Computer and Modernization

Improved Kmeans Segmentation Algorithm for Brain Tumor Based on HMRF

MA Yu-juan, HAN Jian-ning, SHI Shao-jie, CAO Shang-bin, YANG Zhi-xiu

2023, 0(03): 1-5.

Asbtract ( 310 )

PDF (1147KB) ( 76 )

References | Related Articles | Metrics

In order to solve the problems of misidentification of brain tumor regions in MRI and the uncertainty in segmentation of tumor sites in brain MRI images， an improved Kmeans algorithm combined with hidden Markov random field （HMRF) model is proposed to achieve accurate segmentation of brain tumor images. Firstly， the Euclidean distance of Kmeans algorithm is replaced by Manhattan-Chebyshev distance and the improved Kmeans algorithm is used to estimate the initial parameters and initial segmentation of the image to be segmented. Then the spatial information of the image is obtained by HMRF theory and the clustering center is updated by combining with EM algorithm to obtain more accurate clustering center so as to improve the segmentation performance of the algorithm. The experimental results show that the proposed method has good performance effect of brain tumor segmentation， in which the average values of Dice coefficient and Jaccard coefficient reach 0.9289 and 0.8725， respectively.

Intelligent Foreign Body Detection System of Micro-gel Blood Group Test Card

LI Yu-qing, WEN Yong-jun, ZENG Xiao-wei, TANG Li-jun, ZHOU Qin-hua, ZHANG Zhi-gang

2023, 0(03): 6-10.

Asbtract ( 134 )

PDF (1409KB) ( 46 )

References | Related Articles | Metrics

The foreign matter in the micro-gel blood group test card seriously affects the clinical test results， so it must be strictly tested before it is distributed and boxed. At present， it mainly depends on manual lamp test， which has low detection efficiency and poor accuracy. In this paper， an intelligent detection system for micro-column gel foreign bodies is designed. In this system， the infrared LED light source with a specific wavelength band is designed to obtain the gel card image， and the improved Gaussian filtering algorithm and quadratic Gaussian difference algorithm for extracting the impurity characteristic information of the gel card image are studied and realized. The automatic control device for identification， tracking and rejection is designed， and the automatic separation of blood type test cards with foreign bodies is realized. Samples provided by Nevil Intelligent Technology Co.， Ltd. are used for testing， the detection accuracy of the proposed system is higher than 98%， and the minimum diameter of foreign bodies detected is 50 microns.

Algorithm of Multi-scale Dense Receptive Domain GAN lmage Dehazing

YIN Xiang-chen, CHEN Si-long, LI Zhen-kai, ZHANG Wen-jin, LI Gui-qing

2023, 0(03): 11-15.

Asbtract ( 181 )

PDF (3211KB) ( 53 )

References | Related Articles | Metrics

The image dehazing based on prior relies on the atmospheric scattering model， which is susceptible to incomplete defogging and color distortion. Based on deep learning， this paper proposes a multi-scale dense receptive domain GAN image dehazing algorithm. Firstly， a multi-scale learning generator network is constructed to extract local details and global information of images through three different scales for feature fusion. Then， receptive dense blocks are used to increase receptive fields and obtain rich context information， and the extracted feature maps are further refined in multiple receptive dense blocks. Then，a multi-scale GAN discriminator is used， which consists of two identical sub-discriminators D1 and D2， and the two sub-discriminators jointly guide the generator training. Finally， L1 loss， perception loss and adversarial loss are combined to design a multivariate loss function to converge the network. The proposed algorithm is evaluated subjectively and objectively on SOTS test sets. The experimental results show that the proposed algorithm achieves better results and effectively， which improves the phenomenon of incomplete dehazing.

Lightweight Object Detection Model for Underwater Sonar Images

FAN Xin-nan, CHEN Xin-yang, SHI Peng-fei, SUN Huan-ru, LU Liang, ZHOU Zhong-kai

2023, 0(03): 16-22.

