Unmanned aerial vehicle radio frequency interference signal multi-task processing method and device

Abstract

The invention discloses a multi-task processing method for radio frequency interference signals of an unmanned aerial vehicle. The multi-task processing method comprises the following steps: constructing a training data set according to the obtained radio frequency interference signals on an electric power inspection line, and preprocessing the training data set; constructing a multi-task model based on a signal detection task and a modulation identification task of the radio frequency interference signal; designing a multi-task computing resource dynamic allocation scheme in combination with the task relationship, and constructing a multi-target reward function; 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; a multi-target reward function is calculated to be maximized, and a trained multi-task model is obtained; dynamically optimizing the multi-task model parameters according to the multi-dimensional indexes to obtain an optimized multi-task model; and inputting to-be-tested radio frequency interference signal data into the optimized multi-task model to obtain detection and modulation results. According to the scheme, computing resources are dynamically allocated, and the model is automatically optimized.

Description

Technical Field

[0001] This invention belongs to the field of multi-task learning technology, and specifically relates to a multi-task processing method and apparatus for radio frequency interference signals of unmanned aerial vehicles. Background Technology

[0002] During drone inspections of high-voltage power transmission lines, illegal radio frequency interference often exists around the lines, causing communication interruptions between the drone and ground base stations, leading to risks such as data loss and flight loss of control. To provide ground personnel with information on interference source location and ensure continuous and safe inspection operations, drones need to simultaneously handle multiple tasks, including radio frequency signal detection, modulation identification, and spectrum sensing. However, the drone's onboard environment presents significant challenges, including limited CPU computing power, restricted memory capacity, energy sensitivity due to battery power, and the impact of turbulence, vibration, and electromagnetic interference on data stability during flight.

[0003] In traditional technology systems, multitasking mainly relies on two types of solutions, neither of which can meet the needs of drone scenarios:

[0004] Single-task independent modeling: Designing a separate model architecture for each task can avoid interference between tasks, but the total number of model parameters increases linearly with the number of tasks. In the UAV airborne environment, this approach is prone to memory overflow, and tasks cannot share data and feature knowledge, resulting in the repeated consumption of computing power and battery energy, low resource utilization, and difficulty in meeting the energy consumption and performance requirements of UAVs for long-term flight.

[0005] Early multi-task learning techniques reduced parameter redundancy by sharing the underlying feature extractor, alleviating resource pressure to some extent. However, they faced three major bottlenecks in adaptability to drone scenarios:

[0006] The lack of task relationship modeling leads to a high risk of negative transfer: Existing technologies mostly adopt fixed sharing mechanisms and do not design dynamic relationship analysis schemes for the heterogeneity of UAV tasks. For example, in signal detection and modulation recognition during UAV inspection, signal strength fluctuates due to changes in flight altitude, and feature requirements vary significantly. Fixed sharing mechanisms are prone to "negative transfer"—that is, irrelevant features are transferred between tasks, causing a decrease in modulation recognition accuracy. This cannot be avoided by manual adjustment and seriously affects the multi-task collaborative effect.

[0007] Low feature discriminative power and lightweight design: Existing solutions are limited to a single dimension for feature extraction, failing to combine multi-dimensional feature extraction such as time domain, frequency domain, and higher-order features, resulting in low feature discriminative power; at the same time, feature selection relies on simple variance filtering or a single feature. Regularization cannot accurately filter common key features across multiple tasks, increasing model computational overhead and consuming a large amount of battery energy.

[0008] Lack of dynamic resource allocation and model optimization mechanisms: The resource status of the UAV airborne environment is volatile, such as CPU utilization fluctuating with task load and battery power continuously decreasing during flight. However, existing multi-task learning techniques use static model architectures and fixed resource allocation strategies, which cannot adjust model parameters or task weights according to resource changes. In addition, existing technologies have not designed efficient feedback and iteration mechanisms for UAV scenarios. The performance indicators and resource consumption indicators of model training cannot be fed back in real time for model optimization, resulting in a continuous decline in the model's adaptability to the UAV airborne environment and making it difficult to run stably for a long time. Summary of the Invention

[0009] Objective: In order to overcome the shortcomings of the existing technology, the present invention provides a multi-task processing method and device for UAV radio frequency interference signals, which can dynamically allocate computing resources and dynamically optimize the model according to the actual changes in airborne environmental resources.

[0010] Technical solution: To solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0011] Firstly, a multi-task processing method for UAV radio frequency interference signals is provided, including:

[0012] A training dataset is constructed and preprocessed based on the radio frequency interference signals obtained from the power line inspection lines.

[0013] A multi-task model is constructed based on signal detection and modulation recognition tasks for radio frequency interference signals; a dynamic allocation scheme for multi-task computing resources is designed by combining task relationships, and a multi-objective reward function is constructed.

[0014] The multi-task model is trained by inputting the preprocessed training dataset, and the signal detection accuracy, modulation recognition accuracy, and resource utilization are calculated based on the output prediction results and the true labels. The multi-objective reward function is maximized to obtain the trained multi-task model.

[0015] The parameters of the multi-task model are dynamically optimized based on multi-dimensional indicators to obtain the optimized multi-task model.

[0016] The RF interference signal data to be tested is input into the optimized multi-task model to obtain signal detection and modulation identification results.

