Abstract:Object detection and parameter estimation are two fundamental problems of radars. The process firstly detects existence states of objects， and then estimates parameters to identify the existence state. In this paper， we establish a generalized system model combining detection and estimation. Radar information， detection information and estimation information are strictly defined. It is proved that the radar information is the sum of the detection information and the estimation information. The joint posterior probability distribution of the existence state and time delay is derived for constant amplitude scattering objects. Radar entropy error is defined as power of posterior entropy. Both radar information and entropy error can be used to evaluate radar performance. A stochastic method of joint object detection and parameter estimation is proposed by sampling a posteriori （SAP） probability distribution. The main contribution is the proof of radar theorem which points out the theoretical radar limit for comparison of various methods. The radar theorem shows that the entropy error is achievable. Conversely， the empirical entropy error of any radar is no less than the theoretical entropy error. It is further proved that the concatenated system with the optimal detector and the optimal estimator can approach the theoretical limit of radar. The radar theorem and separation theorem answer the theoretical problem of joint detection-estimation， and will give a great promotion to the development of radar technology.

Abstract:Spacecraft faces a complex environment of repeated takeoff and landing， large temperature difference， strong vibration and severe impact. In order to improve its usability and maintainability， and enhance the acquisition of the structural capacity data of the spacecraft in the whole life cycle， it is necessary to monitor the health status of the spacecraft structure all day long and carry out research on structural health management technology. The data acquisition， transmission， collection and processing of the sensitive measuring points inside and outside the spacecraft are realized by wireless， and the all-round all-weather health monitoring function of the spacecraft structure is realized. The design of wireless health monitoring system for spacecraft is studied. The prototype implementation of wireless health monitoring system for spacecraft is realized.

Abstract:Integrated satellite and terrestrial networks are the integrated networks in which the satellite network and the terrestrial network are interconnected to realize the complementary advantages of the satellite network and the terrestrial network. With the rapid development of wireless communication， the co-frequency multiplexing technology of integrated satellite and terrestrial networks can effectively improve the resource utilization efficiency and meet the growing demands of wireless data services. However， due to the diversity of wireless data services and the scarcity of spectrum resources， how to improve the quality of experience （QoE） and fairness has become a key problem to be solved in the integrated satellite and terrestrial networks. We investigate a general QoE-aware dynamic resource allocation strategy with consideration of the time-varying channels， diversified user’s quality of experience and fairness for the integrated satellite and terrestrial networks. To do so， we formulate the objective problem as the maximization of the network fairness utility based on the time-averaged QoE. By using the Lyapunov optimization theory， we further transform and decompose the initial problem into three subproblems which can be independently solved at each slot. Meanwhile， we derive a low complexity two-step algorithm to solve the third resource allocation subproblem which is a non-concave mix combination optimization problem. The simulation results reveal the tradeoff between user’s long-term time-averaged QoE and fairness performance under different fairness parameters.

Abstract:In order to improve the performance of the spectrum detection， and reliable communication in the spectrum congestion and competition complex electromagnetic environment of satellite system， the spectrum detection is converted to a binary classification problem by employing the support vector machine （SVM） algorithm. Specifically， the feature vector， which is used to characterize the signals， is obtained by removing the central and basis vectors from the energy vector and the SVM model for determining the spectrum status is then constructed. Moreover， the optimal parameter of the Gaussian kernel is determined by adopting the simulated annealing （SA） algorithm. Simulation results show that the proposed scheme can achieve better spectrum detection accuracy and increase the detection robustness as well as improve the system throughput and energy efficiency as compared to the existing single threshold and double-threshold base spectrum sensing schemes. The work conducted in this paper could support the construction and development of future cognitive satellite communications systems.

Abstract:To solve the problem that the unlabeled data in the text classification task cannot be immediately marked and the cost is too high， this paper proposes an active learning method for uncertainty based on text classification. The MSDL （Measure sample density by LDA） algorithm is proposed to calculate the unlabeled sample density， and the new metric sample aggregation situation is introduced. The initial training set sample is selected in the densely sampled region， thus making the initial The training set is more representative. The more uncertain samples from the unlabeled samples are added to the training set， the samples are weighted based on the information entropy， and the classifier model is iteratively updated until the expected termination condition is reached. Experimental results show that this method is better than other traditional active learning algorithms in text classification tasks.

