cloud particle volume fraction in the far-field free stream.

[0069] like Figure 4 As shown in Figure 2, different types of ice have different formation mechanisms. The formation mechanisms of clear ice, frost ice, and mixed ice are related to the freezing process of cloud particles, environmental conditions, and impact efficiency ( ) are related. Rime ice is often formed in low-temperature environments (e.g., below -15°C). Ice crystal impact efficiency is low, and some particles evaporate before impact or are deflected by air currents. The remaining ice crystals condense upon collision, rapidly cooling to form a porous, opaque, loose ice crystal structure. Clear ice typically forms in clouds with temperatures slightly below 0°C and a high liquid water content. Large, supercooled water droplets are more likely to impact aircraft surfaces due to their higher impact efficiency. Droplet diameter and temperature have a significant impact, creating a heat transfer effect. Upon impact, the droplets flow, then quickly spread and freeze completely, forming a dense, transparent, smooth ice layer. Mixed ice arises from the coexistence of multi-scale particles in the cloud. Large droplet impacts form localized clear ice, while ice crystal impacts form rime ice, ultimately forming a layered ice body with alternating dense and porous layers.

[0070] Preferably, the impact rates of frost ice, clear ice, and mixed ice are respectively expressed as:

[0071]

[0072] in, The impact efficiency when frost ice is generated; The impact efficiency when the ice is produced; The impact efficiency when mixed ice is produced; 、 is a coefficient related to the volume ratio of ice crystals and water droplets. is a function of temperature. As follows:

[0073]

[0074] Wherein, k is a slope parameter, and in this embodiment, k=0.5 is taken to control temperature sensitivity; T0 is a characteristic temperature, and in this embodiment, it is taken to be -5°C.

[0075] Step 2.3: The impact efficiency of cloud particles on the wing surface calculated in step 2.2 The ice mass flow rate per unit time on the wing surface m is expressed as:

[0076]

[0077] in, is the impact efficiency of the particle, S is the effective area of the wing that the particle impacts, is the density of cloud particles (water droplets / ice crystals), is the velocity vector of the cloud particles, is the concentration of cloud particles, V(D) is the volume of the particles, D is the particle diameter, and d represents the sign of the variable of integration.

[0078] Step 3, construct a hybrid model based on BP neural network and recurrent neural network (RNN), including BP neural network and RNN neural network, such as Figure 5 As shown. The BP neural network is responsible for processing multidimensional data, including cloud particle data and aircraft flight parameters; the RNN neural network models time series data through long short-term memory (LSTM) units to capture the temporal evolution of the icing process. This embodiment uses the ice thickness and ice type data observed in real time by the planar electrode icing detector as physical constraints, constructs a constraint loss function based on physical equations, and jointly optimizes the measured data of the planar electrode icing detector and the predicted ice thickness and ice type data ultimately output by the hybrid model. Specifically:

[0079] In step 3.1, the BP neural network extracts key feature vectors such as temperature, cloud particle type, cloud particle diameter, cloud particle concentration, flight altitude, impact rate, ice mass flow rate, and atmospheric pressure through a multi-layer feedforward structure. The data obtained in steps 1 and 2 are initialized and input into the BP neural network. The initialization process involves converting the cloud particle data, aircraft flight parameter data, capacitance, excitation frequency in step 1, and the cloud particle impact efficiency and ice mass flow rate in step 2 into multi-dimensional input vectors. A five-layer fully connected structure is set up, the multi-dimensional input vectors are input, and the predicted ice type and ice thickness data are output. The BP neural network uses an adaptive activation function, which is:

[0080]

[0081] in, is the activation function, is the output of the current neuron, z is the weighted linear combination of the current neuron, that is, the input value of the activation function. a and b in the activation function are learnable parameters that change with the training process. In this embodiment, the training learning parameters a are 1.5 and b are 0.3.