[0017] This solution categorizes task relationships into four classes based on interference levels using cosine similarity, accurately identifying heterogeneity between tasks and avoiding the impact of negative transfer. It provides five weight allocation strategies, selecting different strategies for different task relationships and scenarios such as fluctuations in airborne environmental resources or changes in task priority, thereby dynamically allocating computing resources to adapt to changes in environmental resources.

[0018] By combining a shared backbone network with private branch networks, resource utilization is improved, and energy consumption caused by redundant computing power consumption is reduced. The model can dynamically adjust multi-task model parameters and automatically optimize the model to adapt to the onboard environment, ensuring long-term stable operation of the model.

[0019] In some embodiments, the step of constructing a training dataset based on the acquired radio frequency interference signals on the power inspection line and performing preprocessing includes:

[0020] The radio frequency interference signal data is labeled to obtain tagged radio frequency interference signal data;

[0021] The labeled radio frequency interference signal data is divided proportionally to obtain the training dataset;

[0022] The training dataset is cleaned and normalized.

[0023] Multi-dimensional feature extraction and stepwise feature selection are performed on the processed training dataset to obtain effective feature extraction results.

[0024] In some embodiments, the cleaning process uses the interquartile range method to calculate the difference between the 25th and 75th percentiles of the training data to determine the outlier range. The calculation formula is shown in equation (1):

[0025] (1)

[0026] Where Q3 is the 75th percentile of the training data, Q1 is the 25th percentile of the training data, and IQR is the difference between Q3 and Q1; (The last part, "less than," appears to be a typo and can be omitted.) or greater than The training data was identified as outliers, and the median was used to replace the outliers;

[0027] And / or, the normalization is performed using the min-max normalization method, and the calculation formula is shown in equation (2):

[0028] (2)

[0029] in, X represents the normalized training data, and X represents the original training data. The maximum value of the original training data. This is the minimum value of the original training data.

[0030] And / or, the multi-dimensional feature extraction includes time-domain feature extraction, frequency-domain feature extraction, and higher-order feature extraction;

[0031] The time-domain features include the mean, standard deviation, variance, skewness, and kurtosis of the I component; and the mean, standard deviation, variance, skewness, and kurtosis of the Q component.

[0032] Mean instantaneous amplitude, standard deviation of instantaneous amplitude, maximum instantaneous amplitude, and minimum instantaneous amplitude;

[0033] Instantaneous phase mean and instantaneous phase standard deviation;

[0034] Signal power, peak power, and peak-to-average power ratio;

[0035] The frequency domain features include the spectral centroid, spectral bandwidth, spectral roll-off point, and spectral flatness.

[0036] Mean power spectral density, standard deviation of power spectral density, maximum power spectral density, and power spectral entropy;

[0037] The higher-order features include second-order cumulants c20, second-order cumulants c21, fourth-order cumulants c40, fourth-order cumulants c41, fourth-order cumulants c42, second-order cyclic moments, and fourth-order cyclic moments.

[0038] And / or, the hierarchical feature selection includes first-level selection, second-level selection, and third-level selection; first-level selection is... Norm feature selection, secondary selection is mutual information feature selection, and tertiary selection is variance threshold feature selection.

[0039] After stepwise feature selection, the following 17-dimensional effective feature extraction results were obtained:

[0040] Time-domain characteristics include the standard deviation and kurtosis of the I component; and the standard deviation and kurtosis of the Q component.

[0041] Instantaneous amplitude mean, instantaneous amplitude standard deviation, instantaneous phase standard deviation, signal power and peak-to-average power ratio;

[0042] Frequency domain features include spectral centroid, spectral flatness, mean power spectral density, and power spectral entropy;

[0043] Higher-order features include second-order cumulants c20, fourth-order cumulants c40, fourth-order cumulants c42, and second-order cyclic moments.

[0044] Multi-dimensional feature extraction enables the model to perceive the environment more comprehensively, improving the accuracy of the model's computational analysis; step-by-step feature selection progresses layer by layer, filtering out redundant features and fully retaining key features, reducing computational resource consumption and memory usage.

[0045] In some embodiments, the construction of a multi-task model based on the signal detection task and modulation identification task of the radio frequency interference signal includes:

[0046] Construct a shared backbone network and a private branch network; the shared backbone network includes a shared Conv1D layer, a shared GlobalAveragePooling1D layer, and a shared Dense layer; the private branch network includes a signal detection branch network and a modulation identification branch network; the signal detection branch network and the modulation identification branch network each include a private Dense layer.

[0047] Define the initial values ​​of the parameters and the loss function for the multi-task model;

[0048] The multi-task model parameters include the number of shared Conv1D layers in the shared backbone network, the number of private Dense layers in the private branch network, the weights of the signal detection task, the weights of the modulation recognition task, and the optimizer learning rate β. In the initial multi-task model parameters, the shared backbone network includes 4 Conv1D layers (adjustable range 2-6 layers); the private branch network includes 2 Dense layers (adjustable range 2-3 layers); the weight of the signal detection task is 1.0 (adjustable range 0.7-1.5); the weight of the modulation recognition task is 1.2 (adjustable range 0.7-1.5); the learning rate β is 0.001 (adjustable range 0.000001-0.001); the batch size is 128 (adjustable range 64-256); and the random dropout rate is 0.3 (adjustable range 0.2-0.4). The optimizer used is the Adam optimizer, with an initial learning rate β of 0.001.

[0049] The formula for calculating the loss function is shown in equation (3):

[0050] (3)

[0051] in, This is the total loss value. For signal detection task weights, This represents the loss value for the signal detection task. To modulate the weights for the recognition task, To modulate the loss value for the recognition task.