Asbtract ( 262 )

PDF (3518KB) ( 76 )

References | Related Articles | Metrics

With the development of unmanned underwater detection technology， AUV with sonar detection has become the main method of underwater object detection. However， due to the complexity of the underwater environment and the limitation of the sonar imaging mode， the sonar image resolution is low. Therefore， the traditional morphology based on object detection method has the problems of low detection accuracy and poor real-time performance. When deep learning algorithms such as YOLO are directly applied to underwater sonar image target detection， they still face challenges such as few underwater samples and many model parameters. This paper proposes a lightweight object detection model for sonar image datasets. In view of the characteristics of low-resolution sonar image data and the real-time requirements of underwater AUV automatic detection， the YOLOv4 model is used as the main framework to carry out model tailoring， replace the optimized feature fusion module， target prediction K-means clustering and improve the loss function， etc.， and the constructed detection model is applied to sonar target detection. According to the experimental data， the mAP of the proposed model in this paper is 0.0659，0.0214，0.0402 and 0.1701 higher than that of SSD， YOLOv3， YOLOV3-DFPIN and YOLOV4-tiny respectively， Under the conditioms of the mAPs are only 0.0186 lower than that of YOLOv4， only 0.0093 lower than CenterNet， only 0.0074 lower than EfficientdetD0， however， FPS is more than twice as high as YOLOv4 and CenterNet， more than fifth as high as EfficientdetD0. At the same time， the proposed model in this paper has the advantages of both high precision and real time. The experimental results show that the proposed feature extraction network can greatly reduce the redundancy of network parameters and improve the model efficiency and detection speed. Combined with the adaptive spatial feature fusion module， the mutual fusion and reuse of features in different scales are enhanced， and the accuracy of low resolution sonar image target detection is improved.

Traffic Sign Recognition Based on Image Enhancement and SKNet

LIAO Cong , GUO Huang, ZHAO Mao-jun, WANG Yu-song, BAI Jun-feng

2023, 0(03): 23-28.

Asbtract ( 177 )

PDF (4724KB) ( 55 )

References | Related Articles | Metrics

Aiming at the problems that the existing traffic sign recognition systems extract image feature insufficiently and are difficult to recognize under complex situations， a traffic sign recognition model HE-SKNet is designed based on image enhancement and SKNet. Firstly， the histogram equalization is used to enhance the images of traffic signs with too bright or too dark. Then the SKNet network that adaptively adjusts the size of the receptive field is used for feature extraction and classification. The experimental results on the GTSRB dataset show that the recognition accuracy of the proposed HE-SKNet model reaches 98.95%， which enjoys 2.77 percentage points higher than that of ResNet， ResNeXt， SENet and SKNet on average. It verifies that the HE-SKNet model adaptively extracts different scales of feature and is more suitable for complex practical application scenarios with too bright or too dark.

CTR Prediction Model Combining Attention Mechanism and Graph Neural Network

XIA Yi-chun, LI Wang-gen, LI Dou-dou, GE Ying-kui, WANG Zhi-ge

2023, 0(03): 29-37.

Asbtract ( 210 )

PDF (3742KB) ( 70 )

References | Related Articles | Metrics

Most CTR prediction algorithms initialize the feature embedding as a fixed dimension， ignoring the low popularity of the long tail feature. Setting it to the same length as the head object embedding vector will lead to unbalanced model training and affect the final recommendation results. Based on this， this paper first uses an end-to-end differentiable framework， which can automatically select different embedded dimensions according to the popularity of features. Secondly， this paper introduces squeeze excitation network mechanism and multi-head self-attention mechanism with residual connection to dynamically learn the importance of features and identify important feature combinations from different angles， and then uses graph neural network to explicitly model the second-order feature interaction instead of traditional inner product and Hadamard product. Finally， in order to further improve the performance， this paper combines the DNN component with the shallow model to form the depth model， uses the Bayesian optimization algorithm to select a set of super parameters for the depth model to avoid the complex parameter adjustment process， and the experimental results on two benchmark datasets verify the effectiveness of the model.

SVR-based Image Watermarking Algorithm with Adaptive Embedding Intensity

XU Kang-jian, GONG Hong-ning, LI Can, TONG De-yu

2023, 0(03): 38-42.

Asbtract ( 190 )

PDF (1288KB) ( 56 )

References | Related Articles | Metrics

Watermark embedding intensity is a key issue in data watermarking research and closely related to the robustness and imperceptibility of watermarking. This paper proposes a watermarking algorithm of adaptive embedding intensity based on SVR model for image data. In terms of image features， the LBP operator of the image and the low frequency coefficients of the wavelet transform are selected. The appropriate watermark embedding intensity is determined according to the SSIM index. Hence， the training dataset is generated and the SVR model is trained. In the process of watermark embedding， the watermark is encrypted first， then the DCT transform is applied to the image sub-blocks and the trained SVR model is used in watermark embedding to adjust the embedding intensity. Similarly， the extraction of the watermark can be performed in an inverse procedure based on the key and SVR model. The experimental results show that the proposed watermarking algorithm can perform adaptive watermark embedding and extraction on images while ensuring the image quality and watermark invisibility， and also has good robustness to adding noise， compression， image clipping and other types of attack.

