Most of the existing detection methods for concept evolution are essentially based on supervised learning and are often used to solve the problem that only one novel class appears in a period of time. However, they can not handle the task of a class disappearing and recurring in streaming data. To address the above problems, an adaptive detection method for concept evolution based on weakly supervised ensemble (AD_WE) is proposed. The weakly supervised ensemble strategy is used to construct an ensemble learner to make local predictions on the training samples in the data block. Similar data with strong cohesion in the feature space are detected and clustered using local density and relative distance. The similarity of the clustering results is then compared to detect novel class instances and distinguish between different novel classes. And a dynamic decay model is established according to the characteristics of data change over time. The vanished class is eliminated in time, and the recurring class is detected through similarity comparison. Experiments show that the proposed method can respond to concept evolution in a timely manner, effectively identify vanished classes and recurring classes, and improve the generalization performance of the learner.

In order to solve the problem that traditional density peak clustering method without considering similarity of bands in information theory and number of bands in band selection, an adaptive density peak band selection method based on spectral angle mapping and spectral information divergence (SSDPC: Spectral angle mapping and Spectral information divergence Density Peaks Cluster)is proposed. SSDPC combines spectral angle mapping and spectral information divergence for density peak clustering band selection in hyperspectral images, replacing the traditional Euclidean distance to construct a band similarity matrix. By constructing a band scoring strategy, an important subset of spectral bands can be selected automatically and effectively. Using RX(Reed- Xiaoli) algorithm for anomaly detection on three sets of hyper-spectral datasets, the accuracy of anomaly detection is 1. 16% ,1. 18% and 0. 07% higher than that of Euclidean distance measurement under the similarity measure of SSDPC. Under the adaptive SSDPC band selection method, the accuracy of anomaly detection is 6. 49% ,2. 71% and 0. 05% higher than that of the original RX algorithm, respectively. The experimental results show that the SSDPC is robust, can improve the performance of hyper-spectral image anomaly detection and reduce its false alarm rate.

In the field of indoor applications of UWB(Ultra Wide Band) positioning technology, it is important to establish an efficient and accurate 3D coordinate positioning system to overcome signal interference. Machine learning methods are used to investigate the problem of accurate positioning of indoor UWB signals under interference. Firstly, various statistical analysis models are used to clean up invalid or error measurements, then the a priori knowledge of TOF ( Time Of Flight) algorithm is combined with neural network and XGBoost algorithm to build a neural XGB(Exterme Gradient Boosting) 3D oriented system. The system can accurately predict the coordinate value of the target point by “ normal data冶 and “ abnormal data冶 ( disturbed), the coordinates of four anchor points, and the final error is as low as 5. 08 cm in two鄄dimensional plane and 8. 03 cm in three鄄dimensional space. A neural network classification system is established to determine whether the data is disturbed or not, with an accuracy of 0. 88. Finally, by combining the above systems, continuous and regular motion trajectories are obtained, which proves the effectiveness and robustness of the systems.

The optical crossbar chip is the core device used to realize optical routing in the field of optical communication. A control and driver system is designed based on a multi-channel DAC ( Digital to Analog Converter) array to achieve optical routing through the optical crossbar chip. The system consists of a control system module, a multi-channel drive circuit module, and a host computer control module. This system has several advantages, including simple adjustment, bipolar output, more output channels, and higher power accuracy. It solves the problems of the current driving circuit, such as complex operation, single power polarity, fewer output channels, and poor accuracy. The host computer control module can control the driving circuit to apply the control voltage and receive the optical power signal collected from the data acquisition device as the feedback signal of the control driving system. By analyzing the relationship between the control voltage and the received optical power, the best control driving voltage of the optical crossbar chip can be obtained. The system test results show that the system can provide high-precision bipolar driving voltage to effectively drive the optical crossbar chip and can calibrate the control voltage of the optical switch in a short time, fully meeting the requirements of the driving voltage in the active optical crossbar chip control. We believe that this system could be useful for optical crossbar chip control.

To solve the problem that Q function overestimation may cause SAC ( Soft Actor Critic) algorithm trapped in local optimal solution, SCSAC ( Stable Constrained Soft Actor Critic) algorithm is proposed for perfectly resolving the above weakness hidden in maximum entropy objective function improving the stability of Stable Constrained Soft Actor Critic algorithm in trailing process. The result of evaluating Stable Constrained Soft Actor Critic algorithm on the suite of OpenAI Gym Mujoco environments shows less Q value overestimation appearance and more stable results in trailing process comparing with SAC algorithm.

