An intelligent unmanned aerial vehicle power system anomaly detection method

Abstract

The application discloses an intelligent unmanned aerial vehicle power system anomaly detection method, which comprises unmanned aerial vehicle data acquisition, data preliminary processing, pseudo label generation, anomaly detection model construction and power system anomaly detection. The original data is obtained through data acquisition; the improved graph clustering model is used to generate pseudo labels to expand the training data; through semi-supervised learning, the unlabeled data is fully utilized, and the time sequence, causality and physical constraints are fused, so that the abnormal characteristics in the system dynamic evolution can be more comprehensively captured, and the false positives and false negatives can be reduced; the double-branch deep learning model is used as the anomaly detection model, the short-time local features and long-time global trajectory are processed in parallel, the multi-scale evolution mode of the anomaly can be comprehensively captured, the limitation of single perspective is avoided, the unmanned aerial vehicle power system physical priori knowledge is introduced to enhance the feature fusion, and the output conforms to the system dynamics law, so that the detection accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology for anomalies in UAV power systems, specifically to an intelligent method for detecting anomalies in UAV power systems. Background Technology

[0002] Unmanned Aerial Vehicle (UAV) power system anomaly detection refers to the use of technologies such as artificial intelligence, machine learning, and data analysis to monitor various parameters of the UAV power system in real time and automatically identify, diagnose, and predict potential faults or abnormal states. By learning from historical data, it improves the sensitivity and accuracy of fault detection, enabling the UAV to autonomously determine whether the power system is in an abnormal state during flight, thereby improving the safety and reliability of the UAV and reducing accidents and losses caused by power system failures.

[0003] However, traditional UAV power system anomaly detection suffers from several problems. It lacks an effective semi-supervised mechanism, the model's performance degrades significantly when data is scarce, it cannot adapt to the real-time requirements of dynamic flight environments, and it struggles to handle complex relationships between multiple data sources, resulting in insufficient ability to identify abnormal patterns in the power system. Furthermore, traditional UAV power system anomaly detection cannot effectively integrate local transient changes and long-term dependencies, leading to insufficient sensitivity in detecting slowly evolving or sudden anomalies. It also lacks multi-scale feature fusion and embedding of physical constraints of the UAV power system, which may cause the model output to deviate from the actual system behavior. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent method for detecting anomalies in UAV power systems. Traditional UAV power system anomaly detection methods suffer from a lack of effective semi-supervised mechanisms, significant performance degradation when data is scarce, inability to adapt to real-time demands in dynamic flight environments, and difficulty in handling complex relationships between multi-source data, resulting in insufficient ability to identify anomaly patterns in the power system. This solution creatively employs an improved graph clustering model to generate pseudo-labels to expand the training data. Through semi-supervised learning, it fully utilizes unlabeled data, automatically identifies potential patterns in the data, thereby expanding the training samples. Simultaneously, it integrates temporal, causal, and physical constraints, enabling a more comprehensive capture of system dynamics. This solution addresses the shortcomings of traditional UAV dynamic system anomaly detection methods. These methods fail to effectively integrate local transient changes and long-term dependencies, resulting in insufficient sensitivity for detecting slowly evolving or sudden anomalies. Furthermore, the lack of multi-scale feature fusion and embedding of physical constraints within the UAV dynamic system leads to model outputs that may deviate from actual system behavior. This innovative approach employs a dual-branch deep learning model for anomaly detection. By processing short-term local features and long-term global trajectories in parallel, it comprehensively captures the multi-scale evolution patterns of anomalies, avoiding the limitations of a single perspective. Simultaneously, it incorporates prior physical knowledge of the UAV dynamic system to enhance feature fusion, ensuring the output conforms to the laws of system dynamics and improving detection accuracy.

[0005] The technical solution adopted by this invention is as follows: This invention provides an intelligent method for detecting anomalies in the power system of an unmanned aerial vehicle (UAV), which includes the following steps:

[0006] Step S1: UAV data acquisition;

[0007] Step S2: Preliminary data processing;

[0008] Step S3: Pseudo-tag generation;

[0009] Step S4: Anomaly detection model construction;

[0010] Step S5: Power system anomaly detection.

[0011] Further, in step S1, the UAV data acquisition is used to collect the raw data required for realizing the anomaly detection of the UAV power system. Specifically, through data acquisition, an anomaly detection raw dataset is obtained. The anomaly detection raw dataset specifically includes a past anomaly detection raw dataset and a real-time anomaly detection raw dataset. Both the past anomaly detection raw dataset and the real-time anomaly detection raw dataset contain system parameter data, UAV motion state data, environmental data, control command data, and physical embedding data. The past anomaly detection raw dataset also contains UAV power system anomaly pattern labels.

[0012] Further, in step S2, the preliminary data processing is used to perform preliminary processing on the collected raw data, specifically including the following steps:

[0013] Step S21: Multi-source data alignment, used to ensure the time synchronization and data quality of the original data, specifically by aligning the original data with different sampling rates using a dynamic time warping algorithm, based on 3 The principle is to remove erroneous data points and use spline interpolation to fill in missing values, resulting in a clean and time-aligned continuous data sequence.

[0014] Step S22: Noise filtering, used to further clean the data, specifically by using digital signal processing technology to separate and filter out high-frequency noise and transient artifacts caused by electromagnetic interference and vibration in the original data, retaining the data that truly reflects the system state;

[0015] Step S23: Feature standardization, used to eliminate the influence of different physical dimensions and reduce feature dimensions to improve model training efficiency. Specifically, the data is converted into dimensionless values ​​through the min-max scaling method, and key features with 95% variance are retained based on principal component analysis to obtain a standardized feature vector with uniform dimensions.

