Crowd counting has important applications in public safety and pandemic control. A robust and practical crowd counting system has to be capable of continuously learning with the newly incoming domain data in real-world scenarios instead of fitting one domain only. Off-the-shelf methods have some drawbacks when handling multiple domains: (1) the models will achieve limited performance (even drop dramatically) among old domains after training images from new domains due to the discrepancies in intrinsic data distributions from various domains, which is called catastrophic forgetting; (2) the well-trained model in a specific domain achieves imperfect performance among other unseen domains because of domain shift; (3) it leads to linearly increasing storage overhead, either mixing all the data for training or simply training dozens of separate models for different domains when new ones are available. To overcome these issues, we investigate a new crowd counting task in incremental domain training setting called lifelong crowd counting. Its goal is to alleviate catastrophic forgetting and improve the generalization ability using a single model updated by the incremental domains. Specifically, we propose a self-distillation learning framework as a benchmark (forget less, count better, or FLCB) for lifelong crowd counting, which helps the model leverage previous meaningful knowledge in a sustainable manner for better crowd counting to mitigate the forgetting when new data arrive. A new quantitative metric, normalized Backward Transfer (nBwT), is developed to evaluate the forgetting degree of the model in the lifelong learning process. Extensive experimental results demonstrate the superiority of our proposed benchmark in achieving a low catastrophic forgetting degree and strong generalization ability.
Defect inspection, also known as defect detection, is significant in mobile screen quality control. There are some challenging issues brought by the characteristics of screen defects, including the following: (1) the problem of interclass similarity and intraclass variation, (2) the difficulty in distinguishing low contrast, tiny-sized, or incomplete defects, and (3) the modeling of category dependencies for multi-label images. To solve these problems, a graph reasoning module, stacked on a classification module, is proposed to expand the feature dimension and improve low-quality image features by exploiting category-wise dependency, image-wise relations, and interactions between them. To further improve the classification performance, the classifier of the classification module is redesigned as a cosine similarity function. With the help of contrastive learning, the classification module can better initialize the category-wise graph of the reasoning module. Experiments on the mobile screen defect dataset show that our two-stage network achieves the following best performances: 97.7% accuracy and 97.3% F-measure. This proves that the proposed approach is effective in industrial applications.
A novel in-contact three-dimensional (3D) measuring device, called MultiCal, is proposed as a convenient, low-cost (less than US$5000), and robust facility for onsite kinematic calibration and online measurement of robot manipulator accuracy. The device has μm-level accuracy and can be easily embedded in robot cells. During the calibration procedure, the robot manipulator first moves automatically to multiple end-effector orientations with its tool center point (TCP) constrained on a fixed point by a 3D displacement measuring device (single point constraint), and the corresponding joint angles are recorded. Then, the measuring device is precisely mounted at different positions using a well-designed fixture, and the above measurement process is repeated to implement a multi-point constraint. The relative mounting positions are accurately measured and used as prior information to improve calibration accuracy and robustness. The results of theoretical analysis indicate that MultiCal reduces calibration accuracy by 10% to 20% in contrast to traditional non-contact 3D or six-dimensional (6D) measuring devices (such as laser trackers) when subject to the same level of artificial measurement noise. The results of a calibration experiment conducted on a Staubli TX90 robot show that MultiCal has only 7% to 14% lower calibration accuracy compared to a measuring arm with a laser scanner, and 21% to 30% lower time efficiency compared to a 6D binocular vision measuring system, yielding maximum and mean absolute position errors of 0.831 mm and 0.339 mm, respectively.
We propose a novel parameter value selection strategy for the Lü system to construct a chaotic robot to accomplish the complete coverage path planning (CCPP) task. The algorithm can meet the requirements of high randomness and coverage rate to perform specific types of missions. First, we roughly determine the value range of the parameter of the Lü system to meet the requirement of being a dissipative system. Second, we calculate the Lyapunov exponents to narrow the value range further. Next, we draw the phase planes of the system to approximately judge the topological distribution characteristics of its trajectories. Furthermore, we calculate the Pearson correlation coefficient of the variable for those good ones to judge its random characteristics. Finally, we construct a chaotic robot using variables with the determined parameter values and simulate and test the coverage rate to study the relationship between the coverage rate and the random characteristics of the variables. The above selection strategy gradually narrows the value range of the system parameter according to the randomness requirement of the coverage trajectory. Using the proposed strategy, proper variables can be chosen with a larger Lyapunov exponent to construct a chaotic robot with a higher coverage rate. Another chaotic system, the Lorenz system, is used to verify the feasibility and effectiveness of the designed strategy. The proposed strategy for enhancing the coverage rate of the mobile robot can improve the efficiency of accomplishing CCPP tasks under specific types of missions.
A multi-sensor-system cooperative scheduling method for multi-task collaboration is proposed in this paper. We studied the method for application in ground area detection and target tracking. The aim of sensor scheduling is to select the optimal sensors to complete the assigned combat tasks and obtain the best combat benefits. First, an area detection model was built, and the method of calculating the detection risk was proposed to quantify the detection benefits in scheduling. Then, combining the information on road constraints and the Doppler blind zone, a ground target tracking model was established, in which the posterior Carmér-Rao lower bound was applied to evaluate future tracking accuracy. Finally, an objective function was developed which considers the requirements of detection, tracking, and energy consumption control. By solving the objective function, the optimal sensor-scheduling scheme can be obtained. Simulation results showed that the proposed sensor-scheduling method can select suitable sensors to complete the required combat tasks, and provide good performance in terms of area detection, target tracking, and energy consumption control.
