To monitor changes of indoor carbon dioxide concentrations in real time, a carbon dioxide alarm based on a gas sensor with a STC89C52 microcontroller as the core is designed. When the CO2 concentration in the air exceeds the preset value, the sound and light alarm function can be activated, and the indoor CO2 concentration value can be displayed in real-time. The hardware system includes a carbon dioxide sensor, signal conditioning circuit, analog-to-digital conversion circuit, STC89C52 microcontroller, and acousto-optic alarm unit. The software system includes data acquisition, data processing, alarm logic, and other functional units. The alarm can be activated in time when indoor CO2 concentration exceeds 1. 5% . 

Taget Detection of Photovoltatic Remote Sensing Based on Improved Yolov5 Model TONG Xifeng, DU Xin, WANG Zhibao (School of Computer and Information Technology, Northeast Petroleum University, Daqing 163318, China) Abstract: Aiming at high-sensing photovoltaic image resolution, high environmental noise, and complex background, an improved Yolov5 model is proposed to achieve positioning of photovoltaic power plants. First of all, the CA(Coordinate Attention) mechanism is added to the compassionate layer of the main feature extraction network to improve the learning ability of the network characteristics; second, the Ghostconv network structure is added to Backbone, useing the Ghostconv network module to replace the Conv network module, designing a new GhostC3 network network instead of the original C3 network module to improve the learning efficiency of the model; finally, the GIoU_Loss function is changed to the SIoU_Loss function. Compared with the original Yolov5 method, the average accuracy of the improved algorithm mAP, accuracy, and recall rate reached 97. 5% , 98. 9% , and 94. 9% , respectively, which have increased by 1. 8% , 1. 7% , and 5. 8% , respectively. The algorithm has a good effect on photovoltaic detection.

 In response to the chattering phenomenon that traditional SMO(Sliding Mode Observer) faces during switching functions, novel sliding mode observer using saturation function instead of switching function is proposed to weaken chattering. In the process of extracting position information, a phase-locked loop is selected to replace the traditional arctangent method, there by improving the observation accuracy of PMSM(Permanent Magnet Synchronous Motor) rotor position. In the Matlab environment, by comparing traditional SMO and new SMO, it can be observed that the speed error of the rotor has increased by about 14 r/ min, and the position error of the rotor has increased by about 0. 03 rad.

 Microseismic inversion is an important way to complete main task of microseismic monitoring by inverting and inferring information such as the location of the epicenter, the time of occurrence of the earthquake, the true magnitude, the initial amplitude of the hypocenter, the focal mechanism and the medium parameters based on the microseismic monitoring data. Through the study of microseismic inversion technology, more accurate microseismic information can be obtained, thereby improving the reliability of reservoir fracture evaluation, reducing development costs and improving oil and gas recovery. We review microseismic inversion from three aspects: microseismic focal location inversion, microseismic focal mechanism inversion, and microseismic multi-parameter joint inversion. The research progress of microseismic inversion technology in recent years is reviewed, and the principle, advantages and disadvantages of various microseismic inversion methods are summarized, the improvement and application of various microseismic inversion techniques are summarized, and the future research ideas and development directions are prospected. It provide reference for further development of microseismic inversion in the future.

To reduce the influence of complex and changeable electromagnetic environment on vehicles, the cable model is built in the engine compartment by using the three-dimensional electromagnetic field simulation software CST(Computer Simulation Technology) to study the influence of different factors on the electromagnetic coupling effect of vehicles. Through simulation, the peak relationship curves between induced voltage and induced current in the cable are drawn when the parameters such as cable length, cable height from the bottom of the car, cable relative distance, cable terminal resistance, conductor radius, and insulation layer thickness change. These conclusions can provide theoretical guidance for the design of vehicle wire harnesses, and provide a basis for the conductor radius selection, cable relative distance, height from the ground, and cable length in electromagnetic protection design. 

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.