BNNs( Binarized Neural Networks) are popular due to their extremely low memory requirements. While BNNs can be further compressed through pruning techniques, existing BNN pruning methods suffer from low pruning ratios, significant accuracy degradation, and reliance depending on fine-tuning after training. To overcome these limitations, a filter-level BNN pruning method is proposed based on evolution from ternary to binary, named ETB ( Evolution from Terry to Binary). ETB is learning-based, and by introducing trainable quantization thresholds into the quantization function of BNNs, it makes the weights and activation values gradually evolve from ternary to binary or zero, aiming to enable the network to automatically identify unimportant structures during training. And a pruning ratio adjustment algorithm is also designed to regulate the pruning rate of the network. After training, all zero filters and corresponding output channels can be directly pruned to obtain a simplified BNN without fine-tuning. To demonstrate the feasibility of the proposed method and the potential for improving BNN inference efficiency without sacrificing accuracy, experiments are conducted on CIFAR-10. ETB is pruned the VGG-Small model by 46. 3% , compressing the model size to 0. 34 MB, with an accuracy of 89. 97% . The ResNet-18 model is also pruned by 30. 01% , compressing the model size to 1. 33 MB, with an accuracy of 90. 79% . Compared with some existing BNN pruning methods in terms of accuracy and parameter quantity, ETB has certain advantages.

fMRI ( functional Magnetic Resonance imaging) is an efficient research method for brain imaging technique. In order to reduce the redundancy of the fMRI data and transform the fMRI data to the constructed features with more classification potential, a feature construction method based on the siamese network named as SANet(Siamese Network) is proposed. It engineered the brain regions features under multiple scanning points of an fMRI image. The improved AlexNet is used for feature engineering, and the incremental feature selection strategy is used to find the best feature subset for the encephalopathy prediction task. The effects of three different network structures and four classifiers on the SANet model are evaluated for their prediction efficiencies, and the ablation experiment is conducted to verify the classification effect of the incremental feature selection algorithm on the SANet features. The experimental data shows that the SANet model can construct features from the fMRI data effectively, and improve the classification performance of original features.

Aiming at the privacy protection requirements of existing image datasets, a privacy-preserving scenario of image datasets and a privacy-preserving image alternative data generation method is proposed. The scenario is to replace the original image dataset with an alternative image dataset processed by a privacy-preserving method, where the substitute image is in one-to-one correspondence with the original image. And humans can not identify the category of the substitute image, the substitute image can be used to train existing deep learning images classification algorithm, having a good classification effect. For this scenario, the data privacy protection method based on the PGD ( Project Gradient Descent) attack is improved, and the attack target of the original PGD attack is changed from the label to the image, that is the image-to-image attack. A robust model for image-to- image attacks as a method for generating alternative data. On the standard testset, the replaced CIFAR(Canadian Institute For Advanced Research 10)dataset and CINIC dataset achieved 87. 15% and 74. 04% test accuracy on the image classification task. Experimental results show that the method is able to generate an alternative dataset to the original dataset while guaranteeing the privacy of the alternative dataset to humans, and guarantees the classification performance of existing methods on this dataset. 

Aiming at the problem that the traditional named entity recognition method can not fully learn the complex sentence structure information of ancient Chinese and it is easy to cause information loss in the process of long sequence feature extraction, an ancient Chinese fusion of SikuBERT ( Siku Bidirectional Encoder Representation from Transformers) model and MHA (Multi-Head Attention) is proposed. First, the SikuBERT model is used to pre-train the ancient Chinese corpus, the information vector obtained from the training into the BiLSTM (Bidirectional Long Short-Term Memory) network is input to extract features, and then the output features of the BiLSTM layer are assigned different weights through MHA to reduce the information loss problem of long sequences. And finally the predicted sequence labels are obtained through CRF (Conditional Random Field) decoding. Experiments show that compared with commonly used BiLSTM-CRF, BERT-BiLSTM-CRF and other models, the F1 value of this method has been significantly improved, which verifies that this method can effectively improve the effect of ancient Chinese named entity recognition.

