CN-121997235-A - Cloud edge cooperative sensing and autonomous intervention method and system for abnormal state of unmanned aerial vehicle

Abstract

The invention belongs to the technical field of intelligent unmanned system safety monitoring and edge artificial intelligence. According to the method, based on Yun Bian collaborative architecture, real-time analysis and lightweight anomaly detection are carried out on time sequence data of the unmanned aerial vehicle multisource sensor through the edge side, a control instruction is generated immediately to intervene on the unmanned aerial vehicle when anomaly is detected, meanwhile, deep learning model training is carried out on historical flight data by depending on a cloud, a lightweight model suitable for edge equipment is generated by utilizing knowledge distillation or model pruning technology, and continuous self-adaptive evolution of anomaly detection capability is realized. The invention provides a new thought for detecting the abnormal state of the unmanned aerial vehicle, and realizes the abnormal state sensing and autonomous intervention in the flight process of the unmanned aerial vehicle.

Inventors

• SHI GANG
• LI ZHENG

Assignees

• 新疆大学

Dates

Publication Date

20260508

Application Date

20260128

Claims (6)

1. 1. Cloud edge cooperative sensing and autonomous intervention method and system for abnormal state of unmanned aerial vehicle, which are characterized by comprising the following steps: acquiring end-side data, namely acquiring flight state data in real time through a multi-source sensor in the flight process of the unmanned aerial vehicle, wherein the data comprise Inertial Measurement Unit (IMU) data, GPS positioning data and flight control log data, and transmitting the data to an edge-side ground station in real time; The side real-time reasoning and synchronization step comprises the steps of receiving and analyzing the flight state data by a side node, preprocessing the data, and then carrying out real-time analysis by using a lightweight anomaly detection model; The cloud server receives and stores data from the edge side to construct a historical flight database, and deep learning model training is carried out on the historical data by utilizing cloud computing power to generate an updated anomaly detection model; And in the model closed-loop updating step, the cloud can generate a lightweight anomaly detection model applicable to edge side computing resource limitation through knowledge distillation or model pruning technology, and send the lightweight anomaly detection model to an edge side ground station through a downlink to replace an old model, so that the self-adaptive evolution of the anomaly detection system is realized.
2. 2. The method of claim 1, wherein the lightweight anomaly detection model in the edge side module is capable of completing anomaly detection within millisecond time and feeding back to the edge side node, and the edge side node makes immediate decisions including but not limited to operations such as back voyage, hover, forced landing, changing course, etc. to ensure flight safety.
3. 3. The method of claim 1, wherein the cloud server pre-processes the stored historical flight data, constructs a multivariate time series data set, and trains a high-precision depth anomaly detection model based on the data set, wherein the depth anomaly detection model is an unsupervised self-encoder structure and comprises an LSTM-Autoencoder or a transducer-Autoencoder, and the data deviating from the mode is judged to be abnormal by minimizing the reconstruction error of the normal flight data and learning the normal behavior mode of the system.
4. 4. The method of claim 1, wherein the lightweight anomaly detection model is a knowledge-distilled self-encoder model comprising a fully-connected network-based self-encoder or a long-short-term memory network (LSTM) -based self-encoder, having model parameters of less than 100KB, suitable for use in resource-constrained edge computing devices.
5. 5. The method according to claim 1, wherein the process of converting the model after training update into the lightweight model specifically adopts a knowledge distillation method based on a teacher-Student Network (Teacher-Student Network), constructs a heterogeneous model architecture, trains a high-precision teacher model by using full amount of historical data at the cloud end, constructs a lightweight chemo model adapting to edge computing force, combines distillation training, fixes parameters of the teacher model, uses a Soft tag (Soft Targets) output by the teacher model as a supervision signal, and drives the Student model to simulate decision boundaries and feature extraction capability of the teacher model by minimizing KL divergence (Kullback-Leibler Divergence) between output probability distributions of the teacher model and the Student model.
6. 6. The method of claim 1, wherein the model closed-loop updating comprises difficult case mining, wherein an edge side ground station screens out difficult samples with low confidence or indeterminate, the difficult samples are actively uploaded to a cloud after marking, collaborative training, namely, the cloud aggregates multi-machine uploaded difficult samples and historical data, retrains knowledge distillation on a deep learning model, namely, based on a teacher-Student (Teacher-Student) network architecture, knowledge of a cloud high-precision teacher model is migrated to an edge side Student model, and a lightweight model with small parameters and high precision is generated and issued.

