Snow making control method and equipment for airworthiness verification of civil aviation power device and medium

Abstract

The invention relates to the technical field of aircraft testing, in particular to a civil aviation power device airworthiness verification snow making control method and device and a medium, and the method comprises the steps: obtaining multi-mode snow making control data, and constructing a multi-mode snow making control data set; constructing a multi-modal fusion model based on a CNN (Convolutional Neural Network) and a Transform, and training the multi-modal fusion model by utilizing the multi-modal snow making control data set to obtain a trained multi-modal fusion model; deploying the trained multi-modal fusion model, inputting the multi-modal snow making control data obtained in real time into the trained multi-modal fusion model, and adjusting the snow making control system based on the control parameters output by the multi-modal fusion model. According to the method, by fusing the local feature extraction capability of CNN and the global time sequence modeling advantage of Transform, real-time efficient control over snow making parameters can be achieved, the requirements of tests for snow quality and the total water content are effectively met, the problem of nozzle blockage caused by low-temperature icing can be effectively solved, and the method can be adapted to snow making equipment of different models.

Description

Technical Field

[0001] This invention relates to the field of aircraft testing technology, and in particular to a snowmaking control method, apparatus, equipment, and medium for airworthiness verification of civil aviation power plants. Background Technology

[0002] Existing artificial snowmaking control technologies mainly rely on single sensor data (such as temperature and humidity) and preset rules for regulation, which has the following main shortcomings: low data utilization efficiency, resulting in insufficient precision of control strategies; poor dynamic adaptability, making it difficult to cope with complex nonlinear weather changes; the use of manual sampling or simple optical sensors to detect snow quality, making it difficult to achieve high-precision real-time monitoring; lack of collaborative optimization of multiple variables, leading to energy waste; and limited prediction and decision-making capabilities, making it difficult to integrate weather forecasts, equipment status, and historical data to predict the optimal snowmaking time. Summary of the Invention

[0003] To address the technical problems existing in the prior art, this invention provides a snowmaking control method, device, terminal equipment, and computer-readable storage medium for airworthiness verification of civil aviation power units, which can achieve real-time and efficient control of snowmaking parameters and effectively meet the requirements of the test for snow quality and total water content.

[0004] The first objective of this invention is to provide a snowmaking control method for airworthiness verification of civil aviation power plants.

[0005] The second objective of this invention is to provide a device for snowmaking control in civil aviation power plant airworthiness verification.

[0006] A third objective of this invention is to provide a computer device.

[0007] A fourth objective of this invention is to provide a computer-readable storage medium.

[0008] The first objective of this invention can be achieved by adopting the following technical solution: A snowmaking control method for airworthiness verification of civil aircraft power plants, the method comprising: S1. Acquire multimodal snowmaking control data, which includes: meteorological data, snowmaking machine equipment operation data, snow quality monitoring data, acoustic data, and experimental environment data, and construct a multimodal snowmaking control dataset; S2. Construct a multimodal fusion model based on CNN network and Transformer. S3. Use the multimodal snowmaking control dataset to train the multimodal fusion model and obtain the trained multimodal fusion model; S4. Deploy the trained multimodal fusion model, input the real-time acquired multimodal snowmaking control data into the trained multimodal fusion model, and adjust the snowmaking control system based on the control parameters output by the multimodal fusion model.

[0009] Specifically, the meteorological data includes temperature, humidity, wind speed, wind direction, and air pressure, and all meteorological data are time-series data. The equipment operation data of the snowmaking machine includes: water pressure, flow rate, nozzle angle, compressor power, voltage / current, and air intake ratio, which are time-series logs or real-time streaming data of industrial protocols; The snow quality monitoring data includes video images captured by a camera, showing the morphology, distribution density, and snowfall rate of snow crystals. The acoustic data includes: snowmaking machine nozzle jet sound and compressor noise, used for fault diagnosis; The experimental environment data includes: the terrain elevation of the experimental site, GPS coordinates, and historical satellite cloud images.

[0010] Specifically, the multimodal fusion model includes an input layer, a multimodal fusion layer, and a network output layer, wherein: The input layer is used to preprocess the multimodal data. The preprocessed data of different modalities are input into different dedicated network branches for modality-independent feature extraction, so as to obtain the independent features of each different modality data. The multimodal fusion layer is used to perform cross-modal feature fusion on the independent features of the extracted multimodal data through a hierarchical attention mechanism. It uses the temporal features of meteorological data as the query vector, and visual and audio features as key-value pairs. It dynamically calculates the contribution weight of each modal data feature and calculates the cross-modal fusion feature of the multimodal data based on the contribution weight of each modal data feature. The network output layer is used to output control parameters based on the cross-modal fusion features of multimodal data.

[0011] Specifically, the preprocessing of multimodal data includes: standardizing different modal data to align them in time and space; for meteorological data, acquiring data at a preset frequency via a bus interface, extracting time-series segments of preset duration using a sliding window method to obtain meteorological time-series data, and normalizing the meteorological time-series data; for snow quality monitoring data, decoding the input video stream into RGB format via hardware acceleration, performing dynamic white balance correction and region of interest adjustment, and cropping and segmenting it into image blocks of preset size; for acoustic data, converting audio / vibration signals into time-spectrum maps via short-time Fourier transform, and then compressing the dimensions using a Mel filter bank to highlight the vibration characteristics of the equipment.

