Aiming at the problems of a low information rate and the poor bit error performance of traditional chaotic keying schemes, a Phase Quadrature Multi-Carrier Noise Reduction Correlated Delay Shift Keying (PQMC-NRCDSK) scheme is proposed. In this scheme, the superposition of chaotic sequence and its time delay sequence is used as the reference signal, and the product of the chaotic sequence and its time delay sequence with data information is used as the information bearing signal. The reference signal and information bearing signal are transmitted by the carrier with different frequencies and quadrature phase, and multiple antennas are used to transmit and acquire signals at the sender and receiver, respectively. The error performance of this scheme in the multi-path fading channel is studied in detail and verified by theoretical analysis and experimental verification. Experimental results show that the theoretical analysis is consistent with the simulation results, and the correctness of the theoretical analysis is verified. The information rate of the PQMC-NRCDSK scheme is higher than that of the MC-CDSK and CDSK schemes under the same carrier number, and is twice as high as that of the MC-CDSK scheme. On the other hand, the error performance of the PQMC-NRCDSK scheme is obviously better than that of the above schemes.

To address the problems that the pedestrian interaction feature of the Social GAN is simple and that it cannot make full use of the most of pedestrian interaction information, this paper proposes a pedestrian trajectory prediction model with social features and attention mechanism. This model adapts the structure of generative adversarial networks. The generator adapts an encoder-decoder model and the attention model is put between encoder and decoder. Three social features are set to enrich pedestrian interaction information which assists the attention module to make full use of the most of pedestrian interaction information by allocating the influence of pedestrians in the scene, so that the accuracy of the model is improved. Experimental results on multiple datasets show that the accuracy of this model in the pedestrian trajectory prediction task is increased by 15% compared with the previous pedestrian trajectory prediction model based on the pooling module. The improvement effect is most obvious in scenes with dense pedestrians and lots of non-straight tracks, with the accuracy increased by 34%.

A coding parameters identification method is proposed, which is suitable for long constrained non-recursive systematic convolutional codes with a code rate of 1/2 and (n-1)/n obtained by puncturing the 1/2 code as the mother code. First, according to the coding principle, a linear block code with a code length of about 1/n of the original convolutional code constrained length is constructed by using the coding data; then, the check matrix of the linear block code is obtained, and the generator polynomial of the original convolutional code is reconstructed from the check matrix. Simulation experiments are carried out for two convolutional codes involved in IESS309. Compared with the existing method, when the code rate is 1/2, the recognition performance is improved by about 1dB; when the code rate is 2/3 and 3/4, the improvement is more than 2dB. Simulation results show that the proposed method is more effective than the existing method.

A high power density monolithic microwave integrated circuit (MMIC) power amplifier is presented for W band application. The chip is fabricated using the 100 nm GaN high electron mobility transistor (HEMT) technology on a 50 μm SiC substrate. The amplifier is designed for a high gain and high output power with three stage topology and low-loss impedance matching networks designed with high and low characteristic impedance micro-strips and metal-insulator-metal (MIM) capacitors. And quarter-wave micro-strips are employed for the DC bias networks, while the power amplifier is also fully integrated with bias networks on the wafer.Measurement results show that, at the drain bias of 15 V, the amplifier MMIC achieves a typical small signal gain of 20 dB within the frequency range of 88~98 GHz. Moreover, the saturated output power is more than 250 mW at the continuous-wave mode. At 98 GHz, a peak output power of 405 mW has been achieved with an associated power gain of 13 dB and a power-added-efficiency of 14.4%. Thus, this GaN MMIC delivers a corresponding peak power density of 3.4 W/mm at the W band.

This work addresses the signal recovery problem in the presence of impulsive disturbance utilizing l_p-norm optimization. In doing so, the resultant optimization is difficult to solve, especially when 0<p< 1, because it is nonconvex. In this work, the alternating direction method for multipliers steps is developed to efficiently obtain the solution from this optimization. In each step of the alternating direction method for multipliers, the corresponding solutions are respectively obtained by utilizing the iteratively reweighted least squares and interior-point approach. Numerical studies including an application of image enhancement demonstrate the superior performance of the proposed weighted estimation algorithms compared to the l_p-ADM approach.