There is a deficiency of information available on waste classification, and many municipalities and educational institutions struggle with this issue. We address this challenge by utilizing the efficiency and accuracy of the neural networks to classify items and implement waste image classification with a deep residual shrinkage network built on the ResNet network and SENet network. By filtering the Garbage dataset to obtain the data set necessary for the experiment, and by enhancing ResNet, SENet and soft threshold processes are incorporated into the ResNet structure. And by training the network and optimizing its hyperparameters, a greater recognition rate and recognition effect are achieved for the classification of campus waste. The experimental findings indicate that the proposed approach is feasible to a certain extent.

 In order to improve the invulnerability and working ability of the satellite constellation network after being attacked, a simulated annealing wolf pack algorithm is proposed. We use the subjective and objective weight method combined with the TOPSIS( Technique for Order Preference by Similarity) to Ideal Solution to evaluate the importance of nodes in the network, and attack the network according to the order of node importance. The network connection efficiency is the optimization goal, and the satellite constellation network communication limitation is the constraint condition. The idea of motion operator is adopted to realize the walking, summoning and sieging of wolves with adaptive step size. The network structure is optimized using the edge-adding scheme obtained through optimization. Experiments show that compared with other optimization algorithms, this algorithm has superiority. It solves the problem that the satellite constellation network爷s working ability declines after being attacked, and improves its invulnerability after being attacked. Key words: satellite network; invulnerability optimization; simulated annealing algorithm; improved wolf colony algorithm

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.

To meet the need of integrating scientific research into teaching of Emerging Engineering Education, a multi-dimensional and hierarchical integrated experimental platform based on Python is designed. Guided by the talent-training plan, hierarchical modules involving image recognition, machine learning and data analysis is designed from scientific research hotspots. Image recognition module starts from character recognition, then face and license plate recognition are realized by several algorithms. In the machine learning module, commonly used machine learning algorithms are studied and corn disease is identified by various methods based on Python. In the data processing and analysis module, Excel data processing experiment based on Python is designed to analyze the data of workload and bioinformatics data. The platform enables students to learn the application of Python in the experiments, and choose different experimental projects according to professional needs and research directions to realize the goal of teaching students in accordance with their aptitude. By applying the experimental platform to teaching practice, it is demonstrated that students have a deeper understanding of Python’s programming implementation in image recognition, machine learning, and data analysis and enhanced research interest. And the goal of integrating scientific research into teaching and improving the quality of undergraduate teaching could be achieved.

 In order to explore the ability of imaging to diagnose benign and malignant breast lesions, and compare the value of MR(Magnetic Resonance) radiomics and traditional MRI(Magnetic Resonance Imaging) in diagnosing breast diseases, a total of 190 cases with benign or malignant breast lesions confirmed by pathological findings are collected from patients who underwent MR Plain and enhanced examination in the Department of Radiology in First Hospital of Jilin University from January 2019 to January 2022. MR radiomics is performed by building logistic regression model. The traditional MR Diagnosis is performed by a radiologist with an associate senior title. The results show that the sensitivity, specificity and AUC (Area Under Curve) of the MR radiomics test set are 0. 92, 0. 83 and 0. 92 respectively. The above values are higher than the corresponding values of traditional MR diagnosis, and the differences are statistically significant ( P = 0. 00 ). The method of MR radiomics can assist in the diagnosis of benign and malignant breast lesions, and the diagnostic ability is better than the traditional MR diagnostic mode.

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% .

The computation of Weil pairing of elliptic curves over finite fields is of great significance to the application of public-key cryptography. The research of Weil pairing focuses on theoretical research, but pays little attention to practical application, which leads to the need for new methods to support some theoretical research. For this reason, the Weil pairing calculation method is given, the point selection problem of Miller algorithm over finite field is pointed out by examples, and the difference between two different methods using Miller algorithm is analyzed. Through Miller algorithm, the limitation of MOV( Menezes-Okamoto-Vanstone) attack discrete logarithm is pointed out. The practical analysis shows that the Weil pairing of elliptic curves in finite fields are of small order and not very effective for large suborder.

