Abstract:Spare parts demand forecast is one of the important means of lean management and economic benefit promotion.Aiming at the problems of insufficient data utilization and low accuracy of traditional prediction model,a new method is proposed to forecast the demand of spare parts.Firstly,the maximum mean difference(MMD)method is used to filter the data of migrating source domains to reduce the possibility of information redundancy caused by the migration of multiple similar source domains.Then,the Attention mechanism and bidirectional ordered short and long-term memory network are integrated to construct the Atten-tion BI ON-LSTM prediction model.The past and future information can be fully applied through the Attention BI ON-LSTM model.And the weights of different feature vectors are calculated to prevent the feature containing important information from disappearing.Finally,by tensor migration,the tensor of source domain training is extracted to initialize the target domain model,and the target domain prediction results are obtained after fine tuning.Experimental results show that the proposed method is superior to the com-mon time series prediction model.

Abstract:The circular composite piezoelectric structure can convert mechanical energy into electrical energy,so as to supply energy for the intelligent sensing sensors of the running parts of rail vehicles.In this paper,the frequency response characteristics of circular plate composite piezoelectric structures are studied.Based on the COMSOL finite element software and ISIGHT optimization design platform,a simulation model of circular plate composite piezoelectric structures is constructed.The key factors affecting the frequency response characteristics of piezoelectric structures are obtained through DOE orthogonal tests.Based on the analysis re-sults,single parameter analysis and correlation analysis are carried out.Among them,the diameter of the metal substrate and the height of the mass block have a negative effect on the first three modal frequencies of the structure,and the diameter and thickness of the piezoelectric ceramics have a positive effect on the first three modal frequencies of the structure,which lays a certain founda-tion for the subsequent structural optimization design.

Abstract:The traditional crack detection methods are mainly manual and static detection,and the detection results are sub-jective and the data processing speed is slow.Aiming at the problems existing in the traditional detection method,a semantic seg-mentation based method is proposed to rapidly realize end-to-end crack segmentation.The residual structure is introduced to pre-vent network degradation caused by deepening network layers of UNet.On this basis,dilated convolution is used to solve the prob-lem that the feature map size is too small in the late stage of feature extraction,which leads to insufficient feature extraction.Based on the detailed information of the crack edge features,convolution block attention module(CBAM)is added to optimize the weight distribution of features and suppress irrelevant features in the target region.The results show that compared with several common seg-mentation models,the number of parameters is reduced by 11.77%on average,and the test time of CrackTree260 is reduced by 20.92%,which can be well applied to the actual engineering detection scenario.

Abstract:Millimeter wave radar gesture recognition,as a new type of contactless human-computer interaction(HCI),has a wide range of application prospects in the fields of smart driving,smart home,and physical games.Although the existing work has achieved excellent recognition accuracy in a laboratory environment,it is still limited in practical applications due to the intensive data collection effort required,the poor robustness of the system in the new environment,and the recognition scheme based on the offline form.To address the above issues,in this paper,range-doppler images of eight common hand gestures are collected by using the IWR1843 BOOST millimeter wave radar sensor,on the basis of which a gesture dataset is built.The problem of poor cross-envi-ronment recognition performance of the system is solved by using the proposed interference elimination algorithm and data enhance-ment algorithm.In addition,the online segmentation performance of the system is enhanced by the gesture sample segmentation al-gorithm based on the detection window and recognition window.The experimental results show that the system achieves 98.26%rec-ognition accuracy on the dataset with an average inference delay of 24 ms.Compared with the conventional sliding window-based gesture recognition system,it features low inference delay and high recognition accuracy.

Abstract:Aiming at the problem that it is difficult to effectively utilize the global and multi-level feature information in the ex-isting large-scale 3D point cloud semantic segmentation method,in order to make the model perceive more abundant feature infor-mation,this paper proposes a large-scale 3D point cloud semantic segmentation method based on the gate cycle unit.In the large-scale 3D point cloud semantic segmentation network,the gate recurrent unit(GRU)module is used for feature fusion,and the multi-level and global feature information is effectively used for semantic segmentation tasks.At the same time,the soft-max-based softpool pooling operation is used in the pooling layer.Feature fusion is performed on the global feature information.Ex-periments show that this method increases the mIoU of semantic segmentation by 1.0%and 0.5%on the S3DIS and SemanticKITTI datasets,respectively.