[0082] In step 3.2, an RNN neural network is constructed to model the time series data through the long short-term memory (LSTM) unit to capture the temporal evolution law of the icing process. The output of the BP neural network in step 3.1, the predicted ice type and thickness data, and the ice type and thickness data of the planar electrode ice detector in step 1 are used as the input features of the RNN network. The predicted ice type and thickness data and the ice type and thickness data detected by the planar electrode ice detector are dynamically weighted and fused; then, the LSTM (long short-term memory) unit is used to process the time series data, and the Sigmoid activation function is selected to map the input data to between 0 and 1. The calculation formula (13) of the LSTM unit is as follows:

[0083]

[0084] in, and are the weight matrix and bias vector of the forget gate respectively; is the weight matrix of the input gate; is the bias vector of the input gate; is the weight matrix of the candidate state; is the bias vector of the candidate state; and are the weight matrix and bias vector of the output gate respectively; Indicates term-by-term multiplication; C t-1 is the cell state (memory state), which is the long-term memory input; is the input information; is the hidden state input, is the output of the hidden state after time t, which is the short-term memory output; C t Represents the information of the neuron after time t, which is the long-term memory output; Is the output gate, control Which information is passed to the hidden layer?

[0085] In step 3.3, to prevent overfitting of the hybrid model and improve its generalization ability, the ice type and thickness data from the planar electrode ice detector in step 1 are introduced into the hybrid model composed of the BP neural network and the RNN neural network as physical constraints. A constraint loss function based on the physical equation is constructed as a dynamic correction module. The constraint loss function shown is as follows:

[0086]

[0087] in, is the loss of physical equation constraints; T is the total detection time; For ice thickness; is the detection time; is the freezing mass flow rate; is the ice type shape factor function;

[0088] The physical constraints are weightedly integrated with the data-driven loss, which is a measure of the difference between the model prediction results and the actual observed data, reflecting the model's ability to fit the existing data. Balance data fitting and physical consistency. Form the total loss function:

[0089]

[0090] Among them, Loss is the loss function of the total output; is a weight hyperparameter that needs to be adjusted according to the data signal-to-noise ratio and the confidence of the physical model; KL is a divergence term used to constrain the consistency of the predicted distribution and the measured distribution; is the predicted ice thickness output by the neural network model; Ice thickness data for flat electrode ice detection; Ice type distribution predicted by the neural network model; is the ice type distribution actually observed by the planar electrode ice detector; represents the mean square error.

[0091] Step 4: Set up multi-level warning rules. When the prediction result output by the hybrid model exceeds the threshold, a wing icing warning corresponding to the icing risk level will be issued. Specifically:

[0092] Icing risk levels are categorized as low, medium, high, and very high based on ice thickness, ice type, and temperature. When the prediction output by the hybrid neural network exceeds a threshold, the system triggers a low, medium, high, or very high risk warning, enabling real-time monitoring and early warning of wing icing.

[0093] Icing risk levels are categorized based on ice thickness, ice type, and temperature into low risk (<5 mm, frost ice, below -10°C), medium risk (5-15 mm, frost ice / mixed ice, -10°C to -5°C), high risk (15-30 mm, clear ice, -5°C to -1°C), and extremely high risk (>30 mm, clear ice, -1°C to 0°C). When the predicted result exceeds the corresponding threshold, the system triggers a corresponding risk warning, enabling real-time monitoring and early warning of wing icing.

