Search

CN-121996414-A - Unmanned aerial vehicle radio frequency interference signal multitasking method and device

CN121996414ACN 121996414 ACN121996414 ACN 121996414ACN-121996414-A

Abstract

The invention discloses a multi-task processing method for radio frequency interference signals of an unmanned aerial vehicle, which comprises the steps of constructing a training data set according to radio frequency interference signals on an acquired power inspection line and preprocessing, constructing a multi-task model based on signal detection tasks and modulation recognition tasks of the radio frequency interference signals, designing a multi-task calculation resource dynamic allocation scheme by combining task relations, constructing a multi-target reward function, inputting the preprocessed training data set to train the multi-task model, calculating signal detection accuracy, modulation recognition accuracy and resource utilization rate according to output prediction results and real labels, calculating the multi-target reward function to maximize to obtain a trained multi-task model, dynamically optimizing multi-task model parameters according to multi-dimension indexes to obtain an optimized multi-task model, and inputting radio frequency interference signal data to be tested into the optimized multi-task model to obtain detection and modulation results. The scheme of the invention realizes dynamic allocation of computing resources and automatic optimization of the model.

Inventors

  • ZHOU GUOQIANG
  • LIANG SHUWEI

Assignees

  • 南京邮电大学

Dates

Publication Date
20260508
Application Date
20251231

Claims (10)