In traditional statistical methods, the conjoint analysis method is can not estimate variables for a large number of parameters at the same time, therefore, a Bayesian 茁 regression model is proposed. In the newly established model, the Dirichlet distribution is used as the prior distribution of the model parameters, and the relevant MCMC(Markov Chain Monte Carlo) algorithm is designed to fit the model. By analyzing the results of applying the model to the evaluation of discrete index variables, it is shown that the model has a good fitting effect on the data and the algorithm has a fast convergence speed. It shows that the Bayesian hierarchical model makes up for the defects of the traditional conjoint analysis method, and optimizes and improves the conjoint analysis method.

To make AGV ( Automated Guided Vehicle ) work efficiently in various environments, it is necessary to select a suitable path planning algorithm according to the actual terrain. We use the A* and DWA ( Dynamic Window Approach ) hybrid path planning algorithm and build four typical terrains,U-shaped, S-shaped, L-shaped and narrow passage in the simulation environment to conduct pathfinding experiments. Furthermore, we improve the weight recursive formula of Gmapping, remove the dependence on the previous moment data and improve the efficiency of the algorithm. The results show that the hybrid path planning algorithm has faster pathfinding speed and better obstacle avoidance ability than the single algorithm. It has the fastest pathfinding speed in L-shaped terrain and was relatively slow in U-shaped terrain and S-shaped terrain.

To inherit and protect the traditional blue calico patterns, it is necessary and challenging to protect blue calico in a digital manner. Due to the lack of blue calico pattern datasets with original manual characteristics, which limits the application of powerful deep learning technology on pattern recognition field like the recognition of Blue Calico patters, therefore a large-scale dataset called Blue-Calico pattern dataset is provided, which is the first publicly available benchmark for blue calico pattern recognition. This dataset contains about 50 216 blue calico patterns, covering 85 pattern classes, such as animals, plants, myths and legends. The construction of this dataset will concern the digital construction of blue calico, such as calico image retrieval and caption, and enable researchers to design and validate data-driven algorithms. On the basis of the new dataset, the experimental results of four state-of-the-art networks are provided as a baseline for future work.

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.

The traditional mobile robot based on DWA ( Dynamic window Approach ) algorithm exists the following deficiencies: longer obstacle avoidance time and the inability to optimize the local path planning in obstacle-intensive dynamic zone. Aimed at the problems mentioned above, a hybrid path algorithm combined A * algorithm with VO(Velocity Obstacle) is proposed to optimize the velocity of obstacle avoidance for mobile robots. Via the comparative experiment combined the DWA algorithm with the VO hybrid path algorithm in the case of three obstacles, the ROS (Robot Operating System) adopted the modular software design is put into practice to test the obstacle avoidance effect of the hybrid path planning algorithm. The results of simulation experiment in multiple environments clearly indicate that the obstacle avoidance effect will be significantly improved via the VO hybrid path algorithm in the scenarios scattered with multiple dynamic obstacles, and it has high speed of the obstacle movement and low frequency of radar scanning.

Approximately 3% to 10% of hepatitis C cases can develop to hepatocellular carcinoma after viral hepatitis C virus infection. Worldwide, 27% of cirrhosis statistics are due to hepatitis C and 25% are due to hepatocellular carcinoma. Accurate prediction of hepatitis C infection is a matter of urgency. Machine learning is fast and accurate. Hepatitis research often used time series analysis or pathological analysis in the past and did not use machine learning algorithms as an auxiliary diagnosis method for hepatitis C. To select the optimal model for detecting hepatitis C, different machine learning models are compared and analyzed in UCI(University of California Irvine) hepatitis C data. The experimental results show that gradient boosting tree, random forest and light gradient boosting machine perform better, among which the gradient boosting tree is accurate in predicting hepatitis C up to 0. 935 1. The most accurate prediction of hepatitis C infection is performed using gradient boosting tree.