[0016] Step S24: Temporal slicing processing, used to transform continuous temporal data into samples that the model can process, specifically by slicing temporal data through overlapping sliding windows to generate short-time window samples and long-time trajectory samples based on flight phases;

[0017] Step S25: Dataset segmentation, used to obtain training data and test data, specifically, to segment the original dataset of the anomaly detection process.

[0018] The real-time raw dataset for anomaly detection is pre-processed through multi-source data alignment, noise filtering, feature standardization, and temporal slicing to obtain a dataset to be detected. The past raw dataset for anomaly detection is pre-processed through multi-source data alignment, noise filtering, feature standardization, temporal slicing, and dataset segmentation to obtain an anomaly detection training set and an anomaly detection test set.

[0019] Furthermore, in step S3, the pseudo-label generation is used to generate pseudo-labels for unlabeled data to expand the training data. Specifically, it is achieved by constructing an improved graph clustering model and combining it with a multi-graph structure fusion mechanism to obtain an anomaly detection expanded training set.

[0020] The generation of pseudo-tags specifically includes the following steps:

[0021] Step S31: Multi-graph structure construction, used to transform time-series data into a graph structure to comprehensively capture temporal, causal, and physical relationships, includes the following steps:

[0022] Step S311: Temporal adjacency graph construction, used to establish basic temporal connection relationships to capture state changes at consecutive time points. Specifically, by defining the feature vector of each time point as a node and connecting consecutive time points, the exponential decay based on state changes is calculated as the time edge weight to obtain the temporal adjacency graph.

[0023] Step S312: Causal inference graph construction, used to discover causal relationships between data to enhance the representation of fault propagation paths. Specifically, it involves calculating the causal strength between nodes by improving the propagation entropy algorithm, and filtering out causal edges with causal edge weights greater than the causal edge threshold to construct the causal inference graph.

[0024] Step S313: Physical constraint graph construction, used to introduce prior physical knowledge of the UAV power system to ensure the rationality of the constraint graph structure, specifically to construct a physical constraint graph based on the functional correlation of nodes;

[0025] Step S314: Multigraph structure fusion, specifically, weighted fusion of temporally adjacent graphs, causal reasoning graphs, and physical constraint graphs to obtain a fused multigraph structure;

[0026] Step S32: Track encoder design for encoding graph structures into long-term track coding features to capture long-term dependencies, including the following steps:

[0027] Step S321: Design of the spatiotemporal graph attention layer, which is used to simultaneously consider the importance of spatial and temporal neighbors to enhance node representation. Specifically, it obtains node update features by combining spatial attention mechanism with temporal attention mechanism.

[0028] Step S322: Trajectory memory pool design, used to save and utilize historical long-term trajectory patterns to improve long-term dependency modeling, specifically by constructing learnable memory slots and combining them with attention mechanisms to retrieve memory-enhanced features;

[0029] Step S323: Multi-scale graph convolution is used to capture local and global graph structure information to enrich feature representation. Specifically, it aggregates neighbor features of different orders through multi-hop graph convolution, designs multiple scales, and fuses multi-scale outputs to obtain the long-term trajectory encoding features of nodes.

[0030] Step S33: Improve graph clustering to cluster long-term trajectory encoding features to discover anomalous patterns in dynamical systems. The steps include:

[0031] Step S331: Adaptive cluster center learning, used to dynamically adjust cluster centers to improve clustering quality. Specifically, it calculates the probability of each cluster belonging to a node of the fusion multi-graph structure through soft assignment, and optimizes the cluster centers using the target distribution to obtain a graph clustering structure that adapts to the data distribution.

[0032] Step S332: Trajectory continuity constraint, used to ensure the smoothness and causal consistency of clustering results over time. Specifically, it constrains the clustering similarity of adjacent time points through temporal smoothing loss and adds causal consistency loss to obtain a physically reasonable graph clustering structure.

[0033] Step S34: Generate pseudo-labels to transform the clustering results into reliable pseudo-labels. Specifically, by calculating the contour coefficient of each cluster, selecting high-confidence clusters with a contour coefficient greater than a preset threshold, and matching them with known UAV power system anomaly pattern labels, the labels are diffused on the fused multigraph structure based on the label propagation algorithm to generate a pseudo-label set for expanding the anomaly detection training set, thus obtaining the anomaly detection expanded training set.

[0034] Further, in step S4, the anomaly detection model construction is used to construct the model required to realize the anomaly detection of the UAV power system. Specifically, it constructs a dual-branch deep learning model as the anomaly detection model. The dual-branch deep learning model specifically includes a short-time branch and a long-time branch. By fusing short-time local information and long-time trajectory information, it obtains the predicted anomaly mode and severity estimate of the UAV power system.

[0035] The construction of the anomaly detection model specifically includes the following steps:

[0036] Step S41: Input preprocessing, used to prepare input data for the dual-branch deep learning model and ensure uniform format, specifically by obtaining short-term branch input features through short-term window sample encoding and obtaining long-term branch input features through long-term trajectory encoding feature linear mapping;

[0037] Step S42: Short-time branching design for extracting subtle anomaly patterns from a local short-time window, including the following steps:

[0038] Step S421: Multi-scale causal convolution is used to capture local patterns at different temporal granularities to enhance feature diversity. Specifically, it processes short-term branch input features through parallel multi-scale convolution kernels and uses causal padding to maintain temporal sequence to obtain multi-scale local features. The parallel multi-scale convolution kernels specifically include three scales: 3ms, 10ms, and 50ms.

[0039] Step S422: Scale attention fusion, used to adaptively fuse multi-scale local features to highlight important information. Specifically, it calculates the weights of each scale through an attention mechanism and weights the fused features to obtain fused local features.