[0052] In some embodiments, the design of a dynamic allocation scheme for multi-task computing resources based on task relationships and the construction of a multi-objective reward function include:

[0053] The cosine similarity of gradient vectors for different tasks is calculated, and the formula for calculating the cosine similarity is shown in equation (4):

[0054] (4)

[0055] Where Sim represents the cosine similarity. For the weight gradient of the signal detection task, To modulate the gradient of the recognition task weights;

[0056] Task relationships are categorized into four types based on cosine similarity, thus grouping the tasks accordingly.

[0057] Task relationship classification includes: The task relationship is set to positive migration. The task relationship is set to neutral. The task relationship is set as weak negative migration. The task relationship is set as strong negative migration.

[0058] In some embodiments, signal detection accuracy, modulation recognition accuracy, and resource utilization are discretized as follows: The state space;

[0059] Five weight allocation strategies are designed to dynamically allocate computing resources; the weight allocation strategies include bias detection, balanced allocation, bias identification, uniform enhancement, and uniform reduction.

[0060] The bias detection task is weighted at 1.5, and the modulation recognition task is weighted at 0.8.

[0061] The weight of the detection task is set to 1.0, and the weight of the modulation recognition task is set to 1.0.

[0062] The bias recognition task is assigned a weight of 0.8, and the modulation recognition task is assigned a weight of 1.5.

[0063] The uniform enhancement detection task weight is set to 1.2, and the modulation recognition task weight is set to 1.2.

[0064] The uniform attenuation detection task weight is set to 0.7, and the modulation recognition task weight is set to 0.7.

[0065] A multi-objective reward function is constructed to improve the rationality of the strategy selection for weight allocation. The calculation formula of the multi-objective reward function is shown in equation (5):

[0066] (5)

[0067] in, For the overall reward value, For signal detection accuracy, To modulate the recognition accuracy, Resource utilization rate; resource utilization rate is a comprehensive normalized indicator that includes the drone's CPU utilization rate, GPU memory usage rate, battery power, and memory usage rate.

[0068] An ε-greedy strategy is adopted to combine the exploration of new strategy selection with the execution of historically optimal strategy selection. The new strategy selection is evaluated by the multi-objective reward function to avoid getting trapped in local optima.

[0069] The ε-greedy strategy sets the learning rate α to 0.1 and the discount factor γ to 0.9. The learning rate α and the discount factor γ can be dynamically adjusted according to the model training progress.

[0070] The learning rate α and discount factor γ can be fixed or dynamically adjusted. In this embodiment, dynamic adjustment can reduce the number of optimization iterations, thereby shortening the training time.

[0071] In some embodiments, the design of a dynamic allocation scheme for multi-task computing resources based on task relationships and the construction of a multi-objective reward function further include: calculating the clustering loss of task parameters to quantify the rationality of task grouping;

[0072] The task parameters include the learnable weights and biases of the shared GlobalAveragePooling1D layer and the shared Dense layer of the shared backbone network, as well as the learnable weights and biases of the private Dense layer and the output layer of the private branch network.

[0073] The formula for calculating the clustering loss is shown in equation (6):

[0074] (6)

[0075] in, Clustering loss value for task parameters The weighting coefficients are for the global mean loss. For global mean loss, These are the inter-cluster variance weighting coefficients. For inter-cluster variance, The intra-cluster variance weighting coefficient. Let V be the variance within the cluster.

[0076] This embodiment quantifies and verifies the rationality of task grouping after classifying task relationships, accurately assesses whether the task relationship classification is reasonable, and is beneficial to the subsequent design of weight allocation strategies and the selection of optimization schemes.

[0077] In some embodiments, the dynamic optimization of multi-task model parameters based on multi-dimensional indicators includes:

[0078] Obtain multi-dimensional metrics during model training, including cosine similarity, signal detection accuracy, modulation recognition accuracy, and stability score; the stability score is 1 / (1+cosine similarity standard deviation).

[0079] Design a buffer to reduce I / O overhead; the design of the buffer can reduce the number of onboard storage read and write operations, thereby extending storage life, ensuring smooth acquisition of feedback data, and shortening the model optimization cycle.

[0080] Analyze the multi-dimensional indicators and adjust the parameters of the multi-task model:

[0081] Verify whether the multi-dimensional indicators after adjusting the parameters of the multi-task model have been optimized. If they have been optimized, continue training; otherwise, re-analyze the multi-dimensional indicators and optimize the parameters of the multi-task model.

[0082] The radio frequency interference signal to be tested is input into the multi-task model for calculation. The optimal weight allocation strategy is selected based on resource data such as CPU utilization, GPU memory usage, battery level, and memory usage. The system resource consumption data for this calculation is also recorded for subsequent multi-task model parameter optimization.

[0083] Secondly, a multi-task processing device for UAV radio frequency interference signals is provided, comprising:

[0084] Data processing module: used to construct a training dataset and perform preprocessing based on the acquired radio frequency interference signals on the power inspection lines;

[0085] Model building module: used to build multi-task models for signal detection and modulation recognition tasks based on radio frequency interference signals; design a dynamic allocation scheme for multi-task computing resources based on task relationships; and construct a multi-objective reward function.