Abstract:For adversarial attacks， black-box attacks are more challenging and applicable than white-box attacks. Recently， black-box attacks based on the transferability of adversarial examples have become mainstream methods. However， the adversarial examples generated by most existing methods exhibit low efficiency in black-box attacks. In this paper， a combination strategy based on Gaussian noise and flipping is proposed to enhance the transferability of adversarial examples， thus achieving higher black-box attack success rates. Moreover， this strategy can be integrated into any gradient-based method to obtain stronger attacks. Extensive experiments on an ImageNet-compatible dataset show that our proposed method can generate more transferable adversarial examples. In addition， our best attack can fool six state-of-the-art defense models with an average success rate of 86.2%， and deliver 8.0% success rate increasement compared with the state-of-the-art gradient-based attack.

Abstract:Car body surface damage detection is a classic problem in computer vision. The main bottleneck of car body surface damage detection lies in the different scales of damage instances in the image， which affects the accuracy and efficiency of segmentation. In this paper， we use a single-stage semantic segmentation network （YOLACT⁺⁺） for damage detection on the car body surface， combine EfficientNet to design a backbone network to improve segmentation efficiency， and improve the loss function optimization YOLACT⁺⁺ to generate the target instance Mask in the experiment. Experimental data are marked by deep learning， and results show that the improved YOLACT⁺⁺ detection frame rate is increased to 35 frame/s， which reduces the mask generation error and improves the instance segmentation accuracy of YOLACT⁺⁺.

Abstract:The brazing quality of metal honeycomb components is usually evaluated by the brazing rate （the proportion of the welded area detected in the unit area） as an indicator. In actual production， the GH4099 superalloy thin-walled narrow-ribbed honeycomb panel is used as the research object， ultrasonic C-scan amplitude imaging is used for non-destructive testing， and an unsupervised machine learning classification method based on the eigenvalue parameters of ultrasonic A-scan signal is proposed. Firstly， eight eigenvalues are extracted in the time domain and power spectrum of the digital ultrasound signal， respectively. Secondly， the data is standardized and reduced the dimensionality by using principal components analysis （PCA） to obtain the top three groups with six principal component values， which have respective contribution rates of more than 95%. Then these values are used as eigenvalues to perform K-means clustering， Gaussian mixture model clustering， and fuzzy C-means clustering as the input. Finally， the multi-classifier fusion algorithm is used to improve the accuracy of the model， and the classification results are visualized and compared with the ultrasound C-scan amplitude imaging to verify the classification evaluation effect. Experimental results of twelve groups of data show that the imaging results of the three clustering algorithms are consistent with those of the ultrasound C-scan amplitude imaging， in which the fusion voting calculation is more accurate than the single classifier. The study provides new ideas for an unsupervised machine learning method in ultrasound signal for evaluating the quality of honeycomb component brazing.

Abstract:With the development of marine resource exploration and marine pollutant monitoring， monitoring and collection of hydrologic data have become an important research direction.Among them， underwater wireless sensor networks play an important role in hydrological data acquisition. This paper studies the data collection of sensor nodes in the two-dimensional monitoring model network of underwater wireless sensors. The proposed method optimizes the path of sensor nodes by using self-organizing mapping （SOM）.We combine the optimized path graphics and the K-means algorithm to find the path internal aggregation point.Then we find the data acquisition point within the sensor communication radius through the aggregation point and the sensor node. Finally， the optimal path of autonomaous underwater vehicle （AUV） to each data collection point is found through SOM. In the verification experiment， 52 sensor nodes are arranged within a range of 1 200 m×1 750 m under the water， and the path obtained by data collection points is optimized by 6.7% compared with the path planning of sensor nodes in the same acquisition sequence. Compared with the optimal solution of sensor node path， the path obtained by reorganizing the self-organizing path planning of data collection points is improved by 12.2%. The results of increasing the number of sensor nodes are similar， so the proposed method can improve the data acquisition efficiency of AUV.

