Aircraft ice accretion forecasting model training method and aircraft ice accretion forecasting method

Abstract

The invention provides an airplane ice accretion forecasting model training method and an airplane ice accretion forecasting method. The airplane ice accretion forecasting model training method and the airplane ice accretion forecasting method are used for solving the problem that accuracy of forecasting ice accretion frequency data is insufficient. The airplane ice accretion forecasting model training method comprises the following steps: acquiring ice accretion frequency sample data; identifying icing states of frequency data in a plurality of continuous duration windows from the icing frequency sample data; determining a plurality of ice accumulation samples with ice accumulation in the ice accumulation state and a plurality of non-ice accumulation samples with no ice accumulation in the ice accumulation state from the ice accumulation frequency sample data; for each ice accretion sample in the plurality of ice accretion samples, determining an intensity grade of the ice accretion sample according to the frequency drop rate; and taking the ice accretion frequency sample data as training data, taking the ice accretion state and the intensity grade as training labels, training the neural network model, and obtaining an airplane ice accretion forecasting model used for predicting the corresponding ice accretion state and the intensity grade according to the ice accretion frequency data.

Description

Technical Field

[0001] This application relates to the technical fields of aerospace safety, artificial intelligence, atmospheric detection and aircraft environmental control, and more specifically, to a training method for an aircraft icing prediction model and an aircraft icing prediction method. Background Technology

[0002] Current icing forecasting methods (such as the IC index method) mainly rely on static meteorological parameters (such as temperature and relative humidity) to construct fixed formulas for prediction. Their limitation lies in the difficulty of dynamically capturing the complex physical mechanisms involved in the icing process. Furthermore, traditional methods are heavily reliant on key variables based on empirical rules, such as vertical motion and liquid water content. These parameters suffer from insufficient spatiotemporal resolution in actual observations, further exacerbating prediction bias. Therefore, these methods often exhibit insufficient accuracy in forecasting icing frequency data when dealing with rapidly changing meteorological conditions. Summary of the Invention

[0003] The purpose of this application is to provide a training method for an aircraft icing forecast model and an aircraft icing forecast method to improve the problem of insufficient accuracy in forecasting icing frequency data.

[0004] This application provides a method for training an aircraft icing prediction model, comprising: acquiring icing frequency sample data, which is obtained by detecting the aircraft using an airborne icing detector; identifying the icing state of frequency data within multiple consecutive time windows from the icing frequency sample data, the icing state including icing and no icing; determining multiple icing samples with icing and multiple non-icing samples with no icing from the icing frequency sample data; determining the intensity level of each icing sample based on the frequency decrease rate; training a neural network model using the icing frequency sample data as training data and the icing state and intensity level as training labels to obtain an aircraft icing prediction model, which is used to predict the corresponding icing state and intensity level based on the icing frequency data. In the implementation of the above scheme, the neural network model is trained using icing frequency sample data from airborne icing detectors as training data and icing state and intensity level determined based on the frequency drop rate as training labels. The neural network model can automatically learn the complex nonlinear mapping relationship between frequency data and icing state and intensity level, so that the trained aircraft icing forecast model can more accurately capture the dynamic characteristics of the icing process. This breaks through the dependence of traditional methods on key variables of static meteorological parameters and empirical rules, and improves the accuracy of forecasting icing frequency data.

[0005] Optionally, in this embodiment of the application, identifying the icing state of frequency data within multiple consecutive duration windows from the icing frequency sample data includes: for each consecutive duration window, determining whether the frequency data within the consecutive duration window meets preset standard conditions. The preset standard conditions include: the airborne icing detector is inside the cloud layer throughout the consecutive duration window, and the detection temperature is negative; the icing frequency data shows a continuous decreasing trend; the frequency value per second satisfies a non-increasing relationship; and the frequency difference between the initial and final times of the frequency data within the consecutive duration window is greater than a frequency threshold. If so, the icing state of the frequency data within the consecutive duration window is determined to be icy; otherwise, the icing state of the frequency data within the consecutive duration window is determined to be icy-free. In the implementation of the above scheme, by setting multiple continuous duration windows and combining them with multi-dimensional conditions to judge the ice accumulation status, the accuracy and reliability of ice accumulation detection can be significantly improved. Moreover, by comprehensively judging multiple conditions and adopting the analysis method of continuous duration windows, the determination of the ice accumulation status is more comprehensive and systematic, taking into account both the necessary conditions for ice accumulation formation and the dynamic characteristics in the ice accumulation process, thereby improving the accuracy of ice accumulation status identification.

[0006] Optionally, in this embodiment, training the neural network model includes: using the neural network model to infer icing frequency sample data to obtain the predicted state and predicted level of the icing frequency sample data; calculating a first loss value between the icing state and the predicted state, and a second loss value between the intensity level and the predicted level using a loss function; updating the model parameters of the neural network model based on the total loss value determined by the first and second loss values, until the total loss value reaches a preset threshold or the number of training iterations reaches a preset number, thus obtaining an aircraft icing forecast model. In the implementation of the above scheme, by simultaneously calculating the dual loss values ​​of icing state and intensity level, multi-dimensional optimization of the neural network model is achieved, enabling the model to learn more comprehensively the key features of icing forecast during training, thereby improving the accuracy and reliability of the forecast. This dual loss mechanism ensures that the model not only focuses on the existence of icing but also on the severity of icing, thus maintaining high prediction accuracy even under complex weather conditions. Furthermore, by directly inputting the icing frequency sample data into the neural network model for inference, end-to-end icing forecasting is achieved, reducing the cumbersome feature engineering steps in traditional methods. This direct data processing method not only simplifies the forecasting process but also avoids information loss caused by human feature selection, thereby improving the comprehensiveness and objectivity of the forecast.

[0007] Optionally, in this embodiment, the neural network model includes: an input layer, an output layer, and multiple hidden layers. Inference using the neural network model on icing frequency sample data includes: extracting features from the icing frequency sample data through the input layer to obtain data features; processing the data features through multiple hidden layers to obtain processed data features; and predicting the processed data features through the output layer to obtain the prediction state and prediction level corresponding to the icing frequency sample data. In the implementation of the above scheme, the reasonable structural design of the entire neural network model and the collaborative work of the input layer, hidden layers, and output layer enable the model to efficiently process icing frequency data, significantly improving prediction speed and efficiency, and making it suitable for real-time or near-real-time icing prediction applications. Furthermore, by processing the data features layer by layer through multiple hidden layers, the deep features of the icing data can be gradually extracted and optimized, enhancing the model's ability to recognize complex data patterns, making the prediction results more accurate and reliable.

[0008] Optionally, in this embodiment, after obtaining the aircraft icing prediction model, the method further includes: acquiring icing frequency data to be processed, which is obtained by detecting the target aircraft using an airborne icing detector; and using the aircraft icing prediction model to predict the icing frequency data to be processed, thereby obtaining the icing state and intensity level corresponding to the icing frequency data. In the implementation of the above scheme, by using the icing state and intensity level as the model's prediction output, more refined icing information can be provided for flight decision-making, enabling pilots or autopilot systems to take more targeted countermeasures based on the specific severity of icing, thereby improving flight safety. Furthermore, this technical solution achieves a direct combination of airborne detection data and the prediction model, reducing intermediate data processing steps, lowering system complexity, and improving the reliability of prediction results, because the model is directly calculated based on actual detection data, avoiding errors that may be introduced during data conversion.

[0009] This application also provides an aircraft icing prediction method, comprising: acquiring icing frequency data to be processed, wherein the icing frequency data to be processed is obtained by detecting the target aircraft through an airborne icing detector; and predicting the icing frequency data to be processed using an aircraft icing prediction model to obtain the icing state and intensity level corresponding to the target aircraft. In the implementation of the above scheme, by using an aircraft icing prediction model to predict the icing frequency data, the complex physical process of icing can be transformed into a computable mathematical model. Through the model's self-learning and optimization capabilities, the icing state and intensity level can be predicted more accurately, providing a more reliable basis for flight safety decisions. Furthermore, by combining real-time data from the airborne detector with the intelligent analysis of the prediction model, dynamic monitoring and prediction of the aircraft icing state are achieved. This not only enables timely detection of icing risks but also predicts the development trend of icing, thereby providing pilots and ground control personnel with more comprehensive safety warning information.