In order to solve the problem of low accuracy in user similarity calculation of traditional collaborative filtering algorithm, an item attribute and item tag information is introduced into the recommendation system, and proposes a hybrid recommendation algorithm is proposed based on tags and attributes. Firstly, the user's score on the item is transformed into the user's score on the item attribute value and label, and the user-attribute rating matrix and user-tag rating matrix are constructed as user description files. Then the similarity between users is calculated according to the user-attribute rating matrix and user-tag rating matrix, and the results are average weighted to obtain the nearest neighbor list of each user. Finally, the recommendation result is generated according to the neighbor's score on the item. Since the number of item attributes and major tags are much lower than the number of items, the algorithm can effectively solve the sparsity problem of collaborative filtering algorithm, and describe the user preference more intuitively. In the process of constructing the user description file, considering the law that the user preference changes with time, different weights are given to the user's scores at different time points, and the weight increases gradually with the passage of time. Experimental results show that the proposed algorithm can predict users' ratings of unrated items more accurately and improve the accuracy and recall of recommendations.

In order to solve the problems of management, coordination and association of cross regional scientific and technological resource sharing and collaborative innovation, practical application and theoretical research of scientific and technological resource platform are carried out to build an ecological digital display system of scientific and technological innovation. First, the definition, composition and index system of scientific and technological innovation elements are constructed, and innovation map, innovation ability, innovation subject, innovation carrier are proposed. The digital display system with innovation resources and innovation environment is the core. Taking Jilin Province as an example, the index method is used to monitor the level of regional scientific and technological innovation, the text mining method is used to analyze the hot spots of scientific and technological policies, and the system architecture and function design are carried out based on the Vue + Django + MySQL technology architecture. Finally, the ecological digital display of scientific and technological data innovation in the region is realized, so that the system has data sharing, data association and intelligent service functions such as decision support. The application practice shows that the system enriches the visualization of scientific and technological data resources, improves the visualization of scientific and technological innovation ecology, and improves the expansibility, efficiency and performance of the system.

An improved artificial potential field method is proposed to solve the problems of local minimum and unable to reach the target in the path planning of mobile robots. Firstly, in order for the robot to reach the target point when there are obstacles near the target point due to the large repulsive force, a safe distance factor is introduced into the potential field, and this parameter is optimized, so that the robot can maintain a proper distance from the obstacles and reach the target point smoothly. Secondly, in order to solve the local minimum problem, the local minimum discriminant condition is introduced, and the local minimum region is circum- navigated when the condition is triggered, so that the robot can reach the target point smoothly. The simulation results show that the improved algorithm has strong robustness when operating in the map environment with different number of obstacles. The proposed algorithm can make the robot bypass the local minimum area in the U-shaped obstacle environment, and successfully solve the local minimum problem in the mobile robot path planning.

The improvement of consensus mechanism is a key research content in the development process of blockchain technology. In the traditional consensus mechanism, all endorsement nodes participate in endorsement, which consumes a lot of time, and has the possibility of forging and manipulating the consensus process with low security. Based on the verifiable random function, the endorsement nodes in the candidate set of endorsement nodes for trading and any endorsement node for endorsement operation are randomly selected which can effectively improve the processing efficiency and reduce the processing time of the consensus mechanism. Based on theoretical analysis and experimental verification of Hyperledger fabric model, the results show that the optimized consensus mechanism has faster transaction processing speed, lower delay time and higher security.

In order to better protect the ecological environment and increase the economic value of recyclable waste, to solve the problems faced by the existing garbage identification methods, such as the complex classification background and the variety of garbage target forms, a residual semantic enhancement network for garbage image classification is proposed, which can strip foreground semantic objects from complex backgrounds. Based on the backbone residual network, the network uses visual concept sampling, inference and modulation modules to achieve visual semantic extraction, and eliminates the gap between semantic level and spatial resolution and visual concept features through the attention module, so as to be more robust to the morphological changes of garbage targets. Through experiments on the Kaggle open source 12 classified garbage dataset and TrashNet dataset, the results show that compared with the backbone network ResNeXt-50 and some other deep networks, the proposed algorithms have improved performance and have good performance in garbage image classification. 

