Fixed-wing unmanned aerial vehicle time series state prediction method and system fusing multi-scale embedding and grouped channel attention, and medium

Abstract

The application discloses a fixed-wing unmanned aerial vehicle time sequence state prediction method and system fusing multi-scale embedding and grouping channel attention, and a medium, comprising: preprocessing multi-dimensional time sequence state data of a collected fixed-wing unmanned aerial vehicle propulsion system; adopting a multi-scale embedding layer to map the preprocessed multi-dimensional time sequence state data into feature representations of different time scales, and fusing to form multi-time scale fused feature representations; inputting the multi-time scale fused feature representations into a preset backbone network, performing weighted processing on time sequence features through a grouping attention mechanism and residual fusion, and outputting multi-scale time sequence features; and performing feature flattening and linear transformation on the multi-scale time sequence features through a mapping layer to obtain a prediction value of the fixed-wing unmanned aerial vehicle time sequence state. The application widens the breadth of time sequence representation through multi-scale embedding, realizes adaptive filtering and enhancement of feature channels through grouping attention, and maintains optimization stability through residual fusion.

Description

Technical Field

[0001] This application relates to the field of UAV prediction technology, and more specifically, to a method, system, and medium for predicting the temporal state of a fixed-wing UAV by integrating multi-scale embedding and grouped channel attention. Background Technology

[0002] With the widespread application of unmanned aerial vehicle (UAV) technology, fixed-wing UAVs play a crucial role in long-endurance flight missions due to their advantages such as long range, high speed, and strong payload capacity. To ensure the long-term stable flight of these UAVs under complex operating conditions, accurate prediction of the critical states of their propulsion systems (such as motor bus voltage, winding temperature, and power device temperature) has become a key link in achieving fault early warning and building an autonomous health management system. Accurate state prediction can provide a basis for early degradation identification, avoiding performance degradation or safety accidents caused by missed detections, while also reducing unnecessary mission interruptions caused by false alarms, thereby improving flight safety and mission economy.

[0003] Currently, methods for UAV state prediction are mainly divided into two categories: model-based methods and data-driven methods. Model-based methods rely on prior knowledge of the system's physics or dynamics to construct state observers or filters, achieving high-precision state estimation when the model is precisely known. However, fixed-wing UAV propulsion systems are complex systems with strong coupling of multiple physical quantities and significant nonlinear characteristics, making it difficult to establish accurate analytical models. This results in limited generalization ability of such methods in practical applications and a high dependence on modeling accuracy. Data-driven methods, on the other hand, do not rely on accurate physical models but directly learn the temporal patterns of state evolution from historical flight data. Recurrent neural networks and their variants (such as LSTM) are widely used due to their ability to process sequential data. In addition, some studies have introduced attention mechanisms or combined convolutional modules to enhance the model's ability to capture key features, thereby improving prediction performance.

[0004] Nevertheless, existing data-driven prediction methods still have significant limitations: on the one hand, most methods have a relatively singular modeling of the time dimension in the feature extraction stage, making it difficult to simultaneously and effectively capture the multi-time constant characteristics (i.e., short-term rapid response and long-term slow variation trend coexist) exhibited by the electrical and thermodynamic parameters in the propulsion system of fixed-wing UAVs; on the other hand, when processing high-dimensional multivariate time series data, they often fail to fully explore the coupling relationships and differences in feature importance within and between variables, resulting in insufficient attention allocation to key state features and susceptibility to noise interference. Summary of the Invention

[0005] To address at least one deficiency or improvement requirement of existing technologies, this invention provides a method, system, and medium for predicting the temporal state of a fixed-wing UAV by integrating multi-scale embedding and grouped channel attention. This addresses the problems in existing technologies where the modeling of the time dimension is relatively singular, making it difficult to simultaneously and effectively capture the multi-time constant characteristics of the fixed-wing UAV propulsion system. Furthermore, when processing high-dimensional multivariate time-series data, it fails to fully explore the coupling relationships and differences in feature importance within and between variables, resulting in insufficient attention allocation to key state features and susceptibility to noise channel interference.

[0006] To achieve the above objectives, according to a first aspect of the present invention, a method for predicting the temporal state of a fixed-wing unmanned aerial vehicle (UAV) by integrating multi-scale embedding and grouped channel attention is provided, comprising:

[0007] Preprocess the multi-dimensional time-series state data of the collected fixed-wing UAV propulsion system;

[0008] A multi-scale embedding layer is used to map the preprocessed multi-dimensional temporal state data into feature representations at different time scales, and then the representations are fused to form a multi-time-scale fused feature representation.

[0009] The feature representation fused across multiple time scales is input into a pre-defined backbone network. The temporal features are weighted through a group attention mechanism and residual fusion to output multi-scale temporal features.

[0010] By performing feature flattening and linear transformation on multi-scale temporal features through a mapping layer, the predicted values ​​of the temporal state of a fixed-wing UAV are obtained.

[0011] In one possible implementation, a multi-scale embedding layer is used to map the preprocessed multi-dimensional temporal state data into feature representations at different time scales, and these representations are then fused to form a multi-time-scale fused feature representation. This also includes:

[0012] By using several one-dimensional convolution kernels with different preset time window lengths, parallel sliding convolution processing is performed on the preprocessed multi-dimensional temporal state data to obtain feature representations at different time scales;

[0013] Feature representations at different time scales are spliced ​​and fused along the feature channel dimension to form a multi-time scale fused feature representation.

[0014] In one possible implementation, the feature representation fused across multiple time scales is input into a pre-defined backbone network. The temporal features are then weighted using a grouped attention mechanism and residual fusion to output multi-scale temporal features. This also includes:

[0015] Perform temporal-dimensional convolution processing on the feature representation fused from multiple time scales;

[0016] The convolutional feature representations are grouped according to the variable dimension, and the first nonlinear transformation and the first channel attention weighting operation are performed on the features within each group;

[0017] The features after the first channel attention weighting operation are grouped according to the feature dimension, and the second nonlinear transformation and the second channel attention weighting operation are performed on the features within each group.

[0018] The features after the second-channel attention weighting operation are fused with the feature representation fused across multiple time scales to output multi-scale temporal features.