[0010] Optionally, in this embodiment, the aircraft icing prediction model includes: an input layer, an output layer, and multiple hidden layers. Predicting icing frequency data to be processed using the aircraft icing prediction model includes: extracting features from the icing frequency data to be processed through the input layer to obtain data features; processing the data features through multiple hidden layers to obtain processed data features; and predicting the processed data features through the output layer to obtain the icing state and intensity level corresponding to the target aircraft. In the implementation of the above scheme, the hierarchical processing structure of the input layer, hidden layer, and output layer makes the icing prediction process have a clear feature extraction-processing-prediction flow. This structured processing method not only improves the systematic nature of the prediction but also enables the model to better adapt to different types and sources of icing data, enhancing the model's generalization ability. Furthermore, the feature extraction and processing are automatically completed by the deep learning model, avoiding the tedious steps of manually designing features and selecting parameters in traditional methods, greatly improving the efficiency and automation of icing prediction, while reducing errors caused by human factors.

[0011] This application embodiment also provides an aircraft icing forecast model training device, comprising: a sample data acquisition module for acquiring icing frequency sample data, wherein the icing frequency sample data is obtained by detecting the aircraft using an airborne icing detector; an icing state identification module for identifying the icing state of frequency data within multiple consecutive time windows from the icing frequency sample data, wherein the icing state includes icing and no icing; an icing sample determination module for determining multiple icing samples with icing state and multiple non-icing samples with icing state from the icing frequency sample data; an intensity level determination module for determining the intensity level of multiple icing samples based on the frequency decrease rate for each of the multiple icing samples; and a neural network training module for training a neural network model using the icing frequency sample data as training data and the icing state and intensity level as training labels to obtain an aircraft icing forecast model, wherein the aircraft icing forecast model is used to predict the corresponding icing state and intensity level based on the icing frequency data.

[0012] Optionally, in this embodiment, the icing state identification module includes: a frequency data judgment submodule, used to determine whether the frequency data within each of multiple consecutive duration windows meets preset standard conditions, the preset standard conditions including: the airborne icing detector is inside the cloud layer throughout the consecutive duration window, the detection temperature is negative, the icing frequency data shows a continuous decreasing trend, the frequency value per second satisfies a non-increasing relationship, and the frequency difference between the initial and final times of the frequency data within the consecutive duration window is greater than a frequency threshold; and an icing state determination submodule, used to determine the icing state of the frequency data within the consecutive duration window as having icing if the frequency data within the consecutive duration window meets the preset standard conditions, and to determine the icing state of the frequency data within the consecutive duration window as having no icing if the frequency data within the consecutive duration window does not meet the preset standard conditions.

[0013] Optionally, in this embodiment, the neural network training module includes: a sample data inference submodule, used to use a neural network model to infer the icing frequency sample data to obtain the predicted state and predicted level of the icing frequency sample data; a loss value calculation submodule, used to calculate a first loss value between the icing state and the predicted state, and a second loss value between the intensity level and the predicted level through a loss function; and a model parameter update submodule, used to update the model parameters of the neural network model according to the total loss value determined by the first loss value and the second loss value, until the total loss value reaches a preset threshold or the number of training iterations reaches a preset number, to obtain an aircraft icing forecast model.

[0014] Optionally, in this embodiment of the application, the neural network model includes: an input layer, an output layer, and multiple hidden layers; the sample data inference submodule includes: a data feature acquisition unit, used to extract features from the ice accumulation frequency sample data through the input layer to obtain data features; a data feature processing unit, used to process the data features through multiple hidden layers to obtain processed data features; and a data feature prediction unit, used to predict the processed data features through the output layer to obtain the prediction state and prediction level corresponding to the ice accumulation frequency sample data.

[0015] Optionally, in this embodiment of the application, the aircraft icing forecast model training device further includes: a frequency data acquisition module, used to acquire icing frequency data to be processed, the icing frequency data to be processed being obtained by detecting the target aircraft through an airborne icing detector; and a frequency data prediction module, used to predict the icing frequency data to be processed through the aircraft icing forecast model, and obtain the icing state and intensity level corresponding to the icing frequency data to be processed.

[0016] This application also provides an aircraft icing prediction device, including: a frequency data acquisition module for acquiring icing frequency data to be processed, wherein the icing frequency data to be processed is obtained by detecting the target aircraft through an airborne icing detector; and a state level prediction module for predicting the icing frequency data to be processed through an aircraft icing prediction model to obtain the icing state and intensity level corresponding to the target aircraft.

[0017] Optionally, in this embodiment, the aircraft icing prediction model includes: an input layer, an output layer, and multiple hidden layers; the state level prediction module includes: a data input extraction submodule, used to extract features from the icing frequency data to be processed through the input layer to obtain data features; a data feature processing subunit, used to process the data features through multiple hidden layers to obtain processed data features; and a state level prediction submodule, used to predict the processed data features through the output layer to obtain the icing state and intensity level corresponding to the target aircraft.

[0018] This application also provides an electronic device, including a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and the machine-readable instructions are executed by the processor to perform the methods described above.

[0019] This application also provides a computer-readable storage medium storing a computer program, which is executed by a processor to perform the methods described above.

[0020] This application also provides a computer program product, including: a computer program or computer instructions, which are executed by a processor to perform the method described above. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 The diagram shows a flowchart of the aircraft icing prediction model training method provided in an embodiment of this application. Figure 2 The diagram shows the frequency percentile distribution of different probe frequency drop rates provided in the embodiments of this application; Figure 3 The flowchart shown is a schematic diagram of the aircraft icing prediction method provided in an embodiment of this application; Figure 4 The diagram shown is a structural schematic of the aircraft icing prediction model training device provided in an embodiment of this application. Figure 5 The diagram shown is a structural schematic of the aircraft icing prediction device provided in an embodiment of this application. Figure 6 The diagram shows a structural schematic of an electronic device provided in an embodiment of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in the embodiments of this application are for illustrative and descriptive purposes only and are not intended to limit the protection scope of the embodiments of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in the embodiments of this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of the embodiments of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0024] Furthermore, the described embodiments are merely a part of the embodiments of this application, and not all of them. The components of the embodiments of this application described and illustrated herein can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed embodiments of this application, but merely to illustrate selected embodiments of this application.

[0025] It is understood that the terms "first" and "second" in the embodiments of this application are used to distinguish similar objects. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that "first" and "second" do not necessarily imply that they are different. In the description of the embodiments of this application, the term "and / or" is merely a description of the association relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship. The term "multiple" refers to two or more (including two), and similarly, "multiple groups" refers to two or more groups (including two groups).

[0026] It should be noted that the aircraft icing forecast model training method provided in this application embodiment can be executed by an electronic device. Here, electronic device refers to a device terminal or server with the function of executing computer programs. Device terminals include, for example, smartphones, personal computers, tablets, personal digital assistants, or mobile internet devices. Servers refer to devices that provide computing services through a network. Servers include, for example, x86 servers and non-x86 servers. Non-x86 servers include, for example, mainframes, minicomputers, and UNIX servers.

[0027] Current icing forecasting methods (such as the IC index method) primarily rely on static meteorological parameters (such as temperature and relative humidity) to construct fixed formulas for prediction. These methods calculate the icing probability by simplifying the linear relationship between temperature and humidity; for example, the IC index method assumes that the probability of icing increases significantly within a specific temperature range (such as near 0°C) and under high humidity conditions. However, such methods neglect the influence of dynamic factors such as flight speed, cloud droplet concentration distribution, and airflow disturbances on the icing process. For example, flight speed directly affects the collision efficiency between the aircraft and supercooled water droplets in the cloud, while the spatial heterogeneity of cloud droplet concentration leads to local differences in icing intensity. Because these dynamic factors are not included in the model, in actual flight, there may be overestimation of icing intensity (e.g., misjudging a low-risk area as high-risk) or underestimation (e.g., failure to trigger a warning at the edge of a rapidly changing cloud cluster), thus reducing the reliability of the forecast.