In order to ensure that the sensor network data are reliable, and that the efficiency of data processing is not reduced due to the lack of network data, a method for data recovery in the sensor network based on the joint graph model is proposed. First, this paper establishes a joint graph domain model based on the smoothness of network data in the time-domain and spatial-domain, and then an iterative recovery method is proposed to recover the network data, which is based on the association characteristics of network data in the joint graph domain model. Experimental simulation shows that compared with the recovery method based on graph total variation minimization in the graph signal model, the method of data recovery based on the joint graph model improves not only by about thirty percent of the data recovery accuracy, but also by about eighty percent of the iteration efficiency.

Aiming at the problems of small coverage, short lifetime and diversity link channel attenuation in energy-constrained cognitive relay networks, the outage performance of energy harvesting multi-hop cognitive relay networks over complicated fading channel scenarios are studied. First, the κ-μ distribution is adopted to represent various single and mixed line-of-sight and non-line-of-sight fading channel scenarios. Then the dedicated power beacon assisted energy harvesting underlay multi-hop decode-and-forward cognitive relay networks are constructed. The exact and asymptotic unified outage probability expressions for secondary networks are derived over the κ-μ fading channels. Simulation results show that the secondary outage probability can be monotonically decreased with the increase of the power beacon signal’s power or/and primary networks’ interference constraint until it tends to saturation on the condition of various fading channel scenarios, and the secondary networks’ outage probability can be degraded with the deterioration of the channel link condition when the primary networks’ interference constraint is lower, and the secondary networks’ outage performance can also be increased with the improvement of the channel link condition when the primary networks’ interference constraint is higher. The secondary networks’ outage probability can be decreased by reasonable settings of the hop number and energy harvesting ratio according to the type of fading channel scenarios.

In order to effectively suppress the self-interference of the in-band full-duplex communication system, a solution for designing the transceiver antenna based on the simple ring structure and collinear dipole is proposed. In this design, two printed dipole antennas are placed in the zero radiation direction of each other to achieve low port isolation between the two transceiver antennas. Then a simple ring structure is added to improve the isolation between the two antennas. Experimental results show that the two transceiver antennas designed by this scheme have omnidirectional radiation characteristics and can obtain a high port isolation of better than 45.9 dB in 2.4~2.6 GHz band. The scheme for designing the transceiver antenna proposed in this paper can not only effectively reduce the interference between two transceiver antennas, but also provide omnidirectional signal coverage. It is an effective solution for designing transceiver antennas for the in-band full-duplex communication system.

In order to solve the security problem of data transmission in Low Earth Orbit(LEO) satellite networks, a novel secure routing protocol is designed. First, by considering two aspects of satellite behavior attribute and communication capability attribute, a multi-dimensional security evaluation model is established, and the model for evaluating the satellite security communication capability is dynamically adjusted with the network state. Then, combined with the on-demand routing protocol, the secure communication path is established by taking the quantitative evaluation of satellite security communication capability as the routing metric. Finally the forwarding behavior of nodes in the transmission path is monitored in real time, and the routing reconstruction is triggered as needed. Simulation results show that, compared with the existing protocol, the proposed protocol can significantly improve the performance of the packet delivery rate, average end-to-end delay and average hop than the existing protocol when malicious nodes exist in the network. The protocol can effectively improve the security of data transmission in LEO satellite networks.

An improved anti-jamming algorithm for spread spectrum communication based on blind source separation (BSS) is proposed to solve the problem that the separation performance of BSS technology deteriorates at a low signal-to-noise ratio (SNR), especially at the threshold SNR of spread spectrum communication. First, the algorithm detects the ambiguity of the output of the blind source separation algorithm, and then uses the detection result to correct the separation matrix. Finally, the new separation matrix is obtained by accumulating a plurality of modified matrices, and applied to the jamming separation of the system. The algorithm has a simple structure and can realize anti-jamming communication of the spread spectrum system at the demodulation threshold. Simulation results show that the demodulation error performance of the communication signal after removing the jamming by the algorithm at a negative SNR is still close to the theoretical performance.