Automatic analysis of lumbar disc herniation requires precise segmentation of the multifidus muscle’s fatty infiltration site in spinal MRI ( Magnetic Resonance Imaging) images. An attention-based approach for segmenting the multifidus muscle in lumbar disc herniation patients is proposed to address issues including ambiguous boundaries between segmentation targets and adjacent components. The network utilizes an encoder- decoder structure, and the addition of an attention mechanism module to increase the network segmentation accuracy. After feature extraction, an atrous spatial pyramid pooling module is added to combine contextual data improving the performance of the network model. In comparison to the traditional U-Net algorithm, the experimental results demonstrate that this model improves the segmentation accuracy of the fatty infiltrated regions of multifidus muscle by improving the Dice coefficient by 7. 8% , Jaccard similarity coefficient by 10. 1% , and Hausdorff Distance by 69. 5% . 

Federated learning has become increasingly important for modern machine learning, especially for data privacy sensitive scenarios. It is difficult to carry out secure machine learning between heterogeneous data islands. A federated learning communication mode between heterogeneous data islands is proposed, which realizes the hybrid federated learning communication between horizontal and vertical, and breaks the communication barrier of the disunity of model structure between horizontal and vertical participants in traditional federated learning. Based on the special privacy requirements of the government, banks and other institutions, the third party aggregator is further removed on the basis of the hybrid federated learning model, and the calculation is carried out only among the participants, which greatly improves the privacy security of local data. In view of the computational speed bottleneck caused by vertical homomorphic encryption in the communication process in the above model, by increasing the local iteration round q, the encryption time of vertical federation learning is shortened by more than 10 times, and the computational bottleneck between horizontal and vertical participants is reduced, and the accuracy loss is less than 5% .

In order to solve the problem of computer aided diagnosis of novel coronavirus pneumonia (Covid-19: Corona virus disease 2019), a bilinear convolutional neural network model is created and a feature extraction subnetwork with VGG(Visual Geometry Group network) 16 and VGG19 is employed. The algorithm is applied to COVID-19 image classification and compared with the basic image classification algorithm. The results and lesion visualization analyses demonstrate that the bilinear convolutional neural network model outperforms other deep learning network models in terms of accuracy, with an accuracy of 95. 19% . By replacing softmaxlayer and using SVM(Support Vector Machines) classifier, the model classification accuracy is improved to 96. 78% . The study provides a trustworthy tool for the quick and accurate diagnosis and treatment of neonatal pneumonia and a confirmation of the viability of fine-grained imaging algorithms for the categorization of COVID-19 CT images. 

 With the advent of big data era, the development of recommendation systems has become more and more vigorous. It has become a research hotspot to push information that may be of interest to users in a timely manner among massive amounts of information. Traditional recommendation algorithm lack implicit information and contextual information about graph structures. In response to this, a recommendation model is proposed based on graph neural network. The main innovations are: 1) Based on the higher-order connectivity theory of graphs, the graph neural network is used to mine the hidden information in the user-item bipartite graph, and a the order is extended to multiple orders, so as to obtain more accurate embedded representation and recommendation effect; 2 ) Consider context information in the update process, which is conducive to understanding the interaction between contexts. The model is tested on the Yelp-OS, Yelp-NC and Amazon-book datasets, and the results show that it is better than the related comparison algorithms in both HR(Hit Ratio)and NDCG(Normalized Discounted Cumulative Gain) indicators, which proves that the algorithm can optimize the recommendation effect and improve the recommendation quality. 

In view of the safety problems caused by low detection efficiency and illegal operation in traditional gas stations, a modified gas station oil unloading detection method based on YOLOv3 is presented. By introducing RFB (Receptive Field Block ) receptive field module after Darknet-53 backbone output, the model can select appropriate receptive fields to match different scale targets and improve detection accuracy. According to CSP (Cross Stage Partial) network, two RFB_CSP and RFBS_CSP structures are provided to realize cross-level splicing and channel integration and reduce calculation cost. Cluster 9 target reclustering in the field is realized by using K-means++ algorithm to determine appropriate network anchor parameters. The experimental results show that the optimized model contrasts the original YOLOv3 model. The average accuracy is increased by 2. 3% and 2. 9% , indicating that the optimized YOLOv3 model has high practical value in gas station scene detection.