[0028] Traditional icing forecasting methods heavily rely on key variables based on empirical rules, such as vertical motion (e.g., updrafts) and liquid water content (LWC). For example, vertical motion determines the distribution altitude of supercooled water droplets in clouds, while LWC is directly related to the accumulation rate of icing mass. However, these parameters suffer from significant limitations in spatiotemporal resolution in actual observations. Satellite remote sensing or ground-based radar measurements of LWC are typically measured in kilometers, making it difficult to capture minute changes along aircraft flight paths; vertical motion data relies more on model simulations, which deviate from actual atmospheric conditions. These observational limitations distort model input parameters, further exacerbating the bias in icing predictions.

[0029] The above analysis shows that traditional icing forecasting methods are significantly inaccurate in predicting icing frequency data when dealing with rapidly changing weather conditions. For example, when an aircraft crosses the boundary layer of an icing cloud or encounters a sudden change in temperature and humidity gradient, the static formula cannot dynamically adjust the prediction results. Furthermore, the icing process itself has nonlinear and hysteresis characteristics. For instance, in the early stages of ice formation, ice may accumulate rapidly due to collisions with tiny supercooled water droplets, but traditional methods, lacking real-time monitoring of frequency changes, struggle to capture this critical point. Therefore, traditional icing forecasting methods often exhibit insufficient accuracy in predicting icing frequency data when dealing with rapidly changing weather conditions.

[0030] Please see Figure 1 The diagram illustrates a flowchart of the aircraft icing forecast model training method provided in this application embodiment. The main idea of ​​this method is to use icing frequency sample data obtained from real-world aircraft detection by an airborne icing detector as training data, and icing state and intensity level as training labels to train a neural network model, thereby obtaining an aircraft icing forecast model. The neural network model can automatically learn the complex nonlinear mapping relationship between frequency data and icing state and intensity level, enabling the trained model to more accurately capture the dynamic characteristics of the icing process. This overcomes the reliance of traditional methods on key variables such as static meteorological parameters and empirical rules, improving the accuracy of forecasting icing frequency data. The implementation methods of the above-described aircraft icing forecast model training method may include: Step S110: Obtain icing frequency sample data. The icing frequency sample data is obtained by detecting the aircraft using an airborne icing detector.

[0031] Icing frequency sample data refers to frequency signal data reflecting aircraft icing conditions collected by airborne rosemount icing detectors. This data is used for subsequent analysis of icing status and intensity levels. Examples include signal variation data in specific frequency bands recorded by airborne icing detectors, or frequency fluctuation data reflecting icing conditions on aircraft wings or engine inlets. The aforementioned icing frequency sample data can include physical quantities such as temperature, specific humidity, relative humidity, temperature-dew point difference, and vertical airflow, as well as icing indices such as FIP, RAP, NAWAU, and RAOB.

[0032] Airborne icing detectors are sensor devices installed on aircraft to detect icing on the aircraft's surface or critical components (such as wings and engines) and output frequency signal data. These airborne icing detectors can include the Type 871 icing detector mounted on the King Air platform, microwave resonant icing sensors, infrared optical icing detectors, or ultrasonic icing monitoring devices, among others.

[0033] The aforementioned airborne icing detector works by detecting ice thickness or distribution using physical or optical principles (such as microwave reflectivity changes or infrared scattering), converting the detection results into electrical or frequency signals for output. When ice accumulates on the probe of the airborne icing detector, the increased ice mass lowers the probe's resonant frequency, and the icing detector outputs a signal proportional to the accumulated ice mass on the probe. When the ice accumulation on the probe reaches a preset value, the probe will begin a 7-second de-icing process, followed by a new icing process. The sensing area of ​​this icing probe is the entire surface of a cylindrical probe; the same ice mass on the cylindrical surface will result in the same frequency change, unaffected by differences in ice density or geometric distribution.

[0034] Step S120: Identify the icing status of frequency data within multiple consecutive time windows from the icing frequency sample data. The icing status includes icing and no icing.

[0035] Icing status is a binary classification result characterizing the presence of ice accumulation on specific parts of an aircraft, including "with icing" and "without icing." For example, if ice coverage is detected at the leading edge of the wing based on icing frequency sample data, then the status is "with icing," and the corresponding icing frequency sample is called an icing sample. Similarly, if no icing signal is detected at the engine air intake based on icing frequency sample data, then the status is "without icing." In practice, the icing status can be automatically determined through threshold analysis or pattern recognition algorithms (such as mutation detection) of the icing frequency sample data.

[0036] A continuous duration window is a fixed time interval for segmenting ice accumulation frequency data. It is used to analyze the temporal characteristics of frequency changes. For example, the sliding window length can be set to 10 seconds, with a step size of 1 or 2 seconds.

[0037] Step S130: Identify multiple ice-accumulated samples with ice accumulation status and multiple non-ice-accumulated samples with ice accumulation status from the ice accumulation frequency sample data.

[0038] Icing samples are data segments selected from icing frequency sample data and labeled as "iced". For example, a 30-second data segment containing a sudden drop in frequency, or a 1-minute data segment recorded by the wing sensor under icing conditions.

[0039] Understandably, by acquiring sample data of icing frequency directly detected by airborne icing detectors and identifying the icing state and intensity level from them, the authenticity and accuracy of the training data can be ensured, thereby improving the prediction accuracy of the aircraft icing forecast model.

[0040] Step S140: For each of the multiple ice accumulation samples, determine the intensity level of the ice accumulation sample based on the frequency decrease rate.

[0041] The frequency drop rate is the slope of the frequency signal in the ice accumulation sample as a function of time. It is used to quantify the severity of ice formation, such as a steep drop of 50 Hz per second, or a gradual drop (e.g., 5 Hz per second).

[0042] Intensity levels are classifications of the severity of icing based on the rate of frequency decay (e.g., light / moderate / severe). For example, a rate <10Hz / s indicates light icing, a rate of 10-50Hz / s indicates moderate icing, and a rate >50Hz / s indicates severe icing.

[0043] Understandably, by selecting samples with icing from the icing frequency data and determining the intensity level of these samples based on the rate of frequency decline, the severity of icing can be captured more accurately, thereby improving the model's ability to predict icing intensity. Furthermore, by analyzing icing status and intensity levels through frequency data within continuous time windows, the changing trends of icing can be dynamically captured, enabling the model to predict the icing development process and thus improving the timeliness and reliability of forecasts.

[0044] Step S150: Using icing frequency sample data as training data and icing state and intensity level as training labels, train the neural network model to obtain the aircraft icing forecast model. The aircraft icing forecast model is used to predict the corresponding icing state and intensity level based on the icing frequency data.

[0045] Neural network models are machine learning models used to learn the mapping relationship between icing state and intensity level from icing frequency data. In practice, convolutional neural networks (CNNs), long short-term memory networks (LSTMs), or one-way propagation multilayer feedforward neural network models can be used. For ease of understanding and explanation, the following detailed explanation will use the one-way propagation multilayer feedforward neural network model as an example.

[0046] The aircraft icing forecasting model is a neural network model trained on a neural network. It can take icing frequency data as input and output predictions of icing state and intensity levels in real time. This model can be used as an embedded airborne icing risk assessment model. By training the neural network model with icing frequency sample data as training data and icing state and intensity levels as training labels, it can simultaneously predict icing state and intensity levels, achieving multi-dimensional output of icing forecasts and enhancing the model's practicality and comprehensiveness.

[0047] In the implementation of the above scheme, the neural network model is trained using icing frequency sample data from the airborne icing detector as training data, and the icing state and intensity level determined based on the frequency decrease rate as training labels. The neural network model can automatically learn the complex nonlinear mapping relationship between frequency data and icing state and intensity level, improving predictions in certain special cases. These special cases include: small fluctuations in the icing detector frequency due to factors such as its own vibration stability, environmental wind, and cloud particle collisions; the icing detector's frequency remaining at a low value due to prolonged icing coverage; and frequency recovery during icing detector de-icing. Therefore, the trained aircraft icing forecast model can more accurately capture the dynamic characteristics of the icing process, thus breaking through the reliance on static meteorological parameters and empirical rules in traditional methods, ultimately improving the accuracy of icing frequency forecasting.

