Blade icing monitoring method based on meteorological data and deep learning, electronic equipment and medium

Abstract

The invention provides a blade icing monitoring method based on meteorological data and deep learning, an electronic device and a medium, features are extracted from meteorological data directly influencing the blade icing rate and form, and the blade icing probability is predicted. According to the invention, a deep adaptive standardized network module is adopted as an encoder to map meteorological data actually from different distributions into the same distribution interval, so that the non-stationarity in an actual data source can be resisted, and the adaptability and generalization ability of the model to different meteorological environment conditions are improved; according to the method, a parallel structure of the multi-head attention mechanism module and the multi-head one-dimensional convolution module is combined as a feature extractor to extract overall and local icing features in meteorological data, the method has the capability of capturing the overall trend of blade icing along with time change, and the blade icing monitoring precision is improved.

Description

Technical Field

[0001] The present invention belongs to the field of indirect monitoring of wind turbine blade icing, and in particular relates to a blade icing monitoring method, electronic equipment, and medium based on meteorological data and deep learning. Background Art

[0002] Wind turbines are categorized as offshore and onshore. Onshore wind farms are typically located in mountainous areas at high altitudes and latitudes, facing the threat of winter cold snaps. This can lead to large-scale ice formation on turbine blades, impacting power system stability. Therefore, real-time blade ice monitoring is crucial to help mitigate this issue.

[0003] Traditional wind turbine blade icing monitoring methods can be divided into direct monitoring and indirect monitoring methods. The direct detection method focuses on the changes in physical properties caused by blade icing, and identifies the presence of ice by measuring parameters such as blade mass, pressure level, thermal conductivity, and inductance. However, this method is limited by the sensors, which not only increases equipment and maintenance costs, but is also easily interfered with by external conditions. For example, changes in micrometeorology may affect the accuracy of measurements. In contrast, the indirect monitoring method can better overcome these shortcomings and effectively identify the blade icing status by indirectly measuring parameters related to icing, such as output power and noise caused by blade imbalance force.

[0004] Driven by data science and machine learning technologies, researchers are making full use of wind turbine operation and environmental data to monitor the icing status of wind turbine blades. This data-driven approach can respond quickly under various environmental conditions and operating conditions, and has shown broad application prospects. Early researchers have applied some shallow machine learning methods, such as random forest (RF), extreme gradient boosting (XGBoost), kernel extreme learning machine (KELM), etc., to solve the problem of icing monitoring. However, when processing icing data, problems such as high data dimension, data instability, or insufficient expression of icing information are often encountered. Feature engineering is considered a forward-looking tool that is expected to improve the accuracy of data-driven methods in icing status diagnosis. Deep learning methods have autonomous learning capabilities and can be used to automatically extract highly abstract feature representations of sensor data. They have broad application potential in blade icing monitoring.

[0005] Currently, feature engineering mainly extracts features from the perspective of the mechanical state of icing, with less consideration given to meteorological parameters that directly affect the icing mechanism. Various meteorological parameters and the duration of icing will lead to changes in the rate and morphology of icing. Secondly, the generalization ability of the current blade icing monitoring model is insufficient, because the actual icing data includes different wind farms, and the operating environments between wind farms are very different, resulting in the dataset showing a non-stationary change trend. The introduction of new samples may disrupt the regularity of the features in the original dataset, reducing the model's ability to adapt to new samples. Recursive neural networks are often used to deal with time series classification and regression problems, but there are problems such as difficulty in capturing long-term dependencies in the data, resulting in a decrease in prediction accuracy. Summary of the Invention

[0006] The purpose of the present invention is to address the shortcomings of the existing technology and provide a blade icing monitoring method, electronic equipment and medium based on meteorological data and deep learning, which can solve the non-stationary effects in actual data sources and improve the adaptability and generalization ability of the blade icing monitoring model to different meteorological environmental conditions.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] A blade icing monitoring method based on meteorological data and deep learning includes the following steps:

[0009] S1: Obtain meteorological data of the wind farm;

[0010] S2: Encoding the meteorological data of the wind farm using a deep adaptive normalized network module to extract a time feature sequence from the meteorological data;

[0011] S3: Using a multi-head attention mechanism module to capture the global dependency of the time feature sequence and extract the global features in the time feature sequence; using a multi-head one-dimensional convolution module to capture the local dependency of the time feature sequence and extract the local features in the time feature sequence;

[0012] S4: combining the global feature and the local feature to obtain a mixed feature;

[0013] S5: Classify the mixed features to obtain blade icing probability.