Since images in traditional ultrasound-guided puncture surgery are obtained by touching the position of the lesion with a hand-held ultrasound probe, it often causes puncture failure by inaccurate probe positioning and unstable contact force control due to the doctor’s fatigue or poor experience. In this regard, a method for guiding the automatic positioning of the ultrasonic probe by using vision and force is proposed. First, the combined image processing algorithm and matching algorithm are used to detect the small target feature points. Then, the robotic arm is used to guide the ultrasonic probe to the initial desired pose based on the desired image through the positional visual servo and pose transformation. Finally, the force closed-loop control algorithm is adopted to fine-tune the probe position to ensure a constant contact pressure. Experimental results show that the average of the feature point detection matching error and the average visual positioning error of this method is within 2.5 pixels, and that the average contact force of the ultrasonic probe is stable at 3.86N. Compared with the traditional hand-held method, the ultrasonic probe has a higher positioning accuracy and more stable contact with the measured object, which makes the doctor’s operation easier.

A simple and low-cost electroless deposition technique is used to prepare nickel-doped carbon nanotubes under different doping conditions, and to explore the influence of different nickel doped samples on the electrical contact properties of carbon nanotubes. First, the original carbon nanotubes are subjected to mixed acid oxidation, sensitization and activation treatment. Subsequently, the nickel chloride hexahydrate is used as the main salt and dimethylamine borane is used as the reducing agent to prepare the electroless deposition solution. And then, the prepared carbon nanotube dispersion is dropped into the electroless deposition solution to obtain the sample of nickel-doped carbon nanotubes. Morphological characterization indicates that the nanoparticles with different particle sizes and doping amounts could be doped on the surface of carbon nanotubes under different deposition conditions. The X-ray energy spectrum shows that the main component of the doped nanoparticles is nickel, and the further X-ray photoelectron spectroscopy reveals the constituent valence state of nickel. However, the Raman spectra indicate that the doping type of nickel-doped carbon nanotubes by this deposition method is the P-type. Finally, the electrical contact performance test results show that the electrical contact properties between gold electrodes and nickel-doped carbon nanotubes with different particle sizes and doping contents are different, but that all of them can be improved to a certain extent. According to the order of nanoparticles with a small particle size and a moderate amount, those with a moderate particle size and a moderate amount, and those with a large particle size and a large amount in the nickel-doped sample, the average value of the contact resistance decreases by 32.70%, 71.63% and 49.33%, respectively.

When the indoor visible light communication link is interrupted due to occlusion blocking, the signal needs to be forwarded through the auxiliary relay, and therefore, an indoor visible light communication auxiliary relay system model is established. First, the positional constraint relationship of indoor light source on the roof when there is link blockage in the indoor visible light communication system is analyzed by using the office ceiling light as the signal light source and relay, and the asymmetrically multiplexing DC offset optical orthogonal frequency division multiplexing is used to realize the auxiliary relay signal multiplexing, which ensures the occlusion signal relay transmission while maintaining its target terminal. Then the genetic algorithm is used to optimize the distribution of the power at the source and the relay to optimize the system performance, and the BER performance of the equal power allocation scheme is compared in the occlusion case. The results show that when the direct communication link is blocked, the system loss of the auxiliary relay system with the optimal power allocation scheme is 3 dB less than the system with equal power allocation.

The human ear can only accept one sound signal at one time, and the signal with the highest energy will shield other signals with low energy. According to the above principle, this paper combines the self-attention and the multi-head attention to propose a speech enhancement method based on the multi-head self-attention mechanism. By applying multi-head self-attention calculation to the input noisy speech features, the clean speech part and the noise part of the input speech feature can be clearly distinguished, thereby enabling subsequent processing to suppress noise more effectively. Experimental results show that the proposed method significantly outperforms the method based on the recurrent neural network in terms of both speech quality and intelligibility.