In order to improve the subdivision accuracy of the photoelectric encoder and realize the compensation
and correction of the Moire fringe signal, the source of the error that affects the signal quality is analyzed, and
the waveform equation of the Moire fringe signal is established. And a method using the cuckoo search algorithm
combined with the least square method is designed, and using the residual sum of squares as the fitness function
to realize the multi-parameter identification of the waveform equation. And the sinusoidal deviation, DC(Direct
Current) component and amplitude deviation in the signal are corrected by using the identified parameters and
the waveform equation. The identified phase parameters are used to construct a look-up table to compensate the
orthogonality deviation of the signal. The experimental results show that the method effectively reduces the
deviation in the output signal and improves the subdivision accuracy of the photoelectric encoder.

To solve the problem of incomplete information in air combat, research of offensive and defensive game strategy of UAV (Unmanned Aerial Vehicle) based on an improved sparrow algorithm is conducted. By analyzing the air combat information of the enemy and ourselves, the offensive and defensive game model is established in the case of determining the air combat situation advantage function, air combat performance advantage function and profit function of the enemy and ourselves. Then, combining the game payout function given by the game model and calculating the payout matrix of the enemy and ourselves, an improved sparrow algorithm based on reverse learning is proposed to solve the Nash equilibrium solution of the offense and defense game between the enemy and ourselves. Finally, the feasibility and the effectiveness of the proposed algorithm are verified through simulations. This scheme can preliminarily solve the problem of incomplete information encountered in the process of air combat confrontation.

In the work of scientific and technological literature retrieval, it is very important to give a reasonable correlation ranking from a professional point of view. The traditional PageRank algorithm uses the method of evenly distributing similarity weights, but this method will cause the unreasonable results of literature ranking. Therefore, an algorithm combining deep learning method and PageRank is proposed to improve the reliability of literature relevance ranking. Firstly, the Siamese BERT ( Bidirectional Encoder Representation from Transformers) network with attention pooling is used to calculate the similarity between literature and citations, and then the similarity between literature and citations contained in literature is normalized. Finally, the normalized similarity is used as the distribution weight to calculate the ranking results of citation network. The experimental results show that compared with the traditional PageRank algorithm, the correlation of the retrieval results of this method is improved by more than 6% , which is more suitable for citation network analysis of scientific and technological literature.

The traditional deadbeat current predictive control depends on accurate discrete power grid mathematical model, which leads to steady-state error of the current, especially the system instability when the inductance changes beyond a certain range. There is sampling delay in the all digital control, and the system robustness is poor. An improved deadbeat predictive control method combined with repetitive control is proposed to effectively eliminate the command error and disturbance error of grid connected current and provide high-quality steady-state waveform. Simulation results show that the influence of sampling delay caused by traditional control methods is effectively suppressed. It has the advantages of fast response, high steady-state accuracy and small current distortion rate (THD: Total Harmonic Distortion), and avoids the disadvantages of overshoot and oscillation.

At present, the fixed time operation mode is adopted during the job scheduling of tunnel cleaning robot. This would waste the power and the fire resources and affect the driving safety. Aiming at this problem, an intelligent scheduling method of tunnel cleaning robot is presented. Firstly, the time varying operation character of tunnel infrastructure is described according to the influence factors, such as traffic flow, temperature, humidity and so on. Then, considering the operation performance of infrastructure, the operation cost of robot, the impact on tunnel traffic flow and so on, the cascading optimization scheduling model of tunnel cleaning robot is constructed. The scheduling cycle and operation time of cleaning robot are optimized respectively to realize the efficient operation of robot. In the experiment, the proposed method is compared with the commonly used fixed time scheduling method. The results show that at the same operation cost, the infrastructure operation performance increases 6% by using the proposed method compared to the fixed time scheduling method.

When natural disasters such as earthquake and tsunami occur. Base stations on the ground are damaged or satellite signals are interfered, which makes it impossible to provide accurate location information, greatly hinders disaster relief search and post-disaster reconstruction. To solve this problem, a wireless localization scheme for target nodes in the GNSS ( Global Navigation Satellite System) rejection scenario is proposed. In this scheme, the rescue vehicle is selected as the reference node, so that the single UAV (Unmanned Aerial Vehicle) can carry out multi-point wireless ranging around a fixed trajectory centered on the reference point. By using the mobility of the UAV, the traditional multi-UAV floating ranging process is changed to single UAV moving multi-point ranging, which effectively solves the problems of resource waste, difficult clock synchronization and high system complexity. In the ranging process, a two-way ranging method based on weighting is proposed, which can effectively reduce the two-way channel error and clock offset error. Finally, the least square method is used to solve multiple groups of ranging information to obtain the relative position information from the target node to the rescue vehicle. Simulation and experimental results show that the positioning method proposed is reasonable and superior to traditional methods in performance.