[0086] Training module: Used to train the multi-task model by inputting the preprocessed training dataset, and to calculate the signal detection accuracy, modulation recognition accuracy and resource utilization based on the output prediction results and the true labels; calculates the maximization of the multi-objective reward function to obtain the trained multi-task model;

[0087] Optimization module: Used to dynamically optimize the parameters of the multi-task model based on multi-dimensional indicators to obtain the optimized multi-task model;

[0088] Test module: Used to input the RF interference signal data to be tested into the optimized multi-task model to obtain signal detection and modulation identification results.

[0089] Thirdly, a computer-readable storage medium is provided, on which a computer program / instruction is stored, which, when executed by a processor, implements the steps of the UAV radio frequency interference signal multitasking method described in the first aspect.

[0090] Fourthly, a computer device / equipment / system is provided, comprising:

[0091] Memory, used to store computer programs / instructions;

[0092] A processor for executing the computer program / instructions to implement the steps of the UAV radio frequency interference signal multitasking method described in the first aspect.

[0093] Beneficial effects: The UAV radio frequency interference signal multi-task processing method and apparatus provided by the present invention have the following advantages:

[0094] 1. Five weight allocation strategies are provided to match the appropriate weight allocation strategy for different task adaptation scenarios, thereby dynamically allocating computing resources and adapting to changes in the resources of the UAV airborne environment; at the same time, the multi-task model parameters can be dynamically adjusted with changes in environmental resources, thereby automatically optimizing the model to adapt to the airborne environment in the long term and ensuring the long-term stable operation of the model.

[0095] 2. Construct a shared backbone network and private branch networks to improve resource utilization and reduce energy consumption; calculate cosine similarity to classify tasks into four categories, accurately identify task heterogeneity, and avoid the impact of negative transfer; multi-dimensional feature extraction and hierarchical feature selection screen out redundant features, reducing computational resource overhead and memory usage while ensuring the complete preservation of key features, thus achieving lightweight design. Attached Figure Description

[0096] Figure 1 This is a flowchart illustrating the overall process of the model training method according to an embodiment of the present invention.

[0097] Figure 2 This is a schematic diagram of the multi-task model architecture according to an embodiment of the present invention;

[0098] Figure 3 This is a schematic diagram of the dynamic allocation logic of computing resources according to an embodiment of the present invention;

[0099] Figure 4 This is a task interference level classification diagram according to an embodiment of the present invention;

[0100] Figure 5 This is a flowchart illustrating the parameter optimization process for a multi-task model according to an embodiment of the present invention. Detailed Implementation

[0101] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0102] The present invention will be further described below with reference to specific embodiments.

[0103] Example 1: As Figure 1 As shown, a multi-task processing method for UAV radio frequency interference signals includes:

[0104] A training dataset is constructed and preprocessed based on the radio frequency interference signals obtained from the power line inspection lines.

[0105] A multi-task model is constructed based on signal detection and modulation recognition tasks for radio frequency interference signals; a dynamic allocation scheme for multi-task computing resources is designed by combining task relationships, and a multi-objective reward function is constructed.

[0106] The multi-task model is trained by inputting the preprocessed training dataset, and the signal detection accuracy, modulation recognition accuracy, and resource utilization are calculated based on the output prediction results and the true labels. The multi-objective reward function is maximized to obtain the trained multi-task model.

[0107] The parameters of the multi-task model are dynamically optimized based on multi-dimensional indicators to obtain the optimized multi-task model.

[0108] The RF interference signal data to be tested is input into the optimized multi-task model to obtain signal detection and modulation identification results.

[0109] Specifically as follows:

[0110] The UAV-borne radio frequency receiver and monitoring module acquired 220,000 radio frequency interference signal samples from power line inspection lines. Each sample simultaneously recorded one set of resource data, including CPU utilization, GPU memory usage, battery level, and memory usage (a total of 220,000 sets of resource data). The collected data covered suburban lines, mountainous lines, and lines surrounding industrial areas to cover different interference source densities. The samples covered 11 modulation types and 20 signal-to-noise ratio (SNR) levels, with 1000 samples per combination, ensuring a balanced data distribution.

[0111] The 11 modulation types include octal phase shift keying, double-sideband amplitude modulation, single-sideband amplitude modulation, binary phase shift keying, continuous phase shift keying, Gaussian shift keying, quaternary pulse amplitude modulation, hexadecimal quadrature amplitude modulation, 64-ary quadrature amplitude modulation, quaternary phase shift keying, and broadband frequency modulation.

[0112] The 20 SNR levels include -20dB, -18dB, -16dB, -14dB, -12dB, -10dB, -8dB, -6dB, -4dB, -2dB, 0dB, 2dB, 4dB, 6dB, 8dB, 10dB, 12dB, 14dB, 16dB, and 18dB.

[0113] Annotate the radio frequency interference signal data:

[0114] Signal detection tag: dynamically generated based on SNR, 0 = no radio frequency interference, 1 = radio frequency interference exists.

[0115] Modulation identification labels: One-Hot encoding is performed on 11 modulation types to generate 11-dimensional vector labels, ensuring that the proportion of each type of modulation sample is consistent in the training / validation / test sets. Interference types are determined by tracing the source of the signal.

[0116] Hierarchical segmentation: The sampled data is divided into training, validation, and test sets in an 8:1:1 ratio. The stratify parameter is used to control the segmentation ratio based on the modulation type label to ensure that the data distribution of each set is consistent with the original dataset, providing a reliable data foundation for model training and evaluation.