[0040] Step S423: Physical constraint loop unit design, used to process sequential data and ensure that the output conforms to physical laws, specifically by improving the gated loop unit to add physical boundary constraints and state pruning to obtain short-time branch output characteristics;

[0041] Step S43: Long-term branch design, used to extract long-term dependencies and anomalous evolution patterns from long-term trajectory information, includes the following steps:

[0042] Step S431: Multi-scale convolutional coding is used to capture the dependency patterns at different time scales in long-term trajectory sequences and ensure temporal causality. Specifically, it involves designing four causal convolutional layers with dilation factors of 1, 2, 4 and 8 to process the long-term branch input features in parallel, and then concatenating the output features of the four causal convolutional layers to obtain multi-scale trajectory coding features.

[0043] Step S432: State space model design, used to model long-term dependencies in trajectory sequences. Specifically, multi-scale trajectory encoding features are used as input to the structured state space model, and the feature patterns of the multi-scale trajectory encoding features at the first L time steps are used to initialize the initial state of the structured state space model based on the multilayer perceptron to construct the state space model.

[0044] Step S433: Obtain the long-term branch output, which is used to integrate the multi-scale trajectory coding feature information with the output of the state space model to form a comprehensive trajectory representation. Specifically, the multi-scale trajectory coding features and the output features of the state space model are spliced ​​together, and the long-term branch output features are obtained after linear mapping.

[0045] Step S44: Dual-branch feature fusion, used to integrate short-time branch output and long-time branch output, includes the following steps:

[0046] Step S441: Introduce physical attention modulation to embed physical prior knowledge into the attention mechanism. Specifically, this involves nonlinearly mapping and decomposing the physical embedding features obtained from physical embedding data to obtain the physical query mapping vector, physical key mapping vector, and physical value mapping vector.

[0047] Step S442: Branch cross attention design, used to realize mutual verification and information complementarity between short-term local details and long-term global trends. Specifically, it calculates short-term-long-term cross attention features and long-term-short-term cross attention features by using physical query, physical key and physical value respectively, and obtains short-term enhancement features and long-term enhancement features based on cross attention features;

[0048] Step S443: Enhance feature fusion, which is used to adaptively fuse features enhanced by cross attention. Specifically, it uses dynamic gating coefficients calculated based on physically embedded features to combine element-wise multiplication and element-wise summation between features in a weighted manner to obtain fused enhanced features.

[0049] Step S45: Multi-task output acquisition, used to output anomaly detection results, specifically by processing and enhancing the fusion features through a fully connected layer and a softmax activation function to obtain the predicted anomaly pattern of the UAV power system, and by processing the fusion enhancement features through a regression layer to obtain the severity estimate;

[0050] Step S46: Construct and train the model. Specifically, this involves integrating the input preprocessing, short-term branch design, long-term branch design, dual-branch feature fusion, and multi-task output acquisition to construct a dual-branch deep learning model. The model is then trained and its performance verified based on the anomaly detection expanded training set and the anomaly detection test set to obtain the dual-branch deep learning model, which serves as the anomaly detection model.

[0051] Further, in step S5, the power system anomaly detection specifically involves using the dataset to be detected as input to the improved graph clustering model, obtaining a long-term trajectory encoding feature set, and combining it with the dataset to be detected as input to the anomaly detection model to obtain the UAV power system anomaly detection result. The UAV power system anomaly detection result specifically includes a predicted UAV power system anomaly pattern and a severity estimate.

[0052] The beneficial effects achieved by the present invention using the above solution are as follows:

[0053] (1) In view of the technical problems of traditional UAV power system anomaly detection, such as lack of effective semi-supervised mechanism, significant performance degradation of model when data is scarce, inability to adapt to real-time requirements in dynamic flight environment, and difficulty in handling complex relationships between multi-source data, resulting in insufficient ability to identify abnormal patterns of power system, this solution creatively adopts an improved graph clustering model to generate pseudo-labels to expand training data. Through semi-supervised learning, it makes full use of unlabeled data, automatically identifies potential patterns in the data, thereby expanding training samples. At the same time, it integrates temporal, causal and physical constraints, which can more comprehensively capture abnormal features in the dynamic evolution of the system and reduce false alarms and false negatives.

[0054] (2) In view of the technical problems of traditional UAV dynamic system anomaly detection, which cannot effectively integrate local transient changes and long-term dependencies, resulting in insufficient detection sensitivity for slow evolution or sudden anomalies, and lack of multi-scale feature fusion and UAV dynamic system physical constraint embedding, the model output may deviate from the actual system behavior, this solution creatively adopts a dual-branch deep learning model as the anomaly detection model. By processing short-term local features and long-term global trajectories in parallel, it can comprehensively capture the multi-scale evolution pattern of anomalies, avoid the limitations of a single perspective, and introduce prior physical knowledge of UAV dynamic system to enhance feature fusion, ensuring that the output conforms to the system dynamics law and improving the accuracy of detection. Attached Figure Description

[0055] Figure 1 A flowchart illustrating an intelligent method for detecting anomalies in the power system of an unmanned aerial vehicle (UAV) provided by this invention.

[0056] Figure 2 This is a flowchart illustrating the preliminary data processing in step S2.

[0057] Figure 3 A flowchart illustrating the process of generating pseudo-labels in step S3;

[0058] Figure 4 A flowchart illustrating the process of constructing the anomaly detection model in step S4.

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0060] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0061] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0062] Example 1, see Figure 1This invention provides an intelligent method for detecting anomalies in the power system of unmanned aerial vehicles (UAVs), which includes the following steps:

[0063] Step S1: UAV data acquisition;

[0064] Step S2: Preliminary data processing;

[0065] Step S3: Pseudo-tag generation;

[0066] Step S4: Anomaly detection model construction;

[0067] Step S5: Power system anomaly detection.

