An automatic garbage classification system was designed based on machine vision in order to improve the efficiency of front-end collection in garbage classification process. The hardware device of the garbage classification system was designed and manufactured, which mainly included two boxes, the recyclable box and the non-recyclable box. A method of garbage type recognition was proposed based on Inception v3 feature extraction network structure and migration learning aiming at the data lacking problem caused by small garbage data sets. The method was trained and tested on the constructed garbage data set. The test results show that the method can accurately identify garbage types with an average accuracy of 0.99. The trained model was deployed on the raspberry pi 3B+, and tested on the real garbage bin. When the whole system was running stably, the average time for the system to complete the classification of one garbage was 0.95 second. The experimental results show that the automatic garbage classification system can effectively identify the types of garbage and complete the classification and recycling of the garbage.
An improved spatio-temporal graph convolutional networks traffic prediction algorithm, named free-flow reachable matrix-based spatio-temporal graph convolutional networks (FAST-GCN), was proposed, in order to predict real-time traffic flows accurately and improve the sensing and prediction of citywide traffic situation. The characteristics of urban complex road network structure were expressed effectively by the graph convolutional neural network, and the spatio-temporal dependency in complex road networks was explored by introducing free-flow reachable matrices. Thus the accuracy of traffic situation prediction was improved. First, preprocess traffic speeds and sensors location data. Second, with the existing spatio-temporal graph convolutional networks, the graph convolution module based on free flow reachable matrix was integrated to effectively capture the unique spatial characteristics of the urban traffic road networks. Finally, the prediction results were generated through a fully connected output layer. The proposed model was evaluated on a real-world traffic dataset PeMS. The experimental results show that this model could capture physical characteristics of road network and spatio-temporal dependency, and outperform the baselines such as spatio-temporal graph convolutional networks (STGCN), and the prediction accuracy in 45 minutes was improved by up to 5.656%. In addition, compared with baselines, the proposed model can adapt to traffic flow prediction in large-scale road networks and has superior scalability.
A deep convolutional neural network model was proposed aiming at the problem that the current data and features determine the upper limit of the classification accuracy of the sleep staging model. The parallel convolutional neural network automatically learns the time-domain and frequency-domain features of the original signals in terms of model construction. The feature fusion neural network fuses multi-features through dilated convolution and residual connection. The classification neural network recognizes the sleep stages based on fused features. Synthetic minority oversampling technique (SMOTE) method was applied to enhance data in order to reduce the effect of classification imbalance on classification effect, and two-step training method was applied to optimize the model. The original single-channel electroencephalogram (Fpz-Cz channel) of the Sleep-EDF data set was used to evaluate the proposed model by the 20-fold cross-validate scheme. The overall accuracy and macro-averaging F1-score were 86.73% and 81.70% respectively. The proposed deep convolution neural network was an end-to-end deep learning model without any prior knowledge. The experimental results showed that the classification accuracy of the proposed model was better than traditional deep learning models.
A wafer map pattern recognition (WMPR) model was proposed based on transfer learning and deep forest, in order to identify the defect pattern of the wafer maps and to timely diagnose the source of the fault in the manufacturing process. Transfer learning was used to migrate the network weight parameters of the deep CNN DenseNet pre-trained on ImageNet to this model, and the classification layer of the model was redesigned, in order to solve the problems of difficulties of deep learning model training and imbalance in the number of defect types in wafer maps. Thus, the training time of the model was reduced and the feature extraction ability was improved. Deep forest model was introduced to identify the wafer defect pattern, based on the abstract features of the wafer maps extracted by DenseNet. The experimental results on an industrial case demonstrated that the average recognition rate was about 96.8%. This method can improve the recognition efficiency and its performance is better than those well-known CNNs and other typical classifiers.
A new vehicle motion trajectory prediction algorithm was proposed by using the attention mechanism based on the classic convolutional social long-short term memory (LSTM) trajectory prediction algorithm. Firstly, the lateral attention mechanism was introduced to assign different weights to neighboring vehicles. The features obtained from the historical trajectory of the vehicle via LSTM were taken as global features, and the trajectory features were extracted as local features through convolution pooling. The two features were fused as the overall neighbor feature information for trajectory prediction. Secondly, the Encoder-Decoder framework of traditional trajectory prediction was improved, and a vertical attention mechanism on historical position was introduced, so that each moment of prediction could use the historical information, which was most relevant to the current moment. The improved model was verified on the US101 and I80 datasets provided by NGSIM, and the results show that the proposed trajectory prediction algorithm can obtain more accurate future trajectories than other algorithms.