Time-optimal Obstacle Avoidance Path Planning for UAV Inspection

TIAN Xiao-zhuang, LI Song, FU Guo-ping, TAN Qi-yun, SHAN De-shuai, WANG Wei-guang, WANG Zhu

2023, 0(03): 43-47.

Asbtract ( 149 )

PDF (2321KB) ( 52 )

References | Related Articles | Metrics

For real-world missions by quadrotors， complete time is an important index to measure the mission effectiveness. However， the path planning methods usually aim at the minimum path length， which can not reflect the task time accurately and directly. Thus， based on the approximate computation method of path time costs， a minimum-time obstacle-avoidance path planning algorithm is proposed for unmanned aerial vehicles. Through path tracking tests， the flight time of different maneuver styles is acquired and is used as the basis for cost calculation in the path search process. By tailoring the node expansion mode and cost-function computation method， a modified A* algorithm is proposed to achieve obstacle avoidance and path-time optimization. The simulations in radon scenarios are conducted to test the effectiveness and prior performances of the proposed method. Finally， taking substation inspection as an example， it verifies that the improved path planning method can guide the quadrotor to complete the task within a less time.

Micro-expression Recognition Based on AU-GCN and Attention Mechanism

ZHAO Jing-hua, YANG Qiu-xiang

2023, 0(03): 48-53.

Asbtract ( 288 )

PDF (1458KB) ( 72 )

References | Related Articles | Metrics

As a kind of expression with very short duration， micro-expression can implicitly express people ’s true feelings of trying to suppress and hide， which has a good application in national security， judicial system， medical category and political elections. However， since micro-expression has the characteristics of less data sets， short duration and low expression amplitude， there are many difficulties in identifying micro-expressions， such as less data samples， larger calculation， lack of attention to key features， and easy to over-fitting. Therefore， this paper uses facial action units （ AU ） to highlight local features by weighted attention mechanism， and applies graph convolution network to find the dependencies between AU nodes， and aggregates them into global features for micro-expression recognition. The experimental results show that compared with the existing methods， the proposed method improves the accuracy to 79.3 %.

Computer Aided Diagnostic on Microscopic Images Based on Deep Learning

WANG Yan, YANG Feng-wei, ZHAI Xing, WANG Li, TANG Yan, LIU Zhe, HAN Ai-qing

2023, 0(03): 54-59.

Asbtract ( 151 )

PDF (4933KB) ( 61 )

References | Related Articles | Metrics

According to a report released by the World Health Organization， the global incidence of malaria and tuberculosis remains high. Manual microscopy of thick and thin slides and sputum smears are an important method for the diagnosis of malaria and tuberculosis， and one of the disadvantages of this approach is that it is highly dependent on medical inspectors， prone to subjective misjudgment. The lack of highly skilled laboratory personnel in remote areas of low-income and developing countries， coupled with the variable shape， small size and uncertainty of plasmodium and mycobacterium tuberculosis in microscopic images and factors of some cell body uncertainty， resulted in difficult detection of plasmodium and mycobacterium tuberculosis. This paper proposes an improved faster R-CNN-based algorithm for automatic screening of plasmodium and mycobacterium tuberculosis from microscopic images. Firstly， the proposed algorithm addes a convolutional filter layer to the original Faster R-CNN framework and uses a deep residual network to extract features to improve the detection performance of the model. Then， this paper evaluates the performance of the improved model on two different microscopy tasks： AP value reached 94.55% on the thick-blood smear malaria micrograph image dataset， and 97.96% on the sputum smear tuberculosis microscopic image dataset. Compared with the original Faster R-CNN model， the improvement is 7.40 percentage points and 8.04 percentage points. The results show that the modified Faster R-CNN model can detect plasmodium and mycobacterium tuberculosis sites from images captured on a microscope eyepiece on a smartphone， reducing the dependence on manual microscopy and assisting researchers in diagnosis， which shows that the model is suitable for deployment in under-resourced areas.

Review of Electricity Theft Detection in Smart Grid Environment

ZHANG Yun, BAI Kai-feng, WANG Xing, CANG Tian, ZHOU Tong, DUAN Jin-wen, SU Han

2023, 0(03): 60-65.