[0117] Data cleaning uses the interquartile range method to calculate the difference between the 25th and 75th quantiles of the data to determine the range of outliers. The calculation formula is shown in equation (1):

[0118] (1)

[0119] Where Q3 is the 75th quantile of the data, Q1 is the 25th quantile of the data, and IQR is the difference between Q3 and Q1; (The last part, "less than," appears to be a typo and can be omitted.) or greater than The data was identified as outliers, with a total of 62,517 outliers detected, representing an outlier rate of 0.8079%. The median was used to replace the outliers to avoid interference from extreme values ​​in subsequent feature extraction and model training, resulting in a more even data distribution.

[0120] Data normalization is performed using the min-max normalization method, and the calculation formula is shown in equation (2):

[0121] (2)

[0122] in, Here, X represents the normalized data, and X represents the original data. The maximum value of the original data. This represents the minimum value of the original data. After normalization, the difference in feature variance is verified. This ensures that features of different dimensions contribute equally to model training.

[0123] Multidimensional feature extraction includes time-domain feature extraction, frequency-domain feature extraction, and higher-order feature extraction;

[0124] The time-domain features include the mean, standard deviation, variance, skewness, and kurtosis of the I component; and the mean, standard deviation, variance, skewness, and kurtosis of the Q component.

[0125] Mean instantaneous amplitude, standard deviation of instantaneous amplitude, maximum instantaneous amplitude, and minimum instantaneous amplitude;

[0126] Instantaneous phase mean and instantaneous phase standard deviation;

[0127] Signal power, peak power, and peak-to-average power ratio;

[0128] The frequency domain features include the spectral centroid, spectral bandwidth, spectral roll-off point, and spectral flatness.

[0129] Mean power spectral density, standard deviation of power spectral density, maximum power spectral density, and power spectral entropy;

[0130] The higher-order features include second-order cumulants c20, second-order cumulants c21, fourth-order cumulants c40, fourth-order cumulants c41, fourth-order cumulants c42, second-order cyclic moments, and fourth-order cyclic moments.

[0131] The hierarchical feature selection includes first-level selection, second-level selection, and third-level selection; first-level selection is... Norm feature selection, based on sparse regularization to optimize the objective function, filters features strongly related to the task, initially reducing dimensionality. Secondary selection involves mutual information feature selection, calculating the mutual information value between features and labels, selecting the top 30% of features with the highest information gain to capture non-linear relationships. Tertiary selection uses variance thresholding to filter redundant features with variance less than 0.01. Ultimately, 17 effective features are selected from 34 features, achieving a feature retention rate of 50%, reducing computational complexity while maintaining the integrity of key information. The resulting 17 effective features after this step-by-step feature selection process include:

[0132] Time-domain characteristics include the standard deviation and kurtosis of the I component; and the standard deviation and kurtosis of the Q component.

[0133] Instantaneous amplitude mean, instantaneous amplitude standard deviation, instantaneous phase standard deviation, signal power and peak-to-average power ratio;

[0134] Frequency domain features include spectral centroid, spectral flatness, mean power spectral density, and power spectral entropy;

[0135] Higher-order features include second-order cumulants c20, fourth-order cumulants c40, fourth-order cumulants c42, and second-order cyclic moments.

[0136] like Figure 2 As shown, a multi-task model is constructed, including:

[0137] Shared backbone network construction:

[0138] Construct a shared Conv1D layer (4 layers): The first two layers use 3×1 convolutional kernels, Rectified Linear Unit (ReLU) activation, and Same Padding, with 64 output channels. Each layer is followed by MaxPooling1D (pooling kernel 2), and the output shapes are (64,64) and (64,32) respectively. The last two layers also use 3×1 convolutional kernels, ReLU activation, and Same Padding, with 128 output channels. Each layer is followed by MaxPooling1D (pooling kernel 2), and the output shapes are (128,16) and (128,8) respectively.

[0139] Construct a shared GlobalAveragePooling1D layer: Perform global average pooling on the output of the last Conv1D layer to compress the feature dimensions;

[0140] Construct a shared Dense layer: containing 256 neurons, using the ReLU activation function, to provide a general feature representation for private branches of subsequent tasks.

[0141] Building a private branch network:

[0142] Construct a signal detection branch: containing one private Dense layer (64 neurons, ReLU activation) and one output layer (1 neuron, Sigmoid activation), adapted for binary classification tasks, outputting the probability of signal presence;

[0143] Construct a modulation recognition branch: containing one private Dense layer (64 neurons, ReLU activation) and one output layer (11 neurons, Softmax activation), adapted to 11 classification tasks, outputting the probability distribution of each modulation type.

[0144] The loss function is defined and calculated as shown in equation (3):

[0145] (3)

[0146] in, This is the total loss value. For signal detection task weights, This represents the loss value for the signal detection task. To modulate the weights for the recognition task, To modulate the loss value for the recognition task.

[0147] Initial multi-task model parameters are loaded: 4 shared Conv1D layers, 2 private Dense layers, batch size of 128, and Dropout rate of 0.3. The Adam optimizer is used with an initial learning rate β of 0.001, which can be dynamically adjusted based on training stability.

[0148] Set a random seed (np.random.seed (42)) to ensure the reproducibility of the experiment.

[0149] Training is divided into 10 epochs, with a visualization interval of 2 epochs, and 128 samples are processed in each batch;

[0150] After each batch of data undergoes data cleaning, normalization, feature extraction, and feature selection, it is input into the multi-task model. The processing time for a single batch is [not specified]. .

