A method for predicting wireless charging efficiency of drone hangar

Abstract

The present invention proposes a method for predicting the efficiency of wireless charging of drone hangars, and relates to the technical field of machine learning. The present invention proposes an MDPM prediction model, which includes a multi-scale feature extraction module, a dynamic mode fusion module, and a prediction decoding layer. Specifically, the multi-scale feature extraction module is used to process the long-term dependency and complex multi-periodic characteristics in the wireless charging data of drone hangars, the dynamic mode fusion module is used to process the combination of short-term dynamic changes and long-term historical patterns in the wireless charging data of drone hangars, and the prediction decoding layer enables the model to predict in multiple steps. Each module has good effects in capturing multi-periodicity, processing complex nonlinear wireless charging data of drone hangars, and integrating multiple features for prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and in particular relates to a method for predicting the wireless charging efficiency of an unmanned aerial vehicle hangar. Background Art

[0002] With the development of drone technology, the scope of use of drones has been continuously expanded, and they play an important role in many fields such as logistics, agriculture, and security. However, the drone's battery life is limited, and frequent charging has become a major obstacle in the use process. Traditional charging methods usually require manual intervention, which affects the efficiency of drone automation. In order to solve this problem, wireless charging technology has been gradually introduced into drone hangar systems, and the operation continuity and maneuverability of drones are improved through automated charging methods.

[0003] In practical applications, the wireless charging efficiency of drones is affected by multiple factors, such as environmental conditions, drone attitude, electromagnetic interference, etc. These factors cause the charging efficiency to present complex multi-periodicity and non-stationary characteristics. Secondly, during the wireless charging process, the charging efficiency changes show strong nonlinear characteristics, and the distribution deviates from the Gaussian distribution. Traditional linear models are difficult to cope with such complex characteristics. At the same time, there may be significant differences between the long-term historical data and recent observation data of drone wireless charging efficiency. Traditional methods are difficult to take into account these two parts of information at the same time. These factors make it difficult to accurately predict the wireless charging efficiency of drone hangars.

[0004] Methods in the field of machine learning, such as multi-scale time series analysis, can extract multi-period features at different time scales from long-term historical data through model representation learning, and can also handle non-Gaussian and nonlinear distributions of data, thereby achieving accurate pressure prediction. Machine learning methods can not only effectively cope with the complexity of pressure data, but also provide strong data support for the operation optimization and maintenance of wireless charging equipment in drone hangars, thereby improving the efficiency and safety of energy management. Summary of the invention

[0005] The present invention provides a method for predicting the wireless charging efficiency of a UAV hangar. Aiming at the wireless charging data of a UAV hangar with multi-periodicity and long-term dependence, an MDPM model is proposed, which consists of a multi-scale feature extraction module (MFEM), a dynamic pattern fusion module (DMFM) and a prediction decoding layer (PDL).

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose specifically includes the following steps:

[0007] S1. Collect relevant data of wireless charging of drone hangar through sensors, including environmental data of wireless charging of drone hangar and parameter data of wireless charging equipment of drone hangar;

[0008] S2, preprocessing the data related to wireless charging in the drone hangar, including data outlier processing and normalization, and dividing the data set;

[0009] S3. Construct a multi-scale feature extraction module to handle the long-term dependencies and complex multi-cycle characteristics in the wireless charging data of the drone hangar. The specific steps are as follows:

[0010] S31. Input the wireless charging sequence data of the drone hangar ,in is the time step, The feature number of the wireless charging data in the drone hangar is randomly selected, and two overlapping subsequences are mapped to the potential representation through a fully connected layer. , using time series masking for data augmentation;

[0011] S32, Extracting contextual feature representation of drone hangar wireless charging data using multi-layer time domain convolutional network , and then extract the time domain representation of the drone hangar wireless charging data through a set of 1D dilated convolutional layers and random pooling operations ;

[0012] S33, context features are represented by short-time Fourier transform and fully connected layers Convert to frequency domain to get frequency domain features , proposed frequency feature attention , calculate the attention score and get the weighted frequency domain representation , the time domain is represented by and frequency domain representation Fusion to obtain multi-scale representation ;

[0013] S4. Build a dynamic pattern fusion module to process the combination of short-term dynamic changes and long-term historical patterns in the wireless charging data of the drone hangar. The specific steps are as follows:

[0014] S41. Encoding short-term drone hangar wireless charging data through a gated neural network to obtain contextual latent representation ;

[0015] S42, multi-scale representation and contextual latent representation Mapped to the same dimension through the fully connected layer and fused using sparse multi-head attention to obtain the fused feature representation ;

[0016] S43. Using radial flow to approximate the probability density of wireless charging data in drone hangars ;

[0017] S5. Construct a prediction decoding layer, use the gated neural network as a decoder, perform autoregressive decoding and implement multi-step prediction.