In this paper, physical layer security techniques are investigated for cooperative multi-input multi-output (C-MIMO), which operates as an underlaid cognitive radio system that coexists with a primary user (PU). The underlaid secrecy paradigm is enabled by improving the secrecy rate towards the C-MIMO receiver and reducing the interference towards the PU. Such a communication model is especially suitable for implementing Industrial Internet of Things (IIoT) systems in the unlicensed spectrum, which can trade off spectral efficiency and information secrecy. To this end, we propose an eigenspace-adaptive precoding (EAP) method and formulate the secrecy rate optimization problem, which is subject to both the single device power constraint and the interference power constraint. This precoder design is enabled by decomposing the original optimization problem into eigenspace selection and power allocation sub-problems. Herein, the eigenvectors are adaptively selected by the transmitter according to the channel conditions of the underlaid users and the PUs. In addition, a simplified EAP method is proposed for large-dimensional C-MIMO transmission, exploiting the additional spatial degree of freedom for a low-complexity secrecy precoder design. Numerical results show that by transmitting signal and artificial noise in the properly selected eigenspace, C-MIMO can eliminate the secrecy outage and outperforms the fixed eigenspace precoding methods. Moreover, the proposed simplified EAP method for the large-dimensional C-MIMO can significantly improve the secrecy rate.
Asymmetric massive multiple-input multiple-output (MIMO) systems have been proposed to reduce the burden of data processing and hardware cost in sixth-generation mobile networks (6G). However, in the asymmetric massive MIMO system, reciprocity between the uplink (UL) and downlink (DL) wireless channels is not valid. As a result, pilots are required to be sent by both the base station (BS) and user equipment (UE) to predict double-directional channels, which consumes more transmission and computational resources. In this paper we propose an ensemble-transfer-learning-based channel parameter prediction method for asymmetric massive MIMO systems. It can predict multiple DL channel parameters including path loss (PL), multipath number, delay spread (DS), and angular spread. Both the UL channel parameters and environment features are chosen to predict the DL parameters. Also, we propose a two-step feature selection algorithm based on the SHapley Additive exPlanations (SHAP) value and the minimum description length (MDL) criterion to reduce the computation complexity and negative impact on model accuracy caused by weakly correlated or uncorrelated features. In addition, the instance transfer method is introduced to support the prediction model in new propagation conditions, where it is difficult to collect enough training data in a short time. Simulation results show that the proposed method is more accurate than the back propagation neural network (BPNN) and the 3GPP TR 38.901 channel model. Additionally, the proposed instance-transfer-based method outperforms the method without transfer learning in predicting DL parameters when the beamwidth or the communication sector changes.
In this paper, two different n-order topological circuit networks are connected by diodes to establish a unified network model, which is a previously unexplored problem. The network model includes not only five resistive elements but also diode devices, so the network contains many different network types. This problem can be solved through three main steps: First, the network is simplified into two different equivalent circuit models. Second, the nonlinear difference equation model is established by applying Kirchhoff’s law. Finally, the two equations with similar structures are processed uniformly, and the general solutions of the nonlinear difference equations are obtained by using the transformation technique. As an example, several interesting specific results are deduced. Our study on the network model has significant value, as it can be applied to relevant interdisciplinary research.
Overcharging is an important safety issue in the charging process of electric vehicle power batteries, and can easily lead to accelerated battery aging and serious safety accidents. It is necessary to accurately predict the vehicle’s charging time to effectively prevent the battery from overcharging. Due to the complex structure of the battery pack and various charging modes, the traditional charging time prediction method often encounters modeling difficulties and low accuracy. In response to the above problems, data drivers and machine learning theories are applied. On the basis of fully considering the different electric vehicle battery management system (BMS) charging modes, a charging time prediction method with charging mode recognition is proposed. First, an intelligent algorithm based on dynamic weighted density peak clustering (DWDPC) and random forest fusion is proposed to classify vehicle charging modes. Then, on the basis of an improved simplified particle swarm optimization (ISPSO) algorithm, a high-performance charging time prediction method is constructed by fully integrating long short-term memory (LSTM) and a strong tracking filter. Finally, the data run by the actual engineering system are verified for the proposed charging time prediction algorithm. Experimental results show that the new method can effectively distinguish the charging modes of different vehicles, identify the charging characteristics of different electric vehicles, and achieve high prediction accuracy.
A compact input-reflectionless balanced bandpass filter (BPF) with flexible bandwidth (BW) using a three-line coupled structure (TLCS) is presented in this paper. For the differential mode (DM), the TLCS is applied to achieve the bandpass response; meanwhile, the input coupled-feed line of the TLCS is reused in the input absorption network. This design shows a good fusion of the absorptive and BPF sections, effectively reducing the circuit size, and the BWs of the two sections that can be controlled separately result in a flexibly controllable DM response BW of the proposed input-reflectionless balanced BPF. Detailed analyses of the ratio of the two-part BWs have been given for the first time, which is vital for the passband flatness and reflectionless feature. In the codesign of this work, the input-reflectionless DM bandpass response can be optimized easily, while wideband common mode (CM) noise absorption is achieved by the input absorption network. To verify the design method, a prototype with a compact layout (0.52λ×0.36λ) is designed and measured in the 0–7.0 GHz range. The DM center frequency (f0) is 2.45 GHz with a measured 3 dB fractional bandwidth of 31.4%. The simulation and measurement results with good agreement are presented, showing good performance, e.g., low insertion loss (0.43 dB), wide upper stopband for the DM bandpass response (over 20 dB rejection level up to 2.72f0), and wideband DM reflectionless and CM noise absorption (fractional absorption bandwidth of 285.7%).