[0094] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for full-cycle wing icing sensing and early warning, wherein the full wing icing cycle is the entire monitoring process from cloud particle detection to cloud particle impact to ice adhesion to the aircraft surface; characterized in that: The method first acquires cloud particle data, flight parameter sensor data, and planar electrode icing detection data; then, preprocesses the acquired cloud particle data and flight parameter data to obtain the impact efficiency of cloud particles hitting the wing surface and the icing mass flow rate; finally, constructs a hybrid model based on a BP neural network and a recurrent neural network, processes the time series data through long short-term memory units, and predicts icing conditions. The method includes the following steps: Step 1: During the flight of the aircraft, cloud particle imaging sensors are used to collect cloud particle data in real time, wherein the cloud particle data includes the particle size distribution and phase distribution of cloud particles in front of the wing; flight parameter sensing data are obtained using a flight parameter sensing module, including aircraft flight parameter data and wing surface temperature, wherein the aircraft flight parameters include speed, altitude, atmospheric density, air pressure, speed of sound, and dynamic viscosity; and planar electrode ice detection data are obtained using a planar electrode ice detector, including ice type and thickness data, excitation frequency, capacitance at the corresponding excitation frequency, and surface temperature of the planar electrode ice detector; the ice type and thickness data include ice type and ice thickness changes of ice on the aircraft wing; Step 2: Perform preliminary data processing on the cloud particle data and aircraft flight parameters obtained in Step 1. First, use fluid dynamics calculation methods to obtain the ambient flow field distribution. Then, based on the air flow field, solve the trajectory equation of the cloud particle motion to obtain the cloud particle impact efficiency. Finally, combined with the cloud particle impact efficiency, calculate the mass flow rate of the cloud particles impacting the wing surface. Step 3: Construct a hybrid model based on a BP neural network and a recurrent neural network, including a BP neural network and an RNN neural network. The BP neural network is used to process multidimensional data, including cloud particle data and aircraft flight parameters. The RNN neural network uses long short-term memory (LSTM) units to model time series data to capture the temporal evolution of the icing process. The ice thickness and ice type data observed in real time by the planar electrode icing detector is used as a physical constraint condition, and a constraint loss function based on the physical equation is constructed. The real-time ice thickness and ice type data measured by the planar electrode icing detector and the predicted ice thickness and ice type data ultimately output by the hybrid model are jointly optimized. Step 4: Set up multi-level warning rules. When the prediction result output by the hybrid model exceeds the threshold, a wing icing warning corresponding to the icing risk level will be issued.

2. The method for full-cycle wing icing perception and early warning according to claim 1, characterized in that: The step 2 is specifically as follows: In step 2.1, the air flow field is first calculated based on CFD computational fluid dynamics theory. The turbulence model uses the k-ε equation, and the fluid dynamics control theory uses the 3-dimensional Navier-Stokes NS equation. Substituting the cloud particle data and aircraft flight parameters obtained in step 1 into the turbulence model and the three-dimensional Navier-Stokes equations can solve the environmental flow field distribution to obtain the air flow field, wherein the environmental flow field distribution includes air pressure distribution, velocity distribution, temperature distribution, and shear stress distribution; Step 2.2: Based on the air flow field obtained in step 2.1, the cloud particle Euler governing equation is established as follows: ， in, is the cloud particle volume fraction; is the velocity vector of the cloud particles; is the dynamic viscosity of the cloud particles; is the cloud particle density; is the diameter of the cloud particles; is the drag coefficient; is the Reynolds number; is the velocity vector of the air; is the air density; is the acceleration due to gravity; represents divergence; By solving the cloud particle data in step 1, the aircraft flight parameters and the ambient flow field distribution in step 2.1, equations (6) and (7) can be used to obtain the impact efficiency of cloud particles on the wing surface. The impact efficiency for: ， in, is the velocity of the impacting cloud particles; is the unit normal vector of the surface; is the far-field velocity; is the volume fraction of cloud particles on the surface; is the volume fraction of cloud particles in the far-field free stream; Step 2.3, combine the cloud particle impact efficiency calculated in step 2.2 with the wing surface The ice mass flow rate m per unit time on the wing surface is expressed as: ， Where S is the effective area of cloud particles impacting the wing, is the density of cloud particles, is the velocity vector of the cloud particles, is the concentration of cloud particles, V(D) is the volume of the particles, D is the diameter of the cloud particles, and d represents the sign of the variable of integration.

3. The method for full-cycle wing icing perception and early warning according to claim 2, characterized in that: In step 2.1: The k-ε equation of the turbulence model is: ,in, is the fluid density; is the turbulent kinetic energy; It’s time; is the average speed; is the fluid dynamic viscosity; is the turbulent viscosity; is related to turbulent kinetic energy The corresponding Prandtl number; is the turbulent kinetic energy due to the mean velocity gradient The generated items; is the turbulent kinetic energy due to buoyancy The generated items; is the turbulent dissipation rate; is the contribution of pulsation expansion in turbulent flow; is a user-defined source term; is the spatial coordinate component in the j direction; is the spatial coordinate component in the i direction; The 3-dimensional Navier-Stokes NS equations are: ， in, is the fluid density; It’s time; is the fluid dynamic viscosity; u, v, and w are velocity vectors Components in the x, y, and z directions; p is the pressure on the fluid element; div() is the divergence; grad() is the gradient; is the generalized source term in the u direction of the momentum conservation equation; is the generalized source term in the v direction of the momentum conservation equation; It is the generalized source term in the w direction of the momentum conservation equation.