[0018] Optimally, in step S1, the environmental data of the drone hangar during wireless charging and the parameter data of the drone hangar wireless charging device itself include temperature data, humidity data, internal and external wind speed data, and electromagnetic environment data in the hangar, and the parameter data of the hangar wireless charging device includes charging power data, charging efficiency data, voltage data, current data, device temperature data, and device withstand voltage data.

[0019] Preferably, in step S2, the drone hangar wireless charging related data is preprocessed, including data outlier processing and normalization, and the data set is divided proportionally. The specific formula for outlier processing is as follows:

[0020] ;

[0021] In the formula, Wireless charging data points for drone hangars, The average value of wireless charging data in the drone hangar. The standard deviation of the wireless charging data in the drone hangar, is the standardized deviation.

[0022] Preferably, in the steps S3 and S31, the wireless charging sequence data of the drone hangar is input. ,in is the time step, The number of features of the wireless charging data in the drone hangar, randomly selecting two overlapping subsequences and ,in , and are the starting and ending indexes of the first subsequence, and are the start and end indexes of the second subsequence, mapped to the potential representation through a fully connected layer , and use time series masking for data enhancement. The specific formula is:

[0023] ;

[0024] In the formula, is the potential representation after masking by time series stamps, is the fully connected layer, is the input drone hangar wireless charging subsequence data, is a binary masking matrix, It is an element-by-element multiplication operation.

[0025] Preferably, by randomly selecting two overlapping subsequences, the model can better capture the complex periodic features in the data and map them to potential representations using a fully connected layer, while using time series masking for data enhancement operations to make the model more robust when processing data and reduce the risk of overfitting. This method can capture the multi-periodic features in the wireless charging data of drone hangars, and at the same time improve the generalization ability of the model through data enhancement.

[0026] Preferably, in steps S3 and S32, the inputted drone hangar wireless charging sequence data , using a multi-layer temporal convolutional network to extract contextual feature representation of drone hangar wireless charging data , the specific formula is as follows:

[0027] ;

[0028] In the formula, is the ReLU activation function, is the layer index of the time domain convolutional network, is the total number of layers of the time domain convolutional network, is the index of the convolution kernel, is the size of the convolution kernel, For the In the convolution operation, The convolution kernel weights at each position, For the Layer at time step The output at For the The dilation factor of the layer, For the The bias term of the layer, For the Layer at time step The output of is used to extract the time domain representation of the drone hangar wireless charging data through a set of 1D dilated convolutional layers and random pooling operations. The specific formula is as follows:

[0029] ;

[0030] ;

[0031] In the formula, is a 1D dilated convolution operation, Contextual feature representation for drone hangar wireless charging data, is the size of the convolution kernel, is the expansion rate, For the The output of the dilated convolution layer, is a random pooling operation, The time domain representation of the wireless charging data in the drone hangar. is a hyperparameter that determines the number of dilated convolutional layers used in the network.

[0032] Preferably, a multi-layer time domain convolutional network is used to extract contextual feature representation, and a set of 1D dilated convolutional layers and random pooling operations are used to extract time domain representation. Time domain convolution can capture the contextual information of drone charging data, and the combination of dilated convolution and random pooling further enhances the model's ability to express long-term dependent features. This combination significantly improves the model's ability to handle complex time-dependent features, especially in the face of long time series. It can effectively capture the changing patterns of charging data, thereby improving prediction accuracy and making the model more advantageous in long time series prediction.

[0033] Preferably, in steps S3 and S33, the context feature representation is represented by a short-time Fourier transform and a fully connected layer. Switch to the frequency domain to obtain the frequency domain features. The specific formula is as follows:

[0034] ;

[0035] In the formula, is the fully connected layer, is the short-time Fourier transform operation, Contextual feature representation for drone hangar wireless charging data, is the length of the sliding window in the short-time Fourier transform , proposed frequency feature attention , for frequency domain features Application, first extract the amplitude of the frequency component, the specific formula is as follows:

[0036] ;

[0037] In the formula, and are the real and imaginary parts of the frequency domain features, respectively. is the frequency domain feature, and then the multi-head spectrum scaling mechanism is used to scale the query vector and key vector Projection, value vector It is generated by linear mapping, and the specific formula is as follows:

[0038] ;

