CN-121808718-B - Fixed wing unmanned aerial vehicle time sequence state prediction method, system and medium integrating multi-scale embedding and grouping channel attention

CN121808718BCN 121808718 BCN121808718 BCN 121808718BCN-121808718-B

Abstract

The application discloses a method, a system and a medium for predicting a time sequence state of a fixed wing unmanned aerial vehicle by fusing multi-scale embedding and grouping channel attention, which comprise the steps of preprocessing collected multi-dimensional time sequence state data of a propulsion system of the fixed wing unmanned aerial vehicle, mapping the preprocessed multi-dimensional time sequence state data into characteristic representations of different time scales by adopting a multi-scale embedding layer, fusing to form the characteristic representations of multi-time scale fusion, inputting the characteristic representations of the multi-time scale fusion into a preset backbone network, weighting the time sequence characteristics through grouping attention mechanism and residual fusion, outputting the multi-scale time sequence characteristics, and carrying out characteristic flattening and linear transformation on the multi-scale time sequence characteristics through a mapping layer to obtain a predicted value of the time sequence state of the fixed wing unmanned aerial vehicle. The application widens the breadth of time sequence characterization through multi-scale embedding, realizes the self-adaptive screening and enhancement of the characteristic channels through grouping attention, and maintains the optimal stability through residual fusion.

Inventors

WANG LIANQING
HONG HUAJIE
XIANG XIAOJIA
WANG NAN
WANG WEI
HE KEYAN
GAN ZIHAO
HUANG JIE

Assignees

中国人民解放军国防科技大学

Dates

Publication Date: 20260508
Application Date: 20260312

Claims (8)