Abstract:As the mainstream technical method of modern weather forecasting,numerical weather forecasting is playing an in-creasingly important role.According to the using status and data characteristics of numerical forecast in meteorological business ap-plication,a numerical forecast data service system based on distributed storage environment is proposed.In order to solve the prob-lem of reading and writing massive data,a 4-layer system framework and a 3-layer functional model are designed,and the data for-mat is uniformly converted.In order to extract meteorological time series and spatial data conveniently,it designs and implements the technology of time sequence retrieval and space extraction.At the same time,it provides the meteorological service application platform with the interface of obtaining time sequence data according to any latitude and longitude and the interface of obtaining ele-ment data according to any latitude and longitude range.The test results show that the interface provided by this system can meet the requirements of high timeliness,high stability and high concurrency of large-scale meteorological data in the business application of Zhejiang meteorological big data cloud test platform.

Abstract:As the voltage provided by the vehicle fuel cell is lower than the bus voltage,DC/DC converter is required to boost the voltage.In order to improve the poor anti-interference and slow response of traditional DC/DC converters,a nonlinear distur-bance observer is designed based on the interleaved boost converter(IBC)to estimate the load disturbance and input voltage change in real time.Since the DC/DC boost converter is a nonlinear system,in order to avoid singularity,a nonsingular fast terminal sliding mode surface is established from the perspective of system energy,its convergence is proved by using Lyapunov criterion,and the hyperbolic tangent function is used to replace the symbolic function to reduce chattering.The simulation results show that compared with the nonsingular terminal sliding mode control(NTSMC),it reduces the steady-state error and has better anti-interference per-formance when the input voltage and output load change suddenly.

Abstract:With the increase of users'personalized needs,from the production mode of massive single varieties to the produc-tion mode characterized by multiple varieties and small batches,resulting in the production delay problem is more likely to occur,in order to solve the above problems,this paper proposes a multi-variety and small batch scheduling method based on improved ge-netic algorithm,with the goal of minimizing the maximum completion time and minimizing the delay.Firstly,the OS coding initial-ization method based on roulette probability and EDD strategy and the MS coding initialization method based on GLC strategy opti-mize the initial population.Then,an adaptive crossover and mutation probability is designed to optimize the cross-mutation process to improve the global search ability and local search ability of the algorithm.Finally,by testing the Brandimarte standard study,the feasibility and effectiveness of the algorithm are verified from the aspects of maximum completion time and total delay.

Abstract:Considering that temporal information is an important attribute in knowledge graph,researchers have proposed a va-riety of efficient schemes for storing and querying temporal RDF.However,the existing storage and query of temporal RDF fail to ful-ly utilize the temporal semantic relationship between triples,which can affect the semantic query performance of temporal RDF.Aiming at the above problems,a temporal RDF index structure based on interval index HINT is proposed.The index structure com-bines the hierarchical index of HINT and the coupled bitmap index of RDF,so that the index structure can contain the temporal se-mantic relationship between triples.On the basis of HINT and coupled bitmap index,SPARQL-oriented temporal operations are im-plemented,which can support the semantic queries of thirteen Allen temporal relations.The experimental results show that the se-mantic indexing of temporal RDF based on HINT can improve the query performance of temporal RDF.

Abstract:Gated recurrent unit(GRU)network has a weak ability to capture information in time series dimension.To solve this problem,a time series prediction model based on the integration of improved attention mechanism and GRU network is pro-posed.The model adopts the sequence to sequence(Seq2Seq)structure,and the improved attention mechanism uses distance corre-lation coefficient as the evaluation function to assign weight to the output of each historical moment of the coding layer,so as to adap-tively enhance the influence of key historical moments,thus strengthening the information capture ability of GRU network in the time series dimension.By using the gas load data of a certain area in Shanghai for prediction analysis,the experimental results show that the prediction effect of this model is better than other common models,which proves the feasibility of this model and provides a new idea for time series prediction based on GRU network.