Aiming at the problem of evaluation of the accuracy of track fusion algorithms, a new evaluation index is proposed. The index is based on the cardinalized optimal linear assignment metric. The Hellinger distance is used to replace the Euclidean distance of the distance error part of the metric, so that the error covariance of the state is considered and the cutoff distance is removed, making the cutoff parameter only play the role of adjusting the weight between the distance error and the cardinal error. A variable is added in the potential error part, and can be set according to the need so as to achieve the function of whether the index considers the cardinal error. Simulation experiments show that the new index can not only correctly evaluate the fusion precision, but also evaluate the pros and cons of the fusion algorithm more sensitively by considering the state uncertainty and cardinal error, and that it can effectively evaluate the fusion precision when the fusion error expectation is unknown.

Aiming at the low definition and poor details of synchronous multi-band image fusion, a novel method based on attention generative adversarial networks is proposed. First, the attention weight map is constructed using the difference between the multi-band feature map and its mean, then the feature enhancement map is obtained by the point multiplication and addition of the feature map and the attention weight map to construct the feature enhancement module. Second, the feature-level fusion module is designed, which connects the multi-band feature enhancement map and reconstructs the fused image through normalization, upsampling, convolution, etc. Finally, the feature enhancement module and the feature-level fusion module are cascaded to build the generator, and the VGG-16 is used as a discriminator to establish a Generative Adversarial Network, thereby implementing multi-band image end-to-end fusion. Experimental results show that the proposed fusion method can lead to the most prominent average gradient compared with classical fusion methods, and that the effectiveness of the proposed method is verified.

The performance, capacity and reliability of the flash memory on devices such as smartphones and the Internet of Things are limited. Deduplication can resolve these restrictions by removing the duplicate I/O, but must be done under various resource constraints on the device. This paper proposes the M-Dedupe deduplication, which applies content-aware deduplication I/O requests on critical paths, and improves the performance and efficiency of mobile phones by improving the efficiency of flash garbage collection. The prototype system verification results show that the M-Dedupe reduces the write data by an average of 23.7%~42.5%, the average write response time by 21.2%~39.8%, and the average erased block by 16.8%~43.9%. Besides, it can achieve high-deciding duplication in the mobile flash system, improve the deduplication efficiency, and save the storage space.

Aiming at the time-consuming problem of image registration in video mosaic, a combined algorithm based on the BRISK and GMS is designed and optimized. Based on the fast feature extraction feature of the BRISK, the grid image method is used to divide the overlap region image to make the feature point distribution more uniform, thus reducing the number of feature points that need to be operated. In the feature matching stage after feature extraction, the GMS is used to eliminate the mismatching pair, and improves the matching accuracy according to the two-way matching strategy. In order to make the image fusion more natural and smooth, this paper designs a regionalized feather blending weighted fusion algorithm. To reduce the error in solving the overlapping area, the fusion area is constructed and divided into regions. The best suture method is used to get the seam. The gradual in and out weighting algorithm is used in different regions to achieve image fusion, so as to obtain the panoramic image with higher fusion quality. Finally, a complete prototype of the panoramic video stitching system is designed. The feasibility and practicability of the system scheme are verified by experiments which show that, compared with the traditional video splicing technology, this algorithm ensures the real-time performance of video splicing while eliminating the ghosting and seaming problems that occur during splicing.

Considering the shortcomings of existing research methods in the Chinese medical health questions classification task, this paper proposes a new health questions classification method that incorporates the health questions’ local semantic information and global structural information. We first obtain the questions’ local semantic representation and global structural representation by the convolutional neural network (CNN) and independent recurrent neural network (IndRNN). Then, we extract the questions’ semantic representation, and especially we get the questions’ semantic representation by fusing the local semantic representation and global structural representation using a self-attention mechanism. Finally, we classify the semantic representation of the medical health question through the softmax layer and output classification result. Experimental results show that this method leads to a good performance in the Chinese medical health questions dataset, and that it effectively improves the semantic representation ability of the model and significantly resolves the gradient vanishing and gradient explosion problems.