[0048] Optionally, as an alternative implementation of step S110 above, icing frequency sample data can be extracted from the icing cloud environment feature dataset, expert knowledge base dataset, airborne detection data, and / or civil aviation aircraft report dataset. The civil aviation aircraft report dataset contains subjective reports from pilots, which are subjective, while the airborne detection data is objective data collected by the probes of airborne icing detectors. The aforementioned icing frequency sample data can be obtained through real-time detection of the aircraft during flight using airborne icing detectors. For example, the airborne icing detector can be a vibration detector with an operating frequency range of 20-40kHz and a sampling frequency of not less than 100Hz. After extracting the icing frequency sample data, the original detection data can be preprocessed, including: filtering: using a Butterworth low-pass filter to eliminate high-frequency noise, with a cutoff frequency set to 50kHz; normalization: normalizing the frequency data to the [0,1] interval; and time alignment: synchronizing the detection data with flight parameters (altitude, temperature, etc.).

[0049] As an optional implementation of step S120 above, the implementation of identifying the icing state of frequency data within multiple consecutive time windows from icing frequency sample data may include: Step S121: For each of the multiple consecutive duration windows, determine whether the frequency data within the consecutive duration window meets the preset standard conditions. The preset standard conditions include: the airborne icing detector is inside the cloud layer within the consecutive duration window, the detection temperature is negative, the icing frequency data shows a continuous decreasing trend, the frequency value per second satisfies a non-increasing relationship, and the frequency difference between the initial and final times of the frequency data within the consecutive duration window is greater than the frequency threshold.

[0050] The implementation of step S121 above can be exemplified by, for example, the process of airborne aircraft icing detection, the correlation between aircraft platform flight speed and icing formation time (when an aircraft is flying in clouds, flight speed directly affects the dynamic characteristics of the icing process), and the working principle of the response characteristics of airborne icing detection equipment (the response time of airborne icing detectors is usually in the millisecond range, but changes in their output frequency need to accumulate continuously for a certain period of time to reflect the true icing state). Combined with the mechanism of aircraft icing formation (icing formation depends on the collision and freezing process of supercooled water droplets with the aircraft surface), a continuous duration window is set to 10 seconds (represented by 10S). Then, it is determined whether the frequency data within the continuous 10S window meets preset standard conditions. These preset standard conditions include: the airborne icing detector... Within a 10-second window, the data is all within the cloud layer, the detected temperature is negative, and the icing frequency data shows a continuous decreasing trend (e.g., assuming the initial time is T, then the icing probe frequency at time T+9s is less than the icing probe frequency at time T). Furthermore, the frequency values ​​per second satisfy a non-increasing relationship (e.g., in a 10-second sample, the icing probe frequency in the next second must be less than or equal to the previous second, i.e., the icing probe frequency at T+1s ≤ the icing probe frequency at time T, and so on, with the icing probe frequency at T+2s ≤ the icing probe frequency at T+1s, and the icing probe frequency at T+3s ≤ the icing probe frequency at T+2s). Additionally, the frequency difference between the initial and final times within the continuous window is greater than a frequency threshold (e.g., the difference between the icing probe frequencies at time T and time T+9s reaches 30Hz or more).

[0051] It is understandable that the aforementioned T+9s timeframe can be determined based on aircraft speed. For example, at a flight speed of 360 km / h, approximately 1 km is traveled every 10 seconds, representing the icing distribution within a 1 km range. Due to the high flight speed of aircraft and the high sampling frequency of icing detection, the atmospheric environmental parameters provided by the neural network model for inference on the icing frequency sample data have a resolution at the kilometer level. Therefore, a resolution for the atmospheric environmental parameters that is as small as possible can be selected, and this resolution can serve as a standard for determining the time window for icing samples.

[0052] Step S122: If the frequency data within the continuous duration window meets the preset standard conditions, then the icing state of the frequency data within the continuous duration window is determined to be icing.

[0053] For example, the implementation of step S122 above is as follows: assuming the continuous duration window is set to 10 seconds (represented by 10S), then for each continuous duration window in the multiple continuous duration windows, if the frequency data in the continuous duration window meets all the preset standard conditions, the icing state of the frequency data in the continuous duration window can be determined as having icing. In other words, assuming the initial time is T, the airborne icing detector is inside the cloud layer for a continuous 10-second window, the detection temperature is negative, the icing probe frequency at time T+9s is less than the icing probe frequency at time T, and the 10-second sample satisfies that the icing probe frequency in the next second is less than or equal to the previous second, i.e., the icing probe frequency at T+1s ≤ the icing probe frequency at time T, and so on, the icing probe frequency at T+2s ≤ the icing probe frequency at T+1s, the icing probe frequency at T+3s ≤ the icing probe frequency at T+2s, and the difference between the icing probe frequencies at time T and T+9s reaches 30Hz or more, then the icing status of the frequency data within the continuous time window can be determined as having icing. The following is a detailed description and explanation of each condition.

[0054] When icing only occurs when an aircraft passes through clouds containing supercooled water droplets, effective icing is unlikely to form if the aircraft is outside the cloud or at the cloud edge, even if the detected temperature is below 0°C. By requiring the sample to remain inside the cloud for a continuous 10 seconds, false alarms caused by brief passages through cloud boundaries or localized weather disturbances can be eliminated. Furthermore, by requiring the airborne icing detector to remain inside the cloud and at a negative temperature for a continuous time window, the physical conditions for icing formation are ensured, eliminating the possibility of false alarms in non-icing conditions.

[0055] The formation of ice requires a temperature below 0°C. If the ambient temperature is above 0°C, even if liquid water droplets are present (such as freezing rain), they may not freeze due to the surface dynamic heating effect. Continuous negative temperature detection for 10 seconds ensures that the necessary conditions for ice formation are met, avoiding erroneous markings caused by instantaneous temperature fluctuations (such as equipment errors or local airflow disturbances).

[0056] The remaining conditions above eliminate various interference factors through the monotonicity, continuity and threshold of frequency changes, ensuring the reliability of the icing samples. Specifically, the frequency at T+9s is less than the frequency at T (overall downward trend) to exclude instantaneous frequency rebounds caused by equipment vibration, environmental wind disturbance or particle collision. During the icing process, the probe frequency will gradually decrease due to the increase in ice mass (knowledge base [1]). If the frequency at T+9s is not significantly lower than that at T, it indicates that the frequency change may be caused by non-icing factors (such as probe vibration or short-term collision of particles in the cloud). For example, equipment vibration may cause the frequency to rise briefly and then fall back, but the overall trend does not meet the requirement of decreasing, so it is filtered out. By requiring the icing frequency data to show a continuous downward trend and the frequency value per second to meet the non-increasing relationship, the typical characteristics of frequency change during the icing process can be effectively captured, avoiding misjudgment caused by instantaneous interference or noise, and improving the stability of icing state identification.

[0057] The non-incremental frequency per second (T+n+1s ≤ T+ns) above is to ensure the continuity and stability of frequency changes and avoid misjudgment caused by instantaneous noise or local disturbances. The icing process is gradual, and the increase in the mass of the ice layer will cause the probe frequency to gradually decrease (Knowledge Base [6]). If the frequency rebounds (such as T+1s>T), it may be caused by the following reasons: strong winds may cause airflow disturbance on the probe surface, causing frequency fluctuations; ice crystals or raindrops in the clouds may hit the probe, temporarily changing its vibration characteristics; or the sensor itself may have measurement errors or electronic noise. Therefore, the non-incremental constraint requires that the frequency can only remain stable or decrease, thereby filtering out the above interference factors.

[0058] The aforementioned frequency difference of ≥30Hz between time T and T+9s (threshold condition) is to distinguish between valid icing and invalid disturbances, ensuring that frequency changes reach a identifiable cumulative icing mass threshold. Frequency fluctuations caused by environmental wind, particle collisions, or equipment vibration are usually small (e.g., within a few Hz), failing to meet the 30Hz threshold requirement. If the probe surface is covered by ice for a long time (e.g., residual ice after de-icing), the frequency may stabilize at a low value without further change; in this case, the frequency difference is close to 0Hz and will be excluded. During the de-icing stage, the probe frequency may briefly rise (e.g., the ice melts after heating), but the overall trend does not meet the non-increasing requirement (e.g., T+1s > time T), thus being filtered out. It is understandable that by setting a frequency difference threshold, minute frequency fluctuations can be filtered out; only when the frequency change reaches a certain magnitude is icing detected, thereby reducing the false alarm rate and ensuring the reliability of icing detection.

[0059] Step S123: If the frequency data within the continuous duration window does not meet the preset standard conditions, then the icing status of the frequency data within the continuous duration window is determined to be no icing.

