When large-scale data is labeled artificially, labeling errors are easy to occur, which leads to the existence of noise labels in data sets, and further affects the generalization of deep neural network models. The sample selection mechanism in the existing research methods such as Co-teaching makes the noise samples easy to flow into the selected clean label sample subset, and it is difficult to control the deep neural network model's fitting to the selected clean sample subset in training. Therefore, this study presents a novel algorithm that improves upon Co-teaching. In this method, two regularization losses are added to prevent the model from placing too much trust in a single class and falling into a local optimal solution respectively. Additionally, the introduction of high learning rate attenuation training method makes the model more inclined to learn clean label sample features in the initial training to get better model parameters. Compared with the results of Co-teaching, the performance of the proposed model is improved on MNIST, CIFAR-10 synthetic noise data set and Animal10N realistic data set under 20% and 50% symmetric noise and 45% asymmetric noise environment.

With the increasingly mature communication networks in the fields of air, space and ground, cross-domain heterogeneous converged technology has become an important direction for the integrated development of future communication networks. Driven by the demand for cross-domain heterogeneous in the converged network of air, space and ground, this study aims to solve the problem of low spectrum resource utilization in heterogeneous networks. It uses deep reinforcement learning method to establish a heterogeneous converged network system model and designs intelligent agent access algorithm with fair scale. The system throughput is selected as the maximization objective. The communication network standards that meet the characteristics of air, space and ground integration are selected and corresponding access protocols are extracted. Non-dimensional channel parameters are set according to the principle of fairness and simulation scenarios are established. Multiple comparison strategies are introduced in the simulation to statistically analyze system throughput, collision rate, utilization rate and channel selection ratio. The simulation results show that the system throughput of cross-domain heterogeneous fusion network is increased by more than 60%, system channel utilization efficiency is increased by 20%, and the collision rate of service packets is maintained at 10%, which verifies the adaptability of the algorithm to different business scenarios.

Hybrid fuzzy testing combines fuzzy testing and symbolic execution, uses fuzzy testing to explore the path in the program, and uses symbolic execution to solve complex constraints that are difficult to break through fuzzy testing. However, the existing hybrid fuzzy testing has not considered the task cooperation between the two technologies and the solution benefits of symbolic execution when selecting the solution target of symbolic execution. To solve these problems, a hybrid fuzzy testing method based on multivariate seed selection is proposed.The program control flow diagram is used to analyze the program state and quantify the ability of seeds to discover the path. The seeds thatare difficult to explore the new path in fuzzy testingare solved by symbolic execution, so as to form task cooperation between them. The ability to use target-point oriented ideas to quantify seed mining vulnerabilities makes symbolic execution solutions more likely to find the seeds of vulnerabilities.The experimental results show that compared with the existing hybrid fuzzy testing work, the overall path discovery number of the proposed method increases by 8.35% and the overall vulnerability discovery number increases by 28.69%.

In foggy weather, the collected pictures have the problems of reduced clarity and color distortion. In order to obtain haze-free images with high quality, a hybrid dark channel prior algorithm is proposed in this study. The proposed algorithm employs Retinex algorithm to remove the interference of the illumination component. The variant chicken swarm optimization algorithm is used to obtain the guidance image required by the guided filter to optimize the atmospheric transmittance. The improved dark channel prior algorithm is used to obtain the fog removal image. Compared with other dark channel prior defogging algorithms, the mean standard deviation of the proposed method is reduced by 28.3%, the mean peak signal-to-noise ratio is increased by 10.3% and the mean entropy is increased by 8.0%. In this study, the pictures of different haze levels under the same scene are tested. The results show that the pictures are clear, the details are intact, and the evaluation standard values are basically stable. The above test results indicate that the proposed algorithm has high robustness and color fidelity capabilities.