1. A fixed wing unmanned aerial vehicle time sequence state prediction method integrating multiscale embedding and grouping channel attention is characterized by comprising the following steps: Preprocessing the collected multi-dimensional time sequence state data of the fixed wing unmanned aerial vehicle propulsion system, wherein the multi-dimensional time sequence state data comprises an indicated airspeed, a vacuum speed, a ground speed, an air pressure altitude, an altitude, an atmospheric static temperature, a battery voltage, a motor command rotating speed, a motor actual rotating speed, a bus voltage, a bus current, a control unit temperature, a winding temperature, a power device temperature, a rotor position, a resolving moment, a D-axis current, a Q-axis current, a motor system state and a communication state word; mapping the preprocessed multi-dimensional time sequence state data into feature representations of different time scales by adopting a multi-scale embedding layer, and fusing to form a multi-time scale fused feature representation; inputting the characteristic representation of multi-time scale fusion into a preset backbone network, and carrying out weighting treatment on the time sequence characteristics through a grouping attention mechanism and residual fusion to output multi-scale time sequence characteristics; Performing characteristic flattening and linear transformation on the multi-scale time sequence characteristics through a mapping layer to obtain a predicted value of the time sequence state of the fixed-wing unmanned aerial vehicle; the method for processing the multi-time scale fusion features comprises the steps of inputting the multi-time scale fusion feature representation into a preset backbone network, carrying out weighting processing on time sequence features through a grouping attention mechanism and residual fusion, outputting the multi-time scale time sequence features, and further comprising: performing convolution processing of time dimension on the feature representation fused by the multiple time dimensions; grouping the convolved feature representations by variable dimension and performing a first nonlinear transformation and a first channel attention weighting operation on the features within each group; Grouping the features subjected to the first channel attention weighting operation according to feature dimensions, and executing second nonlinear transformation and second channel attention weighting operation on the features in each group; Fusing the characteristics subjected to the second channel attention weighting operation with the characteristic representation fused in multiple time scales, and outputting the multi-scale time sequence characteristics; The first nonlinear transformation and the second nonlinear transformation are both realized through grouping convolution operation, and the grouping number of the grouping convolution operation is the same as the corresponding variable dimension grouping number and characteristic dimension grouping number.
2. The method for predicting the timing state of a fixed wing unmanned aerial vehicle by fusing attention of a multi-scale embedding and grouping channel according to claim 1, wherein the method for mapping the preprocessed multi-dimensional timing state data into feature representations of different time scales by using a multi-scale embedding layer and fusing the feature representations to form the multi-time scale fused feature representation further comprises: Adopting a plurality of one-dimensional convolution kernels with different preset time window lengths to perform parallel sliding convolution processing on the preprocessed multi-dimensional time sequence state data to obtain feature representations with different time scales; And splicing and fusing the feature representations of different time scales along the feature channel dimension to form the feature representation fused by multiple time scales.
3. The fixed wing drone timing state prediction method of merging multi-scale embedded and grouped channel attentions of claim 1, wherein the first channel attentions weighting operation and the second channel attentions weighting operation each comprise: carrying out global statistics on the characteristics in the group in the time dimension to obtain a group channel description vector; Processing the grouping channel description vector through a full connection layer and a nonlinear activation function to generate scaling weights of all characteristic channels in the grouping; And carrying out channel-by-channel scaling on the characteristics in the corresponding group by utilizing the scaling weight to realize weighting operation.
4. The method for predicting the timing state of the fixed-wing unmanned aerial vehicle by fusing the attention of a multi-scale embedding and grouping channel according to claim 1, wherein the feature flattening and linear transformation are performed on the multi-scale timing features through a mapping layer to obtain a predicted value of the timing state of the fixed-wing unmanned aerial vehicle, and the method further comprises: flattening and splicing the multi-scale time sequence features in a variable dimension and a time dimension to form a one-dimensional feature vector; and mapping the one-dimensional feature vector to a space with the same dimension as the target state by linear transformation to obtain a predicted value of the time sequence state of the fixed-wing unmanned aerial vehicle.
5. The method for predicting the timing state of a fixed-wing unmanned aerial vehicle by fusing multi-scale embedding and grouping channel attention as set forth in claim 1, wherein the preprocessing the collected multi-dimensional timing state data of the fixed-wing unmanned aerial vehicle propulsion system further comprises: Performing anomaly detection on the multi-dimensional time sequence state data based on a preset statistical distribution interval, marking the identified anomaly value as a missing value, and filling by adopting an interpolation method; And adopting a standard scaler to perform standardized transformation on the filled complete time sequence state data sequence.
6. A fixed wing unmanned aerial vehicle time sequence state prediction device integrating multiscale embedding and grouping channel attention is characterized by comprising: The system comprises a preprocessing module, a power device module and a communication module, wherein the preprocessing module is configured to preprocess collected multi-dimensional time sequence state data of the fixed wing unmanned aerial vehicle propulsion system, wherein the multi-dimensional time sequence state data comprises indicating airspeed, vacuum speed, ground speed, air pressure altitude, atmospheric static temperature, battery voltage, motor instruction rotating speed, motor actual rotating speed, bus voltage, bus current, control unit temperature, winding temperature, power device temperature, rotor position, resolving moment, D-axis current, Q-axis current, motor system state and communication state word; The mapping fusion module is configured to map the preprocessed multi-dimensional time sequence state data into feature representations of different time scales by adopting a multi-scale embedding layer, and fuse the feature representations to form a multi-time scale fused feature representation; the grouping weighting module is configured to input the characteristic representation of multi-time scale fusion into a preset backbone network, weight the time sequence characteristic through grouping attention mechanism and residual fusion, and output the multi-scale time sequence characteristic; The state prediction module is configured to perform characteristic flattening and linear transformation on the multi-scale time sequence characteristics through a mapping layer to obtain a predicted value of the time sequence state of the fixed-wing unmanned aerial vehicle; the method for processing the multi-time scale fusion features comprises the steps of inputting the multi-time scale fusion feature representation into a preset backbone network, carrying out weighting processing on time sequence features through a grouping attention mechanism and residual fusion, outputting the multi-time scale time sequence features, and further comprising: performing convolution processing of time dimension on the feature representation fused by the multiple time dimensions; grouping the convolved feature representations by variable dimension and performing a first nonlinear transformation and a first channel attention weighting operation on the features within each group; Grouping the features subjected to the first channel attention weighting operation according to feature dimensions, and executing second nonlinear transformation and second channel attention weighting operation on the features in each group; Fusing the characteristics subjected to the second channel attention weighting operation with the characteristic representation fused in multiple time scales, and outputting the multi-scale time sequence characteristics; The first nonlinear transformation and the second nonlinear transformation are both realized through grouping convolution operation, and the grouping number of the grouping convolution operation is the same as the corresponding variable dimension grouping number and characteristic dimension grouping number.
7. The fixed wing unmanned aerial vehicle time sequence state prediction device integrating the attention of the multi-scale embedding and grouping channels is characterized by 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 executes the steps of the fixed wing unmanned aerial vehicle time sequence state prediction method integrating the attention of the multi-scale embedding and grouping channels according to any one of claims 1-5.
8. A storage medium, characterized in that it stores a computer program executable by a fixed wing unmanned aerial vehicle time sequence state prediction device that merges attention of a multi-scale embedding and grouping channel, when the computer program runs on the fixed wing unmanned aerial vehicle time sequence state prediction device that merges attention of a multi-scale embedding and grouping channel, the fixed wing unmanned aerial vehicle time sequence state prediction device that merges attention of a multi-scale embedding and grouping channel is caused to execute the steps of the fixed wing unmanned aerial vehicle time sequence state prediction method that merges attention of a multi-scale embedding and grouping channel as claimed in any one of claims 1 to 5.