Abstract:Aiming at the problems of low recognition rate in facial expression recognition and more interference information in feature extraction,a network model for facial expression recognition based on improved ResNet50 network is proposed.By embed-ding the coordinate attention mechanism module into the network model,the model can improve the extraction ability of the expres-sion strongly related feature information,reduce the problem of information overload,and improve the robustness and recognition accuracy of the model.The experiment uses the Adam optimizer and improves it,and combines the exponential decay learning rate to further improve the model training effect.The weighted cross-entropy loss function is used to deal with the problem that the accu-racy of model recognition decreases due to the small amount of data in the face dataset and the uneven distribution of categories.Through experimental verification on the CK+dataset and Fer2013 dataset,the accuracy of facial expression recognition of the im-proved network model reaches 98.77%and 73.51%respectively,which has certain advantages over some similar algorithms.

Abstract:Extracting valuable medical information from Chinese electronic medical records has become a popular research topic.The combination of BERT and neural networks has become the mainstream in the field of named entity recognition.Previous Chinese pre-training models have ignored two important features of Chinese characters,which are glyph and pinyin,they contain important grammatical and semantic information in language understanding.Therefore,ChineseBERT pre-training model is used,it integrates Chinese glyph and pinyin information into the model pre-training,and inputs the obtained word vectors into bidirection-al long short-term memory Network(BiLSTM)to obtain contextual features after adversarial training,and finally inputs conditional random field(CRF)decoding to get the final prediction result.The experimental results on the CCKS2019 dataset show that Chine-seBERT-BiLSTM-CRF model gets a F1 value of 84.96%,which can be applied to the task of Chinese electronic medical record named entity recognition.

Abstract:In order to improve the amount of road information acquired during remote driving and improve the safety of remote driving,the"Cloud Edge V2X"technical architecture based on vehicle to X,edge computing and cloud platform is proposed for re-mote driving of intelligent buses,that is,the driver realizes the interaction between people and vehicles by 5G network in the simu-lation cockpit,relying on conventional cameras,laser radar and ultrasonic sensors capture road sensing information,they can ob-tain real-time road event status prompts through the"Cloud-Bide-V2X"architecture to provide security for remote driving.On this basis,a simulation cockpit is built to operate and control vehicles on the remote road.The existing research results show that the pro-posed method does not affect the original remote driving function,while adding road event information,the performance of the re-mote driving system is not affected.

Abstract:Chinese medical question answer are made more challenging by their language and domain specificity.To better represent the Sentence meaning of medical question answer,this paper proposes a method for fusing char-word to enhance semantics with multi-attentive for medical question answering.Firstly,the pre-trained models BERT and WordBERT are used to extract the vector representations of the text at the word level and character level respectively,and the two are fused to obtain more complete se-mantic information of the sentence vector.Then,an attention mechanism is added to generate an answer representation containing information about the question,which is input to bidirectional gating recurrent unit to obtain the overall semantic features of the sen-tence.Finally,the multi-attention pooling module enables the interaction of relevance between questions and answers,and finds the best matching answer by calculating the similarity of question-answer pairs.Experimental notes on the cMedQA medical dataset,the method in this paper has an improved on ACC@1 compared to other deep learning-based methods,it is demonstrated that fusing vector representations of char-words with the addition of attention in front of the neural network can improve the performance of auto-matic question answer models.

Abstract:Traffic target detection technology is of great significance for realizing road traffic safety and automatic driving tech-nology.At present,the most advanced target detection algorithm is a kind of target detection algorithm,such as YOLOv5.However,YOLOv5 has some problems,such as large amount of computation and parameters,and high detection error rate.The solution pro-posed in this paper is to replace the C3 module in the neck of YOLOv5 with the C3Ghost module and replace the convolution module with the Ghost module to reduce the model parameters,and integrate the CBAM attention mechanism module into the C3 module of the backbone network to highlight the key information of the object,increase the network feature extraction ability,and reduce the false detection rate.The experimental results show that the improved algorithm improves the mAP by 1%,reduces the computational load FLOPs by 2.4 G,and reduces the parameter quantity by 1.3×106 on the basis of YOLOv5.The model is lightweight while ensur-ing the accuracy.