  1. 1. The unmanned aerial vehicle radio frequency interference signal multitasking method is characterized by comprising the following steps: constructing a training data set according to the radio frequency interference signals on the acquired power inspection line and preprocessing the training data set; A multi-task model is built based on a signal detection task and a modulation recognition task of the radio frequency interference signal, a multi-task calculation resource dynamic allocation scheme is designed by combining a task relation, and a multi-objective rewarding function is built; Inputting the preprocessed training data set to train the multi-task model, and calculating the signal detection accuracy, the modulation recognition accuracy and the resource utilization rate according to the output prediction result and the real label; The method comprises the steps of dynamically optimizing the parameters of the multi-task model according to the multi-dimension index to obtain an optimized multi-task model, inputting the data of the radio frequency interference signals to be tested into the optimized multi-task model to obtain signal detection and modulation recognition results.
  2. 2. The unmanned aerial vehicle radio frequency interference signal multitasking method according to claim 1, wherein constructing a training data set according to the radio frequency interference signal on the acquired power line and preprocessing comprises: labeling the radio frequency interference signal data to obtain radio frequency interference signal data with labels; The radio frequency interference signal data with the tag are proportionally divided to obtain a training data set; Cleaning and normalizing the training data set; And carrying out multidimensional feature extraction and step-by-step feature selection on the processed training data set to obtain an effective feature extraction result.
  3. 3. The unmanned aerial vehicle radio frequency interference signal multitasking method according to claim 2, wherein the cleaning adopts a four-way bit distance method to calculate the difference between 25% bit number and 75% bit number of training data, and determines the abnormal value judgment range, and the calculation formula is shown in formula (1): (1) Wherein Q3 is 75% quantile of the training data, Q1 is 25% quantile of the training data, IQR is the difference between Q3 and Q1, and the difference is smaller than Or is greater than The training data of the model is judged to be an abnormal value, and the abnormal value is replaced by a median; and/or, the normalization is performed by adopting a minimum-maximum normalization method, and the calculation formula is shown as formula (2): (2) Wherein, the For normalized training data, X is the original training data, As the maximum value of the original training data, Minimum value of original training data; And/or the multi-dimensional feature extraction comprises time domain feature extraction, frequency domain feature extraction and high-order feature extraction; the time domain features comprise the mean value, standard deviation, variance, skewness and kurtosis of the I component, and the mean value, standard deviation, variance, skewness and kurtosis of the Q component; the instantaneous amplitude mean value, the instantaneous amplitude standard deviation, the instantaneous amplitude maximum value and the instantaneous amplitude minimum value; instantaneous phase mean and instantaneous phase standard deviation; signal power, peak power and peak-to-average power ratio; the frequency domain features comprise a spectrum centroid, a spectrum bandwidth, a spectrum roll-off point and a spectrum flatness; the power spectrum density average value, the power spectrum density standard deviation, the power spectrum density maximum value and the power spectrum entropy; wherein the high-order features include a second-order cumulative amount c20, a second-order cumulative amount c21, a fourth-order cumulative amount c40, a fourth-order cumulative amount c41, a fourth-order cumulative amount c42, a second-order cyclic moment, and a fourth-order cyclic moment; and/or the step-by-step feature selection comprises a first-stage selection, a second-stage selection and a third-stage selection, wherein the first-stage selection is And the norm characteristic selection, the second level selection, the third level selection and the variance threshold characteristic selection are adopted as mutual information characteristic selection.
  4. 4. The unmanned aerial vehicle radio frequency interference signal multitasking method according to claim 1, wherein the constructing a multitasking model based on the signal detection task and the modulation recognition task of the radio frequency interference signal comprises: Constructing a shared backbone network and a private branch network, wherein the shared backbone network comprises a shared Conv1D layer, a shared GlobalAveragePooling D layer and a shared Dense layer, and the private branch network comprises a signal detection branch network and a modulation identification branch network; Defining a parameter initial value and a loss function of the multitasking model; The multi-task model parameters comprise the number of shared Conv1D layers of a shared backbone network, the number of private Dense layers of a private branch network, signal detection task weights, modulation recognition task weights and an optimizer learning rate beta; the calculation formula of the loss function is shown in formula (3): (3) Wherein, the As a value of the total loss, The task weights are detected for the signals, The task loss value is detected for the signal, For the modulation to identify the task weights, A task loss value is identified for the modulation.
  5. 5. The unmanned aerial vehicle radio frequency interference signal multitasking method of claim 1, wherein the designing a multitasking computing resource dynamic allocation scheme in combination with the task relationship, constructing a multi-objective rewarding function comprises: the cosine similarity of gradient vectors of different tasks is calculated, and a calculation formula is shown as formula (4): (4) wherein Sim is cosine similarity, The task weight gradient is detected for the signal, Identifying a task weight gradient for the modulation; Dividing the task relationship into four classes according to cosine similarity, so as to group the tasks; task relationship classification includes classifying Is set to be in positive migration, Is set to a neutral relationship in relation to the task, Is set to a weak negative migration, Setting the task relation of the (E) to be strong negative migration; Discretizing the signal detection accuracy, the modulation recognition accuracy and the resource utilization into Is a state space of (2); designing 5 weight distribution strategies, namely dynamically distributing computing resources, wherein the weight distribution strategies comprise deflection detection, balance distribution, deflection identification, uniform enhancement and uniform weakening; constructing a multi-objective rewarding function to improve policy selection rationality of a weight allocation policy, wherein a calculation formula of the multi-objective rewarding function is shown in a formula (5): (5) Wherein, the In order to integrate the prize values, For the accuracy of the signal detection, In order to modulate the recognition accuracy rate of the recognition, The resource utilization rate is a comprehensive normalization index comprising the CPU utilization rate of the unmanned aerial vehicle, the GPU video memory occupancy rate, the battery power and the memory occupancy rate; Combining exploration of new strategy selection and execution history optimal strategy selection by adopting an epsilon-greedy strategy, wherein the new strategy selection is evaluated by the multi-objective rewarding function so as to avoid sinking into local optimal; the epsilon-greedy strategy is used for setting the learning rate alpha to be 0.1, the discount factor gamma to be 0.9, and the learning rate alpha and the discount factor gamma can be dynamically adjusted according to the model training progress.
  6. 6. The unmanned aerial vehicle radio frequency interference signal multitasking method of claim 5, wherein said designing a multitasking computing resource dynamic allocation scheme in combination with task relationships, constructing a multi-objective rewarding function further comprises: Calculating the clustering loss of the task parameters to quantify the task grouping rationality; The task parameters comprise the learnable weight and bias of a shared backbone network shared Global Average Pooling D layer and a shared Dense layer and the learnable weight and bias of a private branch network private Dense layer and an output layer; The calculation formula of the clustering loss is shown in formula (6): (6) Wherein, the For the cluster loss value of the task parameter, The weight coefficients are lost for the global average value, For the global average value loss, As the inter-cluster variance weight coefficient, For the inter-cluster variance to be the same, As the intra-cluster variance weight coefficient, Is the intra-cluster variance.
  7. 7. The unmanned aerial vehicle radio frequency interference signal multitasking method of claim 4, wherein dynamically optimizing multitasking model parameters according to a multidimensional index comprises: obtaining multidimensional indexes in the model training process, wherein the multidimensional indexes comprise cosine similarity, signal detection accuracy, modulation identification accuracy and stability score, and the stability score is 1/(1+cosine similarity standard deviation); Designing a buffer area for reducing I/O overhead; Analyzing the multi-dimensional index and adjusting the parameters of the multi-task model; and verifying whether the multidimensional index after the parameter adjustment of the multitasking model is optimized, if so, continuing training, and if not, re-analyzing the multidimensional index and optimizing the parameter of the multitasking model.
  8. 8. An unmanned aerial vehicle radio frequency interference signal multitasking device, characterized by comprising: the data processing module is used for constructing a training data set according to the radio frequency interference signals on the acquired power inspection line and preprocessing the training data set; The model construction module is used for constructing a multi-task model based on a signal detection task and a modulation recognition task of the radio frequency interference signal, designing a multi-task computing resource dynamic allocation scheme by combining a task relation, and constructing a multi-objective rewarding function; The training module is used for inputting the preprocessed training data set to train the multi-task model, and calculating the signal detection accuracy, the modulation recognition accuracy and the resource utilization rate according to the output prediction result and the real label; The optimization module is used for dynamically optimizing the parameters of the multi-task model according to the multi-dimensional index to obtain an optimized multi-task model; And the test module is used for inputting the data of the radio frequency interference signals to be tested into the optimized multitask model to obtain the signal detection and modulation recognition results.
  9. 9. A computer readable storage medium having stored thereon a computer program/instruction, which when executed by a processor, implements the steps of the unmanned aerial vehicle radio frequency interference signal multitasking method of any of claims 1-7.
  10. 10. A computer apparatus/device/system comprising: a memory for storing computer programs/instructions; a processor for executing the computer program/instructions to implement the steps of the unmanned aerial vehicle radio frequency interference signal multiplexing method of any of claims 1-7.