In view of the problem of collision and soft tissue injury caused by the traction of dual robotic arms in robot-assisted head and neck surgery, this study uses the traditional Informed-RRT*(Informed-Rapidly Exploring Random Tree*) path planning algorithm superimposes gravitational field to reduce the blindness of path search. In order to optimize the poor guidance and low efficiency of the traditional Informed-RRT* path planning algorithm, a regression filtering mechanism is introduced to avoid the local optimal search and dynamically adjust the step size. At the same time, the planning path is optimized and redundant node removal strategy is adopted to remove redundant nodes and improve the smoothness of the path. The dual-arm coordinated path planning method is investigated based on the dual-arm collision detection method and the improved Informed-RRT* algorithm. Simulation experiments show that compared with the original algorithm, the iteration time of the proposed algorithm is reduced by 72.76%, the number of iterations is reduced by 46.39%, the average path length is reduced by 6%, and the number of nodes is reduced by 45%, which verifies the effectiveness of the improved planning algorithm.

Network protocol grey box fuzzing AFLNET technology based on overlay guidance has attracted more attention in the field of network security testing, and there are many excellent research results. After the analysis of AFLNET and its derivatives, it is found that AFLNET has shortcomings in three aspects: message sequence attribution, message sequence evaluation and the selection of message sequence variation points, and a network protocol grey box fuzzy method based on message sequence attribution optimization is proposed. This method defines the concept of preference degree to measure the fuzzy benefit that message sequence can bring to each state, and proposes a new message sequence assignment algorithm combined with preference degree to re-assign interesting message sequences. Then, an evaluation function is constructed using the feedback information of multiple dimensions, which is used to more accurately calculate the true potential of each message sequence. In addition, a new mutation point analysis algorithm is proposed to help the fuzzer filter out the already mutated positions and mutate other more interesting mutated positions instead. The experimental results show that compared with the mainstream method, the QFuzzer implemented based on the proposed method increases the number of path coverage by 6.94%~11.04%, and the number of vulnerabilities found increases by 7.24%~30.70%.

In view of the problems such as low exploration efficiency, slow convergence speed or even non-convergence of deep reinforcement learning algorithm in the motion planning task of robot arm under the requirement of high dimensional state space and high precision, this study introduces asynchronous advantage mechanism based on SAC(Soft Actor-Critic) algorithm, and proposes an AA-SAC(Asynchronous Advantage Soft Actor-Critic) algorithm integrating asynchronous advantage. This algorithm replaces the original V network with a Q_target network,which effectively reduces the variance of the Q network. The n independent processes can be trained in parallel, which improves the training efficiency. The study also divides the experience playback pool of the AA-SAC algorithm into two parts, store and sample high-quality empirical data separately to improve the utilization of effective empirical data. The simulation results show that AA-SAC algorithm has the best performance in convergence speed, success rate and stability. Compared with the SAC algorithm, the convergence time of AA-SAC algorithm is 3 000 rounds earlier. After convergence, the success rate of AA-SAC algorithm reaches 96%, which is 6% higher than SAC algorithm and 26% higher than DDPG(Deep Deterministic Policy Gradient) algorithm.

In view of the problem of accurate estimation of symbol rate in non-collaborative communication, a high-precision symbol rate estimation algorithm based on the improved phase difference correction method is proposed on the basis of the envelope-square method. The algorithm compares the positions of the peaks of the envelope squared spectra of two sequences with the minimum shift relationship, discards the sequences that are seriously disturbed by noise, and selects the two sequences with the same peak positions to accurately estimate the symbol rate of the signal by the phase-difference correction method. Compared with the envelope-leveling method, this algorithm breaks the limitation that the estimation accuracy of the symbol rate depends on the number of sampling points of the signal, and has a very high estimation accuracy. Simulation experiment results show that the proposed algorithm can achieve accurate estimation of MPSK (Multiple Phase Shift Keying), MPAM (Multiple Pulse Amplitude Modulation) and MQAM (Multiple Quadrature Amplitude Modulation) signal element rates at low signal-to-noise ratios, and the complexity of the proposed algorithm is lower than that of existing high-precision estimation algorithms, making it suitable for application in engineering practice.