[0060] Understandably, the aforementioned preset standard conditions can effectively reduce misjudgments of ice samples caused by some special circumstances. These special circumstances include: slight fluctuations in the frequency of the ice probe due to factors such as its own vibration stability, environmental wind, and collisions of particles in the cloud; the frequency of the ice probe remaining at a low value without change due to long-term ice coverage; and the frequency rebounding during the de-icing process of the ice probe.

[0061] For example, if the frequency data within the continuous duration window does not meet any preset standard conditions, i.e., it cannot meet one of these conditions, these conditions include: the airborne icing detector is inside the cloud layer for the entire continuous duration of 10 seconds, and the detection temperature is negative, and the icing probe frequency at time T+9s is less than the icing probe frequency at time T, and the 10-second sample satisfies that the icing probe frequency in the next second is less than or equal to the previous second, i.e., the icing probe frequency at T+1s ≤ the icing probe frequency at time T, and so on, the icing probe frequency at T+2s ≤ the icing probe frequency at T+1s, the icing probe frequency at T+3s ≤ the icing probe frequency at T+2s, and the difference between the icing probe frequencies at time T and T+9s reaches 30Hz or more, then the icing state of the frequency data within the continuous duration window can be determined as ic-free. In the implementation of the above scheme, by comprehensively judging multiple conditions and adopting the analysis method of continuous time window, the determination of the icing state is more comprehensive and systematic. It not only considers the necessary conditions for icing formation, but also pays attention to the dynamic characteristics in the icing process, thereby achieving high-precision identification of the icing state.

[0062] Understandably, the aforementioned pre-defined standard conditions, through the synergistic effect of environmental constraints and frequency dynamic analysis, achieve high-precision screening of icing samples, thereby ensuring that the sample data originates from real icing environments and avoiding data contamination caused by misjudgment of environmental conditions. For example, relying solely on frequency change conditions might misjudge equipment vibration during flight outside clouds as icing. The triple constraints on frequency changes (decreasing trend, non-increasing nature, and threshold) effectively distinguish between real icing and spurious disturbances. Strict time windows and frequency constraints filter out misjudgments caused by equipment noise and environmental disturbances. Furthermore, ensuring the reliability of training data provides high-quality input for subsequent icing prediction models (such as neural networks), improving the model's generalization ability.

[0063] In the above-described process, the synergistic effect of environmental constraints and frequency dynamic analysis enables high-precision screening of icing samples, ensuring that the sample data originates from real icing environments and avoiding data contamination caused by misjudgment of environmental conditions. This technical effect stems from the complementarity of environmental constraints and frequency dynamic analysis. Environmental constraints provide macroscopic verification of environmental conditions, while frequency dynamic analysis captures icing characteristics at the microscopic level. The combination of the two can more comprehensively identify real icing samples and significantly reduce the false positive rate. Furthermore, the triple constraints on frequency changes (decreasing trend, non-increasing nature, and threshold) can effectively distinguish between real icing and spurious disturbances. This technical effect is achieved through strict constraints on frequency changes. The decreasing trend and non-increasing nature ensure the regularity of frequency changes, while the threshold excludes minor disturbances, thus accurately identifying the unique frequency change patterns of icing and avoiding misjudging equipment vibration or other noise as icing. Through strict time windows and frequency constraints, false positives caused by equipment noise and environmental disturbances are filtered out. This technical effect is achieved through dual constraints on time windows and frequency. The time window ensures the continuity of icing events, while the frequency constraint eliminates instantaneous noise interference, effectively filtering out frequency variations unrelated to icing and improving data accuracy. Furthermore, ensuring the reliability of training data provides high-quality input for subsequent icing prediction models (such as neural networks), enhancing the model's generalization ability. This technical effect is achieved through high-precision sample selection; high-quality input data reduces noise interference during model training, allowing the model to focus more on learning the true characteristics of icing, thereby improving its predictive performance on unknown data.

[0064] As an optional implementation of step S130 above, for example: The icing states of frequency data within multiple consecutive time windows have already been identified from the icing frequency sample data, and the icing states include icing and no icing. Therefore, an executable program written in a preset programming language can be used to filter out multiple icing samples with the icing state of icing from the icing frequency sample data based on the icing state of icing and no icing. Assume that the number of filtered icing samples is 570, and the number of filtered non-icing samples is 3253. Optionally, a quality check can also be performed on the filtered samples. This step may include: checking the smoothness of the frequency curve (absolute value of the second derivative < 0.05); and excluding abnormal samples affected by interference (such as frequency abrupt changes > 2kHz).

[0065] Please see Figure 2The illustration shows a frequency distribution diagram of the percentile distribution of different probe frequency drop rates provided in the embodiments of this application. As an optional implementation of step S140 above, for example: for each ice accumulation sample among multiple ice accumulation samples, theoretically, a smaller frequency drop rate corresponds to a lower ice accumulation intensity level, and correspondingly, a larger frequency drop rate corresponds to a higher ice accumulation intensity level. Therefore, the intensity level of frequency data within multiple consecutive time windows can be determined based on the frequency drop rate (e.g., the probe frequency drop rate over 10 seconds). Specifically, after statistically analyzing the percentile distribution of the ice accumulation probe frequency drop rate over 10 seconds, ice accumulation intensity at or below the 25th percentile can be defined as light ice accumulation, i.e., an ice accumulation probe frequency drop rate less than or equal to 4. In practice, ice accumulation intensity above the 25th percentile and below the 75th percentile can also be defined as moderate ice accumulation, i.e., an ice accumulation probe frequency drop rate greater than 4 and less than or equal to 9. In practice, ice accumulation intensity above the 75th percentile can also be defined as severe ice accumulation, i.e., an ice accumulation probe frequency greater than 9.

[0066] As an optional implementation of step S150 above, the above-described implementation of training the neural network model may include: Step S151: Use a neural network model to infer the ice accumulation frequency sample data to obtain the predicted state and predicted level of the ice accumulation frequency sample data.

[0067] It is understood that the implementation of step S151 above is relatively complex. In order to explain the implementation of step S151 above more clearly, it will be described in detail below.

[0068] Step S152: Calculate the first loss value between the icing state and the predicted state, and the second loss value between the intensity level and the predicted level using the loss function.

[0069] An optional implementation of step S152 above includes, for example, calculating a first loss value between the icing state and the predicted state using a binary cross-entropy loss function. This first loss value can be obtained by using... This is represented by the method described above, and a second loss value is calculated between the intensity level and the predicted level using the categorical cross entropy loss function. This second loss value can be expressed as... To express.

[0070] Step S153: Based on the total loss value determined by the first loss value and the second loss value, update the model parameters of the neural network model until the total loss value reaches a preset threshold or the number of training iterations reaches a preset number, and obtain the aircraft icing prediction model.

[0071] One implementation of step S153 above is, for example, using the formula The first and second loss values ​​are calculated to obtain the total loss value. Then, the model parameters of the neural network model are updated using the total loss value until the total loss value reaches a preset threshold or the number of training iterations reaches a preset number, thus obtaining the aircraft icing prediction model. This represents the total loss value. This represents the first loss value. This represents the weight value of the first loss value. This represents the second loss value. This represents the weight value of the second loss value. In the above implementation, by employing a dual termination condition of a preset threshold and the number of training iterations, overfitting or underfitting problems during model training are avoided, ensuring that the model stops training in a timely manner when it reaches optimal performance. This intelligent training termination mechanism not only improves training efficiency but also guarantees the model's generalization ability in practical applications, enabling it to adapt to various meteorological conditions and flight environments. Furthermore, by directly inputting icing frequency sample data into the neural network model for inference, end-to-end icing forecasting is achieved, reducing the cumbersome feature engineering steps in traditional methods. This direct data processing method not only simplifies the forecasting process but also avoids information loss caused by manual feature selection, thereby improving the comprehensiveness and objectivity of the forecast.

[0072] As an optional implementation of step S151 above, the neural network model can be a unidirectional propagation multilayer feedforward neural network model. Specifically, the neural network model may include: an input layer, an output layer, and multiple hidden layers. The input and output layers can be fully connected layers, while the hidden layers can be either fully connected layers or LSTM layers. The implementation method of using the neural network model to infer from icing frequency sample data may include: Step S151a: Extract features from the ice accumulation frequency sample data through the input layer to obtain data features.