[0019] In one possible implementation, both the first-channel attention weighting operation and the second-channel attention weighting operation include:

[0020] Global statistics are performed on the features within the group along the time dimension to obtain the group channel description vector;

[0021] The grouped channel description vectors are processed through a fully connected layer and a non-linear activation function to generate scaling weights for each feature channel within the group.

[0022] The weighting operation is achieved by scaling the features within the corresponding group one channel at a time using scaling weights.

[0023] In one possible implementation, both the first and second nonlinear transformations are implemented through grouped convolution operations, where the number of groups in the grouped convolution operation is the same as the number of groups in the corresponding variable dimension and feature dimension.

[0024] In one possible implementation, the predicted values ​​of the time-series state of the fixed-wing UAV are obtained by flattening and linearly transforming the multi-scale temporal features through a mapping layer, and also include:

[0025] Multi-scale temporal features are flattened and stitched together in both the variable and time dimensions to form a one-dimensional feature vector.

[0026] By mapping a one-dimensional feature vector to a space with the same dimension as the target state through linear transformation, the predicted value of the time-series state of the fixed-wing UAV is obtained.

[0027] In one possible implementation, preprocessing the collected multi-dimensional time-series state data of the fixed-wing UAV propulsion system also includes:

[0028] Anomaly detection is performed on multi-dimensional time-series state data based on a preset statistical distribution interval. The identified outliers are marked as missing values ​​and filled using interpolation.

[0029] A standard scaler is used to standardize the padded complete time-series state data sequence.

[0030] According to a second aspect of the present invention, a time-series state prediction system for a fixed-wing unmanned aerial vehicle that integrates multi-scale embedding and grouped channel attention is also provided, comprising:

[0031] The preprocessing module is configured to preprocess the multi-dimensional time-series state data of the acquired fixed-wing UAV propulsion system;

[0032] The mapping and fusion module is configured to use a multi-scale embedding layer to map the preprocessed multi-dimensional temporal state data into feature representations at different time scales, and then fuse them to form a multi-time scale fused feature representation.

[0033] The group weighting module is configured to input the feature representation fused from multiple time scales into a preset backbone network, and to weight the temporal features through a group attention mechanism and residual fusion to output multi-scale temporal features.

[0034] The state prediction module is configured to flatten and linearly transform multi-scale temporal features through a mapping layer to obtain predicted values ​​of the temporal state of a fixed-wing UAV.

[0035] According to a third aspect of the present invention, a fixed-wing UAV timing state prediction device integrating multi-scale embedding and grouped channel attention is also provided, comprising at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the computer program is executed by the processing unit, the processing unit performs the steps of any of the above-described fixed-wing UAV timing state prediction methods integrating multi-scale embedding and grouped channel attention.

[0036] According to a fourth aspect of the present invention, a storage medium is also provided, which stores a computer program executable by a fixed-wing UAV timing state prediction device that integrates multi-scale embedding and grouped channel attention. When the computer program is run on the fixed-wing UAV timing state prediction device that integrates multi-scale embedding and grouped channel attention, the device performs the steps of any of the above-described fixed-wing UAV timing state prediction methods that integrate multi-scale embedding and grouped channel attention.

[0037] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0038] This invention provides a time-series state prediction method for fixed-wing UAVs that integrates multi-scale embedding and grouped channel attention. By introducing a multi-scale embedding layer, it employs convolutional kernels with different time windows for parallel processing in the initial stage of feature extraction. This allows for the simultaneous acquisition of short-term fluctuations and long-term evolution information in the time series, effectively characterizing the rapid response characteristics of electrical parameters such as voltage and current, and the slow accumulation characteristics of thermal parameters such as temperature in the fixed-wing UAV propulsion system. This overcomes the limitation of single-time-scale modeling, which cannot adapt to physical processes with multiple time constants. The introduction of the grouped attention mechanism achieves non-uniform, adaptive weighting of feature channels, dynamically identifying and strengthening feature subsets (channel groups) that are more important to the current prediction task, while suppressing secondary or interfering information. Residual fusion ensures that the original multi-scale information flow is preserved during deep feature transformation, enhancing gradient flow and training stability, and effectively optimizing the complex multi-scale and grouped attention structure. This invention constructs an end-to-end feature learning mode from multi-scale initialization, temporal convolution, group attention weighting and residual fusion. Its output multi-scale temporal features are adaptively selected and enhanced deep representations, which are finally transformed into accurate state prediction values ​​through a mapping layer. This improves the ability to capture multi-rate dynamics and identify key features, thereby improving the accuracy and robustness of state prediction under complex working conditions of fixed-wing UAVs. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating an embodiment of the fixed-wing UAV temporal state prediction method that integrates multi-scale embedding and grouped channel attention provided by the present invention.

[0041] Figure 2 This is a schematic diagram of the structure of an embodiment of the temporal state prediction model that integrates multi-scale embedding and grouped channel attention provided by the present invention.

[0042] Figure 3 A schematic diagram illustrating the construction of an embodiment of the sample structure provided by the present invention;

[0043] Figure 4 A schematic diagram of a structure of an embodiment of the fixed-wing UAV temporal state prediction system that integrates multi-scale embedding and grouped channel attention provided by the present invention;

[0044] Figure 5This is a schematic diagram of the structure of a fixed-wing UAV time-series state prediction device that integrates multi-scale embedding and grouped channel attention, as provided in an embodiment of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0046] The terms "first," "second," "third," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0047] This invention provides a method, system, and medium for predicting the temporal state of a fixed-wing UAV by integrating multi-scale embedding and grouped channel attention, which will be described below.

[0048] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the fixed-wing UAV temporal state prediction method integrating multi-scale embedding and grouped channel attention provided by the present invention. In a specific embodiment of the present invention, a fixed-wing UAV temporal state prediction method integrating multi-scale embedding and grouped channel attention is disclosed, including:

[0049] S101. Preprocess the multi-dimensional time-series state data of the collected fixed-wing UAV propulsion system;

[0050] S102. A multi-scale embedding layer is used to map the preprocessed multi-dimensional temporal state data into feature representations at different time scales, and then the data is fused to form a multi-time scale fused feature representation.

[0051] S103. Input the feature representation fused from multiple time scales into the preset backbone network, and perform weighted processing on the temporal features through group attention mechanism and residual fusion to output multi-scale temporal features.