The connection between mobile internet devices based on the Android platform and people’s lives is becoming increasingly close, and the security issues of mobile devices have become a major research hotspot. Currently, many visual Android malware detection methods based on convolutional neural networks have been proposed and have shown good performance. In order to better utilize deep learning frameworks to prevent malicious software attacks on the Android platform, a new application visualization method is proposed, which to some extent compensates for the information loss problem caused by traditional sampling methods. In order to obtain more accurate software representation vectors, this study uses the Swin Transformer architecture instead of the traditional CNN(Convolutional Neural Network) architecture as the backbone network for feature extraction. The samples used in the research experiment are from the Drebin and CICCalDroid 2020 datasets. The research experimental results show that the proposed visualization method is superior to traditional visualization methods, and the detection system can achieve an accuracy of 97. 39% , with a high ability to identify malicious software.

 In the existing researches on closed frequent item set mining algorithms, pruning strategies are relatively single, most of which are for 1item set pruning, and there are relatively few pruning strategies for 2item set and nitem set (n逸3). However, effective pruning strategies can find and cut off a large number of hopeless item sets in advance. Therefore, improving the pruning strategy of closed frequent item set is of great help to improve the efficiency of this kind of algorithm. On the basis of ESCS(Estimated Support Cooccurrence Structure) structure, an ESCS pruning strategy for 2itemsets is proposed, and the classical closed frequent itemset mining algorithm DCI_Closed(Direct Count Intersect Closed) is improved to DCI_ESCS(Direct Count Intersect Estimated Support Cooccurrence Structure) algorithm, and the effect of ESCS pruning strategy is verified. On multiple public datasets and under different minimum support thresholds, experiments are conducted to compare the time performance of the algorithm before and after the improvement. The experimental results show that the improved DCI_ESCS algorithm performs well on long and dense data sets with long transaction and itemsets, and the time efficiency is improved to a certain extent.

The rock image recognition algorithm model based on traditional deep learning is cumbersome and requires certain computing power when it is applied to mobile terminals, so it is difficult to realize real-time and accurate identification of rock types. Based on the ShuffleNetV2 network, we insert the ECA (Efficient Channel Attention) module of the channel connection attention mechanism, use the Mish activation function to replace the ReLU activation function, and introduce the depthwise separable convolution in the lightweight network components. Experiments are performed on rock images with this method. Experiments show that the recognition accuracy of the algorithm reaches 94. 74% . The improved algorithm structure is not complex and maintains the characteristics of lightweight, which lays a foundation for its application in limited resource environments such as mobile terminals.

In the power Internet of Things, user-side data is in a relatively isolated position, which makes it more difficult to mine data association rules. Therefore, a deep mining method of user-side data in the power Internet of Things based on association rule mapping is proposed. Based on the directed graph structure of user-side data mesh topology, the association mapping relationship among data sets is analyzed according to the association attribute group. And the association rules among data sets are mined using the correlation matrix. Extreme value standardization strategy and radial basis function neural network are introduced, and the dimensionless method and discrete clustering method are built. Through the hidden layer neural network is obtained. According to the K-means clustering process, data preprocessing, data types according to the different users of dominant and recessive side matrix, score matrix and users-project scale, deep data mining is realized. Experimental results show that this method can complete the mining task in a relatively short time, the processing effect of different data sets is better, and the data depth mining can be completed in a small memory space.

A network intrusion detection algorithm that combines systematic data pre-processing and hybrid sampling is proposed for the problem of class imbalance in intrusion detection datasets. Based on the feature distribution of the intrusion detection dataset, the feature values are systematically processed as follows: for the three categorical features, “Proto’’,“Service’’ and “State’’, minor categories within each feature are combined to reduce the total dimension of one-hot encoding; the 18 extremely distributed numerical features are processed with logarithm and then standardized according to the numerical distribution. The class imbalance processing technology, which combines Nearmiss-1 under-sampling and SMOTE ( Synthetic Minority Over-sampling Technique) is designed. Each class of samples in the training dataset is divided into sub-classes based on the “Proto’’,“ Service’’ and “ State’’ categorical features, and each sub-class is under-sampled or oversampled in equal proportion. The intrusion detection model PSSNS-RF ( Nearmiss and SMOTE based on Proto, Service, State-Random Forest) is built, which achieves a 97. 02% multiclass detection rate in the UNSW-NB15 dataset, resolving the data imbalance problem and significantly improving the detection rate of minority classes.