[0014] In practical applications, meteorological datasets usually come from different wind farms. The geographical differences between different wind farms and wind turbines can lead to significant differences in the micro-meteorological characteristics of the wind turbines. In particular, due to differences in altitude, the meteorological data of the actual operation of wind turbines, such as temperature and liquid water content, may show large distribution deviations in the time dimension, making the model trained for a certain wind turbine unable to be directly applied to ice detection for another wind turbine. At the same time, wind turbine icing is a dynamic process that is constrained by time, and there are long-term dependencies between meteorological data. The blade icing monitoring model needs to be able to capture the overall trend of blade icing over time. In addition, ice formation is accompanied by specific local patterns, such as a sharp drop in temperature or an increase in humidity, which also affects the judgment of the model. The present invention adopts the Deep Adaptive Normalized Network module (DAIN) as an encoder to map meteorological data actually derived from different distributions into the same distribution interval, which can solve the problem of non-stationary influence in the actual data source and improve the adaptability and generalization ability of the model to different meteorological environmental conditions; the present invention combines the multi-head attention mechanism module (Transformer) and the multi-head one-dimensional convolution module (TCN) parallel structure as a feature extractor to capture icing information from both global and local perspectives, extract overall and local icing features, corresponding to the actual blade icing process, and has the ability to capture the overall trend of blade icing changes over time, thereby improving the accuracy of blade icing monitoring.

[0015] Furthermore, the meteorological data includes ambient temperature, wind speed, liquid water content in the air and relative humidity.

[0016] Furthermore, the deep adaptive standard network module includes an adaptive mobile layer, an adaptive scaling layer and an adaptive gating layer. The adaptive mobile layer is connected to the adaptive scaling layer, and the adaptive scaling layer is connected to the adaptive gating layer. The adaptive mobile layer maps the meteorological data to the feature interval, and the adaptive scaling layer linearly scales the meteorological data in the feature interval. The adaptive gating layer filters out the non-icing features in the meteorological data after linear scaling to obtain a time feature sequence.

[0017] Furthermore, the time characteristic series is calculated by the following formula:

[0018]

[0019]

[0020]

[0021] Time feature series

[0022] The meteorological data {X( i) ∈Rd×L ; i=1,.....,N} is a set of N time series, each time series consists of L-dimensional features, represents the d features observed in time series i at time point j, W a is the weight matrix of the adaptive mobile layer, b (i) yes The standard deviation, W b is the weight matrix of the adaptive scaling layer, W c is the weight matrix of the adaptive gating layer, yes Temporal features obtained after processing by the deep adaptive standard network module.

[0023] Furthermore, the multi-head attention mechanism module includes multiple attention mechanism modules, a splicing layer, and an average pooling layer. The attention mechanism module includes a fully connected layer and an attention mechanism layer. The multiple attention mechanism modules are connected in parallel to the splicing layer. The splicing layer is connected to the average pooling layer. The weight matrix of each attention mechanism module is different.

[0024] Furthermore, the global feature acquisition process includes:

[0025] (1) Time feature series in Perform linear transformation to obtain the query vector q j , key vector k j Sum value vector v j , W q 、W k and W v is the weight matrix of the attention mechanism module;

[0026] (2) Calculate the weighted sum of attention weights d k It is k j Dimensions;

[0027] (3) Perform steps (1)-(2) on each feature in the time feature sequence to obtain the output sequence of the attention mechanism module;

[0028] (4) Steps (1)-(3) are performed simultaneously in each attention mechanism module, and the weight matrix W of each attention mechanism module is q 、W k and W v different;

[0029] (5) The output sequences of all attention mechanism modules are spliced ​​together, and the spliced ​​output sequences are processed by average pooling to obtain global features.