Asbtract ( 202 )

PDF (1073KB) ( 58 )

References | Related Articles | Metrics

The non-technical loss （NTL） on the power consumption side caused by malicious power theft by users has always been one of the problems that power companies around the world expect to solve. With the rapid development of artificial intelligence algorithms and the popularization of smart meters， modeling and detection of electricity theft will effectively reduce the occurrence of such situations. Firstly， this article introduces the methods of collecting， processing， and sampling electricity consumption behavior data. Secondly， it analyzes and compares the characteristics of various algorithms and summarizes existing work on outlier detection， machine learning methods， and deep learning methods for mining abnormal electricity behavior. Finally， by discussing the problems of intelligent methods in the research of electricity theft detection and future research works， it provides some reference for researchers in this field.

Association Rule Mining of Undergraduate Physical Test Items Based on Apriori Algorithm

WANG Shao-hua, OUYANG Hui-dan, SUN Dan, WANG Kang, WU Hong-ping, ZHONG Xun, CHU Xing-ping, YANG Song-tao

2023, 0(03): 66-70.

Asbtract ( 230 )

PDF (954KB) ( 56 )

References | Related Articles | Metrics

Physical quality of college students is the comprehensive performance of strength， speed， endurance， flexibility and agility. In order to measure the physical quality of college students and analyze the improvement of their physical quality， the various indexes of male students （pull-up， 50 m， 1000 m， sitting forward and long jump） and the various indexs of female students （one-minute sit-up， 50 m， 800 m， sitting forward and long jump） are measured. In this paper， Apriori algorithm is used to carry out three groups of experiments respectively，namely under the premise of support of 50% and confidence of 70%， support of 60% and confidence of 70%， support of 70% and confidence of 70%， association rule mining is carried out on various indicators of male and female students in a certain university in recent five years. The experimental results show that the confidence of normal weight students to pass each body test is more than 70%， and the highest confidence of all related items is 87.7% for passing vital capacity and normal weight， while the confidence of abnormal weight students to pass each body test is less than 70%. There is no significant difference in confidence between height and body measurements among all related items， which verifies that Apriori algorithm plays an important role in the mining of association rules of college students’ physical fitness. Using the mined frequent item set， it can well assist universities to improve the physical fitness of college students.

Classification Algorithm for Goods Names Based on Enhanced Semantic Model

LI Xiao-feng, MA Jing, ZHOU Yan

2023, 0(03): 71-78.

Asbtract ( 80 )

PDF (4176KB) ( 46 )

References | Related Articles | Metrics

Customs declaration is the process of the owner of the import and export goods to the customs . The process of customs declaration mainly includes： filling customs declaration， document inspection， cargo inspection， and others. This paper primarily focuses on the name of goods in the customs declaration depends on manual filling and the problems of high declaration cost， low efficiency， unstable accuracy， and other to be optimized， proposes to take the short texts of goods description at customs declaration as corpus， and extractes the word-frequency features and semantic features using the TF-IDF and the BERT models. According to the characteristics of the corpus， this paper innovatively enhances semantic features with word-frequency features. Secondly， the ViT model extractes image features and fuses them with text features under the cross-attention mechanism. Finally， the multi-grain cascade forest classifier realizes the classification of goods names and achieves the purpose of accurately obtaining goods names. The experimental results show that the precision is 0.92， the recall is 0.90， and the F1-score is 0.91， which fully demonstrates the rationality and superiority of the algorithm in solving this problem and helps solve the existing problems.

Review of Abnormal Service Data Detection Methods in Power Grid

BAI Kai-feng, ZHAO Hong-bin, ZHANG Yun, LI Yan, CUI Jing-an, LIU Qian-jin, YANG Hua, NI Na

2023, 0(03): 79-83.

Asbtract ( 247 )

PDF (1027KB) ( 64 )

References | Related Articles | Metrics

The concept of smart grid has promoted the development of grid intelligence and informatization. The amount of various types of business data generated by the power system has also increased exponentially， and the abnormal data therein has decisive influnce on the analysis of power data and the stability of grid operation to a large extent. This article classifies， aralyzes and summarizes the methods for detecting abnormal business data in the power grid. The methods for detecting abnormal business data based on traditional technology and artificial intelligence technology are introduced respectively， and the basic principles and characteristics of each method are analyzed and explained. The challenges and future development trends of abnormal business data detection in power grids are summarized and prospected. This article which provides a certain reference for follow-up research.

Semantic Loss Degree of Text Summarization Evaluation Method

JIN Du-liang, FAN Yong-sheng, ZHANG Qi

2023, 0(03): 84-89.