[0073] The implementation of step S151a above is as follows: It can be understood that the first layer of the aforementioned neural network model is the input layer. The variable parameters of this input layer can be constructed by fusing icing-related atmospheric environmental parameters from multiple icing diagnostic products. The core is to construct a comprehensive and physically meaningful input feature space by integrating the output results of multiple icing diagnostic algorithms and physical environmental parameters. Specifically, physical quantities such as temperature, specific humidity, relative humidity, temperature-dew point difference, and vertical airflow directly reflect the atmospheric thermodynamic state and are necessary conditions for icing formation. Vertical airflow (rising or sinking motion) affects the distribution height and concentration of water droplets in clouds, thus determining the intensity of icing. Feature extraction from icing frequency sample data through the input layer demonstrates the feature extraction capability of the neural network. The model can automatically learn the nonlinear relationship between physical quantities and icing indices such as FIP, RAP, NAWAU, and RAOB, avoiding the subjectivity of human experience. Feature extraction from icing frequency sample data through the input layer effectively captures the original features of icing data, providing a high-quality data foundation for subsequent processing, thereby improving the accuracy of predictions.

[0074] Step S151b: Perform feature processing on the data features through multiple hidden layers to obtain the processed data features.

[0075] The implementation of step S151b above can be illustrated as follows: It is understood that the combination of neurons in the hidden layers (e.g., 2-4 layers) needs to be optimized in conjunction with task complexity and data dimensionality. Its design goal is to capture key features in icing forecasting and suppress noise. The number of hidden layers can be selected based on specific circumstances (e.g., determined through cross-validation). If the number of hidden layers is relatively small, the neural network is a shallow network. Shallow networks (e.g., a single hidden layer) are suitable for simple tasks but struggle to model complex nonlinear relationships in the icing process. However, if the number of hidden layers is relatively large, the neural network is a deep network. Such deep networks (e.g., 3-4 layers) can extract higher-order features through layer-by-layer abstraction (e.g., the synergistic effect of "temperature-vertical airflow-cloud droplet concentration"). The number of hidden layers can be determined through cross-validation. For example: layer 2 has 128 neurons (extracting basic features, such as the interaction between temperature and humidity); layer 3 has 64 neurons (learning the correlation pattern between the icing index and physical quantities); layer 4 has 32 neurons (compressing redundant information and retaining discriminative features). In other words, the data features can be processed through multiple hidden layers from layer 2 to layer 5 to obtain the processed data features. By processing the data features layer by layer through multiple hidden layers, the deep features of the ice accumulation data can be gradually extracted and optimized, enhancing the model's ability to recognize complex data patterns and making the prediction results more accurate and reliable.

[0076] Step S151c: Predict the processed data features through the output layer to obtain the prediction status and prediction level corresponding to the ice accumulation frequency sample data.

[0077] For example, the implementation of step S151c described above involves the output layer being used to predict the processed data features. This output layer can employ two fully connected layers. Therefore, the predicted state corresponding to the ice accumulation frequency sample data can be obtained through the first fully connected layer, and the predicted level corresponding to the ice accumulation frequency sample data can be obtained through the second fully connected layer. By predicting the processed data features through the output layer, not only can the predicted state of the ice accumulation frequency be output, but it can also be further refined to the prediction level, providing more comprehensive prediction information and meeting the needs of different application scenarios.

[0078] It is understandable that the process described above, which involves feature processing of data through multiple hidden layers and prediction of the processed data features through the output layer, can be represented as follows: , , and ;in, This represents the ice accumulation frequency sample data input from the input layer. This represents the data features input to the first hidden layer. This represents the data features input to the second hidden layer. This represents the data features input to the third hidden layer. This represents the prediction results (including prediction state and prediction level) of the aforementioned neural network model. Indicates the connection weights of the first hidden layer. This indicates the connection weights of the second hidden layer. Indicates the connection weights of the third hidden layer. This represents the bias value of the first hidden layer. This represents the bias value of the second hidden layer. This represents the bias value of the third hidden layer. This represents the activation function of the first hidden layer. This represents the activation function of the second hidden layer. This represents the activation function of the third hidden layer. The activation function mentioned above can be the ReLU function.

[0079] Optionally, during the training of the aircraft icing prediction model, the Adam algorithm can be used to optimize the model's weight parameters. Furthermore, the training can be performed using a mini-batch method. For example, a portion of all data can be selected as a representative of all data, and the loss function value can be calculated for this portion of data to find the set of parameters that minimizes this value.

[0080] As an optional implementation of the above-mentioned aircraft icing prediction model training method, after obtaining the aircraft icing prediction model, it may further include: Step S160: Obtain the icing frequency data to be processed. The icing frequency data to be processed is obtained by detecting the target aircraft through an airborne icing detector.

[0081] For example, in the implementation of step S160 above, the electronic device can acquire temperature, specific humidity, relative humidity, temperature-dew point difference, vertical airflow, and FIP, RAP, NAWAU, RAOB, etc., from the airborne icing detector as icing frequency data to be processed. Alternatively, after the airborne icing detector sends the icing frequency data to be processed to the electronic device, the electronic device can directly receive the icing frequency data sent by the airborne icing detector. The aforementioned icing frequency data to be processed can be acquired in real time using an airborne icing detector of model ICE-2020 at a sampling frequency of 10Hz, collecting icing characteristic signals from the leading edge of the target aircraft wing. The detection frequency band can also be set to the 18-26GHz millimeter wave band, the sampling interval can be set to 0.1 seconds, and the resolution can be set to 0.5dBz. Optionally, after acquiring the icing frequency data to be processed, the electronic device can also preprocess the icing frequency data to be processed, such as denoising (using a wavelet threshold denoising algorithm) and normalization (mapping the data to the [0,1] interval). By directly using real-time detection data obtained from airborne icing detectors as input and combining it with a pre-established icing forecast model for prediction, real-time dynamic monitoring and assessment of aircraft icing status is achieved, significantly improving the timeliness and accuracy of icing warnings and avoiding the lag problem caused by relying on meteorological forecast data in traditional methods.

[0082] Step S170: Predict the icing frequency data to be processed using the aircraft icing prediction model to obtain the icing state and intensity level corresponding to the icing frequency data to be processed.

[0083] The implementation of step S170 above can be exemplified as follows: Assuming the aircraft icing prediction model includes an input layer, an output layer, and multiple hidden layers, the implementation of predicting the icing frequency data to be processed using the aircraft icing prediction model can be exemplified as follows: Feature extraction is performed on the icing frequency data to be processed through the input layer to obtain data features; feature processing is performed on the data features through multiple hidden layers to obtain processed data features; and data prediction is performed on the processed data features through the output layer to obtain the icing state and intensity level corresponding to the target aircraft. In the implementation of the above scheme, by using the icing state and intensity level as the model's prediction output, more refined icing information can be provided for flight decision-making, enabling pilots or autopilot systems to take more targeted countermeasures based on the specific severity of icing, thereby improving flight safety. Furthermore, this technical solution achieves direct integration of airborne detection data and the prediction model, reducing intermediate data processing steps, lowering system complexity, and improving the reliability of prediction results, because the model directly calculates based on actual detection data, avoiding errors that may be introduced during data conversion.

[0084] Please see Figure 3 The illustrated diagram shows a flowchart of the aircraft icing forecasting method provided in an embodiment of this application; this application also provides an aircraft icing forecasting method executed on an electronic device, the implementation of which may include: Step S210: Obtain the icing frequency data to be processed. The icing frequency data to be processed is obtained by detecting the target aircraft using an airborne icing detector.

[0085] For example, in the implementation of step S210 above, the electronic device can acquire the icing frequency data to be processed from the airborne icing detector. Alternatively, after the airborne icing detector sends the icing frequency data to be processed to the electronic device, the electronic device can directly receive the icing frequency data sent by the airborne icing detector. The aforementioned icing frequency data to be processed can be acquired in real time using an airborne icing detector of model ICE-2020 at a sampling frequency of 10Hz, collecting icing characteristic signals from the leading edge of the target aircraft wing. The detection frequency band can also be set to the 18-26GHz millimeter wave band, the sampling interval can be set to 0.1 seconds, and the resolution can be set to 0.5dBz. Optionally, after acquiring the icing frequency data to be processed, the electronic device can also preprocess the icing frequency data to be processed, such as denoising (using a wavelet threshold denoising algorithm) and normalization (mapping the data to the [0,1] interval). By directly acquiring the icing frequency data detected by the airborne icing detector, the current icing situation of the target aircraft can be reflected in real time and accurately, avoiding the lag problem of relying on ground meteorological data or remote sensing data in traditional methods, thus significantly improving the timeliness and accuracy of icing forecasts.