[0030] Furthermore, the multi-head one-dimensional convolution module includes multiple parallel one-dimensional convolution modules, the one-dimensional convolution module includes a one-dimensional convolution layer, a flattening layer, and a fully connected layer, the one-dimensional convolution layer is connected to the flattening layer, the flattening layer is connected to the fully connected layer, and the convolution kernel of each one-dimensional convolution module is different.

[0031] Furthermore, the local features are obtained through the following process:

[0032] (1) The time feature sequence X is processed in parallel by each one-dimensional convolution module to obtain the output H of each one-dimensional convolution module i ,

[0033] H i =ReLU(h i W i.o +b i,o )

[0034] h i =Flatten(ReLU(X*W i,k +b i ))

[0035] (2) Concatenate the outputs of all one-dimensional convolution modules to obtain the local feature Y.

[0036] Y=Concat(H1,H2,...,H k )W O

[0037] Concat is to concatenate the outputs of all one-dimensional convolution modules according to the feature dimension, W O is the weight matrix, W i,k is the convolution kernel weight matrix of the i-th one-dimensional convolution module, b i is the bias term of the i-th one-dimensional convolution module. * represents the one-dimensional convolution operation, ReLU is the activation function, Flatten is the flattening operation, W i.o is the fully connected layer weight matrix of the i-th one-dimensional convolution module, b i,o is the fully connected layer bias term of the i-th one-dimensional convolution module.

[0038] Based on the same concept, the present invention provides an electronic device, comprising:

[0039] one or more processors;

[0040] A memory having one or more programs stored thereon, which, when executed by the one or more processors, enables the one or more processors to implement the steps of a blade icing monitoring method based on meteorological data and deep learning.

[0041] Based on the same concept, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a blade icing monitoring method based on meteorological data and deep learning.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] (1) The present invention uses DAIN as an encoder to map meteorological data sets from different distributions into the same distribution interval, which can solve the problem of non-stationary influence in the actual data source and improve the adaptability and generalization ability of the model to different meteorological environmental conditions.

[0044] (2) The present invention combines the Transformer and TCN parallel structures as feature extractors to extract icing features globally and locally, and has the ability to capture the overall trend of blade icing changes over time, thereby improving the accuracy of blade icing prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a structural diagram of the blade icing monitoring method based on meteorological data and deep learning of the present invention;

[0046] Figure 2 This is a structural diagram of the deep adaptive standardized network module of the present invention;

[0047] Figure 3 This is a structural diagram of the multi-head attention mechanism module of the present invention;

[0048] Figure 4 This is a structural diagram of the multi-head one-dimensional convolution module of the present invention. DETAILED DESCRIPTION

[0049] The present invention will be described in detail below with reference to the following embodiments. It should be noted that the embodiments and features of the embodiments may be combined unless they conflict. For ease of description, the words "upper," "lower," "left," and "right" appearing below merely indicate the directions of upper, lower, left, and right relative to the accompanying drawings and do not limit the structure.

[0050] Example

[0051] In this example, 16 wind farms distributed in different areas of Guilin were selected as data sources. To ensure data quality, 52 wind turbines under different operating conditions were selected. The dataset spans from November 25, 2022, to February 3, 2023, which basically covers the entire winter in Guilin and can fully reflect the operating conditions of wind turbines under conditions where icing may occur. The ice sample labels in the dataset are collected in real time by on-site staff, with a sampling frequency of 3 hours. The sample features are derived from the ERA5 meteorological dataset, with a sampling frequency of 1 hour. This meteorological dataset provides information on the vertical distribution of meteorological data in space, maps the altitude of each wind turbine to the corresponding pressure range, and can relatively accurately restore the actual meteorological conditions during the operation of each wind turbine. In this example, meteorological data from 42 wind turbines in 12 wind farms were selected as the training set, while meteorological data from 10 wind turbines in another 4 wind farms were selected as the test set. In addition, when conducting model testing, the four wind farms in the test set were treated as independent entities for testing. The purpose was to preserve the independence of each wind farm while studying the overall performance of the test set, so as to more comprehensively understand the performance of the model in different wind farms.