[0151] like Figure 3 As shown, dynamic computing resource allocation based on Q-learning includes:

[0152] The cosine similarity between the gradient vectors of the signal detection task and the modulation recognition task is calculated for every 10 batches, and the calculation formula is shown in Equation (4):

[0153] (4)

[0154] Where Sim represents the cosine similarity. For the weight gradient of the signal detection task, To modulate the gradient of the recognition task weights;

[0155] like Figure 4 As shown, based on cosine similarity, task relationships are divided into four levels according to interference level: The task relationships are set as positive migration (accounting for 60%). The task relationship is set as neutral (accounting for 37%). The task relationship is set as weak negative migration (accounting for 2.5%). The task relationship is set as strong negative migration (accounting for 0.5%), and the judgment result provides a basis for the allocation of computing resources.

[0156] Every 5 epochs, the clustering loss of the task parameters is calculated to quantify the rationality of task grouping. The task parameters include learnable weights and biases of the shared backbone network GlobalAveragePooling1D layer and shared Dense layer, as well as learnable weights and biases of the private branch network's private Dense layer and output layer. The input feature dimension is determined by the output of the GlobalAveragePooling1D layer. The weight shape of the shared Dense layer is (128, 256), and the bias shape is (256,), with each neuron corresponding to a bias value. The weight shape of the private Dense layer is (256, 128), and the bias shape is (128,), initially all 0. The weight shape of the output layer is (128, 1), and the bias shape is (1,), initially all 0.

[0157] The formula for calculating the clustering loss is shown in equation (5):

[0158] (5)

[0159] in, Clustering loss value for task parameters The weighting coefficients are for the global mean loss. For global mean loss, These are the inter-cluster variance weighting coefficients. For inter-cluster variance, The intra-cluster variance weighting coefficient. Let V be the variance within the cluster.

[0160] Define the state space: Discretize the detection accuracy, modulation accuracy, and resource utilization as follows: The three-dimensional state;

[0161] Define the action space: Design five weight allocation strategies, including biased detection (detection task weight is 1.5, modulation recognition task weight is 0.8), balanced allocation (balanced allocation sets detection task weight to 1.0, modulation recognition task weight to 1.0), biased recognition (biased recognition sets detection task weight to 0.8, modulation recognition task weight to 1.5), uniform enhancement (uniform enhancement sets detection task weight to 1.2, modulation recognition task weight to 1.2), and uniform reduction (uniform reduction sets detection task weight to 0.7, modulation recognition task weight to 0.7). These strategies cover the main computing resource allocation scenarios and are matched according to task relationships and actual application scenario requirements (such as environmental resource fluctuations and changes in task priority).

[0162] A multi-objective reward function is constructed to improve the rationality of the strategy selection for weight allocation. The calculation formula of the reward function is shown in equation (6):

[0163] (6)

[0164] in, For the overall reward value, For signal detection accuracy, To modulate the recognition accuracy, Resource utilization rate is a comprehensive normalized indicator that includes the drone's CPU utilization rate, GPU memory usage rate, battery power, and memory usage rate; the reward function encourages high task performance while avoiding resource waste.

[0165] An ε-greedy strategy (ε=0.1) is employed to balance exploration and exploitation. The historically optimal matching scheme is selected with a 90% probability, while a weight allocation strategy is randomly selected with a 10% probability to explore new matching schemes. These new schemes are further evaluated by the reward function to avoid getting trapped in local optima. A learning rate α of 0.1 is set to control the policy update step size, and a discount factor γ of 0.9 is used to balance current and future rewards, ensuring the stability of the optimization process. The learning rate α and discount factor γ can be dynamically adjusted according to the model training progress. The Q-value is updated iteratively; after 100 iterations, the Q-table converges, and the hit rate of the optimal solution for the weight allocation strategy is determined. This ensures that the GPU memory usage remains stable at 75%-85%.

[0166] like Figure 5 As shown, training metrics are collected every 50 batches, including total loss, accuracy, precision, recall, and F1 score (including macro / micro average), with I / O overhead reduced by using a 50-record buffer.

[0167] Each epoch collects validation metrics, including total loss, accuracy, precision, recall, and F1 score, which are used as evaluation metrics for model performance and written to a CSV file in real time.

[0168] The system collects system metrics every second, including CPU utilization, GPU memory usage, battery level, memory usage, and I / O throughput, to ensure real-time monitoring of the airborne environment's resource status.

[0169] Using a sliding window method, the most recent 100 training records and 10 validation records are analyzed to calculate signal detection accuracy, modulation recognition accuracy, cosine similarity, stability score, and performance bottleneck (total loss decrease over 3 consecutive epochs). Resource bottleneck (GPU memory usage) ) and interference bottleneck ( ).

[0170] Multi-task model parameter adjustment:

[0171] Negative migration response: At the same time, the shared Conv1D layer was reduced from 4 layers to 2 layers, and the private Dense layer was increased from 2 layers to 3 layers to avoid negative migration effects;

[0172] Positive migration utilization: At the same time, the shared Conv1D layer was increased from 4 layers to 5 layers, promoting knowledge sharing between tasks;

[0173] Performance balance: signal detection accuracy At that time, the weight increased from 1.0 to 1.1; modulation recognition accuracy. At that time, the weight increased from 1.2 to 1.3;

[0174] Stability optimization: Stability score At that time, the learning rate β decreased from 0.001 to 0.0009, controlling the oscillation amplitude of the weights. ;

[0175] CPU / GPU bottlenecks: The final effective feature dimension was reduced from 17 to 12; the weight of the shared layer was reduced from 32-bit floating-point to 16-bit floating-point; the batch size was adjusted from 128 to 64; and the number of neurons in the private Dense layer was reduced from 64 to 32.