[0039] ;

[0040] ;

[0041] In the formula, and It is a multi-head spectrum scaling mechanism used to generate query vectors and key vectors. is the amplitude of the frequency component, is a linear mapping function, The input is the wireless charging sequence data of the drone hangar. The multi-head spectrum scaling mechanism captures the correlation between different frequencies by scaling the frequency components. The specific formula is as follows:

[0042] ;

[0043] In the formula, is the amplitude of the frequency component, is a learnable scaling matrix, is element-by-element multiplication, representing and The corresponding elements are multiplied one by one, and then the attention score is calculated to obtain the weighted frequency domain representation. The specific formula is as follows:

[0044] ;

[0045] In the formula, is the activation function, is the query vector, is the transpose of the key vector, is the key vector The dimension of As a value vector, the time domain is finally represented and frequency domain representation Fusion to obtain multi-scale representation .

[0046] Preferably, the context feature representation is converted to the frequency domain through short-time Fourier transform, and a frequency feature attention mechanism is proposed, combined with multi-head spectrum scaling to capture the correlation between different frequencies. The frequency domain conversion enables the model to identify and capture the periodic changes in the charging process of the drone, while the multi-head spectrum scaling mechanism enhances the model's understanding and utilization of different frequency features, improves the ability to capture complex periodic data, and improves the model's ability to capture and predict changing trends in charging efficiency.

[0047] Preferably, in the steps S4 and S41, a gated neural network is used to encode the short-term wireless charging data of the drone hangar to generate a contextual latent representation, which can capture the change pattern of the wireless charging data of the drone hangar in a short period of time. The specific formula is as follows:

[0048] ;

[0049] In the formula, is a gated recurrent unit, is the input short-term UAV hangar wireless charging data, representing the time step To time step The data between is the input short-term UAV hangar wireless charging covariate data, representing the time step To time step The covariate data between for The learnable parameters of .

[0050] Preferably, a gated neural network is used to encode short-term charging data to generate contextual latent representations. The gated neural network can effectively capture short-term dynamic changes and improve the model's sensitivity to short-time series data. Specifically, it can process short-term change patterns during the charging process of the drone, enhance the ability to capture short-term data fluctuations, and ensure the prediction accuracy of the model.

[0051] Preferably, in steps S4 and S42, the multi-scale representation and contextual latent representation Through the full connection layer, they are integrated into the same dimension, and the sparse multi-head attention fusion is used to obtain the fusion feature representation. The specific formula is as follows:

[0052] ;

[0053] In the formula, For sparse multi-head attention, is the contextual latent representation, For multi-scale representation, are the learnable parameters of the sparse multi-head attention mechanism.

[0054] Preferably, the multi-scale representation is fused with the contextual latent representation through a sparse multi-head attention mechanism to obtain a fused feature representation. The sparse multi-head attention mechanism enables the model to focus on important features and ignore irrelevant noise data, thereby improving the efficiency and accuracy of feature fusion and enhancing the ability to process complex data. This mechanism can effectively process the relationship between high-dimensional features and enhance the model's ability to express complex features.

[0055] Preferably, in the steps S4 and S43, radial flow is used to approximate the probability density of the wireless charging data of the drone hangar. First, radial flow transformation is performed. The specific formula is as follows:

[0056] ;

[0057] ;

[0058] In the formula, is the front after radial flow transformation The output of dimensions, which constitutes part of the predicted value, is sampled from a standard normal distribution The latent vector of dimension After radial flow transformation The output of the dimension constitutes another part of the predicted value, is sampled from a standard normal distribution The latent vector of dimension and are learnable parameters in radial flow, is the fusion feature representation, is a learnable vector of the same dimension as the input variable latent vector, For point With reference point The Euclidean distance between them is then used to calculate the corresponding density. The specific formula is as follows:

[0059] ;

[0060] In the formula, To input potential variables Given a fusion feature representation The logarithmic probability density under is a latent vector sampled from a standard normal distribution, is the fusion feature representation, for The Euclidean distance of and are learnable parameters in radial flow.

[0061] Preferably, radial flow is introduced to approximate the probability density of drone charging data. The radial flow transformation can better capture the non-Gaussian distribution characteristics of the data, so that the model performs better when processing nonlinear and non-stationary data, enhances the reliability of the prediction, and the model can better fit the distribution characteristics of the data and improve the prediction ability of future time series.