Asbtract ( 109 )

PDF (1492KB) ( 46 )

References | Related Articles | Metrics

In the current field of text summarization automatic generation， the traditional ROUGE evaluation method has been repeatedly found by researchers that the gap between its evaluation results and artificial evaluation results is too large， but the gap has not been numerical and cannot be measured. Based on this situation， this paper uses multiple public Chinese summary datasets of different types and lengths to measure the degree of semantic loss generated by ROUGE in the evaluation by defining the calculation method of semantic loss rate. At the same time， it comprehensively considers the influence of summary length and internal factors of datasets on the generation of summary evaluation， and the specific values of errors between ROUGE evaluation and artificial evaluation are visualized finally. The experimental results show that the ROUGE evaluation score is weakly correlated with the artificial evaluation score. ROUGE method has a certain degree of semantic loss for different length datasets， and the length of the summary and the original annotation error of the datasets will also have an important impact on the final evaluation score. The calculation method of semantic loss rate defined in this paper can provide a certain reference for better selection of datasets and evaluation methods， provide a direction of thinking for improving evaluation methods， and also provide certain a guidance and help for the effectiveness of the final objective evaluation model.

Person-post Matching Adj-LightGBM Algorithm Based on SMOTE and Bayesian Optimization

LIU Fu-qian, QIN Hua-ni, LAI Hui-hui

2023, 0(03): 90-95.

Asbtract ( 235 )

PDF (1436KB) ( 60 )

References | Related Articles | Metrics

COVID-19 has a significant impact on all walks of life during the last two years. The traditional recruitment tactics are difficult to put into practice. On the one hand， the recruitment gap is large， on the other hand， job seekers have nowhere to apply for a job. The emergence of online recruitment has brought some convenience to job seekers and recruitment units， but there are still issues such as low efficiency and unbalanced matching betheen person-post. How to execute job matching effectively and swiftly has become an urgent issue that need to be addressed. To solve this problem， a person-posts matching algorithm of Adj-LightGBM based on SMOTE and Bayesian optimization is proposed. Firstly， the post data set is preprocessed. Secondly， SMOTE algorithm is used to over sample the successfully matched samples with a positive-to-negative sample ratio of 1：3. Then， Bayesian optimization is used to find the optimal LightGBM model on the verification set. Finally， the model is tested and evaluated. The optimal Auc and F1-score of the model is 0.974 and 0.970. Compared with support vector machine， random forest and XGBoost algorithm， it is discovered that the proposed algorithm not only has higher accuracy in person-post matching prediction， but also has substantial benefits in model training efficiency.

Load Forecasting Based on Decomposition and Multi-component Ensemble Learning

ZHANG Zi-sen, XU Xiao-zhong

2023, 0(03): 96-101.

Asbtract ( 117 )

PDF (1928KB) ( 35 )

References | Related Articles | Metrics

The time series of energy load has a dual trend of periodic fluctuation and growth. By integrating multiple classification and regression trees（CART）， the existing gradient boosting tree（GBDT） can well fit the periodic fluctuation trend， but it is poor for the growth trend. The related researches decompose the load first， and then use the combined model for prediction. This paper studies the prediction using multi-component ensemble learning after decomposition. Firstly， the two trend components and residual components of load are decomposed， and the leaf node of CART is changed into three prediction models， so that it can predict the three components. At the same time， the CART loss function is optimized as the sum of the square errors of the prediction results of each component， so that it can consider the loss of the three component prediction models. Then， the prediction results are reconstructed based on gradient boosting， so that it can fit the dual trend of load in the way of multi-component ensemble learning. Finally， a load forecasting method based on decomposition and multi-component ensemble learning is proposed. In a regional power load forecasting experiment， compared with other forecasting methods， the error evaluation values of the proposed method are reduced. The results of experiments show that in the prediction of double trend load， the method proposed in this paper has better performance， and also provides an improvement for GBDT to predict other types of data.

Multi-time Scale Scheduling Simulation of Microgrid Considering Demand Response

HAN Mu-feng

2023, 0(03): 102-106.

Asbtract ( 80 )

PDF (1528KB) ( 33 )

References | Related Articles | Metrics

The existing scheduling methods of microgrid do not consider the generation characteristics and demand response side of the power supply， resulting in the power of the microgrid after dispatching cannot reach the ideal value. Therefore， a multi-time scale scheduling method of microgrid taking the demand response into account is proposed. In the analysis of the generation characteristics of different distributed power sources in microgrid and the characteristics of different loads on the demand side， the load is classified according to the characteristics of loads. According to load classification results， the overall problem of microgrid is converted into resident residents optimization problems and multiple time scale problems. Taking the minimum power cost and the minimum charge and discharge cycles as the target， constructs a multiple time scale micro power grid scheduling model. The improved firefly algorithm is used to solve the model to obtain the optimal scheduling scheme. The experimental results show that the proposed method can obtain the optimal scheduling scheme.