[0052] S104. The multi-scale temporal features are flattened and linearly transformed by the mapping layer to obtain the predicted values ​​of the temporal state of the fixed-wing UAV.

[0053] In the above embodiments, the preprocessing process typically includes: anomaly detection and cleaning, such as identifying and processing abnormal sampling points that deviate significantly from the normal range based on statistical methods (e.g., interquartile range); reasonable imputation of missing values ​​caused by temporary sensor failure or communication interruption to maintain the continuity of the time series; and data standardization or normalization, transforming state variables with different physical dimensions (e.g., voltage in volts, temperature in degrees Celsius) and large differences in numerical range to a similar numerical range (e.g., a distribution with a mean of 0 and a variance of 1) to eliminate the influence of dimensions, accelerate model training convergence, and improve numerical stability.

[0054] This invention establishes a temporal state prediction model (MS-GSE-ModernTCN) that integrates multi-scale embedding and grouped channel attention to achieve prediction. The multi-scale embedding layer typically consists of multiple parallel one-dimensional convolutional branches, each using a convolutional kernel of different widths (i.e., temporal receptive field sizes). For example, narrow convolutional kernels (capturing short-term transient fluctuations), medium-width kernels (capturing medium-term change patterns), and wide kernels (capturing long-term trends) can be used simultaneously. Each branch independently performs sliding convolution on the input multivariate temporal data, transforming the time series of each variable into a new set of feature maps. The feature maps output by all branches, which have different temporal granularities but are spatially aligned, are concatenated along the feature channel dimension to form a unified feature representation that integrates multi-scale temporal information. This allows for the simultaneous perception and utilization of fast and slow-changing components in the data, facilitating the characterization of the multi-time constant characteristics of fixed-wing UAV propulsion systems.

[0055] The pre-defined backbone network (the MS-GSE-ModernTCN Block within MS-GSE-ModernTCN) consists of multiple cascaded, structurally identical processing blocks stacked together. Within each processing block, the following sub-operations are executed sequentially: First, deep convolutions are performed on the input features along the temporal dimension to further integrate cross-temporal contextual dependencies and expand temporal modeling capabilities. Subsequently, a grouped channel attention mechanism is introduced: This mechanism first groups the feature tensors according to a certain dimension (e.g., variable index or feature type), and then within each independent group, calculates the global statistics (e.g., global average) of all feature channels in that group along the temporal dimension. Based on this, a set of channel attention weights is dynamically generated through a small neural network (typically containing fully connected layers and non-linear activation functions). These weights are used to adaptively scale (recalibrate) the individual feature channels within the group, thereby enhancing important features and suppressing minor or noisy features. This grouping design allows the model to independently learn the attention distribution within different semantic groups, enabling more refined feature selection. Residual fusion refers to the process of adding the features processed by convolution and attention weighting to the original input features of each processing module (Shortcut Connection). This helps alleviate the vanishing or exploding gradient problem in deep network training, ensuring that information and gradients can be effectively propagated, allowing the network to be built deeper and more powerful. After layer-by-layer processing by multiple such modules, the final output is multi-scale temporal features that have undergone deep abstraction and adaptive enhancement.

[0056] Feature flattening reshapes or concatenates the feature tensors output by the backbone network, which typically retain a multi-dimensional structure (e.g., containing multiple dimensions such as variables, time segments, and channels), into a one-dimensional feature vector for subsequent fully connected processing. Linear transformation is usually implemented through one or more fully connected layers (linear layers), which project the flattened high-dimensional feature vector onto the dimension of the target prediction space. For example, if it is necessary to predict 16 states, such as the bus voltage of four motors and the temperature of the control unit, at a future moment, the output dimension of this linear layer would be 16. The weight parameters of this layer are learned during training, thus establishing a mapping relationship from complex features to specific state values. Ultimately, the output of this linear layer is the predicted value of the time-series state of the fixed-wing UAV, which can be directly used for downstream tasks such as state monitoring and fault early warning.

[0057] Compared with existing technologies, the fixed-wing UAV time-series state prediction method fused with multi-scale embedding and grouped channel attention provided in this embodiment introduces a multi-scale embedding layer. This layer uses convolutional kernels with different time windows for parallel processing from the initial feature extraction stage, thereby simultaneously acquiring short-term fluctuations and long-term evolution information in the time series. This effectively characterizes the rapid response characteristics of electrical parameters such as voltage and current, and the slow accumulation characteristics of thermal parameters such as temperature, in the fixed-wing UAV propulsion system, overcoming the limitation of single-time-scale modeling in adapting to physical processes with multiple time constants. The introduction of the grouped attention mechanism achieves non-uniform, adaptive weighting of feature channels, dynamically identifying and strengthening feature subsets (channel groups) that are more important to the current prediction task, while suppressing secondary or interfering information. Residual fusion ensures that the original multi-scale information flow is preserved during deep feature transformation, enhancing gradient flow and training stability, and effectively optimizing the complex multi-scale and grouped attention structure. This invention constructs an end-to-end feature learning mode from multi-scale initialization, temporal convolution, group attention weighting and residual fusion. Its output multi-scale temporal features are adaptively selected and enhanced deep representations, which are finally transformed into accurate state prediction values ​​through a mapping layer. This improves the ability to capture multi-rate dynamics and identify key features, thereby improving the accuracy and robustness of state prediction under complex working conditions of fixed-wing UAVs.

[0058] In some embodiments of the present invention, a multi-scale embedding layer is used to map the preprocessed multi-dimensional temporal state data into feature representations at different time scales, and these representations are then fused to form a multi-time-scale fused feature representation. The method further includes:

[0059] By using several one-dimensional convolution kernels with different preset time window lengths, parallel sliding convolution processing is performed on the preprocessed multi-dimensional temporal state data to obtain feature representations at different time scales;

[0060] Feature representations at different time scales are spliced ​​and fused along the feature channel dimension to form a multi-time scale fused feature representation.