[0068] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the UAV data acquisition is used to collect the raw data required to realize the anomaly detection of the UAV power system. Specifically, through data acquisition, an anomaly detection raw dataset is obtained. The anomaly detection raw dataset specifically includes an anomaly detection past raw dataset and an anomaly detection real-time raw dataset. Both the anomaly detection past raw dataset and the anomaly detection real-time raw dataset contain system parameter data, UAV motion state data, environmental data, control command data, and physical embedding data. The anomaly detection past raw dataset also contains UAV power system anomaly mode labels.

[0069] The system parameter data specifically includes drone battery voltage data, drone battery current data, drone motor speed data, drone ESC temperature data, and drone propeller thrust data.

[0070] The drone motion state data specifically includes drone three-axis acceleration data, drone angular velocity data, drone position change data, and drone velocity vector data;

[0071] The environmental data specifically includes flight environment temperature and humidity data, flight environment wind speed and direction data, and air pressure and altitude data;

[0072] The control command data specifically includes UAV throttle command data, UAV attitude angle command data, and UAV control allocation matrix;

[0073] The physical embedded data specifically includes dynamic estimation data of battery internal resistance, motor efficiency coefficient data, power system energy conversion efficiency data, and vibration harmonic characteristic data.

[0074] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S2, the preliminary data processing is used to perform preliminary processing on the collected raw data, specifically including the following steps:

[0075] Step S21: Multi-source data alignment, used to ensure the time synchronization and data quality of the original data, specifically by aligning the original data with different sampling rates using a dynamic time warping algorithm, based on 3 The principle is to remove erroneous data points and use spline interpolation to fill in missing values, resulting in a clean and time-aligned continuous data sequence.

[0076] Step S22: Noise filtering, used to further clean the data, specifically by using digital signal processing technology to separate and filter out high-frequency noise and transient artifacts caused by electromagnetic interference and vibration in the original data, retaining the data that truly reflects the system state;

[0077] Step S23: Feature standardization, used to eliminate the influence of different physical dimensions and reduce feature dimensions to improve model training efficiency. Specifically, the data is converted into dimensionless values ​​through the min-max scaling method, and key features with 95% variance are retained based on principal component analysis to obtain a standardized feature vector with uniform dimensions.

[0078] Step S24: Temporal slicing processing, used to transform continuous temporal data into samples that the model can process, specifically by slicing temporal data through overlapping sliding windows to generate short-time window samples and long-time trajectory samples based on flight phases;

[0079] Step S25: Dataset segmentation, used to obtain training data and test data, specifically, to segment the original dataset of the anomaly detection process.

[0080] The real-time raw dataset for anomaly detection is pre-processed through multi-source data alignment, noise filtering, feature standardization, and temporal slicing to obtain a dataset to be detected. The past raw dataset for anomaly detection is pre-processed through multi-source data alignment, noise filtering, feature standardization, temporal slicing, and dataset segmentation to obtain an anomaly detection training set and an anomaly detection test set.

[0081] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S3, the pseudo-label generation is used to generate pseudo-labels for unlabeled data to expand the training data. Specifically, it is to obtain an anomaly detection expanded training set by constructing an improved graph clustering model and combining it with a multi-graph structure fusion mechanism.

[0082] The generation of pseudo-tags specifically includes the following steps:

[0083] Step S31: Multi-graph structure construction, used to transform time-series data into a graph structure to comprehensively capture temporal, causal, and physical relationships, includes the following steps:

[0084] Step S311: Temporal adjacency graph construction, used to establish basic temporal connectivity to capture state changes at consecutive time points. Specifically, this involves defining the feature vector of each time point as a node and connecting consecutive time points, calculating the exponential decay based on state changes as the time edge weights, and obtaining the temporal adjacency graph. The formula used for calculating the time edge weights is as follows:

[0085] ;

[0086] In the formula, This represents the time edge weight between node t and node t+1. Represent the node characteristics of node t. The node characteristics of node t+1 are represented. Indicates the scale parameter of feature change;

[0087] Step S312: Construction of a causal inference graph, used to discover causal relationships between data to enhance the representation of fault propagation paths. Specifically, this involves calculating the causal strength between nodes by improving the propagation entropy algorithm, and selecting causal edges with weights greater than a causal edge threshold to construct the causal inference graph. The formula used is as follows:

[0088] ;

[0089] In the formula, This represents the improved transfer entropy value from variable X to variable Y, where X and Y represent distinct variables. This represents the value of variable Y at time point t+1. This represents the value of variable Y at time point t. This represents the value of variable X at time point t. express , and Simultaneous joint probability distribution Indicates that in the known and Under the conditions, The conditional probability of occurrence Indicates that in the known Under the conditions, The conditional probability of occurrence Indicates the causal boundary threshold. This represents the arithmetic mean of the improved propagation entropy values. The standard deviation of the improved propagation entropy value is represented by... Represents the relationship between node t and node The causal edge weights between them Indicates the distance from node t to node Improved transfer entropy value, This represents the maximum improved propagation entropy value. This represents the maximum time lag parameter;

[0090] Step S313: Physical constraint graph construction, used to introduce prior physical knowledge of the UAV power system to ensure the rationality of the constraint graph structure. Specifically, based on the functional correlation of nodes, a physical constraint graph is constructed, and the physical edge weights are calculated using the following formula:

[0091] ;

[0092] In the formula, Represents the relationship between node t and node Physical edge weights between them Represents the relationship between node t and node The functional correlation value between them is obtained based on the physical prior knowledge of the UAV's power system. This indicates the smallest positive value that is prevented from being divided by zero.