[0012] Specifically, the input layer includes: a meteorological data branch, an equipment operation data branch, a visual data branch, an acoustic data branch, and an experimental environment data branch, wherein: The meteorological data branch uses a CNN network to extract the temporal features of the meteorological data. The equipment operation data branch adopts a hybrid structure of CNN network and self-attention mechanism. The CNN network extracts short-term local features of equipment operation data, and the self-attention mechanism captures long-term dependencies of equipment operation data. The visual data branch adopts a dual-path structure. It extracts local features of snow quality monitoring data through a CNN network and processes global features of snow quality monitoring data through a ViT encoder. The extracted local features and global features are then fused together. The acoustic data branch uses a 5-layer 1D-CNN structure with max pooling layers to convolve the original audio waveform and extract the spectral features of the acoustic data in conjunction with the max pooling layers. The experimental environment data branch is used to process GPS coordinates, elevation, and historical cloud map data, and to perform two-dimensional convolution dimensionality reduction on the historical cloud map data.

[0013] Specifically, the control parameters include snowmaking machine control parameter signals, fault status signals, snow quality assessment, and optimal snowmaking time. The output adopts a fully connected network, and the snowmaking machine control parameters include water pressure, flow rate, nozzle angle, and compressor power.

[0014] Specifically, training the multimodal fusion model using a multimodal snowmaking control dataset includes: A phased optimization strategy is adopted. First, each modal branch of the input layer of the multimodal fusion model is pre-trained independently. After the convolutional neural network of the visual data branch is pre-trained on a public dataset, the input layer modal branches are trained and fine-tuned using a multimodal snowmaking control dataset. For sequence data, a one-dimensional CNN network is used to predict sensor data, and the Transformer part is trained jointly. When training the multimodal fusion model as a whole, a dynamic weighted loss function is adopted, with Focal Loss for classification tasks and Huber Loss for regression tasks.

[0015] The second objective of this invention can be achieved by adopting the following technical solution: A device for snowmaking control during airworthiness verification of civil aircraft power plants, the device comprising: The multimodal snowmaking control data unit is used to acquire multimodal snowmaking control data, which includes: meteorological data, snowmaking machine equipment operation data, snow quality monitoring data, acoustic data, and experimental environment data, to construct a multimodal snowmaking control dataset. The multimodal fusion model unit is used to build a multimodal fusion model based on CNN network and Transformer. The multimodal fusion model is trained using a multimodal snowmaking control dataset to obtain a trained multimodal fusion model. The output unit is used to deploy the trained multimodal fusion model, input the real-time acquired multimodal snowmaking control data into the trained multimodal fusion model, and adjust the snowmaking control system based on the control parameters output by the multimodal fusion model.

[0016] The third objective of this invention can be achieved by adopting the following technical solution: A computing device includes a processor and a memory for storing processor-executable programs, wherein when the processor executes the programs stored in the memory, it implements the above-described snowmaking control method for airworthiness verification of civil aviation power units.

[0017] The fourth objective of this invention can be achieved by adopting the following technical solution: A computer-readable storage medium storing a program that, when executed by a processor, implements the above-described snowmaking control method for airworthiness verification of civil aviation power units.

[0018] The present invention has the following advantages over the prior art: This invention provides a snowmaking control method, equipment, and medium for airworthiness verification of civil aviation power units. By constructing a multimodal fusion model based on CNN networks and Transformers, it achieves a deep fusion of the local feature extraction capabilities of CNNs and the global temporal modeling advantages of the self-attention mechanism of Transformers, significantly improving the control accuracy and adaptability of the snowmaking system. First, multimodal data collaborative analysis enhances environmental perception capabilities: the visual branch accurately identifies snowflake morphology and nozzle icing status, meteorological sensor data dynamically predicts optimal snowmaking parameters, and audio signals monitor equipment health status in real time, enabling the system to operate stably in extreme environments ranging from -30°C to 60°C. Second, the hierarchical attention mechanism realizes intelligent weight allocation between modalities. For example, it automatically enhances the decision weight of visual features during blizzards, while prioritizing audio features when mechanical vibrations are abnormal, ensuring the physical rationality of the control strategy. It enables real-time and efficient control of snowmaking parameters, effectively meeting the experimental requirements for snow quality and total water content. It effectively solves the nozzle clogging problem caused by low-temperature icing and can be adapted to different models of snowmaking equipment through parameter migration, possessing significant engineering application value. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the snowmaking control method for airworthiness verification of civil aviation power units according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the structure of the CNN+Transformer multimodal fusion model in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the encoder and decoder in Embodiment 1 of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments, and the implementation of the present invention is not limited thereto. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Airworthiness verification of civil aviation power plants (such as aircraft engines and auxiliary power units, APUs) is a core step in ensuring their safety and reliability. It requires verifying their performance in complex scenarios such as low temperatures, icing, and snow accumulation by simulating extreme environmental conditions. Among these, snowmaking environment simulation is a necessary means to verify key characteristics of the power plant in icy and snowy weather, such as the anti-icing capability of the air intake system, the compressor's resistance to foreign object damage, and combustion stability.