The real-time prediction model that can predict the onset of paroxysmal atrial fibrillation (PAF) 45 min in advance on the one minute electrocardiogram (ECG) segment with 8 Hz sampling frequency was proposed, for real-time and data-intensive application scenarios such as long-term ECG monitoring and intensive care units (ICU). The probabilistic symbolic pattern recognition method was used to extract the pattern transition features within one minute window of down sampled ECG sequence, reducing the calculation complexity of the model and the demand for storage space, so as to ensure the effect of real-time prediction. A hybrid model (CNN-LSTM) of the convolutional neural network (CNN) and the long short-term memory (LSTM) was proposed to extract local spatial features and time-dependent features implied in pattern transition features. An ensemble classifier based on CNN-LSTM was constructed to improve the generalization ability of the model. Spark Streaming technology was used to read, write and calculate ECG streaming data, and low latency communication between data and model was realized. The accuracy, sensitivity, and specificity of the proposed model were 91.26%, 82.21%, and 95.79% respectively. The average delay of model processing was 2 s, which can meet the real-time PAF prediction demand.
The user’s stable long-term preferences and dynamic instant interests were obtained by modeling on the user’s historical behavior records, and the user preferences were aggregated for personalized recommendation. Firstly, the users’ reviews on the items were extracted to represent the characteristics of the items. Secondly, users’ historical behavior records were used to represent their stable long-term preferences, and query data was used to model their instant interests. Third, the user’s final preferences were aggregated by assigning different weights to the long-term preferences and instant interests through the attention mechanism. Experiments on real data sets of Amazon were conducted to evaluate the performance of SeqRec model, and results show that it is superior to the current state-of-the-art sequential recommendation methods more than 10% in recall rate and percision ratio. Meanwhile, SeqRec model proves that the long-term preferences and instant interests of different users have different influences on their next purchases.
An improved single-shot-multibox-detector (SSD) algorithm was proposed. Referring to the feature pyramid networks (FPN) algorithm, the features of the Conv4-3 layer were merged with the features of Conv7 and Conv3-3 layers, and the number of default boxes at each location in merged feature map was increased. The squeeze-and-excitation networks (SENet) was added to the network structure; the feature channels of each layer were weighted, in order to enhance the useful feature weights and suppress the invalid feature weights. A series of enhancements were performed on the training data to enhance the generalization performance of the network. The experimental results show that the improved algorithm has a better performance on the VOC (07+12) dataset; the mean average precision (mAP) value of the improved algorithm is 80.4%, which is 2.7% higher than that of the original algorithm; the mAP value of the improved algorithm on COCO dataset (2017) is 42.5%, which is 2.3% higher than that of the original algorithm. Thus, the proposed algorithm can accurately detect the target with a size of at least 16×16 pixels.
There have been some research in different image background modeling methods to detect abandoned objects in highway scenes. However, traditional fixed background modeling methods easily generate foreground noises because of the environmental changes, and dynamic background modeling methods quickly integrate the motionless foreground abandoned objects into the background model. A dynamic background modeling method was proposed based on background separation Gaussian mixture model (BS-GMM) for detecting abandoned objects in highway to solve this problem. Background division method and model matching method were improved in traditional Gaussian mixture model. The weight attenuation of the Gaussian distribution models per pixel was utilized to dynamically model and update image background model. The background update frequency of the traditional Gaussian mixture model method was retained, and the stationary target was continuously detected by the method of background separation. The method can reduce the impact of environmental noises easily generated in the open environment of highway, and effectively detect the long-time motionless abandoned objects. The method can achieve the effect of real-time detection in terms of computing performance. The experimental results show that our BS-GMM method produces less foreground noises than other methods, and detects abandoned objects which are motionless for more than 20 seconds. BS-GMM method can be effectively applied to detect abandoned objects in highway.