[0052] like Figure 1-Figure 4 As shown in Figure 2, the blade icing monitoring method based on meteorological data and deep learning includes:

[0053] Step 1: According to the Makkonen ice growth model, ice can be understood as the collision, collection and freezing of liquid water in the air with the blades in three stages, which can be expressed as:

[0054] dM=αβfWVAdt

[0055] Where dM is the ice mass, α is the local impact coefficient, β is the supercooled droplet capture coefficient, f is the local freezing coefficient, W is the water content of the particle swarm, V is the effective relative velocity of the particles, A is the effective cross-sectional area of ​​the object with which the droplets collide, and dt represents time. This numerical simulation model incorporates the air flow field, droplet impact characteristics, and liquid water freezing heat balance calculations. It also investigates the impact of various meteorological factors on the type and intensity of icing. Increased wind speed increases wind turbine rotation speed, accelerating heat exchange on the blade surface and increasing the rate of icing. As temperature decreases, the temperature difference of the water film on the blade surface increases, accelerating the freezing of droplets. Increased liquid water content and humidity lead to an increase in the number of supercooled droplets, intensifying the impact force on the blade surface and further exacerbating icing. It is believed that the main meteorological parameters affecting icing on wind turbine blade surfaces are ambient temperature, wind speed, liquid water content in the air, and relative humidity. Icing is considered a dynamic process, and a time series is used to represent the temporal evolution of the data.

[0056] Step 2: The encoding module is a learnable Deep Adaptive Normalization Network (DAIN) module, which is used to map meteorological data sets from different distributions into a common distribution interval. This aims to combat non-stationarity in the actual data source and improve the model's adaptability and generalization capabilities to complex environmental conditions. The specific implementation includes three key adaptive layers: an adaptive shifting layer, an adaptive scaling layer, and an adaptive gating layer. These layers can participate in directional propagation and achieve self-update. After passing through the encoder, the meteorological data retains its time series features and is output as an embedding vector, which provides input for the subsequent feature extraction module. The adaptive shifting layer maps the meteorological data to the feature interval, the adaptive scaling layer linearly scales the meteorological data within the feature interval, and the adaptive gating layer filters out non-ice features in the linearly scaled meteorological data, resulting in a time feature series.

[0057] Step 3: The feature extraction module is divided into two parts, aiming to extract different forms of ice cover features in the time series of meteorological data. First, one-dimensional convolution is used to capture the local features of meteorological data in the time series. On this basis, a multi-head one-dimensional convolution module is constructed to extract ice cover information under different receptive fields. Secondly, a Transformer network is used to perceive the global representation of temporal information, and a multi-head self-attention mechanism module is used to capture the global dependencies of meteorological data in the time series. The two parts are stacked and average pooled and output as hidden states. Then, a fully connected layer is used to extract higher-level abstract features of the meteorological data. Finally, a mixed ice cover feature is formed by splicing.

[0058] The mixed icing features are classified and processed by the fully connected layer to obtain the blade icing probability. A blade icing probability greater than 0.5 indicates ice coverage, while a blade icing probability less than or equal to 0.5 indicates no ice coverage.

[0059] The multi-head attention mechanism module includes multiple attention mechanism modules, splicing layers, and average pooling layers. The attention mechanism module includes a fully connected layer and an attention mechanism layer. Multiple attention mechanism modules are connected in parallel with the splicing layer, and the splicing layer is connected to the average pooling layer. The weight matrix of each attention mechanism is different.

[0060] The multi-head one-dimensional convolution module includes multiple parallel one-dimensional convolution modules. The one-dimensional convolution module includes a one-dimensional convolution layer, a flattening layer, and a fully connected layer. The one-dimensional convolution layer is connected to the flattening layer, and the flattening layer is connected to the fully connected layer. The convolution kernel of each one-dimensional convolution module is different.

[0061] Step 4: Use comparative experiments to verify the practical significance of the two modules in the deep learning network of this embodiment. To verify the effectiveness of the proposed model, it is compared with traditional recurrent neural networks (LSTM and GRU). In addition, ablation experiments are performed on the encoder module to gain a deeper understanding of its contribution to the overall performance. To further verify the effectiveness of the feature extraction module in the model, ablation experiments are performed on this combined model. To ensure the best performance of the model after ablation, ablation experiments of the feature extraction module are all performed on the basis of the encoder module.