Abstract:Through non-intrusive load monitoring technology， a more detailed understanding of the electricity consumption information of residents at various time periods can be obtained， This can help us to develop a reasonable electricity consumption plan for scientific electricity use. The focus of non-intrusive load monitoring technology is the detection of transient events. The proposed Hilbert vibration decomposition （HVD） algorithm detects the transient events in the sudden change of electrical parameters such as power and current when the electrical equipment is turned on. Compared with the double-sliding window CUSUM change point detection algorithm， the HVD algorithm load detection does not need to set a threshold， so the possibility of missed and false detection is greatly reduced. The corresponding circuit model is built by MATLAB / Simulink simulation software， and the simulation analysis shows that the HVD algorithm can effectively identify transient events.

Abstract:From the perspective of complex networks， this paper constructs a network model of the search index of artificial intelligence online translation based on the time series data， and then analyzes its structure characteristics based on actual data from China. The results show that： although the search index of online translation shows significant fluctuation characteristics， it is still dominated by small fluctuations in most of the time； the distribution of the shortest path lengths of the online translation network is approximately skewed distribution， and the conversion from one symbol to another in the network requires three intermediate symbols on average； the symbols with less volatility have larger clustering coefficient； online translation shows a downward trend as a whole， and has experienced a process from early immature to gradually mature.

Abstract:Mechanomyogram （MMG）-based pattern recognition refers to the process of collecting MMG bands and applying machine learning algorithms to perform motion recognition. To realize the real-time classification of sign language motions， six-channel MMG signals of three muscles on forearm are collected by dual-axis acceleration sensors which are controlled by lightweight embedded device with STM32 chip. The back propagation neural network （BPNN） algorithm is used to establish recognition classification models， where the parameters are extracted and put into the embedded system to transfer BPNN algorithm. The embedded device can accomplish real-time recognition of 30 kinds of sign language motions， with the self-test accuracy up to 99.6% and the accuracy of real-time recognition up to 97.5%. Moreover， the classification time for each motion is less than 0.52 ms， satisfying the real-time recognition condition. The results can be applied to the fields of rehabilitation engineering， sign language translator， and prosthetic control， etc.

Abstract:SimRank is a method based on the topological structure information of the graph to measure the similarity between any two objects. However， in real large-scale social networks， the iterative computation between nodes is time-consuming. Here we propose a hierarchical community detection algorithm based on global matrix smooth convergence using SimRank， called SGSC. First， the SGSC algorithm identifies the initial core nodes in a network by classical measurement.Then， it smoothly converges a matrix to calculate SimRank to obtain original core nodes. Based on the global convergence matrix， we cluster the communities around the core nodes and use a closeness index to merge two communities. By recursively repeating the process， a dendrogram of the communities is eventually constructed. We validate the performance of SGSC by comparing its results with those of two representative methods for three real-world networks with different scales， and comparison results show that the proposed SGSC algorithm improves the accuracy in community division and reduces running time in social networks of different scales.

Abstract:Three-dimensional （3-D） human target detection has important application value in intelligent security， robot， automatic driving and other fields. At present， the 3-D human target detection method based on radar and image data fusion mainly adopts two-stage network structure， which respectively completes the selection of candidate boundary boxes with high target probability and the target classification/regression of target candidate boxes. Although the preselection of target candidate bounding box enables the two-stage network structure to achieve higher detection accuracy and positioning accuracy， the complexity of the network structure leads to the limitation of the operation speed， which cannot be applied in scenarios with high real-time requirements. In order to solve the above problem， this paper studies a real-time detection method of 3-D human targets based on improved RetinaNet. The backbone network and feature pyramid network are combined for point cloud and image feature extraction， and the fused feature anchors are input into the functional network to output the 3-D boundary boxes and target category information. By using the one-stage network structure， the method directly regresses the category probability and position coordinates of the targets， solving the imbalance problem of positive and negative samples in the process of one-stage network training by introducing focal loss function. Experiments on KITTI dataset show that the proposed method outperforms the contrast algorithms in terms of average accuracy and time-consuming， and can effectively balance the accuracy and real-time performance of target detection.