[0061] In some embodiments of the present invention, the feature representation fused across multiple time scales is input into a preset backbone network, and the temporal features are weighted through a group attention mechanism and residual fusion to output multi-scale temporal features. The method further includes:

[0062] Perform temporal-dimensional convolution processing on the feature representation fused from multiple time scales;

[0063] The convolutional feature representations are grouped according to the variable dimension, and the first nonlinear transformation and the first channel attention weighting operation are performed on the features within each group;

[0064] The features after the first channel attention weighting operation are grouped according to the feature dimension, and the second nonlinear transformation and the second channel attention weighting operation are performed on the features within each group.

[0065] The features after the second-channel attention weighting operation are fused with the feature representation fused across multiple time scales to output multi-scale temporal features.

[0066] In some embodiments of the present invention, both the first channel attention weighting operation and the second channel attention weighting operation include:

[0067] Global statistics are performed on the features within the group along the time dimension to obtain the group channel description vector;

[0068] The grouped channel description vectors are processed through a fully connected layer and a non-linear activation function to generate scaling weights for each feature channel within the group.

[0069] The weighting operation is achieved by scaling the features within the corresponding group one channel at a time using scaling weights.

[0070] In some embodiments of the present invention, both the first nonlinear transformation and the second nonlinear transformation are implemented through grouped convolution operations, and the number of groups in the grouped convolution operation is the same as the number of groups in the corresponding variable dimension and the number of groups in the feature dimension.

[0071] Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of an embodiment of the temporal state prediction model that integrates multi-scale embedding and grouped channel attention provided by the present invention, as shown below. Figure 2 As shown, the architecture includes a multi-scale embedding layer (MSE), a backbone composed of convolutional blocks (MS-GSE-ModernTCN Blocks), and a projection layer that varies according to the prediction task. The multi-scale embedding layer maps the original input sequence into a high-dimensional latent space representation, enabling subsequent networks to capture temporal trends. The backbone consists of several stacked MS-GSE-ModernTCN Blocks, each containing five parts: depthwise convolution (DWConv), grouped feedforward convolutional networks (ConvFFN), grouped channel attention (GroupSE), inter-layer residual connections, and a batch normalization layer (BatchNorm). Each grouped feedforward convolutional network (ConvFFN) and grouped channel attention (GroupSE) requires grouping. Finally, the projection layer flattens the extracted multi-scale temporal features and outputs the predicted value for the target step through a linear layer.

[0072] Model input is ,in B, M and L These represent the batch size, variable dimension (i.e., the number of drone state variables), and time window length, respectively. During the embedding phase, the model uses... Parallel sliding window feature extraction is achieved using three one-dimensional convolutions with kernels of different sizes. P With stride S Controlling the compression ratio of the time dimension, dividing each variable into... N Each variable has a time slice. To maintain the independence of each variable, the convolution operation is only expanded in the time dimension, and the channel directions do not share weights, thus obtaining variable-independent embedding representations. ,in D For feature embedding dimension, N For the number of time slices, .

[0073] The MS-GSE-ModernTCN Block in the backbone is responsible for extracting data features at different semantic levels. DWConv is used to capture feature changes along the time dimension, and its convolutional kernel size is typically set to a value greater than 50 to expand the effective receptivity, thereby enhancing the ability to model long-term dependencies without increasing network depth. After extracting features in the time dimension, BatchNorm batch normalization is performed to maintain consistency in data scale and accelerate training. Next, the model achieves "decoupling-recoupling" of variables and channel dimensions through a grouped feedforward convolutional network and its closely integrated grouped channel attention. ConvFFN1 groups the variables based on the feature dimensions within each variable (groups= M The input performs nonlinear mapping on the multidimensional feature channels of a single variable to achieve nonlinear reconstruction of the intra-variable representation. GroupSE1, with the same grouping configuration, is responsible for redistributing feature weights within the same variable, i.e., learning the relative importance of features within each variable. ConvFFN2, on the other hand, transposes the input along the feature dimension and groups it into multiple variables along the same feature dimension (groups= D The MS-GSE-ModernTCN Block learns the coupling relationships between variables, while GroupSE2, with the same grouping configuration, is responsible for learning the dynamic weights of the same feature dimension across different variables. This hierarchical structure of progressive convolution combined with intervariate convolution and weight recalibration ensures the separate modeling of temporal relationships, intervariate correlations, and feature abstraction, significantly improving the model's learning ability and generalization performance. Simultaneously, to alleviate gradient instability issues during deep training, residual connections are configured in the MS-GSE-ModernTCN Block. After stacking multiple MS-GSE-ModernTCN Blocks, the model obtains multi-scale time-series representations. Then, the mapping layer rearranges the embedded features of each variable to obtain... Finally, the data is mapped to the target output dimension through a linear layer. By structurally introducing MsE and GroupSE, MS-GSE-ModernTCN shifts from uniform modeling to differentiated adaptive modeling, which is more in line with the data characteristics of fixed-wing UAV propulsion systems.

[0074] Specifically:

[0075] The multi-scale embedding layer extracts temporal features across different temporal receptive fields using three parallel one-dimensional convolutions, aiming to enrich the scale representation of temporal features and enable the model to simultaneously capture short- and long-term dynamic changes. The input to this layer is represented as follows: The embedding layer of ModernTCN can be represented as:

[0076] (1)

[0077] in D The feature embedding dimension, and the kernel size. P and stride S Fixed, single-scale operation creates a resolution bottleneck for temporal features, reducing their richness. MsE addresses this by introducing a multi-scale convolutional kernel set. Features are extracted in parallel under different time windows to obtain three embedding representations at different time scales:

[0078] (2)

[0079] (3)

[0080] Next, the outputs of each branch are aligned in the time dimension and then concatenated in the channel direction to form a multi-timescale fused embedded representation. :

[0081] (4)

[0082] Multi-scale embedding can be viewed as an implicit multi-band filtering mechanism. Small convolutional kernels act as high-pass filters in the time domain, emphasizing local differences; large convolutional kernels are similar to low-pass filters, smoothing long-term trends; and mesoscale kernels act as a bridge between the two. This multi-band fusion helps capture the multi-time-constant characteristics of electrical and thermal parameters, enabling the model to simultaneously model both sudden responses and gradual accumulations.