[0176] Optimization effect verification:

[0177] After optimization, the modulation recognition accuracy improved from 44.5% to 46.1%, and the key feature retention rate increased. ;

[0178] If the verification passes, training continues; if it fails, the feedback data is re-analyzed to ensure that an iteration is completed every 5 epochs, forming a closed loop of "collection-analysis-optimization-verification".

[0179] The multi-task model is output and embedded in the UAV computing module. The radio frequency interference signal to be tested is input into the multi-task model for calculation. Based on resource data such as CPU utilization, GPU memory usage, battery level, and memory usage, the optimal weight allocation strategy is selected (e.g., a uniform attenuation strategy is chosen to reduce computational accuracy when the battery is low). Simultaneously, the system resource consumption data for this calculation is recorded for subsequent multi-task model parameter optimization.

[0180] Example 2: A multi-task processing device for UAV radio frequency interference signals, comprising:

[0181] Data processing module: used to construct a training dataset and perform preprocessing based on the acquired radio frequency interference signals on the power inspection lines;

[0182] Model building module: used to build multi-task models for signal detection and modulation recognition tasks based on radio frequency interference signals; design a dynamic allocation scheme for multi-task computing resources based on task relationships; and construct a multi-objective reward function.

[0183] Training module: Used to train the multi-task model by inputting the preprocessed training dataset, and to calculate the signal detection accuracy, modulation recognition accuracy and resource utilization based on the output prediction results and the true labels; calculates the maximization of the multi-objective reward function to obtain the trained multi-task model;

[0184] Optimization module: Used to dynamically optimize the parameters of the multi-task model based on multi-dimensional indicators to obtain the optimized multi-task model.

[0185] Test module: Used to input the RF interference signal data to be tested into the optimized multi-task model to obtain signal detection and modulation identification results.

[0186] Example 3: A computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the UAV radio frequency interference signal multitasking method described in Example 1.

[0187] Example 4: A computer device / equipment / system, comprising:

[0188] Memory, used to store computer programs / instructions;

[0189] A processor is used to execute the computer program / instructions to implement the steps of the UAV radio frequency interference signal multitasking method described in Embodiment 1.

[0190] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-task processing method for radio frequency interference signals from unmanned aerial vehicles (UAVs), characterized in that, include: A training dataset is constructed and preprocessed based on the radio frequency interference signals obtained from the power line inspection lines. A multi-task model is constructed based on signal detection and modulation identification tasks for radio frequency interference signals; A dynamic allocation scheme for multi-task computing resources is designed based on task relationships, and a multi-objective reward function is constructed. The multi-task model is trained by inputting the preprocessed training dataset, and the signal detection accuracy, modulation recognition accuracy, and resource utilization are calculated based on the output prediction results and the true labels. Calculate and maximize the multi-objective reward function to obtain the trained multi-task model; The parameters of the multi-task model are dynamically optimized based on multi-dimensional indicators to obtain the optimized multi-task model; the radio frequency interference signal data to be tested is input into the optimized multi-task model to obtain the signal detection and modulation identification results.

2. The multi-task processing method for UAV radio frequency interference signals according to claim 1, characterized in that, The step of constructing a training dataset based on the acquired radio frequency interference signals on the power inspection line and performing preprocessing includes: The radio frequency interference signal data is labeled to obtain tagged radio frequency interference signal data; The labeled radio frequency interference signal data is divided proportionally to obtain the training dataset; The training dataset is cleaned and normalized. Multi-dimensional feature extraction and stepwise feature selection are performed on the processed training dataset to obtain effective feature extraction results.

3. The multi-task processing method for UAV radio frequency interference signals according to claim 2, characterized in that, The cleaning process uses the interquartile range method to calculate the difference between the 25th and 75th percentiles of the training data to determine the outlier range. The calculation formula is shown in equation (1): （1） Where Q3 is the 75th percentile of the training data, Q1 is the 25th percentile of the training data, and IQR is the difference between Q3 and Q1; (The last part, "less than," appears to be a typo and can be omitted.) or greater than The training data was identified as outliers; the median was used to replace the outliers. And / or, the normalization is performed using the min-max normalization method, and the calculation formula is shown in equation (2): （2） in, X represents the normalized training data, and X represents the original training data. The maximum value of the original training data. The minimum value of the original training data; And / or, the multi-dimensional feature extraction includes time-domain feature extraction, frequency-domain feature extraction, and higher-order feature extraction; The time-domain features include the mean, standard deviation, variance, skewness, and kurtosis of the I component; and the mean, standard deviation, variance, skewness, and kurtosis of the Q component. Mean instantaneous amplitude, standard deviation of instantaneous amplitude, maximum instantaneous amplitude, and minimum instantaneous amplitude; Instantaneous phase mean and instantaneous phase standard deviation; Signal power, peak power, and peak-to-average power ratio; The frequency domain features include the spectral centroid, spectral bandwidth, spectral roll-off point, and spectral flatness. Mean power spectral density, standard deviation of power spectral density, maximum power spectral density, and power spectral entropy; The higher-order features include second-order cumulants c20, second-order cumulants c21, fourth-order cumulants c40, fourth-order cumulants c41, fourth-order cumulants c42, second-order cyclic moments, and fourth-order cyclic moments. And / or, the hierarchical feature selection includes first-level selection, second-level selection, and third-level selection; first-level selection is... Norm feature selection, secondary selection is mutual information feature selection, and tertiary selection is variance threshold feature selection.