In the real working environment,the mobile robots have a poor recognition performance to speech control commands due to the noise effect. Aiming at this issue,this paper proposes a new algorithm based on the gammatone frequency cepstral coefficient and the mixed feature extraction of the Teager energy operator. This algorithm replaces the common Mel filter with the Gammatone filter. In the process of extracting gammatone frequency cepstral coefficients,the Teager energy operator reflecting the energy of speech signal is added to form a new feature, with the dynamic characteristics of the speech signal considered. It is combined with the first-order difference parameters to form a mixed feature. And the principal component analysis is made to reduce the dimension,and the final mixed features are used to the speech recognition system for control command of the mobile robot. Experimental results show that,in the environment of the workshop noise and signal-to-noise ratio of 10dB,the recognition rate of mixed features is improved by 12.20% compared with the mel frequency cepstrum coefficient. The recognition rate of the mixed feature is increased by 1.02% when the dimension is reduced by principal component analysis.

An optimization algorithm is proposed utilizing the video data and point cloud data captured by the depth camera to solve the problems such as error-proneness and incoherence of motion sequence caused by the existing human pose estimation algorithms based on the morphable model. For video data, the neural network is first used in extracting the model parameters from each color image frame. Next, the human key-points and contour constraint are considered to optimize the above parameters. Then the coherence between every two consecutive frames is utilized to correct the error of pose estimation, thus making the resulting motion sequence smoother. In addition, the point cloud and the model obtained from the corresponding color image frame are used as the joint input to further improve the estimation accuracy. Finally, the distance between the point cloud and the corresponding point of the model is constrained to be as small as possible to obtain a more reasonable solution. The proposed algorithm and the state-of-the-art algorithms are compared qualitatively and quantitatively on the data set and real video set. Experimental results show that the algorithm can effectively correct the error and incoherence in the single-frame pose estimation results and greatly improve the accuracy when using point cloud data optimization.

The relatively low piezoelectric constant of the aluminum nitride piezoelectric film limits the development and application of the piezoelectric micro-machined ultrasonic transducer based on the aluminum nitride. Therefore, the directivity of the piezoelectric micro-machined ultrasonic transducer array is studied for this problem. First, the far-field sound pressure and normalized directivity function of the transducer array are calculated according to the Rayleigh principle. Furthermore, the effects of array element radius, array element spacing, array element number and operating frequency on the beam width, direction sharpness angle and sidelobe level of the array are analyzed, and the transducer array is optimized. Finally, based on the optimization results, area and filling efficiency, the final structure size of the transducer array is determined and the sound pressure distribution visually simulated. The results show that the optimized transducer array has an ideal sound pressure distribution, good directivity and sharp main lobe. The-3dB beamwidth of the main lobe is about 9 ° and the sidelobe level is about 0.228.

In order to solve the classification problem on the urine-forming sub-cell images, a urine cell image classification algorithm based on the Squeeze-and-Excitation GoogLeNet is proposed. The algorithm uses the feature recalibration mechanism and brings about significant improvement in the useful feature for the current task through squeeze and excitation operations, which explicitly models interdependencies between cell feature channels learned by the Inception architecture during the training process. On the urine cell datasets, comparative experimental results show that the algorithm provides a better classification effect, which improves the accuracy of classification by 3% and the recall rate by 1% at the similar speed of the GoogLeNet network.

For point cloud classification, deep learning based methods use operations like voxelization to generate regular 3D grids or render the 3D mesh into a collection of images from multiple angles. However, the conversion will introduce additional computing and storage consumption. Some methods directly consume the raw point cloud. But their network scale and computational complexity make it difficult for them to deploy in embedded environments. On the basis of intensive studies of these algorithms, a novel lightweight dual path way network is proposed in this paper. Without additional conversion, our network attains a comparable performance but has 0.8 million floating parameters only. With point-wise and neighbor-wise representations, our approach incorporates global and local features of the point cloud. Experimental results on ModelNet40 and MNIST data-set demonstrate that our method achieves a good accuracy, and prove the effectiveness of our design.