[0093] Step S314: Multigraph structure fusion, specifically, weighted fusion of temporally adjacent graphs, causal reasoning graphs, and physical constraint graphs to obtain a fused multigraph structure;

[0094] Step S32: Track encoder design for encoding graph structures into long-term track coding features to capture long-term dependencies, including the following steps:

[0095] Step S321: Design of the spatiotemporal graph attention layer, which is used to simultaneously consider the importance of spatial and temporal neighbors to enhance node representation. Specifically, it obtains node update features by combining spatial attention mechanisms with temporal attention mechanisms. The formula used is as follows:

[0096] ;

[0097] In the formula, Represents the relationship between node t and node Spatial attention between Let Sl represent the LeakyReLU activation function, and let Sl represent the spatial attention learnable vector. Indicates the spatial mapping weights. Represents a node Node characteristics, Describes the set of spatial neighbors of node t. This represents the node characteristics of node a. Represents the relationship between node t and node Temporal attention, where Tl represents the learnable vector for temporal attention. Indicates the time-mapped weights. Represents the time decay function, Represents the time neighbor set of node t. Represents the node characteristics of node b. This represents the node update characteristics of node t. This represents the ELU activation function. This represents the spatial attention between node t and node c. This represents the temporal attention between node t and node c. This represents the attention weights in the spatiotemporal graph. Represents the node characteristics of node c;

[0098] Step S322: Trajectory memory pool design, used to save and utilize historical long-term trajectory patterns to improve long-term dependency modeling, specifically by constructing learnable memory slots and combining them with attention mechanisms to retrieve memory-enhanced features;

[0099] Step S323: Multi-scale graph convolution is used to capture local and global graph structure information to enrich feature representation. Specifically, it aggregates neighbor features of different orders through multi-hop graph convolution, designs multiple scales, and fuses multi-scale outputs to obtain the long-term trajectory encoding features of nodes.

[0100] Step S33: Improve graph clustering to cluster long-term trajectory encoding features to discover anomalous patterns in dynamical systems. The steps include:

[0101] Step S331: Adaptive cluster center learning, used to dynamically adjust cluster centers to improve clustering quality. Specifically, it calculates the probability of nodes belonging to each cluster by fusing multiple graph structure nodes through soft assignment, and optimizes the cluster centers using the target distribution to obtain a graph clustering structure that adapts to the data distribution. The formula used is as follows:

[0102] ;

[0103] In the formula, This represents the probability that node t belongs to cluster m. This represents the long-term trajectory encoding features of node t. Denotes the cluster center of cluster m. Represents the cluster centers of cluster n. This represents the target distribution of node t for cluster m. This represents the probability that node t belongs to cluster n. This represents the clustering loss value;

[0104] Step S332: Trajectory continuity constraint, used to ensure the smoothness and causal consistency of clustering results over time. Specifically, it constrains the cluster similarity between adjacent time points through temporal smoothing loss and adds causal consistency loss to obtain a physically reasonable graph clustering structure. The formula used is as follows:

[0105] ;

[0106] In the formula, This represents the time series smoothing loss value. This represents the causal consistency loss value. Let represent the probability distribution of node t's affiliation. Let represent the probability distribution of node t+1, and Nc represent the set of node pairs with connecting edges in the causal inference graph. Represents the probability distribution of node t's affiliation and the node The KL divergence between the probability distributions of attribution;

[0107] Step S34: Generate pseudo-labels to transform the clustering results into reliable pseudo-labels. Specifically, by calculating the contour coefficient of each cluster, selecting high-confidence clusters with a contour coefficient greater than a preset threshold, and matching them with known UAV power system anomaly pattern labels, the labels are diffused on the fused multigraph structure based on the label propagation algorithm to generate a pseudo-label set for expanding the anomaly detection training set, thus obtaining the anomaly detection expanded training set.

[0108] By performing the above operations, this solution addresses the technical problems of traditional UAV power system anomaly detection, such as the lack of an effective semi-supervised mechanism, significant performance degradation of the model when data is scarce, inability to adapt to real-time requirements in dynamic flight environments, and difficulty in handling complex relationships between multi-source data, resulting in insufficient ability to identify abnormal patterns in the power system. This solution creatively adopts an improved graph clustering model to generate pseudo-labels to expand the training data. By fully utilizing unlabeled data through semi-supervised learning, it automatically identifies potential patterns in the data, thereby expanding the training samples. At the same time, by integrating temporal, causal, and physical constraints, it can more comprehensively capture abnormal features in the dynamic evolution of the system and reduce false alarms and false negatives.

[0109] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S4, the anomaly detection model is constructed to build the model required to realize the anomaly detection of the UAV power system. Specifically, a dual-branch deep learning model is constructed as the anomaly detection model. The dual-branch deep learning model specifically includes a short-time branch and a long-time branch. By fusing short-time local information and long-time trajectory information, the predicted anomaly mode and severity estimate of the UAV power system are obtained.

[0110] The construction of the anomaly detection model specifically includes the following steps:

[0111] Step S41: Input preprocessing, used to prepare input data for the dual-branch deep learning model and ensure uniform format, specifically by obtaining short-term branch input features through short-term window sample encoding and obtaining long-term branch input features through long-term trajectory encoding feature linear mapping;

[0112] Step S42: Short-time branching design for extracting subtle anomaly patterns from a local short-time window, including the following steps:

[0113] Step S421: Multi-scale causal convolution is used to capture local patterns at different temporal granularities to enhance feature diversity. Specifically, it processes short-term branch input features through parallel multi-scale convolution kernels and uses causal padding to maintain temporal sequence to obtain multi-scale local features. The parallel multi-scale convolution kernels specifically include three scales: 3ms, 10ms, and 50ms.