[0083] GroupSE, while maintaining the decoupled structure, assigns differentiated weights to intra-variable and inter-variable features, achieving hierarchical allocation of attention. In the MS-GSE-ModernTCN Block, after processing by ConvFFN1 and ConvFFN2, the channel response strengths of different variables and feature dimensions often differ. The original channel attention (SE) generates uniform weights through global pooling and nonlinear mapping, applying calibration of the same scale to all channels. In multivariate time series prediction scenarios, the feature correlations of different variable groups differ significantly. Temperature features are highly correlated, while voltage features are more sensitive to external disturbances. Uniform weighting may lead to the masking of important features. Therefore, this paper proposes grouped channel attention. GroupSE introduces a grouping mechanism on the basis of SE, that is, the weights of the channels are independently learned in groups, achieving independent calibration within variables and independent modulation between variables, thereby assigning an independent attention distribution to each channel group without destroying the decoupled structure.

[0084] GroupSE i The input of the module is represented as ,in , GroupSE i Divide the channel into There are 1 independent groups, among which , The number of channels in each group can be represented as:

[0085] (5)

[0086] First, analyze the input data. Perform global average pooling along the time dimension to obtain a global statistical vector along the time dimension:

[0087] (6)

[0088] In the Squeeze-Excitation stage, the feature information within the group is compressed and activated. Each channel group independently learns a non-linear weight vector through two pointwise convolutional layers PWConv with a kernel size of 1 (equivalent to two fully connected layers).

[0089] (7)

[0090] The weight matrix , , r The compression ratio is... Represents the ReLU activation function. This represents Sigmoid normalization. Next, using... Perform feature weight recalibration, extend the attention weights to the entire time dimension, and multiply them element-wise with the input data:

[0091] (8)

[0092] The entire process employs an intra-group channel adaptive scaling mechanism, where weights are learned independently by the group to which each channel belongs, preserving the independence and semantic consistency between variables. The introduction of GroupSE can be viewed as adding a channel-weighted filter after the convolutional output, using the learned attention weights... Dynamically adjusting the signal gain of different channels is equivalent to forming an adaptive energy allocation mechanism within the model.

[0093] In some embodiments of the present invention, the predicted values ​​of the temporal state of a fixed-wing UAV are obtained by performing feature flattening and linear transformation on multi-scale temporal features through a mapping layer, and further include:

[0094] Multi-scale temporal features are flattened and stitched together in both the variable and time dimensions to form a one-dimensional feature vector.

[0095] By mapping a one-dimensional feature vector to a space with the same dimension as the target state through linear transformation, the predicted value of the time-series state of the fixed-wing UAV is obtained.

[0096] In the above embodiment, the structured multidimensional feature tensor is converted into a one-dimensional vector. This is not a simple overall flattening, but a logical concatenation. For each sample, the feature channel dimension remains unchanged. Then, the feature representations of all variables across all time segments are concatenated in a specific order (e.g., first iterating through the features of all variables in the first time segment, then the second time segment, and so on; or first iterating through all time segments of the first variable, then the second variable). The variable and time structural dimensions are removed from the feature tensor, and all the information it contains is encoded into the extension of the feature channel dimension, thus forming a long one-dimensional feature vector that integrates the information of all variables at all historical moments after being refined by the model.

[0097] Mapping the aforementioned one-dimensional feature vector to the physical quantity space, the input dimension equals the length of the one-dimensional feature vector obtained in the previous step, and the output dimension equals the dimension of the predicted target state value. For example, if the final goal is to predict four states—bus voltage, control unit temperature, winding temperature, and power device temperature—for four motors at a future moment, totaling 16 specific values, then the output dimension of this linear layer is 16. The weight matrix learned through training essentially establishes a mapping relationship from complex fused features to each specific physical state quantity.

[0098] In some embodiments of the present invention, preprocessing of the collected multi-dimensional time-series state data of the fixed-wing unmanned aerial vehicle propulsion system further includes:

[0099] Anomaly detection is performed on multi-dimensional time-series state data based on a preset statistical distribution interval. The identified outliers are marked as missing values ​​and filled using interpolation.

[0100] A standard scaler is used to standardize the padded complete time-series state data sequence.

[0101] In the above embodiments, the interquartile range (IQR) method is commonly used, based on a preset statistical distribution interval. This method first calculates the first quartile (Q1) and the third quartile (Q3) of each data sequence, then defines the normal value interval as [Q1 - k×IQR, Q3 + k×IQR], where IQR = Q3 - Q1 is the IQR, and k is a preset constant (usually 1.5 or 3). Data points falling outside this interval are identified as outliers.

[0102] After identifying outliers, they are not deleted directly, but first marked as missing values ​​(NaN). Then, interpolation is used to fill them in. Common interpolation methods include linear interpolation or time series interpolation, which uses the values ​​of normal data points adjacent to the outlier to estimate and fill the missing value. This can remove obvious outlier interference while maintaining the continuity and trend of the time series to the greatest extent.

[0103] Because the dimensions and numerical ranges of different state variables (such as voltage, temperature, and current values) in the original data vary greatly, directly inputting them into the model can lead to optimization difficulties (inconsistent gradient scales) and cause the model to become biased towards features with large values. Standardization can eliminate this effect. The standard scaler usually refers to the StandardScaler, whose standardization transformation is performed independently on each feature dimension (i.e., each state variable). For the entire data sequence of a variable, the mean (μ) and standard deviation (σ) of all its samples are calculated. Then, each data point of that variable is transformed: new value = (original value - μ) / σ.

[0104] After this transformation, the data distribution across each feature dimension becomes a standard normal distribution with a mean of 0 and a standard deviation of 1. This ensures all features are on the same numerical scale, accelerating model convergence (especially for gradient descent-based neural networks) and improving numerical stability and generalization performance. In practice, the mean μ and standard deviation σ used to fit the StandardScaler are calculated only from the training set, and the same transformation parameters are applied to the validation and test sets to ensure consistency in data distribution.

[0105] In a specific embodiment of the present invention, the data used comes from the system operation records of a certain type of fixed-wing UAV under a typical cruise flight mission. The system sampling frequency is 1Hz, and data is collected continuously for 9 days, resulting in a total of 806,629 samples. The fixed-wing UAV propulsion system sends the main state parameters of the motor operation to the host computer communication system according to the sampling frequency. Through data framing, the raw time-series data is transmitted to the ground station via the telemetry link to perform state parameter prediction, thereby realizing early warning of propulsion system faults.