A Poisson noise image denoising method based on Bayesian probability model was proposed. An image denoising model was constructed based on Bayesian maximum a posteriori probability model and with combination of Poisson probability distribution. Considering that Markov random fields cannot represent complex natural images effectively, a higher-order Markov fields of experts was introduced as a prior regular term of the model to represent the probability distribution of the image. The quadratic penalty function was used to optimize the denoising model and restore clear images. The proposed method was compared with other denoising algorithms; the denoising effect was evaluated objectively by using two evaluation indexes: peak signal-to-noise ratio and structural similarity. The experimental results show that, compared with the traditional denoising methods, the peak signal-to-noise ratio of this method increased by at least 0.18 dB, and the denoising performance is significantly better than that of other methods. Thus, the details of the image can be retained better by using this mothed.
Double-layer fusion for cooperative localization was used combined with lidar in car and roadside binocular camera in order to achieve high-precision localization aiming at the problem of large localization error of unmanned vehicles in unstructured scenes. The down layer was two parallel pose estimations. Switching dual map was achieved through short-term and long-term estimation of pose error based on the adaptive Monte Carlo localization, and the cumulative error of scan matching for lidar was corrected. Kalman filter based on probability data association was used to eliminate non-detected targets’ interference for roadside cameras, and tracking was achieved. The upper layer fused pose estimations of two down layer as a global fusion estimation, and the result feedback was used to achieve autoregulation. The vehicle experiments showed that the localization accuracy of double-layer fusion cooperative localization was 0.199 m, and the yaw angle accuracy was 2.179°. It was greatly improved compared with localization by lidar on car or tight fusion without feedback. The localization accuracy can reach 7.8 cm as the number of roadside cameras increases.
The direct pullout tests under different corrosion rates were conducted to evaluate the bond performance between the corroded stainless steel bar and the concrete. The influences of corrosion rate and surface crack width on the bond degradation were analyzed, and then predictive formulas of bond strength between the stainless steel bar and the concrete were given respectively based on the corrosion rate and the surface crack width. The differences of bond properties between ordinary and stainless steel reinforced concretes were compared. Test results show that the bond strength between corroded stainless steel bar and concrete increases firstly and then decreases as the increase of corrosion rate, which is similar to that of the ordinary reinforced concrete, however the critical corrosion rate of stainless steel affecting the bond strength is higher than that of ordinary reinforcement. Pitting corrosion has strong randomness, which leads to the discrete relationship between the corrosion crack width and corrosion rate. Compared with the model based on the corrosion rate, the predictive model based on the surface crack width has a better agreement with the test results, and is of better practicability. The degradation amount of bond strength of stainless steel reinforced concrete is smaller than that of ordinary reinforced concrete at a same corrosion degree. The existing bond degradation model of ordinary reinforced concrete can be used to descript the stainless steel reinforced concrete directly and has a certain safety stock.
Firstly, the research background and the significance of supernumerary robotic limbs (SRLs) were introduced. Secondly, the concept of SRLs was provided, and SRLs were divided into two parts, i.e., auxiliary operation SRLs and auxiliary support SRLs, according to their functions. Thirdly, the current research status and progress for different structure types and structural flexibility SRLs were outlined. Besides, the main research points about man-machine integration for the SRLs were analyzed in terms of lightweight design and security, robotic-human-environment interaction and cooperation and anti-interference capability. Finally, the prospect of the SRLs was summarized.
In motor electromagnetic noise analysis, the main focus is the radial electromagnetic force and other electromagnetic force components are ignored, which causes the analysis to be inaccurate. The electromagnetic simulation was conducted aiming at the problem to analyze the magnetic field and electromagnetic force distribution of 8-pole 24-slot permanent magnet synchronous drive motor. Radial electromagnetic force alone and radial, tangential and axial electromagnetic forces were respectively applied on the stator model of the motor in order to analyze the electromagnetic noise. The electromagnetic noise of the motor was calculated by the boundary element method, and the accuracy of the simulation analysis was verified by experiments. The parameterized structural model of the motor was established, and the influence of different stator slot widths and magnet fillet radii on radial, and tangential electromagnetic force and electromagnetic noise was analyzed. The results show that the tangential electromagnetic force has a certain effect on the electromagnetic noise, and the motor stator model loaded with radial, tangential and axial electromagnetic forces is more reliable. Reasonable reduction of stator slot width and magnet fillet radius can effectively reduce electromagnetic noise of the motor.