In the deep Reinforcement Learning, the PPO ( Proximal Policy Optimization) performs very well in many experimental tasks. However, KL(Kullback-Leibler) -PPO with adaptive KL divergence affects the update efficiency of KL-PPO strategy because of its asymmetry. In order to solve the negative impact of this asymmetry, Proximal Policy Optimization algorithm based on CIM( Correntropy Induced Metric) is proposed characterize the difference between the old and new strategies, update the policies more accurately, and then the experimental test of OpenAI gym shows that compared with the mainstream near end strategy optimization algorithms clip PPO and KL PPO, the proposed algorithm can obtain more than 50% reward, and the convergence speed is accelerated by about 500 ~ 1 100 episodes in different environments. And it also has good robustness.

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.

The frequency of highway congestion is increasing. In order to provide convenient travel path for drivers and slow down road traffic congestion, according to traffic statistics a highway traffic congestion prediction model based on improved genetic algorithm is designed. The fixed and mobile detection technology is used to collect macro traffic flow data such as flow, density and speed. Different identification and processing methods are adopted for abnormal parameters such as redundant data, missing data and error data to obtain effective and complete traffic flow data. The back propagation neural network and support vector machine regression network are used to improve the genetic algorithm. Two sub prediction models are established, and a hybrid prediction model is constructed by weighting the weights of the two models. According to the congestion prediction deviation of the sub prediction model, the weight coefficient of the hybrid prediction model is modified combined with the optimal weight combination strategy. The experimental results show that the design model can divide the traffic congestion level of the target expressway, and predict the congestion status according to the data of traffic flow, speed and occupancy rate, and the model has high prediction accuracy and ideal prediction effectiveness and accuracy.

In order to improve the recognition performance and accuracy of traditional object detection and to accelerate the computation speed, a CNN( Convolutional Neural Network) model with more powerful feature learning and representation capabilities and with related deep learning training algorithms is adopted and applied to large-scale recognition tasks in the field of computer vision. The characteristics of traditional object detection algorithms, such as the V-J(Viola-Jones) detector, HOG(Histogram of Oriented Gradients) features combined with SVM( Support Vector Machine) classifier, and DPM ( Deformable Parts Model) detector are analyzed. Subsequently, the deep learning algorithms that emerged after 2013, such as the RCNN ( Region-based Convolutional Neural Networks) algorithm and YOLO(You Only Look Once) algorithm are introduced, and their application status in object detection tasks is analyzed. To detect fallen individuals, the YOLOv5(You Only Look Once version 5) model is used to train the behavior of individuals with different heights and body types. By using evaluation metrics such as IoU(Intersection over Union), Precision, Recall, and PR curves, the YOLOv5 model is analyzed and evaluated for its performance in detecting both standing and fallen activities. In addition, by pre- training and data augmentation, the number of training samples is increased, and the recognition accuracy of the network is improved. The experimental results show that the recognition rate of fallen individuals reaches 86% . The achievements of this study will be applied to the design of disaster detection and rescue robots, assisting in the identification and classification of injured individuals who have fallen, and improving the efficiency of disaster area rescue.

As one of the important components of the detection system, the performance of photoelectric detector is directly related to the quality of system data acquisition. In order not to affect the final detection result, it is essential to ensure the detector performance. The performance of high performance PtS₂ / MoTe₂ heterojunction infrared photodetector is studied. First, the materials, reagents and equipment are prepared to make PtS₂ / MoTe₂ heterojunction infrared photodetectors. The detector performance test environment, the four indicators of light response, detection rate, response time and photoconductivity gain are set up, and the detector performance is analyzed. The results show that the optical responsivity of PtS₂ / MoTe₂ heterojunction infrared photodetector is always above the 5 A/ W limit with the passage of test time. The detection rate of the detector is greater than 10 cm·Hz1 / 2 W ^-1 regardless of the infrared light reflected from any material. Whether the photocurrent is in the rising time or the falling time, its response time is always below the limit of 150 μs; The photoconductivity gain value has been kept above 80% . 