4. The method for full-cycle wing icing perception and early warning according to claim 2, characterized in that: The step 3 is specifically as follows: In step 3.1, a BP neural network extracts key feature vectors through a multi-layer feedforward structure. The data obtained in steps 1 and 2 are initialized and processed by the BP neural network and input into the BP neural network. The initialization process converts the cloud particle data, aircraft flight parameter data, capacitance, and excitation frequency from step 1, and the cloud particle impact efficiency and icing mass flow rate from step 2 into multi-dimensional input vectors. A five-layer fully connected structure is set up to input the multi-dimensional input vectors and output predicted ice type and ice thickness data. Step 3.2: Construct an RNN neural network and use long short-term memory (LSTM) units to model the time series data and capture the temporal evolution of the icing process. The predicted ice type and thickness data from the BP neural network in step 3.1 and the ice type and thickness data from the planar electrode ice detector in step 1 are used as input features of the RNN neural network. Dynamic weighted fusion of the predicted ice type and thickness data and the ice type and thickness data detected by the planar electrode ice detector is performed. Then, the LSTM unit is used to process the time series data, and an activation function is selected to map the input data to a value between 0 and 1. In step 3.3, the ice type and thickness data of the planar electrode ice detector in step 1 are introduced into the hybrid model composed of the BP neural network and the RNN neural network as physical constraints. A constraint loss function based on the physical equation is constructed as a dynamic correction module. The constraint loss function shown is as follows: ,in, is the loss of physical equation constraints; T is the total detection time; For ice thickness; is the detection time; is the freezing mass flow rate; is the ice type shape factor function; The physical constraints are weightedly fused with the data-driven loss. The data-driven loss refers to the difference between the prediction results of the hybrid model and the real observation data obtained by the planar electrode ice detector, reflecting the fitting ability of the hybrid model to the existing data. Balance data fitting and physical consistency; form the total loss function: , where Loss is the total output loss function; is a weight hyperparameter that needs to be adjusted according to the data signal-to-noise ratio and the confidence of the physical model; KL is a divergence term used to constrain the consistency of the predicted distribution and the measured distribution; is the predicted ice thickness output by the neural network model; Ice thickness data for flat electrode ice detection; Ice type distribution predicted by the neural network model; is the ice type distribution actually observed by the planar electrode ice detector; represents the mean square error.

5. The method for full-cycle wing icing perception and early warning according to claim 4, characterized in that: In step 3.1, the BP neural network adopts an adaptive activation function, which is: , where a and b in the activation function are learnable parameters that change with the BP neural network training process; is the activation function, is the output of the current neuron, z is the weighted linear combination of the current neuron, and the input value of the activation function.

6. The method for full-cycle wing icing perception and early warning according to claim 4, characterized in that: In step 3.2, the Sigmoid activation function is used to map the input data to a range between 0 and 1.

7. The method for full-cycle wing icing perception and early warning according to claim 1, characterized in that: The step 4 is specifically as follows: The icing risk level is divided into low risk, medium risk, high risk and extremely high risk according to ice thickness, ice type and temperature. When the prediction result output by the hybrid model exceeds a threshold, a low risk warning, medium risk warning, high risk warning or extremely high risk warning is triggered according to the icing risk level, realizing real-time monitoring and warning processing of wing icing.

8. The method for full-cycle wing icing perception and early warning according to claim 7, characterized in that: The principles for dividing the icing risk levels are as follows: ice thickness <5 mm, frost ice, and temperatures below -10°C are low risk; ice thickness 5-15 mm, frost ice / mixed ice, and temperatures between -10°C and -5°C are medium risk; ice thickness 15-30 mm, clear ice, and temperatures between -5°C and -1°C are high risk; ice thickness >30 mm, clear ice, and temperatures between -1°C and 0°C are extremely high risk.

9. The method for full-cycle wing icing perception and early warning according to claim 7, characterized in that: The thresholds are ice thickness, ice type, and temperature range corresponding to each icing risk level.

Images