[0023] Example 1 This invention provides a snowmaking control method for airworthiness verification of civil aviation power units, used in snow verification tests of civil aviation power units. The system features strong control capabilities and high computational accuracy, and calculates the optimal test nodes on-site. Through a CNN+Transformer cross-modal attention mechanism, dynamic collaborative decision-making based on meteorological, visual, and acoustic data is achieved, improving control accuracy by 40% compared to traditional PID control. Based on TensorRT's INT8 quantization technology, a 35ms real-time response is achieved on Xavier, with power consumption <15W, and stable operation in extreme environments down to -30℃.

[0024] like Figure 1 As shown, this embodiment provides a snowmaking control method for airworthiness verification of civil aircraft power plants, including the following steps: S1. Acquire multimodal snowmaking control data and construct a multimodal snowmaking control dataset. Multimodal snowmaking control data includes meteorological data, snowmaking machine operation data, snow quality monitoring data, acoustic data, and experimental environment data.

[0025] Multimodal snowmaking control datasets play a crucial role in the intelligent control of snowmaking machines. This primarily involves fusing multi-source data (meteorology, equipment status, snow quality, etc.) to optimize key indicators such as water-to-air ratio and spray parameters, thereby improving snowmaking efficiency and snow quality stability. Through collaborative analysis of multimodal data, complex meteorological conditions (such as sudden humidity changes and critical temperatures) can be dynamically addressed, reducing human intervention. By combining time-series data on equipment vibration and current, component failures can be predicted, reducing downtime risks. Finally, by analyzing the relationship between energy consumption and snow quality, optimal operating strategies can be generated to reduce energy waste.

[0026] Specifically, current meteorological data, including temperature, humidity, wind speed, wind direction, and air pressure, is all time-series data, typically stored in CSV or database tables. This data is acquired via meteorological sensors using a bus protocol to ensure time synchronization accuracy at the millisecond level.

[0027] The snowmaking machine's equipment operation data, including water pressure, flow rate, nozzle angle, compressor power, voltage / current, and air intake ratio, are collected in real-time as time-series logs or industrial protocol real-time streaming data through sensors installed in the equipment's fluid piping system, such as those for pressure, flow rate, angle, and current / voltage.

[0028] Specifically, meteorological data and equipment operation data are collected at a frequency of 20Hz through the bus interface, and a 1.5-second time segment is extracted using a sliding window method.

[0029] Snow quality monitoring data includes video images of snow crystal morphology, distribution density, and snowfall rate captured by high-definition high-speed cameras. Based on these video images, the water content, particle size, and snowfall rate in the snow are analyzed.

[0030] Acoustic data, including the jet sound of snowmaking machine nozzles and compressor noise, is used for fault diagnosis; the acoustic data structure uses waveform audio or vibration data, which is acquired through acoustic array equipment and vibration sensors.

[0031] Experimental environment data includes the experimental site's topographic elevation (without convolution), GPS coordinates (without convolution), and historical satellite cloud images. Before each experiment, this data is acquired via the network and stored in the system in a pre-processed form. The experimental environment data is already processed and does not require further preprocessing.

[0032] Based on collected meteorological data, snowmaking machine equipment operation data, snow quality monitoring data, acoustic data, and experimental environment data, a multimodal snowmaking control dataset is constructed. Combining expert knowledge and physical models, the equipment status under different operating conditions is labeled, such as "icing risk level" or "equipment health." The final dataset should include various extreme operating conditions.

[0033] S2. A multimodal fusion model is constructed based on CNN network and Transformer. The multimodal fusion model is used to preprocess multimodal data, extract independent features of multimodal data, and perform cross-modal feature fusion based on the extracted independent features of multimodal data through a hierarchical attention mechanism, and map the fused features into control commands.

[0034] The input layer is used to preprocess the multimodal data. The preprocessed data of different modalities are input into different dedicated network branches for modality-independent feature extraction, so as to obtain the independent features of each different modality data.

[0035] Specifically, the preprocessing of multimodal data includes: standardizing the different modal data in the input data to align them in time and space. In this embodiment, for meteorological data, it is acquired at a frequency of 20Hz via a bus interface, a 1.5-second time segment is extracted using a sliding window method, and then normalized. For snow quality monitoring data, the video stream input from the camera is hardware-accelerated and decoded into RGB format, then subjected to dynamic white balance correction and region of interest, and cropped into 16×16 image blocks. For acoustic data, the audio / vibration signals are converted into a time-spectrum graph through short-time Fourier transform, and then compressed using a Mel filter bank to highlight the vibration characteristics of the equipment.

[0036] Visual data is processed by a CNN network to extract multi-scale features. The low-level convolutional layers capture detailed textures such as nozzle icing, while the high-level features identify global structures such as snowflake shapes.

[0037] The time-series data is first pre-encoded using a CNN network, and then the dynamic correlation between sensor data is analyzed through a causal self-attention mechanism. The network audio spectrogram uses a lightweight CNN to extract features, focusing on frequency band features related to mechanical vibration.