[0062] Preferably, in step S5, a prediction decoding layer is constructed, and the conditional probability distribution of future drone hangar wireless charging data is obtained by combining historical drone hangar wireless charging data, covariates and model learning parameters, so as to complete multi-step prediction. Specifically, a gated neural network is used as a decoder to perform autoregressive decoding and realize multi-step prediction. The specific formula is as follows:

[0063] ;

[0064] In the formula, For wireless charging data of drone hangars given historical , covariates Future drone hangar wireless charging data under conditions Probability density, , , , are the learnable parameters of gated recurrent unit, sparse multi-head attention, and radial flow, respectively. From the time step arrive Prediction value of wireless charging data in drone hangars, for Historical drone hangar wireless charging data for time steps, for The covariate of the historical data of wireless charging in the drone hangar for time steps, is the fusion feature representation, is the probability density function.

[0065] Preferably, a gated neural network is used as a decoder, combined with historical charging data and covariates, to perform autoregressive decoding to achieve multi-step prediction. Autoregressive decoding enables the model to rely on historical data for continuous prediction, enhances the ability to capture long-term dependencies, and is able to adjust and optimize the prediction results in each step of prediction. The model can dynamically adjust the prediction results during the decoding process, taking into account the impact of historical data on the future.

[0066] In summary, due to the adoption of the technical solution, the beneficial effects of the present invention are as follows: the present invention proposes an MDPM prediction model, which is applied to the prediction scenario of wireless charging efficiency of unmanned aerial vehicle hangars, including a multi-scale feature extraction module, a dynamic pattern fusion module and a prediction decoding layer. The multi-scale feature extraction module is used to process the long-term dependencies and complex multi-period characteristics in the wireless charging data of unmanned aerial vehicle hangars. The dynamic pattern fusion module is used to process the combination of short-term dynamic changes and long-term historical patterns in the wireless charging data of unmanned aerial vehicle hangars. The prediction decoding layer enables the model to make multi-step predictions. Each module has good effects in capturing multi-periodicity, processing complex nonlinear wireless charging data of unmanned aerial vehicle hangars, and integrating multiple features for prediction, thereby improving the accuracy of model prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A step diagram of a method for predicting wireless charging efficiency in a drone hangar.

[0068] Figure 2 This is the structural diagram of the MDPM prediction model.

[0069] Figure 3 This is the structural diagram of the multi-scale feature extraction module.

[0070] Figure 4 This is the structural diagram of the dynamic mode fusion module.

[0071] Figure 5 The fitting effect diagram of the MDPM prediction model for the wireless charging efficiency prediction of the drone hangar. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0073] See also Figure 1-Figure 5 The present invention provides a technical solution: a method for predicting the wireless charging efficiency of a drone hangar, which can process the long-term dependence and complex multi-period characteristics in the wireless charging data of the drone hangar by constructing a multi-scale feature extraction module, and the dynamic pattern fusion module can effectively combine the short-term dynamic changes and the long-term historical patterns. The prediction decoding layer combines the historical data, covariates and model learning parameters to obtain the conditional probability distribution of the future wireless charging data of the drone hangar, thereby completing the accurate prediction of the future charging efficiency. The specific steps are as follows: Figure 1 shown.

[0074] Construct the MDPM prediction model, whose structure is as follows Figure 2 As shown, the specific steps are as follows:

[0075] S1. Collect relevant data of wireless charging of drone hangar through sensors, including environmental data of wireless charging of drone hangar and parameter data of wireless charging equipment of drone hangar itself.

[0076] Furthermore, the environmental data of the drone hangar during wireless charging and the parameter data of the drone hangar wireless charging equipment itself include temperature data, humidity data, internal and external wind speed data, and electromagnetic environment data in the hangar. The parameter data of the hangar wireless charging equipment includes charging power data, charging efficiency data, voltage data, current data, equipment temperature data, and equipment withstand voltage data.

[0077] S2. Preprocess the data related to wireless charging in the drone hangar, including data outlier processing and normalization, and divide the data set into a ratio of 7:3.

[0078] Furthermore, the data related to wireless charging in the drone hangar is preprocessed, including data outlier processing and normalization, and the data set is divided proportionally. The specific formula for outlier processing is as follows:

[0079] ;

[0080] In the formula, Wireless charging data points for drone hangars, The average value of wireless charging data in the drone hangar. The standard deviation of the wireless charging data in the drone hangar, is the standardized deviation.

[0081] S31. Input the wireless charging sequence data of the drone hangar ,in is the time step, The feature number of the wireless charging data in the drone hangar is randomly selected, and two overlapping subsequences are mapped to the potential representation through a fully connected layer. , using time series masking for data augmentation.