Visual SLAM Algorithm Based on Adaptive Fusion of Point and Edge Features

GE Hong-fei, LI Yi-ran

2023, 0(03): 107-112.

Asbtract ( 140 )

PDF (3034KB) ( 38 )

References | Related Articles | Metrics

Visual SLAM algorithm based on point features cannot estimate camera motion reliably， because it cannot extract point features sufficiently in weak texture environments. Edge features have richer environmental information than point features. However， they will affect the real-time performance of the system by introducing edge features directly. Therefore， this paper proposes a visual SLAM algorithm based on adaptive fusion of point and edge features. In the front end， a method based on grid method is proposed to evaluate the quality of point features， which is used to judge the texture of the external environment. In the back end， different visual constraints are constructed according to the external environment to optimize the camera pose. In addition， a distance transformation algorithm is introduced to construct the distance error function of edge features， which improves the speed of iterative optimization. This paper evaluates the visual SLAM algorithm on the most popular public datasets， and compares with the state-of-the-art methods. The experimental results show that the average positioning accuracy of the proposed algorithm is 22.3% higher than that of the state-of-the-art ORBSLAM algorithm in the weak texture environment， and it also achieves better positioning accuracy in the rich texture environment.

Dynamic Particle Swarm Optimization without Velocity Based on Opposition-based Learning and Elite Promotion

ZENG Yi-pu, DAI Yi-ru, CHEN Yu-tian

2023, 0(03): 113-120.

Asbtract ( 105 )

PDF (1209KB) ( 46 )

References | Related Articles | Metrics

To resolve the problem that particle swarm optimization algorithm has low search accuracy， slow convergence speed and easy to fall into local optimization when dealing with complex optimization problems， a dynamic multi population particle swarm without velocity based on opposition-based learning and elite promotion is proposed. Firstly， based on the particle position update mode without velocity term， this algorithm dynamically divides subpopulations and adopts different evolutionary strategies. Opposition-based learning is used to broaden the search range for subgroups， ensuring population diversity and avoid particles falling into local optimization too early. Then， in order to make full use of the information of excellent particles and improve the search accuracy， the elite promotion strategy is improved to optimize the individual historical optimal particles. The differential evolution algorithm is used to update the optimal particle of the population. Finally， the performance is tested by 22 test functions proposed by CEC2006. The experimental results show that the proposed algorithm has more excellent performance in search accuracy and stability compared with other algorithms. In addition， the proposed algorithm can effectively improve the convergence speed.

Fuel Consumption Prediction Method of Heavy Trucks with Different Accelerating Driving Behaviors Based on Shared-LSTM

WANG Yi-ting, XING Ben-bei, LI Bin, LIU Ge, ZHANG Xiang-yu

2023, 0(03): 121-126.

Asbtract ( 181 )

PDF (2032KB) ( 57 )

References | Related Articles | Metrics

Standardizing the driving behavior of heavy trucks can effectively reduce fuel consumption， which plays a positive role in energy conservation and emission reduction in the transportation industry and the national “Double Carbon” strategy. This paper collects CAN bus data of several heavy trucks for one month， and quantitatively analyzes the relationship between different accelerating behaviors （rapid acceleration， normal driving， rapid deceleration） and fuel consumption. To the problem of the low efficiency and poor accuracy of the existing fuel consumption prediction methods， this paper improves the LSTM model and proposes a shared weight LSTM model （Shared-LSTM）. Based on the collected vehicle CAN bus data， this paper compares and analyzes the fuel consumption prediction effects of Shared-LSTM， GRU and BP neural network models under the same vehicle type， the same road condition and multiple behaviors. The experimental results show that the prediction efficiency of the improved LSTM model is improved by more than 3% under different accelerating driving behaviors， and the prediction indexes in all aspects are better than other models. Taking the rapid acceleration driving behavior as an example， the Shared-LSTM model is reduced by more than 5% compared with GRU and BP neural network in terms of mean absolute error， mean square error and mean root error. Therefore， the Shared-LSTM model can be widely used to predict fuel consumption under a variety of driving behaviors.

Table of Content