[0106] To fully describe the environment in which the propulsion system operates and the status of its key components, the original time-series data encompasses the multi-physical coupling characteristics and environmental parameters of the fixed-wing UAV propulsion system, totaling 72 dimensions, forming typical multi-time-constant time-series data. For each time step, the environment of the UAV is reflected by the indicated airspeed, vacuum speed, ground speed, atmospheric pressure altitude, altitude, and the static temperature of the left and right sides; the operating status of the power supply system is indicated by the voltage of batteries 1#-8#; and the motor status is characterized by the motor number, commanded speed, actual speed, bus voltage, bus current, control unit temperature, winding temperature, power device temperature, rotor position, calculated torque, D-axis current, Q-axis current, motor system status, and communication status word of the four motors (1# to 4#). Finally, a frame counter counts the time steps. Based on expert experience, the established end-to-end mapping focuses on modeling and predicting the key feature information of the four motors (1# to 4#) of the propulsion system from the original 72-dimensional time-series data, including the bus voltage, control unit temperature, winding temperature, and power device temperature, totaling 16 dimensions.

[0107] A unified time index is generated based on the date and time fields output by the data acquisition system. All non-numeric fields are converted to floating-point numbers, and outlier time points are removed. The anomaly detection section employs a statistical discrimination method based on interquartile range. When the sample points exceed... Data within a specified interval is considered an anomaly and marked as missing before interpolation. To mitigate transient disturbances introduced by noise, linear interpolation is used to fill in missing data bidirectionally. Anomaly detection and interpolation iteration are performed again after interpolation to ensure the continuity and smoothness of the sequence. Finally, all numerical columns undergo unified data type optimization, compressing 64-bit floating-point numbers to 32 bits to reduce memory usage.

[0108] In terms of data partitioning, the time-order partitioning principle was strictly followed to ensure that future information was not leaked into the training process. The training set, validation set, and test set were partitioned in an 8:1:1 ratio as shown in Table 1. Subsequently, the training set was standardized and fitted using StandardScaler, and the same transformation was applied to the validation set and test set.

[0109] Specifically as follows:

[0110] Table 1 Dataset Partitioning

[0111]

[0112] Please see Figure 3 , Figure 3 This is a schematic diagram illustrating an embodiment of the sample construction provided by the present invention. The temporal state prediction of a fixed-wing UAV propulsion system is a supervised learning problem, requiring the construction of samples and labels on a pre-divided dataset. The construction method is as follows: Figure 3 In this model, rows represent time steps, and columns represent features of samples or labels. Starting from the first time step, the time window length is set to 96 seconds (time_step=96). All 72-dimensional features are used as the first sample, sample1. The time interval between the time step to be predicted and the last time step of the sample is 10 seconds (horizon=10), and the 16-dimensional propulsion system features are taken as the label, y1. To ensure full utilization of the dataset, the sliding interval S between the sample and label selection boxes is set to 1, thereby constructing samples and corresponding labels in batches.

[0113] In the experimental design, to verify the effectiveness of the proposed model, several typical time series prediction models were selected for comparison, including LSTM, CNN-LSTM, TCN, PatchTST, and DLinear. During model training, the loss function used was MSE, and the optimization algorithm was Adam (learning rate...). The training batch size was set to 128, with 100 training iterations per cycle. The experimental environment configuration is shown in Table 2.

[0114] Table 2 Experimental Environment Configuration

[0115]

[0116] To measure model performance from different perspectives, the experiment used three typical regression metrics: root mean square error (RMSE), mean squared error (MSE), and mean absolute error (MAE).

[38] Among them, RMSE and MSE reflect the energy level of prediction error, while MAE measures the actual magnitude of the mean deviation. The formulas for these three are as follows:

[0117] (9)

[0118] (10)

[0119] (11)

[0120] in, Represents time step i The predicted feature vector, and Represents time step iThe true feature vectors. In the experiments, all metrics were calculated independently on the test set. Smaller MSE, MAE, and RMSE values ​​indicate that the predicted curve is closer to the true value curve, suggesting that the model has higher state prediction accuracy.

[0121] Table 3 Comparison of Model Prediction Accuracy

[0122]

[0123] In the feature column, A represents the complete 16-dimensional features of motors 1# to 4#, while B, C, W, and P represent the bus voltage, control unit temperature, winding temperature, and power device temperature of motors 1# to 4#, respectively.

[0124] To compare prediction accuracy under different flight scenarios during day and night, the daytime and nighttime segments were divided based on two fuzzy transition periods: 05:10:00-06:00:00 and 19:15:00-20:00:00. Three error metrics were calculated for each segment, as detailed in Table 3. In comparison, MS-GSE-ModernTCN achieved the best error metrics for daytime, nighttime, and all-day scenarios. Compared to DLinear, this model achieved significant reductions of 17.4% and 26.2% in MAE and MSE for 16-dimensional features throughout the day, demonstrating superior prediction accuracy. Furthermore, unlike LSTM, CNN-LSTM, and TCN, which show significant differences in winding temperature prediction accuracy between day and night, this model maintains consistent prediction accuracy across different time periods, demonstrating robustness to changes in operating conditions. Furthermore, this model not only maintains its advantage in predicting the complete 16-dimensional features (A) of motors 1#-4#, but also consistently leads in single feature prediction for four categories (B, C, W, P). The improvement in accuracy is not concentrated in a single feature category, demonstrating a universal predictive ability for physical quantities with different time constants.

[0125] Actual deployment results show that this model has approximately 27.9M parameters, about 637MMACs of computation per inference, and an average single-sample inference time of about 1.79ms. While this model does not offer an advantage over DLinear's 1.4K parameters and 22KMACs of computation per inference, it is important to emphasize that the MS-GSE-ModernTCN's single-sample inference time is significantly shorter than the sampling time. Its time and space complexity allow for real-time state prediction on most PC platforms, meeting the needs of engineering applications.

[0126] To verify the specific contributions of each module to the model performance, this paper designed a series of ablation experiments under the same dataset and training configuration. Using ModernTCN as the baseline model, multi-scale embedding layers and grouped channel attention were gradually introduced, and the differences in prediction accuracy of the model under different combinations were compared. The experimental results are detailed in Table 4.