[0082] Furthermore, a multi-scale feature extraction module is constructed, whose structure is as follows: Figure 3 As shown, input the wireless charging sequence data of the drone hangar ,in is the time step, The number of features of the wireless charging data in the drone hangar, randomly selecting two overlapping subsequences and ,in , and are the starting and ending indexes of the first subsequence, and are the start and end indexes of the second subsequence, mapped to the potential representation through a fully connected layer , and use time series masking for data enhancement. The specific formula is:

[0083] ;

[0084] In the formula, is the potential representation after masking by time series stamps, is the fully connected layer, is the input drone hangar wireless charging subsequence data, is a binary masking matrix, It is an element-by-element multiplication operation.

[0085] S32, Extracting contextual feature representation of drone hangar wireless charging data using multi-layer time domain convolutional network , and then extract the time domain representation of the drone hangar wireless charging data through a set of 1D dilated convolutional layers and random pooling operations .

[0086] Furthermore, the input UAV hangar wireless charging sequence data , using a multi-layer temporal convolutional network to extract contextual feature representation of drone hangar wireless charging data , the specific formula is as follows:

[0087] ;

[0088] In the formula, is the ReLU activation function, is the layer index of the time domain convolutional network, is the total number of layers of the time domain convolutional network, is the index of the convolution kernel, is the size of the convolution kernel, For the In the convolution operation, The convolution kernel weights at each position, For the Layer at time step The output at For the The dilation factor of the layer, For the The bias term of the layer, For the Layer at time step The output of is used to extract the time domain representation of the drone hangar wireless charging data through a set of 1D dilated convolutional layers and random pooling operations. The specific formula is as follows:

[0089] ;

[0090] ;

[0091] In the formula, is a 1D dilated convolution operation, Contextual feature representation for drone hangar wireless charging data, is the size of the convolution kernel, is the expansion rate, For the The output of the dilated convolution layer, is a random pooling operation, The time domain representation of the wireless charging data in the drone hangar. is a hyperparameter that determines the number of dilated convolutional layers used in the network.

[0092] S33, context features are represented by short-time Fourier transform and fully connected layers Convert to frequency domain to get frequency domain features , proposed frequency feature attention , calculate the attention score and get the weighted frequency domain representation , the time domain is represented by and frequency domain representation Fusion to obtain multi-scale representation .

[0093] Furthermore, the context features are represented by short-time Fourier transform and fully connected layers. Switch to the frequency domain to obtain the frequency domain features. The specific formula is as follows:

[0094] ;

[0095] In the formula, is the fully connected layer, is the short-time Fourier transform operation, Contextual feature representation for drone hangar wireless charging data, is the length of the sliding window in the short-time Fourier transform , proposed frequency feature attention , for frequency domain features Application, first extract the amplitude of the frequency component, the specific formula is as follows:

[0096] ;

[0097] In the formula, and are the real and imaginary parts of the frequency domain features, respectively. is the frequency domain feature, and then the multi-head spectrum scaling mechanism is used to scale the query vector and key vector Projection, value vector It is generated by linear mapping, and the specific formula is as follows:

[0098] ;

[0099] ;

[0100] ;

[0101] In the formula, and It is a multi-head spectrum scaling mechanism used to generate query vectors and key vectors. is the amplitude of the frequency component, is a linear mapping function, The input is the wireless charging sequence data of the drone hangar. The multi-head spectrum scaling mechanism captures the correlation between different frequencies by scaling the frequency components. The specific formula is as follows:

[0102] ;

[0103] In the formula, is the amplitude of the frequency component, is a learnable scaling matrix, is element-by-element multiplication, representing and The corresponding elements are multiplied one by one, and then the attention score is calculated to obtain the weighted frequency domain representation. The specific formula is as follows:

[0104] ;

[0105] In the formula, is the activation function, is the query vector, is the transpose of the key vector, is the key vector The dimension of As a value vector, the time domain is finally represented and frequency domain representation Fusion to obtain multi-scale representation .

[0106] S41. Encoding short-term drone hangar wireless charging data through a gated neural network to obtain contextual latent representation .

[0107] Furthermore, a dynamic mode fusion module is constructed, whose structure is as follows: Figure 4 As shown in the figure, a gated neural network is used to encode the short-term UAV hangar wireless charging data to generate a contextual latent representation, which can capture the change pattern of the UAV hangar wireless charging data in the short term. The specific formula is as follows:

[0108] ;

[0109] In the formula, is a gated recurrent unit, is the input short-term UAV hangar wireless charging data, representing the time step To time step The data between is the input short-term UAV hangar wireless charging covariate data, representing the time step To time step The covariate data between for The learnable parameters of .