In order to demonstrate the feasibility of hydroxyethyl cellulose hydrogel as an optical fiber sensor moisture sensitive material, a new optical fiber humidity sensor based on hydroxyethyl cellulose is proposed. First, a hollow-core fiber(150μm) is fused to a single mode fiber. Second, a 10μm thick hydroxyethyl cellulose hydrogel film is coated on the end face of the hollow-core fiber. Finally, the proposed humidity sensor is placed in a constant temperature and humidity chamber for a humidity response test. Experimental results show that this sensor has good humidity response characteristics and only needs 2.75s from 35%-85% RH, with the RH sensitivity being 224.5pm/%RH. As a good moisture sensitive material, hydroxyethyl cellulose has a good application prospect in various humidity sensors.

In order to reduce the coding complexity of quality Scalable High efficiency Video Coding (SHVC), a fast algorithm for intra prediction is proposed. First, inter-layer correlation and spatial correlation are combined to predict the coding unit depth range of the enhancement layer. Then, at each depth level, the residual coefficients are tested by Jarque-Bera to determine whether the inter layer reference mode is optimal. If the inter layer reference mode is optimal, the high-complexity intra prediction can be skipped directly. Finally, in order to terminate the depth division ahead of time, hypothesis tests are carried out to determine whether the residual coefficients in the coding unit show significant differences. Experimental results demonstrate that compared with standard SHVC encoder, the proposed algorithm reduces the coding time of the enhancement layer by 79% on average with almost the same coding efficiency.

In order to improve the ability of the non-volatile memory storage device system to concurrently execute access requests, aiming at the diverse nature between read and write access requests and the different properties of file data and metadata in the storage device, we have designed a file-based parallel write-based file data concurrent write strategy, RCU based file data read and write concurrency strategy and a minimum spin lock-based metadata synchronization strategy to improve the degree of concurrency of requests execution. And then we have implemented a prototype of the asymmetric lock-based high concurrent non-volatile memory storage system, which has been tested and analyzed by common test tools and methods, the result shows that compared with the PMFS, the prototype system can increase the throughput by 40%~162% and input/output operations per second by 61%~159%.

In order to improve the long-term in orbit flight reliability of the aircraft control system, a multi-mode control scheme is proposed based on reinforcement learning. This system includes a sensor module, a control module and an execution module. The sensor module is used to input the sensitive flight data of the aircraft to the control module in real time. This data is divided into multidimensional structured floating point data with historical relevance that can be directly used for aircraft control and the unique physical representation quantity of a particular sensor. The control module is divided into an input layer, a feature extraction layer and a full connection layer. The execution module is used to receive the driving data from the control module in real time, which includes the optimal state value for decision-making and the action output value for evaluation. The system decides which specific execution modules to use based on the optimal return value for decision making, with the output value of a selected specific execution module depending on the output value of the action used for evaluation. The system enables the aircraft to complete a long-term orbit operation in the multi-mode input and output state with 15ms fast response and 5.23GOP/s/W Performance per Watt.

In view of the adverse effect of the random initial value on the performance and convergence speed of the gravitation search algorithm, a quasi-oppositional gravity search algorithm (QOGSA) is proposed. The quasi-oppositional based learning OBL is embedded into the GSA algorithm, the number of iteration is divided into multiple learning cycle, the oppositional probability is adjusted according to the success rate of the past learning cycle, and an adjustable oppositional probability is designed to optimize the timing of the mechanism in the evolution, which improves the speed of the algorithm to search for the optimal solution greatly. On this basis, in order to improve the population diversity, elite particles are retained to the next generation population. They replace the particles with a poor fitness value and acquire a higher optimization accuracy. Compared with the existing algorithms in the literature, the optimization accuracy of the QOGSA for the average optimal value of the single-peak and multi-peak test functions can be improved by 1016. For the shaping results of different types of beam, the optimization accuracy of the improved algorithm for the sidelobe can be improved from 1.26dB to 5.99dB. On the premise of the fastest convergence speed, the QOGSA can greatly avoid the problem that other optimization algorithms tend to fall into local optimization, with the overall performance being the best.