4. The multi-task processing method for UAV radio frequency interference signals according to claim 1, characterized in that, The multi-task model for signal detection and modulation identification tasks based on radio frequency interference signals includes: Construct a shared backbone network and a private branch network; the shared backbone network includes a shared Conv1D layer, a shared GlobalAveragePooling1D layer, and a shared Dense layer; the private branch network includes a signal detection branch network and a modulation identification branch network; the signal detection branch network and the modulation identification branch network each include a private Dense layer. Define the initial values ​​of the parameters and the loss function for the multi-task model; The multi-task model parameters include the number of shared Conv1D layers in the shared backbone network, the number of private Dense layers in the private branch network, the weight of the signal detection task, the weight of the modulation recognition task, and the optimizer learning rate β. The formula for calculating the loss function is shown in equation (3): （3） in, This is the total loss value. For signal detection task weights, This represents the loss value for the signal detection task. To modulate the weights for the recognition task, To modulate the loss value for the recognition task.

5. The multi-task processing method for UAV radio frequency interference signals according to claim 1, characterized in that, The design of a dynamic allocation scheme for multi-task computing resources based on task relationships, and the construction of a multi-objective reward function, include: The cosine similarity of gradient vectors for different tasks is calculated using the formula shown in equation (4): （4） Where Sim represents the cosine similarity. For the weight gradient of the signal detection task, To modulate the gradient of the recognition task weights; Task relationships are categorized into four types based on cosine similarity, thus grouping the tasks accordingly. Task relationship classification includes: The task relationship is set to positive migration. The task relationship is set to neutral. The task relationship is set as weak negative migration. The task relationship is set as strong negative migration; The signal detection accuracy, modulation recognition accuracy, and resource utilization are discretized into... The state space; Five weight allocation strategies are designed to dynamically allocate computing resources; the weight allocation strategies include bias detection, balanced allocation, bias identification, uniform enhancement, and uniform reduction. A multi-objective reward function is constructed to improve the rationality of the strategy selection for weight allocation. The calculation formula of the multi-objective reward function is shown in equation (5): （5） in, For the overall reward value, For signal detection accuracy, To modulate the recognition accuracy, Resource utilization rate; resource utilization rate is a comprehensive normalized indicator that includes the drone's CPU utilization rate, GPU memory usage rate, battery power, and memory usage rate. An ε-greedy strategy is adopted to combine the exploration of new strategy selection with the execution of historically optimal strategy selection. The new strategy selection is evaluated by the multi-objective reward function to avoid getting trapped in local optima. The ε-greedy strategy sets the learning rate α to 0.1 and the discount factor γ to 0.

9. The learning rate α and the discount factor γ can be dynamically adjusted according to the model training progress.

6. The multi-task processing method for UAV radio frequency interference signals according to claim 5, characterized in that, The design of a dynamic allocation scheme for multi-task computing resources by combining task relationships and the construction of a multi-objective reward function also include: Calculate the clustering loss of task parameters to quantify the rationality of task grouping; The task parameters include the learnable weights and biases of the shared Global Average Pooling 1D layer and the shared Dense layer of the shared backbone network, as well as the learnable weights and biases of the private Dense layer and the output layer of the private branch network. The formula for calculating the clustering loss is shown in equation (6): （6） in, The clustering loss value of the task parameters. The weighting coefficients are for the global mean loss. For global mean loss, These are the inter-cluster variance weighting coefficients. For inter-cluster variance, The intra-cluster variance weighting coefficient. Let V be the variance within the cluster.

7. The multi-task processing method for UAV radio frequency interference signals according to claim 4, characterized in that, The dynamic optimization of multi-task model parameters based on multi-dimensional indicators includes: Obtain multi-dimensional metrics during model training, including cosine similarity, signal detection accuracy, modulation recognition accuracy, and stability score; the stability score is 1 / (1+cosine similarity standard deviation). Design a buffer to reduce I / O overhead; Analyze the multi-dimensional indicators and adjust the parameters of the multi-task model; Verify whether the multi-dimensional indicators after adjusting the parameters of the multi-task model have been optimized. If they have been optimized, continue training; otherwise, re-analyze the multi-dimensional indicators and optimize the parameters of the multi-task model.

8. A multi-task processing device for radio frequency interference signals of unmanned aerial vehicles, characterized in that, include: Data processing module: used to construct a training dataset and perform preprocessing based on the acquired radio frequency interference signals on the power inspection lines; Model building module: used to build multi-task models for signal detection and modulation recognition tasks based on radio frequency interference signals; A dynamic allocation scheme for multi-task computing resources is designed based on task relationships, and a multi-objective reward function is constructed. Training module: Used to train the multi-task model by inputting the preprocessed training dataset, and to calculate the signal detection accuracy, modulation recognition accuracy and resource utilization based on the output prediction results and the true labels; Calculate and maximize the multi-objective reward function to obtain the trained multi-task model; Optimization module: Used to dynamically optimize the parameters of the multi-task model based on multi-dimensional indicators to obtain the optimized multi-task model; Test module: Used to input the RF interference signal data to be tested into the optimized multi-task model to obtain signal detection and modulation identification results.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the UAV radio frequency interference signal multitasking method according to any one of claims 1-7.

10. A computer device / equipment / system, characterized in that, include: Memory, used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of the UAV radio frequency interference signal multitasking method according to any one of claims 1-7.

Images