[0062] The Deep Adaptive Normalization Network (DAIN) module is used as an encoding module to map meteorological data sets from different distributions into the same distribution interval. Time series is used to represent the changing pattern of meteorological data in the time dimension. (i) ∈R d×L ; i=1,.....,N} is a set of N time series, each time series consists of L-dimensional features, represents the d features observed at time point j in time series i. First, set a measurement indicator to the average value of the aggregation over the current time series:

[0063]

[0064] The mean of the current time series provides an initial estimate that can be used to estimate the distribution generated by the current time series in order to modify the normalization process appropriately. Then define the shift operator α (i) for:

[0065] α (i) =W a a (i) ∈R d (2)

[0066] Where W a ∈R d×d is the weight matrix of the first neural layer (adaptive shift layer), which is responsible for shifting the measurements in each dimension. The meteorological data is then appropriately scaled. An updated aggregate representation is calculated, which corresponds to The standard deviation of the kth feature data in:

[0067]

[0068] express The kth feature in yes The shift operator of the kth feature in .

[0069] Define the scaling function as a linear transformation of the standard deviation above and apply it to the observations at each location:

[0070] β (i) =W b b (i) ∈R d (4)

[0071] Where W b ∈R d×d is the weight matrix of the scaling layer, b (i) yes This layer is called the adaptive scaling layer, which estimates how the weather data should be scaled before feeding it into the network. Its output is:

[0072]

[0073] Finally, the data is fed into an adaptive gating layer that is able to suppress irrelevant or useless features:

[0074]

[0075] γ (i) =sigmoid(W c c (i) +d)∈R d (7)

[0076]

[0077] Where W c ∈R d×d is the gating layer parameter.

[0078] The multi-head self-attention mechanism can focus on the information of the entire time series at each time step, allowing the model to fully capture the long-term ice cover trend. Specifically:

[0079] The input sequence is in is the jth vector in the input sequence. Obtain the corresponding query vector q through linear transformation j , key vector k j Sum value vector v j .

[0080]

[0081]

[0082]

[0083] Where W q 、W k and W vis the weight matrix used to map the input vector to q j 、k j and v j By calculating the attention weights, we get the weighted sum output:

[0084]

[0085] Where d k It is k j The attention distribution is randomly masked to prevent the model from focusing on irrelevant time points, thereby improving its robustness. The softmax function is used to calculate the attention weights. The final output is the weighted sum of the attention weights. This process is repeated for all vectors in the input sequence X to form the final output sequence.

[0086] In the multi-head attention mechanism, the above operations are applied in parallel to multiple attention mechanism modules, each of which has a different weight matrix W q 、W k and W v ,Finally, the outputs of multiple attention mechanism modules are concatenated and ,through an average pooling to obtain the final output.

[0087] The multi-head one-dimensional convolution module enables each convolution kernel to act as an independent "head" responsible for learning local features of different scales in time series. The specific principle is:

[0088] The input time series is X∈R d×L , the output of the multi-head one-dimensional convolution module is:

[0089] Y=Concat(H1,H2,...,H k )W O (13)

[0090] Where H i Represents the output of the i-th one-dimensional convolution module, which is obtained by one-dimensional convolution operation and full connection. Concat is the operation of connecting the outputs of all one-dimensional convolution modules according to the feature dimension. W O is the weight matrix.

[0091] H i The calculation process is:

[0092] H i =ReLU(h i W i.o +b i,o ) (14)

[0093] h i =Flatten(ReLU(X*Wi,k +b i )) (15)

[0094] In the formula is the convolution kernel weight matrix of the i-th one-dimensional convolution module, where K i is the size of the convolution kernel, F i is the number of features. is the bias term for the i-th one-dimensional convolution module. * denotes a one-dimensional convolution operation, ReLU is the activation function, typically a rectified linear unit. Flatten is a flattening operation that converts the multidimensional tensor output by the convolution into a one-dimensional vector. is the fully connected layer weight matrix of the i-th one-dimensional convolution module, b i,o ∈R O is the fully connected layer bias term of the i-th one-dimensional convolution module.