[0114] Step S422: Scale-attention fusion, used to adaptively fuse multi-scale local features to highlight important information. Specifically, it calculates the weights of each scale through an attention mechanism and then weights and fuses the features to obtain the fused local features. The formula used is as follows:

[0115] ;

[0116] In the formula, Let Ak represent the attention weights at the k-th scale, and let Ai represent the multi-scale attention learnable vector. Represents the hyperbolic tangent function. Represents the weights of the multi-scale mapping. Represents the local features at the k-th scale. Represents the local features at the l-th scale. Indicates the fusion of local features;

[0117] Step S423: Design of a physical constraint loop unit to process sequential data and ensure that the output conforms to physical laws. Specifically, by improving the gated loop unit and adding physical boundary constraints and state pruning, short-time branch output characteristics are obtained. The formula used is as follows:

[0118] ;

[0119] In the formula, This indicates that the gate output is reset at time u. This indicates that the gate output is updated at time u. This indicates resetting the gate weight. This indicates resetting the door bias term. This indicates updating the gate weights. This indicates that the gate bias term is being updated. This represents the hidden state at time u-1. This represents the sigmoid activation function. This represents the candidate hidden state at time u. Indicates the weights calculated for candidate hidden states. This indicates the calculation of bias terms for candidate hidden states. This indicates the hidden state at time u. Let Pbound represent the clipping function, and let Pbound represent the physical boundary determined based on the physical laws of the UAV's dynamic system. This means cropping the candidate hidden states at time u to the range of the physical boundary. This represents element-wise multiplication.

[0120] Step S43: Long-term branch design, used to extract long-term dependencies and anomalous evolution patterns from long-term trajectory information, includes the following steps:

[0121] Step S431: Multi-scale convolutional coding is used to capture the dependency patterns at different time scales in long-term trajectory sequences and ensure temporal causality. Specifically, it involves designing four causal convolutional layers with dilation factors of 1, 2, 4 and 8 to process the long-term branch input features in parallel, and then concatenating the output features of the four causal convolutional layers to obtain multi-scale trajectory coding features.

[0122] Step S432: State-space model design, used to model long-term dependencies in trajectory sequences. Specifically, multi-scale trajectory encoding features are used as input to the structured state-space model. The feature patterns of the multi-scale trajectory encoding features at the first L time steps are used to initialize the initial state of the structured state-space model based on a multilayer perceptron, and the state-space model is constructed. The formula used is as follows:

[0123] ;

[0124] In the formula, This represents the state of the state-space model at time u. Represents the state transition matrix. This represents the state of the state-space model at time u-1. Represents the input matrix, This represents the output of the state-space model at time u. Indicates the output matrix. Represents a directly connected matrix. This represents multi-scale trajectory encoding features. This represents the initial state of the state-space model. This represents the function that runs the multilayer perceptron. The feature patterns representing the multi-scale trajectory coding features at the first L time steps;

[0125] Step S433: Obtain the long-term branch output, which is used to integrate the multi-scale trajectory coding feature information with the output of the state space model to form a comprehensive trajectory representation. Specifically, the multi-scale trajectory coding features and the output features of the state space model are spliced ​​together, and the long-term branch output features are obtained after linear mapping.

[0126] Step S44: Dual-branch feature fusion, used to integrate short-time branch output and long-time branch output, includes the following steps:

[0127] Step S441: Introduce physical attention modulation to embed physical prior knowledge into the attention mechanism. Specifically, this involves nonlinearly mapping and decomposing the physical embedding features obtained from physical embedding data to obtain the physical query mapping vector, physical key mapping vector, and physical value mapping vector.

[0128] Step S442: Branch cross-attention design, used to achieve mutual verification and information complementarity between short-term local details and long-term global trends. Specifically, it calculates short-term-long-term cross-attention features and long-term-short-term cross-attention features using physical queries, physical keys, and physical values ​​respectively, and obtains short-term and long-term enhancement features based on the cross-attention features. The formulas used are as follows:

[0129] ;

[0130] In the formula, Indicates a short-term query. Indicates a long-term query. Indicates a long-time key. Indicates a short-time key. Indicates a long time value. Indicates short-time value, This indicates the output characteristics of the short-time branch. This indicates the output characteristics of long-term branches. Represents the physical query mapping vector. Represents the physical bond mapping vector. Represents a physical value mapping vector. This represents short-term and long-term cross-attention features. This represents long-term and short-term cross-attention features. This represents the softmax activation function. The dimension representing the key. This represents a bias term based on physically embedded features. Indicates short-term enhancement features, Indicates long-term enhancement features;

[0131] Step S443: Enhanced feature fusion, used to adaptively fuse features enhanced by cross-attention. Specifically, it uses dynamic gating coefficients calculated based on physically embedded features, and combines element-wise product and element-wise summation between features in a weighted manner to obtain fused enhanced features. The formula used is as follows:

[0132] ;

[0133] In the formula, This indicates enhanced fusion weights. Represents the physical embedding mapping matrix. Let Pf represent the physical embedding mapping bias term, and Pf represent the physical embedding feature. Indicates fusion enhancement features, This represents element-wise addition.

[0134] Step S45: Multi-task output acquisition, used to output anomaly detection results, specifically by processing and enhancing the fusion features through a fully connected layer and a softmax activation function to obtain the predicted anomaly pattern of the UAV power system, and by processing the fusion enhancement features through a regression layer to obtain the severity estimate;

[0135] Step S46: Construct and train the model. Specifically, this involves integrating the input preprocessing, short-term branch design, long-term branch design, dual-branch feature fusion, and multi-task output acquisition to construct a dual-branch deep learning model. The model is then trained and its performance verified based on the anomaly detection expanded training set and the anomaly detection test set to obtain the dual-branch deep learning model, which serves as the anomaly detection model.

[0136] By performing the above operations, this solution addresses the technical problems of traditional UAV dynamic system anomaly detection, which suffers from insufficient sensitivity to detect slowly evolving or sudden anomalies due to the inability to effectively integrate local transient changes and long-term dependencies, as well as the lack of multi-scale feature fusion and embedding of physical constraints of the UAV dynamic system, potentially causing the model output to deviate from the actual system behavior. This solution creatively adopts a dual-branch deep learning model as the anomaly detection model. By processing short-term local features and long-term global trajectories in parallel, it can comprehensively capture the multi-scale evolution patterns of anomalies, avoiding the limitations of a single perspective. At the same time, it introduces prior physical knowledge of the UAV dynamic system to enhance feature fusion, ensuring that the output conforms to the laws of system dynamics and improving the accuracy of detection.