[0127] Table 4. Impact of different modules on prediction accuracy

[0128]

[0129] Ablation experiments show that introducing multi-scale embedding (Model1) alone can reduce RMSE by 33.3%, indicating that frequency division extraction and expanding the effective receptive field are effective for multi-temporal state prediction. Therefore, focusing on the multi-feature single-step prediction task of fixed-wing UAV propulsion system, a temporal state prediction model that integrates multi-scale embedding and grouped channel attention is proposed—MS-GSE-ModernTCN.

[0130] To better implement the fixed-wing UAV temporal state prediction method integrating multi-scale embedding and grouped channel attention in the embodiments of the present invention, based on the corresponding method for fixed-wing UAV temporal state prediction integrating multi-scale embedding and grouped channel attention, please refer to... Figure 4 , Figure 4 This is a schematic diagram of an embodiment of the fixed-wing UAV temporal state prediction system integrating multi-scale embedding and grouped channel attention provided by the present invention. The embodiment of the present invention provides a fixed-wing UAV temporal state prediction system 400 integrating multi-scale embedding and grouped channel attention, comprising:

[0131] The preprocessing module 410 is configured to preprocess the multi-dimensional time-series state data of the acquired fixed-wing UAV propulsion system.

[0132] The mapping and fusion module 420 is configured to use a multi-scale embedding layer to map the preprocessed multi-dimensional temporal state data into feature representations at different time scales, and then fuse them to form a multi-time scale fused feature representation.

[0133] The group weighting module 430 is configured to input the feature representation fused across multiple time scales into a preset backbone network, and to weight the temporal features through a group attention mechanism and residual fusion to output multi-scale temporal features.

[0134] The state prediction module 440 is configured to perform feature flattening and linear transformation on multi-scale temporal features through a mapping layer to obtain the predicted value of the temporal state of the fixed-wing UAV.

[0135] It should be noted that the system 400 provided in the above embodiments can implement the technical solutions described in the above method embodiments. The specific implementation principles of the above modules or units can be found in the corresponding content in the above method embodiments, and will not be repeated here.

[0136] Please see Figure 5 , Figure 5This is a schematic diagram of the structure of a fixed-wing UAV timing state prediction device integrating multi-scale embedding and grouped channel attention, provided in an embodiment of the present invention. Based on the above-described fixed-wing UAV timing state prediction method integrating multi-scale embedding and grouped channel attention, the present invention also provides a corresponding fixed-wing UAV timing state prediction device integrating multi-scale embedding and grouped channel attention. This device can be a computing device such as a mobile terminal, desktop computer, laptop, PDA, or server. The fixed-wing UAV timing state prediction device 500 integrating multi-scale embedding and grouped channel attention includes a processor 510, a memory 520, and a display 530. Figure 5 Only a portion of the components of the fixed-wing UAV temporal state prediction device that integrates multi-scale embedding and grouped channel attention are shown. However, it should be understood that implementation of all the components shown is not required, and more or fewer components may be implemented instead.

[0137] In some embodiments, memory 520 may be an internal storage unit of the fixed-wing UAV timing state prediction device 500 that integrates multi-scale embedding and grouped channel attention, such as a hard disk or memory of the device. In other embodiments, memory 520 may be an external storage device of the device, such as a pluggable hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., provided on the device. Furthermore, memory 520 may include both internal storage units and external storage devices of the device. The memory 520 is used to store application software and various types of data installed on the fixed-wing UAV timing state prediction device 500 that integrates multi-scale embedding and grouped channel attention, such as the program code for the fixed-wing UAV timing state prediction device 500. The memory 520 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 520 stores a fixed-wing UAV timing state prediction program 540 that integrates multi-scale embedding and grouped channel attention. This program 540 can be executed by the processor 510 to implement the fixed-wing UAV timing state prediction methods integrating multi-scale embedding and grouped channel attention according to the embodiments of this application.

[0138] In some embodiments, processor 510 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 520 or process data, such as executing a fixed-wing UAV time-series state prediction method that integrates multi-scale embedding and grouped channel attention.

[0139] In some embodiments, display 530 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 530 is used to display information from the fixed-wing UAV timing state prediction device 500 that integrates multi-scale embedding and grouped channel attention, as well as to display a user interface for visualization. Components 510-530 of the fixed-wing UAV timing state prediction device 500 that integrates multi-scale embedding and grouped channel attention communicate with each other via a system bus.

[0140] In one embodiment, when the processor 510 executes the fixed-wing UAV timing state prediction program 540 fused with multi-scale embedding and grouped channel attention stored in the memory 520, it implements the steps in the above-described fixed-wing UAV timing state prediction method fused with multi-scale embedding and grouped channel attention.

[0141] This embodiment also provides a computer-readable storage medium storing a fixed-wing UAV timing state prediction program that integrates multi-scale embedding and grouped channel attention. When executed by a processor, the fixed-wing UAV timing state prediction program that integrates multi-scale embedding and grouped channel attention implements the following steps:

[0142] Preprocess the multi-dimensional time-series state data of the collected fixed-wing UAV propulsion system;

[0143] A multi-scale embedding layer is used to map the preprocessed multi-dimensional temporal state data into feature representations at different time scales, and then the representations are fused to form a multi-time-scale fused feature representation.

[0144] The feature representation fused across multiple time scales is input into a pre-defined backbone network. The temporal features are weighted through a group attention mechanism and residual fusion to output multi-scale temporal features.

[0145] By performing feature flattening and linear transformation on multi-scale temporal features through a mapping layer, the predicted values ​​of the temporal state of a fixed-wing UAV are obtained.