[0095] Table 1 shows the evaluation indicators of each model in the encoding module ablation experiment. It can be seen that compared with the LSTM and GRU architectures represented by the traditional recursive neural network for processing time series problems, the model of this embodiment is more stable in classification performance and has stronger generalization ability. Overall, in wind farm 1, each model showed high classification performance. In other data sets (wind farm 2, wind farm 3, wind farm 4), the classification performance of the model of this embodiment has been improved to varying degrees. In particular, in wind farm 3, compared with the LSTM model, the accuracy, precision, recall and F1 parameters have been improved by 1.76%, 2.57%, 12.45% and 4.52% respectively. The reason is that when processing ice cover features in a global and local combination of perception, it is more sensitive to data than the recursive neural network.

[0096] On the other hand, after removing the encoding module, the performance indicators of the model in this embodiment declined to varying degrees. The decline was most significant in Wind Farm 3 and Wind Farm 4, where the F1 parameter decreased by 11.38% and 6.58%, respectively. More notably, the model showed a general decline in recall, meaning that some samples that were actually iced were misclassified, resulting in a high false negative rate. This is because the presence of the encoding module enhances the representation of ice in the data, facilitating the model's identification of these samples. Furthermore, the LSTM model's prediction accuracy declined most significantly in Wind Farm 1, with the F1 parameter decreasing by approximately 12.52%.

[0097] Table 1 Evaluation indicators of each model in the encoding module ablation experiment

[0098]

[0099]

[0100] Table 2 lists the evaluation metrics of each model in the feature extraction module ablation experiment. To further verify the effectiveness of the feature extraction module in this embodiment, ablation experiments were conducted on the model. To ensure optimal performance after ablation, ablation experiments on the feature extraction module were conducted on models with an encoding module. It can be seen that the combined approach proposed in this embodiment for extracting icing features results in more robust model classification performance. Overall, the best-performing models for each wind farm in the test set were the model of this embodiment, DAIN-Transformer, and DAIN-TCN. All three models demonstrated high prediction accuracy for Wind Farm 1, indicating a strong ability to capture icing information in the data from this wind farm. The ablation experiments revealed that DAIN-TCN exhibited significant degradation in classification performance compared to the model of this embodiment, particularly in accuracy, with a decrease of 10.75%. While DAIN-Transformer and the combined model performed similarly, prediction performance was poor in Wind Farms 3 and 4, with decreases in the F1 parameter of 13.01% and 4.87%, respectively.

[0101] Table 2 Evaluation indicators of each model in the feature extraction module ablation experiment

[0102]

[0103] This embodiment provides an electronic device, including:

[0104] one or more processors;

[0105] A memory having one or more programs stored thereon, which, when executed by one or more processors, enables the one or more processors to implement the steps of a blade icing monitoring method based on meteorological data and deep learning.

[0106] In some implementations, the memory may be a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk storage.

[0107] In other implementations, the processor may be a central processing unit (CPU), a digital signal processor (DSP), or other general-purpose processors, which are not limited here.

[0108] This embodiment provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a blade icing monitoring method based on meteorological data and deep learning.

[0109] The contents illustrated in the above embodiments should be understood as these embodiments are only used to more clearly illustrate the present invention, and are not used to limit the scope of the present invention. After reading the present invention, various equivalent modifications of the present invention by those skilled in the art shall fall within the scope defined by the claims attached to this application.

Claims

1. A blade icing monitoring method based on meteorological data and deep learning, characterized in that: The following steps are involved: S1: Obtain meteorological data of the wind farm; S2: Encoding the meteorological data of the wind farm using a deep adaptive normalized network module to extract a time feature sequence from the meteorological data; S3: Using a multi-head attention mechanism module to capture the global dependency of the time feature sequence and extract the global features in the time feature sequence; using a multi-head one-dimensional convolution module to capture the local dependency of the time feature sequence and extract the local features in the time feature sequence; S4: combining the global feature and the local feature to obtain a mixed feature; S5: Classify the mixed features to obtain blade icing probability.