Helical milling tool wear monitoring methods based on traditional machine learning require complex feature extraction and rich experience, and different wear stages have the same misclassification cost. A new method based on the one-dimensional convolutional neural network (1D CNN) and cost-sensitive learning was introduced, aiming at the above problems and considering the characteristics of one dimension of current signals. Current signals are acquired through spindle, revolution shaft and feed shaft of robotic helical milling end-effector respectively as monitoring signals, and samples are divided using sliding window method to reduce network capacity and increase sample numbers and diversity at the same time. The cost matrix is introduced into the network loss function and the misclassification cost of severe wear stage is increased to make 1D CNN cost sensitive. The time domain current signals are input into 1D CNN directly, and the network can extract tool wear features automatically and unify feature extraction and classification of different wear stages together. Experiment results demonstrate that the tool wear monitoring accuracy of the proposed method is 99.29% and the recall of severe wear stage is 99.60% in the robotic helical milling system.
A punching-shear capacity model for concrete slabs reinforced with fiber-reinforced polymer (FRP) bars was proposed based on critical shear crack theory. The moment-curvature relationship of FRP concrete slabs was established by a sectional analysis method, in which the tensile stress-strain relationship of concrete was incorporated with the tension stiffening effect. The demanding curve of FRP concrete slabs was determined by a simplified deformation of concrete slabs, the moment-curvature relationship of FRP concrete slabs, and the equilibrium equations of slabs. The validity of the model was verified with experimental results collected from the literature. At the same time, the accuracy of existing punching-shear models of FRP concrete slabs was compared and evaluated. The parametric analysis results indicate that the nominal punching shear stress in slabs increases but the critical rotation angle of slabs decreases with the increase of FRP reinforcement ratio and effective depth. As the strength of concrete and the loading area increase, the nominal punching shear stress in slabs decreases but the critical rotation angle of slabs increases. The effect of slab thickness and reinforcing ratio on the durability of concrete slabs should be considered in the design of slabs.
The workflow task, coverage of wireless communication, medical wisdom scenario and moving path were modeled, respectively. According to the location and the velocity of the mobile terminal, the execution time and energy consumption model based on the moving path were constructed. Based on the wireless communication model of different edge servers, task deferred execution and task migration were introduced to guarantee the service continuity and execution time constraint. Then, considering the execution benefits of the cloud, the edge server and the mobile terminal from a global viewpoint, the priority segmentation algorithm and the task offloading optimization algorithm were proposed. Meanwhile, the genetic algorithm was used to find the optimal path and solve the energy consumption optimization problem with the constraint of response time. The experimental results show that the proposed algorithm lowered the mobile energy consumption by 19.8%, compared with the offloading algorithm without considering terminal mobility. Thus, the algorithm can effectively reduce the energy consumption of edge device with the constraint of response time.
A front multi-vehicle target tracking algorithm optimized by DeepSort was proposed in order to improve the awareness of autonomous vehicles to the surrounding environment. Gaussian YOLO v3 model was adopted as the front-end target detector, and training was based on DarkNet-53 backbone network. Gaussian YOLO v3-Vehicle, a detector specially designed for vehicles was obtained, which improved the vehicle detection accuracy by 3%. The augmented VeRi data set was proposed to conduct the re-recognition pre-training in order to overcome the shortcomings that the traditional pre-training model doesn't target vehicles. A new loss function combining the central loss function and the cross entropy loss function was proposed, which can make the target features extracted by the network become better in-class aggregation and inter-class resolution. Actual road videos in different environments were collected in the test part, and CLEAR MOT evaluation index was used for performance evaluation. Results showed a 1% increase in tracking accuracy and a 4% reduction in identity switching times compared with the benchmark DeepSort YOLO v3.
The objective function of phase time allocation based on vehicle demand was proposed, and the basic model for phase time allocation of intersections with pedestrian crossing constraints was established, considering the traffic demands of non-critical movement, overlap phase movement, interrupted phase movement and pedestrian. The optimization model of phase design scheme was established, aiming at the optimization problem of multiple optional signal phase design schemes and combined with the basic model for phase time allocation of intersections. The proposed model can realize the synchronous optimization of intersection phase design and signal timing. A multi-round phase time allocation model was presented, for the phase time allocation of multiple overlap phase movement, interrupted phase movement and non-critical movement. The phase time allocation process was given and a simultaneous optimization method of signal phase design and timing at intersection was proposed based on the proposed allocation model. Case analysis shows that the model can optimize signal phase structure and allocate phase time in multiple rounds. It can deal with the complex phase design conditions such as overlap phase, interrupted phase and repeated phase, and take into account the demand of pedestrian crossing, so as to better ensure the overall operation efficiency of intersections.