Abstract:Aiming at the problems of stacking and positioning of unloading targets,the target guidance system is designed ac-cording to the existing rectangular coordinate system of unloading robots.Based on the in-depth learning target detection model Mask R-CNN,an improvement of the fusion CBAM attention mechanism is proposed according to the feature pyramid FPN,which optimizes the target area and channel weight,and completes the target cargo identification based on the fusion features.The 3D in-formation of the input image is calculated by SGBM binocular stereo vision algorithm,and the 3D coordinates of the target are ex-tracted.The experimental results show that the average precision of the proposed method for cargo target recognition is 90.20%,and the maximum error of the depth direction positioning accuracy is not more than 15 mm.The method meets the guidance requirements of the unloading robot for stacking cargo.

Abstract:The existing robot social navigation methods have the problems of exploration capability needs more improvement and the risk assessment mechanism of pedestrian's behavior is incomplete.The both may cause occasional collision and robot naviga-tion policy being in local optimal value.To this end,this paper proposes a robot crowd environment navigation algorithm based on LSTM Optimization.Specifically,a danger factor is defined to optimize LSTM in deep reinforcement learning.Then,the risk assess-ment of pedestrian's behavior is more consistent with human social norms,which improves the safety of robot navigation.By combin-ing with SAC,a LSTM-SAC framework is constructed to improve the stochastic exploration capability and accelerate convergence of this algorithm.Finally,the algorithm is tested and compared with the latest algorithm in the simulation environment.

Abstract:Relationship extraction is a key part of knowledge graph construction,meanwhile,TCM information extraction has been a research focus in the field of natural language processing in TCM.In this paper,entity relationship extraction of Huangdi Nei-jing is studied by the Pipeline method based on RoBERTa-wwm-BiLSTM-CRF and the Joint algorithm based on TPLinker.After the comparison study,the TPLinker model is proved to have better performance in the TCM relationship extraction task with accura-cy,recall and F1 values of 94.54%,95.59%and 95.04%.This shows the superiority of TPLinker in the relationship extraction task,and it can also provide a reference for other TCM information extraction tasks and lay the foundation for the construction of TCM knowledge graphs.

Abstract:In order to deal with the COVID-19 that is still raging,many researchers analyze the gene sequence of COV-ID-19,and cluster analysis is an effective means of analysis.This paper analyzes the gene sequence data of COVID-19 with cluster-ing algorithm as the core,and uses a combination of stepwise dimension reduction and clustering to solve the problem that high-di-mensional gene sequence mutation data is not suitable for direct clustering.By using two common datasets to train several commonly used algorithms after dimension reduction and clustering combination,the best combination of PCA+UMAP dimension reduction and Birch clustering algorithm is selected.Finally,the mutation data of COVID-19 gene sequence is input into the algorithm model for classification.At the same time,further analysis also confirmed that the mutation frequency of S protein of COVID-19 is very high,and S protein is closely related to the high infectivity of the virus,which is consistent with the conclusions reached by relevant researchers.

Abstract:Adversarial examples are an important means to measure the robustness of deep learning models.In the field of Chi-nese text adversarial example generation,text adversarial examples have the problems of low coverage situations,weakness attack effect and low quality.To address the above problems,a character-level text adversarial example generation approach SCGA(Simi-lar Character Generation Adversarial Example)is proposed,SCAG constructs a phonetic-form similarity dictionary by calculating the similarity of pronunciation and shape.This method effectively locates important word positions,using similar characters to re-place the original examples,and finally performing an adversarial example attack under black-box.This paper verifies the method's effectiveness on several datasets compared with several methods.This method has high quality and can effectively mislead the classi-fication results of the model.