[0086] Step S220: Predict the icing frequency data to be processed using the aircraft icing prediction model to obtain the icing state and intensity level corresponding to the target aircraft.

[0087] As an optional implementation of step S220 above, the aircraft icing prediction model may include: an input layer, an output layer, and multiple hidden layers; the implementation of predicting the icing frequency data to be processed using the aircraft icing prediction model may include: Step S221: Extract features from the ice accumulation frequency data to be processed through the input layer to obtain data features.

[0088] The implementation of step S221 above can be illustrated as follows: The first layer of the aforementioned neural network model is the input layer. The variable parameters of this input layer can be constructed by fusing icing-related atmospheric environmental parameters from multiple icing diagnostic products. The core of this approach lies in integrating the outputs of various icing diagnostic algorithms with physical environmental parameters to construct a comprehensive and physically meaningful input feature space. Specifically, physical quantities such as temperature, specific humidity, relative humidity, temperature-dew point difference, and vertical airflow directly reflect the atmospheric thermodynamic state and are necessary conditions for icing formation. Vertical airflow (rising or descending motion) affects the distribution height and concentration of water droplets in clouds, thus determining the intensity of icing. Feature extraction of the icing frequency data to be processed through the input layer demonstrates the feature extraction capability of the neural network. The model can automatically learn the nonlinear relationship between physical quantities and icing indices such as FIP, RAP, NAWAU, and RAOB, avoiding the subjectivity of human experience. Feature extraction of the icing frequency data to be processed through the input layer effectively captures the original features of the icing data, providing a high-quality data foundation for subsequent processing and thus improving the accuracy of predictions.

[0089] Step S222: Perform feature processing on the data features through multiple hidden layers to obtain the processed data features.

[0090] The implementation of step S222 above can be illustrated as follows: It is understood that the combination of neurons in the hidden layers (e.g., 2-4 layers) needs to be optimized in conjunction with task complexity and data dimensionality. Its design goal is to capture key features in icing forecasting and suppress noise. The number of hidden layers can be selected based on specific circumstances (e.g., determined through cross-validation). If the number of hidden layers is relatively small, the neural network is a shallow network. Shallow networks (e.g., a single hidden layer) are suitable for simple tasks but struggle to model complex nonlinear relationships in the icing process. However, if the number of hidden layers is relatively large, the neural network is a deep network. Such deep networks (e.g., 3-4 layers) can extract higher-order features through layer-by-layer abstraction (e.g., the synergistic effect of "temperature-vertical airflow-cloud droplet concentration"). The number of hidden layers can be determined through cross-validation. For example: layer 2 has 128 neurons (extracting basic features, such as the interaction between temperature and humidity); layer 3 has 64 neurons (learning the correlation pattern between the icing index and physical quantities); layer 4 has 32 neurons (compressing redundant information and retaining discriminative features). In other words, the data features can be processed through multiple hidden layers from the second to the fifth layer to obtain the processed data features.

[0091] Understandably, by processing data features layer by layer through multiple hidden layers, deeper features of icing data can be gradually extracted and optimized, enhancing the model's ability to recognize complex data patterns and making predictions more accurate and reliable. Furthermore, by employing a deep neural network structure with multiple hidden layers, this technique enables multi-level nonlinear feature extraction and processing of icing frequency data, significantly improving the accuracy and reliability of icing prediction. The stacked structure of multiple hidden layers allows the model to learn more complex data feature relationships, which is difficult to achieve with traditional single-layer or shallow models.

[0092] Step S223: Perform data prediction on the processed data features through the output layer to obtain the icing state and intensity level corresponding to the target aircraft.

[0093] For example, the implementation of step S223 described above involves the output layer being used to predict the processed data features. This output layer can employ two fully connected layers. Therefore, the predicted state corresponding to the icing frequency sample data can be obtained through the first fully connected layer, and the predicted level can be obtained through the second fully connected layer. By predicting the processed data features through the output layer, not only can the predicted state of the icing frequency be output, but it can also be further refined to the prediction level, providing more comprehensive prediction information to meet the needs of different application scenarios. By using the icing state and intensity level as the prediction targets of the output layer, this technical solution achieves a comprehensive assessment of the icing situation, simultaneously providing a judgment on the presence or absence of icing and a quantitative indicator of the severity of icing. This dual-output structure provides a more comprehensive reference for flight safety decisions. The feature extraction and processing process is automatically completed through a deep learning model, avoiding the tedious steps of manually designing features and selecting parameters required in traditional methods. This greatly improves the efficiency and automation of icing prediction while reducing errors caused by human factors.

[0094] In implementing the above scheme, by using an aircraft icing prediction model to predict icing frequency data, the complex physical process of icing can be transformed into a computable mathematical model. Through the model's self-learning and optimization capabilities, the icing state and intensity level can be predicted more accurately, providing a more reliable basis for flight safety decisions. Furthermore, by combining real-time data from airborne detectors with intelligent analysis of the prediction model, dynamic monitoring and prediction of aircraft icing states are achieved. This not only enables timely detection of icing risks but also predicts icing development trends, thus providing pilots and ground control personnel with more comprehensive safety warning information.

[0095] Please see Figure 4 The diagram shown is a structural schematic of the aircraft icing forecasting model training device provided in this application embodiment; this application embodiment provides an aircraft icing forecasting model training device 300, including: The sample data acquisition module 310 is used to acquire icing frequency sample data, which is obtained by detecting the aircraft using an airborne icing detector.

[0096] The icing state identification module 320 is used to identify the icing state of frequency data within multiple consecutive time windows from the icing frequency sample data. The icing state includes icing and no icing.

[0097] The state sample determination module 330 is used to determine multiple ice-accumulated samples with ice accumulation and multiple non-ice-accumulated samples with ice accumulation from the ice accumulation frequency sample data.

[0098] The intensity level determination module 340 is used to determine the intensity level of an ice sample based on the rate of frequency decrease for each of a plurality of ice samples.

[0099] The neural network training module 350 is used to train the neural network model using icing frequency sample data as training data and icing state and intensity level as training labels to obtain an aircraft icing forecast model. The aircraft icing forecast model is used to predict the corresponding icing state and intensity level based on the icing frequency data.

[0100] As an optional implementation of the above-mentioned device, the ice accumulation state recognition module includes: The frequency data judgment submodule is used to determine whether the frequency data within each of multiple consecutive duration windows meets preset standard conditions. The preset standard conditions include: the airborne icing detector is inside the cloud layer throughout the consecutive duration window, the detection temperature is negative, the icing frequency data shows a continuous downward trend, the frequency value per second satisfies a non-increasing relationship, and the frequency difference between the initial and final moments of the frequency data within the consecutive duration window is greater than the frequency threshold.

[0101] The icing state determination submodule is used to determine the icing state of the frequency data within a continuous time window as having icing if the frequency data within the continuous time window meets the preset standard conditions, and to determine the icing state of the frequency data within the continuous time window as having no icing if the frequency data within the continuous time window does not meet the preset standard conditions.

[0102] As an optional implementation of the above-mentioned device, the neural network training module includes: The sample data inference submodule is used to infer the icing frequency sample data using a neural network model to obtain the predicted state and predicted level of the icing frequency sample data.

[0103] The loss value calculation submodule is used to calculate the first loss value between the icing state and the predicted state, and the second loss value between the intensity level and the predicted level, using a loss function.

[0104] The model parameter update submodule is used to update the model parameters of the neural network model based on the total loss value determined by the first loss value and the second loss value, until the total loss value reaches a preset threshold or the number of training iterations reaches a preset number, thus obtaining the aircraft icing prediction model.

[0105] As an optional implementation of the above-mentioned device, the neural network model includes: an input layer, an output layer, and multiple hidden layers; the sample data inference submodule includes: The data feature acquisition unit is used to extract features from the ice accumulation frequency sample data through the input layer to obtain data features.