[0110] S42, multi-scale representation and contextual latent representation Mapped to the same dimension through the fully connected layer and fused using sparse multi-head attention to obtain the fused feature representation .

[0111] Furthermore, the multi-scale representation and contextual latent representation Through the full connection layer, they are integrated into the same dimension, and the sparse multi-head attention fusion is used to obtain the fusion feature representation. The specific formula is as follows:

[0112] ;

[0113] In the formula, For sparse multi-head attention, is the contextual latent representation, For multi-scale representation, are the learnable parameters of the sparse multi-head attention mechanism.

[0114] S43. Using radial flow to approximate the probability density of wireless charging data in drone hangars .

[0115] Furthermore, radial flow is used to approximate the probability density of wireless charging data in the drone hangar. First, radial flow transformation is performed. The specific formula is as follows:

[0116] ;

[0117] ;

[0118] In the formula, is the front after radial flow transformation The output of dimensions, which constitutes part of the predicted value, is sampled from a standard normal distribution The latent vector of dimension After radial flow transformation The output of the dimension constitutes another part of the predicted value, is sampled from a standard normal distribution The latent vector of dimension and are learnable parameters in radial flow, is the fusion feature representation, is a learnable vector of the same dimension as the input variable latent vector, For point With reference point The Euclidean distance between them is then used to calculate the corresponding density. The specific formula is as follows:

[0119] ;

[0120] In the formula, To input potential variables Given a fusion feature representation The logarithmic probability density under is a latent vector sampled from a standard normal distribution, is the fusion feature representation, for The Euclidean distance of and are learnable parameters in radial flow.

[0121] S5. Construct a prediction decoding layer, use the gated neural network as a decoder, perform autoregressive decoding and implement multi-step prediction.

[0122] Furthermore, by combining historical drone hangar wireless charging data, covariates and model learning parameters, the conditional probability distribution of future drone hangar wireless charging data is obtained, thereby completing multi-step prediction. Specifically, the gated neural network is used as a decoder to perform autoregressive decoding and realize multi-step prediction. The specific formula is as follows:

[0123] ;

[0124] In the formula, For wireless charging data of drone hangars given historical , covariates Future drone hangar wireless charging data under conditions Probability density, , , , are the learnable parameters of gated recurrent unit, sparse multi-head attention, and radial flow, respectively. From the time step arrive Prediction value of wireless charging data in drone hangars, for Historical drone hangar wireless charging data for time steps, for The covariate of the historical data of wireless charging in the drone hangar for time steps, is the fusion feature representation, is the probability density function.

[0125] Furthermore, the MDPM prediction model is written in Python. The experiment runs on the Linux operating system. Pytorch is selected as the framework in the CUDA11.27 environment. The training is performed on the NVIDIA A800 80GB GPU. The data set is 9 months of wireless charging data of drone hangars. After data preprocessing, it is input into the MDPM prediction model. The training batch is set to 128. Optimizer, with the learning rate set to .

[0126] Furthermore, the MDPM prediction model realizes the prediction fitting effect of wireless charging efficiency of drone hangar as shown in the figure below. Figure 5 As shown in the figure, the horizontal axis is the date, the vertical axis is the charging efficiency (percentage), the black solid line and dots represent the actual charging efficiency, and the gray dotted line and cross represent the predicted charging efficiency. It can be seen from the figure that the changing trends of the actual value and the predicted value are roughly similar, especially in the peak and trough parts, the changing directions of the two are consistent. The experimental results show that the MDPM model can effectively capture the changing trend of the wireless charging efficiency of the drone hangar, especially in the time period with large fluctuations, it can better predict the charging efficiency.