Recently, with the rapid improvement of Chinese economy and people's living standards, the demand for the exploitation and transportation of energy is continuously increasing. As one of the important transportation modes, the length of petroleum and gas pipelines in China has increased greatly. Pipeline leakage detection technology is one of the basis for efficient and safe pipeline transportation. This article firstly introduces the development history and existing research achievements of pipeline leakage detection and location technology, and analyzes various common methods. Secondly, the fundamentals and operation manners of numerous routine pipeline leakage detection approaches are described. Then, based on the actual industrial situation, the characteristics and application scenarios of each detection method are described. At last this article analyzes the challenges of the mentioned detection technologies, and looks forward to the development direction of pipeline leak detection technology.

Edge detection is a basic problem in digital image processing. Its purpose is to detect the pixels whose gray level changes obviously in the neighborhood. The Canny edge detector is currently the most popular edge detection tool. The specific implementation of Canny detector in the quantum computing paradigm is studied. For Gaussian smoothing filtering and Sobel sharpening operators, we have designed a new method called Translation, Stacking and Weighted Summation, which can make full use of the parallelism of quantum computing to accelerate its classical counterpart and avoid convolution operation. For the gradient and angle calculations required in edge detection, we design new operators such as addition,multiplication and division of signed number by introducing the binary complement description of gray-scale value. For the non-maximum suppression and double threshold processing required in edge detection, we have separately designed the quantum circuits that implement these tasks by introducing quantum complement comparators. Complexity analysis shows that the quantum Canny edge detector has exponential speedup compared to its classical counterpart. The simulation results on the classical computer verify the effectiveness of the proposed method, and reveal that the research idea of integrating quantum computing and image processing is feasible.

In order to give full play to the overall combat superiorities of the aircraft cluster to obtain optimal target allocation plan, we use an improved wolf pack algorithm to solve the battlefield situation model. In order to improve the global optimization ability of the algorithm and ensure the optimization efficiency of the algorithm,the concept of the second wolf is introduced to improve the calling and siege behavior of the wolf pack, and the update mechanism of the wolf pack algorithm is optimized. The simulation results show that the proposed method can quickly and accurately find the optimal objective function value, and to a certain extent improves the situation that the traditional wolf pack algorithm is easy to fall into the local optimum.

Aiming at the shortcomings of traditional pipeline welding robots such as long start-up time, slow speed response, insensitive speed change, and slow climbing speed, and the problem of variable parameters and multiple disturbances during the orbital motion of the welding robot, an improved FSMC ( Fuzzy Sliding Mode Control) is proposed. The discontinuous intelligent control method combined with fuzzy PI(Proportional Integral) realizes double closed loop control. The inner loop is the current loop control based on fuzzy PI, and the outer loop is the speed loop control based on the improved fuzzy sliding mode algorithm. This kind of control effectively reduces the external disturbance of the orbital motion system, and increases the response speed and robustness of the system. For the speed setting and excessive speed change under variable load, variable speed, and variable welding mode, a follow-up is proposed. The system adopts a tracking and setting system of subregional fine target speed, which effectively divides the range of different welding methods of the welding robot and the precise feedback mechanism. After simulation and experimental testing, the entire control system has the functions of parameter self-tuning and self-adjustment during the welding process. The walking robot has smooth speed transition, fast speed response, good robustness, and stable and reliable pipeline all-position welding.

With the rapid development of the fifth-generation communication technology, the amount of communication data has grown rapidly. In order to solve the problem of low communication efficiency, a data receiving system with higher data transmission rate and higher frequency efficiency is proposed. By analyzing the application of AD9361 technology in wireless communication systems, a MIMO ( Multiple-Input Multiple-Output)、 OFDM (Orthogonal Frequency Division Multiplexing) synchronous timing receiving system based on AD9361 is designed. This system has the advantages of high spectrum utilization, flexible demodulation mode and fast communication speed having important significance in 5G construction. The test result shows that MIMO- OFDM synchronous timing receiving system based on AD9361 has high convergence and the FPGA ( Field- Programmable Gate Array) resource occupancy of the system is generally good, and the collected data is relatively stable. In addition, during the synchronization of data receiving the system can effectively avoid the influence of bit error rate, without ISI(Inter-Symbol Interference) and ICI(Inter-Channel Interference), and the data receiving system meets the transmission channel standard which has high wireless communication stability.