In order to handle the low accuracy of the algorithm DCS-SOMP, two underdetermined wideband direction of arrival estimation methods with a high accuracy are proposed from the perspective of multiple iteration. First, the wideband signal processing model using the sparse array is established and transformed into a distributed compressive sensing problem through sparse representation. Then, the noises are eliminated through matrix transformation. After that, an algorithm utilizing proximity searching is proposed to improve the estimation accuracy and an algorithm based on the refined grid is proposed to compensate grid mismatch. Simulation results show that the proposed algorithms outperform the DCS-SOMP with a higher estimation accuracy and possess the advantage in computational speed.

Batch Normalization (BN) can effectively speed up deep neural network training, while its complex data dependence leads to the serious "memory wall" bottleneck. Aiming at the "memory wall" bottleneck for the training of the convolutional neural network(CNN) with BN layers, an effective memory access optimization method is proposed through BN reconstruction and fused-layers computation. First, through detailed analysis of BN’s data dependence and memory access features during training, some key factors for large amounts of memory access are identified. Second, the “Convolution + BN + ReLU (Rectified Linear Unit)” block is fused as a computational block to reduce memory access with re-computing strategy in training. Besides, the BN layer is split into two sub-layers which are respectively fused with its adjacent layers, and this approach further reduces memory access during training and effectively improves the accelerator’s computational efficiency. Experimental results show that the amount of memory access is decreased by 33%, 22% and 31% respectively, and the actual computing efficiency of the V100 is improved by 20.5%, 18.5% and 18.1% respectively when the ResNet-50, Inception V3 and DenseNet are trained on the NVIDIA TELSA V100 GPU with the optimization method. The proposed method exploits the characteristics of memory access during training, and can be used in conjunction with other optimization methods to further reduce the amount of memory access during training.

The positioning accuracy of a satellite image is mainly affected by the estimation accuracy of the rational polynomial coefficients (RPCs). Image point compensation or ground control point correction methods are usually used in the existing algorithms. Because the error characteristics of the design matrix elements are not considered, there are problems such as incomplete systematic error elimination and low parameter estimation accuracy. Considering the influence of the model systematic error, a heteroscedastic estimation method is proposed in this paper. First, the random model of matrix elements is established in the algorithm to describe the system characteristics more accurately. Taking into account the system deviations of the design matrix elements, the least square model is constructed using the Mahalanobis distance as the metric, and parameters are solved using the generalized eigenvalue method. The systematic error can be reduced theoretically. Experiment on different terrain images of TH-1 shows that the image correction accuracy of the proposed method is improved by more than 36 times compared with the traditional method, and the precision consistency is superior, which is of great significance to improving the accuracy of RPC parameters estimation and satellite imagery positioning.

In order to solve the problem of over-fitting of traditional supervised learning methods in anomaly detection of unbalanced datasets, an unsupervised adversarial learning method is proposed for hard disk failure prediction. This method uses the long short-term memory neural network and fully connected layer to design an Autoencoder that can be used for secondary coding. Only normal samples are used for training. By reducing the reconstruction error and the distance between potential vectors, the model can learn the data distribution of normal samples, thus improving the generalization ability of the model. The model also introduces the generative adversarial network to enhance the effect of unsupervised learning. Experiments on several datasets show that the recall rate and precision of the proposed method are higher than those of traditional supervised learning and semi-supervised learning classifiers, and that its generalization ability is stronger. Therefore, the unsupervised adversarial learning method is effective in hard disk failure prediction.

Mobile Edge Computing (MEC) can perform computational task offloading with the help of edge servers, and is no longer limited by the power of mobile terminals (MTs). When the edge server is overloaded, it often chooses to queue, postpone or reject the MT’s offloading request. QoS (Quality of Service) of users will deteriorate greatly due to service disruption and extended waiting, but the existing research work does not consider how the MEC-BS can relieve load pressure at this time. In this paper, we study how to enhance the computing offloading service of the MEC-BS by offloading the task of the overloaded base station to the other MEC-BS in the same collaboration space. Combining the penalty function with the two-step quasi-newton method, an optimization algorithm is proposed to minimize the joint utility function including the total delay and energy consumption of the edge computing network. Empirical factors are used to adjust the optimization deviation according to the different needs of the optimization target for time delay or energy efficiency. Simulation results show that the proposed scheme is better than two other schemes in improving the system performance and convergence speed.