Abstract:To solve the problem of the registration position deviation for multi-repeat texture images，a double-match image registration （DMIR） algorithm is proposed. The DMIR algorithm not only considers the matching result of one graph with another graph，but also considers the matching result of a graph with its own features. Firstly，the key points are matched by the K-nearest neighbor （KNN） algorithm after extracting the feature points by the scale-invariant feature transform （SIFT） algorithm. As a result，the self-matching point pairs of the same image and the initial matching point pairs between different images are obtained respectively. Secondly，the best matching point pairs are obtained by computing the correlation between different points of the initial matching point pairs. Thirdly，the correct matching point pairs of the two images are determined，which depend on the positional relationship between the best matching point pairs and the self-matching point pairs. Lastly，the affine matrix is calculated according to the matching point pairs，and the image stitching is performed. The experimental results show that the matching point pairs obtained by the DMIR algorithm are more accurate， and the stitched images are better than others.

Abstract:This paper proposes a method combining space-time coding and mixed-ADC architecture to reduce system power while increasing transmit diversity gain. Meanwhile， this paper analyzes the performance of mixed-ADCs massive MIMO system with space-time coding. By using the additive quantization noise model， we get the expression of the uplink spectral efficiency of the system. And random matrix theory is applied to derive approximate expression. Based on the approximate expression， we further analyze the impact on key system parameters. The simulation result shows that mixed-ADCs architecture can effectively reduce the power and cost of the massive MIMO system with space-time coding， and the transmit diversity gain of the system with space-time coding is double.

Abstract:Aiming at the problem of resource allocation and interference of D2D （Device to device） communication technology in cellular system， a D2D resource allocation strategy based on the improved genetic algorithm is proposed. Firstly， the power range to ensure the communication quality between cellular users and D2D users is determined， and then an improved genetic algorithm is proposed to determine the optimal transmit power of D2D to maximize the system throughput. The algorithm guarantees the quality-of-service-e （QoS） of cellular system， and makes crossover operator and mutation operator change adaptively with evolution algebra， so as to achieve the global optimization. Simulation results show that the proposed algorithm can effectively improve system throughput and channel utilization of D2D users.

Abstract:Traditional PID controllers have problems of the inability to adjust online control parameters， poor control effects， etc. so this paper proposes an intelligent PID controller based on an improved fuzzy neural network. The controller not only combines the reasoning ability of fuzzy control and the learning ability of neural network， but also creatively parameterizes the fuzzy rules so that the fuzzy rules can also be adjusted online， thereby improving the accuracy of control. At the same time， by constructing a new type of activation function—IThLU function， it can effectively avoid the occurrence of gradient disappearance and gradient explosion， and improve the responsiveness of control. The final simulation experiment results show that the intelligent PID controller of improved fuzzy neural network can realize online real-time adjustment of control parameters， improve the responsiveness， stability and accuracy of the system， and is an effective improvement to the PID control algorithm.

Abstract:A versatile data display module based on a stream-processing data acquisition architecture is designed and implemented to solve the reusing problem of the data display module in different DAQ system and to reduce the design workload. This data display module is divided into two parts： display data generation （DG） node and data display （DS） node. As the standard stream-processing nodes in the new DAQ architecture， the DG and DS nodes have same data format and data interface definition with other stream-processing nodes， and can be connected freely to other nodes in the same data domain， while the difference properties of each data domain can be configured by different work parameters. Through the abstract of the display data generation and data display mode in a DAQ system， in the DG node a universal model is used to extract and gather information to be displayed from the input data stream， and in the DS node some visualization methods are provided for different data information display based on the intrinsic dimensions of the data. Thus the universal display demand can be fulfilled by inserting and configuring these two nodes in the DAQ system. The versatile data display module has been used in the DAQ system for offshore seismic exploration and high-energy physics experiment.

Abstract:Aiming at the problems of data discontinuity and slow transmission speed in the traditional flash storage process， a storage method combining double FIFO ping-pong operation reading and writing with four-line flash writing is designed to improve the data storage rate. Through the accurate analysis of chip operation time， the resource utilization rate is improved. The system uses FPGA as the control chip， as well as creates two FIFOs through IP core， which is used for ping-pong reading and writing of data， and uses four chips of two NAND FLASH chips to form four-line operation. The feasibility， storage rate and storage continuity of the pipelined operation are verified by Modelsim simulation and FPGA generated pseudo-random code data write-read experiment and read-data correlation detection test. Combining with the infrared camera to collect and store data， and then through the host computer to read， the correct and continuous infrared image is displayed. By expanding the number of caches and flash chips， the continuity can be ensured and the storage rate can be improved， showing that the system has the characteristics of high storage rate and strong adaptability.