[0137] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, the power system anomaly detection specifically involves using the dataset to be detected as the input of the improved graph clustering model, obtaining a long-term trajectory encoding feature set, and combining it with the dataset to be detected as the input of the anomaly detection model to obtain the UAV power system anomaly detection result. The UAV power system anomaly detection result specifically includes the predicted UAV power system anomaly pattern and the severity estimate.

[0138] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0139] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0140] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent method for detecting anomalies in the power system of an unmanned aerial vehicle (UAV), characterized in that: The method includes the following steps: S1: UAV data acquisition. Through data acquisition, an anomaly detection raw dataset is obtained. The anomaly detection raw dataset specifically includes past anomaly detection raw datasets and real-time anomaly detection raw datasets. S2: Preliminary data processing: The collected raw data is preliminarily processed to obtain the dataset to be detected, the anomaly detection training set, and the anomaly detection test set; S3: Pseudo-label generation, used to generate pseudo-labels for unlabeled data to augment the training data. Specifically, it involves constructing an improved graph clustering model and combining it with a multi-graph structure fusion mechanism to obtain an anomaly detection-enlarged training set. The pseudo-label generation process includes multi-graph structure construction, trajectory encoder design, improved graph clustering, and pseudo-label generation. The multi-graph structure construction process includes the following steps: Step S311: Temporal adjacency graph construction, used to establish basic temporal connection relationships to capture state changes at consecutive time points. Specifically, by defining the feature vector of each time point as a node and connecting consecutive time points, the exponential decay based on state changes is calculated as the weight of the time edge to obtain the temporal adjacency graph. Step S312: Causal inference graph construction, used to discover causal relationships between data to enhance the representation of fault propagation paths. Specifically, it involves calculating the causal strength between nodes by improving the propagation entropy algorithm, and filtering out causal edges with causal edge weights greater than the causal edge threshold to construct the causal inference graph. Step S313: Physical constraint graph construction, used to introduce prior physical knowledge of the UAV power system to ensure the rationality of the constraint graph structure, specifically to construct a physical constraint graph based on the functional correlation of nodes; Step S314: Multigraph structure fusion, specifically, weighted fusion of temporally adjacent graphs, causal reasoning graphs, and physical constraint graphs to obtain a fused multigraph structure; S4: Anomaly detection model construction, used to construct the model required to realize the anomaly detection of the UAV power system. Specifically, it constructs a dual-branch deep learning model as the anomaly detection model. The dual-branch deep learning model specifically includes a short-time branch and a long-time branch. By fusing short-time local information and long-time trajectory information, it obtains the predicted anomaly mode and severity estimate of the UAV power system. S5: Power system anomaly detection, specifically, the dataset to be detected is used as the input of the improved graph clustering model, and the resulting long-term trajectory encoding feature set is combined with the dataset to be detected as the input of the anomaly detection model to obtain the UAV power system anomaly detection result. The UAV power system anomaly detection result specifically includes the predicted UAV power system anomaly pattern and the severity estimate.

2. The intelligent method for detecting anomalies in the power system of an unmanned aerial vehicle (UAV) according to claim 1, characterized in that: The trajectory encoder design is used to encode graph structures into long-term trajectory coding features to capture long-term dependencies, and the steps include: Step S321: Design of the spatiotemporal graph attention layer, which is used to simultaneously consider the importance of spatial and temporal neighbors to enhance node representation. Specifically, it obtains node update features by combining spatial attention mechanism with temporal attention mechanism. Step S322: Trajectory memory pool design, used to save and utilize historical long-term trajectory patterns to improve long-term dependency modeling, specifically by constructing learnable memory slots and combining them with attention mechanisms to retrieve memory-enhanced features; Step S323: Multi-scale graph convolution is used to capture local and global graph structure information to enrich feature representation. Specifically, it aggregates neighbor features of different orders through multi-hop graph convolution, designs multiple scales, and fuses multi-scale outputs to obtain the long-term trajectory encoding features of nodes. The improved graph clustering, used to cluster long-term trajectory encoding features to discover abnormal patterns in dynamic systems, includes the following steps: Step S331: Adaptive cluster center learning, used to dynamically adjust cluster centers to improve clustering quality. Specifically, it calculates the probability of each cluster belonging to a node of the fusion multi-graph structure through soft assignment, and optimizes the cluster centers using the target distribution to obtain a graph clustering structure that adapts to the data distribution. Step S332: Trajectory continuity constraint, used to ensure the smoothness and causal consistency of clustering results over time. Specifically, it constrains the clustering similarity of adjacent time points through temporal smoothing loss and adds causal consistency loss to obtain a physically reasonable graph clustering structure. The generation of pseudo-labels is used to transform the clustering results into reliable pseudo-labels. Specifically, it involves calculating the contour coefficient of each cluster, selecting high-confidence clusters with a contour coefficient greater than a preset threshold, matching them with known UAV power system anomaly pattern labels, and spreading the labels on the fused multigraph structure based on the label propagation algorithm to generate a pseudo-label set for expanding the anomaly detection training set, thus obtaining the anomaly detection expanded training set.

3. The intelligent method for detecting anomalies in the power system of an unmanned aerial vehicle (UAV) according to claim 1, characterized in that: The construction of the anomaly detection model specifically includes the following steps: Step S41: Input preprocessing; Step S42: Short-time branch design; Step S43: Long-term branch design; Step S44: Dual-branch feature fusion; Step S45: Multi-task output acquisition; Step S46: Build and train the model.