[0038] Snow quality monitoring data is processed using a CNN network to extract multi-scale features. Low-level convolutional layers capture detailed textures such as nozzle icing, while high-level features identify global structures such as snowflake shapes. Based on an improved ResNet34, standard convolutions are replaced with depthwise separable convolutions to reduce computation, and a multi-scale feature fusion module is introduced. The temporal processing branch uses a CNN to extract local temporal features, followed by a Transformer to capture long-range dependencies. The self-attention layer uses a causal mask to ensure temporal causality.

[0039] The audio spectrograms in the acoustic data are extracted using a lightweight CNN network, with a focus on frequency band features related to mechanical vibrations.

[0040] Meteorological data, equipment operation data, and geographical and environmental data, after preprocessing, require feature extraction using a CNN network. After independent feature extraction, the system employs a hierarchical attention mechanism to achieve cross-modal information interaction.

[0041] like Figure 2 As shown, a multimodal fusion model is constructed based on a CNN network and a Transformer. The multimodal fusion control model includes an input layer, a multimodal fusion layer, and an output layer. Specifically, the input layer includes: a meteorological data branch, an equipment operation data branch, a visual data branch, an acoustic data branch, and a vibration data branch, wherein: The meteorological data branch uses a CNN network to extract the temporal features of meteorological data. Meteorological data is typically multivariate time series data, such as data containing multiple dimensions including temperature, humidity, air pressure, precipitation, wind speed, and wind direction; it is highly temporally correlated, meaning the current meteorological conditions are closely related to several past time steps.

[0042] like Figure 3The diagram shows the structure of the encoder and decoder, with the encoder (left) and decoder (right). The encoder includes input embedding, positional encoding, and a multi-head self-attention mechanism. Since the Transformer model lacks time-series information like RNNs, positional encoding is needed to provide the positional information of words within a sentence. The positional encoding and input embedding are added together and used as the encoder's input. The multi-head self-attention mechanism specifically includes: Self-attention: Calculating the attention score between each word in the input sequence and other words. Through these attention scores, the model can capture the dependencies between words. Multi-head: Capturing different attention patterns through multiple attention heads. Each head independently calculates attention, and their outputs are concatenated. Add & Normalization: The output of each multi-head self-attention and feedforward neural network is added to the input, and then layer normalization is performed. Feedforward Neural Network: A fully connected layer containing two linear transformations and an activation function, which processes the input at each location independently.

[0043] The decoder specifically includes: Output Embedding, which embeds the words of the target sequence into a high-dimensional vector space; Positional Encoding, the same as in the encoder, which adds the positional encoding and output embedding as the input to the decoder; Masked Multi-Head Self-Attention, which masks the self-attention mechanism in the decoder to ensure that future words are not seen when predicting the next word. This is achieved through a masking matrix; Multi-Head Attention, with an additional multi-head attention layer at each layer of the decoder, which performs attention calculations on the encoder's output. This allows the decoder to refer to information from the input sequence when generating words; Add & Normalization, the same as in the encoder; and Feed Forward, the same as in the encoder. The final output includes: a Linear Layer, which maps the decoder's output to a vector the size of the vocabulary; and a Softmax layer, which transforms the output of the Linear Layer into a probability distribution.

[0044] The device operation data branch employs a hybrid structure of CNN network and self-attention mechanism. The CNN network extracts short-term local features from the device operation data, while the self-attention mechanism captures long-term dependencies. The CNN convolutional kernels slide along the time dimension (or spatiotemporal dimension), effectively extracting short-term local features. The self-attention mechanism captures long-term dependencies; attention mechanisms (such as Self-Attention in Transformers) allow the representation at each time step to interact with all other time steps, thereby capturing dependencies between distant time steps in the device operation data.

[0045] The visual data branch employs a dual-path structure. It extracts local features from snow quality monitoring data using a CNN (Convolutional Neural Network) and processes global features using a ViT (Vision Transformer) encoder. The local features from the CNN and the global features from the ViT encoder are then fused. CNN excels at capturing details such as local textures, edges, and shapes in images, exhibiting strong spatial inductive bias (prior knowledge of local regions). The output typically consists of local feature maps at each spatial location, providing rich, fine-grained information. ViT models the relationship between any two regions in an image through a self-attention mechanism, thereby obtaining global contextual information and being more sensitive to long-distance dependencies, overall structure, and semantic relationships. The input typically involves segmenting the image into multiple patches, which are then processed through linear projection and a Transformer Encoder. The output is a global context-dependent feature representation for each patch or token. The local features from the CNN and the global features from the ViT are fused along appropriate dimensions to combine the advantages of both.

[0046] The acoustic data branch uses a 5-layer 1D-CNN structure with max pooling layers to convolve the original audio waveform (1D time-series) and extract the spectral features of the acoustic data with max pooling layers.

[0047] The experimental environment data branch processes the geographical location information of the experimental environment, including GPS coordinates, elevation, and historical cloud image data. Convolutional dimensionality reduction is not required for the coordinate and elevation data, but two-dimensional convolutional dimensionality reduction is necessary for the historical cloud image data.

[0048] The multimodal fusion layer is used to perform cross-modal feature fusion on the independent features of the extracted multimodal data through a hierarchical attention mechanism. The temporal features of meteorological data are used as query vectors, and visual and audio features are used as key-value pairs. The contribution weight of each modal data feature is dynamically calculated, and cross-modal fusion feature information is calculated based on the contribution weight of each modal data feature.