[0146] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method. The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, as well as magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0147] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0148] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0149] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of embodiments of this disclosure upon considering the specification and practicing the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

[0150] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0151] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the temporal state of a fixed-wing UAV by integrating multi-scale embedding and grouped channel attention, characterized in that, include: The multi-dimensional time-series status data of the fixed-wing UAV propulsion system collected is preprocessed. The multi-dimensional time-series status data includes indicated airspeed, vacuum speed, ground speed, barometric altitude, altitude, atmospheric static temperature, battery voltage, motor command speed, motor actual speed, bus voltage, bus current, control unit temperature, winding temperature, power device temperature, rotor position, calculated torque, D-axis current, Q-axis current, motor system status, and communication status words. A multi-scale embedding layer is used to map the preprocessed multi-dimensional temporal state data into feature representations at different time scales, and then the representations are fused to form a multi-time-scale fused feature representation. The feature representation fused across multiple time scales is input into a pre-defined backbone network. The temporal features are weighted through a group attention mechanism and residual fusion to output multi-scale temporal features. The multi-scale temporal features are flattened and linearly transformed by a mapping layer to obtain the predicted values ​​of the temporal state of the fixed-wing UAV. The step of inputting the multi-timescale fused feature representation into a preset backbone network, weighting the temporal features through a group attention mechanism and residual fusion, and outputting multi-scale temporal features further includes: The feature representation fused across multiple time scales is subjected to temporal-dimensional convolution processing. The convolutional feature representations are grouped according to the variable dimension, and the first nonlinear transformation and the first channel attention weighting operation are performed on the features within each group; The features after the first channel attention weighting operation are grouped according to the feature dimension, and the second nonlinear transformation and the second channel attention weighting operation are performed on the features within each group. The features after the second channel attention weighting operation are fused with the feature representation fused across multiple time scales to output multi-scale temporal features; Both the first nonlinear transformation and the second nonlinear transformation are implemented through grouped convolution operations, and the number of groups in the grouped convolution operation is the same as the number of groups for the corresponding variable dimension and the number of groups for the feature dimension.

2. The method for predicting the temporal state of a fixed-wing UAV by fusing multi-scale embedding and grouped channel attention as described in claim 1, characterized in that, The method of using a multi-scale embedding layer to map preprocessed multi-dimensional temporal state data into feature representations at different time scales, and then fusing them to form a multi-time-scale fused feature representation, further includes: By using several one-dimensional convolution kernels with different preset time window lengths, parallel sliding convolution processing is performed on the preprocessed multi-dimensional temporal state data to obtain feature representations at different time scales; The feature representations at different time scales are spliced ​​and fused along the feature channel dimension to form the multi-time scale fused feature representation.

3. The method for predicting the temporal state of a fixed-wing UAV by fusing multi-scale embedding and grouped channel attention as described in claim 1, characterized in that, Both the first channel attention weighting operation and the second channel attention weighting operation include: Global statistics are performed on the features within the group along the time dimension to obtain the group channel description vector; The grouped channel description vectors are processed through a fully connected layer and a nonlinear activation function to generate scaling weights for each feature channel within the group. The scaling weights are used to perform channel-by-channel scaling of the features within the corresponding groups to achieve a weighted operation.

4. The method for predicting the temporal state of a fixed-wing UAV by fusing multi-scale embedding and grouped channel attention as described in claim 1, characterized in that, The step of flattening and linearly transforming the multi-scale temporal features through a mapping layer to obtain the predicted value of the temporal state of the fixed-wing UAV also includes: The multi-scale temporal features are flattened and stitched together in the variable dimension and the time dimension to form a one-dimensional feature vector. By mapping the one-dimensional feature vector to a space with the same dimension as the target state through linear transformation, the predicted value of the temporal state of the fixed-wing UAV is obtained.

5. The method for predicting the temporal state of a fixed-wing UAV by fusing multi-scale embedding and grouped channel attention as described in claim 1, characterized in that, The preprocessing of the multi-dimensional time-series state data of the collected fixed-wing UAV propulsion system also includes: Anomaly detection is performed on the multi-dimensional time-series state data based on a preset statistical distribution interval. The identified outliers are marked as missing values ​​and filled in using interpolation. A standard scaler is used to standardize the padded complete time-series state data sequence.

6. A fixed-wing UAV temporal state prediction device integrating multi-scale embedding and grouped channel attention, characterized in that, include: The preprocessing module is configured to preprocess the multi-dimensional time-series state data of the collected fixed-wing UAV propulsion system. The multi-dimensional time-series state data includes indicated airspeed, vacuum speed, ground speed, barometric altitude, altitude, atmospheric static temperature, battery voltage, motor command speed, motor actual speed, bus voltage, bus current, control unit temperature, winding temperature, power device temperature, rotor position, calculated torque, D-axis current, Q-axis current, motor system status, and communication status words. The mapping and fusion module is configured to use a multi-scale embedding layer to map the preprocessed multi-dimensional temporal state data into feature representations at different time scales, and then fuse them to form a multi-time scale fused feature representation. The group weighting module is configured to input the feature representation fused from multiple time scales into a preset backbone network, and to weight the temporal features through a group attention mechanism and residual fusion to output multi-scale temporal features. The state prediction module is configured to perform feature flattening and linear transformation on the multi-scale temporal features through a mapping layer to obtain the predicted value of the temporal state of the fixed-wing UAV. The step of inputting the multi-timescale fused feature representation into a preset backbone network, weighting the temporal features through a group attention mechanism and residual fusion, and outputting multi-scale temporal features further includes: The feature representation fused across multiple time scales is subjected to temporal-dimensional convolution processing. The convolutional feature representations are grouped according to the variable dimension, and the first nonlinear transformation and the first channel attention weighting operation are performed on the features within each group; The features after the first channel attention weighting operation are grouped according to the feature dimension, and the second nonlinear transformation and the second channel attention weighting operation are performed on the features within each group. The features after the second channel attention weighting operation are fused with the feature representation fused across multiple time scales to output multi-scale temporal features; Both the first nonlinear transformation and the second nonlinear transformation are implemented through grouped convolution operations, and the number of groups in the grouped convolution operation is the same as the number of groups for the corresponding variable dimension and the number of groups for the feature dimension.

7. A fixed-wing UAV temporal state prediction device integrating multi-scale embedding and grouped channel attention, characterized in that, It includes at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the steps of the fixed-wing UAV temporal state prediction method according to any one of claims 1 to 5, which integrates multi-scale embedding and grouped channel attention.

8. A storage medium, characterized in that, It stores a computer program that can be executed by a fixed-wing UAV timing state prediction device that integrates multi-scale embedding and grouped channel attention. When the computer program is run on the fixed-wing UAV timing state prediction device that integrates multi-scale embedding and grouped channel attention, the device performs the steps of the fixed-wing UAV timing state prediction method that integrates multi-scale embedding and grouped channel attention as described in any one of claims 1 to 5.

Images