The decomposition components inside SF₆ electrical equipment can be detected by tunable absorption spectroscopy technique, in which the concentration of CO₂ reflects the insulation defect situation inside the equipment. Therefore, potential insulation faults of the equipment can be found in time by measuring the CO₂ concentration accurately. To overcome the problem of poor stability of traditional least squares concentration inversion model, ISFO-KELM gas concentration inversion model based on ISFO (Improved Sailed Fish Optimizer) and KELM (Kernel Based Extreme Learning Machine) is established in this study. The optimization ability and the ability to jump out of local optimal solution of ISFO are improved by using multi-strategy initialization method, Levy random step length, Cauchy mutation and adaptive t-distribution mutation techniques. The experimental results show that this model has high accuracy and robustness, and is superior to traditional methods such as least squares method, extreme learning machine, BP (Back Propagation) neural network in stability and generalization ability, which has important significance for evaluating the operation state of SF₆ electrical equipment.

In view of the problems of high cost and complicated parameters, the magnetic field simulation model of φ100 mm cold crucible experimental device is established by COMSOL simulation software, and the simulation experiments of high-frequency power supply and coil parameters are carried out. The optimal combination of high-frequency power supply parameters is predicted by neural network and simulated annealing algorithm. The magnetic field distribution, eddy current loss and energy utilization of cold crucible are simulated under different current intensity, frequency, coil height, turns, diameter and spacing. Based on the above simulation results, the neural network can quickly predict the magnetic field and eddy current loss under various power supply parameters, and finally use the simulated annealing algorithm to predict the optimal combination of power supply parameters. The simulation results show that after parameter optimization, the magnetic field uniformity can be greatly improved, the melting effect can be improved, the energy utilization rate can be increased by 68%, and energy waste can be avoided. These results preliminarily verify the improvement effect of high-frequency power supply and coil parameter optimization on magnetic field distribution and energy use efficiency, which can provide an important basis for the optimization design of power supply parameters and coil parameters of the following cold crucible engineering prototype.

Revealing human brain activity through machine learning EEG(Electroencephalogram) has become an important scheme to explore the inner emotional state of humans. Because the change of emotion state is dynamic rather than constant, it is difficult to predict the change of emotion state in the field of emotion recognition. This study proposes a label generation framework for instantaneous emotion intensity. A set of supervised labels is generated by having subjects watch videos that stimulate and capture their instantaneous emotional intensity, and combine the supervised labels with EEG features to generate three sets of semi-supervised labels to correspond to the instantaneous emotional state changes of subjects. In this study, EEG features and various machine learning methods are used to analyze the applicability of four groups of labels to emotional state changes. The support vector machine model achieves 80.02%, 54.76% and 56.14% classification accuracy for two-class, three-class and four-class sentiment intensities on supervised label sets. The experimental results show that the supervised instantaneous emotion intensity labels are more universal for EEG data and emotional state changes across different subjects.

In order to solve the optimal scheduling problem of large-scale electric vehicles into the network, a two-layer optimal scheduling strategy of integrated energy system with electric vehicles is proposed in this study. The upper layer is the optimal dispatching layer, in which electric vehicle agents group electric vehicles into clusters according to the dispatchable time and upload the cluster information to the system dispatching center, which cooperates with electric vehicle clusters and energy systems and builds an economic dispatching model with the goal of minimizing the dispatching cost by considering integrated demand response and ladder-type carbon trading mechanism. The lower layer is the power allocation layer, where electric vehicle agents build a power allocation model with the goal of satisfying users' travel demand, and guide electric vehicles to participate in system scheduling in an orderly manner. The simulation algorithm is constructed and solved by using CPLEX solver. The simulation results show that the proposed strategy can not only effectively reduce the scheduling cost of integrated energy system, smooth the system load curves and reduce carbon emissions, but also significantly reduce the cost of electricity consumption of customers on the basis of securing their travel demand, thus achieving a win-win situation for both supply and demand.