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.

In order to improve the torque ripple and flux ripple in the model predictive control system of PMSM (Permanent Magnet Synchronous Motor), a control system scheme is designed by learning the basic structure and control methods of PMSM. The scheme adjusts the duty cycle and voltage vector synchronously. The optimal expected voltage vector and action time at a certain sampling time are selected, and the optimal expected voltage vector and action time at the current sampling time are added to adjust the duty cycle coefficient of the sampling time. The feasibility and effectiveness of this method in improving the control performance of PMSM are verified by comparative analysis of the simulation model.

To avoid the gradient vanishing or exploding issue in the RNN(Recurrent Neural Network)-based sequential recommenders, a gated recurrent unit based sequential recommender DeepGRU is proposed which introduces the residual connection, layer normalization and feed forward neural network. The proposed algorithm is verified on three public datasets, and the experimental results show that DeepGRU has superior recommendation performance over several state-of-the-art sequential recommenders ( averagely improved by 8. 68% ) over all compared metrics. The ablation study verifies the effectiveness of the introduced residual connection, layer normalization and feedforward layer. It is empirically demonstrated that DeepGRU effectively alleviates the unstable training issue when dealing with long sequences. 

 FBG(Fiber Bragg Grating) sensing technology has achieved rapid development in scientific research and engineering applications, but it is rarely used in undergraduate experimental teaching. Currently, there are few devices available in the market that can be directly used for FBG sensing experimental teaching, and cutting- edge scientific research technology is disconnected from undergraduate experimental teaching. To address this situation, a teaching experiment device for temperature stress sensing based on FBG has been designed. The device consists of three parts: a fiber laser, a spectrometer, and upper computer control software. The fiber laser enables laser output of about 1 550 nm. The spectrometer measures the change of FBG center wavelength and collects data into the computer. The upper computer control software is used for graphic display and data storage. The experimental device has the advantages of simple operation, flexible assembly, good repeatability, and stability, and can be used for undergraduate experimental teaching. We introduce cutting-edge science and technology into undergraduate experimental teaching, promote the integration of scientific research and experimental teaching, and realize the synchronous improvement of scientific research and teaching levels. 

In order to optimize the CCHP(Combined Cooling Heating and Power) supply system reasonably, a system of efficient installations of saving is established. It improves system efficiency, and realizes the cascade utilization of energy. On the basis of existing research, a waste heat recovery system is built. It contains a generator set and refrigeration unit. According to the partial load characteristic of the system on the unit in the system modeling alone, the original particle swarm optimization algorithm is optimized, and an example is introduced to verify it. The results show that the system has a good effect on the optimization of the algorithm, and the improved algorithm ensures the stability and practicability of the system operation.

 Due to the influence of changes in lighting conditions and pedestrian height differences, there are large cross modal differences in surveillance video images at different times. In order to accurately identify pedestrians in cross modal images, a pedestrian recognition algorithm based on generalized transfer depth learning is proposed. The cross modal image is formed through Cyele GAN(Cycle Generative Adversarial Network), and the reference map is segmented using single object image processing to obtain candidate human body regions. The matching regions are searched in the matching map to obtain the disparity of human body regions, and the depth and perspective features of human body regions are extracted through the disparity. The attention mechanism and cross modal pedestrian recognition are combined to analyze the differences between the two types of images. The two subspaces are mapped to the same feature space. And the generalized migration depth learning algorithm is introduced to learn the loss function measurement, automatically screen the pedestrian features of the cross modal images, and finally complete pedestrian recognition through the modal fusion module to fuse the filtered features. The experimental results show that the proposed algorithm can quickly and accurately extract pedestrians from different modal images, and the recognition effect is good.