4. The intelligent method for detecting anomalies in the power system of an unmanned aerial vehicle (UAV) according to claim 3, characterized in that: The input preprocessing is used to prepare the input data for the dual-branch deep learning model and ensure that the format is consistent. Specifically, it obtains short-term branch input features through short-term window sample encoding and obtains long-term branch input features through long-term trajectory encoding feature linear mapping. The short-time branching design is used to extract subtle anomaly patterns from a local short-time window, and the steps include: Step S421: Multi-scale causal convolution is used to capture local patterns at different temporal granularities to enhance feature diversity. Specifically, it processes short-term branch input features through parallel multi-scale convolution kernels and uses causal padding to maintain temporal sequence to obtain multi-scale local features. The parallel multi-scale convolution kernels specifically include three scales: 3ms, 10ms, and 50ms. Step S422: Scale attention fusion, used to adaptively fuse multi-scale local features to highlight important information. Specifically, it calculates the weights of each scale through an attention mechanism and weights the fused features to obtain fused local features. Step S423: Physical constraint loop unit design, used to process sequential data and ensure that the output conforms to physical laws, specifically by improving the gated loop unit to add physical boundary constraints and state pruning to obtain short-time branch output characteristics; The long-term branching design is used to extract long-term dependencies and anomalous evolution patterns from long-term trajectory information, and the steps include: Step S431: Multi-scale convolutional coding is used to capture the dependency patterns at different time scales in long-term trajectory sequences and ensure temporal causality. Specifically, it involves designing four causal convolutional layers with dilation factors of 1, 2, 4 and 8 to process the long-term branch input features in parallel, and then concatenating the output features of the four causal convolutional layers to obtain multi-scale trajectory coding features. Step S432: State space model design, used to model long-term dependencies in trajectory sequences. Specifically, multi-scale trajectory encoding features are used as input to the structured state space model, and the feature patterns of the multi-scale trajectory encoding features at the first L time steps are used to initialize the initial state of the structured state space model based on the multilayer perceptron to construct the state space model. Step S433: Obtain the long-term branch output, which is used to integrate the multi-scale trajectory coding feature information with the output of the state space model to form a comprehensive trajectory representation. Specifically, the multi-scale trajectory coding features and the output features of the state space model are spliced ​​together, and the long-term branch output features are obtained after linear mapping. The dual-branch feature fusion, used to integrate short-time branch outputs and long-time branch outputs, includes the following steps: Step S441: Introduce physical attention modulation to embed physical prior knowledge into the attention mechanism. Specifically, this involves nonlinearly mapping and decomposing the physical embedding features obtained from physical embedding data to obtain the physical query mapping vector, physical key mapping vector, and physical value mapping vector. Step S442: Branch cross attention design, used to realize mutual verification and information complementarity between short-term local details and long-term global trends. Specifically, it calculates short-term-long-term cross attention features and long-term-short-term cross attention features by using physical query, physical key and physical value respectively, and obtains short-term enhancement features and long-term enhancement features based on cross attention features; Step S443: Enhance feature fusion, which is used to adaptively fuse features enhanced by cross attention. Specifically, it uses dynamic gating coefficients calculated based on physically embedded features to combine element-wise multiplication and element-wise summation between features in a weighted manner to obtain fused enhanced features. The multi-task output acquisition is used to output the anomaly detection results. Specifically, it processes and enhances the features through a fully connected layer and a softmax activation function to obtain the predicted abnormal mode of the UAV power system, and processes the fused and enhanced features through a regression layer to obtain the severity estimate. The construction and training of the model specifically involves integrating the input preprocessing, short-term branch design, long-term branch design, dual-branch feature fusion, and multi-task output acquisition to construct a dual-branch deep learning model. The model is then trained and its performance verified based on the anomaly detection expanded training set and the anomaly detection test set, resulting in a dual-branch deep learning model that serves as the anomaly detection model.

5. The intelligent method for detecting anomalies in the power system of an unmanned aerial vehicle (UAV) according to claim 1, characterized in that: Both the past raw dataset for anomaly detection and the real-time raw dataset for anomaly detection contain system parameter data, UAV motion state data, environmental data, control command data, and physical embedding data. The past raw dataset for anomaly detection also contains anomaly pattern labels for the UAV power system.

6. The intelligent method for detecting anomalies in the power system of an unmanned aerial vehicle (UAV) according to claim 1, characterized in that: The preliminary data processing specifically includes the following steps: Step S21: Multi-source data alignment, used to ensure the time synchronization and data quality of the original data, specifically by aligning the original data with different sampling rates using a dynamic time warping algorithm, based on 3 The principle is to remove erroneous data points and use spline interpolation to fill in missing values, resulting in a clean and time-aligned continuous data sequence. Step S22: Noise filtering, used to further clean the data, specifically by using digital signal processing technology to separate and filter out high-frequency noise and transient artifacts caused by electromagnetic interference and vibration in the original data, retaining the data that truly reflects the system state; Step S23: Feature standardization, used to eliminate the influence of different physical dimensions and reduce feature dimensions to improve model training efficiency. Specifically, the data is converted into dimensionless values ​​through the min-max scaling method, and key features with 95% variance are retained based on principal component analysis to obtain a standardized feature vector with uniform dimensions. Step S24: Temporal slicing processing, used to transform continuous temporal data into samples that the model can process, specifically by slicing temporal data through overlapping sliding windows to generate short-time window samples and long-time trajectory samples based on flight phases; Step S25: Dataset segmentation, used to obtain training data and test data, specifically, to segment the original dataset of the anomaly detection process. The real-time raw dataset for anomaly detection is pre-processed through multi-source data alignment, noise filtering, feature standardization, and temporal slicing to obtain a dataset to be detected. The past raw dataset for anomaly detection is pre-processed through multi-source data alignment, noise filtering, feature standardization, temporal slicing, and dataset segmentation to obtain an anomaly detection training set and an anomaly detection test set.

Images