[0049] Specifically, after extracting independent features from the multimodal data, the multimodal fusion layer employs a hierarchical attention mechanism to achieve cross-modal information interaction. The first stage involves feature enhancement within each modality: the visual branch aggregates global information through global labeling, while the temporal branch uses self-attention to highlight key time points. The second stage establishes intermodal relationships through a cross-attention mechanism, using temporal features as query vectors and visual and audio features as key-value pairs. The contribution weights of each modality are dynamically calculated, and residual connections are introduced during the fusion process to preserve the unique features of the original modalities.

[0050] In this implementation, meteorological data is used as the query vector, and other modalities are used as the key and value. The key and value represent a mapping relationship, that is, accessing the corresponding value through a unique key. Attention weights are calculated, while feature splicing channels are retained as a supplement. The dimensionality of the fused features is reduced to 512 dimensions through a 3-layer MLP (Multilayer Perceptron).

[0051] Meteorological data (Query) is directly used as the query vector, denoted as Q∈RT×d (T is the time step or number of samples of meteorological data); each other modality m (m=1,2,...,M) is projected through a linear layer as Key Km∈RNm×d and ValueVm∈RNm×d (Nm is the number of features of this modality, such as the number of pixel blocks in an image or the number of words in a text).

[0052] Cross-modal attention weight calculation, for example: for each position i of the meteorological data (corresponding to the query vector qi∈ Rd), calculate its similarity score with the key vector kjm∈ Rd (j=1,...,Nm) of all modes m, and obtain the attention weight through Softmax normalization.

[0053] The weighted fusion of Value values ​​is achieved by summing the Value vector and Key vector of the corresponding mode m using attention weights, thus obtaining the fused feature information of meteorological location i from mode m. This process is repeated for all M modes to finally obtain the cross-modal fusion features for each meteorological location i. Cross-modal attention mechanisms dynamically calculate the similarity weights of "Query-Key-Value" pairs, adaptively focusing on the most relevant parts to other modalities, thus explicitly modeling fine-grained relationships between modalities. However, attention mechanisms may overlook low-level features of the modality itself (such as absolute numerical thresholds in meteorological data or basic textures in images). Therefore, it is necessary to retain the feature stitching channels as a supplement to ensure the integrity of the original information. Finally, dimensionality reduction of the stitched fusion result through a 3-layer MLP can reduce the computational burden of subsequent tasks while preserving key information.

[0054] The network output layer is used to output control parameters based on the cross-modal fusion features of multimodal data.

[0055] Specifically, the control parameters include snowmaking machine control parameter signals, fault status signals, snow quality assessment, and optimal snowmaking time. The output uses a fully connected network. The snowmaking machine control parameters include four continuous output values: water pressure, flow rate, nozzle angle, and compressor power. The snow quality assessment uses a 3-layer MLP+Softmax approach, outputting four snow quality classifications. Fault warnings use a binary classifier to judge based on comprehensive equipment data and acoustic / vibration characteristics.

[0056] Specifically, the network output layer maps the fused features to snowmaking machine control parameters. Physical constraints are added to the network output layer to ensure that parameters such as water pressure and nozzle angle are within reasonable ranges. The control signal is smoothed using a Kalman filter to eliminate fluctuations caused by sensor noise. The system monitors the inference confidence level in real time, automatically switching to a backup PID controller when it falls below a threshold and triggering an alarm signal via the GPIO interface. The entire processing flow achieves a real-time response within 35ms on the edge computing platform, meeting the control accuracy requirements under harsh environments.

[0057] This architecture achieves efficient processing of multimodal data through modular design. CNN excels at extracting local spatial features, while Transformer captures long-range dependencies, with the two complementing each other. A hierarchical attention mechanism dynamically adjusts the weights of each modality, enabling the system to adapt to complex environmental changes and maintain stable control performance even under extreme conditions.

[0058] S3. Use the multimodal snowmaking control dataset to train the multimodal fusion model and obtain the trained multimodal fusion model.

[0059] Specifically, in training the multimodal fusion model using a multimodal snowmaking control dataset, a phased optimization strategy was adopted. First, each modal branch of the input layer was pre-trained independently. The convolutional neural network for the visual data branch was pre-trained on a public dataset, and then fine-tuned using the multimodal snowmaking control dataset to train each modal branch of the input layer. For sequence data, a one-dimensional CNN network (one-dimensional convolutional neural network) was used to predict sensor data, and the Transformer part was jointly trained. The one-dimensional CNN network is used to extract local temporal features from the original sensor sequence data and is suitable for processing one-dimensional sequences such as vibration, sound, and sensor readings. Second, the multimodal fusion model was trained as a whole, using a dynamic weighted loss function. For the classification task, the Focal Loss scheme was used to address class imbalance, and for the regression task, the Huber Loss scheme was used to enhance robustness. The optimizer was Adam, with an initial learning rate of 3e-4, and a cosine annealing scheduler was used.

[0060] Specifically, the dynamic weighted loss function is expressed as: L total = w1⋅L1+ w2⋅L 2； Where: L1 and L2 are the losses for Task 1 (such as classification) and Task 2 (such as regression), respectively (e.g., Focal Loss and Huber Loss); w1 and w2 are weights that are manually set or dynamically calculated.