To improve the ability of the integrated circuit to resist reverse engineering, we study the logic obfuscation technology and propose a logic obfuscation scheme based on the Reed-Muller camouflage gate. First, different virtual hole configurations are adopted to realize XOR/AND logical functions on the same layout, and feature information of the logical obfuscating circuit is extracted to make the standard cell physical library. Then, the obfuscation physical library is applied in the circuit netlist by the random insertion algorithm. Finally, the ISCAS benchmark is used to verify the effectiveness of the proposed scheme. Simulation results reveal that the similarity of the Reed-Muller logic camouflage layout is improved by 14.36%, and that the power consumption overhead is about 2.36% under the larger scale benchmark. Experiment indicates that the designed obfuscation gate can effectively resist reverse engineering and improve the hardware security of the circuit.

The visual characteristics of low-altitude drones are less obvious and the scale changes during the detection process. Traditional detection methods are susceptible to interference during detection, and most of those methods cannot work quickly and robustly. To solve this problem, a real-time drone detection algorithm combined with the improved YOLOv3 model and the super resolution method is proposed in this paper. First, frame difference is used to propose the candidate area, and the super-resolution method is used to strengthen the details. Then the dimensional clustering algorithm is used to regenerate the anchors for the model, and the model is slightly adjusted. Finally, we use the improved YOLOv3 to scan both the whole frame and the processed candidate area so as to detect the drones. The frame relationship is also used to implement tracking of drones by real-time detection. With GPU (GTX 1070Ti) acceleration, the method works at a speed of about 20FPS and has an accuracy rate of 96.8% and a recall rate of 95.6%. The results prove that the method can detect drones in different complex backgrounds with a considerable effective detection distance. Compared with the traditional method or normal machine learning method, our method is of a certain theoretical and practical value.

Because existing ultrasonic transducers mostly use PZT and ZnO materials as piezoelectric thin films, while the PZT contains lead and ZnO has the problem of contaminating CMOS manufacturing, a piezoelectric ultrasonic micromechanical transducer with circular bi-laminate bending vibration which uses the aluminium nitride as the piezoelectric layer is designed. The working principle of the transducer is analyzed, the finite element model is established, and the finite element simulation is carried out for the size parameters of the transducer. It is found that the resonant frequency of the transducer is proportional to the thickness of each layer and inversely proportional to the square of the radius of the transducer; when the radius of the upper electrode is about 65% of the radius of the transducer, the resonant amplitude of the transducer is the largest; when the thickness ratio of the silicon and the aluminum nitride of the piezoelectric layer is about 0.6, the resonant amplitude is also the largest. The optimized transducer is simulated and compared with the original model. The results show that the working frequency in air is 9.21MHz, the electromechanical coupling coefficient increases from 21.44% to 27.16% in air and from 3.55% to 11.93% in water. These conclusions provide basic data for the research on the medical imaging probe.

Aiming at the low prediction accuracy of traditional network security situation prediction technology, a network security situation adaptive prediction model (NAP) is proposed. First, it extracts alarm elements and calculate network security situation time sequences based on the entropy correlation method. Then, the sequences are taken as the input of the sliding adaptive cubic exponential smoothing method with initial security situation predicted value sequences generated. Finally, the time-varying weighted Markov chain is used to predict the error value based on the error state and the initial predicted values are modified. Experimental results show that the NAP has a better prediction accuracy than other existing models.

Due to the moving platforms, the clutters in distributed airborne MIMO radar are non-Gaussian and non-homogeneous, which leads to having no independent and identically distributed training data to estimate the clutter covariance matrix. To solve the problem, we propose that the covariance of the clutter should be modeled as an inverse complex Wishart distribution whose average value is a Hadamard product of the covariance matrix taper (CMT) and the clutter Doppler spectrum component. Based on this clutter model, a novel detector combing the Bayesian approach and the generalized likelihood ratio test(GLRT) is proposed. Numerical simulation results show that the proposed detector has a better detection performance compared with two current commonly used non-Bayesian detectors.