2. The blade icing monitoring method based on meteorological data and deep learning according to claim 1, characterized in that: The meteorological data includes ambient temperature, wind speed, liquid water content in the air and relative humidity.

3. The blade icing monitoring method based on meteorological data and deep learning according to claim 1, characterized in that: The deep adaptive standard network module includes an adaptive mobile layer, an adaptive scaling layer and an adaptive gating layer. The adaptive mobile layer is connected to the adaptive scaling layer, which is connected to the adaptive gating layer. The adaptive mobile layer maps meteorological data to a feature interval, the adaptive scaling layer linearly scales the meteorological data in the feature interval, and the adaptive gating layer filters out non-icing features in the meteorological data after linear scaling to obtain a time feature sequence.

4. The blade icing monitoring method based on meteorological data and deep learning according to claim 3 is characterized in that: The time characteristic series is calculated by the following formula: Time feature series The meteorological data {X (i) ∈R d×L ; i=1,.....,N} is a set of N time series, each time series consists of L-dimensional features, represents the d features observed in time series i at time point j; W a is the weight matrix of the adaptive mobile layer, b (i) yes The standard deviation, W b is the weight matrix of the adaptive scaling layer, W c is the weight matrix of the adaptive gating layer, yes Temporal features obtained after processing by the deep adaptive standard network module.

5. The blade icing monitoring method based on meteorological data and deep learning according to claim 1, characterized in that: The multi-head attention mechanism module includes multiple attention mechanism modules, a splicing layer, and an average pooling layer. The attention mechanism module includes a fully connected layer and an attention mechanism layer. The multiple attention mechanism modules are connected in parallel to the splicing layer. The splicing layer is connected to the average pooling layer. The weight matrix of each attention mechanism module is different.

6. The blade icing monitoring method based on meteorological data and deep learning according to claim 5, characterized in that: The global feature acquisition process includes: (1) Time feature series in Perform linear transformation to obtain the query vector q j , key vector k j Sum value vector v j , W q 、W k and W v is the weight matrix of the attention mechanism module; (2) Calculate the weighted sum of attention weights d k It is k j Dimensions; (3) Perform steps (1)-(2) on each feature in the time feature sequence to obtain the output sequence of the attention mechanism module; (4) Steps (1)-(3) are performed simultaneously in each attention mechanism module, and the weight matrix W of each attention mechanism module is q 、W k and W v different; (5) The output sequences of all attention mechanism modules are spliced ​​together, and the spliced ​​output sequences are processed by average pooling to obtain global features.

7. The blade icing monitoring method based on meteorological data and deep learning according to claim 1, characterized in that: The multi-head one-dimensional convolution module includes multiple parallel one-dimensional convolution modules, and the one-dimensional convolution module includes a one-dimensional convolution layer, a flattening layer, and a fully connected layer. The one-dimensional convolution layer is connected to the flattening layer, and the flattening layer is connected to the fully connected layer. The convolution kernel of each one-dimensional convolution module is different.

8. The blade icing monitoring method based on meteorological data and deep learning according to claim 7, characterized in that: The local features are obtained through the following process: (1) The time feature sequence X is processed in parallel by each one-dimensional convolution module to obtain the output H of each one-dimensional convolution module i , A i =ReLU(h i ·W i.o +b i,o ) h i =Flatten(ReLU(X*W i,k +b i )) (2) Concatenate the outputs of all one-dimensional convolution modules to obtain the local feature Y. Y=Concat(H1,H2,...,H k )W O Concat is to concatenate the outputs of all one-dimensional convolution modules according to the feature dimension, W O is the weight matrix, W i,k is the convolution kernel weight matrix of the i-th one-dimensional convolution module, b i is the bias term of the i-th one-dimensional convolution module. * represents the one-dimensional convolution operation, ReLU is the activation function, Flatten is the flattening operation, W i.o is the fully connected layer weight matrix of the i-th one-dimensional convolution module, b i,o is the fully connected layer bias term of the i-th one-dimensional convolution module.

9. An electronic device, characterized in that: include: one or more processors; A memory having one or more programs stored thereon, which, when executed by the one or more processors, enables the one or more processors to implement the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that The computer program is stored therein, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images