[0106] The data feature processing unit is used to process data features through multiple hidden layers to obtain processed data features.

[0107] The data feature prediction unit is used to predict the processed data features through the output layer to obtain the prediction status and prediction level corresponding to the ice accumulation frequency sample data.

[0108] As an optional implementation of the above-mentioned device, the aircraft icing prediction model training device further includes: The frequency data acquisition module is used to acquire the icing frequency data to be processed. The icing frequency data to be processed is obtained by detecting the target aircraft through an airborne icing detector.

[0109] The frequency data prediction module is used to predict the icing frequency data to be processed using the aircraft icing prediction model, and to obtain the icing state and intensity level corresponding to the icing frequency data to be processed.

[0110] Please see Figure 5 The diagram shown is a structural schematic of the aircraft icing forecasting device provided in this application embodiment; this application embodiment provides an aircraft icing forecasting device 400, including: The frequency data acquisition module 410 is used to acquire the icing frequency data to be processed, which is obtained by detecting the target aircraft through an airborne icing detector.

[0111] The state level prediction module 420 is used to predict the icing frequency data to be processed through the aircraft icing prediction model, and obtain the icing state and intensity level corresponding to the target aircraft.

[0112] As an optional implementation of the above-mentioned device, the aircraft icing prediction model includes: an input layer, an output layer, and multiple hidden layers; the state level prediction module includes: The data input extraction submodule is used to extract features from the ice accumulation frequency data to be processed through the input layer to obtain data features.

[0113] The data feature processing subunit is used to process data features through multiple hidden layers to obtain processed data features.

[0114] The State Level Prediction Submodule is used to predict the icing state and intensity level of the target aircraft by using the processed data features through the output layer.

[0115] It should be understood that this device corresponds to the above-described aircraft icing forecast model training method embodiment and is capable of performing the various steps involved in the above method embodiment. The specific functions of this device can be found in the description above, and detailed descriptions are appropriately omitted here. The device includes at least one software functional module that can be stored in memory or embedded in the device's operating system (OS) in the form of software or firmware.

[0116] Please see Figure 6 The diagram shows a structural schematic of an electronic device provided in an embodiment of this application. An electronic device 500 provided in this application includes a processor 510 and a memory 520. The memory 520 stores machine-readable instructions executable by the processor 510. When the machine-readable instructions are executed by the processor 510, the method described above is performed.

[0117] This application embodiment also provides a computer-readable storage medium 530, on which a computer program is stored. This computer program is executed by a processor 510 to perform the methods described above. The computer-readable storage medium 530 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0118] This application also provides a computer program product, including: a computer program or computer instructions, which are executed by a processor to perform the method described above.

[0119] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For apparatus embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0120] It should be understood that the disclosed apparatus and methods can also be implemented in other ways, as provided in the embodiments of this application. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in the flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the accompanying drawings. For example, two consecutive blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending primarily on the functions involved.

[0121] Furthermore, the functional modules of each embodiment in this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part. In addition, in the description of this specification, the reference to terms such as "one embodiment," "some embodiments," "example," "specific example," "some examples," etc., means that the specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0122] The above description is only an optional implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of this application should be covered within the protection scope of the embodiments of this application.

Claims

1. A training method for an aircraft icing prediction model, characterized in that, include: Acquire icing frequency sample data, which is obtained by detecting the aircraft using an airborne icing detector; The icing state of frequency data within multiple consecutive time windows is identified from the icing frequency sample data, and the icing state includes icing and no icing. From the ice accumulation frequency sample data, a plurality of ice accumulation samples with ice accumulation status and a plurality of non-ice accumulation samples with ice accumulation status are determined. For each of the plurality of ice accumulation samples, the intensity level of the ice accumulation sample is determined based on the frequency decrease rate; Using the icing frequency sample data as training data, and the icing state and intensity level as training labels, a neural network model is trained to obtain an aircraft icing prediction model. The aircraft icing prediction model is used to predict the corresponding icing state and intensity level based on the icing frequency data.

2. The method according to claim 1, characterized in that, The step of identifying the icing state of frequency data within multiple consecutive time windows from the icing frequency sample data includes: For each of the multiple consecutive duration windows, it is determined whether the frequency data within the consecutive duration window meets preset standard conditions. The preset standard conditions include: the airborne icing detector is inside the cloud layer within the consecutive duration window, and the detection temperature is negative; the icing frequency data shows a continuous decreasing trend; the frequency value per second satisfies a non-increasing relationship; and the frequency difference between the initial and final times of the frequency data within the consecutive duration window is greater than a frequency threshold. If so, the icing state of the frequency data within the continuous duration window is determined to be icing-free; otherwise, the icing state of the frequency data within the continuous duration window is determined to be icing-free.

3. The method according to claim 1, characterized in that, The training of the neural network model includes: The neural network model is used to infer the ice accumulation frequency sample data to obtain the predicted state and predicted level of the ice accumulation frequency sample data. The first loss value between the icing state and the predicted state, and the second loss value between the intensity level and the predicted level are calculated using a loss function. Based on the total loss value determined by the first loss value and the second loss value, the model parameters of the neural network model are updated until the total loss value reaches a preset threshold or the number of training iterations reaches a preset number, thus obtaining the aircraft icing prediction model.

4. The method according to claim 3, characterized in that, The neural network model includes: an input layer, an output layer, and multiple hidden layers; the inference using the neural network model on the icing frequency sample data includes: The input layer is used to extract features from the ice accumulation frequency sample data to obtain data features; The data features are processed by the multiple hidden layers to obtain the processed data features. The output layer predicts the processed data features to obtain the prediction status and prediction level corresponding to the ice accumulation frequency sample data.

5. The method according to any one of claims 1-4, characterized in that, After obtaining the aircraft icing prediction model, the following is also included: The icing frequency data to be processed is obtained by detecting the target aircraft using an airborne icing detector. The aircraft icing prediction model is used to predict the icing frequency data to be processed, thereby obtaining the icing state and intensity level corresponding to the icing frequency data.

6. A method for predicting aircraft icing, characterized in that, include: The icing frequency data to be processed is obtained by detecting the target aircraft using an airborne icing detector. The icing frequency data to be processed is predicted by the aircraft icing prediction model to obtain the icing state and intensity level corresponding to the target aircraft.

7. The method according to claim 6, characterized in that, The aircraft icing prediction model includes: an input layer, an output layer, and multiple hidden layers; the prediction of the icing frequency data to be processed using the aircraft icing prediction model includes: The input layer is used to extract features from the ice accumulation frequency data to be processed, and the data features are obtained. The data features are processed by the multiple hidden layers to obtain the processed data features. The output layer performs data prediction on the processed data features to obtain the icing state and intensity level of the target aircraft.

8. A training device for an aircraft icing prediction model, characterized in that, include: The sample data acquisition module is used to acquire icing frequency sample data, which is obtained by detecting the aircraft using an airborne icing detector. An icing state identification module is used to identify the icing state of frequency data within multiple consecutive time windows from the icing frequency sample data, wherein the icing state includes icing and no icing; The state sample determination module is used to determine, from the ice accumulation frequency sample data, multiple ice accumulation samples in which the ice accumulation state is ice accumulation, and multiple non-ice accumulation samples in which the ice accumulation state is no ice accumulation. An intensity level determination module is used to determine the intensity level of each of the plurality of ice accumulation samples based on the frequency decrease rate. The neural network training module is used to train the neural network model using the icing frequency sample data as training data and the icing state and intensity level as training labels to obtain an aircraft icing prediction model. The aircraft icing prediction model is used to predict the corresponding icing state and intensity level based on the icing frequency data.

9. An aircraft icing prediction device, characterized in that, include: The frequency data acquisition module is used to acquire the icing frequency data to be processed, which is obtained by detecting the target aircraft through an airborne icing detector. The state level prediction module is used to predict the icing frequency data to be processed using an aircraft icing prediction model, so as to obtain the icing state and intensity level corresponding to the target aircraft.

10. An electronic device, characterized in that, include: A processor and a memory, the memory storing machine-readable instructions executable by the processor, the machine-readable instructions being executed by the processor to perform the method of any one of claims 1 to 7.

Images