Ontology fusion has become an important way to reconstruct knowledge map and share knowledge in the subject field. In order to solve the problem of lack of standard fusion framework in ontology fusion field, we propose an ontology fusion model based on international standard MFI4OR. This model provides a standard ontology information partition standard, namely ontology-ontology component-ontology atomic component, and realizes the management and mapping of ontology information. In the process of fusion calculation, we choose editing distance algorithm and introduces external resources WordNet dictionary to calculate similarity. Finally, based on the application background of learner model construction requirements in provincial fund projects, the FOAF (Friend-Of-A-Friend)and RELATIONSHIP ontologies are fused. The results show that the standardized fusion model can achieve fusion and has good recall rate and accuracy.

Real natural images are often contaminated by various kinds of image noise. Traditional denoising methods are generally designed for only one type of noise, so the denoising effect is not good when dealing with mixed noise. To solve this problem, a fast two-step cross non-local hybrid filtering algorithm is proposed. Firstly, the extremum point of pixel gray level is located. The non local median filter is used to remove salt and pepper noise, and the difference integral image of the image is obtained. And then the improved non mean filter is used to further remove the noise. Finally, the non local median filter is used to further remove the noise. The experimental results show that the proposed algorithm achieves higher measurement index and better visual effect in the case of high intensity mixed noise pollution.

Aiming at the large number of defects at the electron transport layer/ perovskite layer interface of perovskite solar cells, an optimized strategy employing an inorganic salt for interface passivation is proposed. This strategy selects low-cost LiCl(Lithium Chloride) as the interface passivation material between the electron transport layer and perovskite layer, and prepares the perovskite solar cell with the device structure of ITO/ TiO₂ / LiCl / CH₃NH₃PbI₃ / spiro-OMeTAD/ Ag. After the introduction of LiCl with an optimized concentration, the short- circuit current density and fill factor of the perovskite solar cell achieve 21. 05 mA/ cm ² and 72. 55% , and the energy conversion efficiency is 16. 95% , which shows an increase of 23. 00% compared with the device without the introduction of LiCl. After characterizing the device and the film, it is found that LiCl can passivate the defects and traps at the interface and increase the conductivity of TiO₂ , thereby reducing the interface recombination loss and promoting the charge transport.

In order to alleviate the new load peak situation caused by the existing time-of-use pricing mechanism, a power consumption incentive mechanism which can contribute to the load regulation is proposed. At first, all kinds of household electric equipment load model, load model for the battery electric vehicle charging and discharging model are set up. Secondly the load side of indirect adjustment grid peak valley incentive mechanism model is established. Finally to electricity cost and user comfort as the objective function, based on niche chaotic particle swarm optimization algorithm is used for solving multi-objective Pareto solutions. The simulation results show that the proposed incentive mechanism can meet the comfort requirements, and significantly reduce the cost of electricity consumption and the peak-valley difference of electricity consumption, which can relieve the power grid pressure and improve its operation stability to a certain extent.

With the development of information technology, a large amount of text classification is needed in many industries. In order to improve the accuracy and the efficiency of classification at the same time, a kind of CNN-BiLSTM/ BiGRU mixed text classification model based on the attention mechanism(CBLGA) is proposed, in which parallel CNN(Convolution Neural Networks) with different window sizes to extract a variety of text characteristics, then input the data in BiLSTM/ BiGRU parallel model. BiLSTM/ BiGRU combination model is used to extract global characteristics relate with the whole text context, finally the characteristics of two models are fused and the Attention mechanism is introduced. Secondly, another kind of Attention of CNN-BiLSTM/ CNN mixed text classification model based on the attention mechanism(CBLCA) is proposed, and its feature is divided CNN's output into two parts. One part is input to the BiLSTM network, another is integrated to the output of BiLSTM network. Successfully retaining the partial text features extracted by CNN and the global text features extracted by BiLSTM. After several experiments, the CBLGA model and CBLCA model is achieved effective improvements in accuracy and efficiency. Finally, a set of preprocessing methods for texts with different lengths is established, so the model can improve the accuracy and efficiency of text classification target in long text and short text.