[0061] The Focal Loss function is expressed as follows: ; [ is the Focal Loss function, the model's predicted probability of the target class, is the class balancing weight (static or dynamic), is the static focusing parameter or adaptive parameter, in the early stage of training (t≈0) it is close to the standard cross-entropy, and in the later stage (t≈T) it strengthens the learning of hard samples.

[0062] The Huber Loss function is expressed as: ; The Huber Loss function, with a threshold (typically 1.5 times the standard deviation of the data), controls the switching point between loss smoothing and linearity, or is adaptively set as follows: .

[0063] The formulas for calculating weights w1 and w2 are as follows: ; Where are the feature vectors of CNN and Transformer, respectively. This is the Sigmoid function.

[0064] Specifically, during training, multimodal data augmentation techniques can be used to apply random lighting variations and occlusions to video data, and Gaussian noise injection and random sampling can be performed on time-series data. For the audio spectrograms of acoustic vibration data, random band masking is required. After each training round, the model is evaluated on a validation set, employing an early stopping mechanism to prevent overfitting. The final model is then evaluated against a test set collected from field comparative experiments, including accuracy and real-time performance metrics, to obtain a well-trained modal fusion model.

[0065] S4. Deploy the trained multimodal fusion model, input the real-time acquired multimodal snowmaking control data into the trained multimodal fusion model, and adjust the snowmaking control system based on the output control parameters. S41. Deploy a pre-trained multimodal fusion model based on the NVIDIA Xavier edge computing platform.

[0066] In this embodiment, based on the NVIDIA Xavier edge computing platform, real-time decision-making for the intelligent snowmaking control system is achieved through multimodal fusion and lightweight deployment technologies. First, the PyTorch-trained CNN+Transformer hybrid model undergoes INT8 quantization and TensorRT optimization. TensorRT is used to optimize the trained model: after converting the PyTorch model to ONNX format, graph optimization and layer fusion are performed, and INT8 quantization is applied to the convolutional and LSTM layers. This generates a high-efficiency inference engine, reducing inference latency from 120ms to 35ms while keeping power consumption below 15W. On the NVIDIA Xavier edge computing platform, a heterogeneous computing strategy is adopted: the vision branch runs on the GPU, and temporal processing is deployed to the DLA accelerator, achieving zero-copy data transmission through the NvMedia API. For memory management, pre-allocate GPU memory pools for inputs of different resolutions, supporting dynamic batch processing. The real-time inference service adopts a multi-threaded architecture: the main thread is responsible for task scheduling, and worker threads handle specific inference tasks, avoiding I / O blocking through a double-buffering mechanism. The system integrates a fault detection module. When the inference confidence level falls below a threshold or the hardware status is abnormal, it automatically switches to the backup PID controller and triggers an alarm signal via GPIO. A 72-hour stress test is required after deployment to ensure stable operation in an ambient temperature range of -30°C to 60°C.

[0067] In actual operation, the multimodal fusion model simultaneously processes four types of input data: time-series data from meteorological sensors, images, equipment audio signals, and equipment parameters. It dynamically fuses features through a cross-modal attention mechanism, ultimately outputting three sets of key parameters: 1) snowmaking machine control parameters; 2) equipment health status; and 3) optimal snowmaking time. All outputs drive the actuators in real time via CAN2.0B (serial communication protocol) and PWM (Pulse Width Modulation) interfaces, while simultaneously recording JSON logs containing timestamps, inference time, and sensor checksums. The system remains stable even at -30℃, preventing icing through heating element start / stop and nozzle vibration commands (5Hz micro-vibration), meeting the requirements for all-weather outdoor operation. Control accuracy is improved by 40% compared to traditional PID control, while energy consumption is reduced by 50%.

[0068] S42. Input the real-time acquired multimodal snowmaking control data into the trained multimodal fusion model, and use the output control parameters to adjust the snowmaking control system.

[0069] The water pressure control signal is smoothed using a Kalman filter to eliminate fluctuations caused by high-frequency noise. The nozzle angle parameter is converted into a PWM duty cycle through piecewise linear interpolation to drive the stepper motor. All output parameters are transmitted to the PLC via the Modbus TCP protocol, with the sampling rate strictly synchronized with the control system clock. The system provides a two-level API interface: the lower-level API directly exposes the raw inference results for use by advanced control algorithms; the upper-level API encapsulates safety logic, such as automatically increasing water pressure and triggering a de-icing process when an ice-causing risk is detected. The operation and maintenance interface displays a heatmap of the importance of each modal characteristic, assisting engineers in analysis and decision-making. Historical data storage uses a circular buffer design, retaining high-frequency sampling data (100Hz) from the most recent 24 hours for post-fault analysis and model iteration optimization.