Claims

1. A method for predicting wireless charging efficiency of a drone hangar, characterized in that: The following steps are involved: S1. Collect relevant data of wireless charging of drone hangar through sensors, including environmental data of wireless charging of drone hangar and parameter data of wireless charging equipment of drone hangar; S2, preprocessing the data related to wireless charging in the drone hangar, including data outlier processing and normalization, and dividing the data set; S3. Construct a multi-scale feature extraction module to handle the long-term dependencies and complex multi-cycle characteristics in the wireless charging data of the drone hangar. The specific steps are as follows: S31, input the wireless charging sequence data y∈R of the drone hangar T×F , where T is the time step, F is the number of features of the drone hangar wireless charging data, and two overlapping subsequences are randomly selected and mapped to the potential representation z through a fully connected layer. t , using time series masking for data augmentation; S32, Extracting contextual feature representation of drone hangar wireless charging data using multi-layer time domain convolutional network Then, a set of 1D dilated convolutional layers and random pooling operations are used to extract the time domain representation of the drone hangar wireless charging data. S33, context features are represented by short-time Fourier transform and fully connected layers Convert to frequency domain to get frequency domain features Propose frequency feature attention FCatte n , and obtain the weighted frequency domain representation Representing the time domain and frequency domain representation The fusion obtains the multi-scale representation rt; S4. Build a dynamic pattern fusion module to process the combination of short-term dynamic changes and long-term historical patterns in the wireless charging data of the drone hangar. The specific steps are as follows: S41. Encode the short-term UAV hangar wireless charging data through a gated neural network to obtain the contextual latent representation C t ; S42, multi-scale representation r t and the contextual latent representation c t Through the fully connected layer mapping and sparse multi-head attention fusion, the fused feature representation h is obtained t ; S43. Using radial flow to approximate the probability density log p(y|h t ); S5. Construct a prediction decoding layer, use the gated neural network as a decoder, perform autoregressive decoding and implement multi-step prediction.

2. The method for predicting the wireless charging efficiency of a drone hangar according to claim 1 is characterized in that: In the step s31, the wireless charging data y ∈ R of the drone hangar is input T×F , where T is the time step and F is the number of features of the wireless charging data of the drone hangar. Two overlapping subsequences Segment1 = y[a1, a2] and Segment2 = y[b1, b2] are randomly selected, where 0 < a1 < b1 < a2 < b2 ≤ T, a1 and a2 are the start and end indices of the first subsequence, and b1 and b2 are the start and end indices of the second subsequence, and are mapped to the latent representation z through the fully connected layer t , and a data augmentation operation is performed using time series masking. The specific formula is as follows: In the formula, z t is the potential representation after time series stamp masking, MLP is the fully connected layer, y t is the input drone hangar wireless charging subsequence data, m is the binary masking matrix, It is an element-by-element multiplication operation.

3. A method for predicting wireless charging efficiency of a drone hangar according to claim 2, characterized in that: In step S32, the inputted drone hangar wireless charging sequence data y∈R T×F Extracting contextual feature representation of drone hangar wireless charging data using multi-layer time-domain convolutional networks The specific formula is as follows: Where σ is the ReLU activation function, l is the layer index of the time domain convolutional network, L is the total number of layers of the time domain convolutional network, k is the index of the convolution kernel, K is the size of the convolution kernel, is the convolution kernel weight at the kth position in the lth layer convolution operation, is the l-1th layer at time step td (l) The output at .k, where d (l) is the dilation factor of the lth layer, b (l) is the bias term of the lth layer, is the output of the l-1th layer at time step t. The time domain representation of the drone hangar wireless charging data is extracted through a set of 1D dilated convolutional layers and random pooling operations. The specific formula is as follows: Where DilatedCOnv1D is a 1D dilated convolution operation. Contextual feature representation for drone hangar wireless charging data,2 i is the size of the convolution kernel, dilation_rate=d is the dilation rate, is the output of the i-th layer dilated convolution, RandomPool is the random pooling operation, r T is the time domain representation of the wireless charging data in the drone hangar. L is a hyperparameter that determines the number of dilated convolutional layers used in the network.

4. A method for predicting wireless charging efficiency of a drone hangar according to claim 3, characterized in that: In step S33, the context feature representation is obtained by short-time Fourier transform and full connection layer. Switch to the frequency domain to obtain the frequency domain features. The specific formula is as follows: In the formula, MLP is the fully connected layer, STFT is the short-time Fourier transform operation, It is the context feature representation of the wireless charging data of the drone hangar. Window_size = W is the length of the sliding window in the short-time Fourier transform. The frequency feature attention FCatten is proposed to Application, first extract the amplitude of the frequency component, the specific formula is as follows: In the formula, and are the real and imaginary parts of the frequency domain features, respectively. It is a frequency domain feature. Then the query vector Q and the key vector K are projected using the multi-head spectrum scaling mechanism, and the value vector V is generated through a linear mapping operation. The specific formula is as follows: Q=MSS Q (A k ); K=MSS K (A k ); V=Linear V (y); In the formula, MSS Q and MSS K A is a multi-head spectrum scaling mechanism used to generate query vectors and key vectors. k is the amplitude of the frequency component, Linear V is a linear mapping function, y is the input drone hangar wireless charging sequence data, where the multi-head spectrum scaling mechanism captures the correlation between different frequencies by scaling the frequency components. The specific formula is as follows: In the formula, A k is the amplitude of the frequency component, W h is a learnable scaling matrix, is element-by-element multiplication, representing A k and W h The corresponding elements are multiplied one by one, and then the attention score is calculated to obtain the weighted frequency domain representation. The specific formula is as follows: In the formula, softmax is the activation function, Q is the query vector, and K T is the transpose of the key vector, d k is the dimension of the key vector K, V is the value vector, and finally the time domain is represented and frequency domain representation Fusion to obtain multi-scale representation 5. A method for predicting wireless charging efficiency of a drone hangar according to claim 4, characterized in that: In step S41, a gated neural network is used to encode the short-term UAV hangar wireless charging data to generate a contextual latent representation, which can capture the change pattern of the UAV hangar wireless charging data in a short period of time. The specific formula is as follows: c t =GRU(y t-L：t ,x t-L：t ;i); In the formula, GRU is the gated recurrent unit, y t-L：t is the input short-term wireless charging data of the drone hangar, representing the data from time step tL to time step t, x t-L：t is the input short-term UAV hangar wireless charging covariate data, represents the covariate data from time step tL to time step t, and θ is the learnable parameter of GRU.