Aiming at the problem of insufficient concentration and effective utilization of the existing tourist data of Changbai Mountain scenic spots, through reasonable development of a Python-based web crawler, we have realized the crawling of some Changbai Mountain tourism data, and used Tableau tools to make a visual analysis of the data. Accurately excavating the potential relationship between the number of tourists and various factors from multiple dimensions presents a more intuitive effect, which is conducive to the observation of trend distribution, and lays the foundation for the further development of reasonable tourism strategies in the Changbai Mountain area.

According to the teaching objectives and requirements of the electronic measurement experiment, an experimental teaching platform for electronic measurement is developed, which introduces the advanced technology of electronic measurement. It solves the problem that the function of the existing electronic measurement experimental teaching platformcan does not satisfy the multi-level personnel training. The platform adopts modular and open design idea, in which a signal source module based on DDS(Direct Digital Freqiaency Synthesizers) technology, a bandpass filter module, an AC(Alternating Cunent) voltage parameter measurement module, a digital frequency meter module based on DDS technology are provided. It provides a practical platform for students to verify theoretical knowledge. And it provides an open hardware platform and resources for comprehensive design experiments. This platform enables students to master the theoretical knowledge of electronic measurement in depth, and exercises innovative thinking and practical ability, and finally realizes a high level multi-level talented person training goal.

A variable gain iterative learning control law is proposed to solve the problem that the tracking efficiency of nonlinear 3-DOF(three-Degree-Of-Freedom) manipulator's trajectory tracking control system is low in the presence of disturbance. First, the Lagrangian method is used to establish the kinetic equation. Then, the variable gain iterative learning controller with three degrees of freedom manipulator structure is designed and its convergence is analyzed. Finally, through Matlab Simulink simulation module, the three-degree-of-freedom manipulator control system simulation diagram is constructed, and the comparative simulation test of closed-loop fixed constant and closed-loop variable gain PD(Proportion Differentiation) iterative learning control is carried out. Simulation results show that the three-degree-of-freedom manipulator based on variable gain iterative learning can converge to the desired trajectory in a short time in the presence of interference.

With the widespread use of knowledge graphs, in order to improve the accuracy and efficiency of extracting knowledge data and product data from them, the method and system use a knowledge graph as a data source, and formulates business data extraction and organization rules based on actual business requirements (the extraction rules are the mapping rules in the title, the expression description methods and specification constraints of the design rules are designed by us, and the actual requirements can be filled out by the business demander), and support the extraction of subgraphs that meet the rules from the knowledge graph according to the rules. Because the subgraph conforms to the rules of the business demander, the subgraph contains the data and organizational structure that meet the business requirements. Further, through data product generation rules (generating data products that are ultimately required by business users, such as report files and statistical tables, from subgraph data with relatively fixed structure and actual business meaning), generate the required data products from the extracted subgraphs ( report documents, statistical tables, etc). Rapid and automatic generation of data products such as text, charts, and report documents are achieved by using SPARQL query language, natural language generation, and other technologies to use knowledge graphs as data sources, which has substantially improved efficiency.

Aiming at the problem of poor performance of traditional Faster R-CNN(Region-Convolutional Neural Networks) in detecting small target defects on steel surface, a new method for steel surface defect detection based on improved Faster R-CNN is proposed. First, the GA-RPN(Guided Anchoring Region Proposal Network) is introduced to predict the position and shape of the anchor points, and an adjustable mechanism is designed to solve the problem that the shape offset of network anchors exceed the region of interest, thereby solving the influence of irrelevant features. Then, a multi-task FPN (Feature Pyramid Network) structure is used to shorten the high-level feature location information mapping path, and can solve the insufficient features fusion of adjacent layers features fusion and re-sampling, and to improve the performance of small target detection. The results show that the recall rate and accuracy of the network are improved. Therefore, this method has better performance and can effectively detect steel surface defects.

The traditional LE(Laplacian Eigenmaps) algorithm uses Euclidean distance to measure the position relationship between sample points, which is only applicable to linear data sets. However, the data in practical engineering often show strong non-linearity, which makes the final embedding results difficult to reflect the essential characteristics of the original data. An algorithm for D-LE(Double metric constraint Laplace Eigenmaps) based on Double metric constraint is proposed. The algorithm uses cosine similarity to evaluate the similarity between samples, and combines the measurement relations between samples and between samples and local manifolds to build dimensionality reduction model. Experiments on three bearing datasets show that this method can significantly improve the dimensionality reduction effect for processing nonlinear datasets.