Description

Unmanned aerial vehicle radio frequency interference signal multitasking method and device Technical Field The invention belongs to the technical field of multi-task learning, and particularly relates to a method and a device for multi-task processing of radio-frequency interference signals of an unmanned aerial vehicle. Background In the process of inspection of the unmanned aerial vehicle of the high-voltage transmission line, illegal radio frequency interference exists around the transmission line, so that communication between the unmanned aerial vehicle and a ground base station is interrupted, and risks such as inspection data loss and flight control are caused. In order to provide an interference source positioning basis for ground personnel and ensure continuous safety of inspection operation, the unmanned aerial vehicle is required to process multiple tasks such as radio frequency signal detection, modulation identification, spectrum sensing and the like. However, the unmanned aerial vehicle has prominent problems in the airborne environment, including limited CPU computing power, limited memory capacity, sensitive energy consumption caused by battery power supply, and influence of jolt vibration and electromagnetic interference on data stability in the flight process. In the traditional technical system, the multitasking mainly depends on two schemes, and the unmanned airport scene requirement cannot be adapted: Independent modeling of single task, namely independently designing a model architecture for each task, wherein the total quantity of model parameters linearly increases along with the quantity of the tasks although the interference between the tasks can be avoided. In unmanned aerial vehicle airborne environment, this kind of scheme easily leads to the memory to spill over, and unable shared data and characteristic knowledge between the task cause calculated power and battery energy's repetition consumption, and resource utilization is low, is difficult to satisfy unmanned aerial vehicle long-time flight's energy consumption and performance demand. The early multi-task learning technology reduces parameter redundancy by sharing a bottom layer feature extractor, relieves resource pressure to a certain extent, but has three major core bottlenecks in unmanned airport scene suitability: The lack of relationship modeling among tasks leads to high negative migration risk, and the prior art mostly adopts a fixed sharing mechanism and does not design a dynamic relationship analysis scheme aiming at unmanned aerial vehicle task isomerism. For example, signal detection and modulation recognition during unmanned aerial vehicle inspection cause signal intensity fluctuation due to flight altitude change, characteristic demand difference is obvious, a fixed sharing mechanism is easy to cause 'negative migration' -i.e. irrelevant characteristics are transferred among tasks, so that modulation recognition accuracy is reduced, and the unmanned aerial vehicle inspection cannot be avoided through manual adjustment, thereby seriously affecting the multi-task synergistic effect. The characteristic distinguishing degree and the light weight degree are low, the characteristic extraction of the prior scheme is limited to single dimension, and the characteristic extraction of the prior scheme is not combined with the multi-dimension characteristic extraction of time domain, frequency domain, higher order characteristic and the like, so that the characteristic distinguishing degree is low, and simultaneously, the characteristic selection depends on simple variance filtering or single variance filteringRegularization, the general key features among the multitasks cannot be accurately screened, the calculation cost of the model is increased, and the battery energy is greatly consumed. The resource state of the unmanned aerial vehicle airborne environment has volatility, such as the fluctuation of CPU utilization rate along with task load and the continuous decline of battery electric quantity in the flying process. However, the existing multi-task learning technology adopts a static model architecture and a fixed resource allocation strategy, and cannot adjust model parameters or task weights according to resource changes. In addition, the prior art does not design a high-efficiency feedback and iteration mechanism aiming at an unmanned aerial vehicle scene, and performance indexes and resource consumption indexes of model training cannot be returned in real time for model optimization, so that suitability of a model in an unmanned aerial vehicle-mounted environment is continuously reduced, and long-term stable operation is difficult. Disclosure of Invention The invention aims to overcome the defects in the prior art, and provides a method and a device for multitasking radio frequency interference signals of an unmanned aerial vehicle, which can dynamically allocate computing resources and dynamically optimize a mod