6. A method for predicting wireless charging efficiency of a drone hangar according to claim 5, characterized in that: In step S42, the multi-scale representation r t and the contextual latent representation c t Through the full connection layer, they are integrated into the same dimension, and the sparse multi-head attention fusion is used to obtain the fusion feature representation. The specific formula is as follows: h t =SparseAttention(c t ,r t ; φ); In the formula, SparseAttention is sparse multi-head attention, c t is the contextual latent representation, r t is a multi-scale representation and φ is a learnable parameter of the sparse multi-head attention mechanism.

7. A method for predicting wireless charging efficiency of a drone hangar according to claim 6, characterized in that: In step S43, radial flow is used to approximate the probability density of the wireless charging data of the drone hangar. First, radial flow transformation is performed. The specific formula is as follows: In the formula, is the output of the first d dimensions after the radial flow transformation, which constitutes part of the predicted value, z 1：d is a latent vector sampled from the first d dimensions of a standard normal distribution, is the output of Dd dimensions after radial flow transformation, which constitutes another part of the predicted value, z d+1：D is a latent vector of Dd dimensions sampled from a standard normal distribution, β and α are learnable parameters in radial flow, and h t is the fusion feature representation, u is a learnable vector with the same dimension as the input variable latent vector, ||z d+1：D -z0|| is point z d+1：D The Euclidean distance between the reference point z0 and the reference point z0 is calculated, and then the corresponding density is calculated. The specific formula is as follows: In the formula, logp(z|h t ) is the input latent variable z in the given fusion feature representation h t The logarithmic probability density under , z is the potential vector sampled from the standard normal distribution, h t is the fusion feature representation, r is ||z d+1：D -z0||, β and α are learnable parameters in radial flow, and h is the radial transformation function in radial flow.

8. A method for predicting wireless charging efficiency of a drone hangar according to claim 7, characterized in that: In step S5, a prediction decoding layer is constructed, and the conditional probability distribution of future wireless charging data of drone hangars is obtained by combining historical drone hangar wireless charging data, covariates and model learning parameters, so as to complete multi-step prediction. Specifically, a gated neural network is used as a decoder to perform autoregressive decoding and realize multi-step prediction. The specific formula is as follows: In the formula, Given the historical drone hangar wireless charging data y t-L：t , covariate x t-L：t Future drone hangar wireless charging data under conditions Y t：T The probability density of θ, φ, α, and β are the learnable parameters of gated recurrent unit, sparse multi-head attention, and radial flow, respectively. t：T is the predicted value of the wireless charging data of the drone hangar from time step t to T, y t-L：t is the historical drone hangar wireless charging data of L time steps, x t-L：t is the covariate of the wireless charging history data of the drone hangar for L time steps, h t is the fusion feature representation, and p is the probability density function.

9. The method for predicting wireless charging efficiency of a drone hangar according to claim 8 is characterized in that: In order to predict the wireless charging efficiency of drone hangar, the relevant data collected by the hangar wireless charging equipment are used, including hangar temperature data, humidity data, internal and external wind speed data, electromagnetic environment data, charging power data, charging efficiency data, voltage data, current data, equipment temperature data, and equipment withstand voltage data. The collected relevant data are preprocessed to ensure data quality, and then the processed data are divided into training sets and test sets for training and evaluating the performance of the drone hangar wireless charging efficiency prediction model.

Images