Description

Cloud edge cooperative sensing and autonomous intervention method and system for abnormal state of unmanned aerial vehicle Technical Field The invention relates to the technical field of intelligent unmanned system safety monitoring and edge artificial intelligence, in particular to an abnormal state sensing and autonomous intervention method and system for unmanned plane flight process. Background Along with the rapid development of low-altitude economy, unmanned aerial vehicles are deployed on a large scale in the scenes of electric power inspection, logistics distribution, emergency rescue, agricultural plant protection and the like. However, its flight safety is highly dependent on the ability to perceive its own state in real time. In complex environments, sensor drift, motor failure, communication interruption, external disturbances, etc. may lead to runaway and even crash accidents. Therefore, the construction of a set of highly reliable, low-delay and evolutionary abnormal state detection and response mechanism has become a key technical requirement for guaranteeing the safe operation of the unmanned aerial vehicle. Disclosure of Invention The invention provides a cloud edge cooperative sensing and autonomous intervention method and a cloud edge cooperative sensing and autonomous intervention system for an abnormal state of an unmanned aerial vehicle, which are used for improving the capability of detecting the abnormal state of the unmanned aerial vehicle and guaranteeing the safe flight of the unmanned aerial vehicle. The invention provides a cloud edge cooperative sensing and autonomous intervention method and system for an abnormal state of an unmanned aerial vehicle. The specific technical scheme is as follows: A cloud edge end cooperative closed-loop sensing system is constructed, the method mainly comprises four core steps: The unmanned plane collects flight data and flight control logs in real time through multisource sensors (IMU, GPS and the like) and sends the flight data and the flight control logs to an edge side ground station; The edge side preprocessing and synchronization comprises the steps that after data are preprocessed by an edge side ground station, an abnormal state of the unmanned aerial vehicle is detected in real time by using a lightweight model, and instant intervention is performed when the abnormality is found, and meanwhile, the unmanned aerial vehicle data are synchronized to a cloud; Cloud training, namely training and updating a high-precision deep learning abnormal detection model by using massive historical data by the cloud; and model updating, namely generating a light abnormal detection model by the cloud through knowledge distillation or pruning, and transmitting the light abnormal detection model to the edge side through a downlink to replace the old model, so as to realize system evolution. The autonomous intervention of the millisecond level of the edge side is realized, and the lightweight abnormal detection model in the edge side module can complete abnormal detection in millisecond level time. Once an abnormality is detected, the edge side node does not depend on a cloud instruction, but directly makes instant decisions (including returning, hovering, forced landing, changing a route and the like) according to a preset strategy, so that flight safety is guaranteed under extreme conditions such as communication interruption and the like. And constructing a cloud high-precision unsupervised model, namely constructing a multivariate time sequence data set on the historical data by the cloud server, and training a high-precision deep anomaly detection model. The model adopts an unsupervised self-encoder structure (such as LSTM-Autoencoder or transducer-Autoencoder) and learns the normal behavior mode of the system by minimizing the reconstruction error of normal flight data, thereby effectively identifying abnormal data deviating from the mode. Lightweight model design adapting edge computing force lightweight anomaly detection model issued to edge side is a knowledge distillation derived self-encoder model (e.g. fully connected network or LSTM based self-encoder). The model parameters are compressed to be within 100KB, so that the method can be perfectly matched with edge computing equipment with limited computing resources, and can be ensured to run under low power consumption. In the model light-weight process, the invention constructs a heterogeneous model framework, wherein the cloud end is a high-precision teacher model, and the edge end is a light-weight student model. The teacher model parameters are fixed during training, the soft labels and the real labels which are output by the teacher model parameters are used as supervision signals, and the student model is driven to simulate the decision boundary of the teacher model by minimizing KL divergence between probability distribution output by the soft labels and the real labels, so that the student