[0070] This embodiment provides a snowmaking control method for airworthiness verification of civil aviation power units. By constructing a multimodal fusion model based on CNN networks and Transformers, it combines the local feature extraction capabilities of CNNs with the global temporal modeling advantages of the self-attention mechanism of Transformers, significantly improving the control accuracy and adaptability of the snowmaking system. First, the multimodal data collaborative analysis of the multimodal fusion model enhances environmental perception: the visual branch accurately identifies snowflake morphology and nozzle icing status, meteorological sensor data dynamically predicts optimal snowmaking parameters, and audio signals monitor equipment health status in real time, enabling the system to operate stably in extreme environments ranging from -30°C to 60°C. Second, the hierarchical attention mechanism achieves intelligent weight allocation between modalities. For example, it automatically enhances the decision weight of visual features during blizzards, while prioritizing audio features when mechanical vibrations are abnormal, ensuring the physical rationality of the control strategy. Finally, edge computing optimization keeps the model inference latency within 35ms and the water pressure regulation error below ±0.5MPa, achieving energy savings of over 15% compared to traditional PID control. This technology has been successfully applied to high-altitude ski resorts, effectively solving the problem of nozzle clogging caused by low-temperature icing. It can also be adapted to different models of snowmaking equipment through parameter migration, and has significant engineering promotion value.

[0071] Example 2 Based on the above-mentioned snowmaking control method for airworthiness verification of civil aviation power plants, this embodiment provides a device for snowmaking control during airworthiness verification of civil aviation power plants, the device specifically including: The multimodal snowmaking control data unit is used to acquire multimodal snowmaking control data, which includes: meteorological data, snowmaking machine equipment operation data, snow quality monitoring data, acoustic data, and experimental environment data, to construct a multimodal snowmaking control dataset. The multimodal fusion model unit is used to build a multimodal fusion model based on CNN network and Transformer. The multimodal fusion model is trained using a multimodal snowmaking control dataset to obtain a trained multimodal fusion model. The output unit is used to deploy the trained multimodal fusion model, input the real-time acquired multimodal snowmaking control data into the trained multimodal fusion model, and adjust the snowmaking control system based on the control parameters output by the multimodal fusion model.

[0072] Example 3 This embodiment provides a computer device, which may be a server, computer, etc., including a processor, memory, input device, display, and network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. When the processor executes the computer programs stored in the memory, it implements the snowmaking control method for airworthiness verification of a civil aviation power plant as described in Embodiment 1 above, as follows: S1. Acquire multimodal snowmaking control data, which includes: meteorological data, snowmaking machine equipment operation data, snow quality monitoring data, acoustic data, and experimental environment data, and construct a multimodal snowmaking control dataset; S2. Construct a multimodal fusion model based on CNN network and Transformer; S3. Use the multimodal snowmaking control dataset to train the multimodal fusion model and obtain the trained multimodal fusion model; S4. Deploy the trained multimodal fusion model, input the real-time acquired multimodal snowmaking control data into the trained multimodal fusion model, and adjust the snowmaking control system based on the control parameters output by the multimodal fusion model.

[0073] Specifically, the meteorological data includes temperature, humidity, wind speed, wind direction, and air pressure, and all meteorological data are time-series data. The equipment operation data of the snowmaking machine includes: water pressure, flow rate, nozzle angle, compressor power, voltage / current, and air intake ratio, which are time-series logs or real-time streaming data of industrial protocols; The snow quality monitoring data includes video images captured by a camera, showing the morphology, distribution density, and snowfall rate of snow crystals. The acoustic data includes: snowmaking machine nozzle jet sound and compressor noise, used for fault diagnosis; The experimental environment data includes: the terrain elevation of the experimental site, GPS coordinates, and historical satellite cloud images.

[0074] Specifically, the multimodal fusion model includes an input layer, a multimodal fusion layer, and a network output layer, wherein: The input layer is used to preprocess the multimodal data. The preprocessed data of different modalities are input into different dedicated network branches for modality-independent feature extraction, so as to obtain the independent features of each different modality data. The multimodal fusion layer is used to perform cross-modal feature fusion on the independent features of the extracted multimodal data through a hierarchical attention mechanism. It uses the temporal features of meteorological data as the query vector, and visual and audio features as key-value pairs. It dynamically calculates the contribution weight of each modal data feature and calculates the cross-modal fusion feature of the multimodal data based on the contribution weight of each modal data feature. The network output layer is used to output control parameters based on the cross-modal fusion features of multimodal data.

[0075] Example 4 This embodiment provides a storage medium, which is a computer-readable storage medium that stores a computer program. When the program is executed by a processor, the processor executes the computer program stored in the memory to implement the snowmaking control method for airworthiness verification of civil aviation power units described in Embodiment 1 above.

[0076] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A snowmaking control method for airworthiness verification of civil aviation power plants, characterized in that, Includes the following steps: Acquire multimodal snowmaking control data, which includes meteorological data, snowmaking machine equipment operation data, snow quality monitoring data, acoustic data, and experimental environment data, and construct a multimodal snowmaking control dataset. A multimodal fusion model is built based on CNN networks and Transformers; The multimodal fusion model was trained using a multimodal snowmaking control dataset to obtain a well-trained multimodal fusion model. The trained multimodal fusion model is deployed, and the real-time acquired multimodal snowmaking control data is input into the trained multimodal fusion model. The snowmaking control system is adjusted based on the control parameters output by the multimodal fusion model.