Description

Fixed wing unmanned aerial vehicle time sequence state prediction method, system and medium integrating multi-scale embedding and grouping channel attention Technical Field The application relates to the technical field of unmanned aerial vehicle prediction, in particular to a method, a system and a medium for predicting a time sequence state of a fixed wing unmanned aerial vehicle by fusing multi-scale embedding and grouping channel attention. Background Along with the wide application of unmanned aerial vehicle technology, fixed wing unmanned aerial vehicle plays an important role in long-endurance flight tasks because of the advantages of long range, high speed, strong loading capacity and the like. In order to ensure long-term stable flight of the unmanned aerial vehicle under complex working conditions, accurate prediction of key states (such as motor bus voltage, winding temperature, power device temperature and the like) of a propulsion system of the unmanned aerial vehicle is achieved, and the unmanned aerial vehicle becomes a key link for achieving fault early warning and constructing an autonomous health management system. Accurate state prediction can provide basis for early degradation identification, avoids performance attenuation or safety accidents caused by missed detection, and reduces unnecessary task interruption caused by false alarm at the same time, thereby improving flight safety and task economy. At present, methods for unmanned aerial vehicle state prediction are mainly divided into two types, namely a model-based method and a data-driven method. The model-based method relies on the prior knowledge of physics or dynamics of the system to construct a state observer or filter, and can realize higher-precision state estimation under the scene that the model is accurate and known. However, the fixed-wing unmanned aerial vehicle propulsion system is a complex system with strong coupling of multiple physical quantities and obvious nonlinear characteristics, and an accurate analytical model is difficult to build, so that the generalization capability of the method in practical application is limited, and the modeling accuracy is highly dependent. The data-driven approach does not rely on an accurate physical model, but rather learns the timing law of state evolution directly from historical flight data, where recurrent neural networks and their variants (e.g., LSTM) are widely adopted for their ability to process sequence data. In addition, there have been studies to enhance the capture of key features by models by introducing attention mechanisms or incorporating convolution modules to enhance predictive performance. However, the existing prediction method based on data driving still has obvious limitations that on one hand, most methods model time dimension more singly in a feature extraction stage, are difficult to capture multi-time constant characteristics (namely coexistence of short-term quick response and long-term slow-change trend) of electric and thermodynamic parameters in a fixed-wing unmanned aerial vehicle propulsion system simultaneously, and on the other hand, when high-dimensional multivariable time sequence data are processed, coupling relations and feature importance differences inside variables and among cross variables are often not fully mined, so that insufficient attention distribution to key state features is caused, and noise channel interference is easy to occur. Disclosure of Invention Aiming at least one defect or improvement requirement in the prior art, the invention provides a fixed wing unmanned aerial vehicle time sequence state prediction method, a system and a medium which integrate multi-scale embedding and grouping channel attention, which are used for solving the problems that modeling of time dimension is single, multi-time constant characteristics in a fixed wing unmanned aerial vehicle propulsion system are difficult to capture simultaneously and effectively, and when high-dimensional multivariable time sequence data are processed, coupling relations and feature importance differences between variable interiors and cross variables cannot be fully mined, so that attention distribution to key state features is insufficient and noise channel interference is easy to occur. In order to achieve the above object, according to a first aspect of the present invention, there is provided a fixed wing unmanned aerial vehicle time sequence state prediction method integrating attention of a multi-scale embedding and grouping channel, including: Preprocessing the collected multi-dimensional time sequence state data of the fixed wing unmanned aerial vehicle propulsion system; mapping the preprocessed multi-dimensional time sequence state data into feature representations of different time scales by adopting a multi-scale embedding layer, and fusing to form a multi-time scale fused feature representation; inputting the characteristic representation of multi-time s