2. The snowmaking control method for airworthiness verification of civil aviation power units according to claim 1, characterized in that, The meteorological data includes temperature, humidity, wind speed, wind direction, and air pressure, and all meteorological data are time-series data. The equipment operation data of the snowmaking machine includes: water pressure, flow rate, nozzle angle, compressor power, voltage / current, and air intake ratio, which are time-series logs or real-time streaming data of industrial protocols; The snow quality monitoring data includes video images captured by a camera, showing the morphology, distribution density, and snowfall rate of snow crystals. The acoustic data includes: snowmaking machine nozzle jet sound and compressor noise, used for fault diagnosis; The experimental environment data includes: the terrain elevation of the experimental site, GPS coordinates, and historical satellite cloud images.

3. The snowmaking control method for airworthiness verification of civil aviation power units according to claim 1, characterized in that, The multimodal fusion model includes an input layer, a multimodal fusion layer, and a network output layer, wherein: The input layer is used to preprocess the multimodal data. The preprocessed data of different modalities are input into different dedicated network branches for modality-independent feature extraction, so as to obtain the independent features of each different modality data. The multimodal fusion layer is used to perform cross-modal feature fusion on the independent features of the extracted multimodal data through a hierarchical attention mechanism. It uses the temporal features of meteorological data as the query vector, and visual and audio features as key-value pairs. It dynamically calculates the contribution weight of each modal data feature and calculates the cross-modal fusion feature of the multimodal data based on the contribution weight of each modal data feature. The network output layer is used to output control parameters based on the cross-modal fusion features of multimodal data.

4. The snowmaking control method for airworthiness verification of civil aviation power units according to claim 3, characterized in that, The preprocessing of multimodal data includes: standardizing different modal data to align them in time and space; for meteorological data, acquiring data at a preset frequency via a bus interface, extracting time-series segments of preset duration using a sliding window method to obtain meteorological time-series data, and normalizing the meteorological time-series data; for snow quality monitoring data, decoding the input video stream into RGB format via hardware acceleration, performing dynamic white balance correction and region of interest adjustment, and cropping and segmenting it into image blocks of preset sizes; for acoustic data, converting audio / vibration signals into time-spectrum maps via short-time Fourier transform, and then compressing the dimensions using a Mel filter bank to highlight the vibration characteristics of the equipment.

5. The snowmaking control method for airworthiness verification of civil aviation power units according to claim 3, characterized in that, The input layer includes: a meteorological data branch, an equipment operation data branch, a visual data branch, an acoustic data branch, and an experimental environment data branch, wherein: The meteorological data branch uses a CNN network to extract the temporal features of the meteorological data. The equipment operation data branch adopts a hybrid structure of CNN network and self-attention mechanism. The CNN network extracts short-term local features of equipment operation data, and the self-attention mechanism captures long-term dependencies of equipment operation data. The visual data branch adopts a dual-path structure. It extracts local features of snow quality monitoring data through a CNN network and processes global features of snow quality monitoring data through a ViT encoder. The extracted local features and global features are then fused together. The acoustic data branch uses a 5-layer 1D-CNN structure with max pooling layers to convolve the original audio waveform and extract the spectral features of the acoustic data in conjunction with the max pooling layers. The experimental environment data branch is used to process GPS coordinates, terrain elevation, and historical cloud map data, and to perform two-dimensional convolution dimensionality reduction on the historical cloud map data.

6. The snowmaking control method for airworthiness verification of civil aviation power units according to claim 3, characterized in that, The control parameters include snowmaking machine control parameter signals, fault status signals, snow quality assessment, and optimal snowmaking time. The output adopts a fully connected network. The snowmaking machine control parameters include water pressure, flow rate, nozzle angle, and compressor power.

7. The snowmaking control method for airworthiness verification of civil aviation power units according to claim 3, characterized in that, The training of the multimodal fusion model using a multimodal snowmaking control dataset includes: A phased optimization strategy is adopted. First, each modal branch of the input layer of the multimodal fusion model is pre-trained independently. After the convolutional neural network of the visual data branch is pre-trained on a public dataset, the input layer modal branches are trained and fine-tuned using a multimodal snowmaking control dataset. For sequence data, a one-dimensional CNN network is used to predict sensor data, and the Transformer part is trained jointly. When training the multimodal fusion model as a whole, a dynamic weighted loss function is adopted, with Focal Loss for classification tasks and Huber Loss for regression tasks.

8. A device for snowmaking control in civil aviation power plant airworthiness verification, characterized in that, The device includes: The multimodal snowmaking control data unit is used to acquire multimodal snowmaking control data, which includes: meteorological data, snowmaking machine equipment operation data, snow quality monitoring data, acoustic data, and experimental environment data, to construct a multimodal snowmaking control dataset. The multimodal fusion model unit is used to build a multimodal fusion model based on CNN network and Transformer. The multimodal fusion model is trained using a multimodal snowmaking control dataset to obtain a trained multimodal fusion model. The output unit is used to deploy the trained multimodal fusion model, input the real-time acquired multimodal snowmaking control data into the trained multimodal fusion model, and adjust the snowmaking control system based on the control parameters output by the multimodal fusion model.

9. A computer device comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the snowmaking control method for airworthiness verification of civil aviation power units as described in any one of claims 1-7.

10. A storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the snowmaking control method for airworthiness verification of civil aviation power units as described in any one of claims 1-7.

Images